@oh-my-pi/pi-coding-agent 15.12.4 → 15.13.0

package/src/internal-urls/docs-index.generated.ts

@@ -12,18 +12,18 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"blob-artifact-architecture.md": "# Blob and artifact storage architecture\n\nThis document describes how coding-agent stores large/binary payloads outside session JSONL, how truncated tool output is persisted, and how internal URLs (`artifact://`, `agent://`) resolve back to stored data.\n\n## Why two storage systems exist\n\nThe runtime uses two different persistence mechanisms for different data shapes:\n\n- **Content-addressed blobs** (`blob:sha256:<hash>`): global storage used to externalize large image base64 payloads and provider image data URLs from persisted session entries.\n- **Session-scoped artifacts** (files under `<sessionFile-without-.jsonl>/`): per-session text files used for full tool outputs and subagent outputs.\n\nThey are intentionally separate:\n\n- blob storage optimizes deduplication and stable references by content hash,\n- artifact storage optimizes append-only session tooling and human/tool retrieval by local IDs.\n\n## Storage boundaries and on-disk layout\n\n### Blob store boundary (global)\n\n`SessionManager` constructs `BlobStore(getBlobsDir())`, so blob files live in a shared global blob directory, not in a session folder.\n\nBlob file naming:\n\n- file path: `<blobsDir>/<sha256-hex>`\n- canonical file has no extension; when an extension is supplied (image MIME type), a typed sidecar `<sha256-hex>.<ext>` is hardlinked (or copied) next to it so OS openers can type-detect\n- reference string stored in entries: `blob:sha256:<sha256-hex>`\n\nImplications:\n\n- same binary content across sessions resolves to the same hash/path,\n- writes are idempotent at the content level,\n- blobs can outlive any individual session file.\n\n## Artifact boundary (session-local)\n\n`ArtifactManager` derives artifact directory from session file path:\n\n- session file: `.../<timestamp>_<sessionId>.jsonl`\n- artifacts directory: `.../<timestamp>_<sessionId>/` (strip `.jsonl`)\n\nArtifact types share this directory:\n\n- truncated tool output files: `<numericId>.<toolType>.log` (for `artifact://`)\n- subagent output files: `<outputId>.md` (for `agent://`)\n- subagent session JSONL sidecars: `<outputId>.jsonl` when task execution receives an artifacts directory\n\nSubagents can adopt the parent `ArtifactManager`; in that case parent and subagent tree share one artifact directory and numeric artifact ID space.\n\n## ID and name allocation schemes\n\n### Blob IDs: content hash\n\n`BlobStore.put()` / `putSync()` computes SHA-256 over the bytes it is given and returns:\n\n- `hash`: hex digest,\n- `path`: `<blobsDir>/<hash>`,\n- `displayPath`: `<blobsDir>/<hash>.<ext>` when an extension was supplied, otherwise the canonical path,\n- `ref`: `blob:sha256:<hash>`.\n\nNo session-local counter is used.\n\n### Artifact IDs: session-local monotonic integer\n\n`ArtifactManager` scans existing `*.log` artifact files on first directory-backed allocation to find max existing numeric ID and sets `nextId = max + 1`.\n\nAllocation behavior:\n\n- file format: `{id}.{toolType}.log`\n- IDs are sequential strings (`\"0\"`, `\"1\"`, ...)\n- resume does not overwrite existing artifacts because scan happens before allocation\n- the directory is created lazily on first save/allocation\n\nIf the artifact directory is missing, scanning yields an empty list and allocation starts from `0`.\n\nNon-persistent sessions without an adopted manager can store `saveArtifact(...)` content in memory under numeric IDs, but `artifact://` resolution is file-backed through registered artifact directories.\n\n### Agent output IDs (`agent://`)\n\n`AgentOutputManager` allocates IDs for subagent outputs from the requested name, used verbatim the first time and suffixed (`-2`, `-3`, …) only when the same name repeats (e.g. `Anna`, `Anna-2`). Nested outputs are grouped under the parent prefix (e.g. `Parent.Child`). It scans existing `.md` files on initialization so a resumed session never reuses a name that would clobber a prior output.\n\n## Persistence dataflow\n\n### 1) Session entry persistence rewrite path\n\nBefore session entries are written (`#rewriteFile` / incremental persist), `SessionManager` calls `prepareEntryForPersistence()` / `prepareEntryForPersistenceSync()` through the truncation pipeline.\n\nKey behaviors:\n\n1. **Large string truncation**: oversized strings are cut and suffixed with `\"[Session persistence truncated large content]\"`; signature fields (`thinkingSignature`, `thoughtSignature`, `textSignature`) are cleared instead of truncated.\n2. **Transient field stripping**: `partialJson` and `jsonlEvents` are removed from persisted entries.\n3. **Image externalization to blobs**:\n   - image blocks in `content` arrays are externalized when `data` is not already a blob ref and base64 length is at least threshold (`BLOB_EXTERNALIZE_THRESHOLD = 1024`),\n   - provider-style `image_url` data URLs are externalized when they start with `data:image/` and contain `;base64,`,\n   - image block `data` is stored as decoded binary bytes,\n   - provider data URLs are stored as the original UTF-8 data URL string,\n   - persisted values are replaced with `blob:sha256:<hash>`.\n\nThis keeps session JSONL compact while preserving recoverability.\n\n### 2) Session load rehydration path\n\nWhen opening a session (`setSessionFile`), after migrations, `SessionManager` runs `resolveBlobRefsInEntries()`.\n\nFor message/custom-message image blocks with `blob:sha256:<hash>` and for persisted provider `image_url` fields with blob refs:\n\n- reads blob bytes from blob store,\n- converts image-block bytes back to base64,\n- converts provider `image_url` blobs back to the original string,\n- mutates in-memory entry fields for runtime consumers.\n\nIf a blob is missing:\n\n- image-block resolution logs a warning and keeps the original `blob:sha256:` ref string in memory,\n- provider `image_url` resolution logs a warning and keeps the original ref string,\n- load continues.\n\n### 3) Tool output spill/truncation path\n\n`OutputSink` powers streaming output in bash/python/ssh and related executors.\n\nBehavior:\n\n1. Every chunk is sanitized with `sanitizeWithOptionalSixelPassthrough(..., sanitizeText)` and appended to in-memory accounting.\n2. Optional live `onChunk` receives sanitized pre-column-cap chunks, throttled if configured.\n3. A per-line column cap can drop bytes from long lines in the LLM-facing buffer; when this happens, artifact mirroring starts so the on-disk file keeps the full sanitized stream.\n4. When the in-memory tail buffer would exceed spill threshold (`DEFAULT_MAX_BYTES`, 50KB), sink marks output truncated and starts artifact mirroring if an artifact path is available.\n5. If a file sink is opened, it first writes the current buffer, then all queued/subsequent sanitized chunks.\n6. In-memory buffer is trimmed to a tail window, or to head + elision marker + tail when head retention is configured.\n7. `dump()` returns summary including `artifactId` only when file sink creation succeeded.\n\nPractical effect:\n\n- UI/tool return shows bounded output,\n- full sanitized output is preserved in artifact file and referenced as `artifact://<id>` when file-backed artifact mirroring succeeded.\n\nIf file sink creation fails (I/O error, missing path, etc.), sink falls back to in-memory truncation only; full output is not persisted.\n\n## URL access model\n\n### `blob:` references\n\n`blob:sha256:<hash>` is a persistence reference inside session entry payloads, not an internal URL scheme handled by the router. Resolution is done by `SessionManager` during session load.\n\n### `artifact://<id>`\n\nHandled by `ArtifactProtocolHandler` over registered active session artifact directories:\n\n- requires a numeric ID,\n- searches each registered artifacts directory for filename prefix `<id>.`,\n- returns raw text (`text/plain`) from the matched `.log` file,\n- when missing, error includes available numeric artifact IDs from existing artifact files.\n\nFailure behavior:\n\n- if no artifact directories are registered: throws `No session - artifacts unavailable`,\n- if registered directories exist but none are present on disk: throws `No artifacts directory found`,\n- if ID is not numeric: throws `artifact:// ID must be numeric, got: <id>`.\n\n### `agent://<id>`\n\nHandled by `AgentProtocolHandler` over registered active session artifact directories and `<artifactsDir>/<id>.md`:\n\n- plain form returns markdown text,\n- `/path` or `?q=` forms perform JSON extraction,\n- path and query extraction cannot be combined,\n- if extraction requested, file content must parse as JSON.\n\nFailure behavior:\n\n- if no artifact directories are registered: throws `No session - agent outputs unavailable`,\n- if registered directories exist but none are present on disk: throws `No artifacts directory found`,\n- missing output throws `Not found: <id>` with available `.md` output IDs when directory listing succeeds.\n\nRead tool integration:\n\n- `read` supports offset/limit pagination for non-extraction internal URL reads,\n- rejects offset/limit when `agent://` extraction is used.\n\n## Resume, fork, and move semantics\n\n### Resume\n\n- `ArtifactManager` scans existing `{id}.*.log` files on first allocation and continues numbering.\n- `AgentOutputManager` scans existing `.md` output IDs and continues numbering.\n- `SessionManager` rehydrates blob refs to base64/data URLs on load.\n\n### Fork\n\n`SessionManager.fork()` creates a new session file with new session ID and `parentSession` link, then returns old/new file paths. Artifact copying is handled by `AgentSession.fork()`:\n\n- flushes current session first,\n- attempts recursive copy of old artifact directory to new artifact directory,\n- missing old directory is tolerated,\n- non-ENOENT copy errors are logged as warnings and fork still completes.\n\nID implications after fork:\n\n- if copy succeeded, artifact counters in the new session continue after max copied ID when the new `ArtifactManager` first scans,\n- if copy failed/skipped, new session artifact IDs start from `0`.\n\nBlob implications after fork:\n\n- blobs are global and content-addressed, so no blob directory copy is required.\n\n### Move to new cwd\n\n`SessionManager.moveTo()` renames both session file and artifact directory to the new default session directory, with rollback logic if a later step fails. This preserves artifact identity while relocating session scope.\n\n## Failure handling and fallback paths\n\n| Case                                                      | Behavior                                                             |\n| --------------------------------------------------------- | -------------------------------------------------------------------- |\n| Blob file missing during image-block rehydration          | Warn and keep `blob:sha256:` ref string in memory                    |\n| Blob file missing during provider `image_url` rehydration | Warn and keep `blob:sha256:` ref string in memory                    |\n| Blob read ENOENT via `BlobStore.get`                      | Returns `null`                                                       |\n| Artifact directory missing (`ArtifactManager.listFiles`)  | Returns empty list (allocation can start fresh)                      |\n| No registered artifact dirs (`artifact://`)               | Throws `No session - artifacts unavailable`                          |\n| No registered artifact dirs (`agent://`)                  | Throws `No session - agent outputs unavailable`                      |\n| Registered artifact dirs missing on disk                  | Throws explicit `No artifacts directory found`                       |\n| Artifact ID not found                                     | Throws with available IDs listing                                    |\n| OutputSink artifact writer init fails                     | Continues with bounded in-memory output only                         |\n| Non-persistent `saveArtifact`                             | Stores text in `SessionManager` memory map; not file-backed URL data |\n\n## Binary blob externalization vs text-output artifacts\n\n- **Blob externalization** is for image payloads inside persisted session entry content and provider image data URLs; it replaces inline payload strings in JSONL with stable content refs.\n- **Artifacts** are plain text files for execution output and subagent output; file-backed artifacts are addressable by session-local IDs through internal URLs.\n\nThe two systems intersect only indirectly: both reduce session JSONL bloat, but they have different identity, lifetime, and retrieval paths.\n\n## Implementation files\n\n- [`src/session/blob-store.ts`](../packages/coding-agent/src/session/blob-store.ts) — blob reference format, hashing, put/get, externalize/resolve helpers.\n- [`src/session/artifacts.ts`](../packages/coding-agent/src/session/artifacts.ts) — session artifact directory model and numeric artifact ID/path allocation.\n- [`src/session/streaming-output.ts`](../packages/coding-agent/src/session/streaming-output.ts) — `OutputSink` truncation/spill-to-file behavior and summary metadata.\n- [`src/session/session-manager.ts`](../packages/coding-agent/src/session/session-manager.ts) — persistence transforms, blob rehydration on load, session fork/move interactions.\n- [`src/session/agent-session.ts`](../packages/coding-agent/src/session/agent-session.ts) — artifact directory copy during interactive fork.\n- [`src/internal-urls/artifact-protocol.ts`](../packages/coding-agent/src/internal-urls/artifact-protocol.ts) — `artifact://` resolver.\n- [`src/internal-urls/agent-protocol.ts`](../packages/coding-agent/src/internal-urls/agent-protocol.ts) — `agent://` resolver + JSON extraction.\n- [`src/internal-urls/router.ts`](../packages/coding-agent/src/internal-urls/router.ts) — internal URL router wiring.\n- [`src/task/output-manager.ts`](../packages/coding-agent/src/task/output-manager.ts) — session-scoped agent output ID allocation for `agent://`.\n- [`src/task/executor.ts`](../packages/coding-agent/src/task/executor.ts) — subagent output artifact writes (`<id>.md`) and session JSONL sidecars.\n",
 	"collab.md": "# Collab: Live Session Sharing\n\n`/collab` shares your running session with other omp instances in real time. Guests render the **same session natively in their own TUI** — streaming assistant text, tool-call cards, footer state (cwd, model, context %, cost), ctrl+o expansion, `/dump` — no terminal mirroring. Guests can prompt and interrupt the agent; the host machine runs the agent and all tools.\n\n## Quick start\n\nHost:\n\n```\n/collab\n```\n\nprints\n\n```\nCollab session started!\n • Join from another terminal: omp join \"mgAYTZwEnpRQtca0CTgn-Q.gdJUbTovD94ofDaa8YvhY0-ty16w4fn8PgB6PLnoA30\"\n • or any web browser: my.omp.sh/#mgAYTZwEnpRQtca0CTgn-Q.gdJUbTovD94ofDaa8YvhY0-ty16w4fn8PgB6PLnoA30\n```\n\nThe browser line is click-to-join (an OSC 8 hyperlink to the full `https://` deep link): the relay serves the web guest client at `/`, and the room id + key ride in the URL fragment. From another omp (any directory, any machine), either form works:\n\n```\n/join my.omp.sh/#mgAYTZwEnpRQtca0CTgn-Q.gdJU…\n```\n\nThe guest's previous session is restored on `/leave` (or when the host stops).\n\n### Commands\n\n| Command | Effect |\n|---|---|\n| `/collab` | Start sharing (or re-print the link when already hosting) |\n| `/collab <relay>` | Start sharing through a specific relay (`relay.example.com`, `ws://localhost:7475`) |\n| `/collab view` | Print a read-only (view-only) link (starts sharing first if needed) |\n| `/collab status` | Show link + participants |\n| `/collab stop` | Stop sharing |\n| `/join <link>` | Join a shared session as a guest |\n| `/leave` | Leave (guest) or stop sharing (host) |\n\n## Link format\n\n```\nhttps://host[:port]/#<link>          → browser deep link (printed by /collab; /join accepts it too)\n<roomId>.<key>                       → default relay (my.omp.sh)\nhost[:port]/r/<roomId>.<key>         → custom relay, wss:// inferred\nws://localhost:7475/r/<roomId>.<key> → plain ws, allowed for localhost only\n```\n\nThe trailing `.<key>` part is the room secret, base64url-encoded, in one of two strengths:\n\n- **Full link** — 48 bytes: the 32-byte AES-256-GCM room key followed by a 16-byte write token. Grants prompting, interrupting, and subagent control.\n- **View-only link** — the bare 32-byte key, no write token. Grants live read access only. Pre-token links parse as view-only.\n\nThe room secret is dot-joined rather than `#`-joined: RFC 3986 forbids a raw `#` inside a URL fragment, so strict URL stacks (macOS Foundation behind terminal click-to-open) percent-encode a second `#` to `%23` and break the link. Parsers leniently accept the legacy `#` form and the mangled `%23` form. In the browser deep link, everything after the `#` — room id and key — is a URL fragment: it never appears in any HTTP request, and neither secret is ever sent to the relay.\n\n## End-to-end encryption\n\nEvery session payload (entries, events, state, prompts) is sealed with AES-256-GCM before it touches the socket. The relay sees only:\n\n- room ids and connection counts,\n- opaque ciphertext frames and their sizes,\n- a 4-byte routing prefix (which guest a frame targets).\n\nPossession of the link is the trust boundary: a full link reads and steers the session, a view-only link reads it. Share both like secrets.\n\n## Guest permission model\n\nTwo trust levels, enforced by the link itself — the host verifies the 16-byte write token at join and rejects writes from peers without it (they appear as read-only in the participants list, and the join notice says so).\n\nGuests with a full link can:\n\n- read the entire session (including the back-transcript at join time),\n- prompt the agent (rendered with their name badge on every participant's transcript; the LLM sees the prompt text verbatim — names are display-only),\n- interrupt the agent (Esc),\n- use the Agent Hub against the host's subagents: live table and progress, chat (steers the host's subagent), kill, revive, and transcript viewing (fetched from the host on demand).\n\nGuests with a view-only link can read everything live — back-transcript, streaming text, tool cards, subagent transcripts — but the host rejects prompting, interrupting, and agent control from them.\n\nEverything that mutates the host session or machine is host-only: `/model`, `/compact`, `/resume`, `/branch`, bash (`!`), python (`$`), skills, etc. Guests keep a small local allowlist (`/dump`, `/export`, `/copy`, `/help`, `/hotkeys`, `/theme`, `/settings`, `/leave`, `/collab`, `/exit`).\n\nKnown v1 limit for guests: a turn already streaming when you join becomes visible from its next message boundary.\n\n## Web client\n\n`packages/collab-web` is a standalone browser client for the same links — no omp install needed on the guest side. The relay serves it at `/`, which is what makes the `/collab` deep link click-to-join: `https://<relay>/#<link>` loads the client and auto-connects from the fragment. It renders the live transcript (streaming text, thinking, tool cards), a subagent panel with on-demand transcripts, and a composer with the same guest powers (prompt, interrupt, hub actions). Run `bun run dev` in the package for a local instance, `bun run mock-host` for an offline scripted host to develop against, and `bun run build` to emit a static `dist/` deployable anywhere (HTTPS required for WebCrypto). The client never talks to anything but the relay, and the key stays in the URL fragment.\n\n## Settings\n\n| Setting | Default | Meaning |\n|---|---|---|\n| `collab.relayUrl` | `wss://my.omp.sh` | Relay used by `/collab` when no relay is passed inline |\n| `collab.displayName` | OS username | Name shown to other participants |\n| `share.serverUrl` | `https://my.omp.sh/s` | Share viewer/upload base used by `/share` (links are `<base>/<id>#<key>`) |\n| `share.redactSecrets` | `true` | Run the secret obfuscator over `/share` snapshots before upload |\n\n## Self-hosting the relay\n\nThe relay is a small content-blind Go service. It keeps no state beyond live connections and exposes:\n\n- `GET /` — the static collab-web guest client (target of the `/collab` deep link),\n- `GET /r/<roomId>?role=host|guest` — WebSocket upgrade,\n- `POST /s` / `GET /s/<id>` / `GET /s/<id>/raw` — `/share` blob upload, viewer page, and blob fetch,\n- `GET /healthz` — liveness.\n\n\n## Architecture notes\n\nHub topology — the host is authoritative, guests never peer:\n\n1. `entry` frames — durable session entries, broadcast pre-blob-externalization so images stay inline (guests cannot resolve host blob refs). Guests append them verbatim (ids preserved) to a replica session file under `~/.omp/collab/<roomId>.jsonl` and into the agent's message array, which is why `/dump` and context estimates work.\n2. `event` frames — live agent events, fed straight into the guest's normal event controller; rendering is events-only to prevent double-render.\n3. `state` frames — debounced footer snapshots: streaming flag, the host's full model object and thinking level (applied to the guest's replica agent state, so model display and context-window math are native), host context numbers, and participants.\n4. `bus` frames — mirrored task-subagent lifecycle/progress EventBus traffic, republished on the guest's local bus so the subagent HUD and status-line count work natively.\n5. `agents` frames — agent-registry snapshots feeding a guest-local registry, so the Agent Hub table renders host subagents.\n\nGuest→host: `hello`, `prompt`, `abort`, `agent-cmd` (hub chat/kill/revive), and `fetch-transcript` (incremental subagent-transcript reads answered by targeted `transcript` frames). The replica loads through the regular `/resume` machinery, so theming, ctrl+o, and transcript behavior are native by construction; the guest process never chdirs to host paths.\n",
 	"compaction.md": "# Compaction and Branch Summaries\n\nCompaction and branch summaries are the two mechanisms that keep long sessions usable without losing prior work context.\n\n- **Compaction** rewrites old history into a summary on the current branch.\n- **Branch summary** captures abandoned branch context during `/tree` navigation.\n\nBoth are persisted as session entries and converted back into user-context messages when rebuilding LLM input.\n\n## Key implementation files\n\n- `packages/agent/src/compaction/compaction.ts` (context-full summarization and handoff generation)\n- `packages/snapcompact/src/snapcompact.ts` (snapcompact strategy: history archived as dense bitmap images)\n- `packages/agent/src/compaction/branch-summarization.ts`\n- `packages/agent/src/compaction/pruning.ts`\n- `packages/agent/src/compaction/utils.ts`\n- `packages/agent/src/compaction/openai.ts`\n- `packages/coding-agent/src/session/session-manager.ts`\n- `packages/coding-agent/src/session/agent-session.ts`\n- `packages/coding-agent/src/session/messages.ts`\n- `packages/coding-agent/src/extensibility/hooks/types.ts`\n- `packages/coding-agent/src/config/settings-schema.ts`\n\n## Session entry model\n\nCompaction and branch summaries are first-class session entries, not plain assistant/user messages.\n\n- `CompactionEntry`\n  - `type: \"compaction\"`\n  - `summary`, optional `shortSummary`\n  - `firstKeptEntryId` (compaction boundary)\n  - `tokensBefore`\n  - optional `details`, `preserveData`, `fromExtension`\n- `BranchSummaryEntry`\n  - `type: \"branch_summary\"`\n  - `fromId`, `summary`\n  - optional `details`, `fromExtension`\n\nWhen context is rebuilt (`buildSessionContext`):\n\n1. Latest compaction on the active path is converted to one `compactionSummary` message.\n2. Kept entries from `firstKeptEntryId` to the compaction point are re-included.\n3. Later entries on the path are appended.\n4. `branch_summary` entries are converted to `branchSummary` messages.\n5. `custom_message` entries are converted to `custom` messages.\n\nThose custom roles are then transformed into LLM-facing messages in `convertToLlm()`: `compactionSummary` and `branchSummary` become user messages rendered through the static templates\n\n- `packages/agent/src/compaction/prompts/compaction-summary-context.md`\n- `packages/agent/src/compaction/prompts/branch-summary-context.md`\n\nwhile `custom` messages pass through as developer messages with their raw content (no template).\n\n## Compaction pipeline\n\n### Triggers\n\nCompaction/context maintenance can run in five ways:\n\n1. **Manual context compaction**: `/compact [instructions]` calls `AgentSession.compact(...)`.\n2. **Automatic overflow recovery**: after a same-model assistant error that matches context overflow.\n3. **Automatic incomplete-output recovery**: after a same-model assistant message ends with `stopReason === \"length\"` (OpenAI/Codex `response.incomplete`).\n4. **Automatic threshold maintenance**: after a successful turn when context exceeds the resolved threshold.\n5. **Idle maintenance**: `runIdleCompaction()` can invoke the same auto-maintenance path with reason `\"idle\"`.\n\n### Compaction shape (visual)\n\n```text\nBefore compaction:\n\n  entry:  0     1     2     3      4     5     6      7      8     9\n        ┌─────┬─────┬─────┬──────┬─────┬─────┬──────┬──────┬─────┬──────┐\n        │ hdr │ usr │ ass │ tool │ usr │ ass │ tool │ tool │ ass │ tool │\n        └─────┴─────┴─────┴──────┴─────┴─────┴──────┴──────┴─────┴──────┘\n                └────────┬───────┘ └──────────────┬──────────────┘\n               messagesToSummarize            kept messages\n                                   ↑\n                          firstKeptEntryId (entry 4)\n\nAfter compaction (new entry appended):\n\n  entry:  0     1     2     3      4     5     6      7      8     9      10\n        ┌─────┬─────┬─────┬──────┬─────┬─────┬──────┬──────┬─────┬──────┬─────┐\n        │ hdr │ usr │ ass │ tool │ usr │ ass │ tool │ tool │ ass │ tool │ cmp │\n        └─────┴─────┴─────┴──────┴─────┴─────┴──────┴──────┴─────┴──────┴─────┘\n               └──────────┬──────┘ └──────────────────────┬───────────────────┘\n                 not sent to LLM                    sent to LLM\n                                                         ↑\n                                              starts from firstKeptEntryId\n\nWhat the LLM sees:\n\n  ┌────────┬─────────┬─────┬─────┬──────┬──────┬─────┬──────┐\n  │ system │ summary │ usr │ ass │ tool │ tool │ ass │ tool │\n  └────────┴─────────┴─────┴─────┴──────┴──────┴─────┴──────┘\n       ↑         ↑      └─────────────────┬────────────────┘\n    prompt   from cmp          messages from firstKeptEntryId\n```\n\n### Overflow/incomplete recovery vs threshold/idle maintenance\n\nThe automatic paths are intentionally different:\n\n- **Overflow recovery**\n  - Trigger: current-model assistant error is detected as context overflow and the error is not older than the latest compaction.\n  - The failing assistant error message is removed from active agent state before retry.\n  - Context promotion is tried first; if a configured larger model is available, the agent switches model and retries without compacting.\n  - If promotion is unavailable and compaction is enabled, context-full compaction runs with `reason: \"overflow\"` and `willRetry: true`; handoff strategy is not used for overflow because the handoff request would reuse the overflowing input.\n  - On success, `agent.continue()` is scheduled to retry the turn.\n\n- **Incomplete-output recovery**\n  - Trigger: same-model assistant message ends with `stopReason === \"length\"` and the message is not older than the latest compaction.\n  - The incomplete assistant message is removed from active agent state before recovery.\n  - Context promotion is tried first.\n  - If promotion is unavailable and compaction is enabled, auto maintenance runs with `reason: \"incomplete\"` and `willRetry: true`.\n  - Unlike overflow, `compaction.strategy: \"handoff\"` is allowed for incomplete-output recovery because the input context is still usable.\n  - On context-full success, `agent.continue()` is scheduled to retry the turn.\n\n- **Threshold maintenance**\n  - Trigger: successful, non-error assistant message whose adjusted context tokens exceed `resolveThresholdTokens(...)`.\n  - Tool-output pruning can reduce the measured token count before threshold comparison.\n  - Context promotion is tried before compaction.\n  - If promotion is unavailable, auto maintenance runs with `reason: \"threshold\"` and `willRetry: false`.\n  - With `compaction.strategy: \"handoff\"`, threshold maintenance normally schedules a post-prompt auto-handoff task instead of writing a compaction entry; pre-prompt checks run it inline to avoid racing the next turn. If handoff returns no document without aborting, it falls back to context-full compaction.\n  - On success, if `compaction.autoContinue !== false`, schedules an agent-authored developer auto-continue prompt from `prompts/system/auto-continue.md`.\n\n- **Idle maintenance**\n  - Trigger: `runIdleCompaction()` when not streaming or already compacting.\n  - Uses `reason: \"idle\"` and does not auto-continue afterward.\n\n### Snapcompact strategy\n\n`compaction.strategy: \"snapcompact\"` replaces the LLM summarization call with a local, deterministic archival pass (`compact` from `@oh-my-pi/snapcompact`):\n\n- The discarded history is serialized, whitespace-collapsed, and printed onto model-aware PNG frames (frame width fixed per shape; frame height hugs the rows actually printed) using bundled public-domain pixel fonts. The shape — and frame size — resolve from the **model id** when the model line was measured: Claude reads X.org `6x12` glyphs with dimmed stopwords (`6x12-dim`; high-res lines — Opus 4.7+, Fable, Mythos — get 1932px frames under Anthropic's 4,784 visual-token cap, older lines stay at 1568px), Gemini reads two word-wrapped columns of `8x13` glyphs with sentence-hue ink and dimmed stopwords (`doc-8on16-sent-dim` at 2048px — Gemini 3.x bills a fixed 1,120-token budget per image at any pixel size), GPT/Kimi/GLM read `8x13` glyphs on a 16px pitch (`8on16-bw` at 1568px — patch billing is area-proportional, and kimi's processor downscales past 1792px). A Claude routed through Vertex or OpenRouter keeps its Claude shape. Unmeasured models fall back to their wire API family (Anthropic-family/unknown → `6x12-dim`, Google → `doc-8on16-sent-dim`, OpenAI-compatible → `8on16-bw`); billing (per-family patch/budget formulas, OpenAI's `detail: \"original\"` hint) always follows the API carrying the request, computed for the resolved frame size. The `snapcompact.shape` setting (default `auto`) forces one of the research-eval variants instead: square grids (`8x8r`/`8x8u`/`6x6u`/`5x8` × sentence-hue/black ink) or the per-model eval winners (`6x12-dim`, `8x13-bw`, `8on16-bw`, and the two-column word-wrapped `doc-8on16-bw`/`-sent`/`-sent-dim`, where `dim` prints stopwords in gray). A forced variant keeps its geometry but is re-priced for the target provider's image billing. The same setting governs inline system-prompt/tool-result imaging (`snapcompact.systemPrompt`, `snapcompact.toolResults`).\n- Serialization keeps the archive conversation-dense: tool results are truncated head+tail (default 2,000 chars at a 0.6 head ratio), tool-call argument values are capped per value (500) and per call (2,000), and tool output is printed in dim gray ink so conversation reads louder than tool noise. All budgets and the dimming are configurable via `SerializeOptions` (`toolResultMaxChars`, `toolArgMaxChars`, `toolCallMaxChars`, `truncateHeadRatio`, `dimToolResults`).\n- Frames persist under `CompactionEntry.preserveData.snapcompact` and are re-attached to the `compactionSummary` message as image blocks on every context rebuild; the entry's `summary` is a deterministic reading guide (grid geometry, role tags, truncation notes) plus the usual file-operation lists.\n- Later compactions carry earlier frames forward. The frame budget is provider-aware (`providerFrameBudget`): the per-provider image cap clamped to 8 (`MAX_FRAMES`) — OpenRouter hard-caps requests at 8 images and silently drops the excess, unknown providers get a safe floor of 5. Beyond the budget the archive fades from the middle out: the earliest frame (session head — the original request, or the filmed summary of older history) is pinned, and the oldest *unpinned* frames are evicted. Pages of the *current* compaction that no longer fit are never rendered or dropped — the newest unframed slice survives verbatim as a text tail on the summary (`Archive.textTail`, capped at two frame capacities with middle elision) and is folded back into frames by the next compaction. If the previous compaction was text-based, its summary is printed at the head of the frame archive as `[Summary of earlier history]`.\n- No model, API key, or network is involved, so snapcompact is also safe for overflow recovery. It requires a vision-capable current model (`model.input` includes `\"image\"`); otherwise the run falls back to context-full and emits a warning notice (auto and manual paths). Manual `/compact` honors the strategy unless custom instructions are given (those imply a directed LLM summary).\n- Rationale: the shape table comes from the snapcompact 200k-token evals in `packages/snapcompact`, where bitmap frames preserved QA recall at lower billed-token cost than raw text for vision-capable models.\n\n### Display transcript\n\nCompaction no longer visually restarts the conversation. The TUI renders the **display transcript** (`buildSessionContext({ transcript: true })` / `AgentSession.buildTranscriptSessionContext()`): every path entry in chronological order, with each compaction shown inline as a slim divider — `── 📷 compacted · ctrl+o ──` — at the point it fired. Expanding (ctrl+o) reveals the summary. Only the LLM context resets at the compaction boundary; the scrollback above the divider stays intact, including across session resume.\n\n### Pre-compaction pruning\n\nBefore compaction checks, tool-result pruning may run (`pruneToolOutputs`).\n\nDefault prune policy:\n\n- Protect newest `40_000` tool-output tokens.\n- Require at least `20_000` total estimated savings.\n- Never prune `skill` tool results, `read` results of `skill://` paths, or reads of the active plan reference file (added via `AgentSession`'s plan protection).\n\nPruned tool results are replaced with:\n\n- `[Output truncated - N tokens]`\n\nIf pruning changes entries, session storage is rewritten and agent message state is refreshed before compaction decisions.\n\n### Useless-result elision\n\nTools can flag a finished result as contextually useless — a search with zero matches, a `job` poll that timed out with everything still running, an empty `irc` inbox drain. The flag originates on the tool result (`AgentToolResult.useless`, set via `ToolResultBuilder.useless()` or directly on the returned object), is copied by the agent loop onto the persisted `ToolResultMessage` (never together with `isError` — errors always win), and is consumed in three places:\n\n- **Per-turn stale-result pass** (`pruneSupersededToolResults`, gated by `compaction.dropUseless`, default on): flagged results are blanked to the exact placeholder `[Uneventful result elided]` (`USELESS_NOTICE`) with the same cache-aware timing as superseded reads — only when the suffix after the candidate is small (≤ ~8k tokens) or the session has idled past the provider prompt-cache lifetime. Results smaller than the notice itself are never blanked (no savings), and protected tools are exempt.\n- **Threshold prune** (`pruneToolOutputs`): flagged results bypass the protect-recent window, same as superseded reads, and receive `USELESS_NOTICE` instead of the token-count placeholder.\n- **Summary serialization**: `serializeConversation` (agent and snapcompact) drops the whole tool call/result pair from summarizer/archive input — the source region is discarded after summarization anyway, so the exclusion costs no cache.\n\nThe flag never reaches provider wire formats, and flagged pairs are never removed from history (only blanked in place), so tool-call/result pairing and provider-native history replay stay intact.\n\n### Boundary and cut-point logic\n\n`prepareCompaction()` only considers entries since the last compaction entry (if any).\n\n1. Find previous compaction index.\n2. Compute `boundaryStart = prevCompactionIndex + 1`.\n3. Adapt `keepRecentTokens` using measured usage ratio when available.\n4. Run `findCutPoint()` over the boundary window.\n\nValid cut points include:\n\n- message entries with roles: `user`, `assistant`, `bashExecution`, `hookMessage`, `branchSummary`, `compactionSummary`\n- `custom_message` entries\n- `branch_summary` entries\n\nHard rule: never cut at `toolResult`.\n\nIf there are non-message metadata entries immediately before the cut point (`model_change`, `thinking_level_change`, labels, etc.), they are pulled into the kept region by moving cut index backward until a message or compaction boundary is hit.\n\n### Split-turn handling\n\nIf cut point is not at a user-turn start, compaction treats it as a split turn.\n\nTurn start detection treats these as user-turn boundaries:\n\n- `message.role === \"user\"`\n- `message.role === \"bashExecution\"`\n- `custom_message` entry\n- `branch_summary` entry\n\nSplit-turn compaction generates two summaries:\n\n1. History summary (`messagesToSummarize`)\n2. Turn-prefix summary (`turnPrefixMessages`)\n\nFinal stored summary is merged as:\n\n```markdown\n<history summary>\n\n---\n\n**Turn Context (split turn):**\n\n<turn prefix summary>\n```\n\n### Summary generation\n\n`compact(...)` builds summaries from serialized conversation text:\n\n1. Convert messages via `convertToLlm()`.\n2. Serialize with `serializeConversation()`.\n3. Wrap in `<conversation>...</conversation>`.\n4. Optionally include `<previous-summary>...</previous-summary>`.\n5. Optionally inject extension hook context and active memory-backend compaction context as `<additional-context>` entries.\n6. Execute summarization prompt with `SUMMARIZATION_SYSTEM_PROMPT`.\n\nPrompt selection:\n\n- first compaction: `compaction-summary.md`\n- iterative compaction with prior summary: `compaction-update-summary.md`\n- split-turn second pass: `compaction-turn-prefix.md`\n- short UI summary: `compaction-short-summary.md`\n- handoff document: `handoff-document.md` (used by `generateHandoff(...)`, not serialized compaction)\n\nRemote summarization modes:\n\n- If `compaction.remoteEndpoint` is set and remote compaction is enabled, local summary generation POSTs:\n  - `{ systemPrompt, prompt }`\n- Expects JSON containing at least `{ summary }`.\n- For OpenAI/OpenAI Codex models, compaction first tries the provider-native `/responses/compact` endpoint when remote compaction is enabled. It preserves provider replacement history in `preserveData.openaiRemoteCompaction` and falls back to local summarization if that native request fails.\n\n### Handoff generation\n\n`packages/agent/src/compaction/compaction.ts` also exports `generateHandoff(...)`. Handoff generation uses the same `completeSimple(...)` oneshot style as summarization, but it preserves the live agent cache prefix by sending the active system prompt, tool array, and real LLM message history, then appending one agent-attributed `user` message containing the handoff prompt. It forces `toolChoice: \"none\"` and returns joined text blocks directly.\n\nHandoff does not write a `CompactionEntry`. `AgentSession.handoff()` owns the session transition: it starts a new session, injects the generated document as a visible `custom_message` with `customType: \"handoff\"`, and rebuilds agent messages from that new session.\n\n### File-operation context in summaries\n\nCompaction tracks cumulative file activity using assistant tool calls:\n\n- `read(path)` → read set\n- `write(path)` → modified set\n- `edit(path)` → modified set\n\nCumulative behavior:\n\n- Includes prior compaction details only when prior entry is pi-generated (`fromExtension !== true`).\n- In split turns, includes turn-prefix file ops too.\n- `details.readFiles` excludes files also modified; `details.modifiedFiles` carries the rest (persisted shape is unchanged).\n\nSummary text gets one `<files>` tag appended via prompt template: a grouped, prefix-folded directory tree (find-tool shape) with a per-file access marker — `(Read)` for read-only files, `(Write)` for modified files never read, `(RW)` for modified files also present in the cumulative read set. Capped at 20 files with an `… (N more files omitted)` line.\n\n```xml\n<files>\n# packages/agent/src/compaction/\ncompaction.ts (Read)\nutils.ts (RW)\n## prompts/\nfile-operations.md (Write)\n</files>\n```\n\nLegacy `<read-files>`/`<modified-files>` tags from summaries written by earlier versions are stripped (alongside `<files>`) before re-appending, so old summaries self-heal on the next compaction.\n\n### Persist and reload\n\nAfter summary generation (or hook-provided summary), agent session:\n\n1. Appends `CompactionEntry` with `appendCompaction(...)` for context-full maintenance; handoff strategy creates a new session and injects a handoff `custom_message` instead.\n2. Rebuilds display context from the active leaf via `buildDisplaySessionContext()`.\n3. Replaces live agent messages with rebuilt context.\n4. Synchronizes active todo phases from the rebuilt branch and closes provider sessions whose history was rewritten.\n5. Emits `session_compact` hook event.\n\n## Branch summarization pipeline\n\nBranch summarization is tied to tree navigation, not token overflow.\n\n### Trigger\n\nDuring `navigateTree(...)`:\n\n1. Compute abandoned entries from old leaf to common ancestor using `collectEntriesForBranchSummary(...)`.\n2. If caller requested summary (`options.summarize`), generate summary before switching leaf.\n3. If summary exists, attach it at the navigation target using `branchWithSummary(...)`.\n\nOperationally this is commonly driven by `/tree` flow when `branchSummary.enabled` is enabled.\n\n### Branch switch shape (visual)\n\n```text\nTree before navigation:\n\n         ┌─ B ─ C ─ D (old leaf, being abandoned)\n    A ───┤\n         └─ E ─ F (target)\n\nCommon ancestor: A\nEntries to summarize: B, C, D\n\nAfter navigation with summary:\n\n         ┌─ B ─ C ─ D ─ [summary of B,C,D]\n    A ───┤\n         └─ E ─ F (new leaf)\n```\n\n### Preparation and token budget\n\n`generateBranchSummary(...)` computes budget as:\n\n- `tokenBudget = model.contextWindow - branchSummary.reserveTokens`\n\n`prepareBranchEntries(...)` then:\n\n1. First pass: collect cumulative file ops from all summarized entries, including prior pi-generated `branch_summary` details.\n2. Second pass: walk newest → oldest, adding messages until token budget is reached.\n3. Prefer preserving recent context.\n4. May still include large summary entries near budget edge for continuity.\n\nCompaction entries are included as messages (`compactionSummary`) during branch summarization input.\n\n### Summary generation and persistence\n\nBranch summarization:\n\n1. Converts and serializes selected messages.\n2. Wraps in `<conversation>`.\n3. Uses custom instructions if supplied, otherwise `branch-summary.md`.\n4. Calls summarization model with `SUMMARIZATION_SYSTEM_PROMPT`.\n5. Prepends `branch-summary-preamble.md`.\n6. Appends file-operation tags.\n\nResult is stored as `BranchSummaryEntry` with optional details (`readFiles`, `modifiedFiles`).\n\n## Extension and hook touchpoints\n\n### `session_before_compact`\n\nPre-compaction hook.\n\nCan:\n\n- cancel compaction (`{ cancel: true }`)\n- provide full custom compaction payload (`{ compaction: CompactionResult }`)\n\n### `session.compacting`\n\nPrompt/context customization hook for default compaction.\n\nCan return:\n\n- `prompt` (override base summary prompt)\n- `context` (extra context lines injected into `<additional-context>`)\n- `preserveData` (stored on compaction entry)\n\n### `session_compact`\n\nPost-compaction notification with saved `compactionEntry` and `fromExtension` flag.\n\n### `session_before_tree`\n\nRuns on tree navigation before default branch summary generation.\n\nCan:\n\n- cancel navigation\n- provide custom `{ summary: { summary, details } }` used when user requested summarization\n\n### `session_tree`\n\nPost-navigation event exposing new/old leaf and optional summary entry.\n\n## Runtime behavior and failure semantics\n\n- Manual compaction aborts current agent operation first.\n- `abortCompaction()` cancels manual compaction, auto-compaction, and handoff generation controllers.\n- Auto compaction emits start/end session events for UI/state updates.\n- Auto compaction can try multiple model candidates and retry transient failures; long retry delays prefer the next candidate when one is available.\n- Overflow errors are excluded from generic retry path because they are handled by context promotion/compaction.\n- If auto-compaction fails:\n  - overflow path emits `Context overflow recovery failed: ...`\n  - incomplete-output path emits `Incomplete response recovery failed: ...`\n  - threshold/idle paths emit `Auto-compaction failed: ...`\n- Branch summarization can be cancelled via abort signal (e.g., Escape), returning canceled/aborted navigation result.\n\n## Settings and defaults\n\nFrom `settings-schema.ts`:\n\n- `compaction.enabled` = `true`\n- `compaction.strategy` = `\"context-full\"` (`\"handoff\"`, `\"shake\"`, `\"snapcompact\"`, and `\"off\"` are also supported)\n- `compaction.reserveTokens` = `16384`\n- `compaction.keepRecentTokens` = `20000`\n- `compaction.autoContinue` = `true`\n- `compaction.remoteEnabled` = `true`\n- `compaction.remoteEndpoint` = `undefined`\n- `compaction.thresholdPercent` = `-1` and `compaction.thresholdTokens` = `-1`; when no positive override is set, the threshold is `contextWindow - max(15% of contextWindow, reserveTokens)`\n- `compaction.idleEnabled` = `false`\n- `compaction.idleThresholdTokens` = `200000`\n- `compaction.idleTimeoutSeconds` = `300`\n- `branchSummary.enabled` = `false`\n- `branchSummary.reserveTokens` = `16384`\n\nThese values are consumed at runtime by `AgentSession` and compaction/branch summarization modules.\n",
-	"config-usage.md": "# Configuration Discovery and Resolution\n\nThis document describes how the coding-agent resolves configuration today: which roots are scanned, how precedence works, and how resolved config is consumed by settings, skills, hooks, tools, and extensions.\n\n## Scope\n\nPrimary implementation:\n\n- `packages/coding-agent/src/config.ts`\n- `packages/coding-agent/src/config/settings.ts`\n- `packages/coding-agent/src/config/settings-schema.ts`\n- `packages/coding-agent/src/discovery/builtin.ts`\n- `packages/coding-agent/src/discovery/helpers.ts`\n\nKey integration points:\n\n- `packages/coding-agent/src/capability/index.ts`\n- `packages/coding-agent/src/discovery/index.ts`\n- `packages/coding-agent/src/extensibility/skills.ts`\n- `packages/coding-agent/src/extensibility/hooks/loader.ts`\n- `packages/coding-agent/src/extensibility/custom-tools/loader.ts`\n- `packages/coding-agent/src/extensibility/extensions/loader.ts`\n\n---\n\n## Resolution flow (visual)\n\n```text\n         Generic helper order (`config.ts`)\n┌───────────────────────────────────────┐\n│ 1) ~/.omp/agent, ~/.claude, ...       │\n│ 2) <cwd>/.omp, <cwd>/.claude, ...     │\n└───────────────────────────────────────┘\n                    │\n                    ▼\n        capability providers enumerate items\n (native provider scans project .omp before user .omp;\n  other providers have their own loading rules)\n                    │\n                    ▼\n      provider priority sort + capability dedup\n                    │\n                    ▼\n          subsystem-specific consumption\n   (settings, skills, hooks, tools, extensions)\n```\n\n## 1) Config roots and source order\n\n## Canonical roots\n\n`src/config.ts` defines a fixed source priority list:\n\n1. `.omp` (native)\n2. `.claude`\n3. `.codex`\n4. `.gemini`\n\nUser-level bases:\n\n- `~/.omp/agent`\n- `~/.claude`\n- `~/.codex`\n- `~/.gemini`\n\nProject-level bases:\n\n- `<cwd>/.omp`\n- `<cwd>/.claude`\n- `<cwd>/.codex`\n- `<cwd>/.gemini`\n\n`CONFIG_DIR_NAME` is `.omp` (`packages/utils/src/dirs.ts`).\n\n## Important constraint\n\nThe generic helpers in `src/config.ts` do **not** include `.pi` in source discovery order.\n\n---\n\n## 2) Core discovery helpers (`src/config.ts`)\n\n## `getConfigDirs(subpath, options)`\n\nReturns ordered entries:\n\n- User-level entries first (by source priority)\n- Then project-level entries (by same source priority)\n\nOptions:\n\n- `user` (default `true`)\n- `project` (default `true`)\n- `cwd` (default `getProjectDir()`)\n- `existingOnly` (default `false`)\n\nThis API is used for directory-based config lookups (commands, hooks, tools, agents, etc.).\n\n## `findConfigFile(subpath, options)` / `findConfigFileWithMeta(...)`\n\nSearches for the first existing file across ordered bases, returns first match (path-only or path+metadata).\n\n## `findAllNearestProjectConfigDirs(subpath, cwd)`\n\nWalks parent directories upward and returns the **nearest existing directory per source base** (`.omp`, `.claude`, `.codex`, `.gemini`), then sorts results by source priority.\n\nUse this when project config should be inherited from ancestor directories (monorepo/nested workspace behavior).\n\n---\n\n## 3) File config wrapper (`ConfigFile<T>` in `src/config.ts`)\n\n`ConfigFile<T>` is the schema-validated loader for single config files.\n\nSupported formats:\n\n- `.yml` / `.yaml`\n- `.json` / `.jsonc`\n\nBehavior:\n\n- Validates parsed data against a provided Zod schema.\n- Caches load result until `invalidate()`.\n- Returns tri-state result via `tryLoad()`:\n  - `ok`\n  - `not-found`\n  - `error` (`ConfigError` with schema/parse context)\n\nLegacy migration still supported:\n\n- If target path is `.yml`/`.yaml`, a sibling `.json` is auto-migrated once (`migrateJsonToYml`).\n\n---\n\n## 4) Settings resolution model (`src/config/settings.ts`)\n\nThe runtime settings model is layered:\n\n1. Global settings: `~/.omp/agent/config.yml`\n2. Project settings: discovered via settings capability (`settings.json` and `config.yml` from providers)\n3. CLI config overlays: `omp --config <path>` / repeated `--config` files, loaded as `config.yml`-style YAML for this process only\n4. Runtime overrides: in-memory, non-persistent\n5. Schema defaults: from `SETTINGS_SCHEMA`\n\nEffective precedence:\n\n`defaults <- global <- project <- CLI config overlays <- overrides`\n\nWrite behavior:\n\n- `settings.set(...)` writes to the **global** layer (`config.yml`) and queues background save.\n- Project settings are read-only from capability discovery.\n\n## Migration behavior still active\n\nOn startup, if `config.yml` is missing:\n\n1. Migrate from `~/.omp/agent/settings.json` (renamed to `.bak` on success)\n2. Merge with legacy DB settings from `agent.db`\n3. Write merged result to `config.yml`\n\nField-level migrations in `#migrateRawSettings`:\n\n- `queueMode` -> `steeringMode`\n- `ask.timeout` milliseconds -> seconds when old value looks like ms (`> 1000`)\n- Legacy flat `theme: \"...\"` -> `theme.dark/theme.light` structure\n\n---\n\n## 5) Capability/discovery integration\n\nMost non-core config loading flows through the capability registry (`src/capability/index.ts` + `src/discovery/index.ts`).\n\n## Provider ordering\n\nProviders are sorted by numeric priority (higher first). Example priorities:\n\n- Native OMP (`builtin.ts`): `100`\n- Claude: `80`\n- Codex / agents / Claude marketplace: `70`\n- Gemini: `60`\n\n```text\nProvider precedence (higher wins)\n\nnative (.omp)          priority 100\nclaude                 priority  80\ncodex / agents / ...   priority  70\ngemini                 priority  60\n```\n\n## Dedup semantics\n\nCapabilities define a `key(item)`:\n\n- same key => first item wins (higher-priority/earlier-loaded item)\n- no key (`undefined`) => no dedup, all items retained\n\nRelevant keys:\n\n- skills: `name`\n- tools: `name`\n- hooks: `${type}:${tool}:${name}`\n- extension modules: `name`\n- extensions: `name`\n- settings: no dedup (all items preserved)\n\n---\n\n## 6) Native `.omp` provider behavior (`packages/coding-agent/src/discovery/builtin.ts`)\n\nNative provider (`id: native`) reads native config from:\n\n- project: `<cwd>/.omp/...`\n- user: `~/.omp/agent/...`\n\n### Directory admission rules\n\n- Slash commands, rules, prompts, instructions, hooks, tools, extensions, extension modules, and settings use a project/user root only when the root directory exists and is non-empty.\n- Skills scan `<ancestor>/.omp/skills` for each ancestor from the current working directory up to the repo root/home boundary, plus `~/.omp/agent/skills`, without requiring the root `.omp` directory itself to be non-empty.\n- `SYSTEM.md` and `AGENTS.md` read user-level files directly and use nearest-ancestor project `.omp` lookup for project files, but the project `.omp` directory must be non-empty. See [`docs/system-prompt-customization.md`](./system-prompt-customization.md) for the full `SYSTEM.md` / `APPEND_SYSTEM.md` contract (replace vs. append, templating).\n\n### Scope-specific loading\n\n- Skills: `<ancestor>/.omp/skills/*/SKILL.md` and `~/.omp/agent/skills/*/SKILL.md`\n- Slash commands: `commands/*.md`\n- Rules: `rules/*.{md,mdc}`\n- Prompts: `prompts/*.md`\n- Instructions: `instructions/*.md`\n- Hooks: `hooks/pre/*`, `hooks/post/*`\n- Tools: `tools/*.{json,md,ts,js,sh,bash,py}` and `tools/<name>/index.ts`\n- Extension modules: discovered under `extensions/` (+ legacy `settings.json.extensions` string array)\n- Extensions: `extensions/<name>/gemini-extension.json`\n- Settings capability: `settings.json`, then `config.yml`\n\n### Nearest-project lookup nuance\n\n## For `SYSTEM.md` and `AGENTS.md`, native provider uses nearest-ancestor project `.omp` directory search (walk-up) and still requires the project `.omp` dir to be non-empty.\n\n## 7) How major subsystems consume config\n\n## Settings subsystem\n\n- `Settings.init()` loads global `config.yml` + discovered project settings capability items.\n- Only capability items with `level === \"project\"` are merged into project layer.\n\n### Session title prompt override\n\nCreate `TITLE_SYSTEM.md` in the same config locations as `SYSTEM.md` / `APPEND_SYSTEM.md`:\n\n```text\n# ~/.omp/agent/TITLE_SYSTEM.md\nGenerate a session name using lowercase `<type>:<primary-objective>`.\n```\n\n- Missing `TITLE_SYSTEM.md` keeps the bundled title prompts.\n- Discovery uses the same project-then-user config directory pattern as `SYSTEM.md`: project `.omp/TITLE_SYSTEM.md` first, then user `~/.omp/agent/TITLE_SYSTEM.md` and the other supported config bases.\n- The override replaces only the automatic session-title generation system prompt; normal `SYSTEM.md` / `APPEND_SYSTEM.md` prompt customization is unaffected.\n- The online path still forces the `set_title` tool call. The local tiny-title path keeps the `<title>...</title>` prefill/stop wrapper and uses this file as its system turn.\n\n## Skills subsystem\n\n- `extensibility/skills.ts` loads via `loadCapability(skillCapability.id, { cwd })`.\n- Applies source toggles and filters (`ignoredSkills`, `includeSkills`, custom dirs).\n- Legacy-named toggles still exist (`skills.enablePiUser`, `skills.enablePiProject`) but they gate the native provider (`provider === \"native\"`).\n\n## Hooks subsystem\n\n- `discoverAndLoadHooks()` resolves hook paths from hook capability + explicit configured paths.\n- Then loads modules via Bun import.\n\n## Tools subsystem\n\n- `discoverAndLoadCustomTools()` resolves tool paths from tool capability + plugin tool paths + explicit configured paths.\n- Declarative `.md/.json` tool files are metadata only; executable loading expects code modules.\n\n## Extensions subsystem\n\n- `discoverAndLoadExtensions()` resolves extension modules from extension-module capability plus explicit paths.\n- Current implementation intentionally keeps only capability items with `_source.provider === \"native\"` before loading.\n\n---\n\n## 8) Precedence rules to rely on\n\nUse this mental model:\n\n1. Source directory ordering from `config.ts` determines candidate path order.\n2. Capability provider priority determines cross-provider precedence.\n3. Capability key dedup determines collision behavior (first wins for keyed capabilities).\n4. Subsystem-specific merge logic can further change effective precedence (especially settings).\n\n### Settings-specific caveat\n\nSettings capability items are not deduplicated; `Settings.#loadProjectSettings()` deep-merges project items in returned order. Because merge applies later item values over earlier values, effective override behavior depends on provider emission order, not just capability key semantics.\n\n---\n\n## 9) Legacy/compatibility behaviors still present\n\n- `ConfigFile` JSON -> YAML migration for YAML-targeted files.\n- Settings migration from `settings.json` and `agent.db` to `config.yml`.\n- Settings key migrations include `queueMode`, `ask.timeout`, flat `theme`, `task.isolation.enabled`, legacy `task.isolation.mode` values, removed edit modes, `statusLine.plan_mode`, `memories.enabled`, and hindsight scoping/name fields.\n- Legacy setting names `skills.enablePiUser` / `skills.enablePiProject` are still active gates for native skill source.\n\nIf these compatibility paths are removed in code, update this document immediately; several runtime behaviors still depend on them today.\n",
+	"config-usage.md": "# Configuration Discovery and Resolution\n\nThis document describes how the coding-agent resolves configuration today: which roots are scanned, how precedence works, and how resolved config is consumed by settings, skills, hooks, tools, and extensions.\n\n## Scope\n\nPrimary implementation:\n\n- `packages/coding-agent/src/config.ts`\n- `packages/coding-agent/src/config/settings.ts`\n- `packages/coding-agent/src/config/settings-schema.ts`\n- `packages/coding-agent/src/discovery/builtin.ts`\n- `packages/coding-agent/src/discovery/helpers.ts`\n\nKey integration points:\n\n- `packages/coding-agent/src/capability/index.ts`\n- `packages/coding-agent/src/discovery/index.ts`\n- `packages/coding-agent/src/extensibility/skills.ts`\n- `packages/coding-agent/src/extensibility/hooks/loader.ts`\n- `packages/coding-agent/src/extensibility/custom-tools/loader.ts`\n- `packages/coding-agent/src/extensibility/extensions/loader.ts`\n\n---\n\n## Resolution flow (visual)\n\n```text\n         Generic helper order (`config.ts`)\n┌───────────────────────────────────────┐\n│ 1) ~/.omp/agent, ~/.claude, ...       │\n│ 2) <cwd>/.omp, <cwd>/.claude, ...     │\n└───────────────────────────────────────┘\n                    │\n                    ▼\n        capability providers enumerate items\n (native provider scans project .omp before user .omp;\n  other providers have their own loading rules)\n                    │\n                    ▼\n      provider priority sort + capability dedup\n                    │\n                    ▼\n          subsystem-specific consumption\n   (settings, skills, hooks, tools, extensions)\n```\n\n## 1) Config roots and source order\n\n## Canonical roots\n\n`src/config.ts` defines a fixed source priority list:\n\n1. `.omp` (native)\n2. `.claude`\n3. `.codex`\n4. `.gemini`\n\nUser-level bases:\n\n- `~/.omp/agent`\n- `~/.claude`\n- `~/.codex`\n- `~/.gemini`\n\nProject-level bases:\n\n- `<cwd>/.omp`\n- `<cwd>/.claude`\n- `<cwd>/.codex`\n- `<cwd>/.gemini`\n\n`CONFIG_DIR_NAME` is `.omp` (`packages/utils/src/dirs.ts`).\n\n## Important constraint\n\nThe generic helpers in `src/config.ts` do **not** include `.pi` in source discovery order.\n\n---\n\n## 2) Core discovery helpers (`src/config.ts`)\n\n## `getConfigDirs(subpath, options)`\n\nReturns ordered entries:\n\n- User-level entries first (by source priority)\n- Then project-level entries (by same source priority)\n\nOptions:\n\n- `user` (default `true`)\n- `project` (default `true`)\n- `cwd` (default `getProjectDir()`)\n- `existingOnly` (default `false`)\n\nThis API is used for directory-based config lookups (commands, hooks, tools, agents, etc.).\n\n## `findConfigFile(subpath, options)` / `findConfigFileWithMeta(...)`\n\nSearches for the first existing file across ordered bases, returns first match (path-only or path+metadata).\n\n## `findAllNearestProjectConfigDirs(subpath, cwd)`\n\nWalks parent directories upward and returns the **nearest existing directory per source base** (`.omp`, `.claude`, `.codex`, `.gemini`), then sorts results by source priority.\n\nUse this when project config should be inherited from ancestor directories (monorepo/nested workspace behavior).\n\n---\n\n## 3) File config wrapper (`ConfigFile<T>` in `src/config.ts`)\n\n`ConfigFile<T>` is the schema-validated loader for single config files.\n\nSupported formats:\n\n- `.yml` / `.yaml`\n- `.json` / `.jsonc`\n\nBehavior:\n\n- Validates parsed data against a provided Zod schema.\n- Caches load result until `invalidate()`.\n- Returns tri-state result via `tryLoad()`:\n  - `ok`\n  - `not-found`\n  - `error` (`ConfigError` with schema/parse context)\n\nLegacy migration still supported:\n\n- If target path is `.yml`/`.yaml`, a sibling `.json` is auto-migrated once (`migrateJsonToYml`).\n\n---\n\n## 4) Settings resolution model (`src/config/settings.ts`)\n\nThe runtime settings model is layered:\n\n1. Global settings: `~/.omp/agent/config.yml`\n2. Project settings: discovered via settings capability (`settings.json` and `config.yml` from providers)\n3. CLI config overlays: `omp --config <path>` / repeated `--config` files, loaded as `config.yml`-style YAML for this process only\n4. Runtime overrides: in-memory, non-persistent\n5. Schema defaults: from `SETTINGS_SCHEMA`\n\nEffective precedence:\n\n`defaults <- global <- project <- CLI config overlays <- overrides`\n\nWrite behavior:\n\n- `settings.set(...)` writes to the **global** layer (`config.yml`) and queues background save.\n- Project settings are read-only from capability discovery.\n\n## Migration behavior still active\n\nOn startup, if `config.yml` is missing:\n\n1. Migrate from `~/.omp/agent/settings.json` (renamed to `.bak` on success)\n2. Merge with legacy DB settings from `agent.db`\n3. Write merged result to `config.yml`\n\nField-level migrations in `#migrateRawSettings`:\n\n- `queueMode` -> `steeringMode`\n- `ask.timeout` milliseconds -> seconds when old value looks like ms (`> 1000`)\n- Legacy flat `theme: \"...\"` -> `theme.dark/theme.light` structure\n\n---\n\n## 5) Capability/discovery integration\n\nMost non-core config loading flows through the capability registry (`src/capability/index.ts` + `src/discovery/index.ts`).\n\n## Provider ordering\n\nProviders are sorted by numeric priority (higher first). Example priorities:\n\n- Native OMP (`builtin.ts`): `100`\n- Claude: `80`\n- Codex / agents / Claude marketplace: `70`\n- Gemini: `60`\n\n```text\nProvider precedence (higher wins)\n\nnative (.omp)          priority 100\nclaude                 priority  80\ncodex / agents / ...   priority  70\ngemini                 priority  60\n```\n\n## Dedup semantics\n\nCapabilities define a `key(item)`:\n\n- same key => first item wins (higher-priority/earlier-loaded item)\n- no key (`undefined`) => no dedup, all items retained\n\nRelevant keys:\n\n- skills: `name`\n- tools: `name`\n- hooks: `${type}:${tool}:${name}`\n- extension modules: `name`\n- extensions: `name`\n- settings: no dedup (all items preserved)\n\n---\n\n## 6) Native `.omp` provider behavior (`packages/coding-agent/src/discovery/builtin.ts`)\n\nNative provider (`id: native`) reads native config from:\n\n- project: `<cwd>/.omp/...`\n- user: `~/.omp/agent/...`\n\n### Directory admission rules\n\n- Slash commands, rules, prompts, instructions, hooks, tools, extensions, extension modules, and settings use a project/user root only when the root directory exists and is non-empty.\n- Skills scan `<ancestor>/.omp/skills` for each ancestor from the current working directory up to the repo root/home boundary, plus `~/.omp/agent/skills`, without requiring the root `.omp` directory itself to be non-empty.\n- `SYSTEM.md` and `AGENTS.md` read user-level files directly and use nearest-ancestor project `.omp` lookup for project files, but the project `.omp` directory must be non-empty. See [`docs/system-prompt-customization.md`](./system-prompt-customization.md) for the full `SYSTEM.md` / `APPEND_SYSTEM.md` contract (replace vs. append, templating).\n\n### Scope-specific loading\n\n- Skills: `<ancestor>/.omp/skills/*/SKILL.md` and `~/.omp/agent/skills/*/SKILL.md`\n- Slash commands: `commands/*.md`\n- Rules: `rules/*.{md,mdc}`\n- Prompts: `prompts/*.md`\n- Instructions: `instructions/*.md`\n- Hooks: `hooks/pre/*`, `hooks/post/*`\n- Tools: `tools/*.{json,md,ts,js,sh,bash,py}` and `tools/<name>/index.ts`\n- Extension modules: discovered under `extensions/` (+ legacy `settings.json.extensions` string array)\n- Extensions: `extensions/<name>/gemini-extension.json`\n- Settings capability: `settings.json`, then `config.yml`\n\n### Nearest-project lookup nuance\n\n## For `SYSTEM.md` and `AGENTS.md`, native provider uses nearest-ancestor project `.omp` directory search (walk-up) and still requires the project `.omp` dir to be non-empty.\n\n## 7) How major subsystems consume config\n\n## Settings subsystem\n\n- `Settings.init()` loads global `config.yml` + discovered project settings capability items.\n- Only capability items with `level === \"project\"` are merged into project layer.\n\n### Session title prompt override\n\nCreate `TITLE_SYSTEM.md` in the same config locations as `SYSTEM.md` / `APPEND_SYSTEM.md`:\n\n```text\n# ~/.omp/agent/TITLE_SYSTEM.md\nGenerate a session name using lowercase `<type>:<primary-objective>`.\n```\n\n- Missing `TITLE_SYSTEM.md` keeps the bundled title prompts.\n- Discovery uses the same project-then-user config directory pattern as `SYSTEM.md`: project `.omp/TITLE_SYSTEM.md` first, then user `~/.omp/agent/TITLE_SYSTEM.md` and the other supported config bases.\n- The override replaces only the automatic session-title generation system prompt; normal `SYSTEM.md` / `APPEND_SYSTEM.md` prompt customization is unaffected.\n- The online path forces the `set_title` tool call when the title model honors a forced `tool_choice`. Tool-choice-less providers (chat-completions hosts without `tool_choice` support, Claude Fable/Mythos) instead receive a marker-based prompt and emit the title wrapped in `<title>...</title>`, which is parsed leniently (a plain sentence or a truncated/unclosed tag still works). A `TITLE_SYSTEM.md` override is reused in both modes; in marker mode the wrap-in-`<title>` instruction is appended after it. The local tiny-title path keeps the `<title>...</title>` prefill/stop wrapper and uses this file as its system turn.\n\n## Skills subsystem\n\n- `extensibility/skills.ts` loads via `loadCapability(skillCapability.id, { cwd })`.\n- Applies source toggles and filters (`ignoredSkills`, `includeSkills`, custom dirs).\n- Legacy-named toggles still exist (`skills.enablePiUser`, `skills.enablePiProject`) but they gate the native provider (`provider === \"native\"`).\n\n## Hooks subsystem\n\n- `discoverAndLoadHooks()` resolves hook paths from hook capability + explicit configured paths.\n- Then loads modules via Bun import.\n\n## Tools subsystem\n\n- `discoverAndLoadCustomTools()` resolves tool paths from tool capability + plugin tool paths + explicit configured paths.\n- Declarative `.md/.json` tool files are metadata only; executable loading expects code modules.\n\n## Extensions subsystem\n\n- `discoverAndLoadExtensions()` resolves extension modules from extension-module capability plus explicit paths.\n- Current implementation intentionally keeps only capability items with `_source.provider === \"native\"` before loading.\n\n---\n\n## 8) Precedence rules to rely on\n\nUse this mental model:\n\n1. Source directory ordering from `config.ts` determines candidate path order.\n2. Capability provider priority determines cross-provider precedence.\n3. Capability key dedup determines collision behavior (first wins for keyed capabilities).\n4. Subsystem-specific merge logic can further change effective precedence (especially settings).\n\n### Settings-specific caveat\n\nSettings capability items are not deduplicated; `Settings.#loadProjectSettings()` deep-merges project items in returned order. Because merge applies later item values over earlier values, effective override behavior depends on provider emission order, not just capability key semantics.\n\n---\n\n## 9) Legacy/compatibility behaviors still present\n\n- `ConfigFile` JSON -> YAML migration for YAML-targeted files.\n- Settings migration from `settings.json` and `agent.db` to `config.yml`.\n- Settings key migrations include `queueMode`, `ask.timeout`, flat `theme`, `task.isolation.enabled`, legacy `task.isolation.mode` values, removed edit modes, `statusLine.plan_mode`, `memories.enabled`, and hindsight scoping/name fields.\n- Legacy setting names `skills.enablePiUser` / `skills.enablePiProject` are still active gates for native skill source.\n\nIf these compatibility paths are removed in code, update this document immediately; several runtime behaviors still depend on them today.\n",
 	"context-files.md": "# Context files\n\nContext files are Markdown instruction files that `omp` discovers automatically before a session starts and injects into the agent's project context. Use them for repository conventions, architecture notes, test and review expectations, and instructions that should travel with a user account or a project.\n\nYou never have to ask the agent to go read `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, or similar files — the relevant ones are already discovered, loaded, and placed in context when the session begins.\n\n## How context files relate to other concepts\n\nFour similarly named things behave differently. Keep them straight:\n\n- **Context files** are read as plain Markdown and shown to the agent inside a `<context>` block. They are advisory background that stays in the session's opening context.\n- **Sticky rules** come from a top-level `RULES.md`. They are converted into an always-apply rule that is re-attached near the current turn, so they keep their hold even after the visible conversation grows. See \"Sticky rules vs normal context\" below.\n- **Discovery providers** are the config-source adapters (`native`, `claude`, `codex`, `gemini`, `opencode`, `github`, `agents`, `agents-md`) that know where each tool keeps its files. The same provider that contributes context files may also contribute MCP servers, slash commands, skills, hooks, tools, prompts, and settings.\n- **Model providers** are inference backends such as `anthropic`, `openai`, `google`, `groq`, `ollama`, and `openrouter`. They have nothing to do with context files except that both kinds of id share the one `disabledProviders` list — see \"Disabling discovery providers\" below and [Providers](./providers.md).\n\nAuthoring **skills** and **rule** files (as opposed to the sticky `RULES.md`) is covered in [Skills](./skills.md). Customizing the system prompt with `SYSTEM.md` is covered in [System prompt customization](./system-prompt-customization.md).\n\n## Native `.omp` files\n\nThe native provider is the recommended format for new projects. It reads from your user agent directory and from `.omp/` directories inside a project, and it has the highest discovery priority, so its files win over every other convention at the same scope.\n\n| File | Scope | Behavior |\n|---|---|---|\n| `~/.omp/agent/AGENTS.md` | User | User-level context for every session unless the `native` provider is disabled. |\n| `<ancestor>/.omp/AGENTS.md` | Project | Project context. `omp` walks upward from the current directory to the repository root and uses the **nearest** non-empty `.omp/AGENTS.md`. Farther native project files are not also included. |\n| `~/.omp/agent/RULES.md` | User | User-level sticky rule content. Loaded as an always-apply rule, not as a context file. |\n| `<ancestor>/.omp/RULES.md` | Project | Project sticky rule content. Same nearest-ancestor walk-up as above. Loaded as an always-apply rule. |\n\nTwo details matter:\n\n- **Walk-up to the repository root.** Discovery starts in the current working directory and climbs through each ancestor up to the repository root, stopping at the first ancestor that has a usable `.omp/` directory. The *nearest* match wins; ancestors above it are not loaded as native context.\n- **The `.omp/` directory must be non-empty.** An empty `.omp/` directory is skipped during the walk-up, so the search continues to the next ancestor. An empty `AGENTS.md` or `RULES.md` file contributes nothing.\n\n`~/.omp/agent` is the user base. If `PI_CODING_AGENT_DIR` is set, it relocates that base, so the user files become `$PI_CODING_AGENT_DIR/AGENTS.md` and `$PI_CODING_AGENT_DIR/RULES.md`.\n\n### Monorepo example\n\n```text\nrepo/\n  .omp/\n    AGENTS.md\n    RULES.md\n  packages/api/\n    .omp/\n      AGENTS.md\n```\n\nStarting a session in `repo/packages/api`:\n\n- The native context file is `repo/packages/api/.omp/AGENTS.md` (the nearest one). `repo/.omp/AGENTS.md` is **not** also included.\n- The project sticky rule is `repo/packages/api/.omp/RULES.md` if present; otherwise the walk-up continues and `repo/.omp/RULES.md` is used.\n\nPut broad, durable project background in `AGENTS.md`. Reserve `RULES.md` for short, hard requirements that must stay visible across long conversations.\n\n## Other supported context conventions\n\n`omp` also discovers the context files of other agent tools so existing projects keep working without migration. The table lists context-file behavior only; each provider may also contribute other capabilities.\n\n| Provider id | Convention path | Scope | Notes |\n|---|---|---|---|\n| `native` | `.omp/AGENTS.md` | User + project | Recommended `omp` format. User file at `~/.omp/agent/AGENTS.md`; project file is the nearest non-empty `.omp/AGENTS.md` walking up to the repo root. |\n| `claude` | `.claude/CLAUDE.md` | User + project | User file `~/.claude/CLAUDE.md`; project file `<cwd>/.claude/CLAUDE.md` only (no ancestor walk-up). |\n| `codex` | `.codex/AGENTS.md` | User | User file `~/.codex/AGENTS.md` only. Project-level Codex context comes from a standalone `AGENTS.md` via the `agents-md` provider, not from `<cwd>/.codex/AGENTS.md`. |\n| `gemini` | `.gemini/GEMINI.md` | User + project | User file `~/.gemini/GEMINI.md`; project file `<cwd>/.gemini/GEMINI.md` only (no ancestor walk-up). |\n| `opencode` | `.config/opencode/AGENTS.md` | User | User file `~/.config/opencode/AGENTS.md` only. |\n| `github` | `.github/copilot-instructions.md` | Project | Project-only GitHub Copilot instructions from `<cwd>/.github/copilot-instructions.md`. |\n| `agents` | `.agent/AGENTS.md`, `.agents/AGENTS.md` | User + project | User files from `~/.agent/` and `~/.agents/`; project files discovered while walking up from the current directory to the repository root. |\n| `agents-md` | `AGENTS.md` | Project | Standalone (non-config-directory) `AGENTS.md` files, discovered by walking up from the current directory to the repository root (or home when no repo root is known). Files whose parent directory name starts with `.` are ignored — those belong to a config-directory provider instead. |\n\nProviders marked \"(no ancestor walk-up)\" only look in the current working directory's config directory. If you need ancestor walk-up behavior, prefer the native `.omp/AGENTS.md` format or a standalone `AGENTS.md` (the `agents-md` provider), or launch `omp` from the directory that holds the config directory.\n\n## Load order and shadowing\n\nWhen two providers describe the *same* scope, the higher-priority provider wins. Provider priorities:\n\n| Priority | Provider id |\n|---:|---|\n| 100 | `native` |\n| 80 | `claude` |\n| 70 | `agents`, `codex` |\n| 60 | `gemini` |\n| 55 | `opencode` |\n| 30 | `github` |\n| 10 | `agents-md` |\n\nDiscovered files are then deduplicated by scope:\n\n- **One user context file** is kept across all providers. Because `native` has the highest priority, `~/.omp/agent/AGENTS.md` shadows every other user-level context file.\n- **One project context file per directory depth.** Depth is measured from the current directory: the cwd is depth 0, its parent depth 1, and so on. Config subdirectories of an ancestor (`.claude/`, `.github/`, `.gemini/`, …) count as the same depth as that ancestor.\n- **At the same depth, the higher-priority provider shadows the rest.**\n- **Across depths, multiple files survive.** In a monorepo, an ancestor `AGENTS.md` and a package-level one are different depths and both load.\n- **Byte-identical files are collapsed.** If two surviving files have exactly the same content, only the copy closest to the cwd is kept.\n\nAfter deduplication, project files are sorted so **farther ancestors appear first** and files **closer to the cwd appear last**. Later files sit nearer the end of the context block, where they are most prominent.\n\n### Worked shadowing example\n\n```text\nrepo/\n  AGENTS.md\n  packages/api/\n    AGENTS.md\n    .github/copilot-instructions.md\n```\n\nStarting in `repo/packages/api`:\n\n- `repo/AGENTS.md` is found by `agents-md` at depth 2 and kept.\n- `repo/packages/api/AGENTS.md` (`agents-md`, priority 10) and `repo/packages/api/.github/copilot-instructions.md` (`github`, priority 30) both resolve to depth 0. GitHub's higher priority shadows the package-level standalone `AGENTS.md`, so the Copilot file wins at that depth.\n- The two kept files are ordered root-first, package-last, so `packages/api`'s file is the more prominent one.\n- If you add `repo/packages/api/.omp/AGENTS.md`, `native` (priority 100) wins depth 0 outright, shadowing both lower-priority files.\n\n## Injection behavior\n\nDiscovered context files are injected into the opening project prompt as a single `<context>` block, one `<file>` element per surviving file, in the sort order above:\n\n```xml\n<context>\nYou MUST follow the context files below for all tasks:\n<file path=\"/abs/path/to/repo/AGENTS.md\">\n...root content...\n</file>\n<file path=\"/abs/path/to/repo/packages/api/.github/copilot-instructions.md\">\n...package content...\n</file>\n</context>\n```\n\nThe agent sees each file's absolute path and its fully expanded Markdown content (with `@` imports already resolved — see below). Loading is automatic — there is no need to instruct the agent to search for `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, `.cursorrules`, or similar files during a session.\n\nDeeper-directory `AGENTS.md` files that were *not* auto-loaded (for example, ones below the current directory) are surfaced separately in a `<dir-context>` block that lists their paths and tells the agent to read them before editing those directories. Those files are pointers, not full injected content.\n\n## `@` imports\n\nInside any context file, an `@path` token expands inline to the referenced file's content before injection:\n\n```markdown\n# Project notes\n\nRead @docs/architecture.md before changing storage code.\nShared release steps live in @../RELEASE.md and personal aliases in @~/.notes/aliases.md.\n```\n\nThe exact rules:\n\n- **Relative paths resolve from the importing file's own directory**, not the session's working directory.\n- **`~/` and `~`** resolve from the user's home directory; absolute paths are used as-is.\n- **Tokens inside fenced code blocks and inline code spans are left untouched** — useful when you want to *write about* an `@token` without expanding it.\n- **`git@github.com:org/repo.git` and `user@example.com`-style tokens are not treated as imports.** A token only counts when the `@` sits at the start of a line or after a space or tab.\n- **Trailing sentence punctuation is trimmed** off the path (`. , ; : ! ? ) ] } \" '`), so `@docs/setup.md.` imports `docs/setup.md`.\n- **Imports recurse up to five hops.** An imported file may itself contain `@` imports, up to a total depth of five.\n- **Cycles are skipped.** A file already pulled into the current expansion tree is not re-expanded, so mutual imports terminate cleanly.\n- **A missing or unreadable target leaves the original `@token` text in place** rather than erroring.\n\n## Sticky rules vs normal context\n\nUse a normal context file (`AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, `.github/copilot-instructions.md`, …) for the bulk of your guidance: repository overview, code style, build and test commands, review expectations, and local conventions. These load into the opening `<context>` block.\n\nUse a top-level **`RULES.md`** for the handful of hard requirements that must stay active even after a long conversation has pushed the opening context far up the transcript:\n\n```markdown\n# ~/.omp/agent/RULES.md\n\nNever commit or push unless the user explicitly asks.\nDo not edit generated files.\n```\n\n`RULES.md` is special:\n\n- It is read **only** at the native locations — `~/.omp/agent/RULES.md` and the nearest `<ancestor>/.omp/RULES.md` from the cwd up to the repo root. A `RULES.md` anywhere else is not a context-file convention and is ignored.\n- It is loaded as an **always-apply rule**, not as a context file, so it is re-attached near the current turn and keeps its hold across long sessions.\n- It is **always sticky**: frontmatter cannot make it non-sticky. If you want conditional or opt-in behavior, write a normal rule file instead (see [Skills](./skills.md)).\n\nKeep `RULES.md` short. Long background belongs in `AGENTS.md`, where it costs context budget only once.\n\n## Disabling discovery providers\n\nTurn a provider off with the `disabledProviders` setting in `~/.omp/agent/config.yml`, a project's `.omp/config.yml`, or a `--config` overlay:\n\n```yaml\n# .omp/config.yml\ndisabledProviders:\n  - claude\n  - github\n```\n\n`disabledProviders` is a **whole-provider switch with one shared id namespace**, used by two unrelated subsystems:\n\n| Id kind | Examples | Effect when listed |\n|---|---|---|\n| Discovery provider ids | `native`, `claude`, `codex`, `gemini`, `opencode`, `github`, `agents`, `agents-md` | The entire config source is removed — not just its context files, but also any MCP servers, slash commands, skills, hooks, tools, prompts, and settings it would have contributed. |\n| Model provider ids | `anthropic`, `openai`, `google`, `groq`, `ollama`, `openrouter` | The model backend is removed from selection even when its credentials are present. See [Providers](./providers.md). |\n\nIds are exact and the two namespaces do not collide by accident: `google` disables the Google model backend, while `gemini` disables the Gemini CLI discovery files. Disabling a discovery provider is heavier than it looks — disabling `claude`, for instance, also drops Claude-discovered MCP servers, commands, skills, hooks, tools, and settings, not only `CLAUDE.md`.\n\nOnly `enabledModels` and `disabledProviders` support **path-scoped** entries, so you can vary provider availability per subtree:\n\n```yaml\ndisabledProviders:\n  - github            # disabled everywhere\n  - path: ~/work/legacy-claude\n    providers:\n      - claude         # disabled only under this directory\n```\n\nA scoped entry applies when the cwd equals the configured path or sits beneath it; `~` expands to home. Bare string entries apply everywhere.\n\nRemember that higher-precedence settings layers **replace** array settings rather than appending to them. If your global config disables `claude` but a project config sets `disabledProviders: [github]`, then inside that project Claude discovery is re-enabled and only GitHub is disabled. See [Settings](./settings.md) for the full layer precedence, merge rules, and path-scoped array details.\n\n## Troubleshooting\n\n### A file is not loaded\n\n- Native project context must live at `.omp/AGENTS.md`, and the `.omp/` directory must be non-empty; an empty `.omp/` is skipped and the walk-up continues to the next ancestor.\n- A standalone `AGENTS.md` is handled by `agents-md`, not `native`.\n- `.claude/CLAUDE.md`, `.gemini/GEMINI.md`, and `.github/copilot-instructions.md` are read only from the current working directory's config directory — not from every ancestor.\n- `~/.codex/AGENTS.md` and `~/.config/opencode/AGENTS.md` are user-level only and have no project equivalent.\n- Empty files contribute nothing for the native and standalone providers.\n- A disabled discovery provider contributes nothing — check `disabledProviders` across your global, project, and `--config` layers.\n\n### The wrong file wins\n\nAt one user scope or project depth, the higher-priority provider shadows the others (native > claude > agents/codex > gemini > opencode > github > agents-md). To force deterministic behavior, move your guidance into `.omp/AGENTS.md` (native always wins) or disable the competing discovery provider.\n\n### User context disappeared\n\nOnly one user-level context file survives, and `~/.omp/agent/AGENTS.md` has the highest priority. If it exists, it shadows user-level `~/.claude/CLAUDE.md`, `~/.codex/AGENTS.md`, `~/.gemini/GEMINI.md`, `~/.config/opencode/AGENTS.md`, and `~/.agent`/`~/.agents` files. Consolidate user guidance into the native file or remove the native one if you prefer another tool's file.\n\n### A `RULES.md` file is ignored\n\nOnly the native `RULES.md` locations are sticky: `~/.omp/agent/RULES.md` and the nearest `<ancestor>/.omp/RULES.md` from cwd to the repo root. A `RULES.md` in any other directory is not a recognized convention and will not be loaded.\n\n### An `@` import did not expand\n\nConfirm the target exists relative to the importing file (not the cwd). Imports inside fenced code blocks or inline code spans are intentionally left literal, `git@`/email-looking tokens are never imported, cycles are skipped, expansion stops after five hops, and a missing target leaves the original `@path` text unchanged.\n",
 	"custom-tools.md": "# Custom Tools\n\nCustom tools are model-callable functions that plug into the same tool execution pipeline as built-in tools.\n\nA custom tool is a TypeScript/JavaScript module that exports a factory. The factory receives a host API (`CustomToolAPI`) and returns one tool or an array of tools.\n\n## What this is (and is not)\n\n- **Custom tool**: callable by the model during a turn (`execute` + Zod parameter schema).\n- **Extension**: lifecycle/event framework that can register tools and intercept/modify events.\n- **Hook**: external pre/post command scripts.\n- **Skill**: static guidance/context package, not executable tool code.\n\nIf you need the model to call code directly, use a custom tool.\n\n## Integration paths in current code\n\nThere are two active integration styles:\n\n1. **SDK-provided custom tools** (`options.customTools`)\n   - Wrapped into agent tools via `CustomToolAdapter` or extension wrappers.\n   - Always included in the initial active tool set in SDK bootstrap.\n\n2. **Filesystem-discovered modules via loader API** (`discoverAndLoadCustomTools` / `loadCustomTools`)\n   - Exposed as library APIs in `src/extensibility/custom-tools/loader.ts`.\n   - Host code can call these to discover and load tool modules from config/provider/plugin paths.\n\n```text\nModel tool call flow\n\nLLM tool call\n   │\n   ▼\nTool registry (built-ins + custom tool adapters)\n   │\n   ▼\nCustomTool.execute(toolCallId, params, onUpdate, ctx, signal)\n   │\n   ├─ onUpdate(...)  -> streamed partial result\n   └─ return result  -> final tool content/details\n```\n\n## Discovery locations (loader API)\n\n`discoverAndLoadCustomTools(configuredPaths, cwd, builtInToolNames)` merges:\n\n1. Capability providers (`toolCapability`), including:\n   - Native OMP config (`~/.omp/agent/tools`, `.omp/tools`)\n   - Claude config (`~/.claude/tools`, `.claude/tools`)\n   - Codex config (`~/.codex/tools`, `.codex/tools`)\n   - Claude marketplace plugin cache provider\n2. Installed plugin manifests (`~/.omp/plugins/node_modules/*` via plugin loader)\n3. Explicit configured paths passed to the loader\n\n### Important behavior\n\n- Duplicate resolved paths are deduplicated.\n- Tool name conflicts are rejected against built-ins and already-loaded custom tools.\n- `.md` and `.json` files are discovered as tool metadata by some providers, but the executable module loader rejects them as runnable tools.\n- Relative configured paths are resolved from `cwd`; `~` is expanded.\n\n## Module contract\n\nA custom tool module must export a function (default export preferred):\n\n```ts\nimport type { CustomToolFactory } from \"@oh-my-pi/pi-coding-agent\";\n\nconst factory: CustomToolFactory = (pi) => ({\n  name: \"repo_stats\",\n  label: \"Repo Stats\",\n  description: \"Counts tracked TypeScript files\",\n  parameters: pi.zod.object({\n    glob: pi.zod.string().optional().default(\"**/*.ts\"),\n  }),\n\n  async execute(toolCallId, params, onUpdate, ctx, signal) {\n    onUpdate?.({\n      content: [{ type: \"text\", text: \"Scanning files...\" }],\n      details: { phase: \"scan\" },\n    });\n\n    const result = await pi.exec(\n      \"git\",\n      [\"ls-files\", params.glob ?? \"**/*.ts\"],\n      { signal, cwd: pi.cwd },\n    );\n    if (result.killed) {\n      throw new Error(\"Scan was cancelled\");\n    }\n    if (result.code !== 0) {\n      throw new Error(result.stderr || \"git ls-files failed\");\n    }\n\n    const files = result.stdout.split(\"\\n\").filter(Boolean);\n    return {\n      content: [{ type: \"text\", text: `Found ${files.length} files` }],\n      details: { count: files.length, sample: files.slice(0, 10) },\n    };\n  },\n\n  onSession(event) {\n    if (event.reason === \"shutdown\") {\n      // cleanup resources if needed\n    }\n  },\n});\n\nexport default factory;\n```\n\nSchemas are authored with Zod (`pi.zod`) and flow through the shared validation/wire pipeline.\n\nFactory return type:\n\n- `CustomTool`\n- `CustomTool[]`\n- `Promise<CustomTool | CustomTool[]>`\n\n## API surface passed to factories (`CustomToolAPI`)\n\nFrom `types.ts` and `loader.ts`:\n\n- `cwd`: host working directory\n- `exec(command, args, options?)`: process execution helper\n- `ui`: UI context (can be no-op in headless modes)\n- `hasUI`: `false` in non-interactive flows\n- `logger`: shared file logger\n- `typebox`: zod-backed compatibility shim for legacy TypeBox-style schemas\n- `zod`: injected `zod/v4` module (canonical for new schemas)\n- `pi`: injected `@oh-my-pi/pi-coding-agent` exports\n- `pushPendingAction(action)`: register a preview action for hidden `resolve` tool (`docs/resolve-tool-runtime.md`)\n  Loader starts with a no-op UI context and requires host code to call `setUIContext(...)` when real UI is ready.\n\n## Execution contract and typing\n\n`CustomTool.execute` signature:\n\n```ts\nexecute(toolCallId, params, onUpdate, ctx, signal);\n```\n\n- `params` is statically typed from your Zod/TypeBox schema via `Static<TParams>`.\n- Runtime argument validation happens before execution in the agent loop.\n- `onUpdate` emits partial results for UI streaming.\n- `ctx` includes `sessionManager`, `modelRegistry`, current `model`, `isIdle()`, `hasQueuedMessages()`, `abort()`, and optional `settings` / `autoApprove`.\n- `signal` carries cancellation.\n\n`CustomToolAdapter` bridges this to the agent tool interface and forwards calls in the correct argument order.\n\nTool definitions may also declare `strict`, `hidden`, `deferrable`, `mcpServerName`, `mcpToolName`, `approval`, and `formatApprovalDetails`.\n\n## How tools are exposed to the model\n\n- Tools are wrapped into `AgentTool` instances (`CustomToolAdapter` or extension wrappers).\n- They are inserted into the session tool registry by name.\n- In SDK bootstrap, custom and extension-registered tools are force-included in the initial active set.\n- CLI `--tools` currently validates only built-in tool names; custom tool inclusion is handled through discovery/registration paths and SDK options.\n\n## Rendering hooks\n\nOptional rendering hooks:\n\n- `renderCall(args, options, theme)`\n- `renderResult(result, options, theme, args?)`\n\nRuntime behavior in TUI:\n\n- If hooks exist, tool output is rendered inside a `Box` container.\n- `renderResult` receives `{ expanded, isPartial, spinnerFrame? }`.\n- Renderer errors are caught and logged; UI falls back to default text rendering.\n\n## Session/state handling\n\nOptional `onSession(event, ctx)` receives session lifecycle events, including:\n\n- `start`, `switch`, `branch`, `tree`, `shutdown`\n- `auto_compaction_start`, `auto_compaction_end`\n- `auto_retry_start`, `auto_retry_end`\n- `ttsr_triggered`, `todo_reminder`\n\nUse `ctx.sessionManager` to reconstruct state from history when branch/session context changes.\n\n## Failures and cancellation semantics\n\n### Synchronous/async failures\n\n- Throwing (or rejected promises) in `execute` is treated as tool failure.\n- Agent runtime converts failures into tool result messages with `isError: true` and error text content.\n- With extension wrappers, `tool_result` handlers can further rewrite content/details and even override error status.\n\n### Cancellation\n\n- Agent abort propagates through `AbortSignal` to `execute`.\n- Forward `signal` to subprocess work (`pi.exec(..., { signal })`) for cooperative cancellation.\n- `ctx.abort()` lets a tool request abort of the current agent operation.\n\n### onSession errors\n\n- `onSession` errors are caught and logged as warnings; they do not crash the session.\n\n## Real constraints to design for\n\n- Tool names must be globally unique in the active registry.\n- Prefer deterministic, schema-shaped outputs in `details` for renderer/state reconstruction.\n- Guard UI usage with `pi.hasUI`.\n- Treat `.md`/`.json` in tool directories as metadata, not executable modules.\n",
 	"environment-variables.md": "# Environment Variables (Current Runtime Reference)\n\nThis reference is derived from current code paths in:\n\n- `packages/coding-agent/src/**`\n- `packages/ai/src/**` (provider/auth resolution used by coding-agent)\n- `packages/utils/src/**` and `packages/tui/src/**` where those vars directly affect coding-agent runtime\n\nIt documents only active behavior.\n\n## Resolution model and precedence\n\nMost runtime lookups use `$env` from `@oh-my-pi/pi-utils` (`packages/utils/src/env.ts`).\n\n`$env` loading order:\n\n1. Existing process environment (`Bun.env`)\n2. Project `.env` (`$PWD/.env`) for keys not already set\n3. Agent `.env` (`~/.omp/agent/.env`, respecting `PI_CONFIG_DIR` / `PI_CODING_AGENT_DIR`) for keys not already set\n4. Config-root `.env` (`~/.omp/.env`, respecting `PI_CONFIG_DIR`) for keys not already set\n5. Home `.env` (`~/.env`) for keys not already set\n\nAdditional rule inside each `.env` file: `OMP_*` keys are mirrored to `PI_*` keys in that parsed file.\n\n---\n\n## 1) Model/provider authentication\n\nThese are consumed via `getEnvApiKey()` (`packages/ai/src/stream.ts`) unless noted otherwise.\n\n### Core provider credentials\n\n| Variable                        | Used for                                         | Required when                                                  | Notes / precedence                                                                                  |\n| ------------------------------- | ------------------------------------------------ | -------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- |\n| `ANTHROPIC_OAUTH_TOKEN`         | Anthropic API auth                               | Using Anthropic with OAuth token auth                          | Takes precedence over `ANTHROPIC_API_KEY` for provider auth resolution                              |\n| `ANTHROPIC_API_KEY`             | Anthropic API auth                               | Using Anthropic without OAuth token                            | Fallback after `ANTHROPIC_OAUTH_TOKEN`                                                              |\n| `ANTHROPIC_FOUNDRY_API_KEY`     | Anthropic via Azure Foundry / enterprise gateway | `CLAUDE_CODE_USE_FOUNDRY` enabled                              | Takes precedence over `ANTHROPIC_OAUTH_TOKEN` and `ANTHROPIC_API_KEY` when Foundry mode is enabled  |\n| `OPENAI_API_KEY`                | OpenAI auth                                      | Using OpenAI-family providers without explicit apiKey argument | Used by OpenAI Completions/Responses providers                                                      |\n| `GEMINI_API_KEY`                | Google Gemini auth                               | Using `google` provider models                                 | Primary key for Gemini provider mapping                                                             |\n| `GOOGLE_API_KEY`                | Gemini image tool auth fallback                  | Using `gemini_image` tool without `GEMINI_API_KEY`             | Used by coding-agent image tool fallback path                                                       |\n| `GROQ_API_KEY`                  | Groq auth                                        | Using Groq models                                              |                                                                                                     |\n| `CEREBRAS_API_KEY`              | Cerebras auth                                    | Using Cerebras models                                          |                                                                                                     |\n| `FIREWORKS_API_KEY`             | Fireworks auth                                   | Using Fireworks models                                         |                                                                                                     |\n| `FIREPASS_API_KEY`              | Fire Pass auth                                   | Using Fire Pass models                                         |                                                                                                     |\n| `TOGETHER_API_KEY`              | Together auth                                    | Using `together` provider                                      |                                                                                                     |\n| `AIMLAPI_API_KEY`               | AIML API auth                                   | Using `aimlapi` provider                                       | OpenAI-compatible AIML API endpoint at `https://api.aimlapi.com/v1`                                  |\n| `HUGGINGFACE_HUB_TOKEN`         | Hugging Face auth                                | Using `huggingface` provider                                   | Primary Hugging Face token env var                                                                  |\n| `HF_TOKEN`                      | Hugging Face auth                                | Using `huggingface` provider                                   | Fallback when `HUGGINGFACE_HUB_TOKEN` is unset                                                      |\n| `SYNTHETIC_API_KEY`             | Synthetic auth                                   | Using Synthetic models                                         |                                                                                                     |\n| `NVIDIA_API_KEY`                | NVIDIA auth                                      | Using `nvidia` provider                                        |                                                                                                     |\n| `NANO_GPT_API_KEY`              | NanoGPT auth                                     | Using `nanogpt` provider                                       |                                                                                                     |\n| `VENICE_API_KEY`                | Venice auth                                      | Using `venice` provider                                        |                                                                                                     |\n| `LITELLM_API_KEY`               | LiteLLM auth                                     | Using `litellm` provider                                       | OpenAI-compatible LiteLLM proxy key                                                                 |\n| `LM_STUDIO_API_KEY`             | LM Studio auth (optional)                        | Using `lm-studio` provider with authenticated hosts            | Local LM Studio usually runs without auth; any non-empty token works when a key is required         |\n| `OLLAMA_API_KEY`                | Ollama auth (optional)                           | Using `ollama` provider with authenticated hosts               | Local Ollama usually runs without auth; any non-empty token works when a key is required            |\n| `LLAMA_CPP_API_KEY`             | llama.cpp auth (optional)                        | Using `llama.cpp` provider with authenticated hosts            | Local llama.cpp usually runs without auth; any non-empty token works when a key is configured       |\n| `XIAOMI_API_KEY`                | Xiaomi MiMo auth                                 | Using `xiaomi` provider                                        |                                                                                                     |\n| `MOONSHOT_API_KEY`              | Moonshot auth                                    | Using `moonshot` provider                                      |                                                                                                     |\n| `XAI_API_KEY`                   | xAI auth                                         | Using xAI models or as fallback for `xai-oauth`                |                                                                                                     |\n| `XAI_OAUTH_TOKEN`               | xAI OAuth/SuperGrok auth                         | Using `xai-oauth` provider                                     | Takes precedence over `XAI_API_KEY` for `xai-oauth`                                                 |\n| `OPENROUTER_API_KEY`            | OpenRouter auth                                  | Using OpenRouter models                                        | Also used by image tool when preferred/auto provider is OpenRouter                                  |\n| `MISTRAL_API_KEY`               | Mistral auth                                     | Using Mistral models                                           |                                                                                                     |\n| `ZAI_API_KEY`                   | z.ai auth                                        | Using z.ai models                                              | Also used by z.ai web search provider                                                               |\n| `ZHIPU_API_KEY`                 | Zhipu Coding Plan auth                           | Using `zhipu-coding-plan` provider                             |                                                                                                     |\n| `MINIMAX_API_KEY`               | MiniMax auth                                     | Using `minimax` provider                                       |                                                                                                     |\n| `MINIMAX_CODE_API_KEY`          | MiniMax Code auth                                | Using `minimax-code` provider                                  |                                                                                                     |\n| `MINIMAX_CODE_CN_API_KEY`       | MiniMax Code CN auth                             | Using `minimax-code-cn` provider                               |                                                                                                     |\n| `OPENCODE_API_KEY`              | OpenCode auth                                    | Using `opencode-go` / `opencode-zen` models                    |                                                                                                     |\n| `QIANFAN_API_KEY`               | Qianfan auth                                     | Using `qianfan` provider                                       |                                                                                                     |\n| `QWEN_OAUTH_TOKEN`              | Qwen Portal auth                                 | Using `qwen-portal` with OAuth token                           | Takes precedence over `QWEN_PORTAL_API_KEY`                                                         |\n| `QWEN_PORTAL_API_KEY`           | Qwen Portal auth                                 | Using `qwen-portal` with API key                               | Fallback after `QWEN_OAUTH_TOKEN`                                                                   |\n| `ZENMUX_API_KEY`                | ZenMux auth                                      | Using `zenmux` provider                                        | Used for ZenMux OpenAI and Anthropic-compatible routes                                              |\n| `VLLM_API_KEY`                  | vLLM auth/discovery opt-in                       | Using `vllm` provider (local OpenAI-compatible servers)        | Any non-empty value works for no-auth local servers                                                 |\n| `CURSOR_ACCESS_TOKEN`           | Cursor provider auth                             | Using Cursor provider                                          |                                                                                                     |\n| `AI_GATEWAY_API_KEY`            | Vercel AI Gateway auth                           | Using `vercel-ai-gateway` provider                             |                                                                                                     |\n| `CLOUDFLARE_AI_GATEWAY_API_KEY` | Cloudflare AI Gateway auth                       | Using `cloudflare-ai-gateway` provider                         | Base URL must be configured as `https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/anthropic` |\n| `ALIBABA_CODING_PLAN_API_KEY`   | Alibaba Coding Plan auth                         | Using `alibaba-coding-plan` provider                           |                                                                                                     |\n| `DEEPSEEK_API_KEY`              | DeepSeek auth                                    | Using DeepSeek models                                          |                                                                                                     |\n| `KILO_API_KEY`                  | Kilo auth                                        | Using Kilo models                                              |                                                                                                     |\n| `OLLAMA_CLOUD_API_KEY`          | Ollama Cloud auth                                | Using `ollama-cloud` provider                                  |                                                                                                     |\n| `WAFER_PASS_API_KEY`            | Wafer Pass auth                                  | Using `wafer-pass` provider                                    | Flat-rate Wafer subscription; validated against `https://pass.wafer.ai/v1/models`                   |\n| `WAFER_SERVERLESS_API_KEY`      | Wafer Serverless auth                            | Using `wafer-serverless` provider                              | Pay-as-you-go Wafer SKU; validated against `https://pass.wafer.ai/v1/models`                        |\n| `GITLAB_TOKEN`                  | GitLab Duo auth                                  | Using `gitlab-duo` provider                                    |                                                                                                     |\n\n### GitHub/Copilot tokens\n\n| Variable               | Used for                                         | Notes                                      |\n| ---------------------- | ------------------------------------------------ | ------------------------------------------ |\n| `COPILOT_GITHUB_TOKEN` | GitHub Copilot provider auth                     | Generic GitHub tokens are not used here    |\n| `GH_TOKEN`             | GitHub API auth in web scraper                   | Web scraper fallback after `GITHUB_TOKEN`  |\n| `GITHUB_TOKEN`         | GitHub API auth in web scraper                   | Web scraper checks this before `GH_TOKEN`  |\n\n### Auth broker / auth gateway (remote credential vault)\n\nWhen the broker is enabled, the local SQLite credential store is bypassed and all OAuth refresh / access tokens live on the broker host. See [`auth-broker-gateway.md`](./auth-broker-gateway.md) for the full protocol, CLI surface, and 5-min/15-s usage cache layering.\n\n| Variable                | Used for                                                                                     | Required when                                                                                                             | Notes / precedence                                                                                                                                                                         |\n| ----------------------- | -------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |\n| `OMP_AUTH_BROKER_URL`   | Base URL of the remote auth-broker (e.g. `https://broker.tailnet:8765`); selects broker mode | Resolving credentials through a broker; also required by `omp auth-gateway serve` (the gateway is itself a broker client) | Wins over `auth.broker.url` in `config.yml`. When set with no resolvable token, `resolveAuthBrokerConfig()` hard-errors instead of falling back to local SQLite.                           |\n| `OMP_AUTH_BROKER_TOKEN` | Bearer token sent on every broker endpoint except `/v1/healthz`                              | `OMP_AUTH_BROKER_URL` is set and no token is available from `auth.broker.token` or `<config-dir>/auth-broker.token`       | Resolution: this env → `auth.broker.token` (`$ENV_NAME` indirection supported) → `<config-dir>/auth-broker.token` (mode `0600`). `<config-dir>` is `~/.omp/` (respecting `PI_CONFIG_DIR`). |\n| `OMP_AUTH_BROKER_SNAPSHOT_TTL_MS` | Freshness window for the encrypted local broker snapshot cache | Optional in broker mode | Default `3600000` (1 h). Freshness is based on broker `snapshot.generatedAt`; `0` disables cache reads/writes and forces the old blocking fetch every startup. |\n| `OMP_AUTH_BROKER_SNAPSHOT_CACHE` | Path to the encrypted local broker snapshot cache | Optional in broker mode | Defaults to `~/.omp/cache/auth-broker-snapshot.enc` (or XDG cache equivalent). Useful for tests, ephemeral hosts, or relocating the `0600` cache file. |\n\nThe gateway has no dedicated env vars — it inherits `OMP_AUTH_BROKER_*`. Its own inbound bearer token lives at `<config-dir>/auth-gateway.token` and is managed via `omp auth-gateway token`.\n\n---\n\n## 2) Provider-specific runtime configuration\n\n### Anthropic Foundry Gateway (Azure / enterprise proxy)\n\nWhen `CLAUDE_CODE_USE_FOUNDRY` is enabled, Anthropic requests switch to Foundry mode:\n\n- Base URL resolves from `FOUNDRY_BASE_URL` (fallback remains model/default base URL if unset).\n- API key resolution for provider `anthropic` becomes:\n  `ANTHROPIC_FOUNDRY_API_KEY` → `ANTHROPIC_OAUTH_TOKEN` → `ANTHROPIC_API_KEY`.\n- `ANTHROPIC_CUSTOM_HEADERS` is parsed as comma/newline-separated `key: value`\n  pairs and merged into request headers. They are also forwarded when\n  `ANTHROPIC_BASE_URL` points to a non-Anthropic host (e.g. a corporate API\n  gateway), so enterprise gateways requiring proprietary auth headers work\n  without enabling Foundry mode.\n- TLS client/server material can be injected from env values:\n  `NODE_EXTRA_CA_CERTS`, `CLAUDE_CODE_CLIENT_CERT`, `CLAUDE_CODE_CLIENT_KEY`.\n  Each accepts either:\n  - a filesystem path to PEM content, or\n  - inline PEM (including escaped `\\n` sequences).\n\n| Variable                    | Value type                                     | Behavior                                                                      |\n| --------------------------- | ---------------------------------------------- | ----------------------------------------------------------------------------- |\n| `CLAUDE_CODE_USE_FOUNDRY`   | Boolean-like string (`1`, `true`, `yes`, `on`) | Enables Foundry mode for Anthropic provider                                   |\n| `FOUNDRY_BASE_URL`          | URL string                                     | Anthropic endpoint base URL in Foundry mode                                   |\n| `ANTHROPIC_FOUNDRY_API_KEY` | Token string                                   | Used for `Authorization: Bearer <token>`                                      |\n| `ANTHROPIC_CUSTOM_HEADERS`  | Header list string                             | Extra headers; format `header-a: value, header-b: value` or newline-separated. Also forwarded outside Foundry whenever `ANTHROPIC_BASE_URL` is non-Anthropic. |\n| `NODE_EXTRA_CA_CERTS`       | PEM path or inline PEM                         | Extra CA chain for server certificate validation                              |\n| `CLAUDE_CODE_CLIENT_CERT`   | PEM path or inline PEM                         | mTLS client certificate                                                       |\n| `CLAUDE_CODE_CLIENT_KEY`    | PEM path or inline PEM                         | mTLS client private key (must be paired with cert)                            |\n\n### Amazon Bedrock\n\n| Variable                                                                        | Default / behavior                                                                            |\n| ------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------- |\n| `AWS_REGION`                                                                    | Primary region source                                                                         |\n| `AWS_DEFAULT_REGION`                                                            | Fallback if `AWS_REGION` unset                                                                |\n| `AWS_PROFILE`                                                                   | Enables named profile auth path                                                               |\n| `AWS_ACCESS_KEY_ID` + `AWS_SECRET_ACCESS_KEY`                                   | Enables IAM key auth path                                                                     |\n| `AWS_BEARER_TOKEN_BEDROCK`                                                      | Highest-precedence bearer token auth path; skips AWS profile/credential-chain lookup when set |\n| `AWS_CONTAINER_CREDENTIALS_RELATIVE_URI` / `AWS_CONTAINER_CREDENTIALS_FULL_URI` | Marks Bedrock as available in provider detection (credential resolution itself covers env keys, profiles/SSO/`credential_process`, then IMDSv2) |\n| `AWS_WEB_IDENTITY_TOKEN_FILE` + `AWS_ROLE_ARN`                                  | Marks Bedrock as available in provider detection (same caveat as the ECS variables above)     |\n| `AWS_BEDROCK_SKIP_AUTH`                                                         | If `1`, injects dummy credentials (proxy/non-auth scenarios)                                  |\n| `HTTPS_PROXY` / `HTTP_PROXY`                                                    | Honored via Bun's native fetch proxy support (the provider no longer ships an AWS SDK / proxy-agent transport) |\n| `NO_PROXY`                                                                      | Excludes matching hosts from Bun's native proxy routing                                       |\n\nRegion fallback in provider code: `options.region` → `AWS_REGION` → `AWS_DEFAULT_REGION` → `us-east-1`.\n\n### Azure OpenAI Responses\n\n| Variable                           | Default / behavior                                                          |\n| ---------------------------------- | --------------------------------------------------------------------------- |\n| `AZURE_OPENAI_API_KEY`             | Required unless API key passed as option                                    |\n| `AZURE_OPENAI_API_VERSION`         | Default `v1`                                                                |\n| `AZURE_OPENAI_BASE_URL`            | Direct base URL override                                                    |\n| `AZURE_OPENAI_RESOURCE_NAME`       | Used to construct base URL: `https://<resource>.openai.azure.com/openai/v1` |\n| `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` | Optional mapping string: `modelId=deploymentName,model2=deployment2`        |\n\nBase URL resolution: option `azureBaseUrl` → env `AZURE_OPENAI_BASE_URL` → option/env resource name → `model.baseUrl`.\n\n### Google Vertex AI\n\n| Variable                         | Required?                      | Notes                                                                                                                     |\n| -------------------------------- | ------------------------------ | ------------------------------------------------------------------------------------------------------------------------- |\n| `GOOGLE_CLOUD_PROJECT`           | Yes (unless passed in options) | Primary project ID source                                                                                                 |\n| `GCP_PROJECT`                    | Fallback                       | Alternate project ID source                                                                                               |\n| `GCLOUD_PROJECT`                 | Fallback                       | Alternate project ID source                                                                                               |\n| `GOOGLE_CLOUD_PROJECT_ID`        | OAuth login helper only        | Used by Gemini CLI OAuth project discovery                                                                                |\n| `GOOGLE_VERTEX_LOCATION`         | Yes (unless passed in options) | Primary Vertex location source                                                                                            |\n| `GOOGLE_CLOUD_LOCATION`          | Fallback                       | Alternate Vertex location source                                                                                          |\n| `VERTEX_LOCATION`                | Fallback                       | Alternate Vertex location source                                                                                          |\n| `GOOGLE_CLOUD_API_KEY`           | Conditional                    | Direct Vertex API-key auth; otherwise ADC fallback can authenticate when project and location are set                     |\n| `GOOGLE_APPLICATION_CREDENTIALS` | Conditional                    | If set, file must exist; otherwise ADC fallback path is checked (`~/.config/gcloud/application_default_credentials.json`) |\n\n### Kimi\n\n| Variable               | Default / behavior                                       |\n| ---------------------- | -------------------------------------------------------- |\n| `KIMI_CODE_OAUTH_HOST` | Primary OAuth host override                              |\n| `KIMI_OAUTH_HOST`      | Fallback OAuth host override                             |\n| `KIMI_CODE_BASE_URL`   | Overrides Kimi usage endpoint base URL (`usage/kimi.ts`) |\n\nOAuth host chain: `KIMI_CODE_OAUTH_HOST` → `KIMI_OAUTH_HOST` → `https://auth.kimi.com`.\n\n### Gemini CLI compatibility\n\n| Variable                   | Default / behavior                                              |\n| -------------------------- | --------------------------------------------------------------- |\n| `PI_AI_GEMINI_CLI_VERSION` | Overrides Gemini CLI user-agent version tag (`0.35.3` if unset) |\n\n### OpenAI Codex responses (feature/debug controls)\n\n| Variable                                   | Behavior                                             |\n| ------------------------------------------ | ---------------------------------------------------- |\n| `PI_CODEX_DEBUG`                           | `1`/`true` enables Codex provider debug logging      |\n| `PI_CODEX_WEBSOCKET`                       | `1`/`true` enables websocket transport preference    |\n| `PI_OPENAI_STATEFUL`                       | Overrides the stateful-chaining default for the platform OpenAI Responses API (`previous_response_id`, forces `store: true`): on by default against api.openai.com, off elsewhere |\n| `PI_CODEX_WEBSOCKET_IDLE_TIMEOUT_MS`       | Positive integer override (default 300000)           |\n| `PI_CODEX_WEBSOCKET_RETRY_BUDGET`          | Non-negative integer override (default 5)            |\n| `PI_CODEX_WEBSOCKET_RETRY_DELAY_MS`        | Positive integer base backoff override (default 500) |\n| `PI_OPENAI_STREAM_FIRST_EVENT_TIMEOUT_MS`  | Positive integer OpenAI first-event timeout override |\n| `PI_OPENAI_STREAM_IDLE_TIMEOUT_MS`         | Positive integer OpenAI stream idle timeout override |\n\n### Cursor provider debug\n\n| Variable           | Behavior                                                                 |\n| ------------------ | ------------------------------------------------------------------------ |\n| `DEBUG_CURSOR`     | Enables provider debug logs; `2`/`verbose` for detailed payload snippets |\n| `DEBUG_CURSOR_LOG` | Optional file path for JSONL debug log output                            |\n\n### Prompt cache compatibility switch\n\n| Variable             | Behavior                                                                                                          |\n| -------------------- | ----------------------------------------------------------------------------------------------------------------- |\n| `PI_CACHE_RETENTION` | If `long`, enables long retention where supported (`anthropic`, `openai-responses`, Bedrock retention resolution) |\n\n---\n\n## 3) Web search subsystem\n\n### Search provider credentials\n\n| Variable                                            | Used by                                                       |\n| --------------------------------------------------- | ------------------------------------------------------------- |\n| `EXA_API_KEY`                                       | Exa search provider and Exa MCP tools                         |\n| `BRAVE_API_KEY`                                     | Brave search provider                                         |\n| `PERPLEXITY_API_KEY`                                | Perplexity search provider API-key mode                       |\n| `PERPLEXITY_COOKIES`                                | Perplexity cookie-auth search mode                            |\n| `TAVILY_API_KEY`                                    | Tavily search provider                                        |\n| `ZAI_API_KEY`                                       | z.ai search provider (also checks stored OAuth in `agent.db`) |\n| `OPENAI_API_KEY` / Codex OAuth in DB                | Codex search provider availability/auth                       |\n| `PI_CODEX_WEB_SEARCH_MODEL`                         | Codex search provider model override                          |\n| `MOONSHOT_SEARCH_API_KEY` / `KIMI_SEARCH_API_KEY`   | Kimi/Moonshot search provider env auth                        |\n| `MOONSHOT_SEARCH_BASE_URL` / `KIMI_SEARCH_BASE_URL` | Kimi/Moonshot search endpoint override                        |\n| `KAGI_API_KEY`                                      | Kagi search provider                                          |\n| `JINA_API_KEY`                                      | Jina search provider                                          |\n| `PARALLEL_API_KEY`                                  | Parallel search provider                                      |\n| `SEARXNG_ENDPOINT`, `SEARXNG_TOKEN`                 | SearXNG endpoint and optional bearer token                    |\n| `SEARXNG_BASIC_USERNAME`, `SEARXNG_BASIC_PASSWORD`  | SearXNG HTTP Basic Auth credentials                           |\n\nSearXNG also reads the equivalent `searxng.endpoint`, `searxng.token`, `searxng.basicUsername`, and `searxng.basicPassword` settings from `~/.omp/agent/config.yml`; environment variables are fallbacks.\n\n### Anthropic web search auth chain\n\n`searchAnthropic()` resolves credentials in this order:\n\n1. `ANTHROPIC_SEARCH_API_KEY`\n2. `authStorage.getApiKey(\"anthropic\")` fallback credentials (runtime/config overrides, stored API-key credentials, stored OAuth credentials, then generic Anthropic env fallback: `ANTHROPIC_FOUNDRY_API_KEY` in Foundry mode, otherwise `ANTHROPIC_OAUTH_TOKEN` / `ANTHROPIC_API_KEY`)\n\nFor either credential path, base URL resolution is:\n\n1. `ANTHROPIC_SEARCH_BASE_URL`\n2. `FOUNDRY_BASE_URL` when `CLAUDE_CODE_USE_FOUNDRY` is enabled\n3. `ANTHROPIC_BASE_URL`\n4. `https://api.anthropic.com`\n\nRelated vars:\n\n| Variable                    | Default / behavior                                                                                                                                                                                |\n| --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| `ANTHROPIC_SEARCH_API_KEY`  | API key used exclusively for the Anthropic web search provider. Highest-priority search auth; overrides `ANTHROPIC_API_KEY` / OAuth / Foundry for search calls without affecting chat completions. |\n| `ANTHROPIC_SEARCH_BASE_URL` | Base URL used exclusively for the Anthropic web search provider. Applied to either `ANTHROPIC_SEARCH_API_KEY` or fallback Anthropic credentials; overrides `ANTHROPIC_BASE_URL` (and `FOUNDRY_BASE_URL` in Foundry mode) for search calls. |\n| `ANTHROPIC_SEARCH_MODEL`    | Search model override. Defaults to `claude-haiku-4-5`.                                                                                                                                             |\n| `ANTHROPIC_BASE_URL`        | Generic fallback base URL for Anthropic requests when no search-specific base URL is set.                                                                                                          |\n\nUse `ANTHROPIC_SEARCH_BASE_URL` (optionally with `ANTHROPIC_SEARCH_API_KEY`) to keep chat routed through an enterprise gateway (`ANTHROPIC_BASE_URL` or `CLAUDE_CODE_USE_FOUNDRY=true`) while pointing web search at a direct Anthropic endpoint, or vice versa.\n\n### Perplexity OAuth flow behavior flag\n\n| Variable            | Behavior                                                                        |\n| ------------------- | ------------------------------------------------------------------------------- |\n| `PI_AUTH_NO_BORROW` | If set, disables macOS native-app token borrowing path in Perplexity login flow |\n\n---\n\n## 4) Python tooling and kernel runtime\n\n| Variable                | Default / behavior                                                                                                  |\n| ----------------------- | ------------------------------------------------------------------------------------------------------------------- |\n| `PI_PY`                 | Boolean-like override for the Python eval backend: truthy (`1`/`true`/`yes`/`on`) enables, any other value disables; unset defers to the `eval.py` setting (default enabled)        |\n| `PI_JS`                 | Same boolean-like override for the JavaScript eval backend; unset defers to the `eval.js` setting (default enabled)                                                                 |\n| `PI_PYTHON_SKIP_CHECK`  | If `1`, skips Python interpreter availability checks (subprocess runner still starts on demand)                     |\n| `PI_PYTHON_INTEGRATION` | If `1`, opts gated integration tests in (e.g. `python-runner.integration.test.ts`) into running against real Python |\n| `PI_PYTHON_IPC_TRACE`   | If `1`, logs NDJSON frames exchanged with the Python runner subprocess                                              |\n| `VIRTUAL_ENV`           | Highest-priority venv path for Python runtime resolution                                                            |\n\nExtra conditional behavior:\n\n- If `BUN_ENV=test` or `NODE_ENV=test`, Python availability checks are treated as OK and warming is skipped.\n- Python env filtering denies common API keys and allows safe base vars + `LC_`, `XDG_`, `PI_` prefixes.\n\n---\n\n## 5) Agent/runtime behavior toggles\n\n| Variable                     | Default / behavior                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         |\n| ---------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| `PI_SMOL_MODEL`              | Ephemeral model-role override for `smol` (CLI `--smol` takes precedence)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |\n| `PI_SLOW_MODEL`              | Ephemeral model-role override for `slow` (CLI `--slow` takes precedence)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |\n| `PI_PLAN_MODEL`              | Ephemeral model-role override for `plan` (CLI `--plan` takes precedence)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |\n| `PI_NO_TITLE`                | If set (any non-empty value), disables auto session title generation on first user message                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |\n| `PI_TINY_DEVICE`             | ONNX execution provider for local tiny models; overrides the `providers.tinyModelDevice` setting (default: CPU; supports `cpu`, `gpu`, `metal`/`webgpu`, `auto`, `cuda`, `dml`, `coreml`, `wasm`, `webnn`, `webnn-gpu`, `webnn-cpu`, `webnn-npu`)                                                                                                                                                                                                                                                                                                                                                          |\n| `PI_TINY_DTYPE`              | ONNX quantization/precision for local tiny models; overrides the `providers.tinyModelDtype` setting (default: each model's shipped dtype, currently `q4`; supports `auto`, `fp32`, `fp16`, `q8`, `int8`, `uint8`, `q4`, `bnb4`, `q4f16`, `q2`, `q2f16`, `q1`, `q1f16`)                                                                                                                                                                                                                                                                                                                                     |\n| `PI_NO_INTERLEAVED_THINKING` | If `1`, disables Anthropic interleaved thinking budget behavior and uses output-token inflation for older thinking mode                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |\n| `NULL_PROMPT`                | If `true`, system prompt builder returns empty string                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |\n| `PI_BLOCKED_AGENT`           | Blocks a specific subagent type in task tool                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |\n| `PI_SUBPROCESS_CMD`          | Overrides subagent spawn command (`omp` / `omp.cmd` resolution bypass)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |\n| `PI_TASK_MAX_OUTPUT_BYTES`   | Max captured output bytes per subagent (default `500000`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |\n| `PI_TASK_MAX_OUTPUT_LINES`   | Max captured output lines per subagent (default `5000`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |\n| `PI_TIMING`                  | If set (any non-empty value), prints a hierarchical timing-span tree to **stderr** via `logger.printTimings()`. In interactive mode the tree prints once the agent is ready (before the TUI starts); in print mode it prints after the whole prompt batch completes. Print-mode prompts are wrapped in `print:prompt:initial` / `print:prompt:next` spans so each user message shows up as its own row. `PI_TIMING=x` exits the process with code 0 right after printing in interactive mode (use to measure cold startup only). `PI_TIMING=full` lists every module-load entry instead of just the top N. |\n| `PI_DEBUG_STARTUP`           | If set (any non-empty value), streams one synchronous `[startup] <phase>:start` / `:done` marker line to **stderr** as each startup phase begins/ends — including command-module imports (`cli:load:<name>`) and the native addon extraction/`dlopen` (`native:*`). Unlike `PI_TIMING` (which prints only once startup completes), the markers survive a hard hang: the last line on stderr names the phase the process is stuck in. Combine with `PI_TIMING` freely; markers and the span tree share the same phase names.                                                                                |\n| `PI_PACKAGE_DIR`             | Overrides package asset base dir resolution (`docs/`, `examples/`, `CHANGELOG.md`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         |\n| `PI_DISABLE_LSPMUX`          | If `1`, disables lspmux detection/integration and forces direct LSP server spawning                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |\n| `PI_RPC_EMIT_TITLE`          | Boolean-like flag enabling title events in RPC mode                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |\n| `SMITHERY_URL`               | Smithery web URL override (default `https://smithery.ai`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |\n| `SMITHERY_API_URL`           | Smithery API base URL override (default `https://api.smithery.ai`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                         |\n| `SMITHERY_API_KEY`           | Smithery API key for managed MCP auth lookup                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |\n| `PUPPETEER_EXECUTABLE_PATH`  | Browser tool Chromium executable override                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |\n| `LM_STUDIO_BASE_URL`         | Default implicit LM Studio discovery base URL override (`http://127.0.0.1:1234/v1` if unset)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |\n| `OLLAMA_BASE_URL`            | Default implicit Ollama discovery base URL override (`OLLAMA_HOST` if unset, then `http://127.0.0.1:11434`)                                                                                                                                                                                                                                                                                                                                                                                                                                     |\n| `OLLAMA_HOST`                | Ollama host used for implicit Ollama discovery when `OLLAMA_BASE_URL` is unset; accepts Ollama-style values such as `127.0.0.1:11434` or `http://host:11434`                                                                                                                                                                                                                                                                                                                                                                                     |\n| `OLLAMA_CONTEXT_LENGTH`      | Positive integer context-window override for implicit Ollama discovery; affects OMP context budgeting only and does not change Ollama's runtime `num_ctx`                                                                                                                                                                                                                                                                                                                                                                                           |\n| `LLAMA_CPP_BASE_URL`         | Default implicit Llama.cpp discovery base URL override (`http://127.0.0.1:8080` if unset)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  |\n| `PI_EDIT_VARIANT`            | Forces edit tool variant when valid (`patch`, `replace`, `hashline`, `apply_patch`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |\n| `PI_FORCE_IMAGE_PROTOCOL`    | Forces supported image protocol (`kitty`, `iterm2`/`iterm`, `sixel`, `none`) where used                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |\n| `PI_ALLOW_SIXEL_PASSTHROUGH` | Allows SIXEL passthrough when `PI_FORCE_IMAGE_PROTOCOL=sixel`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |\n| `PI_NO_PTY`                  | If `1`, disables interactive PTY path for bash tool                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |\n| `OMP_MCP_TIMEOUT_MS`         | Overrides MCP client request timeout (ms) for every MCP server. `0` disables client-side timeouts (`AbortSignal` never fires). Invalid (negative or non-numeric) values are ignored with a warning and the per-server config or default (`30000`) is used.                                                                                                                                                                                                                                                                                                                                                 |\n\n`PI_NO_PTY` is also set internally when CLI `--no-pty` is used.\n\n---\n\n## 6) Storage and config root paths\n\nThese are consumed via `@oh-my-pi/pi-utils/dirs` and affect where coding-agent stores data.\n\n| Variable              | Default / behavior                                                            |\n| --------------------- | ----------------------------------------------------------------------------- |\n| `PI_CONFIG_DIR`       | Config root dirname under home (default `.omp`)                               |\n| `PI_CODING_AGENT_DIR` | Full override for agent directory (default `~/<PI_CONFIG_DIR or .omp>/agent`) |\n| `PWD`                 | Used when matching canonical current working directory in path helpers        |\n\n---\n\n## 7) Shell/tool execution environment\n\n(From `packages/utils/src/procmgr.ts` and coding-agent bash tool integration.)\n\n| Variable                   | Behavior                                                                       |\n| -------------------------- | ------------------------------------------------------------------------------ |\n| `PI_BASH_NO_CI`            | Suppresses automatic `CI=true` injection into spawned shell env                |\n| `CLAUDE_BASH_NO_CI`        | Legacy alias fallback for `PI_BASH_NO_CI`                                      |\n| `PI_BASH_NO_LOGIN`         | Disables login-shell mode; shell args become `['-c']` instead of `['-l','-c']` |\n| `CLAUDE_BASH_NO_LOGIN`     | Legacy alias fallback for `PI_BASH_NO_LOGIN`                                   |\n| `PI_SHELL_PREFIX`          | Optional command prefix wrapper                                                |\n| `CLAUDE_CODE_SHELL_PREFIX` | Legacy alias fallback for `PI_SHELL_PREFIX`                                    |\n| `VISUAL`                   | Preferred external editor command                                              |\n| `EDITOR`                   | Fallback external editor command                                               |\n\nCurrent implementation: `PI_BASH_NO_LOGIN`/`CLAUDE_BASH_NO_LOGIN` are active; when either is set, `getShellArgs()` returns `['-c']`.\n\n---\n\n## 8) UI/theme/session detection (auto-detected env)\n\nThese are read as runtime signals; they are usually set by the terminal/OS rather than manually configured.\n\n| Variable                                                                                                           | Used for                                                  |\n| ------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------- |\n| `COLORTERM`, `TERM`, `WT_SESSION`                                                                                  | Color capability detection (theme color mode)             |\n| `COLORFGBG`                                                                                                        | Terminal background light/dark auto-detection             |\n| `TERM_PROGRAM`, `TERM_PROGRAM_VERSION`, `TERMINAL_EMULATOR`                                                        | Terminal identity in system prompt/context                |\n| `TMUX_PANE`, `CMUX_SURFACE_ID`, `KITTY_WINDOW_ID`, `TERM_SESSION_ID`, `WT_SESSION`                                 | Stable per-terminal session breadcrumb IDs                |\n| `SHELL`, `ComSpec`, `TERM_PROGRAM`, `TERM`                                                                         | System info diagnostics                                   |\n| `APPDATA`, `XDG_CONFIG_HOME`                                                                                       | lspmux config path resolution                             |\n| `HOME`                                                                                                             | Path shortening in MCP command UI                         |\n\n---\n\n## 9) TUI runtime flags (shared package, affects coding-agent UX)\n\n| Variable                  | Behavior                                                                              |\n| ------------------------- | ------------------------------------------------------------------------------------- |\n| `PI_NOTIFICATIONS`        | `off` / `0` / `false` suppress desktop notifications                                  |\n| `PI_TUI_WRITE_LOG`        | If set, logs TUI writes to file                                                       |\n| `PI_HARDWARE_CURSOR`      | If `1`, enables hardware cursor mode                                                  |\n| `PI_NO_SYNC_OUTPUT`       | If set (any non-empty value), disables DEC 2026 synchronized-output wrappers while keeping TUI autowrap guards |\n| `PI_NO_DECCARA`           | If set (truthy), disables Kitty DECCARA rectangular-SGR background fills (forces padded-string rendering) |\n| `PI_DEBUG_REDRAW`         | If `1`, enables redraw debug logging                                                  |\n| `PI_FORCE_IMAGE_PROTOCOL` | Forces terminal image protocol detection (`kitty`, `iterm2`/`iterm`, `sixel`, `none`) |\n\n---\n\n## 10) Commit generation controls\n\n| Variable                  | Behavior                                                            |\n| ------------------------- | ------------------------------------------------------------------- |\n| `PI_COMMIT_TEST_FALLBACK` | If `true` (case-insensitive), force commit fallback generation path |\n| `PI_COMMIT_NO_FALLBACK`   | If `true`, disables fallback when agent returns no proposal         |\n| `PI_COMMIT_MAP_REDUCE`    | If `false`, disables map-reduce commit analysis path                |\n| `DEBUG`                   | If set, commit agent error stack traces are printed                 |\n\n---\n\n## Security-sensitive variables\n\nTreat these as secrets; do not log or commit them:\n\n- Provider/API keys and OAuth/bearer credentials (all `*_API_KEY`, `*_TOKEN`, OAuth access/refresh tokens)\n- Cloud credentials (`AWS_*`, `GOOGLE_APPLICATION_CREDENTIALS` path may expose service-account material)\n- Search/provider auth vars (`EXA_API_KEY`, `BRAVE_API_KEY`, `PERPLEXITY_API_KEY`, Anthropic search keys)\n- Foundry mTLS material (`CLAUDE_CODE_CLIENT_CERT`, `CLAUDE_CODE_CLIENT_KEY`, `NODE_EXTRA_CA_CERTS` when it points to private CA bundles)\n\nPython runtime also explicitly strips many common key vars before spawning kernel subprocesses (`packages/coding-agent/src/eval/py/runtime.ts`).\n",
 	"extension-loading.md": "# Extension Loading (TypeScript/JavaScript Modules)\n\nThis document covers how the coding agent discovers and loads **extension modules** (`.ts`/`.js`) at startup.\n\nIt does **not** cover `gemini-extension.json` manifest extensions (documented separately).\n\n## What this subsystem does\n\nExtension loading builds a list of module entry files, imports each module with Bun, executes its factory, and returns:\n\n- loaded extension definitions\n- per-path load errors (without aborting the whole load)\n- a shared extension runtime object used later by `ExtensionRunner`\n\n## Primary implementation files\n\n- `src/extensibility/extensions/loader.ts` — path discovery + import/execution\n- `src/extensibility/extensions/index.ts` — public exports\n- `src/extensibility/extensions/runner.ts` — runtime/event execution after load\n- `src/discovery/builtin.ts` — native auto-discovery provider for extension modules\n- `src/config/settings.ts` — loads merged `extensions` / `disabledExtensions` settings\n\n---\n\n## Inputs to extension loading\n\n### 1) Auto-discovered native extension modules\n\n`discoverAndLoadExtensions()` first asks discovery providers for `extension-module` capability items, then keeps only provider `native` items.\n\nNative `extension-module` discovery comes from:\n\n- Project directory: `<cwd>/.omp/extensions`\n- User directory: `~/.omp/agent/extensions`\n- Native legacy/settings JSON entries: `<cwd>/.omp/settings.json#extensions` and `~/.omp/agent/settings.json#extensions`\n\nPath roots come from the native provider (`SOURCE_PATHS.native`). Project lookup is cwd-only for these native roots; it does not walk ancestors.\n\nNotes:\n\n- Native auto-discovery is currently `.omp` based.\n- Legacy `.pi` is still accepted in package manifests (`pi.extensions`) and project override lookup, but `.pi/extensions` is not a native root here.\n\n### 2) Installed plugin extension entries\n\nAfter native auto-discovery, `discoverAndLoadExtensions()` appends extension entry points from enabled installed plugins via `getAllPluginExtensionPaths(cwd)`.\n\nPlugin extension entries come from package `omp.extensions` / `pi.extensions` manifests, including enabled feature entries.\n\n### 3) Explicitly configured paths\n\nAfter plugin extension entries, configured paths are appended and resolved.\n\nConfigured path sources in the main session startup path (`sdk.ts`):\n\n1. CLI-provided paths (`--extension/-e`, and `--hook` is also treated as an extension path)\n2. Merged settings `extensions` array\n\nSettings files:\n\n- User: `~/.omp/agent/config.yml` (or custom agent dir via `PI_CODING_AGENT_DIR`)\n- Project/native settings capability: `<cwd>/.omp/config.yml` and `<cwd>/.omp/settings.json`\n\nNative extension-module discovery also reads legacy JSON extension lists from:\n\n- `~/.omp/agent/settings.json`\n- `<cwd>/.omp/settings.json`\n\nExamples:\n\n```yaml\n# ~/.omp/agent/config.yml\nextensions:\n  - ~/my-exts/safety.ts\n  - ./local/ext-pack\n```\n\n```json\n{\n  \"extensions\": [\"./.omp/extensions/my-extra\"]\n}\n```\n\n---\n\n## Enable/disable controls\n\n### Disable discovery\n\n- CLI: `--no-extensions`\n- SDK option: `disableExtensionDiscovery`\n\nBehavior split:\n\n- SDK: when `disableExtensionDiscovery=true`, it still loads `additionalExtensionPaths` via `loadExtensions()`.\n- CLI path building (`main.ts`) currently clears CLI extension paths when `--no-extensions` is set, so explicit `-e/--hook` are not forwarded in that mode.\n\n### Disable specific extension modules\n\n`disabledExtensions` setting filters by extension id format:\n\n- `extension-module:<derivedName>`\n\n`derivedName` is based on entry path (`getExtensionNameFromPath`), for example:\n\n- `/x/foo.ts` -> `foo`\n- `/x/bar/index.ts` -> `bar`\n\nExample:\n\n```yaml\ndisabledExtensions:\n  - extension-module:foo\n```\n\n---\n\n## Path and entry resolution\n\n### Path normalization\n\nFor configured paths:\n\n1. Normalize unicode spaces\n2. Expand `~`\n3. If relative, resolve against current `cwd`\n\n### If configured path is a file\n\nIt is used directly as a module entry candidate.\n\n### If configured path is a directory\n\nResolution order:\n\n1. `package.json` in that directory with `omp.extensions` (or legacy `pi.extensions`) -> use declared entries\n2. `index.ts`\n3. `index.js`\n4. Otherwise scan one level for extension entries:\n   - direct `*.ts` / `*.js`\n   - subdir `index.ts` / `index.js`\n   - subdir `package.json` with `omp.extensions` / `pi.extensions`\n\nRules and constraints:\n\n- no recursive discovery beyond one subdirectory level\n- declared `extensions` manifest entries are resolved relative to that package directory\n- declared entries are included only if file exists/access is allowed\n- in `*/index.{ts,js}` pairs, TypeScript is preferred over JavaScript\n- symlinks are treated as eligible files/directories\n\n### Ignore behavior differs by source\n\n- Native auto-discovery (`discoverExtensionModulePaths` in discovery helpers) uses native glob with `gitignore: true` and `hidden: false`.\n- Explicit configured directory scanning in `loader.ts` uses `readdir` rules and does **not** apply gitignore filtering.\n\n---\n\n## Load order and precedence\n\n`discoverAndLoadExtensions()` builds one ordered list and then calls `loadExtensions()`.\n\nOrder:\n\n1. Native auto-discovered modules\n2. Installed plugin extension entries\n3. Explicit configured paths (in provided order)\n\nIn `sdk.ts`, configured order is:\n\n1. CLI additional paths\n2. Settings `extensions`\n\nDe-duplication:\n\n- absolute path based\n- first seen path wins\n- later duplicates are ignored\n\nImplication: if the same module path is both auto-discovered and explicitly configured, it is loaded once at the first position (auto-discovered stage).\n\n---\n\n## Module import and factory contract\n\nEach candidate path is loaded with dynamic import:\n\n- `await import(resolvedPath)`\n- factory is `module.default ?? module`\n- factory must be a function (`ExtensionFactory`)\n\nIf export is not a function, that path fails with a structured error and loading continues.\n\n---\n\n## Failure handling and isolation\n\n### During loading\n\nPer extension path, failures are captured as `{ path, error }` and do not stop other paths from loading.\n\nCommon cases:\n\n- import failure / missing file\n- invalid factory export (non-function)\n- exception thrown while executing factory\n\n### Runtime isolation model\n\n- Extensions are **not sandboxed** (same process/runtime).\n- They share one `EventBus` and one `ExtensionRuntime` instance.\n- During load, runtime action methods intentionally throw `ExtensionRuntimeNotInitializedError`; action wiring happens later in `ExtensionRunner.initialize()`.\n\n### After loading\n\nWhen events run through `ExtensionRunner`, handler exceptions are caught and emitted as extension errors instead of crashing the runner loop.\n\n---\n\n## Minimal user/project layout examples\n\n### User-level\n\n```text\n~/.omp/agent/\n  config.yml\n  extensions/\n    guardrails.ts\n    audit/\n      index.ts\n```\n\n### Project-level\n\n```text\n<repo>/\n  .omp/\n    settings.json\n    extensions/\n      checks/\n        package.json\n      lint-gates.ts\n```\n\n`checks/package.json`:\n\n```json\n{\n  \"omp\": {\n    \"extensions\": [\"./src/check-a.ts\", \"./src/check-b.js\"]\n  }\n}\n```\n\nLegacy manifest key still accepted:\n\n```json\n{\n  \"pi\": {\n    \"extensions\": [\"./index.ts\"]\n  }\n}\n```\n",
-	"extensions.md": "# Extensions\n\nPrimary guide for authoring runtime extensions in `packages/coding-agent`.\n\nThis document covers the current extension runtime in:\n\n- `src/extensibility/extensions/types.ts`\n- `src/extensibility/extensions/runner.ts`\n- `src/extensibility/extensions/wrapper.ts`\n- `src/extensibility/extensions/index.ts`\n- `src/modes/controllers/extension-ui-controller.ts`\n\nFor discovery paths and filesystem loading rules, see [`extension-loading.md`](./extension-loading.md).\n\n## What an extension is\n\nAn extension is a TS/JS module exporting a default factory:\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function myExtension(pi: ExtensionAPI) {\n  // register handlers/tools/commands/renderers\n}\n```\n\nExtensions can combine all of the following in one module:\n\n- event handlers (`pi.on(...)`)\n- LLM-callable tools (`pi.registerTool(...)`)\n- slash commands (`pi.registerCommand(...)`)\n- keyboard shortcuts and flags\n- custom message rendering\n- session/message injection APIs (`sendMessage`, `sendUserMessage`, `appendEntry`)\n\n## Runtime model\n\n1. Extensions are imported and their factory functions run.\n2. During that load phase, registration methods are valid; runtime action methods are not yet initialized.\n3. `ExtensionRunner.initialize(...)` wires live actions/contexts for the active mode.\n4. Session/agent/tool lifecycle events are emitted to handlers.\n5. Every tool execution is wrapped with extension interception (`tool_call` / `tool_result`).\n\n```text\nExtension lifecycle (simplified)\n\nload paths\n   │\n   ▼\nimport module + run factory (registration only)\n   │\n   ▼\nExtensionRunner.initialize(mode/session/tool registry)\n   │\n   ├─ emit session/agent events to handlers\n   ├─ wrap tool execution (tool_call/tool_result)\n   └─ expose runtime actions (sendMessage, setActiveTools, ...)\n```\n\nImportant constraint from `loader.ts`:\n\n- calling action methods like `pi.sendMessage()` during extension load throws `ExtensionRuntimeNotInitializedError`\n- register first; perform runtime behavior from events/commands/tools\n\n## Quick start\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function (pi: ExtensionAPI) {\n  const { z } = pi.zod;\n\n  pi.setLabel(\"Safety + Utilities\");\n\n  pi.on(\"session_start\", async (_event, ctx) => {\n    ctx.ui.notify(`Extension loaded in ${ctx.cwd}`, \"info\");\n  });\n\n  pi.on(\"tool_call\", async (event) => {\n    if (event.toolName === \"bash\" && event.input.command?.includes(\"rm -rf\")) {\n      return { block: true, reason: \"Blocked by extension policy\" };\n    }\n  });\n\n  pi.registerTool({\n    name: \"hello_extension\",\n    label: \"Hello Extension\",\n    description: \"Return a greeting\",\n    parameters: z.object({ name: z.string() }),\n    async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {\n      return {\n        content: [{ type: \"text\", text: `Hello, ${params.name}` }],\n        details: { greeted: params.name },\n      };\n    },\n  });\n\n  pi.registerCommand(\"hello-ext\", {\n    description: \"Show queue state\",\n    handler: async (_args, ctx) => {\n      ctx.ui.notify(`pending=${ctx.hasPendingMessages()}`, \"info\");\n    },\n  });\n}\n```\n\n## Extension API surfaces\n\n## 1) Registration and actions (`ExtensionAPI`)\n\nCore methods:\n\n- `on(event, handler)`\n- `registerTool`, `registerCommand`, `registerShortcut`, `registerFlag`\n- `registerMessageRenderer`, `registerAssistantThinkingRenderer`\n- `setLabel`, `getFlag`\n- `sendMessage`, `sendUserMessage`, `appendEntry`, `exec`\n- `getActiveTools`, `getAllTools`, `setActiveTools`\n- `getCommands`\n- `getSessionName`, `setSessionName`\n- `setModel`, `getThinkingLevel`, `setThinkingLevel`\n- `registerProvider`\n- `events` (shared event bus)\n\nIn interactive mode, `input` handlers run before the built-in first-message auto-title check. Extensions that call `await pi.setSessionName(...)` from `input` can set the persisted session name and prevent the default auto-generated title from running for that session.\n\nAlso exposed:\n\n- `pi.logger`\n- `pi.typebox` (zod-backed compatibility shim for legacy TypeBox-style schemas)\n- `pi.zod` (injected `zod/v4` module — canonical for tool parameter schemas)\n- `pi.pi` (package exports)\n\n### Message delivery semantics\n\n`pi.sendMessage(message, options)` supports:\n\n- `deliverAs: \"steer\"` (default) — interrupts current run\n- `deliverAs: \"followUp\"` — queued to run after current run\n- `deliverAs: \"nextTurn\"` — stored and injected on the next user prompt\n- `triggerTurn: true` — starts a turn when idle (also honored with `deliverAs: \"nextTurn\"`: idle prompts immediately; while streaming the queued message schedules an internal continuation)\n\n`pi.sendUserMessage(content, { deliverAs })` always goes through prompt flow; while streaming it queues as steer/follow-up.\n\n## 2) Handler context (`ExtensionContext`)\n\nHandlers and tool `execute` receive `ctx` with:\n\n- `ui`\n- `hasUI`\n- `cwd`\n- `sessionManager` (read-only)\n- `modelRegistry`, `model`\n- `getContextUsage()`\n- `compact(...)`\n- `isIdle()`, `hasPendingMessages()`, `abort()`\n- `shutdown()`\n- `getSystemPrompt()`\n\n## 3) Command context (`ExtensionCommandContext`)\n\nCommand handlers additionally get:\n\n- `waitForIdle()`\n- `newSession(...)`\n- `switchSession(...)`\n- `branch(entryId)`\n- `navigateTree(targetId, { summarize })`\n- `reload()`\n\nUse command context for session-control flows; these methods are intentionally separated from general event handlers.\n\n## Event surface (current names and behavior)\n\nCanonical event unions and payload types are in `types.ts`.\n\n### Session lifecycle\n\n- `session_start`\n- `session_before_switch` / `session_switch`\n- `session_before_branch` / `session_branch`\n- `session_before_compact` / `session.compacting` / `session_compact`\n- `session_before_tree` / `session_tree`\n- `session_shutdown`\n\nCancelable pre-events:\n\n- `session_before_switch` → `{ cancel?: boolean }`\n- `session_before_branch` → `{ cancel?: boolean; skipConversationRestore?: boolean }`\n- `session_before_compact` → `{ cancel?: boolean; compaction?: CompactionResult }`\n- `session_before_tree` → `{ cancel?: boolean; summary?: { summary: string; details?: unknown } }`\n\n### Prompt and turn lifecycle\n\n- `input`\n- `before_agent_start`\n- `before_provider_request` (may replace provider request payload)\n- `after_provider_response`\n- `context`\n- `agent_start` / `agent_end`\n- `turn_start` / `turn_end`\n- `message_start` / `message_update` / `message_end`\n\n### Tool lifecycle\n\n- `tool_call` (pre-exec, may block)\n- `tool_result` (post-exec, may patch content/details/isError)\n- `tool_execution_start` / `tool_execution_update` / `tool_execution_end` (observability)\n\n`tool_result` is middleware-style: handlers run in extension order and each sees prior modifications.\n\n### Reliability/runtime signals\n\n- `auto_compaction_start` / `auto_compaction_end`\n- `auto_retry_start` / `auto_retry_end`\n- `ttsr_triggered`\n- `todo_reminder`\n- `goal_updated`\n- `credential_disabled`\n\n### User command interception\n\n- `user_bash` (override with `{ result }`)\n- `user_python` (override with `{ result }`)\n\n### `resources_discover`\n\n`resources_discover` exists in extension types and `ExtensionRunner`.\nCurrent runtime note: `ExtensionRunner.emitResourcesDiscover(...)` is implemented, but there are no `AgentSession` callsites invoking it in the current codebase.\n\n## Tool authoring details\n\n`registerTool` uses `ToolDefinition` from `types.ts`.\n\nCurrent `execute` signature:\n\n```ts\nexecute(\n\ttoolCallId,\n\tparams,\n\tsignal,\n\tonUpdate,\n\tctx,\n): Promise<AgentToolResult>\n```\n\nTemplate:\n\n```ts\nconst { z } = pi.zod;\n\npi.registerTool({\n  name: \"my_tool\",\n  label: \"My Tool\",\n  description: \"...\",\n  parameters: z.object({}),\n  hidden: false,\n  defaultInactive: false,\n  deferrable: false,\n  async execute(_id, _params, signal, onUpdate, ctx) {\n    if (signal?.aborted) {\n      return { content: [{ type: \"text\", text: \"Cancelled\" }] };\n    }\n    onUpdate?.({ content: [{ type: \"text\", text: \"Working...\" }] });\n    return { content: [{ type: \"text\", text: \"Done\" }], details: {} };\n  },\n  onSession(event, ctx) {\n    // reason: start|switch|branch|tree|shutdown\n  },\n  renderCall(args, options, theme) {\n    // optional TUI render\n  },\n  renderResult(result, options, theme, args) {\n    // optional TUI render\n  },\n});\n```\n\n`tool_call`/`tool_result` intercept all tools once the registry is wrapped in `sdk.ts`, including built-ins and extension/custom tools. `ToolDefinition` also supports optional `hidden`, `defaultInactive`, `deferrable`, `approval`, `mcpServerName`, `mcpToolName`, `renderCall`, and `renderResult` fields.\n\n## UI integration points\n\n`ctx.ui` implements the `ExtensionUIContext` interface. Support differs by mode.\n\n### Interactive mode (`extension-ui-controller.ts`)\n\nSupported:\n\n- dialogs: `select`, `confirm`, `input`, `editor`\n- input editing: `setEditorText`, `getEditorText`, `pasteToEditor`, `editor`\n- terminal title and working message (`setTitle`, `setWorkingMessage`)\n- notifications/status/editor text/terminal input/custom overlays\n- theme listing/loading by name (`setTheme` supports string names)\n- tools expanded toggle\n\nCurrent no-op methods in this controller:\n\n- `setFooter`\n- `setHeader`\n\n`setEditorComponent` is wired to the live editor (`ctx.setEditorComponent(factory)`). `setWidget` renders real widget components above or below the editor via `setHookWidget(...)` (`placement: \"aboveEditor\" | \"belowEditor\"`; string-array content capped at 10 lines).\n\n### RPC mode (`rpc-mode.ts`)\n\n`ctx.ui` is backed by RPC `extension_ui_request` events:\n\n- dialog methods (`select`, `confirm`, `input`, `editor`) round-trip to client responses\n- fire-and-forget methods emit requests (`notify`, `setStatus`, `setWidget` for string arrays, `setEditorText`; `setTitle` emits only when `PI_RPC_EMIT_TITLE=1`)\n\nUnsupported/no-op in RPC implementation:\n\n- `onTerminalInput`\n- `custom`\n- `setFooter`, `setHeader`, `setEditorComponent`\n- `setWorkingMessage`\n- theme switching/loading (`setTheme` returns failure)\n- tool expansion controls are inert\n\n### Print/headless/subagent paths\n\nWhen no UI context is supplied to runner init, `ctx.hasUI` is `false` and methods are no-op/default-returning.\n\n### ACP mode\n\nACP installs an elicitation-bridged UI context (`createAcpExtensionUiContext` in `acp-agent.ts`). `ctx.hasUI` is `true` while only `select`/`confirm`/`input` round-trip (as ACP elicitations; defaults are returned when the client lacks the `elicitation.form` capability). The non-elicitation surface (widgets, editor, theming, terminal input) is stubbed no-op.\n\n## Session and state patterns\n\nFor durable extension state:\n\n1. Persist with `pi.appendEntry(customType, data)`.\n2. Rebuild state from `ctx.sessionManager.getBranch()` on `session_start`, `session_branch`, `session_tree`.\n3. Keep tool result `details` structured when state should be visible/reconstructible from tool result history.\n\nExample reconstruction pattern:\n\n```ts\npi.on(\"session_start\", async (_event, ctx) => {\n  let latest;\n  for (const entry of ctx.sessionManager.getBranch()) {\n    if (entry.type === \"custom\" && entry.customType === \"my-state\") {\n      latest = entry.data;\n    }\n  }\n  // restore from latest\n});\n```\n\n## Rendering extension points\n\n## Custom message renderer\n\n```ts\npi.registerMessageRenderer(\"my-type\", (message, { expanded }, theme) => {\n  // return pi-tui Component\n});\n```\n\nUsed by interactive rendering when custom messages are displayed.\n\n## Assistant thinking renderer\n\n```ts\nimport { Container, Text } from \"@oh-my-pi/pi-tui\";\n\npi.registerAssistantThinkingRenderer((context, theme) => {\n  const container = new Container();\n  container.addChild(new Text(theme.fg(\"dim\", `thinking chars: ${context.text.length}`), 1, 0));\n  return container;\n});\n```\n\nUsed by interactive rendering to add display-only supplemental UI below each visible assistant thinking block. The renderer receives the already-visible thinking text, content/thinking indexes, theme, and a `requestRender()` callback for async renderers. All registered renderers that return a component are appended in registration order. Renderers must not mutate messages; the original thinking block remains the provider/session source of truth.\n\n## Tool call/result renderer\n\nProvide `renderCall` / `renderResult` on `registerTool` definitions for custom tool visualization in TUI.\n\n## Constraints and pitfalls\n\n- Runtime actions are unavailable during extension load.\n- `tool_call` errors block execution (fail-closed).\n- Command name conflicts with built-ins are skipped with diagnostics.\n- Reserved shortcuts are ignored (`ctrl+c`, `ctrl+d`, `ctrl+z`, `ctrl+k`, `ctrl+p`, `ctrl+l`, `ctrl+o`, `ctrl+t`, `ctrl+g`, `ctrl+q`, `alt+m`, `shift+tab`, `shift+ctrl+p`, `alt+enter`, `escape`, `enter`).\n- Treat `ctx.reload()` as terminal for the current command handler frame.\n\n## Extensions vs hooks vs custom-tools\n\nUse the right surface:\n\n- **Extensions** (`src/extensibility/extensions/*`): unified system (events + tools + commands + renderers + provider registration).\n- **Hooks** (`src/extensibility/hooks/*`): separate legacy event API.\n- **Custom-tools** (`src/extensibility/custom-tools/*`): tool-focused modules; when loaded alongside extensions they are adapted and still pass through extension interception wrappers.\n\nIf you need one package that owns policy, tools, command UX, and rendering together, use extensions.\n",
+	"extensions.md": "# Extensions\n\nPrimary guide for authoring runtime extensions in `packages/coding-agent`.\n\nThis document covers the current extension runtime in:\n\n- `src/extensibility/extensions/types.ts`\n- `src/extensibility/extensions/runner.ts`\n- `src/extensibility/extensions/wrapper.ts`\n- `src/extensibility/extensions/index.ts`\n- `src/modes/controllers/extension-ui-controller.ts`\n\nFor discovery paths and filesystem loading rules, see [`extension-loading.md`](./extension-loading.md).\n\n## What an extension is\n\nAn extension is a TS/JS module exporting a default factory:\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function myExtension(pi: ExtensionAPI) {\n  // register handlers/tools/commands/renderers\n}\n```\n\nExtensions can combine all of the following in one module:\n\n- event handlers (`pi.on(...)`)\n- LLM-callable tools (`pi.registerTool(...)`)\n- slash commands (`pi.registerCommand(...)`)\n- keyboard shortcuts and flags\n- custom message rendering\n- session/message injection APIs (`sendMessage`, `sendUserMessage`, `appendEntry`)\n\n## Runtime model\n\n1. Extensions are imported and their factory functions run.\n2. During that load phase, registration methods are valid; runtime action methods are not yet initialized.\n3. `ExtensionRunner.initialize(...)` wires live actions/contexts for the active mode.\n4. Session/agent/tool lifecycle events are emitted to handlers.\n5. Every tool execution is wrapped with extension interception (`tool_call` / `tool_result`).\n\n```text\nExtension lifecycle (simplified)\n\nload paths\n   │\n   ▼\nimport module + run factory (registration only)\n   │\n   ▼\nExtensionRunner.initialize(mode/session/tool registry)\n   │\n   ├─ emit session/agent events to handlers\n   ├─ wrap tool execution (tool_call/tool_result)\n   └─ expose runtime actions (sendMessage, setActiveTools, ...)\n```\n\nImportant constraint from `loader.ts`:\n\n- calling action methods like `pi.sendMessage()` during extension load throws `ExtensionRuntimeNotInitializedError`\n- register first; perform runtime behavior from events/commands/tools\n\n## Quick start\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function (pi: ExtensionAPI) {\n  const { z } = pi.zod;\n\n  pi.setLabel(\"Safety + Utilities\");\n\n  pi.on(\"session_start\", async (_event, ctx) => {\n    ctx.ui.notify(`Extension loaded in ${ctx.cwd}`, \"info\");\n  });\n\n  pi.on(\"tool_call\", async (event) => {\n    if (event.toolName === \"bash\" && event.input.command?.includes(\"rm -rf\")) {\n      return { block: true, reason: \"Blocked by extension policy\" };\n    }\n  });\n\n  pi.registerTool({\n    name: \"hello_extension\",\n    label: \"Hello Extension\",\n    description: \"Return a greeting\",\n    parameters: z.object({ name: z.string() }),\n    async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {\n      return {\n        content: [{ type: \"text\", text: `Hello, ${params.name}` }],\n        details: { greeted: params.name },\n      };\n    },\n  });\n\n  pi.registerCommand(\"hello-ext\", {\n    description: \"Show queue state\",\n    handler: async (_args, ctx) => {\n      ctx.ui.notify(`pending=${ctx.hasPendingMessages()}`, \"info\");\n    },\n  });\n}\n```\n\n## Extension API surfaces\n\n## 1) Registration and actions (`ExtensionAPI`)\n\nCore methods:\n\n- `on(event, handler)`\n- `registerTool`, `registerCommand`, `registerShortcut`, `registerFlag`\n- `registerMessageRenderer`, `registerAssistantThinkingRenderer`\n- `setLabel`, `getFlag`\n- `sendMessage`, `sendUserMessage`, `appendEntry`, `exec`\n- `getActiveTools`, `getAllTools`, `setActiveTools`\n- `getCommands`\n- `getSessionName`, `setSessionName`\n- `setModel`, `getThinkingLevel`, `setThinkingLevel`\n- `registerProvider`\n- `events` (shared event bus)\n\nIn interactive mode, `input` handlers run before the built-in first-message auto-title check. Extensions that call `await pi.setSessionName(...)` from `input` can set the persisted session name and prevent the default auto-generated title from running for that session.\n\nAlso exposed:\n\n- `pi.logger`\n- `pi.typebox` (zod-backed compatibility shim for legacy TypeBox-style schemas)\n- `pi.zod` (injected `zod/v4` module — canonical for tool parameter schemas)\n- `pi.pi` (package exports)\n\n### Message delivery semantics\n\n`pi.sendMessage(message, options)` supports:\n\n- `deliverAs: \"steer\"` (default) — interrupts current run\n- `deliverAs: \"followUp\"` — queued to run after current run\n- `deliverAs: \"nextTurn\"` — stored and injected on the next user prompt\n- `triggerTurn: true` — starts a turn when idle (also honored with `deliverAs: \"nextTurn\"`: idle prompts immediately; while streaming the queued message schedules an internal continuation)\n\n`pi.sendUserMessage(content, { deliverAs })` always goes through prompt flow; while streaming it queues as steer/follow-up.\n\n## 2) Handler context (`ExtensionContext`)\n\nHandlers and tool `execute` receive `ctx` with:\n\n- `ui`\n- `hasUI`\n- `cwd`\n- `sessionManager` (read-only)\n- `modelRegistry`, `model`\n- `models` (read-only model query — see below)\n- `getContextUsage()`\n- `compact(...)`\n- `isIdle()`, `hasPendingMessages()`, `abort()`\n- `shutdown()`\n- `getSystemPrompt()`\n\n### Model selection (`ctx.models`)\n\n`ctx.models` is a read-only facade for picking and comparing models the same way core does:\n\n- `list()` — authenticated models available this session.\n- `current()` — the live session model (read lazily, so it reflects `/model` switches).\n- `resolve(spec)` — a model string (`provider/id`, bare id) or role alias (`pi/slow`, a configured role) → `Model`, honoring the same settings-backed aliases and match preferences as `--model`. Returns `undefined` when nothing matches.\n- `family(model)` — an opaque lineage token for \"same family?\" checks (Claude point releases share a token; Claude and GPT differ). Compare it; don't persist it (the vocabulary tracks new releases).\n\n```ts\n// Pick a model from a different family than the current one (e.g. a cross-family reviewer).\nconst current = ctx.models.current();\nconst contrasting = ctx.models\n  .list()\n  .find(m => current && ctx.models.family(m) !== ctx.models.family(current));\n```\n\n## 3) Command context (`ExtensionCommandContext`)\n\nCommand handlers additionally get:\n\n- `waitForIdle()`\n- `newSession(...)`\n- `switchSession(...)`\n- `branch(entryId)`\n- `navigateTree(targetId, { summarize })`\n- `reload()`\n\nUse command context for session-control flows; these methods are intentionally separated from general event handlers.\n\n## Event surface (current names and behavior)\n\nCanonical event unions and payload types are in `types.ts`.\n\n### Session lifecycle\n\n- `session_start`\n- `session_before_switch` / `session_switch`\n- `session_before_branch` / `session_branch`\n- `session_before_compact` / `session.compacting` / `session_compact`\n- `session_before_tree` / `session_tree`\n- `session_shutdown`\n\nCancelable pre-events:\n\n- `session_before_switch` → `{ cancel?: boolean }`\n- `session_before_branch` → `{ cancel?: boolean; skipConversationRestore?: boolean }`\n- `session_before_compact` → `{ cancel?: boolean; compaction?: CompactionResult }`\n- `session_before_tree` → `{ cancel?: boolean; summary?: { summary: string; details?: unknown } }`\n\n### Prompt and turn lifecycle\n\n- `input`\n- `before_agent_start`\n- `before_provider_request` (may replace provider request payload)\n- `after_provider_response`\n- `context`\n- `agent_start` / `agent_end`\n- `turn_start` / `turn_end`\n- `message_start` / `message_update` / `message_end`\n\n### Tool lifecycle\n\n- `tool_call` (pre-exec, may block)\n- `tool_result` (post-exec, may patch content/details/isError)\n- `tool_execution_start` / `tool_execution_update` / `tool_execution_end` (observability)\n\n`tool_result` is middleware-style: handlers run in extension order and each sees prior modifications.\n\n### Reliability/runtime signals\n\n- `auto_compaction_start` / `auto_compaction_end`\n- `auto_retry_start` / `auto_retry_end`\n- `ttsr_triggered`\n- `todo_reminder`\n- `goal_updated`\n- `credential_disabled`\n\n### User command interception\n\n- `user_bash` (override with `{ result }`)\n- `user_python` (override with `{ result }`)\n\n### `resources_discover`\n\n`resources_discover` exists in extension types and `ExtensionRunner`.\nCurrent runtime note: `ExtensionRunner.emitResourcesDiscover(...)` is implemented, but there are no `AgentSession` callsites invoking it in the current codebase.\n\n## Tool authoring details\n\n`registerTool` uses `ToolDefinition` from `types.ts`.\n\nCurrent `execute` signature:\n\n```ts\nexecute(\n\ttoolCallId,\n\tparams,\n\tsignal,\n\tonUpdate,\n\tctx,\n): Promise<AgentToolResult>\n```\n\nTemplate:\n\n```ts\nconst { z } = pi.zod;\n\npi.registerTool({\n  name: \"my_tool\",\n  label: \"My Tool\",\n  description: \"...\",\n  parameters: z.object({}),\n  hidden: false,\n  defaultInactive: false,\n  deferrable: false,\n  async execute(_id, _params, signal, onUpdate, ctx) {\n    if (signal?.aborted) {\n      return { content: [{ type: \"text\", text: \"Cancelled\" }] };\n    }\n    onUpdate?.({ content: [{ type: \"text\", text: \"Working...\" }] });\n    return { content: [{ type: \"text\", text: \"Done\" }], details: {} };\n  },\n  onSession(event, ctx) {\n    // reason: start|switch|branch|tree|shutdown\n  },\n  renderCall(args, options, theme) {\n    // optional TUI render\n  },\n  renderResult(result, options, theme, args) {\n    // optional TUI render\n  },\n});\n```\n\n`tool_call`/`tool_result` intercept all tools once the registry is wrapped in `sdk.ts`, including built-ins and extension/custom tools. `ToolDefinition` also supports optional `hidden`, `defaultInactive`, `deferrable`, `approval`, `mcpServerName`, `mcpToolName`, `renderCall`, and `renderResult` fields.\n\n## UI integration points\n\n`ctx.ui` implements the `ExtensionUIContext` interface. Support differs by mode.\n\n### Interactive mode (`extension-ui-controller.ts`)\n\nSupported:\n\n- dialogs: `select`, `confirm`, `input`, `editor`\n- input editing: `setEditorText`, `getEditorText`, `pasteToEditor`, `editor`\n- terminal title and working message (`setTitle`, `setWorkingMessage`)\n- notifications/status/editor text/terminal input/custom overlays\n- theme listing/loading by name (`setTheme` supports string names)\n- tools expanded toggle\n\nCurrent no-op methods in this controller:\n\n- `setFooter`\n- `setHeader`\n\n`setEditorComponent` is wired to the live editor (`ctx.setEditorComponent(factory)`). `setWidget` renders real widget components above or below the editor via `setHookWidget(...)` (`placement: \"aboveEditor\" | \"belowEditor\"`; string-array content capped at 10 lines).\n\n### RPC mode (`rpc-mode.ts`)\n\n`ctx.ui` is backed by RPC `extension_ui_request` events:\n\n- dialog methods (`select`, `confirm`, `input`, `editor`) round-trip to client responses\n- fire-and-forget methods emit requests (`notify`, `setStatus`, `setWidget` for string arrays, `setEditorText`; `setTitle` emits only when `PI_RPC_EMIT_TITLE=1`)\n\nUnsupported/no-op in RPC implementation:\n\n- `onTerminalInput`\n- `custom`\n- `setFooter`, `setHeader`, `setEditorComponent`\n- `setWorkingMessage`\n- theme switching/loading (`setTheme` returns failure)\n- tool expansion controls are inert\n\n### Print/headless/subagent paths\n\nWhen no UI context is supplied to runner init, `ctx.hasUI` is `false` and methods are no-op/default-returning.\n\n### ACP mode\n\nACP installs an elicitation-bridged UI context (`createAcpExtensionUiContext` in `acp-agent.ts`). `ctx.hasUI` is `true` while only `select`/`confirm`/`input` round-trip (as ACP elicitations; defaults are returned when the client lacks the `elicitation.form` capability). The non-elicitation surface (widgets, editor, theming, terminal input) is stubbed no-op.\n\n## Session and state patterns\n\nFor durable extension state:\n\n1. Persist with `pi.appendEntry(customType, data)`.\n2. Rebuild state from `ctx.sessionManager.getBranch()` on `session_start`, `session_branch`, `session_tree`.\n3. Keep tool result `details` structured when state should be visible/reconstructible from tool result history.\n\nExample reconstruction pattern:\n\n```ts\npi.on(\"session_start\", async (_event, ctx) => {\n  let latest;\n  for (const entry of ctx.sessionManager.getBranch()) {\n    if (entry.type === \"custom\" && entry.customType === \"my-state\") {\n      latest = entry.data;\n    }\n  }\n  // restore from latest\n});\n```\n\n## Rendering extension points\n\n## Custom message renderer\n\n```ts\npi.registerMessageRenderer(\"my-type\", (message, { expanded }, theme) => {\n  // return pi-tui Component\n});\n```\n\nUsed by interactive rendering when custom messages are displayed.\n\n## Assistant thinking renderer\n\n```ts\nimport { Container, Text } from \"@oh-my-pi/pi-tui\";\n\npi.registerAssistantThinkingRenderer((context, theme) => {\n  const container = new Container();\n  container.addChild(new Text(theme.fg(\"dim\", `thinking chars: ${context.text.length}`), 1, 0));\n  return container;\n});\n```\n\nUsed by interactive rendering to add display-only supplemental UI below each visible assistant thinking block. The renderer receives the already-visible thinking text, content/thinking indexes, theme, and a `requestRender()` callback for async renderers. All registered renderers that return a component are appended in registration order. Renderers must not mutate messages; the original thinking block remains the provider/session source of truth.\n\n## Tool call/result renderer\n\nProvide `renderCall` / `renderResult` on `registerTool` definitions for custom tool visualization in TUI.\n\n## Constraints and pitfalls\n\n- Runtime actions are unavailable during extension load.\n- `tool_call` errors block execution (fail-closed).\n- Command name conflicts with built-ins are skipped with diagnostics.\n- Reserved shortcuts are ignored (`ctrl+c`, `ctrl+d`, `ctrl+z`, `ctrl+k`, `ctrl+p`, `ctrl+l`, `ctrl+o`, `ctrl+t`, `ctrl+g`, `ctrl+q`, `alt+m`, `shift+tab`, `shift+ctrl+p`, `alt+enter`, `escape`, `enter`).\n- Treat `ctx.reload()` as terminal for the current command handler frame.\n\n## Extensions vs hooks vs custom-tools\n\nUse the right surface:\n\n- **Extensions** (`src/extensibility/extensions/*`): unified system (events + tools + commands + renderers + provider registration).\n- **Hooks** (`src/extensibility/hooks/*`): separate legacy event API.\n- **Custom-tools** (`src/extensibility/custom-tools/*`): tool-focused modules; when loaded alongside extensions they are adapted and still pass through extension interception wrappers.\n\nIf you need one package that owns policy, tools, command UX, and rendering together, use extensions.\n",
 	"fs-scan-cache-architecture.md": "# Filesystem Scan Cache Architecture Contract\n\nThis document defines the current contract for the shared filesystem scan cache implemented in Rust (`crates/pi-natives/src/fs_cache.rs`) and consumed by native discovery/search APIs exposed to `packages/coding-agent`.\n\n## What this cache is\n\nThe cache stores full directory-scan entry lists (`GlobMatch[]`) keyed by scan scope, traversal policy, and requested metadata detail. Higher-level operations (`glob` filtering, `fuzzyFind` scoring, and cached `grep` candidate selection) run against those cached entries.\n\nPrimary goals:\n\n- avoid repeated filesystem walks for repeated discovery/search calls\n- keep consistency across native discovery/search flows when they share the same scan policy\n- allow explicit staleness recovery for empty results and explicit invalidation after file mutations\n\n## Ownership and public surface\n\n- Cache implementation and policy: `crates/pi-natives/src/fs_cache.rs`\n- Native consumers:\n  - `crates/pi-natives/src/glob.rs`\n  - `crates/pi-natives/src/fd.rs` (`fuzzyFind`)\n  - `crates/pi-natives/src/grep.rs` (cached directory mode only)\n- JS binding/export:\n  - `packages/natives/native/index.d.ts` (`invalidateFsScanCache`)\n  - `packages/natives/native/index.js`\n- Coding-agent mutation invalidation helpers:\n  - `packages/coding-agent/src/tools/fs-cache-invalidation.ts`\n\n## Cache key partitioning (hard contract)\n\nEach entry is keyed by:\n\n- canonicalized `root` directory path\n- `include_hidden` boolean\n- `use_gitignore` boolean\n- `skip_node_modules` boolean\n- `detail` (`ScanDetail::Minimal` or `ScanDetail::Full`)\n\nImplications:\n\n- Hidden and non-hidden scans do **not** share entries.\n- Gitignore-respecting and ignore-disabled scans do **not** share entries.\n- Scans that prune `node_modules` do **not** share entries with scans that include it.\n- Minimal scans (path + file type only) do **not** share entries with full scans (mtime + regular-file size metadata).\n- `follow_links` is part of `ScanOptions` used to build the walker, but is not currently part of `CacheKey`; calls that differ only by `follow_links` can share a cache entry.\n\nConsumers must pass stable semantics for hidden/gitignore/node_modules/detail behavior; changing any keyed flag creates a different cache partition.\n\n## Scan collection behavior\n\nCache population uses `ignore::WalkBuilder` configured by `include_hidden`, `use_gitignore`, `skip_node_modules`, and `follow_links`:\n\n- sorted by file path\n- `.git` is always pruned\n- `node_modules` is pruned at traversal time when `skip_node_modules=true`\n- cancellation is checked before the walk and every 128 visited entries per parallel visitor\n- `ScanDetail::Minimal` records normalized relative path and file type only\n- `ScanDetail::Full` also records mtime and regular-file size\n\nSearch roots for cache scans are resolved by `fs_cache::resolve_search_path`:\n\n- relative paths are resolved against current cwd\n- target must be an existing directory\n- root is canonicalized when possible\n\n## Freshness and eviction policy\n\nGlobal policy (environment-overridable):\n\n- `FS_SCAN_CACHE_TTL_MS` (default `1000`)\n- `FS_SCAN_EMPTY_RECHECK_MS` (default `200`)\n- `FS_SCAN_CACHE_MAX_ENTRIES` (default `16`)\n\nBehavior:\n\n- `get_or_scan(...)`\n  - if TTL is `0`: bypass cache entirely, always fresh scan (`cache_age_ms = 0`)\n  - on cache hit within TTL: return cloned cached entries + non-zero `cache_age_ms`\n  - on expired hit: evict key, rescan, store fresh entry\n- `force_rescan(..., store=false)`: remove any matching key, scan fresh, and do not repopulate cache\n- `force_rescan(..., store=true)`: remove any matching key, scan fresh, then store the new entry\n- max entry enforcement is oldest-first eviction by `created_at` after insert\n\n## Empty-result fast recheck (separate from normal hits)\n\nNormal cache hit:\n\n- a cache hit inside TTL returns cached entries and does nothing else.\n\nEmpty-result fast recheck:\n\n- this is a **caller-side** policy using `ScanResult.cache_age_ms`\n- if filtered/query result is empty and cached scan age is at least `empty_recheck_ms()`, caller performs one `force_rescan(..., store=true)` and retries\n- intended to reduce stale-negative results when files were added while the cache is still inside TTL\n\nCurrent consumers:\n\n- `glob`: rechecks when filtered matches are empty and scan age exceeds threshold\n- `fuzzyFind` (`fd.rs`): rechecks only when query is non-empty and scored matches are empty\n- `grep`: rechecks when cached directory candidate file list is empty\n\n## Consumer defaults and cache usage\n\nCache is opt-in on exposed scan/search APIs (`cache?: boolean`, default `false`).\n\nCurrent defaults in native APIs:\n\n- `glob`: `hidden=false`, `gitignore=true`, `cache=false`; `node_modules` is included only when `includeNodeModules=true` or the pattern mentions `node_modules`; full detail is used only when `sortByMtime=true`\n- `fuzzyFind`: `hidden=false`, `gitignore=true`, `cache=false`, `node_modules` is skipped, `follow_links=true`, minimal detail\n- `grep`: `hidden=true`, `gitignore=true`, `cache=false`; cached directory mode skips `node_modules` unless the glob mentions `node_modules`; minimal detail\n\nCurrent callers:\n\n- `@`-mention fuzzy file autocomplete enables cache (`fuzzyFind` with `cache: true`):\n  - `packages/tui/src/autocomplete.ts`\n- Mutation flows invalidate through `packages/coding-agent/src/tools/fs-cache-invalidation.ts`.\n- Tool-level search integration (`packages/coding-agent/src/tools/search.ts`) currently calls native `grep` with `cache: false`.\n\n## Invalidation contract\n\nNative invalidation entrypoint:\n\n- `invalidateFsScanCache(path?: string)`\n  - with `path`: remove cache entries whose root is a prefix of the target path\n  - without path: clear all scan cache entries\n\nPath handling details:\n\n- relative invalidation paths are resolved against cwd\n- invalidation attempts canonicalization\n- if target does not exist (for example after delete), fallback canonicalizes the parent and reattaches the filename when possible\n- this preserves invalidation behavior for create/delete/rename where one side may not exist\n\n## Coding-agent mutation flow responsibilities\n\nCoding-agent code must invalidate after successful filesystem mutations.\n\nCentral helpers:\n\n- `invalidateFsScanAfterWrite(path)`\n- `invalidateFsScanAfterDelete(path)`\n- `invalidateFsScanAfterRename(oldPath, newPath)` (invalidates both sides when paths differ)\n\nCurrent mutation callsites include:\n\n- `packages/coding-agent/src/tools/write.ts`\n- `packages/coding-agent/src/edit/hashline/filesystem.ts`\n- `packages/coding-agent/src/edit/modes/patch.ts`\n- `packages/coding-agent/src/edit/modes/replace.ts`\n\nRule: if a flow mutates filesystem content or location and bypasses these helpers, cache staleness bugs are expected.\n\n## Adding a new cache consumer safely\n\nWhen introducing cache use in a new scanner/search path:\n\n1. **Use stable scan policy inputs**\n   - decide hidden/gitignore/node_modules/detail semantics first\n   - pass them consistently to `get_or_scan`/`force_rescan` so cache partitions are intentional\n\n2. **Treat cache data as pre-filtered only by traversal policy**\n   - apply tool-specific filtering (glob patterns, type filters, scoring) after retrieval\n   - never assume cached entries already reflect your higher-level filters\n\n3. **Implement empty-result fast recheck only for stale-negative risk**\n   - use `scan.cache_age_ms >= empty_recheck_ms()`\n   - retry once with `force_rescan(..., store=true, ...)`\n   - keep this path separate from normal cache-hit logic\n\n4. **Respect no-cache mode explicitly**\n   - when caller disables cache, call `force_rescan(..., store=false, ...)` or use an uncached streaming walker\n   - do not populate shared cache in a no-cache request path\n\n5. **Wire mutation invalidation for any new write path**\n   - after successful write/edit/delete/rename, call the coding-agent invalidation helper\n   - for rename/move, invalidate both old and new paths\n\n6. **Do not add per-call TTL knobs**\n   - current contract is global policy only (env-configured), no per-request TTL override\n\n## Known boundaries\n\n- Cache scope is process-local in-memory (`DashMap`), not persisted across process restarts.\n- Cache stores scan entries, not final tool results.\n- `glob`/`fuzzyFind`/cached `grep` share scan entries only when key dimensions (`root`, `hidden`, `gitignore`, `skip_node_modules`, `detail`) match.\n- `.git` is always excluded at scan collection time regardless of caller options.\n",
 	"gemini-manifest-extensions.md": "# Gemini Manifest Extensions (`gemini-extension.json`)\n\nThis document covers how the coding-agent discovers and parses Gemini-style manifest extensions (`gemini-extension.json`) into the `extensions` capability.\n\nIt does **not** cover TypeScript/JavaScript extension module loading (`extensions/*.ts`, `index.ts`, `package.json omp.extensions`), which is documented in `extension-loading.md`.\n\n## Implementation files\n\n- [`packages/coding-agent/src/discovery/gemini.ts`](../packages/coding-agent/src/discovery/gemini.ts)\n- [`packages/coding-agent/src/discovery/builtin.ts`](../packages/coding-agent/src/discovery/builtin.ts)\n- [`packages/coding-agent/src/discovery/helpers.ts`](../packages/coding-agent/src/discovery/helpers.ts)\n- [`packages/coding-agent/src/capability/extension.ts`](../packages/coding-agent/src/capability/extension.ts)\n- [`packages/coding-agent/src/capability/index.ts`](../packages/coding-agent/src/capability/index.ts)\n- [`packages/coding-agent/src/extensibility/extensions/loader.ts`](../packages/coding-agent/src/extensibility/extensions/loader.ts)\n\n---\n\n## What gets discovered\n\nThe Gemini provider (`id: gemini`, priority `60`) registers an `extensions` loader that scans two fixed roots:\n\n- User: `~/.gemini/extensions`\n- Project: `<cwd>/.gemini/extensions`\n\nPath resolution is direct from `ctx.home` and `ctx.cwd` via `getUserPath()` / `getProjectPath()`.\n\nImportant scope rule: project lookup is **cwd-only**. It does not walk parent directories.\n\n---\n\n## Directory scan rules\n\nFor each root (`~/.gemini/extensions` and `<cwd>/.gemini/extensions`), discovery does:\n\n1. `readDirEntries(root)`\n2. keep only direct child directories (`entry.isDirectory()`)\n3. for each child `<name>`, attempt to read exactly:\n   - `<root>/<name>/gemini-extension.json`\n\nThere is no recursive scan beyond one directory level.\n\n### Hidden directories\n\nGemini manifest discovery does **not** filter out dot-prefixed directory names. If a hidden child directory exists and contains `gemini-extension.json`, it is considered.\n\n### Missing/unreadable files\n\nIf `gemini-extension.json` is missing or unreadable, that directory is skipped silently (no warning).\n\n---\n\n## Manifest shape (as implemented)\n\nThe capability type defines this manifest shape:\n\n```ts\ninterface ExtensionManifest {\n  name?: string;\n  description?: string;\n  mcpServers?: Record<string, Omit<MCPServer, \"name\" | \"_source\">>;\n  tools?: unknown[];\n  context?: unknown;\n}\n```\n\nDiscovery-time behavior is intentionally loose:\n\n- JSON parse success is required.\n- There is no runtime schema validation for field types/content beyond JSON syntax.\n- The parsed object is stored as `manifest` on the capability item.\n\n### Name normalization\n\n`Extension.name` is set to:\n\n1. `manifest.name` if it is not `null`/`undefined`\n2. otherwise the extension directory name\n\nNo string-type enforcement is applied here.\n\n---\n\n## Materialization into capability items\n\nA valid parsed manifest creates one `Extension` capability item:\n\n```ts\n{\n\tname: manifest.name ?? <directory-name>,\n\tpath: <extension-directory>,\n\tmanifest: <parsed-json>,\n\tlevel: \"user\" | \"project\",\n\t_source: {\n\t\tprovider: \"gemini\",\n\t\tproviderName: \"Gemini CLI\" // attached by capability registry\n\t\tpath: <absolute-manifest-path>,\n\t\tlevel: \"user\" | \"project\"\n\t}\n}\n```\n\nNotes:\n\n- `_source.path` is normalized to an absolute path by `createSourceMeta()`.\n- Registry-level capability validation for `extensions` only checks presence of `name` and `path`.\n- Manifest internals (`mcpServers`, `tools`, `context`) are not validated during discovery.\n\n---\n\n## Error handling and warning semantics\n\n### Warned\n\n- Invalid JSON in a manifest file:\n  - warning format: `Invalid JSON in <manifestPath>`\n\n### Not warned (silent skip)\n\n- `extensions` directory missing\n- child directory has no `gemini-extension.json`\n- unreadable manifest file\n- manifest JSON is syntactically valid but semantically odd/incomplete\n\nThis means partial validity is accepted: only syntactic JSON failure emits a warning.\n\n---\n\n## Precedence and deduplication with other sources\n\n`extensions` capability is aggregated across providers by the capability registry.\n\nCurrent providers for this capability:\n\n- `native` (`packages/coding-agent/src/discovery/builtin.ts`) priority `100`\n- `gemini` (`packages/coding-agent/src/discovery/gemini.ts`) priority `60`\n\nDedup key is `ext.name` (`extensionCapability.key = ext => ext.name`).\n\n### Cross-provider precedence\n\nHigher-priority provider wins on duplicate extension names.\n\n- If `native` and `gemini` both emit extension name `foo`, the native item is kept.\n- Lower-priority duplicate is retained only in `result.all` with `_shadowed = true`.\n\n### Intra-provider order effects\n\nBecause dedup is “first seen wins”, provider-local item order matters.\n\n- Gemini loader appends **user first**, then **project**.\n- Therefore, duplicate names between `~/.gemini/extensions` and `<cwd>/.gemini/extensions` keep the user entry and shadow the project entry.\n\nBy contrast, native provider builds config dir order differently (`project` then `user` in `getConfigDirs()`), so native intra-provider shadowing is the opposite direction.\n\n---\n\n## User vs project behavior summary\n\nFor Gemini manifests specifically:\n\n- Both user and project roots are scanned every load.\n- Project root is fixed to `<cwd>/.gemini/extensions` (no ancestor walk).\n- Duplicate names inside Gemini source resolve to user-first.\n- Duplicate names against higher-priority providers (notably native) lose by priority.\n\n---\n\n## Boundary: discovery metadata vs runtime extension loading\n\n`gemini-extension.json` discovery currently feeds capability metadata (`Extension` items). It does **not** directly load runnable TS/JS extension modules.\n\nRuntime module loading (`discoverAndLoadExtensions()` / `loadExtensions()`) uses the `extension-module` capability and explicit paths, and currently filters auto-discovered modules to provider `native` only.\n\nPractical implication:\n\n- Gemini manifest extensions are discoverable as capability records.\n- They are not, by themselves, executed as runtime extension modules by the extension loader pipeline.\n\nThis boundary is intentional in current implementation and explains why manifest discovery and executable module loading can diverge.\n",
 	"handoff-generation-pipeline.md": "# `/handoff` generation pipeline\n\nThis document describes how the coding-agent implements `/handoff`: trigger path, oneshot generation, session switch, context reinjection, persistence, and UI behavior.\n\n## Scope\n\nCovers:\n\n- Interactive `/handoff` command dispatch\n- `AgentSession.handoff()` lifecycle and state transitions\n- `generateHandoff(...)` request shape\n- How old/new sessions persist handoff data differently\n- UI behavior for success, cancel, and failure\n\nDoes not cover:\n\n- Generic tree navigation/branch internals\n- Non-handoff session commands (`/new`, `/fork`, `/resume`)\n\n## Implementation files\n\n- [`../src/modes/controllers/input-controller.ts`](../packages/coding-agent/src/modes/controllers/input-controller.ts)\n- [`../src/modes/controllers/command-controller.ts`](../packages/coding-agent/src/modes/controllers/command-controller.ts)\n- [`../src/session/agent-session.ts`](../packages/coding-agent/src/session/agent-session.ts)\n- [`packages/agent/src/compaction/compaction.ts`](../packages/agent/src/compaction/compaction.ts)\n- [`../src/session/session-manager.ts`](../packages/coding-agent/src/session/session-manager.ts)\n- [`../src/slash-commands/builtin-registry.ts`](../packages/coding-agent/src/slash-commands/builtin-registry.ts)\n\n## Trigger path\n\n1. `/handoff` is declared in builtin slash command metadata (`slash-commands/builtin-registry.ts`) with optional inline hint: `[focus instructions]`.\n2. In interactive input handling (`InputController`), submit text matching `/handoff` or `/handoff ...` is intercepted before normal prompt submission.\n3. The editor is cleared and `handleHandoffCommand(customInstructions?)` is called.\n4. `CommandController.handleHandoffCommand` performs a preflight guard using current entries:\n   - Counts `type === \"message\"` entries.\n   - If `< 2`, it warns: `Nothing to hand off (no messages yet)` and returns.\n\nThe same minimum-content guard exists again inside `AgentSession.handoff()` and throws if violated. This duplicates safety at both UI and session layers.\n\n## End-to-end lifecycle\n\n### 1) Start handoff generation\n\n`AgentSession.handoff(customInstructions?)`:\n\n- Reads current branch entries (`sessionManager.getBranch()`).\n- Validates minimum message count (`>= 2`).\n- Creates `#handoffAbortController` and links any caller-provided abort signal to it.\n- Resolves the current model API key through `ModelRegistry`.\n- Calls `generateHandoff(...)` with:\n  - live agent messages (`agent.state.messages`),\n  - the current model and API key,\n  - the base system prompt (`#baseSystemPrompt`),\n  - the live tool array (`agent.state.tools`),\n  - optional focus instructions,\n  - coding-agent message conversion (`convertToLlm`),\n  - provider metadata, current thinking level, and `initiatorOverride: \"agent\"`.\n\n`generateHandoff(...)` lives in `packages/agent/src/compaction/compaction.ts` next to summarization. It renders `packages/agent/src/compaction/prompts/handoff-document.md` via `renderHandoffPrompt(...)` with optional `additionalFocus`.\n\n### 2) Generate and capture output\n\n`generateHandoff(...)` converts the existing `AgentMessage[]` history to real LLM `Message[]` history, then appends one trailing agent-attributed `user` message containing the rendered handoff prompt.\n\nThe request uses `instrumentedCompleteSimple(...)` (the OTEL-instrumented `completeSimple` oneshot wrapper) directly:\n\n```ts\nawait instrumentedCompleteSimple(\n  model,\n  {\n    systemPrompt,\n    messages: requestMessages,\n    tools,\n  },\n  {\n    apiKey,\n    signal,\n    reasoning: resolveCompactionEffort(model, options.thinkingLevel),\n    toolChoice: \"none\",\n    initiatorOverride,\n    metadata,\n  },\n  { telemetry, oneshotKind: \"handoff\" },\n);\n```\n\nImportant generation properties:\n\n- The request preserves the live provider cache prefix by reusing the same system prompt, tool definitions, and real message history shape as the active agent.\n- The handoff instruction is a trailing `user` message, not a developer message, so the cached prefix remains aligned with the prior turn.\n- `toolChoice: \"none\"` prevents intentional tool dispatch.\n- The returned assistant content is filtered to text blocks and joined with `\\n`; stray tool-call blocks are ignored if a provider does not honor `toolChoice: \"none\"`.\n- `stopReason === \"error\"` throws a generation error.\n\nNo agent-loop events are used for capture. The handoff path no longer waits for `agent_end` and no longer scans the latest assistant message.\n\n### 3) Cancellation checks\n\nCancellation throws `Error(\"Handoff cancelled\")`; a completed generation with no text returns `undefined`.\n\n- caller signal aborts `#handoffAbortController`\n- `completeSimple(...)` receives the abort signal\n- aborted handoff signal or provider `AbortError` is normalized to `Error(\"Handoff cancelled\")`\n- empty generated text returns `undefined`\n\n`AgentSession.handoff()` always clears `#handoffAbortController` in `finally`.\n\n### 4) New session creation\n\nIf text was generated and not aborted:\n\n1. Flush current session writer (`sessionManager.flush()`).\n2. Cancel session-owned async jobs.\n3. Start a brand-new session with `parentSession` pointing at the previous session file when one exists.\n4. Reset in-memory agent state (`agent.reset()`).\n5. Rebind `agent.sessionId` to the new session id.\n6. Rekey/reset Hindsight and Mnemopi memory session tracking for the new session.\n7. Clear queued context arrays (`#steeringMessages`, `#followUpMessages`, `#pendingNextTurnMessages`) and any scheduled hidden next-turn generation.\n8. Reset todo reminder counter.\n\n### 5) Handoff-context injection\n\nThe generated handoff document is wrapped by coding-agent session glue and appended to the new session as a `custom_message` entry:\n\n```text\n<handoff-context>\n...handoff text...\n</handoff-context>\n\nThe above is a handoff document from a previous session. Use this context to continue the work seamlessly.\n```\n\nInsertion call:\n\n```ts\nthis.sessionManager.appendCustomMessageEntry(\n  \"handoff\",\n  handoffContent,\n  true,\n  undefined,\n  \"agent\",\n);\n```\n\nSemantics:\n\n- `customType`: `\"handoff\"`\n- `display`: `true` (visible in TUI rebuild)\n- attribution: `\"agent\"`\n- Entry type: `custom_message` (participates in LLM context)\n\n### 6) Rebuild active agent context\n\nAfter injection:\n\n1. `buildDisplaySessionContext()` resolves message list for current leaf.\n2. `agent.replaceMessages(sessionContext.messages)` makes the injected handoff message active context.\n3. Todo phases are synchronized from the new branch.\n4. Method returns `{ document: handoffText, savedPath? }`.\n\nAt this point, the active LLM context in the new session contains the injected handoff message, not the old transcript.\n\n## Persistence model: old session vs new session\n\n### Old session\n\nHandoff generation is a oneshot request, not a visible agent turn. The generated handoff text is not appended to the old session as an assistant message.\n\nResult: the original session keeps its prior transcript unchanged except for data already persisted before handoff began.\n\n### New session\n\nAfter session reset, handoff is persisted as `custom_message` with `customType: \"handoff\"`.\n\n`buildSessionContext()` converts this entry into a runtime custom/user-context message via `createCustomMessage(...)`, so it is included in future prompts from the new session.\n\nAuto-triggered handoffs can additionally write a timestamped `handoff-*.md` artifact under the session artifacts directory when `compaction.handoffSaveToDisk` is enabled. Manual `/handoff` does not write that artifact.\n\n## Controller/UI behavior\n\n`CommandController.handleHandoffCommand` behavior:\n\n- Shows a status loader: `Generating handoff… (esc to cancel)`.\n- Calls `await session.handoff(customInstructions)`.\n- If result is `undefined`: `showError(\"Handoff cancelled\")`.\n- On success:\n  - `rebuildChatFromMessages()` (loads new session context, including injected handoff)\n  - invalidates status line and editor top border\n  - reloads todos\n  - appends success chat line: `New session started with handoff context`\n- On exception:\n  - if message is `\"Handoff cancelled\"` or error name is `AbortError`: `showError(\"Handoff cancelled\")`\n  - otherwise: `showError(\"Handoff failed: <message>\")`\n- Stops the loader, restores the previous Escape handler, and requests render at end.\n\nManual `/handoff` no longer streams the generated document into chat. A cancellable loader remains visible while the oneshot request runs, and the chat is rebuilt after generation completes.\n\n## Cancellation semantics\n\n### Session-level cancellation primitive\n\n`AgentSession` exposes:\n\n- `abortHandoff()` → aborts `#handoffAbortController`\n- `isGeneratingHandoff` → true while controller exists\n\nWhen this abort path is used, the abort signal is passed to `completeSimple(...)`; `handoff()` normalizes the cancellation to `Error(\"Handoff cancelled\")`, and command controller maps it to cancellation UI.\n\n### Interactive `/handoff` path\n\nThe command controller installs a temporary Escape handler for `/handoff` while the loader is visible. Pressing Escape calls `session.abortHandoff()`, which aborts the `completeSimple(...)` request through `#handoffAbortController`.\n\n## Aborted vs failed handoff\n\nCurrent UI classification:\n\n- **Aborted/cancelled**\n  - `abortHandoff()` path triggers `\"Handoff cancelled\"`, or\n  - thrown `AbortError`\n  - UI shows `Handoff cancelled`\n- **Failed**\n  - any other thrown error from `handoff()` / `generateHandoff()` / provider request path\n  - UI shows `Handoff failed: ...`\n\nAdditional nuance: if generation completes but no text is returned, `handoff()` returns `undefined` and controller currently reports **cancelled**, not **failed**.\n\n## Short-session and minimum-content guardrails\n\nTwo guards prevent low-signal handoffs:\n\n- UI layer (`handleHandoffCommand`): warns and returns early for `< 2` message entries\n- Session layer (`handoff()`): throws the same condition as an error\n\nThis avoids creating a new session with empty/near-empty handoff context.\n\n## State transition summary\n\nHigh-level state flow:\n\n1. Interactive slash command intercepted.\n2. Preflight message-count guard.\n3. `#handoffAbortController` created (`isGeneratingHandoff = true`).\n4. `generateHandoff(...)` issues one `instrumentedCompleteSimple(...)` request with live system prompt, tools, message history, current thinking level, and trailing handoff prompt.\n5. Assistant response text blocks are joined; tool-call blocks are discarded.\n6. If missing text → return `undefined`; if aborted → cancellation error path.\n7. If present:\n   - flush old session\n   - cancel async jobs\n   - create new empty session with previous session as parent\n   - reset runtime queues/counters\n   - append `custom_message(handoff)`\n   - optionally save an auto-triggered handoff document under the session artifacts directory when `compaction.handoffSaveToDisk` is enabled\n8. Controller rebuilds chat UI and announces success.\n9. `#handoffAbortController` cleared (`isGeneratingHandoff = false`).\n\n## Known assumptions and limitations\n\n- No structural validation checks that generated markdown follows the requested section format.\n- Missing generated text is reported as cancellation in controller UX.\n- Manual handoff has no streaming visibility; a cancellable loader is shown until the UI updates after generation completes.\n- Auto-triggered handoffs can write a timestamped `handoff-*.md` artifact when `compaction.handoffSaveToDisk` is enabled; write failure is logged and does not fail the handoff.\n",
 	"hooks.md": "# Hooks\n\nThis document describes the **current hook subsystem code** in `src/extensibility/hooks/*`.\n\n## Current status in runtime\n\nThe hook package (`src/extensibility/hooks/`) is still exported and usable as an API surface, but the default CLI runtime now initializes the **extension runner** path. In current startup flow:\n\n- `--hook` is treated as an alias for `--extension` (CLI paths are merged into `additionalExtensionPaths`)\n- tools are wrapped by `ExtensionToolWrapper`, not `HookToolWrapper`\n- context transforms and lifecycle emissions go through `ExtensionRunner`\n\nSo this file documents the hook subsystem implementation itself (types/loader/runner/wrapper), including legacy behavior and constraints.\n\n## Key files\n\n- `src/extensibility/hooks/types.ts` — hook context, event types, and result contracts\n- `src/extensibility/hooks/loader.ts` — module loading and hook discovery bridge\n- `src/extensibility/hooks/runner.ts` — event dispatch, command lookup, error signaling\n- `src/extensibility/hooks/tool-wrapper.ts` — pre/post tool interception wrapper\n- `src/extensibility/hooks/index.ts` — exports/re-exports\n\n## What a hook module is\n\nA hook module must default-export a factory:\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function hook(pi: HookAPI): void {\n  pi.on(\"tool_call\", async (event, ctx) => {\n    if (\n      event.toolName === \"bash\" &&\n      String(event.input.command ?? \"\").includes(\"rm -rf\")\n    ) {\n      return { block: true, reason: \"blocked by policy\" };\n    }\n  });\n}\n```\n\nThe factory can:\n\n- register event handlers with `pi.on(...)`\n- send persistent custom messages with `pi.sendMessage(...)`\n- persist non-LLM state with `pi.appendEntry(...)`\n- register slash commands via `pi.registerCommand(...)`\n- register custom message renderers via `pi.registerMessageRenderer(...)`\n- run shell commands via `pi.exec(...)`\n- author schemas/helpers with injected `pi.zod`, `pi.typebox`, and package exports via `pi.pi`\n\n## Discovery and loading\n\n`discoverAndLoadHooks(configuredPaths, cwd)` does:\n\n1. Load discovered hooks from capability registry (`loadCapability(\"hooks\")`)\n2. Append explicitly configured paths (deduped by absolute path)\n3. Call `loadHooks(allPaths, cwd)`\n\n`loadHooks` then imports each path and expects a `default` function.\n\n### Path resolution\n\n`loader.ts` resolves hook paths as:\n\n- absolute path: used as-is\n- `~` path: expanded\n- relative path: resolved against `cwd`\n\n### Important legacy mismatch\n\nDiscovery providers for `hookCapability` still model pre/post shell-style hook files (for example `.claude/hooks/pre/*`, `.omp/.../hooks/pre/*`).\n\nThe hook loader here uses dynamic module import and requires a default JS/TS hook factory. If a discovered hook path is not importable as a module, load fails and is reported in `LoadHooksResult.errors`.\n\n## Event surfaces\n\nHook events are strongly typed in `types.ts`.\n\n### Session events\n\n- `session_start`\n- `session_before_switch` → can return `{ cancel?: boolean }`\n- `session_switch`\n- `session_before_branch` → can return `{ cancel?: boolean; skipConversationRestore?: boolean }`\n- `session_branch`\n- `session_before_compact` → can return `{ cancel?: boolean; compaction?: CompactionResult }`\n- `session.compacting` → can return `{ context?: string[]; prompt?: string; preserveData?: Record<string, unknown> }`\n- `session_compact`\n- `session_before_tree` → can return `{ cancel?: boolean; summary?: { summary: string; details?: unknown } }`\n- `session_tree`\n- `session_shutdown`\n\n### Agent/context events\n\n- `context` → can return `{ messages?: Message[] }`\n- `before_agent_start` → can return `{ message?: { customType; content; display; details } }`\n- `agent_start`\n- `agent_end`\n- `turn_start`\n- `turn_end`\n- `auto_compaction_start`\n- `auto_compaction_end`\n- `auto_retry_start`\n- `auto_retry_end`\n- `ttsr_triggered`\n- `todo_reminder`\n\n### Tool events (pre/post model)\n\n- `tool_call` (pre-execution) → can return `{ block?: boolean; reason?: string }`\n- `tool_result` (post-execution) → can return `{ content?; details?; isError? }`\n\nThis is the hook subsystem’s core pre/post interception model.\n\n```text\nHook tool interception flow\n\ntool_call handlers\n   │\n   ├─ any { block: true }? ── yes ──> throw (tool blocked)\n   │\n   └─ no\n      │\n      ▼\n   execute underlying tool\n      │\n      ├─ success ──> tool_result handlers can override { content, details }\n      │\n      └─ error   ──> emit tool_result(isError=true) then rethrow original error\n```\n\n## Execution model and mutation semantics\n\n### 1) Pre-execution: `tool_call`\n\n`HookToolWrapper.execute()` emits `tool_call` before tool execution.\n\n- if any handler returns `{ block: true }`, execution stops\n- if handler throws, wrapper fails closed and blocks execution\n- returned `reason` becomes the thrown error text\n\n### 2) Tool execution\n\nUnderlying tool executes normally if not blocked.\n\n### 3) Post-execution: `tool_result`\n\nAfter success, wrapper emits `tool_result` with:\n\n- `toolName`, `toolCallId`, `input`\n- `content`\n- `details`\n- `isError: false`\n\nIf handler returns overrides:\n\n- `content` can replace result content\n- `details` can replace result details\n\nOn tool failure, wrapper emits `tool_result` with `isError: true` and error text content, then rethrows original error.\n\n### What hooks can mutate\n\n- LLM context for a single call via `context` (`messages` replacement chain)\n- tool output content/details on successful tool calls (`tool_result` path)\n- pre-agent injected message via `before_agent_start`\n- cancellation/custom compaction/tree behavior via `session_before_*` and `session.compacting`\n\n### What hooks cannot mutate in this implementation\n\n- raw tool input parameters in-place (only block/allow on `tool_call`)\n- execution continuation after thrown tool errors (error path rethrows)\n- final success/error status in wrapper behavior (returned `isError` is typed but not applied by `HookToolWrapper`)\n\n## Ordering and conflict behavior\n\n### Discovery-level ordering\n\nCapability providers are priority-sorted (higher first). Dedupe is by capability key, first wins.\n\nFor `hooks`, capability key is `${type}:${tool}:${name}`. Shadowed duplicates from lower-priority providers are marked and excluded from effective discovered list.\n\n### Load order\n\n`discoverAndLoadHooks` builds a flat `allPaths` list, deduped by resolved absolute path, then `loadHooks` iterates in that order.\nFile order within each discovered directory depends on `readdir` output; the hook loader does not perform an additional sort.\n\n### Runtime handler order\n\nInside `HookRunner`, order is deterministic by registration sequence:\n\n1. hooks array order\n2. handler registration order per hook/event\n\nConflict behavior by event type:\n\n- `tool_call`: last returned result wins unless a handler blocks; first block short-circuits\n- `tool_result`: last returned override wins (no short-circuit)\n- `context`: chained; each handler receives prior handler’s message output\n- `before_agent_start`: first returned message is kept; later messages ignored\n- `session_before_*`: latest returned result is tracked; `cancel: true` short-circuits immediately\n- `session.compacting`: latest returned result wins\n\nCommand/renderer conflicts:\n\n- `getCommand(name)` returns first match across hooks (first loaded wins)\n- `getMessageRenderer(customType)` returns first match\n- `getRegisteredCommands()` returns all commands (no dedupe)\n\n## UI interactions (`HookContext.ui`)\n\n`HookUIContext` includes:\n\n- `select`, `confirm`, `input`, `editor`\n- `notify`\n- `setStatus`\n- `custom`\n- `setEditorText`, `getEditorText`\n- `theme` getter\n\n`ctx` includes `hasUI`, `cwd`, `sessionManager`, `modelRegistry`, current `model`, `isIdle()`, `abort()`, and `hasQueuedMessages()`.\n\nWhen running with no UI, the default no-op context behavior is:\n\n- `select/input/editor` return `undefined`\n- `confirm` returns `false`\n- `notify`, `setStatus`, `setEditorText` are no-ops\n- `getEditorText` returns `\"\"`\n\n### Status line behavior\n\nHook status text set via `ctx.ui.setStatus(key, text)` is:\n\n- stored per key\n- sorted by key name\n- sanitized (ANSI/VT escape sequences stripped; control characters mapped to spaces; repeated spaces collapsed; trimmed)\n- joined and width-truncated for display\n\n## Error propagation and fallback\n\n### Load-time\n\n- invalid module or missing default export → captured in `LoadHooksResult.errors`\n- loading continues for other hooks\n\n### Event-time\n\n`HookRunner.emit(...)` catches handler errors for most events and emits `HookError` to listeners (`hookPath`, `event`, `error`), then continues.\n\n`emitToolCall(...)` is stricter: handler errors are not swallowed there; they propagate to caller. In `HookToolWrapper`, this blocks the tool call (fail-safe).\n\n## Realistic API examples\n\n### Block unsafe bash commands\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function (pi: HookAPI): void {\n  pi.on(\"tool_call\", async (event, ctx) => {\n    if (event.toolName !== \"bash\") return;\n    const cmd = String(event.input.command ?? \"\");\n    if (!cmd.includes(\"rm -rf\")) return;\n\n    if (!ctx.hasUI) return { block: true, reason: \"rm -rf blocked (no UI)\" };\n    const ok = await ctx.ui.confirm(\"Dangerous command\", `Allow: ${cmd}`);\n    if (!ok) return { block: true, reason: \"user denied command\" };\n  });\n}\n```\n\n### Redact tool output on post-execution\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function (pi: HookAPI): void {\n  pi.on(\"tool_result\", async (event) => {\n    if (event.toolName !== \"read\" || event.isError) return;\n\n    const redacted = event.content.map((chunk) => {\n      if (chunk.type !== \"text\") return chunk;\n      return {\n        ...chunk,\n        text: chunk.text.replaceAll(/API_KEY=\\S+/g, \"API_KEY=[REDACTED]\"),\n      };\n    });\n\n    return { content: redacted };\n  });\n}\n```\n\n### Modify model context per LLM call\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function (pi: HookAPI): void {\n  pi.on(\"context\", async (event) => {\n    const filtered = event.messages.filter(\n      (msg) => !(msg.role === \"custom\" && msg.customType === \"debug-only\"),\n    );\n    return { messages: filtered };\n  });\n}\n```\n\n### Register slash command with command-safe context methods\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function (pi: HookAPI): void {\n  pi.registerCommand(\"handoff\", {\n    description: \"Create a new session with setup message\",\n    handler: async (_args, ctx) => {\n      await ctx.waitForIdle();\n      await ctx.newSession({\n        parentSession: ctx.sessionManager.getSessionFile(),\n        setup: async (sm) => {\n          sm.appendMessage({\n            role: \"user\",\n            content: [\n              { type: \"text\", text: \"Continue from prior session summary.\" },\n            ],\n            timestamp: Date.now(),\n          });\n        },\n      });\n    },\n  });\n}\n```\n\n## Export surface\n\n`src/extensibility/hooks/index.ts` and the package subpath `@oh-my-pi/pi-coding-agent/extensibility/hooks` export:\n\n- loading APIs (`discoverAndLoadHooks`, `loadHooks`)\n- runner and wrapper (`HookRunner`, `HookToolWrapper`)\n- all hook types\n- `execCommand` re-export\n\nThe package root (`@oh-my-pi/pi-coding-agent`) does not re-export `HookAPI`; import legacy hook types from the hooks subpath.\n",
 	"install-id.md": "# Install ID\n\nA persistent per-install UUID that identifies a single oh-my-pi installation across sessions. Used as a stable correlation key for server-side dedup of telemetry-style pushes (currently the auto-QA grievance flush from `report_tool_issue`).\n\n## API\n\nExported from `@oh-my-pi/pi-utils` (`packages/utils/src/dirs.ts`):\n\n| Symbol                                  | Purpose                                                                                                                           |\n| --------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------- |\n| `getInstallId(): string`                | Returns the install ID, generating and persisting one on first call. Result is cached in-process for the lifetime of the runtime. |\n| `__resetInstallIdCacheForTests(): void` | Clears the in-process cache. Test-only — MUST NOT be called from production code.                                                 |\n\nGenerated IDs are lowercase RFC 4122 UUIDs. Existing persisted values are accepted case-insensitively when they match `^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$` with the regex `i` flag, and are returned exactly as stored.\n\n## Storage\n\n- Path: `<config-root>/install-id` — i.e. `~/.omp/install-id` by default, respecting `PI_CONFIG_DIR` via `getConfigRootDir()`.\n- Format: a single UUID line (trailing `\\n`).\n- Permissions: file is created with mode `0o600`.\n- Lifecycle: independent of `~/.omp/agent/`. Wiping agent state (sessions, settings, DB) does NOT regenerate the install ID; only deleting the `install-id` file itself does.\n\n## Generation and lifecycle\n\n1. First call to `getInstallId()` reads the file. If contents parse as a valid UUID, that value is cached and returned.\n2. Otherwise the helper calls `crypto.randomUUID()` (Node's CSPRNG-backed UUID v4) to mint a new ID.\n3. The new value is written via `open(O_WRONLY | O_CREAT | O_EXCL, 0o600)`. The exclusive-create guard means two processes hitting first-call simultaneously cannot both succeed — the loser sees `EEXIST`, re-reads the winner's file, and adopts that ID.\n4. If the existing file contained non-empty garbage (failed UUID regex), it is `unlink`ed before the exclusive create so `O_EXCL` does not trip on stale data.\n5. Any other write failure (read-only FS, permission error) is swallowed: the freshly generated UUID is still cached in-memory so the rest of the process sees a stable value, and subsequent process launches will retry persistence.\n6. Subsequent in-process calls return the cached value without touching disk. Mutating the file on disk after the first call has no effect until the process restarts (or tests call `__resetInstallIdCacheForTests`).\n\n## Consumers\n\n- `packages/coding-agent/src/tools/report-tool-issue.ts` — included as `installId` in the auto-QA grievance push body so the backend can deduplicate repeated reports from the same install. See `dev.autoqaPush.*` settings and `PI_AUTO_QA_PUSH_*` env vars.\n\nNew consumers MUST treat the value as opaque and MUST NOT derive PII from it; the helper does not mix in hostname, username, or any other host-identifying entropy.\n\n## See also\n\n- [environment-variables.md](environment-variables.md) — `PI_CONFIG_DIR` controls where `install-id` lives.\n- [config-usage.md](config-usage.md) — broader config-root layout.\n",
-	"keybindings.md": "# Keybindings\n\nRun `/hotkeys` inside an `omp` session to see the active chords for your current build. The list reflects any remaps loaded from disk and any bindings added by extensions.\n\n## Customize keybindings\n\nUser remaps live in `~/.omp/agent/keybindings.yml`. The file is a YAML mapping whose keys are keybinding action IDs and whose values are either one chord string or an array of chord strings. It is not read from `~/.omp/agent/config.yml`, and there is no nested `keybindings` object.\n\n```yaml\napp.model.cycleForward: Ctrl+P\napp.model.selectTemporary: Alt+P\napp.plan.toggle: Alt+Shift+P\n```\n\nChord names are case-insensitive and use the same notation shown in the UI, such as `Ctrl+P`, `Alt+Shift+P`, `Shift+Enter`, and `Ctrl+Backspace`.\n\nSet an action to an empty array to disable it:\n\n```yaml\napp.stt.toggle: []\n```\n\n## Common action IDs\n\n| Action ID                   | Default                                | Meaning                                       |\n| --------------------------- | -------------------------------------- | --------------------------------------------- |\n| `app.model.cycleForward`    | `Ctrl+P`                               | Cycle role models forward                     |\n| `app.model.cycleBackward`   | `Shift+Ctrl+P`                         | Cycle role models backward                    |\n| `app.model.selectTemporary` | `Alt+P`                                | Pick a model temporarily for this session     |\n| `app.model.select`          | `Alt+M`                                | Open the model selector and set roles         |\n| `app.plan.toggle`           | `Alt+Shift+P`                          | Toggle plan mode                              |\n| `app.history.search`        | `Ctrl+R`                               | Search prompt history                         |\n| `app.tools.expand`          | `Ctrl+O`                               | Toggle tool-output expansion                  |\n| `app.thinking.toggle`       | `Ctrl+T`                               | Toggle thinking-block visibility              |\n| `app.thinking.cycle`        | `Shift+Tab`                            | Cycle thinking level                          |\n| `app.editor.external`       | `Ctrl+G`                               | Edit the draft in `$VISUAL` / `$EDITOR`       |\n| `app.message.followUp`      | `Ctrl+Q`, `Ctrl+Enter`                 | Queue a follow-up message                     |\n| `app.message.dequeue`       | `Alt+Up`                               | Dequeue a queued message back into the editor |\n| `app.display.reset`         | `Ctrl+L`                               | Reset terminal display                        |\n| `app.clipboard.copyLine`    | `Alt+Shift+L`                          | Copy the current line                         |\n| `app.clipboard.copyPrompt`  | `Alt+Shift+C`                          | Copy the whole prompt                         |\n| `app.clipboard.pasteImage`  | `Ctrl+V` (`Alt+V` fallback on Windows) | Paste from the clipboard (image preferred, text fallback) |\n| `app.stt.toggle`            | `Alt+H`                                | Toggle speech-to-text recording               |\n\nOn Windows Terminal, `Ctrl+V` may be handled by the terminal paste command before `omp` sees it; use the `Alt+V` fallback when clipboard image paste appears to do nothing. When the clipboard holds no image, `app.clipboard.pasteImage` pastes the clipboard text instead, so hosts that deliver only this chord (VS Code's integrated terminal when configured to forward `Ctrl+V`, Windows clipboard history via `Win+V`) work for both payload kinds. Windows Terminal also swallows `Ctrl+Enter`, so the follow-up shortcut also binds `Ctrl+Q` — the same chord GitHub Copilot CLI uses. If your existing `keybindings.yml` already assigns `Ctrl+Q` to another action, that user remap wins and follow-up keeps `Ctrl+Enter` unless you explicitly bind `app.message.followUp`.\n\nTerminals that implement OSC 5522 enhanced paste can send clipboard MIME data directly to `omp`; image pastes are attached as `[Image #N]`, while text/plain paste events keep normal paste behavior. When OSC 5522 is unavailable, bracketed paste still handles text, and a pasted single image-file path is loaded as an image when the file is readable from the `omp` host.\n\nOlder unqualified action names are migrated when `keybindings.yml` is loaded, but new docs and new configs should use the namespaced action IDs above. Existing `keybindings.json` files are still accepted and migrated to `keybindings.yml`; `keybindings.yaml` is also accepted.\n",
+	"keybindings.md": "# Keybindings\n\nRun `/hotkeys` inside an `omp` session to see the active chords for your current build. The list reflects any remaps loaded from disk and any bindings added by extensions.\n\n## Customize keybindings\n\nUser remaps live in `~/.omp/agent/keybindings.yml`. The file is a YAML mapping whose keys are keybinding action IDs and whose values are either one chord string or an array of chord strings. It is not read from `~/.omp/agent/config.yml`, and there is no nested `keybindings` object.\n\n```yaml\napp.model.cycleForward: Ctrl+P\napp.model.selectTemporary: Alt+P\napp.plan.toggle: Alt+Shift+P\n```\n\nChord names are case-insensitive and use the same notation shown in the UI, such as `Ctrl+P`, `Alt+Shift+P`, `Shift+Enter`, and `Ctrl+Backspace`.\n\nSet an action to an empty array to disable it:\n\n```yaml\napp.history.search: []\n```\n\n## Common action IDs\n\n| Action ID                   | Default                                | Meaning                                       |\n| --------------------------- | -------------------------------------- | --------------------------------------------- |\n| `app.model.cycleForward`    | `Ctrl+P`                               | Cycle role models forward                     |\n| `app.model.cycleBackward`   | `Shift+Ctrl+P`                         | Cycle role models backward                    |\n| `app.model.selectTemporary` | `Alt+P`                                | Pick a model temporarily for this session     |\n| `app.model.select`          | `Alt+M`                                | Open the model selector and set roles         |\n| `app.plan.toggle`           | `Alt+Shift+P`                          | Toggle plan mode                              |\n| `app.history.search`        | `Ctrl+R`                               | Search prompt history                         |\n| `app.tools.expand`          | `Ctrl+O`                               | Toggle tool-output expansion                  |\n| `app.thinking.toggle`       | `Ctrl+T`                               | Toggle thinking-block visibility              |\n| `app.thinking.cycle`        | `Shift+Tab`                            | Cycle thinking level                          |\n| `app.editor.external`       | `Ctrl+G`                               | Edit the draft in `$VISUAL` / `$EDITOR`       |\n| `app.message.followUp`      | `Ctrl+Q`, `Ctrl+Enter`                 | Queue a follow-up message                     |\n| `app.message.dequeue`       | `Alt+Up`                               | Dequeue a queued message back into the editor |\n| `app.display.reset`         | `Ctrl+L`                               | Reset terminal display                        |\n| `app.clipboard.copyLine`    | `Alt+Shift+L`                          | Copy the current line                         |\n| `app.clipboard.copyPrompt`  | `Alt+Shift+C`                          | Copy the whole prompt                         |\n| `app.clipboard.pasteImage`  | `Ctrl+V` (`Alt+V` fallback on Windows) | Paste from the clipboard (image preferred, text fallback) |\n| `app.stt.toggle`            | Unbound (hold `Space`)                 | Toggle speech-to-text. By default there is no key chord — hold the space bar to record (push-to-talk) and release to transcribe; bind a chord here for a press-to-toggle alternative |\n\nOn Windows Terminal, `Ctrl+V` may be handled by the terminal paste command before `omp` sees it; use the `Alt+V` fallback when clipboard image paste appears to do nothing. When the clipboard holds no image, `app.clipboard.pasteImage` pastes the clipboard text instead, so hosts that deliver only this chord (VS Code's integrated terminal when configured to forward `Ctrl+V`, Windows clipboard history via `Win+V`) work for both payload kinds. Windows Terminal also swallows `Ctrl+Enter`, so the follow-up shortcut also binds `Ctrl+Q` — the same chord GitHub Copilot CLI uses. If your existing `keybindings.yml` already assigns `Ctrl+Q` to another action, that user remap wins and follow-up keeps `Ctrl+Enter` unless you explicitly bind `app.message.followUp`.\n\nTerminals that implement OSC 5522 enhanced paste can send clipboard MIME data directly to `omp`; image pastes are attached as `[Image #N]`, while text/plain paste events keep normal paste behavior. When OSC 5522 is unavailable, bracketed paste still handles text, and a pasted single image-file path is loaded as an image when the file is readable from the `omp` host.\n\nOlder unqualified action names are migrated when `keybindings.yml` is loaded, but new docs and new configs should use the namespaced action IDs above. Existing `keybindings.json` files are still accepted and migrated to `keybindings.yml`; `keybindings.yaml` is also accepted.\n",
 	"local-models.md": "# Embedded Local Tiny-Model Experiments\n\nThis document summarizes the experiments behind the optional **local** tiny-model paths for\nsession-title generation (`providers.tinyModel`), Mnemopi memory extraction/consolidation\n(`providers.memoryModel`), and the `auto` thinking-level difficulty classifier\n(`providers.autoThinkingModel`, which reuses the memory-model registry). It is a factual engineering\nrecord for maintainers: what we measured, which recipes won, and which models we shipped. All three\nsettings default to `online`, so existing users incur no downloads or on-device inference cost unless\nthey opt in.\n\n## Runtime / environment findings\n\n- **Stack**: `@huggingface/transformers` (transformers.js) v4 running under Bun. In Bun the library\n  loads the **native `onnxruntime-node` backend** (not the WASM build).\n- **Device policy**: local tiny models default to CPU-only inference and retry once on CPU if an\n  explicit accelerated provider cannot initialize.\n  - Pick a provider persistently with the `providers.tinyModelDevice` setting (`default` keeps CPU),\n    or per-run with the `PI_TINY_DEVICE` env var (which overrides the setting).\n  - Accepted values are `cpu`, `gpu`, `metal`/`webgpu`, `auto`, `cuda`, `dml`, `coreml`, `wasm`,\n    `webnn`, `webnn-gpu`, `webnn-cpu`, and `webnn-npu`.\n  - Direct `coreml` remains opt-in via `PI_TINY_DEVICE=coreml`; it is not part of the default because\n    cached decoder-LLM ONNX loads can fail during session initialization.\n  - WebGPU/Metal works for the single-process eval harness, but the production worker forces\n    Darwin `gpu`/`webgpu`/`auto` requests back to CPU because ONNX Runtime/Bun currently\n    hard-crashes on worker teardown after WebGPU inference.\n  - Use `providers.tinyModelDevice` or `PI_TINY_DEVICE` only when explicitly opting out of the CPU\n    default.\n- **Quantization: q4 is the sweet spot** — smaller on disk, faster to load, and fast at inference.\n  q8/int8 loads slower _and_ infers slower on CPU. Every shipped model defaults to `q4`; override the\n  precision persistently with the `providers.tinyModelDtype` setting (`default` keeps `q4`, e.g. `fp16`\n  for higher fidelity), or per-run with `PI_TINY_DTYPE` (which overrides the setting). Accepts `auto`,\n  `fp32`, `fp16`, `q8`, `int8`, `uint8`, `q4`, `bnb4`, `q4f16`, `q2`, `q2f16`, `q1`, `q1f16`; an\n  unrecognized value fails loudly at worker startup.\n- **Load-time correction (important).** An earlier belief that \"q4 >=1B models take minutes to load\"\n  was a **measurement artifact** caused by running ~5 multi-GB HuggingFace downloads in parallel\n  (I/O saturation). Clean, isolated **warm** loads are all sub-3s:\n  - TinyLlama-1.1B q4: ~0.5s\n  - Llama-3.2-1B q4: ~2.8s (`graphOpt=all`) / ~0.5s (`disabled`)\n  - LFM2-1.2B q4: ~0.36s\n  - Qwen2.5-1.5B q4: ~1.5s\n  - Qwen3-1.7B q4: ~1.6s\n  - gemma-3-1b q4: ~1.1s\n  - Conclusion: **1B–1.7B models are viable on CPU.**\n- **`session_options.graphOptimizationLevel`** trades load vs inference speed: `disabled` = fastest\n  load, slightly slower inference; `all` = default.\n- **First run** downloads weights from the HF Hub to a cache dir (q4 weights ~200MB–1.1GB depending\n  on model); subsequent **warm** loads are sub-second to ~3s. Inference is async and\n  background-friendly for memory tasks; titles are semi-interactive.\n\n## Task 1: Session title generation (`providers.tinyModel`)\n\n**Task**: turn the first user message into a 3–6 word title. Tiny models (sub-1B) suffice.\n\n**Winning recipe**:\n\n- Plain system prompt (no few-shot).\n- **Prefill** the assistant turn with `<title>` and **stop at `</title>`**, then take the first line.\n- Greedy decoding (`do_sample:false`), `enable_thinking:false` in the chat template.\n\n**What we learned**:\n\n- **Few-shot examples HURT sub-0.6B models** for titles; the tag-prefill rescues even 270M models.\n- **Token biasing (`bad_words_ids`) is a confirmed no-op** here — the prefill already controls the\n  opener.\n\n**Leaderboard** (tag trick, CPU, warm):\n\n| Model         | Verdict                             |\n| ------------- | ----------------------------------- |\n| LFM2-350M     | Best speed/quality balance (~212MB) |\n| Qwen3-0.6B    | Most robust                         |\n| gemma-3-270m  | Smallest viable                     |\n| Qwen2.5-0.5B  | Acceptable                          |\n| SmolLM2-135M  | Too small                           |\n| flan-t5-small | Rejected — just echoes the input    |\n\n**Shipped local options**: `lfm2-350m`, `qwen3-0.6b`, `gemma-270m`, `qwen2.5-0.5b`, `lfm2-700m`.\n**Default**: `online` (pi/smol).\n\n## Task 2: Mnemopi memory (`providers.memoryModel`)\n\nMnemopi runs two small-LLM tasks:\n\n1. **Extraction** — pull durable, structured items from a single message.\n2. **Consolidation** — summarize a list of memories into 1–3 faithful sentences.\n\nThese need **bigger models than titles: 1B–1.7B**. We tested LFM2-1.2B, Qwen2.5-1.5B, Qwen3-1.7B,\nand gemma-3-1b (q4, CPU) via four parallel agents each running 27–31 experiments.\n\n### Extraction findings\n\nThe stock 5-category JSON prompt fails on small models in two ways:\n\n1. The all-empty example `{\"facts\":[],...}` gets **copied verbatim** → 0 facts extracted.\n2. Capable models emit **JSON objects inside arrays**, which Mnemopi's `String(item)` coerces into\n   the literal string `[object Object]`.\n\nThe robust fix is a **one-item-per-line output format** (consumed by Mnemopi's parser line-fallback)\nor a **flat JSON array of strings**. Every model also over-extracts pure small talk; an explicit\nchit-chat → NONE example is the best mitigation.\n\n### Technique polarity flips vs titles\n\n- At 1B+, **few-shot is the dominant quality lever**: e.g. Qwen2.5-1.5B extraction F1 0.52 → 0.83\n  going 1 → 3 shots; gemma recall 0.65 → 0.92 with 2 shots.\n- **Prefill HURTS extraction** — it forces output on small talk, producing false positives.\n- **System-split** (instructions in the system role) helps models that have a system role.\n- **Greedy >= temperature** for both tasks.\n- **Token biasing** is again a no-op.\n\n### Per-model verdicts (head-to-head, 16-fixture set)\n\n- **Qwen3-1.7B** — most disciplined extraction: returns empty on small talk, no buried-fact leak,\n  preserves language, clean flat JSON. Weaknesses: coarse granularity, missed a multi-turn value\n  update.\n- **Qwen2.5-1.5B** — best extraction granularity (atomic facts), caught the value update, zero\n  small-talk leakage. Weaknesses: weakest consolidation (run-on, no dedup) and one degenerate\n  buried-fact output.\n- **gemma-3-1b** — best consolidation (dedup works, faithful, clean single-memory). Weaknesses: leaks\n  small talk and translated German.\n- **LFM2-1.2B** — solid and fastest to load. Weaknesses: `Label: value` noise, small-talk + buried\n  leaks, a fluffy single-memory summary.\n\n### Recommendation\n\nExtraction favors **precision** (do not pollute long-term memory) → **Qwen3-1.7B is the best single\npick** (its consolidation is good enough). If running a second model for consolidation, **gemma-3-1b**\nwins that task.\n\n**Shipped local options**: `qwen3-1.7b` (recommended), `gemma-3-1b`, `qwen2.5-1.5b`, `lfm2-1.2b`.\n**Default**: `online` (the configured smol model).\n\n### Known Mnemopi parser bugs (surfaced by these experiments)\n\n- `String(item)` produces `[object Object]` on object array items.\n- The line-fallback drops items `<=10` chars, so a correct short fact like `Name: Can` is discarded.\n\n\n## Integration notes\n\n- `providers.tinyModel`, `providers.memoryModel`, and `providers.autoThinkingModel` default to\n  `online`, so existing users get **no downloads or on-device inference cost** unless they opt in.\n- Local inference runs **in a worker** (off the main thread); models are cached on disk and\n  downloaded on first use.\n- The memory local path applies the refined recipes (line-format + small-talk-guarded extraction\n  prompt, hardened consolidation prompt) via Mnemopi prompt overrides; the **online path is\n  unchanged**.\n- `providers.autoThinkingModel` uses the same shipped local options as `providers.memoryModel`.\n",
 	"lsp-config.md": "# LSP configuration in OMP\n\nThis guide explains how to configure language servers for the OMP coding agent.\n\nSource of truth in code:\n\n- Server config type: `packages/coding-agent/src/lsp/types.ts` (`ServerConfig`)\n- Config loader: `packages/coding-agent/src/lsp/config.ts`\n- Built-in server definitions: `packages/coding-agent/src/lsp/defaults.json`\n\n## Auto-detection\n\nWhen no LSP config file is present, OMP auto-detects servers by intersecting two conditions:\n\n1. The project directory contains at least one of the server's `rootMarkers`.\n2. The server binary is available — checked in project-local bin directories first (e.g., `node_modules/.bin/`, `.venv/bin/`), then `$PATH`.\n\nNo configuration is required for common setups. The built-in server list covers most popular languages; see [`defaults.json`](../packages/coding-agent/src/lsp/defaults.json) for the full set.\n\n## Config file locations\n\nOMP merges LSP config from multiple files, lowest to highest priority:\n\n| Priority    | Location                                                                                                                    |\n| ----------- | --------------------------------------------------------------------------------------------------------------------------- |\n| 5 (lowest)  | `~/lsp.json`, `~/.lsp.json`, `~/lsp.yaml`, `~/.lsp.yaml`, `~/lsp.yml`, `~/.lsp.yml`                                         |\n| 4           | Plugin LSP configs (marketplace / `--plugin-dir` roots)                                                                     |\n| 3           | User config dirs: `~/.omp/agent/lsp.*`, `~/.claude/lsp.*`, `~/.codex/lsp.*`, `~/.gemini/lsp.*`                              |\n| 2           | Project config dirs: `<project>/.omp/lsp.*`, `<project>/.claude/lsp.*`, `<project>/.codex/lsp.*`, `<project>/.gemini/lsp.*` |\n| 1 (highest) | Project root: `<project>/lsp.*` and `<project>/.lsp.*`                                                                      |\n\nEach location accepts `.json`, `.yaml`, and `.yml` variants, including hidden-file versions (`.lsp.json`, `.lsp.yaml`, `.lsp.yml`). Files are merged in order: higher-priority files override lower-priority fields for the same server. Servers not mentioned in any override file remain at their built-in defaults.\n\n**Recommended locations:**\n\n- User-wide preferences → `~/.omp/agent/lsp.json`\n- Project-specific overrides → `<project>/.omp/lsp.json`\n\n> **Note:** Auto-detection is skipped only when at least one config file contributes server overrides. A config file that only sets `idleTimeoutMs` still lets OMP auto-detect built-in servers. When server overrides exist, OMP merges them with defaults and then loads servers that have matching `rootMarkers`, an available binary, and are not explicitly `disabled`.\n\n## File shape\n\nBoth JSON and YAML are accepted. The top-level object can use either a `servers` wrapper key or a flat map directly:\n\n```json\n{\n  \"servers\": {\n    \"server-name\": { ... }\n  },\n  \"idleTimeoutMs\": 300000\n}\n```\n\nor (flat, without the `servers` wrapper):\n\n```json\n{\n  \"server-name\": { ... },\n  \"idleTimeoutMs\": 300000\n}\n```\n\nTop-level keys:\n\n- `servers` — map of server name to `ServerConfig` (optional wrapper; flat form is equivalent)\n- `idleTimeoutMs` — shut down idle language servers after this many milliseconds; disabled by default\n\n## ServerConfig fields\n\n| Field             | Type       | Required | Description                                                                                                      |\n| ----------------- | ---------- | -------- | ---------------------------------------------------------------------------------------------------------------- |\n| `command`         | `string`   | yes      | Binary name (resolved via PATH/local bins) or absolute path                                                      |\n| `args`            | `string[]` | no       | Arguments passed to the binary                                                                                   |\n| `fileTypes`       | `string[]` | yes      | File extensions this server handles, e.g. `[\".ts\", \".tsx\"]`                                                      |\n| `rootMarkers`     | `string[]` | yes      | Files/dirs that indicate a project root; glob patterns (e.g. `*.cabal`) are supported                            |\n| `initOptions`     | `object`   | no       | Sent as `initializationOptions` during LSP handshake                                                             |\n| `settings`        | `object`   | no       | Workspace settings pushed via `workspace/didChangeConfiguration`                                                 |\n| `disabled`        | `boolean`  | no       | Set to `true` to disable this server entirely                                                                    |\n| `warmupTimeoutMs` | `number`   | no       | Startup timeout in ms for this server (overrides the global default)                                             |\n| `isLinter`        | `boolean`  | no       | Mark server as linter/formatter only; excluded from type-intelligence operations (hover, go-to-definition, etc.) |\n| `capabilities`    | `object`   | no       | Opt-in server-specific features; see [Capabilities](#capabilities)                                               |\n\n`resolvedCommand` is populated automatically at runtime — do not set it manually.\n\n### Capabilities\n\nThe `capabilities` object enables optional server-specific features that OMP supports on a per-server basis:\n\n```json\n{\n  \"capabilities\": {\n    \"flycheck\": true,\n    \"ssr\": true,\n    \"expandMacro\": true,\n    \"runnables\": true,\n    \"relatedTests\": true\n  }\n}\n```\n\nAll fields are boolean and optional. They are currently used by `rust-analyzer`.\n\n## Common recipes\n\n### Override a built-in server's settings\n\nPartial overrides are merged onto the built-in defaults. You only need to specify the fields you want to change.\n\n```json\n{\n  \"servers\": {\n    \"typescript-language-server\": {\n      \"args\": [\"--stdio\", \"--log-level\", \"4\"]\n    }\n  }\n}\n```\n\n```yaml\nservers:\n  gopls:\n    settings:\n      gopls:\n        gofumpt: false\n        staticcheck: false\n```\n\n### Disable a built-in server\n\n```json\n{\n  \"servers\": {\n    \"eslint\": {\n      \"disabled\": true\n    }\n  }\n}\n```\n\n### Register a custom server\n\nNew servers require `command`, `fileTypes`, and `rootMarkers`. All other fields are optional.\n\n```json\n{\n  \"servers\": {\n    \"my-lsp\": {\n      \"command\": \"my-lsp-server\",\n      \"args\": [\"--stdio\"],\n      \"fileTypes\": [\".xyz\"],\n      \"rootMarkers\": [\".xyz-project\", \".git\"]\n    }\n  }\n}\n```\n\n### Set a global idle timeout\n\nShut down language servers that have been inactive for more than five minutes:\n\n```json\n{\n  \"idleTimeoutMs\": 300000\n}\n```\n\n### Disable a server for one project, keep it globally\n\nPlace the override in `<project>/.omp/lsp.json`:\n\n```json\n{\n  \"servers\": {\n    \"pylsp\": {\n      \"disabled\": true\n    }\n  }\n}\n```\n\nThe user-level config in `~/.omp/agent/lsp.json` is unaffected; pylsp is only suppressed in this project.\n\n## Built-in server list\n\nThe following servers ship in `defaults.json` and are eligible for auto-detection:\n\n| Server key                    | Language(s)                   | Binary                            |\n| ----------------------------- | ----------------------------- | --------------------------------- |\n| `rust-analyzer`               | Rust                          | `rust-analyzer`                   |\n| `clangd`                      | C, C++, ObjC                  | `clangd`                          |\n| `zls`                         | Zig                           | `zls`                             |\n| `gopls`                       | Go                            | `gopls`                           |\n| `typescript-language-server`  | TypeScript, JavaScript        | `typescript-language-server`      |\n| `denols`                      | TypeScript, JavaScript (Deno) | `deno`                            |\n| `biome`                       | TS/JS/JSON (linter)           | `biome`                           |\n| `eslint`                      | TS/JS/Vue/Svelte (linter)     | `vscode-eslint-language-server`   |\n| `vscode-html-language-server` | HTML                          | `vscode-html-language-server`     |\n| `vscode-css-language-server`  | CSS, SCSS, Less               | `vscode-css-language-server`      |\n| `vscode-json-language-server` | JSON                          | `vscode-json-language-server`     |\n| `tailwindcss`                 | HTML, CSS, TS/JS              | `tailwindcss-language-server`     |\n| `svelte`                      | Svelte                        | `svelteserver`                    |\n| `vue-language-server`         | Vue                           | `vue-language-server`             |\n| `astro`                       | Astro                         | `astro-ls`                        |\n| `pyright`                     | Python                        | `pyright-langserver`              |\n| `basedpyright`                | Python                        | `basedpyright-langserver`         |\n| `pylsp`                       | Python                        | `pylsp`                           |\n| `ruff`                        | Python (linter)               | `ruff`                            |\n| `jdtls`                       | Java                          | `jdtls`                           |\n| `kotlin-lsp`                  | Kotlin                        | `kotlin-lsp`                      |\n| `metals`                      | Scala                         | `metals`                          |\n| `hls`                         | Haskell                       | `haskell-language-server-wrapper` |\n| `ocamllsp`                    | OCaml                         | `ocamllsp`                        |\n| `elixirls`                    | Elixir                        | `elixir-ls`                       |\n| `expert`                      | Elixir                        | `expert`                          |\n| `erlangls`                    | Erlang                        | `erlang_ls`                       |\n| `gleam`                       | Gleam                         | `gleam`                           |\n| `solargraph`                  | Ruby                          | `solargraph`                      |\n| `ruby-lsp`                    | Ruby                          | `ruby-lsp`                        |\n| `rubocop`                     | Ruby (linter)                 | `rubocop`                         |\n| `bashls`                      | Bash, Zsh                     | `bash-language-server`            |\n| `lua-language-server`         | Lua                           | `lua-language-server`             |\n| `intelephense`                | PHP                           | `intelephense`                    |\n| `phpactor`                    | PHP                           | `phpactor`                        |\n| `omnisharp`                   | C#                            | `omnisharp`                       |\n| `yamlls`                      | YAML                          | `yaml-language-server`            |\n| `terraformls`                 | Terraform                     | `terraform-ls`                    |\n| `dockerls`                    | Dockerfile                    | `docker-langserver`               |\n| `helm-ls`                     | Helm                          | `helm_ls`                         |\n| `nixd`                        | Nix                           | `nixd`                            |\n| `nil`                         | Nix                           | `nil`                             |\n| `ols`                         | Odin                          | `ols`                             |\n| `dartls`                      | Dart                          | `dart`                            |\n| `marksman`                    | Markdown                      | `marksman`                        |\n| `texlab`                      | LaTeX                         | `texlab`                          |\n| `graphql`                     | GraphQL                       | `graphql-lsp`                     |\n| `prismals`                    | Prisma                        | `prisma-language-server`          |\n| `vimls`                       | Vim script                    | `vim-language-server`             |\n| `emmet-language-server`       | HTML, CSS, JSX                | `emmet-language-server`           |\n| `sourcekit-lsp`               | Swift                         | `sourcekit-lsp`                   |\n| `swiftlint`                   | Swift (linter)                | `swiftlint`                       |\n| `tlaplus`                     | TLA+                          | `tlapm_lsp`                       |\n",
 	"macos-signing-notarization.md": "# macOS signing & notarization\n\nThe compiled macOS `omp` binaries shipped on GitHub Releases are signed with a\n**Developer ID Application** certificate and **notarized** by Apple. This makes\nthem Gatekeeper-acceptable and is the prerequisite for an official Homebrew\nsubmission (see [#776](https://github.com/can1357/oh-my-pi/issues/776)).\n\nSigning happens in CI, in the `release_binary` job's darwin matrix legs\n(`.github/workflows/ci.yml`), via `scripts/ci-macos-sign.sh`. It **auto-skips**\nuntil the `APPLE_*` repository secrets below are configured, so releases keep\nworking (ad-hoc signed, as before) in the meantime.\n\n## How it works\n\n1. `ci:release:build-binaries` builds and **ad-hoc** signs the binary (so it can\n   run on the build runner).\n2. `scripts/ci-macos-sign.sh` then:\n   - imports the Developer ID cert into a throwaway keychain;\n   - re-signs with `--options runtime --timestamp` (hardened runtime + secure\n     timestamp) and `--entitlements scripts/macos-entitlements.plist`;\n   - runs `--version` and `--smoke-test` under the new signature to fail fast;\n   - notarizes the binary via `notarytool submit --wait`.\n3. `release_github_verify` re-downloads the published arm64 asset and asserts it\n   is **not** ad-hoc, passes `codesign --verify --strict`, and boots cleanly.\n\n### Why the entitlements are mandatory\n\nThe binary is a Bun single-file executable, so the hardened runtime needs:\n\n| Entitlement | Reason |\n| --- | --- |\n| `com.apple.security.cs.allow-jit` | JavaScriptCore JITs at runtime. |\n| `com.apple.security.cs.allow-unsigned-executable-memory` | JSC executable memory pages. |\n| `com.apple.security.cs.disable-library-validation` | omp extracts its native addon (`pi_natives.<triple>.node`) and other optional dylibs to a runtime cache and `dlopen()`s them. They do not share the main binary's Team ID, so without this the hardened runtime aborts with *\"mapping process and mapped file have different Team IDs\"* — breaking effectively every command. |\n\nWithout `disable-library-validation`, a signed+notarized binary signs and\nnotarizes fine but **fails at first real use**. `scripts/ci-macos-sign.sh` runs\n`--smoke-test` after signing specifically to catch this before notarizing.\n\n### Stapling limitation (important)\n\nA bare Mach-O executable **cannot be stapled** (`stapler` only supports\n`.app`/`.pkg`/`.dmg`). The binary is genuinely notarized — `notarytool` returns\n`Accepted` and the ticket exists on Apple's servers keyed to its cdhash — but\nbecause there is no *stapled* ticket, a direct `spctl -a -t exec` assessment\nreports `rejected / source=Unnotarized Developer ID`. This is expected and is\n**not** a signing or credential failure.\n\nWhat this means in practice:\n\n- `curl https://omp.sh/install | sh` — `curl` sets no quarantine bit, so\n  Gatekeeper is never consulted; the binary just runs. ✅\n- Homebrew **formula** installs — Homebrew does not quarantine formula files, so\n  Gatekeeper is never consulted. ✅\n- Anything that **quarantines** the binary (a browser download, or a Homebrew\n  **cask**) and is assessed offline will be blocked, because there is no stapled\n  ticket. For that route, wrap the binary in a stapleable, notarized **`.pkg` or\n  `.dmg`** (`xcrun stapler staple` works on those). That is a follow-up and is\n  **not** required for the `curl`/formula paths.\n\n## Required GitHub secrets\n\nAdd these under **Settings → Secrets and variables → Actions** (repo secrets).\nAll five secrets (cert, password, and API key trio) must be present for\nsigning to engage.\n\n| Secret | What it is |\n| --- | --- |\n| `APPLE_CERTIFICATE_P12` | base64 of the exported Developer ID Application `.p12` (cert + private key). |\n| `APPLE_CERTIFICATE_PASSWORD` | password you set when exporting the `.p12`. |\n| `APPLE_API_KEY_ID` | App Store Connect API **Key ID**. |\n| `APPLE_API_ISSUER_ID` | App Store Connect API **Issuer ID** (UUID). |\n| `APPLE_API_KEY` | base64 of the App Store Connect `.p8` private key. |\n\n### Producing the credential files\n\nDrop these into a working directory (default `~/omp-signing`):\n\n| File | How |\n| --- | --- |\n| `*.p12` | **Keychain Access** → right-click your *Developer ID Application: …* identity (the entry that expands to a cert **with** a private key) → **Export…** → save as `.p12` and set a password. |\n| `p12-password.txt` | the password you just set on the `.p12`. |\n| `AuthKey_<KEYID>.p8` | App Store Connect → **Users and Access → Integrations → App Store Connect API** → create a key (**Account Holder** role also allows API cert creation; **Developer** is enough for notarization) → **download once** (non-recoverable). |\n| `issuer-id.txt` | the **Issuer ID** (UUID) shown above the keys table. |\n| `key-id.txt` | *optional* — the Key ID; otherwise read from the `.p8` filename. |\n\nThe App Store Connect API key is the one credential that **cannot** be minted\nfrom a CLI — it is the bootstrap credential for the API itself, and the `.p8`\ndownloads exactly once. Everything else is local.\n\n### Uploading (no value leaves disk)\n\n`scripts/ci-macos-upload-secrets.sh` validates the files (opens the `.p12` with\nyour password, sanity-checks the `.p8`) and pipes each value to `gh secret set`\nover stdin — no secret is ever printed to the terminal, argv, or shell history:\n\n```sh\nscripts/ci-macos-upload-secrets.sh ~/omp-signing --dry-run   # validate first\nscripts/ci-macos-upload-secrets.sh ~/omp-signing             # upload all five\ngh secret list --repo can1357/oh-my-pi                       # confirm\n```\n\nRe-run it whenever the certificate is renewed.\n\n### Finding your signing identity / Team ID (sanity check)\n\n```sh\nsecurity find-identity -v -p codesigning\n# e.g. \"Developer ID Application: Your Name (TEAMID1234)\"\n```\n\nThe script selects the first `Developer ID Application` identity automatically;\nyou do not need to store the identity string or Team ID as a secret.\n\n## Local dry run\n\nYou can exercise the full sign+notarize path locally (real cert + API key) by\nexporting the five env vars and running:\n\n```sh\nRELEASE_TARGETS=darwin-arm64 bun run ci:release:build-binaries\nAPPLE_CERTIFICATE_P12=… APPLE_CERTIFICATE_PASSWORD=… \\\nAPPLE_API_KEY_ID=… APPLE_API_ISSUER_ID=… APPLE_API_KEY=… \\\n  bash scripts/ci-macos-sign.sh packages/coding-agent/binaries/omp-darwin-arm64\n```\n",
@@ -34,7 +34,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"mcp-server-tool-authoring.md": "# MCP server and tool authoring\n\nThis document explains how MCP server definitions become callable `mcp__*` tools in coding-agent, and what operators should expect when configs are invalid, duplicated, disabled, or auth-gated.\n\n## Architecture at a glance\n\n```text\nConfig sources (.omp/.claude/.cursor/.vscode/mcp.json, mcp.json, etc.)\n  -> discovery providers normalize to canonical MCPServer\n  -> capability loader dedupes by server name (higher provider priority wins)\n  -> loadAllMCPConfigs converts to MCPServerConfig + skips enabled:false\n  -> MCPManager connects/listTools (with auth/header/env resolution)\n  -> manager best-effort loads resources/prompts and subscribes to resource updates when enabled\n  -> MCPTool/DeferredMCPTool bridge exposes tools as mcp__<server>_<tool>\n  -> AgentSession.refreshMCPTools replaces live MCP tools immediately\n```\n\n## 1) Server config model and validation\n\n`src/mcp/types.ts` defines the authoring shape used by MCP config writers and runtime:\n\n- `stdio` (default when `type` missing): requires `command`, optional `args`, `env`, `cwd`\n- `http`: requires `url`, optional `headers`\n- `sse`: requires `url`, optional `headers` (kept for compatibility)\n- shared fields: `enabled`, `timeout`, `auth`, `oauth`\n\n`validateServerConfig()` (`src/mcp/config.ts`) enforces transport basics:\n\n- rejects configs that set both `command` and `url`\n- requires `command` for stdio\n- requires `url` for http/sse\n- rejects unknown `type`\n\n`config-writer.ts` applies this validation for add/update operations and also validates server names:\n\n- non-empty\n- max 100 chars\n- only `[a-zA-Z0-9_.-]`\n\n### Transport pitfalls\n\n- `type` omitted means stdio. If you intended HTTP/SSE but omitted `type`, `command` becomes mandatory.\n- `sse` is still accepted but treated as HTTP transport internally (`createHttpTransport`).\n- Validation is structural, not reachability: a syntactically valid URL can still fail at connect time.\n\n## 2) Discovery, normalization, and precedence\n\n### Capability-based discovery\n\n`loadAllMCPConfigs()` (`src/mcp/config.ts`) loads canonical `MCPServer` items via `loadCapability(mcpCapability.id)`.\n\nThe capability layer (`src/capability/index.ts`) then:\n\n1. loads providers in priority order\n2. dedupes by `server.name` (first win = highest priority)\n3. validates deduped items\n\nResult: duplicate server names across sources are not merged. One definition wins; lower-priority duplicates are shadowed.\n\n### `.mcp.json` and related files\n\nThe dedicated fallback provider in `src/discovery/mcp-json.ts` reads project-root `mcp.json` and `.mcp.json` (low priority).\n\nIn practice MCP servers also come from higher-priority providers (for example native `.omp/...` and tool-specific config dirs). Authoring guidance:\n\n- Prefer `.omp/mcp.json` (project) or `~/.omp/agent/mcp.json` (user) for explicit control.\n- Use root `mcp.json` / `.mcp.json` when you need fallback compatibility.\n- Reusing the same server name in multiple sources causes precedence shadowing, not merge.\n\n### Normalization behavior\n\n`convertToLegacyConfig()` (`src/mcp/config.ts`) maps canonical `MCPServer` to runtime `MCPServerConfig`.\n\nKey behavior:\n\n- transport inferred as `server.transport ?? (command ? \"stdio\" : url ? \"http\" : \"stdio\")`\n- disabled servers (`enabled === false`) and names in the user `disabledServers` list are dropped before connection\n- optional fields are preserved when present\n\n### Environment expansion during discovery\n\nOMP-native MCP config (`.omp/mcp.json`, `~/.omp/agent/mcp.json`, plus their `.mcp.json` variants) expands `${VAR}` and `${VAR:-default}` placeholders recursively before converting to runtime config. It also accepts boolean/string forms for `enabled` (`true`, `false`, `1`, `0`) and numeric strings for `timeout`.\n\nThe standalone fallback provider in `src/discovery/mcp-json.ts` reads project-root `mcp.json` and `.mcp.json`, expands the same `${...}` placeholders, and type-checks `enabled`/`timeout` without coercing string values.\n\nInvalid `enabled`/`timeout` values are ignored with warnings rather than failing the whole file.\n\n## 3) Auth and runtime value resolution\n\n`MCPManager.prepareConfig()`/`#resolveAuthConfig()` (`src/mcp/manager.ts`) is the final pre-connect pass.\n\n### OAuth credential injection\n\nIf config has:\n\n```ts\nauth: { type: \"oauth\", credentialId: \"...\" }\n```\n\nand credential exists in auth storage:\n\n- `http`/`sse`: injects `Authorization: Bearer <access_token>` header\n- `stdio`: injects `OAUTH_ACCESS_TOKEN` env var\n\nIf credential lookup fails, manager logs a warning and continues with unresolved auth.\n\n### Header/env value resolution\n\nBefore connect, manager resolves stdio `env` values and HTTP/SSE `headers` values via `resolveConfigValue()` (`src/config/resolve-config-value.ts`):\n\n- value starting with `!` => execute shell command, use trimmed stdout (cached)\n- failed, timed-out, or whitespace-only commands produce `undefined`, so that entry is omitted\n- otherwise, treat value as environment variable name first (`process.env[name]`), fallback to literal value\n\nOperational caveat: a mistyped `!` secret command can silently remove that header/env entry, producing downstream 401/403 or server startup failures. A mistyped environment variable name is sent literally unless that literal happens to be meaningful to the server.\n\n## 4) Tool bridge: MCP -> agent-callable tools\n\n`src/mcp/tool-bridge.ts` converts MCP tool definitions into `CustomTool`s.\n\n### Naming and collision domain\n\nTool names are generated as:\n\n```text\nmcp__<sanitized_server_name>_<sanitized_tool_name>\n```\n\nRules:\n\n- lowercases\n- non-`[a-z_]` chars become `_`\n- repeated underscores collapse\n- redundant `<server>_` prefix in tool name is stripped once\n\nThis avoids many collisions, but not all. Different raw names can still sanitize to the same identifier (for example `my-server` and `my.server` both sanitize similarly), and registry insertion is last-write-wins.\n\n### Schema mapping\n\n`tool-bridge.ts` passes each MCP `inputSchema` through `normalizeSchemaForMCP()` before registering it as a `CustomTool` schema.\n\n### Execution mapping\n\n`MCPTool.execute()` / `DeferredMCPTool.execute()`:\n\n- calls MCP `tools/call`\n- flattens MCP content into displayable text\n- returns structured details (`serverName`, `mcpToolName`, provider metadata)\n- maps server-reported `isError` to `Error: ...` text result\n- attempts reconnect + one retry for retriable connection errors\n- maps remaining thrown transport/runtime failures to `MCP error: ...`\n- preserves abort semantics by translating AbortError into `ToolAbortError`\n\n## 5) Operator lifecycle: add/edit/remove and live updates\n\nInteractive mode exposes `/mcp` in `src/modes/controllers/mcp-command-controller.ts`.\n\nSupported operations:\n\n- `add` (wizard or quick-add)\n- `remove` / `rm`\n- `enable` / `disable`\n- `test`\n- `reauth` / `unauth`\n- `reconnect`\n- `reload`\n- `resources`, `prompts`, `notifications`\n- Smithery search/login/logout flows\n\nConfig writes are atomic (`writeMCPConfigFile`: temp file + rename).\n\nAfter changes, controller calls `#reloadMCP()`:\n\n1. `mcpManager.disconnectAll()`\n2. `mcpManager.discoverAndConnect()`\n3. `session.refreshMCPTools(mcpManager.getTools())`\n\n`refreshMCPTools()` replaces all `mcp__` registry entries and immediately re-activates the latest MCP tool set, so changes take effect without restarting the session.\n\n### Mode differences\n\n- **Interactive/TUI mode**: `/mcp` gives in-app UX (wizard, OAuth flow, connection status text, immediate runtime rebinding).\n- **SDK/headless integration**: `discoverAndLoadMCPTools()` (`src/mcp/loader.ts`) returns loaded tools + per-server errors; no `/mcp` command UX.\n\n## 6) User-visible error surfaces\n\nCommon error strings users/operators see:\n\n- add/update validation failures:\n  - `Invalid server config: ...`\n  - `Server \"<name>\" already exists in <path>`\n- quick-add argument issues:\n  - `Use either --url or -- <command...>, not both.`\n  - `--token requires --url (HTTP/SSE transport).`\n- connect/test failures:\n  - `Failed to connect to \"<name>\": <message>`\n  - timeout help text suggests increasing timeout\n  - auth help text for `401/403`\n- auth/OAuth flows:\n  - `Authentication required ... OAuth endpoints could not be discovered`\n  - `OAuth flow timed out. Please try again.`\n  - `OAuth authentication failed: ...`\n- disabled server usage:\n  - `Server \"<name>\" is disabled. Run /mcp enable <name> first.`\n\nBad source JSON in discovery is generally handled as warnings/logs; config-writer paths throw explicit errors.\n\n## 7) Practical authoring guidance\n\nFor robust MCP authoring in this codebase:\n\n1. Keep server names globally unique across all MCP-capable config sources.\n2. Prefer names that remain distinct after MCP tool-name sanitization to avoid generated `mcp__` collisions.\n3. Use explicit `type` to avoid accidental stdio defaults.\n4. Treat `enabled: false` as hard-off: server is omitted from runtime connect set.\n5. For OAuth configs, store a valid `credentialId`; otherwise auth injection is skipped.\n6. If using command-based secret resolution (`!cmd`), verify command output is stable and non-empty.\n\n## Implementation files\n\n- [`src/mcp/types.ts`](../packages/coding-agent/src/mcp/types.ts)\n- [`src/mcp/config.ts`](../packages/coding-agent/src/mcp/config.ts)\n- [`src/mcp/config-writer.ts`](../packages/coding-agent/src/mcp/config-writer.ts)\n- [`src/mcp/tool-bridge.ts`](../packages/coding-agent/src/mcp/tool-bridge.ts)\n- [`src/discovery/mcp-json.ts`](../packages/coding-agent/src/discovery/mcp-json.ts)\n- [`src/modes/controllers/mcp-command-controller.ts`](../packages/coding-agent/src/modes/controllers/mcp-command-controller.ts)\n- [`src/mcp/manager.ts`](../packages/coding-agent/src/mcp/manager.ts)\n- [`src/capability/index.ts`](../packages/coding-agent/src/capability/index.ts)\n- [`src/config/resolve-config-value.ts`](../packages/coding-agent/src/config/resolve-config-value.ts)\n- [`src/mcp/loader.ts`](../packages/coding-agent/src/mcp/loader.ts)\n",
 	"memory.md": "# Autonomous Memory\n\nWhen the local memory backend is enabled, the agent automatically extracts durable knowledge from past sessions and injects a compact summary into future sessions for the same project. Over time it builds a project-scoped memory store — technical decisions, recurring workflows, pitfalls — that carries forward without manual effort.\n\nDisabled by default. Enable the local summary pipeline via `/settings` or `config.yml`:\n\n```yaml\nmemory:\n  backend: local\n```\n\n## Usage\n\n### What gets injected\n\nAt session start, if a memory summary exists for the current project, it is injected into the system prompt as a **Memory Guidance** block. The agent is instructed to:\n\n- Treat memory as heuristic context — useful for process and prior decisions, not authoritative on current repo state.\n- Cite the memory artifact path when memory changes the plan, and pair it with current-repo evidence before acting.\n- Prefer repo state and user instruction when they conflict with memory; treat conflicting memory as stale.\n\n### Reading memory artifacts\n\nThe agent can read memory files directly using `memory://` URLs with the `read` tool:\n\n| URL                                    | Content                             |\n| -------------------------------------- | ----------------------------------- |\n| `memory://root`                        | Compact summary injected at startup |\n| `memory://root/MEMORY.md`              | Full long-term memory document      |\n| `memory://root/skills/<name>/SKILL.md` | A generated skill playbook          |\n\n### `/memory` slash command\n\n| Subcommand            | Effect                                                    |\n| --------------------- | --------------------------------------------------------- |\n| `view`                | Show the current backend injection payload                |\n| `stats`               | Show backend-specific memory statistics, when supported   |\n| `diagnose`            | Show backend-specific diagnostics, when supported         |\n| `clear` / `reset`     | Delete active backend memory data/artifacts               |\n| `enqueue` / `rebuild` | Force consolidation/retention work for the active backend |\n\n## How it works\n\nLocal summary memories are built by a background pipeline that runs at startup; `/memory enqueue` marks consolidation work that the next startup picks up. The pipeline is skipped for subagents and for sessions that are not persisted to a session file.\n\n**Phase 1 — per-session extraction:** For each past session that has changed since it was last processed, a model reads the session history and extracts durable signal: technical decisions, constraints, resolved failures, recurring workflows. Sessions that are too recent, too old, currently active, or beyond the configured scan/age limits are skipped. Each extraction produces a raw memory block and a short synopsis for that session.\n\n**Phase 2 — consolidation:** After extraction, a second model pass reads all per-session extractions and produces three outputs written to disk:\n\n- `MEMORY.md` — a curated long-term memory document\n- `memory_summary.md` — the compact text injected at session start\n- `skills/` — reusable procedural playbooks, each in its own subdirectory\n\nPhase 2 uses a lease and heartbeat to prevent double-running when multiple processes start simultaneously. Stale skill directories from prior runs are pruned automatically.\n\nConsolidated output is redacted for common secret/token patterns before `MEMORY.md`, `memory_summary.md`, or generated skills are written to disk.\n\n### Extraction behavior\n\nMemory extraction and consolidation behavior is driven by static prompt files in `packages/coding-agent/src/prompts/memories/`.\n\n| File                     | Purpose                                      | Variables                                   |\n| ------------------------ | -------------------------------------------- | ------------------------------------------- |\n| `stage_one_system.md`    | System prompt for per-session extraction     | —                                           |\n| `stage_one_input.md`     | User-turn template wrapping session content  | `{{thread_id}}`, `{{response_items_json}}`  |\n| `consolidation_system.md`| System prompt for cross-session consolidation | —                                          |\n| `consolidation.md`       | User-turn prompt for cross-session consolidation | `{{raw_memories}}`, `{{rollout_summaries}}` |\n| `read-path.md`           | Memory guidance injected into live sessions  | `{{memory_summary}}`                        |\n\n### Model selection\n\nMemory piggybacks on the model role system.\n\n| Phase                   | Role                                                                | Purpose                          |\n| ----------------------- | ------------------------------------------------------------------- | -------------------------------- |\n| Phase 1 (extraction)    | `default`                                                           | Per-session knowledge extraction |\n| Phase 2 (consolidation) | `smol` (falls back to `default`, then current/first registry model) | Cross-session synthesis          |\n\nIf the requested memory role is not configured, memory model resolution falls back to the `default` role, then the active session model, then the first model in the registry.\n\n## Configuration\n\n| Setting                               | Default | Description                                                                                                                              |\n| ------------------------------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------------------- |\n| `memory.backend`                      | `off`   | Select `local` for this pipeline; legacy `memories.enabled: true` is migrated to `memory.backend: local` when no explicit backend is set |\n| `memories.maxRolloutAgeDays`          | `30`    | Sessions older than this are not processed                                                                                               |\n| `memories.minRolloutIdleHours`        | `12`    | Sessions active more recently than this are skipped                                                                                      |\n| `memories.maxRolloutsPerStartup`      | `64`    | Cap on sessions processed in a single startup                                                                                            |\n| `memories.summaryInjectionTokenLimit` | `5000`  | Max tokens of the summary injected into the system prompt                                                                                |\n\nAdditional tuning knobs (concurrency, lease durations, token budgets) are available in config for advanced use.\n\n## Key files\n\n- `packages/coding-agent/src/memories/index.ts` — pipeline orchestration, injection, clear/enqueue entry points (the `/memory` command routes here via `packages/coding-agent/src/memory-backend/local-backend.ts`)\n- `packages/coding-agent/src/memories/storage.ts` — SQLite-backed job queue and thread registry\n- `packages/coding-agent/src/prompts/memories/` — memory prompt templates\n- `packages/coding-agent/src/internal-urls/memory-protocol.ts` — `memory://` URL handler\n",
 	"mnemosyne-memory-backend.md": "# Mnemopi memory backend\n\nOh My Pi can use `@oh-my-pi/pi-mnemopi` as a local long-term memory backend.\n\nSet:\n\n```yaml\nmemory:\n  backend: mnemopi\n```\n\nExample:\n\n```yaml\nmemory:\n  backend: mnemopi\nmnemopi:\n  scoping: per-project-tagged\n```\n\nWith this backend enabled, the coding agent:\n\n1. Opens one or more local Mnemopi SQLite databases according to the configured bank scoping.\n2. Recalls relevant memories into a `<memories>` block for the first model turn of a session and refreshes the base prompt if recall happens from the `agent_start` listener.\n3. Retains completed conversation turns into the retain bank after agent turns, no more often than `mnemopi.retainEveryNTurns`.\n4. Adds recalled memory as extra compaction context when compaction asks the memory backend for `preCompactionContext`.\n5. Uses the normal `/memory view`, `/memory stats`, `/memory diagnose`, `/memory clear`, and `/memory enqueue` commands through the shared memory backend interface.\n\nRecalled memory is background context, not instructions. Current user messages and tool output take precedence when they conflict.\n\n## Settings\n\n| Setting                         | Default                | Description                                                                                                                                                             |\n| ------------------------------- | ---------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| `memory.backend`                | `off`                  | Set to `mnemopi` to enable this backend.                                                                                                                              |\n| `mnemopi.dbPath`              | agent memories dir     | Optional SQLite database path.                                                                                                                                          |\n| `mnemopi.bank`                | unset                  | Optional shared bank base name passed to `Mnemopi`; the coding-agent wrapper scopes from this base according to `mnemopi.scoping`. Unset → shared bank `default`; per-project modes derive a project bank from the project root name plus a stable hash. |\n| `mnemopi.scoping`             | `per-project`          | Memory visibility mode: `global` = one shared bank, `per-project` = isolated project memory, `per-project-tagged` = project-local writes plus global recall visibility. |\n| `mnemopi.autoRecall`          | `true`                 | Recall memory on the first turn of a session.                                                                                                                           |\n| `mnemopi.autoRetain`          | `true`                 | Retain completed turns automatically.                                                                                                                                   |\n| `mnemopi.polyphonicRecall`    | `false`                | Enable 4-voice polyphonic recall (vector, graph, fact, temporal) with reciprocal rank fusion; `MNEMOPI_POLYPHONIC_RECALL` overrides when set.                            |\n| `mnemopi.enhancedRecall`      | `false`                | Enable the tiered query result cache for repeated/similar recall queries; `MNEMOPI_ENHANCED_RECALL` overrides when set.                                                  |\n| `mnemopi.retainEveryNTurns`   | `4`                    | Minimum user turns between automatic retain writes.                                                                                                                     |\n| `mnemopi.recallLimit`         | `8`                    | Maximum recalled memories in the prompt block.                                                                                                                          |\n| `mnemopi.recallContextTurns`  | `3`                    | Prior user-bounded turns included in recall queries.                                                                                                                    |\n| `mnemopi.recallMaxQueryChars` | `4000`                 | Maximum composed recall query length.                                                                                                                                   |\n| `mnemopi.injectionTokenLimit` | `5000`                 | Approximate token budget for memory prompt injection.                                                                                                                   |\n| `mnemopi.debug`               | `false`                | Enable debug logging for backend failures.                                                                                                                              |\n| `mnemopi.noEmbeddings`        | `false`                | Pass `noEmbeddings` to `Mnemopi` and force FTS-only recall.                                                                                                           |\n| `mnemopi.embeddingModel`      | env/default            | Embedding model passed to `Mnemopi`.                                                                                                                                  |\n| `mnemopi.embeddingApiUrl`     | env/default            | OpenAI-compatible embedding endpoint passed to `Mnemopi`.                                                                                                             |\n| `mnemopi.embeddingApiKey`     | env/default            | Embedding API key passed to `Mnemopi`.                                                                                                                                |\n| `mnemopi.llmMode`             | `smol`                 | `smol` uses the configured pi-ai smol model, `remote` uses the settings below, and `none` disables LLM calls.                                                           |\n| `mnemopi.llmBaseUrl`          | env/default            | OpenAI-compatible LLM endpoint for `llmMode: remote`.                                                                                                                   |\n| `mnemopi.llmApiKey`           | env/default            | LLM API key for `llmMode: remote`.                                                                                                                                      |\n| `mnemopi.llmModel`            | env/default            | LLM model id for `llmMode: remote`.                                                                                                                                     |\n\n## Scoping\n\nThe coding-agent wrapper applies scoping on top of the underlying `Mnemopi` package:\n\n- `global` uses one shared bank for recall and writes.\n- `per-project` writes to and recalls from a bank derived from the current git repository root (or cwd) plus a stable hash.\n- `per-project-tagged` writes to the project-local bank and recalls from both the project-local bank and the shared global bank, with duplicate recall results merged.\n\nThe combined project-plus-global behavior lives in the wrapper. The `@oh-my-pi/pi-mnemopi` package itself still exposes banks and constructor options directly, including `bank` for selecting a bank name. Project-local banks other than the shared bank are stored as sibling bank databases managed by Mnemopi's `BankManager`.\n\n## LLM and embeddings\n\nThe backend passes these settings to the `Mnemopi` constructor; if a setting is omitted, Mnemopi falls back to its `MNEMOPI_*` environment defaults. The backend does not download or run a local GGUF LLM. LLM-dependent paths use a configured pi-ai model, an opt-in local on-device memory model (`providers.memoryModel`, ONNX — overrides `smol`/`remote` when set to a local model), a dynamic completion function, a remote OpenAI-compatible endpoint, or deterministic no-LLM fallbacks.\n\nFTS-only:\n\n```yaml\nmemory:\n  backend: mnemopi\nmnemopi:\n  noEmbeddings: true\n```\n\nEquivalent constructor shape:\n\n```ts\nnew Mnemopi({ noEmbeddings: true });\n```\n\nRemote embeddings:\n\n```yaml\nmnemopi:\n  embeddingModel: text-embedding-3-small\n  embeddingApiUrl: https://api.openai.com/v1\n  embeddingApiKey: ${OPENAI_API_KEY}\n```\n\nEquivalent constructor shape:\n\n```ts\nnew Mnemopi({\n  embeddingModel: \"text-embedding-3-small\",\n  embeddingApiUrl: \"https://api.openai.com/v1\",\n  embeddingApiKey,\n});\n```\n\nRemote LLM:\n\n```yaml\nmnemopi:\n  llmMode: remote\n  llmBaseUrl: https://api.openai.com/v1\n  llmApiKey: ${OPENAI_API_KEY}\n  llmModel: gpt-4.1-mini\n```\n\nEquivalent constructor shapes:\n\n```ts\nnew Mnemopi({ llm: { baseUrl, apiKey, model } });\nnew Mnemopi({ llmBaseUrl: baseUrl, llmApiKey: apiKey, llmModel: model });\n```\n\nDynamic function LLM for rotating OAuth tokens:\n\n```ts\nnew Mnemopi({\n  llm: async (prompt, opts) => {\n    const token = await getFreshOauthToken();\n    return await completeWithPiAi(prompt, {\n      token,\n      maxTokens: opts?.maxTokens,\n      temperature: opts?.temperature,\n    });\n  },\n});\n```\n\npi-ai smol model LLM:\n\n```yaml\nmnemopi:\n  llmMode: smol\n```\n\nThe coding agent resolves its configured smol role and passes a dynamic completion function so every Mnemopi LLM call can fetch the current provider credentials at call time:\n\n```ts\nnew Mnemopi({\n  llm: async (prompt, opts) => completeSmolWithCurrentAuth(prompt, opts),\n});\n```\n\n## Operational notes\n\n- The default shared database lives under the agent memories directory in `mnemopi/mnemopi.db`; project-scoped banks use sibling database paths under that Mnemopi directory.\n- `/memory clear` removes every scoped Mnemopi SQLite database and sidecar WAL/SHM files for the active configuration.\n- `/memory enqueue` forces retention of the current session, flushes pending fact extractions, and runs Mnemopi sleep/consolidation.\n- `/memory stats` and `/memory diagnose` render backend-specific bank statistics/diagnostics when the Mnemopi backend is active.\n- Subagents do not own separate Mnemopi retain loops; they alias the parent state when a parent Mnemopi state exists, and otherwise remain inert.\n",
-	"models.md": "# Model and Provider Configuration (`models.yml`)\n\nThis document describes how the coding-agent currently loads models, applies overrides, resolves credentials, and chooses models at runtime.\n\n## What controls model behavior\n\nPrimary implementation files:\n\n- `src/config/model-registry.ts` — loads built-in + custom models, provider overrides, runtime discovery, auth integration\n- `src/config/model-resolver.ts` — parses model patterns and selects initial/smol/slow models\n- `src/config/settings-schema.ts` — model-related settings (`modelRoles`, provider transport preferences)\n- `src/session/auth-storage.ts` — API key + OAuth resolution order\n- `packages/catalog/src/models.ts` and `packages/catalog/src/types.ts` — built-in providers/models (`getBundledModels` / `getBundledProviders`) and `Model`/`compat` types\n\n## Config file location and legacy behavior\n\nDefault config path:\n\n- `~/.omp/agent/models.yml`\n\nLegacy behavior still present:\n\n- If `models.yml` is missing and `models.json` exists at the same location, it is migrated to `models.yml`.\n- Explicit `.json` / `.jsonc` config paths are still supported when passed programmatically to `ModelRegistry`.\n\n## `models.yml` shape\n\n```yaml\nproviders:\n  <provider-id>:\n    # provider-level config\nequivalence:\n  overrides:\n    <provider-id>/<model-id>: <canonical-model-id>\n  exclude:\n    - <provider-id>/<model-id>\n```\n\n`provider-id` is the canonical provider key used across selection and auth lookup.\n\n`equivalence` is optional and configures canonical model grouping on top of concrete provider models:\n\n- `overrides` maps an exact concrete selector (`provider/modelId`) to an official upstream canonical id\n- `exclude` opts a concrete selector out of canonical grouping\n\n## Provider-level fields\n\n```yaml\nproviders:\n  my-provider:\n    baseUrl: https://api.example.com/v1\n    apiKey: MY_PROVIDER_API_KEY\n    api: openai-completions\n    headers:\n      X-Team: platform\n    authHeader: true\n    auth: apiKey\n    disableStrictTools: false # set true for Anthropic-compatible endpoints that reject the strict field\n    discovery:\n      type: ollama\n    modelOverrides:\n      some-model-id:\n        name: Renamed model\n    models:\n      - id: some-model-id\n        name: Some Model\n        api: openai-completions\n        reasoning: false\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 128000\n        maxTokens: 16384\n        headers:\n          X-Model: value\n        compat:\n          supportsStore: true\n          supportsDeveloperRole: true\n          supportsReasoningEffort: true\n          maxTokensField: max_completion_tokens\n          openRouterRouting:\n            only: [anthropic]\n          vercelGatewayRouting:\n            order: [anthropic, openai]\n          extraBody:\n            gateway: m1-01\n            controller: mlx\n```\n\n### Allowed provider/model `api` values\n\n- `openai-completions`\n- `openai-responses`\n- `openai-codex-responses`\n- `azure-openai-responses`\n- `anthropic-messages`\n- `google-generative-ai`\n- `google-vertex`\n\n### Allowed auth/discovery values\n\n- `auth`: `apiKey` (default), `none`, or `oauth`; for `models.yml` custom models, `oauth` is accepted by schema but does not waive the `apiKey` requirement\n- `discovery.type`: `ollama`, `llama.cpp`, `lm-studio`, `openai-models-list`, or `proxy`\n- `transport`: `pi-native` only. When set, every model under that provider is sent to an `omp auth-gateway` compatible `baseUrl` via `POST /v1/pi/stream`; `apiKey` is the gateway bearer.\n\n## Validation rules (current)\n\n### Full custom provider (`models` is non-empty)\n\nRequired:\n\n- `baseUrl`\n- `apiKey` unless `auth: none`\n- `api` at provider level or each model\n\n### Override-only provider (`models` missing or empty)\n\nMust define at least one of:\n\n- `baseUrl`\n- `apiKey`\n- `headers`\n- `compat`\n- `disableStrictTools`\n- `modelOverrides`\n- `discovery`\n\n### Discovery\n\n- `discovery` requires provider-level `api`, except `discovery.type: proxy` (per-model wire auto-detected).\n\n### Model value checks\n\n- `id` required\n- `contextWindow` and `maxTokens` must be positive if provided\n\n### Command-resolved secrets\n\nProvider `apiKey` values and provider/model `headers` values may start with `!` to read a secret from command stdout. The command is run with a 10 s timeout, stdout is trimmed, and empty/failing commands are omitted:\n\n```yaml\nproviders:\n  openai:\n    apiKey: \"!op read op://dev/openai/api-key\"\n    headers:\n      X-Team-Key: \"!bw get password omp-team-key\"\n```\n\nSuccessful command outputs are cached for the process lifetime so the command is not re-run for every model.\n\n## Merge and override order\n\nModelRegistry pipeline (on refresh):\n\n1. Load built-in providers/models from `@oh-my-pi/pi-ai`.\n2. Load `models.yml` custom config.\n3. Apply provider overrides (`baseUrl`, `headers`, `disableStrictTools`) to built-in models.\n4. Apply `modelOverrides` (per provider + model id).\n5. Merge custom `models`:\n   - same `provider + id` replaces existing\n   - otherwise append\n6. Load cached/runtime-discovered models (Ollama, llama.cpp, LM Studio, plus built-in provider managers), then re-apply model overrides.\n\n### Provider-model cache and static fingerprint\n\nCached per-provider model lists are persisted in the model-cache SQLite\ndatabase (current schema version 5) with a `static_fingerprint` column that\nhashes the static catalog slice merged into the row. When `resolveProviderModels`\nskips the network fetch and the fingerprint of the in-memory static\ncatalog matches the cached one, the cached rows are returned verbatim —\nthe static + dynamic merge is bypassed entirely. The fingerprint is\nmemoized per process by tagging the static-models array with a symbol\nproperty, so repeated cold-start calls do not re-hash.\n\n## Canonical model equivalence and coalescing\n\nThe registry keeps every concrete provider model and then builds a canonical layer above them.\n\nCanonical ids are official upstream ids only, for example:\n\n- `claude-opus-4-6`\n- `claude-haiku-4-5`\n- `gpt-5.3-codex`\n\n### `models.yml` equivalence config\n\nExample:\n\n```yaml\nproviders:\n  zenmux:\n    baseUrl: https://api.zenmux.example/v1\n    apiKey: ZENMUX_API_KEY\n    api: openai-codex-responses\n    models:\n      - id: codex\n        name: Zenmux Codex\n        reasoning: true\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 200000\n        maxTokens: 32768\n\nequivalence:\n  overrides:\n    zenmux/codex: gpt-5.3-codex\n    p-codex/codex: gpt-5.3-codex\n  exclude:\n    - demo/codex-preview\n```\n\nBuild order for canonical grouping:\n\n1. exact user override from `equivalence.overrides`\n2. bundled official-id matches from built-in model metadata\n3. conservative heuristic normalization for gateway/provider variants\n4. fallback to the concrete model's own id\n\nCurrent heuristics are intentionally narrow:\n\n- embedded upstream prefixes can be stripped when present, for example `anthropic/...` or `openai/...`\n- dotted and dashed version variants can normalize only when they map to an existing official id, for example `4.6 -> 4-6`\n- ambiguous families or versions are not merged without a bundled match or explicit override\n\n### Canonical resolution behavior\n\nWhen multiple concrete variants share a canonical id, resolution uses:\n\n1. availability and auth\n2. `config.yml` `modelProviderOrder`\n3. existing registry/provider order if `modelProviderOrder` is unset\n\nDisabled or unauthenticated providers are skipped.\n\nSession state and transcripts continue to record the concrete provider/model that actually executed the turn.\n\nProvider defaults vs per-model overrides:\n\n- Provider `headers` are baseline.\n- Model `headers` override provider header keys.\n- `modelOverrides` can override model metadata (`name`, `reasoning`, `thinking`, `input`, `cost`, `premiumMultiplier`, `contextWindow`, `maxTokens`, `headers`, `compat`, `contextPromotionTarget`).\n- `compat` is deep-merged for nested routing blocks (`openRouterRouting`, `vercelGatewayRouting`, `extraBody`).\n\n## Runtime discovery integration\n\n### Implicit Ollama discovery\n\nIf `ollama` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `ollama`\n- api: `openai-responses`\n- base URL: `OLLAMA_BASE_URL`, or `OLLAMA_HOST`, or `http://127.0.0.1:11434`\n- context window: `OLLAMA_CONTEXT_LENGTH` if set, otherwise Ollama `/api/show` metadata, otherwise `128000`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls Ollama endpoints and normalizes discovered OpenAI-compatible models to `openai-responses`.\n\n`OLLAMA_CONTEXT_LENGTH` does not configure Ollama's runtime `num_ctx`; set that in Ollama/model configuration separately.\n\n### Implicit llama.cpp discovery\n\nIf `llama.cpp` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `llama.cpp`\n- api: `openai-responses`\n- base URL: `LLAMA_CPP_BASE_URL` or `http://127.0.0.1:8080`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls llama.cpp model endpoints and synthesizes model entries with local defaults.\n\n### Implicit LM Studio discovery\n\nIf `lm-studio` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `lm-studio`\n- api: `openai-completions`\n- base URL: `LM_STUDIO_BASE_URL` or `http://127.0.0.1:1234/v1`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery fetches models (`GET /models`) and synthesizes model entries with local defaults.\n\nThis path also works for local OpenAI-compatible servers that are not LM Studio. For example, if oMLX is bound to Ollama's usual port, set `LM_STUDIO_BASE_URL=http://127.0.0.1:11434/v1` to discover it through the existing `/v1/models` flow. Running oMLX and Ollama side by side requires assigning a different port to one of them. Do not configure oMLX as `ollama`: Ollama discovery uses native `/api/tags` and `/api/show` endpoints, not OpenAI `/v1/models`.\n\n### Explicit provider discovery\n\nYou can configure discovery yourself:\n\n```yaml\nproviders:\n  ollama:\n    baseUrl: http://127.0.0.1:11434\n    api: openai-responses\n    auth: none\n    discovery:\n      type: ollama\n\n  llama.cpp:\n    baseUrl: http://127.0.0.1:8080\n    api: openai-responses\n    auth: none\n    discovery:\n      type: llama.cpp\n```\n\n### Proxy discovery (`discovery.type: proxy`)\n\nFor Anthropic+OpenAI-compatible proxies (new-api / one-api / similar)\nthat expose both `/v1/messages` and `/v1/chat/completions` behind the same\nhost. Discovery hits `GET /v1/models` (10s timeout, OpenAI-style payload) and\nderives each model's `api` from the entry's `supported_endpoint_types`:\n\n- contains `\"anthropic\"` -> `api: anthropic-messages` (routes via `/v1/messages`)\n- contains `\"openai\"` -> `api: openai-completions` (routes via `/v1/chat/completions`)\n- otherwise -> falls back to provider-level `api` if set, else dropped\n\nProvider-level `api` is **optional** with `discovery.type: proxy` because the\nper-model wire is auto-detected. The Anthropic SDK strips a trailing `/v1`\nfrom `baseUrl` before appending `/v1/messages`, so a single discovery `baseUrl`\n(ending in `/v1`) round-trips correctly to both wires.\n\n```yaml\nproviders:\n  newapi-reseller:\n    baseUrl: https://api.example.com/v1\n    apiKey: xxxx\n    authHeader: true # injects Authorization: Bearer for openai models\n    disableStrictTools: true # most anthropic-fronted proxies reject `strict`\n    discovery:\n      type: proxy\n```\n\n### Extension provider registration\n\nExtensions can register providers at runtime (`pi.registerProvider(...)`), including:\n\n- model replacement/append for a provider\n- custom stream handler registration for new API IDs\n- custom OAuth provider registration\n\n## Auth and API key resolution order\n\nWhen requesting a key for a provider, effective order is:\n\n1. Runtime override (CLI `--api-key`)\n2. Stored API key credential in `agent.db`\n3. Stored OAuth credential in `agent.db` (with refresh)\n4. Environment variable mapping (`OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, etc.)\n5. ModelRegistry fallback resolver (provider `apiKey` from `models.yml`, env-name-or-literal semantics)\n\n`models.yml` `apiKey` behavior:\n\n- Value is first treated as an environment variable name.\n- If no env var exists, the literal string is used as the token.\n\nIf `authHeader: true` and provider `apiKey` is set, models get:\n\n- `Authorization: Bearer <resolved-key>` header injected.\n\nKeyless providers:\n\n- Providers marked `auth: none` are treated as available without credentials.\n- `getApiKey*` returns `kNoAuth` for them.\n\n### Broker mode\n\nWhen `OMP_AUTH_BROKER_URL` (or `auth.broker.url`) is set, the local SQLite credential store is replaced by `RemoteAuthCredentialStore`. Layers 2 and 3 above (stored API key / OAuth in `agent.db`) are served from a broker-supplied snapshot whose `refresh` tokens are redacted; expiry triggers `POST /v1/credential/:id/refresh` on the broker rather than a local refresh.\n\n`AuthStorage.setConfigApiKey` lets a `models.yml` `apiKey` win over a broker-resolved OAuth token without overriding a runtime `--api-key`. See [`auth-broker-gateway.md`](./auth-broker-gateway.md) for the full broker / gateway design and env surface (`OMP_AUTH_BROKER_URL`, `OMP_AUTH_BROKER_TOKEN`, `auth.broker.url`, `auth.broker.token`).\n\n## Model availability vs all models\n\n- `getAll()` returns the loaded model registry (built-in + merged custom + discovered).\n- `getAvailable()` filters to models that are keyless or have resolvable auth.\n\nSo a model can exist in registry but not be selectable until auth is available.\n\n## Runtime model resolution\n\n### CLI and pattern parsing\n\n`model-resolver.ts` supports:\n\n- exact `provider/modelId`\n- exact canonical model id\n- exact model id (provider inferred)\n- fuzzy/substring matching\n- glob scope patterns in `--models` (e.g. `openai/*`, `*sonnet*`)\n- optional `:thinkingLevel` suffix (`off|minimal|low|medium|high|xhigh`)\n\n`--provider` is legacy; `--model` is preferred.\n\nResolution precedence for exact selectors:\n\n1. exact `provider/modelId` bypasses coalescing\n2. exact canonical id resolves through the canonical index\n3. exact bare concrete id still works\n4. fuzzy and glob matching run after the exact paths\n\n### Initial model selection priority\n\n`findInitialModel(...)` uses this order:\n\n1. explicit CLI provider+model\n2. first scoped model (if not resuming)\n3. saved default provider/model\n4. known provider defaults (e.g. OpenAI/Anthropic/etc.) among available models\n5. first available model\n\n### Role aliases and settings\n\nSupported model roles:\n\n- `default`, `smol`, `slow`, `vision`, `plan`, `designer`, `commit`, `task`\n\nRole aliases like `pi/smol` expand through `settings.modelRoles`. Each role value can also append a thinking selector such as `:minimal`, `:low`, `:medium`, or `:high`.\n\nIf a role points at another role, the target model still inherits normally and any explicit suffix on the referring role wins for that role-specific use.\n\nRelated settings:\n\n- `modelRoles` (record)\n- `enabledModels` (scoped pattern list)\n- `modelProviderOrder` (global canonical-provider precedence)\n- `providers.kimiApiFormat` (`openai` or `anthropic` request format)\n- `providers.openaiWebsockets` (`auto|off|on` websocket preference for OpenAI Codex transport)\n\n`modelRoles` may store either:\n\n- `provider/modelId` to pin a concrete provider variant\n- a canonical id such as `gpt-5.3-codex` to allow provider coalescing\n\nFor `enabledModels` and CLI `--models`:\n\n- exact canonical ids expand to all concrete variants in that canonical group\n- explicit `provider/modelId` entries stay exact\n- globs and fuzzy matches still operate on concrete models\n\nGlobal `enabledModels` and `disabledProviders` entries may also be scoped to a path prefix:\n\n```yaml\nenabledModels:\n  - claude-sonnet-4-5\n  - path: ~/work\n    models:\n      - anthropic/claude-opus-4-5\ndisabledProviders:\n  - ollama\n  - path: ~/private\n    providers:\n      - anthropic\n```\n\nString entries apply everywhere. Scoped entries apply when the current working directory is the configured path or one of its subdirectories. Use `path`, `paths`, `pathPrefix`, or `pathPrefixes`; use `models` for `enabledModels`, `providers` for `disabledProviders`, or `values` for either.\n\n## `/model` and `--list-models`\n\nBoth surfaces keep provider-prefixed models visible and selectable.\n\nThey now also expose canonical/coalesced models:\n\n- `/model` includes a canonical view alongside provider tabs\n- `--list-models` prints a canonical section plus the concrete provider rows\n\nSelecting a canonical entry stores the canonical selector. Selecting a provider row stores the explicit `provider/modelId`.\n\n## Context promotion (model-level fallback chains)\n\nContext promotion is an overflow recovery mechanism for small-context variants (for example `*-spark`) that automatically promotes to a larger-context sibling when the API rejects a request with a context length error.\n\n### Trigger and order\n\nWhen a turn fails with a context overflow error (e.g. `context_length_exceeded`), `AgentSession` attempts promotion **before** falling back to compaction:\n\n1. If `contextPromotion.enabled` is true, resolve a promotion target (see below).\n2. If a target is found, switch to it and retry the request — no compaction needed.\n3. If no target is available, fall through to auto-compaction on the current model.\n\n### Target selection\n\nSelection is model-driven, not role-driven:\n\n1. `currentModel.contextPromotionTarget` (if configured)\n2. smallest larger-context model on the same provider + API\n\nCandidates are ignored unless credentials resolve (`ModelRegistry.getApiKey(...)`).\n\n### OpenAI Codex websocket handoff\n\nIf switching from/to `openai-codex-responses`, session provider state key `openai-codex-responses` is closed before model switch. This drops websocket transport state so the next turn starts clean on the promoted model.\n\n### Persistence behavior\n\nPromotion uses temporary switching (`setModelTemporary`):\n\n- recorded as a temporary `model_change` in session history\n- does not rewrite saved role mapping\n\n### Configuring explicit fallback chains\n\nConfigure fallback directly in model metadata via `contextPromotionTarget`.\n\n`contextPromotionTarget` accepts either:\n\n- `provider/model-id` (explicit)\n- `model-id` (resolved within current provider)\n\nExample (`models.yml`) for an explicit OpenAI fallback:\n\n```yaml\nproviders:\n  openai-codex:\n    modelOverrides:\n      gpt-5.5:\n        contextPromotionTarget: openai-codex/gpt-5.4\n```\n\nThe built-in model policy currently links OpenAI `codex-spark` variants to `gpt-5.5`, and `gpt-5.5` to `gpt-5.4`, when that target exists on the same provider/API.\n\n## Compatibility and routing fields\n\nThe `compat` block on a provider or model overrides the URL-based auto-detection in `packages/catalog/src/compat/openai.ts` (`buildOpenAICompat`). It is validated by `OpenAICompatSchema` in `packages/coding-agent/src/config/models-config-schema.ts` and consumed by every `openai-completions` transport (`packages/ai/src/providers/openai-completions.ts`). The canonical type is `OpenAICompat` in `packages/catalog/src/types.ts`.\n\n`models.yml` accepts the following keys (all optional; unset falls back to URL detection):\n\nRequest shaping:\n\n- `supportsStore` — emit `store: false` on requests. Default: auto (off for non-standard endpoints).\n- `supportsDeveloperRole` — use the `developer` system role for reasoning models instead of `system`. Default: auto.\n- `supportsMultipleSystemMessages` — preserve separate leading system/developer messages instead of coalescing them. Default: auto (known OpenAI-compatible hosted APIs preserve; strict-template/local hosts coalesce).\n- `supportsUsageInStreaming` — send `stream_options: { include_usage: true }` to receive token usage on streaming responses. Default: `true`.\n- `maxTokensField` — `\"max_completion_tokens\"` or `\"max_tokens\"`. Default: auto.\n- `supportsToolChoice` — emit the `tool_choice` parameter when the caller forces a specific tool. Default: `true`. Set `false` for endpoints that 400 on `tool_choice` (e.g. DeepSeek when reasoning is on).\n- `disableReasoningOnForcedToolChoice` — drop `reasoning_effort` / OpenRouter `reasoning` whenever `tool_choice` forces a call. Default: auto (Kimi/Anthropic-fronted endpoints).\n- `disableReasoningOnToolChoice` — drop reasoning fields whenever any `tool_choice` is sent. Default: auto (DeepSeek reasoning models).\n- `alwaysSendMaxTokens` — always send a max-token field when the caller did not provide one. Default: auto (Kimi-family models derive TPM limits from `max_tokens`).\n- `strictResponsesPairing` — Responses-API tool-call/result history must be strictly paired. Default: auto (Azure OpenAI, GitHub Copilot).\n- `streamIdleTimeoutMs` — stream-watchdog idle-timeout floor in ms for slow reasoning hosts. Default: auto (GLM coding-plan hosts, direct DeepSeek reasoning).\n- `cacheControlFormat` — `\"anthropic\"` to include Anthropic-style prompt-cache markers in chat-completions payloads. Default: auto (OpenRouter `anthropic/*` models).\n- `supportsLongPromptCacheRetention` — host honors `prompt_cache_retention: \"24h\"` on the Responses API. Default: auto (api.openai.com).\n- `extraBody` — extra top-level fields merged into every request body (gateway hints, controller selectors, etc.).\n\nReasoning / thinking:\n\n- `supportsReasoningEffort` — accept `reasoning_effort`. Default: auto (off for Grok, Z.ai/Zhipu, and Xiaomi MiMo).\n- `supportsReasoningParams` — whether request shaping may send reasoning params at all. Default: auto (off for GitHub Copilot chat-completions).\n- `reasoningEffortMap` — partial map from internal effort levels (`minimal|low|medium|high|xhigh`) to provider-specific strings (e.g. DeepSeek maps `xhigh -> \"max\"`).\n- `thinkingFormat` — request shape for thinking: `\"openai\"` (`reasoning_effort`), `\"openrouter\"` (`reasoning: { effort }`), `\"zai\"` (`thinking: { type: \"enabled\" }`), `\"qwen\"` (top-level `enable_thinking`), or `\"qwen-chat-template\"` (`chat_template_kwargs.enable_thinking`). Default: `\"openai\"`.\n- `reasoningContentField` — assistant field carrying chain-of-thought: `\"reasoning_content\"`, `\"reasoning\"`, or `\"reasoning_text\"`. Default: auto.\n- `requiresReasoningContentForToolCalls` — assistant tool-call turns must round-trip the reasoning field (DeepSeek-R1, Kimi, OpenRouter when reasoning is on). Default: `false`.\n- `allowsSyntheticReasoningContentForToolCalls` — allow a placeholder reasoning field when a prior assistant tool-call turn lacks provider reasoning content. Default: `true`; set `false` for providers that validate the exact reasoning value.\n- `requiresAssistantContentForToolCalls` — assistant tool-call turns must include non-empty text content (Kimi). Default: `false`.\n- `whenThinking` — partial compat overrides applied only when a request actually engages thinking mode (deep-merged over the baseline compat).\n\nTool / message normalization:\n\n- `requiresToolResultName` — tool-result messages need a `name` field (Mistral). Default: auto.\n- `requiresAssistantAfterToolResult` — a user message after a tool result needs an assistant turn in between. Default: auto.\n- `requiresThinkingAsText` — convert thinking blocks to text wrapped in `<thinking>` delimiters (Mistral). Default: auto.\n- `requiresMistralToolIds` — normalize tool-call ids to exactly 9 alphanumeric chars. Default: auto.\n- `supportsStrictMode` — accept the per-tool `strict` field on tool schemas. Default: conservative auto-detect per provider/baseUrl.\n- `toolStrictMode` — `\"all_strict\"` forces strict on every tool, `\"none\"` forces it off; unset keeps the existing per-tool mixed behavior.\n\nGateway routing (only applied when `baseUrl` matches the gateway):\n\n- `openRouterRouting.only` / `openRouterRouting.order` — provider routing on `openrouter.ai` (see <https://openrouter.ai/docs/provider-routing>).\n- `vercelGatewayRouting.only` / `vercelGatewayRouting.order` — provider routing on `ai-gateway.vercel.sh` (see <https://vercel.com/docs/ai-gateway/models-and-providers/provider-options>).\n\nProvider-level `compat` is the baseline; per-model `compat` is deep-merged on top, with `openRouterRouting`, `vercelGatewayRouting`, and `extraBody` merged as nested objects.\n\n### Anthropic compatibility (`anthropic-messages`)\n\nFor `anthropic-messages` models the runtime uses a separate `AnthropicCompat` shape (`packages/catalog/src/types.ts`). The `models.yml` schema exposes the strict-tools opt-out as a top-level provider field (see below) plus two Anthropic-side flags in the same `compat` slot — `requiresToolResultId` (non-standard `id` alias on `tool_result` blocks for Z.AI-style proxies) and `replayUnsignedThinking` (replay unsigned thinking blocks as native thinking instead of demoting them to text); the remaining Anthropic-side knobs (`disableAdaptiveThinking`, `supportsEagerToolInputStreaming`, `supportsLongCacheRetention`, `supportsMidConversationSystem`, `supportsForcedToolChoice`, `supportsSamplingParams`) are set by built-in catalog metadata and are not user-configurable from `models.yml`.\n\n### Strict tool schemas (`disableStrictTools`)\n\nAnthropic's API supports a `strict` field on tool definitions that forces the model to always follow the provided schema exactly. OMP enables it by default for a small allowlist of high-frequency built-in `anthropic-messages` tools (`bash`, `python`, `edit`, and `find`) whose schemas fit Anthropic's strict grammar limits; other tools still send normalized schemas but omit `strict`.\n\nThird-party providers that front the Anthropic API (AWS Bedrock, Azure, self-hosted proxies) do not always implement this field and will reject requests that include it. Set `disableStrictTools: true` at the provider level to opt out of strict mode for the allowlisted tools:\n\n```yaml\nproviders:\n  bedrock-anthropic:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com/anthropic\n    apiKey: AWS_BEARER_TOKEN\n    api: anthropic-messages\n    disableStrictTools: true\n    models:\n      - id: claude-sonnet-4-20250514\n        name: Claude Sonnet 4 (Bedrock)\n        input: [text, image]\n        contextWindow: 200000\n        maxTokens: 16384\n        cost:\n          input: 3.00\n          output: 15.00\n          cacheRead: 0.30\n          cacheWrite: 3.75\n```\n\n`disableStrictTools` is a provider-level flag that applies to all models in the provider. It disables the Anthropic `strict` marker only for tools that OMP would otherwise mark strict; it does not change runtime tool argument validation. OMP can automatically retry without strict tools after Anthropic reports a strict-grammar-too-large error before the first streamed token, but proxies that reject the `strict` field for other reasons should set this flag explicitly.\n\nTool schemas going on the wire are normalized by the unified flow in\n`packages/ai/src/utils/schema/normalize.ts` (Google/CCA/MCP dispatchers\nplus the OpenAI strict-mode sanitize+enforce pipeline). See\n[`ai-schema-normalize.md`](./ai-schema-normalize.md) for the strict-mode\nedge cases (local `$ref` inlining, single-item `allOf` collapse,\n`anyOf`-wrapper description hoist, enum/const primitive-type inference)\nand the per-provider dispatcher mapping.\n\n## Practical examples\n\n### Local OpenAI-compatible endpoint (no auth)\n\n```yaml\nproviders:\n  local-openai:\n    baseUrl: http://127.0.0.1:8000/v1\n    auth: none\n    api: openai-completions\n    models:\n      - id: Qwen/Qwen2.5-Coder-32B-Instruct\n        name: Qwen 2.5 Coder 32B (local)\n```\n\nFor oMLX or another local OpenAI-compatible server with a discoverable `/v1/models` endpoint, prefer discovery instead of listing models by hand. Set `api` to the endpoint family your server actually exposes: `openai-completions` uses `/v1/chat/completions`; servers that expose `/v1/responses` need `openai-responses` instead.\n\n```yaml\nproviders:\n  omlx:\n    baseUrl: http://127.0.0.1:11434/v1\n    auth: none\n    api: openai-completions\n    discovery:\n      type: openai-models-list\n```\n\n### Hosted proxy with env-based key\n\n```yaml\nproviders:\n  anthropic-proxy:\n    baseUrl: https://proxy.example.com/anthropic\n    apiKey: ANTHROPIC_PROXY_API_KEY\n    api: anthropic-messages\n    authHeader: true\n    disableStrictTools: true # if the proxy doesn't support strict tool schemas\n    models:\n      - id: claude-sonnet-4-20250514\n        name: Claude Sonnet 4 (Proxy)\n        reasoning: true\n        input: [text, image]\n```\n\n### Override built-in provider route + model metadata\n\n```yaml\nproviders:\n  openrouter:\n    baseUrl: https://my-proxy.example.com/v1\n    headers:\n      X-Team: platform\n    modelOverrides:\n      anthropic/claude-sonnet-4:\n        name: Sonnet 4 (Corp)\n        compat:\n          openRouterRouting:\n            only: [anthropic]\n```\n\n## Legacy consumer caveat\n\nMost model configuration now flows through `models.yml` via `ModelRegistry`. Explicit `.json` / `.jsonc` paths remain supported only when passed programmatically to `ModelRegistry`; the default user config is `~/.omp/agent/models.yml`.\n\n## Failure mode\n\nIf `models.yml` fails schema or validation checks:\n\n- registry keeps operating with built-in models\n- error is exposed via `ModelRegistry.getError()` and surfaced in UI/notifications\n",
+	"models.md": "# Model and Provider Configuration (`models.yml`)\n\nThis document describes how the coding-agent currently loads models, applies overrides, resolves credentials, and chooses models at runtime.\n\n## What controls model behavior\n\nPrimary implementation files:\n\n- `src/config/model-registry.ts` — loads built-in + custom models, provider overrides, runtime discovery, auth integration\n- `src/config/model-resolver.ts` — parses model patterns and selects initial/smol/slow models\n- `src/config/settings-schema.ts` — model-related settings (`modelRoles`, provider transport preferences)\n- `src/session/auth-storage.ts` — API key + OAuth resolution order\n- `packages/catalog/src/models.ts` and `packages/catalog/src/types.ts` — built-in providers/models (`getBundledModels` / `getBundledProviders`) and `Model`/`compat` types\n\n## Config file location and legacy behavior\n\nDefault config path:\n\n- `~/.omp/agent/models.yml`\n\nLegacy behavior still present:\n\n- If `models.yml` is missing and `models.json` exists at the same location, it is migrated to `models.yml`.\n- Explicit `.json` / `.jsonc` config paths are still supported when passed programmatically to `ModelRegistry`.\n\n## `models.yml` shape\n\n```yaml\nproviders:\n  <provider-id>:\n    # provider-level config\nequivalence:\n  overrides:\n    <provider-id>/<model-id>: <canonical-model-id>\n  exclude:\n    - <provider-id>/<model-id>\n```\n\n`provider-id` is the canonical provider key used across selection and auth lookup.\n\n`equivalence` is optional and configures canonical model grouping on top of concrete provider models:\n\n- `overrides` maps an exact concrete selector (`provider/modelId`) to an official upstream canonical id\n- `exclude` opts a concrete selector out of canonical grouping\n\n## Provider-level fields\n\n```yaml\nproviders:\n  my-provider:\n    baseUrl: https://api.example.com/v1\n    apiKey: MY_PROVIDER_API_KEY\n    api: openai-completions\n    headers:\n      X-Team: platform\n    authHeader: true\n    auth: apiKey\n    disableStrictTools: false # set true for Anthropic-compatible endpoints that reject the strict field\n    discovery:\n      type: ollama\n    modelOverrides:\n      some-model-id:\n        name: Renamed model\n    models:\n      - id: some-model-id\n        name: Some Model\n        api: openai-completions\n        reasoning: false\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 128000\n        maxTokens: 16384\n        headers:\n          X-Model: value\n        compat:\n          supportsStore: true\n          supportsDeveloperRole: true\n          supportsReasoningEffort: true\n          maxTokensField: max_completion_tokens\n          openRouterRouting:\n            only: [anthropic]\n          vercelGatewayRouting:\n            order: [anthropic, openai]\n          extraBody:\n            gateway: m1-01\n            controller: mlx\n```\n\n### Allowed provider/model `api` values\n\n- `openai-completions`\n- `openai-responses`\n- `openai-codex-responses`\n- `azure-openai-responses`\n- `anthropic-messages`\n- `google-generative-ai`\n- `google-vertex`\n\n### Allowed auth/discovery values\n\n- `auth`: `apiKey` (default), `none`, or `oauth`; for `models.yml` custom models, `oauth` is accepted by schema but does not waive the `apiKey` requirement\n- `discovery.type`: `ollama`, `llama.cpp`, `lm-studio`, `openai-models-list`, or `proxy`\n- `transport`: `pi-native` only. When set, every model under that provider is sent to an `omp auth-gateway` compatible `baseUrl` via `POST /v1/pi/stream`; `apiKey` is the gateway bearer.\n\n## Validation rules (current)\n\n### Full custom provider (`models` is non-empty)\n\nRequired:\n\n- `baseUrl`\n- `apiKey` unless `auth: none`\n- `api` at provider level or each model\n\n### Override-only provider (`models` missing or empty)\n\nMust define at least one of:\n\n- `baseUrl`\n- `apiKey`\n- `headers`\n- `compat`\n- `disableStrictTools`\n- `modelOverrides`\n- `discovery`\n\n### Discovery\n\n- `discovery` requires provider-level `api`, except `discovery.type: proxy` (per-model wire auto-detected).\n\n### Model value checks\n\n- `id` required\n- `contextWindow` and `maxTokens` must be positive if provided\n\n### Command-resolved secrets\n\nProvider `apiKey` values and provider/model `headers` values may start with `!` to read a secret from command stdout. The command is run with a 10 s timeout, stdout is trimmed, and empty/failing commands are omitted:\n\n```yaml\nproviders:\n  openai:\n    apiKey: \"!op read op://dev/openai/api-key\"\n    headers:\n      X-Team-Key: \"!bw get password omp-team-key\"\n```\n\nSuccessful command outputs are cached for the process lifetime so the command is not re-run for every model.\n\n## Merge and override order\n\nModelRegistry pipeline (on refresh):\n\n1. Load built-in providers/models from `@oh-my-pi/pi-ai`.\n2. Load `models.yml` custom config.\n3. Apply provider overrides (`baseUrl`, `headers`, `disableStrictTools`) to built-in models.\n4. Apply `modelOverrides` (per provider + model id).\n5. Merge custom `models`:\n   - same `provider + id` replaces existing\n   - otherwise append\n6. Load cached/runtime-discovered models (Ollama, llama.cpp, LM Studio, plus built-in provider managers), then re-apply model overrides.\n\n### Provider-model cache and static fingerprint\n\nCached per-provider model lists are persisted in the model-cache SQLite\ndatabase (current schema version 5) with a `static_fingerprint` column that\nhashes the static catalog slice merged into the row. When `resolveProviderModels`\nskips the network fetch and the fingerprint of the in-memory static\ncatalog matches the cached one, the cached rows are returned verbatim —\nthe static + dynamic merge is bypassed entirely. The fingerprint is\nmemoized per process by tagging the static-models array with a symbol\nproperty, so repeated cold-start calls do not re-hash.\n\n## Canonical model equivalence and coalescing\n\nThe registry keeps every concrete provider model and then builds a canonical layer above them.\n\nCanonical ids are official upstream ids only, for example:\n\n- `claude-opus-4-6`\n- `claude-haiku-4-5`\n- `gpt-5.3-codex`\n\n### `models.yml` equivalence config\n\nExample:\n\n```yaml\nproviders:\n  zenmux:\n    baseUrl: https://api.zenmux.example/v1\n    apiKey: ZENMUX_API_KEY\n    api: openai-codex-responses\n    models:\n      - id: codex\n        name: Zenmux Codex\n        reasoning: true\n        input: [text]\n        cost:\n          input: 0\n          output: 0\n          cacheRead: 0\n          cacheWrite: 0\n        contextWindow: 200000\n        maxTokens: 32768\n\nequivalence:\n  overrides:\n    zenmux/codex: gpt-5.3-codex\n    p-codex/codex: gpt-5.3-codex\n  exclude:\n    - demo/codex-preview\n```\n\nBuild order for canonical grouping:\n\n1. exact user override from `equivalence.overrides`\n2. bundled official-id matches from built-in model metadata\n3. conservative heuristic normalization for gateway/provider variants\n4. fallback to the concrete model's own id\n\nCurrent heuristics are intentionally narrow:\n\n- embedded upstream prefixes can be stripped when present, for example `anthropic/...` or `openai/...`\n- dotted and dashed version variants can normalize only when they map to an existing official id, for example `4.6 -> 4-6`\n- ambiguous families or versions are not merged without a bundled match or explicit override\n\n### Canonical resolution behavior\n\nWhen multiple concrete variants share a canonical id, resolution uses:\n\n1. availability and auth\n2. `config.yml` `modelProviderOrder`\n3. existing registry/provider order if `modelProviderOrder` is unset\n\nDisabled or unauthenticated providers are skipped.\n\nSession state and transcripts continue to record the concrete provider/model that actually executed the turn.\n\nProvider defaults vs per-model overrides:\n\n- Provider `headers` are baseline.\n- Model `headers` override provider header keys.\n- `modelOverrides` can override model metadata (`name`, `reasoning`, `thinking`, `input`, `cost`, `premiumMultiplier`, `contextWindow`, `maxTokens`, `headers`, `compat`, `contextPromotionTarget`).\n- `compat` is deep-merged for nested routing blocks (`openRouterRouting`, `vercelGatewayRouting`, `extraBody`).\n\n## Runtime discovery integration\n\n### Implicit Ollama discovery\n\nIf `ollama` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `ollama`\n- api: `openai-responses`\n- base URL: `OLLAMA_BASE_URL`, or `OLLAMA_HOST`, or `http://127.0.0.1:11434`\n- context window: `OLLAMA_CONTEXT_LENGTH` if set, otherwise Ollama `/api/show` metadata, otherwise `128000`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls Ollama endpoints and normalizes discovered OpenAI-compatible models to `openai-responses`.\n\n`OLLAMA_CONTEXT_LENGTH` does not configure Ollama's runtime `num_ctx`; set that in Ollama/model configuration separately.\n\n### Implicit llama.cpp discovery\n\nIf `llama.cpp` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `llama.cpp`\n- api: `openai-responses`\n- base URL: `LLAMA_CPP_BASE_URL` or `http://127.0.0.1:8080`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls llama.cpp model endpoints and synthesizes model entries with local defaults.\n\n### Implicit LM Studio discovery\n\nIf `lm-studio` is not explicitly configured, registry adds an implicit discoverable provider:\n\n- provider: `lm-studio`\n- api: `openai-completions`\n- base URL: `LM_STUDIO_BASE_URL` or `http://127.0.0.1:1234/v1`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery fetches models (`GET /models`) and synthesizes model entries with local defaults.\n\nThis path also works for local OpenAI-compatible servers that are not LM Studio. For example, if oMLX is bound to Ollama's usual port, set `LM_STUDIO_BASE_URL=http://127.0.0.1:11434/v1` to discover it through the existing `/v1/models` flow. Running oMLX and Ollama side by side requires assigning a different port to one of them. Do not configure oMLX as `ollama`: Ollama discovery uses native `/api/tags` and `/api/show` endpoints, not OpenAI `/v1/models`.\n\n### Explicit provider discovery\n\nYou can configure discovery yourself:\n\n```yaml\nproviders:\n  ollama:\n    baseUrl: http://127.0.0.1:11434\n    api: openai-responses\n    auth: none\n    discovery:\n      type: ollama\n\n  llama.cpp:\n    baseUrl: http://127.0.0.1:8080\n    api: openai-responses\n    auth: none\n    discovery:\n      type: llama.cpp\n```\n\n### Proxy discovery (`discovery.type: proxy`)\n\nFor Anthropic+OpenAI-compatible proxies (new-api / one-api / similar)\nthat expose both `/v1/messages` and `/v1/chat/completions` behind the same\nhost. Discovery hits `GET /v1/models` (10s timeout, OpenAI-style payload) and\nderives each model's `api` from the entry's `supported_endpoint_types`:\n\n- contains `\"anthropic\"` -> `api: anthropic-messages` (routes via `/v1/messages`)\n- contains `\"openai\"` -> `api: openai-completions` (routes via `/v1/chat/completions`)\n- otherwise -> falls back to provider-level `api` if set, else dropped\n\nProvider-level `api` is **optional** with `discovery.type: proxy` because the\nper-model wire is auto-detected. The Anthropic SDK strips a trailing `/v1`\nfrom `baseUrl` before appending `/v1/messages`, so a single discovery `baseUrl`\n(ending in `/v1`) round-trips correctly to both wires.\n\n```yaml\nproviders:\n  newapi-reseller:\n    baseUrl: https://api.example.com/v1\n    apiKey: xxxx\n    authHeader: true # injects Authorization: Bearer for openai models\n    disableStrictTools: true # most anthropic-fronted proxies reject `strict`\n    discovery:\n      type: proxy\n```\n\n### Extension provider registration\n\nExtensions can register providers at runtime (`pi.registerProvider(...)`), including:\n\n- model replacement/append for a provider\n- custom stream handler registration for new API IDs\n- custom OAuth provider registration\n\n## Auth and API key resolution order\n\nWhen requesting a key for a provider, effective order is:\n\n1. Runtime override (CLI `--api-key`)\n2. Stored API key credential in `agent.db`\n3. Stored OAuth credential in `agent.db` (with refresh)\n4. Environment variable mapping (`OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, etc.)\n5. ModelRegistry fallback resolver (provider `apiKey` from `models.yml`, env-name-or-literal semantics)\n\n`models.yml` `apiKey` behavior:\n\n- Value is first treated as an environment variable name.\n- If no env var exists, the literal string is used as the token.\n\nIf `authHeader: true` and provider `apiKey` is set, models get:\n\n- `Authorization: Bearer <resolved-key>` header injected.\n\nKeyless providers:\n\n- Providers marked `auth: none` are treated as available without credentials.\n- `getApiKey*` returns `kNoAuth` for them.\n\n### Broker mode\n\nWhen `OMP_AUTH_BROKER_URL` (or `auth.broker.url`) is set, the local SQLite credential store is replaced by `RemoteAuthCredentialStore`. Layers 2 and 3 above (stored API key / OAuth in `agent.db`) are served from a broker-supplied snapshot whose `refresh` tokens are redacted; expiry triggers `POST /v1/credential/:id/refresh` on the broker rather than a local refresh.\n\n`AuthStorage.setConfigApiKey` lets a `models.yml` `apiKey` win over a broker-resolved OAuth token without overriding a runtime `--api-key`. See [`auth-broker-gateway.md`](./auth-broker-gateway.md) for the full broker / gateway design and env surface (`OMP_AUTH_BROKER_URL`, `OMP_AUTH_BROKER_TOKEN`, `auth.broker.url`, `auth.broker.token`).\n\n## Model availability vs all models\n\n- `getAll()` returns the loaded model registry (built-in + merged custom + discovered).\n- `getAvailable()` filters to models that are keyless or have resolvable auth.\n\nSo a model can exist in registry but not be selectable until auth is available.\n\n## Runtime model resolution\n\n### CLI and pattern parsing\n\n`model-resolver.ts` supports:\n\n- exact `provider/modelId`\n- exact canonical model id\n- exact model id (provider inferred)\n- fuzzy/substring matching\n- glob scope patterns in `--models` (e.g. `openai/*`, `*sonnet*`)\n- optional `:thinkingLevel` suffix (`off|minimal|low|medium|high|xhigh`)\n\n`--provider` is legacy; `--model` is preferred.\n\nResolution precedence for exact selectors:\n\n1. exact `provider/modelId` bypasses coalescing\n2. exact canonical id resolves through the canonical index\n3. exact bare concrete id still works\n4. fuzzy and glob matching run after the exact paths\n\n### Initial model selection priority\n\n`findInitialModel(...)` uses this order:\n\n1. explicit CLI provider+model\n2. first scoped model (if not resuming)\n3. saved default provider/model\n4. known provider defaults (e.g. OpenAI/Anthropic/etc.) among available models\n5. first available model\n\n### Role aliases and settings\n\nSupported model roles:\n\n- `default`, `smol`, `slow`, `vision`, `plan`, `designer`, `commit`, `task`\n\nRole aliases like `pi/smol` expand through `settings.modelRoles`. Each role value can also append a thinking selector such as `:minimal`, `:low`, `:medium`, or `:high`.\n\nIf a role points at another role, the target model still inherits normally and any explicit suffix on the referring role wins for that role-specific use.\n\nRelated settings:\n\n- `modelRoles` (record)\n- `enabledModels` (scoped pattern list)\n- `modelProviderOrder` (global canonical-provider precedence)\n- `providers.kimiApiFormat` (`openai` or `anthropic` request format)\n- `providers.openaiWebsockets` (`auto|off|on` websocket preference for OpenAI Codex transport)\n\n`modelRoles` may store either:\n\n- `provider/modelId` to pin a concrete provider variant\n- a canonical id such as `gpt-5.3-codex` to allow provider coalescing\n\nFor `enabledModels` and CLI `--models`:\n\n- exact canonical ids expand to all concrete variants in that canonical group\n- explicit `provider/modelId` entries stay exact\n- globs and fuzzy matches still operate on concrete models\n\nGlobal `enabledModels` and `disabledProviders` entries may also be scoped to a path prefix:\n\n```yaml\nenabledModels:\n  - claude-sonnet-4-5\n  - path: ~/work\n    models:\n      - anthropic/claude-opus-4-5\ndisabledProviders:\n  - ollama\n  - path: ~/private\n    providers:\n      - anthropic\n```\n\nString entries apply everywhere. Scoped entries apply when the current working directory is the configured path or one of its subdirectories. Use `path`, `paths`, `pathPrefix`, or `pathPrefixes`; use `models` for `enabledModels`, `providers` for `disabledProviders`, or `values` for either.\n\n## `/model` and `omp models`\n\nBoth surfaces keep provider-prefixed models visible and selectable.\n\nThey now also expose canonical/coalesced models:\n\n- `/model` includes a canonical view alongside provider tabs\n- `omp models` prints provider-grouped tables of every concrete model, and `omp models canonical` prints the coalesced canonical view\n\nSelecting a canonical entry stores the canonical selector. Selecting a provider row stores the explicit `provider/modelId`.\n\n## Context promotion (model-level fallback chains)\n\nContext promotion is an overflow recovery mechanism for small-context variants (for example `*-spark`) that automatically promotes to a larger-context sibling when the API rejects a request with a context length error.\n\n### Trigger and order\n\nWhen a turn fails with a context overflow error (e.g. `context_length_exceeded`), `AgentSession` attempts promotion **before** falling back to compaction:\n\n1. If `contextPromotion.enabled` is true, resolve a promotion target (see below).\n2. If a target is found, switch to it and retry the request — no compaction needed.\n3. If no target is available, fall through to auto-compaction on the current model.\n\n### Target selection\n\nSelection is model-driven, not role-driven:\n\n1. `currentModel.contextPromotionTarget` (if configured)\n2. smallest larger-context model on the same provider + API\n\nCandidates are ignored unless credentials resolve (`ModelRegistry.getApiKey(...)`).\n\n### OpenAI Codex websocket handoff\n\nIf switching from/to `openai-codex-responses`, session provider state key `openai-codex-responses` is closed before model switch. This drops websocket transport state so the next turn starts clean on the promoted model.\n\n### Persistence behavior\n\nPromotion uses temporary switching (`setModelTemporary`):\n\n- recorded as a temporary `model_change` in session history\n- does not rewrite saved role mapping\n\n### Configuring explicit fallback chains\n\nConfigure fallback directly in model metadata via `contextPromotionTarget`.\n\n`contextPromotionTarget` accepts either:\n\n- `provider/model-id` (explicit)\n- `model-id` (resolved within current provider)\n\nExample (`models.yml`) for an explicit OpenAI fallback:\n\n```yaml\nproviders:\n  openai-codex:\n    modelOverrides:\n      gpt-5.5:\n        contextPromotionTarget: openai-codex/gpt-5.4\n```\n\nThe built-in model policy currently links OpenAI `codex-spark` variants to `gpt-5.5`, and `gpt-5.5` to `gpt-5.4`, when that target exists on the same provider/API.\n\n## Compatibility and routing fields\n\nThe `compat` block on a provider or model overrides the URL-based auto-detection in `packages/catalog/src/compat/openai.ts` (`buildOpenAICompat`). It is validated by `OpenAICompatSchema` in `packages/coding-agent/src/config/models-config-schema.ts` and consumed by every `openai-completions` transport (`packages/ai/src/providers/openai-completions.ts`). The canonical type is `OpenAICompat` in `packages/catalog/src/types.ts`.\n\n`models.yml` accepts the following keys (all optional; unset falls back to URL detection):\n\nRequest shaping:\n\n- `supportsStore` — emit `store: false` on requests. Default: auto (off for non-standard endpoints).\n- `supportsDeveloperRole` — use the `developer` system role for reasoning models instead of `system`. Default: auto.\n- `supportsMultipleSystemMessages` — preserve separate leading system/developer messages instead of coalescing them. Default: auto (known OpenAI-compatible hosted APIs preserve; strict-template/local hosts coalesce).\n- `supportsUsageInStreaming` — send `stream_options: { include_usage: true }` to receive token usage on streaming responses. Default: `true`.\n- `maxTokensField` — `\"max_completion_tokens\"` or `\"max_tokens\"`. Default: auto.\n- `supportsToolChoice` — emit the `tool_choice` parameter when the caller forces a specific tool. Default: `true`. Set `false` for endpoints that 400 on `tool_choice` (e.g. DeepSeek when reasoning is on).\n- `disableReasoningOnForcedToolChoice` — drop `reasoning_effort` / OpenRouter `reasoning` whenever `tool_choice` forces a call. Default: auto (Kimi/Anthropic-fronted endpoints).\n- `disableReasoningOnToolChoice` — drop reasoning fields whenever any `tool_choice` is sent. Default: auto (DeepSeek reasoning models).\n- `alwaysSendMaxTokens` — always send a max-token field when the caller did not provide one. Default: auto (Kimi-family models derive TPM limits from `max_tokens`).\n- `strictResponsesPairing` — Responses-API tool-call/result history must be strictly paired. Default: auto (Azure OpenAI, GitHub Copilot).\n- `streamIdleTimeoutMs` — stream-watchdog idle-timeout floor in ms for slow reasoning hosts. Default: auto (GLM coding-plan hosts, direct DeepSeek reasoning).\n- `cacheControlFormat` — `\"anthropic\"` to include Anthropic-style prompt-cache markers in chat-completions payloads. Default: auto (OpenRouter `anthropic/*` models).\n- `supportsLongPromptCacheRetention` — host honors `prompt_cache_retention: \"24h\"` on the Responses API. Default: auto (api.openai.com).\n- `extraBody` — extra top-level fields merged into every request body (gateway hints, controller selectors, etc.).\n\nReasoning / thinking:\n\n- `supportsReasoningEffort` — accept `reasoning_effort`. Default: auto (off for Grok, Z.ai/Zhipu, and Xiaomi MiMo).\n- `supportsReasoningParams` — whether request shaping may send reasoning params at all. Default: auto (off for GitHub Copilot chat-completions).\n- `reasoningEffortMap` — partial map from internal effort levels (`minimal|low|medium|high|xhigh`) to provider-specific strings (e.g. DeepSeek maps `xhigh -> \"max\"`).\n- `thinkingFormat` — request shape for thinking: `\"openai\"` (`reasoning_effort`), `\"openrouter\"` (`reasoning: { effort }`), `\"zai\"` (`thinking: { type: \"enabled\" }`), `\"qwen\"` (top-level `enable_thinking`), or `\"qwen-chat-template\"` (`chat_template_kwargs.enable_thinking`). Default: `\"openai\"`.\n- `reasoningContentField` — assistant field carrying chain-of-thought: `\"reasoning_content\"`, `\"reasoning\"`, or `\"reasoning_text\"`. Default: auto.\n- `requiresReasoningContentForToolCalls` — assistant tool-call turns must round-trip the reasoning field (DeepSeek-R1, Kimi, OpenRouter when reasoning is on). Default: `false`.\n- `allowsSyntheticReasoningContentForToolCalls` — allow a placeholder reasoning field when a prior assistant tool-call turn lacks provider reasoning content. Default: `true`; set `false` for providers that validate the exact reasoning value.\n- `requiresAssistantContentForToolCalls` — assistant tool-call turns must include non-empty text content (Kimi). Default: `false`.\n- `whenThinking` — partial compat overrides applied only when a request actually engages thinking mode (deep-merged over the baseline compat).\n\nTool / message normalization:\n\n- `requiresToolResultName` — tool-result messages need a `name` field (Mistral). Default: auto.\n- `requiresAssistantAfterToolResult` — a user message after a tool result needs an assistant turn in between. Default: auto.\n- `requiresThinkingAsText` — convert thinking blocks to text wrapped in `<thinking>` delimiters (Mistral). Default: auto.\n- `requiresMistralToolIds` — normalize tool-call ids to exactly 9 alphanumeric chars. Default: auto.\n- `supportsStrictMode` — accept the per-tool `strict` field on tool schemas. Default: conservative auto-detect per provider/baseUrl.\n- `toolStrictMode` — `\"all_strict\"` forces strict on every tool, `\"none\"` forces it off; unset keeps the existing per-tool mixed behavior.\n\nGateway routing (only applied when `baseUrl` matches the gateway):\n\n- `openRouterRouting.only` / `openRouterRouting.order` — provider routing on `openrouter.ai` (see <https://openrouter.ai/docs/provider-routing>).\n- `vercelGatewayRouting.only` / `vercelGatewayRouting.order` — provider routing on `ai-gateway.vercel.sh` (see <https://vercel.com/docs/ai-gateway/models-and-providers/provider-options>).\n\nProvider-level `compat` is the baseline; per-model `compat` is deep-merged on top, with `openRouterRouting`, `vercelGatewayRouting`, and `extraBody` merged as nested objects.\n\n### Anthropic compatibility (`anthropic-messages`)\n\nFor `anthropic-messages` models the runtime uses a separate `AnthropicCompat` shape (`packages/catalog/src/types.ts`). The `models.yml` schema exposes the strict-tools opt-out as a top-level provider field (see below) plus two Anthropic-side flags in the same `compat` slot — `requiresToolResultId` (non-standard `id` alias on `tool_result` blocks for Z.AI-style proxies) and `replayUnsignedThinking` (replay unsigned thinking blocks as native thinking instead of demoting them to text); the remaining Anthropic-side knobs (`disableAdaptiveThinking`, `supportsEagerToolInputStreaming`, `supportsLongCacheRetention`, `supportsMidConversationSystem`, `supportsForcedToolChoice`, `supportsSamplingParams`) are set by built-in catalog metadata and are not user-configurable from `models.yml`.\n\n### Strict tool schemas (`disableStrictTools`)\n\nAnthropic's API supports a `strict` field on tool definitions that forces the model to always follow the provided schema exactly. OMP enables it by default for a small allowlist of high-frequency built-in `anthropic-messages` tools (`bash`, `python`, `edit`, and `find`) whose schemas fit Anthropic's strict grammar limits; other tools still send normalized schemas but omit `strict`.\n\nThird-party providers that front the Anthropic API (AWS Bedrock, Azure, self-hosted proxies) do not always implement this field and will reject requests that include it. Set `disableStrictTools: true` at the provider level to opt out of strict mode for the allowlisted tools:\n\n```yaml\nproviders:\n  bedrock-anthropic:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com/anthropic\n    apiKey: AWS_BEARER_TOKEN\n    api: anthropic-messages\n    disableStrictTools: true\n    models:\n      - id: claude-sonnet-4-20250514\n        name: Claude Sonnet 4 (Bedrock)\n        input: [text, image]\n        contextWindow: 200000\n        maxTokens: 16384\n        cost:\n          input: 3.00\n          output: 15.00\n          cacheRead: 0.30\n          cacheWrite: 3.75\n```\n\n`disableStrictTools` is a provider-level flag that applies to all models in the provider. It disables the Anthropic `strict` marker only for tools that OMP would otherwise mark strict; it does not change runtime tool argument validation. OMP can automatically retry without strict tools after Anthropic reports a strict-grammar-too-large error before the first streamed token, but proxies that reject the `strict` field for other reasons should set this flag explicitly.\n\nTool schemas going on the wire are normalized by the unified flow in\n`packages/ai/src/utils/schema/normalize.ts` (Google/CCA/MCP dispatchers\nplus the OpenAI strict-mode sanitize+enforce pipeline). See\n[`ai-schema-normalize.md`](./ai-schema-normalize.md) for the strict-mode\nedge cases (local `$ref` inlining, single-item `allOf` collapse,\n`anyOf`-wrapper description hoist, enum/const primitive-type inference)\nand the per-provider dispatcher mapping.\n\n## Practical examples\n\n### Local OpenAI-compatible endpoint (no auth)\n\n```yaml\nproviders:\n  local-openai:\n    baseUrl: http://127.0.0.1:8000/v1\n    auth: none\n    api: openai-completions\n    models:\n      - id: Qwen/Qwen2.5-Coder-32B-Instruct\n        name: Qwen 2.5 Coder 32B (local)\n```\n\nFor oMLX or another local OpenAI-compatible server with a discoverable `/v1/models` endpoint, prefer discovery instead of listing models by hand. Set `api` to the endpoint family your server actually exposes: `openai-completions` uses `/v1/chat/completions`; servers that expose `/v1/responses` need `openai-responses` instead.\n\n```yaml\nproviders:\n  omlx:\n    baseUrl: http://127.0.0.1:11434/v1\n    auth: none\n    api: openai-completions\n    discovery:\n      type: openai-models-list\n```\n\nThe built-in vLLM provider can be pointed at a non-default endpoint without declaring a custom discovery type. OMP uses vLLM's `/v1/models` metadata and preserves vLLM's `max_model_len` field as the discovered context window.\n\n```yaml\nproviders:\n  vllm:\n    baseUrl: http://192.168.5.3:8085/v1\n    auth: none\n```\n\nFor multiple vLLM endpoints, use arbitrary provider IDs with the generic OpenAI-compatible discovery path. Set `auth: none` for local no-auth servers or `apiKey` for authenticated ones. Generic discovery reads `max_model_len` first and then `context_length` as a generic OpenAI-compatible fallback.\n\n```yaml\nproviders:\n  vllm-fast:\n    baseUrl: http://host-a:8000/v1\n    auth: none\n    api: openai-completions\n    discovery:\n      type: openai-models-list\n  vllm-long:\n    baseUrl: http://host-b:8000/v1\n    auth: none\n    api: openai-completions\n    discovery:\n      type: openai-models-list\n```\n\n### Hosted proxy with env-based key\n\n```yaml\nproviders:\n  anthropic-proxy:\n    baseUrl: https://proxy.example.com/anthropic\n    apiKey: ANTHROPIC_PROXY_API_KEY\n    api: anthropic-messages\n    authHeader: true\n    disableStrictTools: true # if the proxy doesn't support strict tool schemas\n    models:\n      - id: claude-sonnet-4-20250514\n        name: Claude Sonnet 4 (Proxy)\n        reasoning: true\n        input: [text, image]\n```\n\n### Override built-in provider route + model metadata\n\n```yaml\nproviders:\n  openrouter:\n    baseUrl: https://my-proxy.example.com/v1\n    headers:\n      X-Team: platform\n    modelOverrides:\n      anthropic/claude-sonnet-4:\n        name: Sonnet 4 (Corp)\n        compat:\n          openRouterRouting:\n            only: [anthropic]\n```\n\n## Legacy consumer caveat\n\nMost model configuration now flows through `models.yml` via `ModelRegistry`. Explicit `.json` / `.jsonc` paths remain supported only when passed programmatically to `ModelRegistry`; the default user config is `~/.omp/agent/models.yml`.\n\n## Failure mode\n\nIf `models.yml` fails schema or validation checks:\n\n- registry keeps operating with built-in models\n- error is exposed via `ModelRegistry.getError()` and surfaced in UI/notifications\n",
 	"natives-addon-loader-runtime.md": "# Natives Addon Loader Runtime\n\nThis document covers the runtime loader shipped by `@oh-my-pi/pi-natives`: how `native/index.js` decides which `.node` file to require, how compiled-binary embedded payloads are extracted, and what startup failures report.\n\n## Implementation files\n\n- `packages/natives/native/index.js`\n- `packages/natives/native/loader-state.js`\n- `packages/natives/native/embedded-addon.js`\n- `packages/natives/scripts/embed-native.ts`\n- `packages/natives/package.json`\n\n## Scope and responsibility\n\nThe loader is intentionally narrow:\n\n- Build a platform/CPU-aware candidate list for addon filenames and directories.\n- Treat an embedded-addon manifest as a compiled-binary signal when present.\n- Optionally materialize embedded addon archive contents into a versioned per-user cache directory.\n- On Windows `node_modules` installs, stage addon files into the versioned cache to avoid locked-DLL update failures.\n- Attempt candidates in deterministic order and return the first addon that `require(...)` loads and validates.\n\nFor install and compiled-binary paths, the loader verifies a release sentinel export named from `package.json#version` (for example `__piNativesV15_7_2`). Workspace-dev loads skip this validation so a local checkout can rebuild after a pull. The loader does not validate the full export surface; stale same-version or incomplete binaries still surface as missing members or native errors at use sites.\n\n## Runtime inputs and derived state\n\nAt module initialization, `native/index.js` computes:\n\n- **Platform tag**: `${process.platform}-${process.arch}` (for example `darwin-arm64`).\n- **Package version**: from `packages/natives/package.json`.\n- **Core directories**:\n  - `leafPackageDir`: directory of the platform leaf package, resolved via `require.resolve(\"@oh-my-pi/pi-natives-<tag>/package.json\")`; `null` when no leaf is installed (e.g. local dev) and forced to `null` in compiled-binary mode.\n  - `nativeDir`: package-local `packages/natives/native`.\n  - `execDir`: directory containing `process.execPath`.\n  - `versionedDir`: `<getNativesDir()>/<packageVersion>`.\n  - `userDataDir` fallback:\n    - Windows: `%LOCALAPPDATA%/omp` or `%USERPROFILE%/AppData/Local/omp`.\n    - Non-Windows: `~/.local/bin`.\n- **Natives cache root** (`getNativesDir()`):\n  - if `$XDG_DATA_HOME/omp` exists, `$XDG_DATA_HOME/omp/natives`;\n  - otherwise `~/.omp/natives`.\n- **Compiled-binary mode** (`detectCompiledBinary`): true if any of:\n  - embedded-addon manifest is non-null,\n  - `PI_COMPILED` env var is set,\n  - `import.meta.url` contains Bun embedded markers (`$bunfs`, `~BUN`, `%7EBUN`).\n- **Windows staging mode** (`shouldStageNodeModulesAddon`): true only on Windows, in non-compiled mode, when `nativeDir` is inside `node_modules`.\n- **Variant override**: `PI_NATIVE_VARIANT` (`modern`/`baseline` only; invalid values ignored).\n- **Selected variant**: explicit override, otherwise runtime AVX2 detection on x64 (`modern` if AVX2, else `baseline`).\n\n## Platform support and tag resolution\n\n`SUPPORTED_PLATFORMS` is fixed to:\n\n- `linux-x64`\n- `linux-arm64`\n- `darwin-x64`\n- `darwin-arm64`\n- `win32-x64`\n\nUnsupported platforms are not rejected before probing. The loader first tries the computed candidate paths. If all fail and `platformTag` is unsupported, it throws an unsupported-platform error listing supported tags.\n\n## Variant selection (`modern` / `baseline` / default)\n\n### x64 behavior\n\n1. `PI_NATIVE_VARIANT=modern|baseline` wins when valid.\n2. Otherwise AVX2 support is detected:\n   - Linux: scan `/proc/cpuinfo` for `avx2`.\n   - macOS: `sysctl -n machdep.cpu.leaf7_features`, then `machdep.cpu.features`.\n   - Windows: PowerShell `[System.Runtime.Intrinsics.X86.Avx2]::IsSupported`.\n3. AVX2 selects `modern`; unavailable or undetectable AVX2 selects `baseline`.\n\n### Non-x64 behavior\n\nNo variant suffix is used; the filename is `pi_natives.<platform>-<arch>.node`.\n\n### Filename construction\n\n`loader-state.js#getAddonFilenames` returns:\n\n- Non-x64 or no variant: `pi_natives.<tag>.node`\n- x64 + `modern`:\n  1. `pi_natives.<tag>-modern.node`\n  2. `pi_natives.<tag>-baseline.node`\n  3. `pi_natives.<tag>.node`\n- x64 + `baseline`:\n  1. `pi_natives.<tag>-baseline.node`\n  2. `pi_natives.<tag>.node`\n\nThe default unsuffixed fallback remains part of the x64 candidate list.\n\n## Candidate path construction and fallback ordering\n\n`resolveLoaderCandidates(...)` expands every filename across directories, then de-duplicates while preserving first occurrence order.\n\n### Non-compiled runtime\n\nCandidates are grouped by directory class, in order:\n\n1. `<leafPackageDir>/<filename>` for every filename (omitted when `leafPackageDir` is `null`)\n2. `<nativeDir>/<filename>` then `<execDir>/<filename>`, per filename\n\nThe leaf package dir comes first so the optional-dependency binary published with the release is preferred over any `.node` left in the core package's `native/` (e.g. a stale local-dev build).\n\nOn Windows installs where `nativeDir` is inside a `node_modules` segment (`shouldStageNodeModulesAddon`), `<versionedDir>/<filename>` staging candidates are prepended ahead of the leaf candidates so a locked `node_modules` binary can be sidestepped during `bun install -g` updates. The staged file is copied from `leafPackageDir ?? nativeDir` before probing.\n\n### Compiled runtime\n\nCandidates are grouped, in order:\n\n1. `<versionedDir>/<filename>` then `<userDataDir>/<filename>`, per filename\n2. `<nativeDir>/<filename>` then `<execDir>/<filename>`, per filename\n\nAt load time, an extracted embedded candidate, or a staged Windows candidate when no embedded candidate exists, is prepended ahead of these de-duplicated candidates.\n\n## Embedded addon extraction lifecycle\n\n`embedded-addon.js` is generated by `scripts/embed-native.ts`. The reset stub exports `embeddedAddon = null`. A populated manifest has:\n\n- `platformTag`\n- `version`\n- `archive`: `{ format: \"tar.gz\", filename, filePath }`\n- `files[]` entries with `variant`, `filename`, and `size`\n\nExtraction (`maybeExtractEmbeddedAddon`) runs only when:\n\n1. compiled-binary mode is true,\n2. `embeddedAddon` is non-null,\n3. manifest `platformTag` equals the runtime platform tag,\n4. manifest `version` equals the package version,\n5. a variant-appropriate embedded file exists.\n\nVariant file selection:\n\n- Non-x64: prefer `default`, then first available file.\n- x64 + `modern`: prefer `modern`, fallback to `baseline`.\n- x64 + `baseline`: require `baseline`.\n\nMaterialization:\n\n1. Ensure `<versionedDir>` exists.\n2. Select `<versionedDir>/<selected filename>`.\n3. If the current cached file exists and its size matches manifest metadata, reuse it.\n4. Otherwise extract `embeddedAddon.archive.filePath` into `<versionedDir>` using the manifest `files[]` allowlist.\n5. Verify the selected target by size and return it as the first candidate.\n\nArchive, directory, or write failures are appended to the loader error list; probing continues through normal candidates.\n\n## Lifecycle and state transitions\n\n```text\nInit\n  -> Load package metadata and embedded-addon manifest\n  -> Compute platform/version/variant/filenames/candidate paths\n  -> (compiled + embedded manifest matches?)\n       yes -> extract archive to versionedDir when needed (record errors, continue)\n       no  -> skip extraction\n  -> (Windows non-compiled node_modules install and no embedded candidate?)\n       yes -> stage leaf/core addon to versionedDir (record errors, continue)\n       no  -> skip staging\n  -> For each runtime candidate in order:\n       require(candidate)\n       -> sentinel validation passes or is workspace-dev: return addon exports (READY)\n       -> failure: record error, continue\n  -> none loaded:\n       if unsupported platform tag -> throw Unsupported platform\n       else -> throw Failed to load (tried-path diagnostics + hints)\n```\n\n## Failure behavior and diagnostics\n\n### Unsupported platform\n\nIf all candidates fail and `platformTag` is not supported, the loader throws:\n\n- `Unsupported platform: <tag>`\n- supported platform list\n- issue-reporting guidance\n\n### No loadable candidate\n\nIf the platform is supported but no candidate can be loaded, the final error includes:\n\n- `Failed to load pi_natives native addon for <platformTag>` or `<platformTag> (<variant>)`\n- every attempted path with the corresponding `require(...)` or sentinel-validation error\n- mode-specific remediation hints\n\n### Compiled-binary startup failures\n\nCompiled mode diagnostics include:\n\n- expected versioned cache target paths (`<versionedDir>/<filename>`),\n- remediation to delete the versioned cache and rerun,\n- direct release download `curl` commands for each expected filename.\n- release sentinel mismatch details when a loadable `.node` belongs to another `@oh-my-pi/pi-natives` version.\n\n### Non-compiled startup failures\n\nNormal package/runtime diagnostics include:\n\n- reinstall hint (`bun install @oh-my-pi/pi-natives`),\n- local rebuild command (`bun --cwd=packages/natives run build`),\n- optional x64 variant build hint (`TARGET_VARIANT=baseline|modern bun --cwd=packages/natives run build`).\n",
 	"natives-architecture.md": "# Natives Architecture\n\n`@oh-my-pi/pi-natives` is a two-layer package around an ESM loader:\n\n1. **ESM loader/package entrypoint** resolves and loads the correct `.node` addon with `createRequire`, validates the release sentinel outside workspace-dev loads, and re-exports generated classes/functions plus enum runtime objects as explicit named ESM exports.\n2. **Rust N-API module layer** implements the exported functions/classes and emits the generated TypeScript declarations.\n\nThis document is the foundation for deeper module-level docs.\n\n## Implementation files\n\n- `packages/natives/native/index.js`\n- `packages/natives/native/index.d.ts`\n- `packages/natives/native/loader-state.js`\n- `packages/natives/native/embedded-addon.js`\n- `packages/natives/scripts/build-native.ts`\n- `packages/natives/scripts/embed-native.ts`\n- `packages/natives/scripts/gen-enums.ts`\n- `packages/natives/package.json`\n- `crates/pi-natives/src/lib.rs`\n\n## Package entrypoint and public surface\n\n`packages/natives/package.json` points at generated native artifacts:\n\n- `main`: `./native/index.js`\n- `types`: `./native/index.d.ts`\n- `exports[\".\"].types`: `./native/index.d.ts`\n- `exports[\".\"].import`: `./native/index.js`\n\nThere is no current `packages/natives/src` TypeScript wrapper layer. Consumers import functions/classes/enums directly from `@oh-my-pi/pi-natives`; the type contract is the generated `native/index.d.ts` plus the explicit named exports generated into `native/index.js` by `scripts/gen-enums.ts`.\n\nCurrent capability groups in the generated API include:\n\n- **Search/text/code primitives**: `grep`, `search`, `hasMatch`, `fuzzyFind`, `glob`, `astGrep`, `astEdit`, `blockRangeAt`, `summarizeCode`, text width/slicing/wrapping/sanitization, syntax highlighting, token counting.\n- **Execution/process/terminal primitives**: `executeShell`, `Shell`, `PtySession`, `Process`, key parsing, bash fixups.\n- **System/media/isolation/conversion primitives**: clipboard, SIXEL encoding, HTML-to-Markdown, macOS appearance/power helpers, work profiling, workspace scanning, isolation backend helpers (`iso*`).\n\n## Loader layer\n\n`packages/natives/native/index.js` owns runtime addon selection and optional embedded extraction.\n\n### Candidate resolution model\n\n- Platform tag is `${process.platform}-${process.arch}`.\n- Supported tags are currently:\n  - `linux-x64`\n  - `linux-arm64`\n  - `darwin-x64`\n  - `darwin-arm64`\n  - `win32-x64`\n- x64 can use CPU variants:\n  - `modern` (AVX2-capable)\n  - `baseline` (fallback)\n- Non-x64 uses the default filename without a variant suffix.\n\nFilename strategy:\n\n- Default: `pi_natives.<platform>-<arch>.node`\n- x64 variant: `pi_natives.<platform>-<arch>-modern.node` or `...-baseline.node`\n- x64 runtime fallback includes the unsuffixed default filename after variant candidates.\n\n### Platform-specific variant detection\n\nFor x64, variant selection uses:\n\n- Linux: `/proc/cpuinfo`\n- macOS: `sysctl -n machdep.cpu.leaf7_features`, then `machdep.cpu.features`\n- Windows: PowerShell check for `System.Runtime.Intrinsics.X86.Avx2`\n\n`PI_NATIVE_VARIANT` can force `modern` or `baseline`; invalid values are ignored.\n\n### Binary distribution and extraction model\n\nThe published `@oh-my-pi/pi-natives` package ships **only** the loader layer in `native/`: the ESM loader (`index.js`), generated declarations (`index.d.ts`), the `loader-state.js`/`.d.ts` helpers, and the embedded-addon manifest stub (`embedded-addon.js`). It carries no `.node` binaries.\n\nEach platform's prebuilt `.node` is published as a separate optional-dependency leaf package — `@oh-my-pi/pi-natives-<platform>-<arch>`, one per supported tag — which the core lists in `optionalDependencies` at the lockstep version during publish. npm/bun install only the leaf whose `os`/`cpu` match the host. The working-tree package keeps built `.node` files under `native/` for local dev; the release-publish rewrite (`prepareNativeCorePackage` in `scripts/ci-release-publish.ts`) strips them from the core tarball, and the leaves are generated by `packages/natives/scripts/gen-npm-packages.ts` (`LEAF_TARGETS`). Adding a build target therefore requires a matching `LEAF_TARGETS` entry, or the binary never reaches npm users.\n\nFor compiled binaries, loader behavior is:\n\n1. Check versioned user cache path: `<getNativesDir()>/<packageVersion>/...`.\n2. Check legacy compiled-binary location:\n   - Windows: `%LOCALAPPDATA%/omp` (fallback `%USERPROFILE%/AppData/Local/omp`)\n   - non-Windows: `~/.local/bin`\n3. Fall back to packaged `native/` and executable directory candidates.\n\n`getNativesDir()` uses `$XDG_DATA_HOME/omp/natives` when `$XDG_DATA_HOME/omp` exists; otherwise it uses `~/.omp/natives`.\n\nIf a populated embedded addon manifest is present, it is also treated as a compiled-binary signal. Current embedded manifests point at a gzip-compressed tar archive (`embedded-addons.<tag>.tar.gz`) that contains one or more matching `.node` files. The loader extracts the archive into the versioned cache directory, validates the selected file by size, and prepends that cache path before normal candidate probing.\n\nFor npm/bun installs (non-compiled), `loader-state.js` resolves the platform leaf directory via `require.resolve(\"@oh-my-pi/pi-natives-<tag>/package.json\")` and probes its `.node` **before** the core package's `native/` directory and the executable directory. The optional-dependency binary is therefore preferred over any `.node` left in the core (e.g. a stale local-dev build). On Windows `node_modules` installs, the loader first stages the selected leaf/core addon into `<getNativesDir()>/<packageVersion>/...` and prepends that staged path so running processes do not lock the `node_modules` copy during global updates.\n\n### Failure modes\n\nLoader failures are explicit:\n\n- **Unsupported platform tag**: after failed probing, throws with supported platform list.\n- **No loadable candidate**: throws with all attempted paths and remediation hints.\n- **Embedded/staging errors**: directory/write/archive/staging failures are recorded and included in final load diagnostics if no candidate loads.\n- **Release mismatch**: outside workspace-dev loads, a candidate that loads but lacks the version sentinel export for `package.json#version` is rejected with a reinstall hint.\n\n## Rust N-API module layer\n\n`crates/pi-natives/src/lib.rs` declares exported module ownership:\n\n- `appearance`\n- `ast`\n- `block`\n- `clipboard`\n- `crash_handler`\n- `fd`\n- `fs_cache`\n- `glob`\n- `glob_util`\n- `grep`\n- `highlight`\n- `html`\n- `iso`\n- `keys`\n- `language` (re-exported from `pi_ast`)\n- `power`\n- `prof`\n- `ps`\n- `pty`\n- `shell`\n- `sixel`\n- `snapcompact`\n- `summary`\n- `task`\n- `text`\n- `tokens`\n- `utils` (crate-private helpers)\n- `workspace`\n\nN-API exports are generated from Rust `#[napi]` functions/classes/objects/enums. Snake_case Rust names are exposed as camelCase JavaScript names unless explicitly configured by napi-rs.\n\n## Ownership boundaries\n\n- **Loader/package ownership (`packages/natives/native`, `packages/natives/scripts`)**\n  - runtime binary selection\n  - CPU variant selection and override handling\n  - compiled-binary embedded archive extraction\n  - Windows `node_modules` addon staging\n  - generated TypeScript declarations and explicit ESM export/enum patching\n- **Rust ownership (`crates/pi-natives/src`)**\n  - algorithmic and system-level implementation\n  - platform-native behavior and performance-sensitive logic\n  - N-API symbol implementation consumed directly by package callers\n- **Consumer ownership (`packages/coding-agent`, `packages/tui`)**\n  - user-facing policy and fallbacks that are not built into the native API\n  - higher-level rendering, artifact, shell-session, and command behavior\n\n## Runtime flow (high level)\n\n1. Consumer imports from `@oh-my-pi/pi-natives`.\n2. `native/index.js` computes platform/arch/variant and candidate paths.\n3. Optional embedded archive extraction or Windows `node_modules` staging can prepend a versioned-cache candidate.\n4. Each candidate is `require(...)`d; install/compiled loads must expose the package-version sentinel.\n5. The loaded addon object is bound to explicit named ESM exports, including generated enum objects.\n6. Caller invokes generated N-API functions/classes directly.\n\n## Glossary\n\n- **Native addon**: A `.node` binary loaded via Node-API (N-API).\n- **Platform tag**: Runtime tuple `platform-arch` (for example `darwin-arm64`).\n- **Platform leaf package**: Per-platform npm package `@oh-my-pi/pi-natives-<tag>` that carries one platform's prebuilt `.node`. The core depends on every leaf via `optionalDependencies`; the package manager installs only the host-matching one (`os`/`cpu`).\n- **Variant**: x64 CPU-specific build flavor (`modern` AVX2, `baseline` fallback).\n- **Generated binding declaration**: `native/index.d.ts` emitted by napi-rs during `build-native.ts`.\n- **Version sentinel**: Rust export named from the package version (for example `__piNativesV15_7_2`) that lets the loader reject a `.node` from a different release.\n- **Compiled binary mode**: Runtime mode where the CLI is bundled and native addons are resolved from embedded/cache paths before package-local paths.\n- **Embedded addon**: Build artifact metadata and archive reference generated into `native/embedded-addon.js` so compiled binaries can extract matching `.node` payloads.\n",
 	"natives-binding-contract.md": "# Natives Binding Contract (JavaScript/TypeScript Side)\n\nThis document defines the JS/TS contract between `@oh-my-pi/pi-natives` callers and the loaded N-API addon.\n\nCurrent package shape is direct-to-native: there is no `packages/natives/src/<module>` TypeScript wrapper layer. The public API is the generated `packages/natives/native/index.d.ts` declaration file, the ESM loader/export wrapper in `packages/natives/native/index.js`, and the Rust `#[napi]` exports in `crates/pi-natives/src`.\n\n## Implementation files\n\n- `packages/natives/native/index.js`\n- `packages/natives/native/index.d.ts`\n- `packages/natives/native/loader-state.js`\n- `packages/natives/scripts/build-native.ts`\n- `packages/natives/scripts/gen-enums.ts`\n- `packages/natives/package.json`\n- `crates/pi-natives/src/lib.rs`\n- Rust modules under `crates/pi-natives/src/*.rs`\n\n## Contract model\n\nThe contract has three parts:\n\n1. **ESM runtime loader/export wrapper** (`native/index.js`)\n   - calls `loadNative()` from `loader-state.js`, which `require(...)`s the `.node` addon;\n   - binds generated classes/functions as explicit named ESM exports;\n   - emits enum runtime objects generated by `scripts/gen-enums.ts`.\n2. **Generated TypeScript declarations** (`native/index.d.ts`)\n   - generated by napi-rs during `scripts/build-native.ts`;\n   - declares exported functions, classes, object interfaces, and native enums;\n   - is the package `types` entry.\n3. **Rust N-API exports** (`crates/pi-natives/src`)\n   - `#[napi]` functions/classes/objects/enums are the source of generated declarations and runtime symbols;\n   - snake_case Rust names become camelCase JavaScript names by napi-rs convention.\n\nThere is no current `NativeBindings` declaration-merging lifecycle and no full required-export list in the loader. Install/compiled loads do validate the package-version sentinel export; workspace-dev loads skip that check.\n\n## Public export surface organization\n\n`packages/natives/package.json` exposes the package root only:\n\n```json\n{\n  \"main\": \"./native/index.js\",\n  \"types\": \"./native/index.d.ts\",\n  \"exports\": {\n    \".\": {\n      \"types\": \"./native/index.d.ts\",\n      \"import\": \"./native/index.js\"\n    }\n  }\n}\n```\n\nConsumers in `packages/coding-agent` and `packages/tui` import directly from `@oh-my-pi/pi-natives`.\n\n## JS API ↔ native export mapping (representative)\n\n| Category          | Public JS API                                                                                             | Rust source                                      | Return style               |\n| ----------------- | --------------------------------------------------------------------------------------------------------- | ------------------------------------------------ | -------------------------- |\n| Grep              | `grep(options, onMatch?)`                                                                                 | `grep.rs`                                        | `Promise<GrepResult>`      |\n| Grep              | `search(content, options)`                                                                                | `grep.rs`                                        | `SearchResult`             |\n| Grep              | `hasMatch(content, pattern, ignoreCase?, multiline?)`                                                     | `grep.rs`                                        | `boolean`                  |\n| Fuzzy path search | `fuzzyFind(options)`                                                                                      | `fd.rs`                                          | `Promise<FuzzyFindResult>` |\n| Glob/workspace    | `glob(options, onMatch?)`, `listWorkspace(options)`                                                       | `glob.rs`, `workspace.rs`                        | `Promise<...>`             |\n| Glob cache        | `invalidateFsScanCache(path?)`                                                                            | `fs_cache.rs`                                    | `void`                     |\n| AST/block/summary | `astGrep(options)`, `astEdit(options)`, `blockRangeAt(options)`, `summarizeCode(options)`                 | `ast.rs`, `block.rs`, `summary.rs`               | mixed                      |\n| Shell             | `executeShell(options, onChunk?)`                                                                         | `shell.rs`                                       | `Promise<ShellRunResult>`  |\n| Shell             | `new Shell(options?)`, `shell.run(...)`, `shell.abort()`                                                  | `shell.rs`                                       | class / promises           |\n| PTY               | `new PtySession()`, `start/write/resize/kill`                                                             | `pty.rs`                                         | class / promises           |\n| Process           | `Process.fromPid/fromPath`, `status/children/killTree/terminate/waitForExit`                              | `ps.rs`                                          | class / mixed              |\n| Keys              | `parseKey`, `matchesKey`, Kitty/legacy helpers                                                            | `keys.rs`                                        | sync                       |\n| Text              | `wrapTextWithAnsi`, `truncateToWidth`, `sliceWithWidth`, `extractSegments`, `visibleWidth`                | `text.rs`                                        | sync                       |\n| Highlight         | `highlightCode`, `supportsLanguage`, `getSupportedLanguages`                                              | `highlight.rs`                                   | sync                       |\n| HTML              | `htmlToMarkdown(html, options?)`                                                                          | `html.rs`                                        | `Promise<string>`          |\n| SIXEL             | `encodeSixel`                                                                                             | `sixel.rs`                                       | sync                       |\n| Clipboard         | `copyToClipboard`, `readImageFromClipboard`                                                               | `clipboard.rs`                                   | sync / promise             |\n| Tokens            | `countTokens(input, encoding?)`                                                                           | `tokens.rs`                                      | sync                       |\n| System/isolation  | `detectMacOSAppearance`, `MacAppearanceObserver`, `MacOSPowerAssertion`, `getWorkProfile`, `iso*` helpers | `appearance.rs`, `power.rs`, `prof.rs`, `iso.rs` | mixed                      |\n\n## Sync vs async contract differences\n\nThe contract preserves Rust/N-API call style:\n\n- **Promise-returning exports** for worker-thread or async runtime work (`grep`, `glob`, `fuzzyFind`, `astGrep`, `astEdit`, `htmlToMarkdown`, shell/PTY runs, `isoStart`/`isoStop`/`isoDiff`, clipboard image read, workspace scan).\n- **Synchronous exports** for deterministic in-memory transforms/parsers or direct system calls (`search`, `hasMatch`, highlighting, text utilities, token counting, process construction/status, `copyToClipboard`, `encodeSixel`, isolation probe/resolve helpers).\n- **Constructor exports** for stateful runtime objects (`Shell`, `PtySession`, `Process`, macOS observer/power handles).\n\nChanging sync ↔ async for an existing export is a breaking public API change because consumers call these exports directly.\n\n## Object and enum typing patterns\n\n### Object patterns\n\n`#[napi(object)]` Rust structs become TS interfaces, for example:\n\n- `GrepResult`, `SearchResult`, `GlobResult`, `FuzzyFindResult`\n- `ShellRunResult`, `PtyRunResult`, `MinimizerResult`\n- `AstFindResult`, `AstReplaceResult`, `BlockRange`, `SummaryResult`\n- `System`/media/isolation payloads such as `ClipboardImage`, `WorkProfile`, `ParsedKittyResult`, `IsoResolveResult`\n\nRuntime shape correctness is owned by napi-rs and the Rust implementation.\n\n### Enum patterns\n\nNative enums are represented in generated declarations and also emitted as runtime objects by `scripts/gen-enums.ts`, because napi-rs string enums are TS-only without explicit JS exports. Current enum objects include:\n\n- `AstMatchStrictness`\n- `Ellipsis`\n- `Encoding`\n- `FileType`\n- `GrepOutputMode`\n- `IsoBackendKind`\n- `IsoChangeKind`\n- `KeyEventType`\n- `MacOSAppearance`\n- `ProcessStatus`\n\n## Error behavior and caveats\n\n- Addon load failure or unsupported platform throws during package import from `native/index.js`.\n- The loader rejects install/compiled candidates that lack the package-version sentinel export. It does not verify the full export set after `require(...)`; stale same-version or incomplete binaries surface as native load errors or missing members at use sites.\n- N-API conversion validates basic argument conversion, but TS optional fields do not guarantee semantic validity for untyped callers.\n- Numeric enum declarations do not prevent out-of-range numeric values from untyped callers unless the Rust function rejects them during conversion.\n- Callback exports use napi-rs `ThreadsafeFunction` shape: `(error: Error | null, value) => void`. Native code generally emits successful values; hard failures reject/throw through the owning call.\n\n## Maintainer checklist for binding changes\n\nWhen adding/changing an export, update all of:\n\n1. Rust `#[napi]` implementation in the owning `crates/pi-natives/src/<module>.rs`.\n2. `crates/pi-natives/src/lib.rs` if a new module is added.\n3. Any consumer imports/callsites in `packages/coding-agent` or `packages/tui`.\n4. Build output by running the natives build so `native/index.d.ts` and `native/index.js` stay in sync.\n5. `scripts/gen-enums.ts` if enum runtime export patching needs to change.\n\nDo not add a parallel TS wrapper convention unless the package design intentionally moves back to wrappers; current consumers depend on the direct generated API.\n",
@@ -49,11 +49,11 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"porting-from-pi-mono.md": "# Porting From pi-mono: A Practical Merge Guide\n\nThis guide is a repeatable checklist for porting changes from pi-mono into this repo.\nUse it for any merge: single file, feature branch, or full release sync.\n\n## Last Sync Point (historical upstream marker)\n\n**Commit:** `b21b42d032919de2f2e6920a76fa9a37c3920c0a`\n**Date:** 2026-03-22\n\nUpdate this section after each sync; do not reuse the previous range. This commit is an upstream pi-mono marker and may not exist in this repo's local object database.\n\nWhen starting a new sync, generate patches from this commit forward in a pi-mono checkout or remote that contains the commit:\n\n```bash\ngit format-patch b21b42d032919de2f2e6920a76fa9a37c3920c0a..HEAD --stdout > changes.patch\n```\n\n## 0) Define the scope\n\n- Identify the upstream reference (commit, tag, or PR).\n- List the packages or folders you plan to touch.\n- Decide which features are in-scope and which are intentionally skipped.\n\n## 1) Bring code over safely\n\n- Prefer a clean, focused diff rather than a wholesale copy.\n- Avoid copying built artifacts or generated files.\n- If upstream added new files, add them explicitly and review contents.\n\n## 2) Match import extension conventions\n\nMost runtime TypeScript sources omit `.js` in internal imports, but several current entrypoints and tool modules keep `.js` for ESM/runtime compatibility. Follow the surrounding file and package export style; do not blanket-strip or blanket-add extensions.\n\n- In `packages/coding-agent` runtime sources, prefer extensionless internal imports when the surrounding module does, but preserve existing `.js` imports in files that already require them.\n- In `packages/tui/test` and `packages/natives/bench`, keep `.js` where surrounding files already use it.\n- Keep real file extensions when required by tooling or import assertions (e.g., `.json`, `.css`, `.md` text embeds).\n- Example: `import { x } from \"./foo.js\";` → `import { x } from \"./foo\";` only when that package/file convention is extensionless.\n\n## 3) Replace import scopes\n\nUpstream uses different package scopes. Replace them consistently.\n\n- Replace old scopes with the local scope used here.\n- Examples (adjust to match the actual packages you are porting):\n  - `@mariozechner/pi-coding-agent` → `@oh-my-pi/pi-coding-agent`\n  - `@mariozechner/pi-agent-core` → `@oh-my-pi/pi-agent-core`\n  - `@mariozechner/pi-tui` → `@oh-my-pi/pi-tui`\n  - `@mariozechner/pi-ai` → `@oh-my-pi/pi-ai`\n  - `@mariozechner/pi-utils` → `@oh-my-pi/pi-utils`\n- Some upstream packages publish under the `@earendil-works/*` scope instead of `@mariozechner/*`. Map it the same way (`@earendil-works/pi-coding-agent` → `@oh-my-pi/pi-coding-agent`, and so on).\n- The bare `typebox` package is not an `@oh-my-pi/*` scope; do not rewrite it as one. See the Extensions divergence in section 15 for how tool-parameter schemas map.\n\n## 4) Use Bun APIs where they improve on Node\n\nWe run on Bun, but the current source intentionally mixes Bun APIs with small Node standard-library APIs. Replace Node APIs only when Bun provides a clearer, safer, or simpler implementation; do not mechanically rewrite every Node import.\n\n**Prefer replacing when porting new code:**\n\n- Process spawning: prefer Bun Shell `$` for simple commands; use `Bun.spawn`/`Bun.spawnSync` for streaming or process control. Keep existing `child_process` only where its exact semantics are needed.\n- HTTP clients: `node-fetch`, `axios` → native `fetch`\n- SQLite: `better-sqlite3` → `bun:sqlite`\n- Env loading: `dotenv` → Bun loads `.env` automatically\n- Runtime text/assets: prefer Bun imports such as `with { type: \"text\" }` or `Bun.file()` over copy steps or bundled fallback file reads.\n\n**DO NOT replace (these work fine in Bun):**\n\n- `os.homedir()` — do NOT replace with `Bun.env.HOME` or literal `\"~\"`\n- `os.tmpdir()` — do NOT replace with `Bun.env.TMPDIR || \"/tmp\"` or hardcoded paths\n- `fs.mkdtempSync()` — do NOT replace with manual path construction\n- `path.join()`, `path.resolve()`, etc. — these are fine\n\n**Import style:** Use the `node:` prefix for Node standard-library imports. Namespace imports are common, but named imports are acceptable where the surrounding code already uses them.\n\n**Additional Bun conventions:**\n\n- Prefer Bun Shell `$` for short, non-streaming commands; use `Bun.spawn` only when you need streaming I/O or process control.\n- Use `Bun.file()`/`Bun.write()` for simple files and `node:fs/promises` for directory-oriented operations. Existing synchronous `node:fs` calls are acceptable when the calling flow is intentionally synchronous.\n- Avoid `Bun.file().exists()` checks; use `isEnoent` handling in try/catch.\n- Prefer `Bun.sleep(ms)` over `setTimeout` wrappers.\n\n**Wrong:**\n\n```typescript\n// BROKEN: env vars may be undefined, \"~\" is not expanded\nconst home = Bun.env.HOME || \"~\";\nconst tmp = Bun.env.TMPDIR || \"/tmp\";\n```\n\n**Correct:**\n\n```typescript\nimport * as os from \"node:os\";\nimport * as fs from \"node:fs\";\nimport * as path from \"node:path\";\n\nconst configDir = path.join(os.homedir(), \".config\", \"myapp\");\nconst tempDir = fs.mkdtempSync(path.join(os.tmpdir(), \"myapp-\"));\n```\n\n## 5) Prefer Bun embeds (no copying)\n\nDo not add new runtime asset copy steps. Keep assets in repo and prefer Bun embeds/imports; preserve existing explicit generation workflows such as `packages/coding-agent/src/export/html/tool-views.generated.js` (built from collab-web sources via `bun run build-tool-views`).\n\n- If upstream copies assets into a dist folder, replace with Bun-friendly embeds.\n- Prompts are static `.md` files; use Bun text imports (`with { type: \"text\" }`) and Handlebars instead of inline prompt strings.\n- Use `import.meta.dir` + `Bun.file` to load adjacent non-text resources.\n- Keep assets in-repo and let the bundler include them.\n- Eliminate copy scripts unless the user explicitly requests them or the package already has an intentional generation step.\n- If upstream reads a bundled fallback file at runtime, replace filesystem reads with a Bun text embed import unless the current package already uses a generated asset pipeline.\n  - Example (Codex instructions fallback):\n    - `const FALLBACK_PROMPT_PATH = join(import.meta.dir, \"codex-instructions.md\");` -> removed\n    - `import FALLBACK_INSTRUCTIONS from \"./codex-instructions.md\" with { type: \"text\" };`\n    - Use `return FALLBACK_INSTRUCTIONS;` instead of `readFileSync(FALLBACK_PROMPT_PATH, \"utf8\")`\n\n## 6) Port `package.json` carefully\n\nTreat `package.json` as a contract. Merge intentionally.\n\n- Keep existing `name`, `version`, `type`, `exports`, and `bin` unless the port requires changes.\n- Replace npm/node scripts with Bun equivalents (e.g., `bun check`, `bun test`).\n- Ensure dependencies use the correct scope.\n- Do not downgrade dependencies to fix type errors; upgrade instead.\n- Validate workspace package links and `peerDependencies`.\n\n## 7) Align code style and tooling\n\n- Keep existing formatting conventions.\n- Do not introduce `any` unless required.\n- Avoid dynamic imports unless they are required for optional dependencies, startup cost, or runtime-only modules; prefer top-level imports otherwise.\n- Never build prompts in code; prompts are static `.md` files rendered with Handlebars.\n- In `packages/coding-agent`, use `logger` from `@oh-my-pi/pi-utils` for internal/runtime logging; CLI command files may use `console.*` for intentional user-facing output.\n- Use `Promise.withResolvers()` instead of `new Promise((resolve, reject) => ...)`.\n- Prefer ES `#` private fields for new encapsulated state. Constructor parameter properties already exist in current code and are acceptable; do not churn unrelated access modifiers while porting.\n- Prefer existing helpers and utilities over new ad-hoc code.\n  Preserve Bun-first infrastructure changes already made in this repo:\n  - Runtime is Bun (no Node entry points for the main CLI).\n  - Package manager is Bun (no npm lockfiles).\n  - Heavy Node APIs should not be introduced casually; current source still uses selected Node APIs (`node:crypto`, `node:readline`, synchronous `node:fs`, and `child_process`) where they fit provider, CLI, or process-control semantics.\n  - Lightweight Node APIs (`os.homedir`, `os.tmpdir`, `fs.mkdtempSync`, `path.*`) are kept.\n  - CLI shebangs use `bun` (not `node`, not `tsx`).\n  - TypeScript packages generally use source files directly; `@oh-my-pi/pi-natives` exports generated native bindings from `packages/natives/native`.\n  - CI workflows run Bun for install/check/test.\n\n## 8) Remove old compatibility layers\n\nUnless requested, remove upstream compatibility shims.\n\n- Delete old APIs that were replaced.\n- Update all call sites to the new API directly.\n- Do not keep `*_v2` or parallel versions.\n\n## 9) Update docs and references\n\n- Replace pi-mono repo links where appropriate.\n- Update examples to use Bun and correct package scopes.\n- Ensure README instructions still match the current repo behavior.\n\n## 10) Validate the port\n\nRun the standard checks after changes:\n\n- `bun check`\n\nIf the repo already has failing checks unrelated to your changes, call that out.\nTests use Bun's runner (not Vitest), but only run `bun test` when explicitly requested.\n\n## 11) Protect improved features (regression trap list)\n\nIf you already improved behavior locally, treat those as **non‑negotiable**. Before porting, write down\nthe improvements and add explicit checks so they don’t get lost in the merge.\n\n- **Freeze the expected behavior**: add a short “before/after” note for each improvement (inputs, outputs,\n  defaults, edge cases). This prevents silent rollback.\n- **Map old → new APIs**: if upstream renamed concepts (hooks → extensions, custom tools → tools, etc.),\n  ensure every old entry point still wires through. One missed flag or export equals lost functionality.\n- **Verify exports**: check `package.json` `exports`, public types, and barrel files. Upstream ports often\n  forget to re-export local additions.\n- **Cover non‑happy paths**: if you fixed error handling, timeouts, or fallback logic, add a test or at\n  least a manual checklist that exercises those paths.\n- **Check defaults and config merge order**: improvements often live in defaults. Confirm new defaults\n  didn’t revert (e.g., new config precedence, disabled features, tool lists).\n- **Audit env/shell behavior**: if you fixed execution or sandboxing, verify the new path still uses your\n  sanitized env and does not reintroduce alias/function overrides.\n- **Re-run targeted samples**: keep a minimal set of \"known good\" examples and run them after the port\n  (CLI flags, extension registration, tool execution).\n\n## 12) Detect and handle reworked code\n\nBefore porting a file, check if upstream significantly refactored it:\n\n```bash\n# Compare the file you're about to port against what you have locally\ngit diff HEAD upstream/main -- path/to/file.ts\n```\n\nIf the diff shows the file was **reworked** (not just patched):\n\n- New abstractions, renamed concepts, merged modules, changed data flow\n\nThen you must **read the new implementation thoroughly** before porting. Blind merging of reworked code loses functionality because:\n\nNote: interactive mode was recently split into controllers/utils/types. When backporting related changes, port updates into the individual files we created and ensure `interactive-mode.ts` wiring stays in sync.\n\n1. **Defaults change silently** - A new variable `defaultFoo = [a, b]` may replace an old `getAllFoo()` that returned `[a, b, c, d, e]`.\n\n2. **API options get dropped** - When systems merge (e.g., `hooks` + `customTools` → `extensions`), old options may not wire through to the new implementation.\n\n3. **Code paths go stale** - A renamed concept (e.g., `hookMessage` → `custom`) needs updates in every switch statement, type guard, and handler—not just the definition.\n\n4. **Context/capabilities shrink** - Old APIs may have exposed `{ logger, typebox, pi }` that new APIs forgot to include.\n\n### Semantic porting process\n\nWhen upstream reworked a module:\n\n1. **Read the old implementation** - Understand what it did, what options it accepted, what it exposed.\n\n2. **Read the new implementation** - Understand the new abstractions and how they map to old behavior.\n\n3. **Verify feature parity** - For each capability in the old code, confirm the new code preserves it or explicitly removes it.\n\n4. **Grep for stragglers** - Search for old names/concepts that may have been missed in switch statements, handlers, UI components.\n\n5. **Test the boundaries** - CLI flags, SDK options, event handlers, default values—these are where regressions hide.\n\n### Quick checks\n\n```bash\n# Find all uses of an old concept that may need updating\nrg \"oldConceptName\" --type ts\n\n# Compare default values between versions\ngit show upstream/main:path/to/file.ts | rg \"default|DEFAULT\"\n\n# Check if all enum/union values have handlers\nrg \"case \\\"\" path/to/file.ts\n```\n\n## 13) Quick audit checklist\n\nUse this as a final pass before you finish:\n\n- [ ] Import extensions follow the local package convention (no blanket `.js` stripping)\n- [ ] No newly introduced Node-only APIs unless they match an existing justified pattern\n- [ ] All package scopes updated\n- [ ] `package.json` scripts use Bun\n- [ ] Prompts are `.md` text imports (no inline prompt strings)\n- [ ] No internal/runtime `console.*` in coding-agent; CLI user-facing output is intentional\n- [ ] Assets load via Bun embed/import patterns, or through an existing intentional generation pipeline\n- [ ] Tests or checks run (or explicitly noted as blocked)\n- [ ] No functionality regressions (see sections 11-12)\n\n## 14) Commit message format\n\nWhen committing a backport, follow the repo format `<type>(scope): <past-tense description>` and keep the commit\nrange in the title.\n\n```\nfix(coding-agent): backported pi-mono changes (<from>..<to>)\n\npackages/<package>:\n- <type>: <description>\n- <type>: <description> (#<issue> by @<contributor>)\n\npackages/<other-package>:\n- <type>: <description>\n```\n\n**Example:**\n\n```\nfix(coding-agent): backported pi-mono changes (9f3eef65f..52532c7c0)\n\npackages/ai:\n- fix: handle \"sensitive\" stop reason from Anthropic API\n- fix: normalize tool call IDs with special characters for Responses API\n- fix: add overflow detection for Bedrock, MiniMax, Kimi providers\n- fix: 429 status is rate limiting, not context overflow\n\npackages/tui:\n- fix: refactored autocomplete state tracking\n- fix: file autocomplete should not trigger on empty text\n- fix: configurable autocomplete max visible items\n- fix: improved table column width calculation with word-aware wrapping\n\npackages/coding-agent:\n- fix: preserve external config.yml edits on save (#1046 by @nicobailonMD)\n- fix: resolve macOS NFD and curly quote variants in file paths\n```\n\n**Rules:**\n\n- Group changes by package\n- Use conventional commit types (`fix`, `feat`, `refactor`, `perf`, `docs`)\n- Include upstream issue/PR numbers and contributor attribution for external contributions\n- The commit range in the title helps track sync points\n\n## 15) Intentional Divergences\n\nOur fork has architectural decisions that differ from upstream. **Do not port these upstream patterns:**\n\n### UI Architecture\n\n| Upstream                                    | Our Fork                                                  | Reason                                                                |\n| ------------------------------------------- | --------------------------------------------------------- | --------------------------------------------------------------------- |\n| `FooterDataProvider` class                  | `StatusLineComponent`                                     | Simpler, integrated status line                                       |\n| `ctx.ui.setHeader()` / `ctx.ui.setFooter()` | No-op stubs in current extension contexts                 | Not currently wired to replace the TUI status/header UI               |\n| `ctx.ui.setEditorComponent()`               | No-op stubs in current extension contexts                 | Custom editor replacement is not currently wired                      |\n| `InteractiveModeOptions` options object     | Positional constructor args (options type still exported) | Keep constructor signature; update the type when upstream adds fields |\n\n### Component Naming\n\n| Upstream                     | Our Fork                |\n| ---------------------------- | ----------------------- |\n| `extension-input.ts`         | `hook-input.ts`         |\n| `extension-selector.ts`      | `hook-selector.ts`      |\n| `ExtensionInputComponent`    | `HookInputComponent`    |\n| `ExtensionSelectorComponent` | `HookSelectorComponent` |\n\n### API Naming\n\n| Upstream                                 | Our Fork                                 | Notes                                     |\n| ---------------------------------------- | ---------------------------------------- | ----------------------------------------- |\n| `sessionManager.appendSessionInfo(name)` | `sessionManager.setSessionName(name)`    | We use `sessionName` throughout           |\n| `sessionManager.getSessionName()`        | `sessionManager.getSessionName()`        | Same (we unified to match upstream's RPC) |\n| `agent.sessionName` / `setSessionName()` | `agent.sessionName` / `setSessionName()` | Same                                      |\n\n### File Consolidation\n\n| Upstream                                           | Our Fork                                                  | Reason                                        |\n| -------------------------------------------------- | --------------------------------------------------------- | --------------------------------------------- |\n| `clipboard.ts` + `clipboard-image.ts` (tool files) | `src/utils/clipboard.ts` backed by `@oh-my-pi/pi-natives` | Native implementation with a small TS wrapper |\n\n### Test Framework\n\n| Upstream                  | Our Fork                      |\n| ------------------------- | ----------------------------- |\n| `vitest` with `vi.mock()` | `bun:test` with `vi` from bun |\n| `node:test` assertions    | `expect()` matchers           |\n\n### Tool Architecture\n\n| Upstream                            | Our Fork                                                                                                      | Notes                                                     |\n| ----------------------------------- | ------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------- |\n| `createTool(cwd: string, options?)` | `createTools(session: ToolSession)` via `BUILTIN_TOOLS` registry                                              | Tool factories accept `ToolSession` and can return `null` |\n| Per-tool `*Operations` interfaces   | Only current per-tool override interfaces remain (for example `FindOperations`)                               | Used for SSH/remote overrides where present               |\n| Node.js `fs/promises` everywhere    | Bun file APIs for simple file writes/reads, `node:fs/promises` for dirs, selected sync `node:fs` where needed | Prefer Bun APIs when they simplify                        |\n\n### Auth Storage\n\n| Upstream                        | Our Fork                                    | Notes                                        |\n| ------------------------------- | ------------------------------------------- | -------------------------------------------- |\n| `proper-lockfile` + `auth.json` | `agent.db` (bun:sqlite)                     | Credentials stored exclusively in `agent.db` |\n| Single credential per provider  | Multi-credential with round-robin selection | Session affinity and backoff logic preserved |\n\n### Extensions\n\n| Upstream                                                         | Our Fork                                                                                          |\n| ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- |\n| `jiti` for TypeScript loading                                    | Native Bun `import()`                                                                              |\n| `pkg.pi` manifest field                                          | `pkg.omp` preferred; fallback to `pkg.pi` remains                                                  |\n| `StringEnum` from `pi-ai`                                        | `Type.Enum` from the `pi.typebox` shim (or author the schema with `pi.zod`); `pi-ai` no longer exports `StringEnum` |\n| `formatSize` from `pi-coding-agent`                              | `formatBytes` from `@oh-my-pi/pi-utils`                                                            |\n| `DefaultResourceLoader` / `DefaultPackageManager` / `SettingsManager` / `createEventBus` | Capability-based discovery (`loadCapability(...)`) plus the `Settings` singleton and `EventBus` |\n\n### Skip These Upstream Features\n\nWhen porting, **skip** these files/features entirely:\n\n- `footer-data-provider.ts` — we use StatusLineComponent\n- `clipboard-image.ts` — image clipboard support is exposed through `src/utils/clipboard.ts` backed by `@oh-my-pi/pi-natives`\n- GitHub workflow files — we have our own CI\n- `models.generated.ts` — auto-generated, regenerate locally (as models.json instead)\n\n### Features We Added (Preserve These)\n\nThese exist in our fork but not upstream. **Never overwrite:**\n\n- `StatusLineComponent` in interactive mode\n- Multi-credential auth with session affinity\n- Capability-based discovery system (`defineCapability`, `registerProvider`, `loadCapability`, `skillCapability`, etc.)\n- MCP/Exa/SSH integrations\n- LSP writethrough for format-on-save\n- Bash interception (`checkBashInterception`)\n- Fuzzy path suggestions in read tool\n",
 	"porting-to-natives.md": "# Porting to pi-natives (N-API) — Field Notes\n\nThis is a practical guide for moving hot paths into `crates/pi-natives` and wiring them through the generated native package entrypoint. It exists to avoid the same failures happening twice.\n\n## When to port\n\nPort when any of these are true:\n\n- The hot path runs in render loops, tight UI updates, or large batches.\n- JS allocations dominate (string churn, regex backtracking, large arrays).\n- You already have a JS baseline and can benchmark both versions side by side.\n- The work is CPU-bound or blocking I/O that can run on the libuv thread pool.\n- The work is async I/O that can run on Tokio's runtime (for example shell execution).\n\nAvoid ports that depend on JS-only state or dynamic imports. N-API exports should be data-in/data-out. Long-running work should go through `task::blocking` (CPU-bound/blocking I/O) or `task::future` (async I/O) with cancellation where the caller needs `timeoutMs` or `AbortSignal`.\n\n## Current package shape\n\n`@oh-my-pi/pi-natives` no longer has a `packages/natives/src/<module>` TypeScript wrapper layer. The package root points at generated native artifacts:\n\n- runtime entry/export wrapper: `packages/natives/native/index.js`\n- types entry: `packages/natives/native/index.d.ts`\n- loader helpers: `packages/natives/native/loader-state.js`\n- embedded manifest: `packages/natives/native/embedded-addon.js`\n\nConsumers import directly from `@oh-my-pi/pi-natives`. The generated declarations and explicit ESM exports are produced during `bun --cwd=packages/natives run build`.\n\n## Anatomy of a native export\n\n**Rust side:**\n\n- Implementation lives in `crates/pi-natives/src/<module>.rs`.\n- If you add a new module, register it in `crates/pi-natives/src/lib.rs`.\n- Export with `#[napi]`; snake_case exports are converted to camelCase automatically. Use explicit JS names only for true aliases/non-default names. Use `#[napi(object)]` for object-shaped structs.\n- For CPU-bound or blocking work, use `task::blocking(tag, cancel_token, work)`.\n- For async work that needs Tokio, use `task::future(env, tag, work)`.\n- Pass a `CancelToken` when the API exposes `timeoutMs` or `AbortSignal`, and call `heartbeat()` inside long loops.\n\n**Package/build side:**\n\n- `packages/natives/scripts/build-native.ts` runs napi-rs, installs the `.node` artifact, copies generated `index.d.ts`, and regenerates explicit ESM class/function exports plus enum runtime exports in the checked-in `native/index.js`.\n- `packages/natives/native/index.js` is the ESM entrypoint that calls the loader, exposes named exports, and rejects install/compiled `.node` files that do not expose the package-version sentinel.\n- `packages/natives/package.json` exposes only the package root (`@oh-my-pi/pi-natives`) as the import surface. At publish time the binaries are split out: the core ships the loader only (no `.node`), and each platform's `.node` is published as an optional-dependency leaf package `@oh-my-pi/pi-natives-<tag>` (`scripts/ci-release-publish.ts` + `packages/natives/scripts/gen-npm-packages.ts`). This is transparent to importers — you still `import` from `@oh-my-pi/pi-natives`.\n\n**Consumer side:**\n\n- Update direct imports/callsites in `packages/coding-agent` or `packages/tui` when the new export replaces a JS implementation.\n- Keep higher-level policy in consumers unless it belongs in the native primitive itself.\n\n## Porting checklist\n\n1. **Add the Rust implementation**\n\n- Put the core logic in a plain Rust function.\n- If it is a new module, add it to `crates/pi-natives/src/lib.rs`.\n- Expose it with `#[napi]` so the default snake_case -> camelCase mapping stays consistent.\n- Keep signatures owned and simple: `String`, `Vec<String>`, `Uint8Array`, `Either<JsString, Uint8Array>`, or `#[napi(object)]` structs.\n- For CPU-bound or blocking work, use `task::blocking`; for async work, use `task::future`.\n- If exposing cancellation, include `timeout_ms: Option<u32>` and `signal: Option<Unknown<'env>>` in options, create `CancelToken::new(...)`, and heartbeat in long loops.\n\n2. **Build generated bindings**\n\n- Run `bun --cwd=packages/natives run build`.\n- Confirm the generated `packages/natives/native/index.d.ts` includes the new export with the intended JS name/signature.\n- Confirm `packages/natives/native/index.js` has generated explicit ESM exports for the new class/function and enum objects when enum changes are involved.\n\n3. **Update consumers**\n\n- Import the new export directly from `@oh-my-pi/pi-natives`.\n- Replace only callsites where the native implementation is faster/equivalent and preserves behavior.\n- Remove obsolete JS implementation code in the same change when the native path becomes canonical.\n\n4. **Add benchmarks**\n\n- Put benchmarks next to the owning package (`packages/tui/bench`, `packages/natives/bench`, or `packages/coding-agent/bench`).\n- Include a JS baseline and native version in the same run.\n- Use `Bun.nanoseconds()` and a fixed iteration count.\n- Keep benchmark inputs realistic for the hot path.\n\n5. **Run focused verification**\n\n- Build the native package.\n- Run the benchmark.\n- Run the narrow tests or scenario covering the changed export/callsites.\n\n## Pain points and how to avoid them\n\n### 1) Stale platform/variant artifacts\n\nThe loader probes platform-tagged artifacts in deterministic order. For x64, selected variant candidates are tried before the unsuffixed default fallback:\n\n- `modern`: `pi_natives.<tag>-modern.node`, then `...-baseline.node`, then `pi_natives.<tag>.node`.\n- `baseline`: `pi_natives.<tag>-baseline.node`, then `pi_natives.<tag>.node`.\n\nNon-x64 uses `pi_natives.<tag>.node`.\n\nCompiled binaries also probe `<getNativesDir()>/<version>/...` and a legacy user-data directory before package/executable locations. Windows `node_modules` installs stage leaf/core addons into the same versioned directory before probing. If any earlier candidate is stale, a new export may appear missing unless the version sentinel rejects it first.\n\n**Fix:** remove stale candidate/cache files and rebuild.\n\n```bash\nrm packages/natives/native/pi_natives.<platform>-<arch>.node\nrm packages/natives/native/pi_natives.<platform>-<arch>-modern.node\nrm packages/natives/native/pi_natives.<platform>-<arch>-baseline.node\nbun --cwd=packages/natives run build\n```\n\nFor compiled binaries or Windows staging, delete the versioned addon cache shown in the loader error (normally under `~/.omp/natives/<version>` unless `$XDG_DATA_HOME/omp` is used).\n\n### 2) Generated types do not match loaded binary\n\nThis can happen when `native/index.d.ts` was regenerated but the `.node` file being loaded is stale, same-version incomplete, or from a different platform/variant. Different-version install/compiled binaries should be rejected by the version sentinel during loading.\n\nVerify the loaded export set from the actual candidate path reported by the loader:\n\n```bash\nbun -e 'import { createRequire } from \"node:module\"; const require = createRequire(import.meta.url); const mod = require(process.argv[2]); console.log(Object.keys(mod).sort())' -- /path/from/loader/error/pi_natives.<tag>[-variant].node\n```\n\nFix the build/candidate mismatch. Do not paper over it with optional consumer checks if the export is required.\n\n### 3) Rust signature mismatch\n\nKeep N-API signatures simple and owned. Avoid borrowed references like `&str` in public exports. If you need structured data, use `#[napi(object)]` structs. If you need callbacks, use napi-rs `ThreadsafeFunction` and keep callback error/value behavior explicit.\n\n### 4) Enum runtime exports and ESM named exports\n\nnapi-rs declarations alone are not enough for JS callers that import named symbols or use enum objects at runtime. `scripts/gen-enums.ts` reads `native/index.d.ts`, writes explicit `export const ... = nativeBindings...` entries for public classes/functions, and emits enum objects in `native/index.js`. If you add or change a native export, verify both `native/index.d.ts` and the generated export block in `native/index.js`.\n\n### 5) Benchmarking mistakes\n\n- Do not compare different inputs or allocations.\n- Keep JS and native using identical input arrays.\n- Run both in the same benchmark file to avoid skew.\n- Include enough iterations to smooth startup noise, but keep inputs realistic.\n\n## Benchmark template\n\n```ts\nconst ITERATIONS = 2000;\n\nfunction bench(name: string, fn: () => void): number {\n  const start = Bun.nanoseconds();\n  for (let i = 0; i < ITERATIONS; i++) fn();\n  const elapsed = (Bun.nanoseconds() - start) / 1e6;\n  console.log(\n    `${name}: ${elapsed.toFixed(2)}ms total (${(elapsed / ITERATIONS).toFixed(6)}ms/op)`,\n  );\n  return elapsed;\n}\n\nbench(\"feature/js\", () => {\n  jsImpl(sample);\n});\n\nbench(\"feature/native\", () => {\n  nativeImpl(sample);\n});\n```\n\n## Verification checklist\n\n- Generated `native/index.d.ts` includes the new export and intended TS signature.\n- `native/index.js` includes the generated named export; enum objects are present when the change adds/changes enums.\n- The loaded `.node` file's `Object.keys(require(candidate))` includes the new export and the package-version sentinel.\n- Bench numbers are recorded in the PR/notes.\n- Call sites are updated only if native is faster/equal and behavior-compatible.\n- Obsolete JS code is removed when the native implementation becomes canonical.\n\n## Rule of thumb\n\n- If native is slower, do not switch callsites. Keep or remove the export based on whether it has a near-term owner.\n- If native is faster and behavior-compatible, switch callsites and keep a benchmark to catch regressions.\n",
 	"provider-streaming-internals.md": "# Provider streaming internals\n\nThis document explains how token/tool streaming is normalized in `@oh-my-pi/pi-ai`, then propagated through `@oh-my-pi/pi-agent-core` and `coding-agent` session events.\n\n## End-to-end flow\n\n1. `streamSimple()` (`packages/ai/src/stream.ts`) maps generic options and dispatches to a provider stream function.\n2. Provider stream functions translate provider-native stream events into the unified `AssistantMessageEvent` sequence. Current built-ins include Anthropic, OpenAI Responses/Completions/Codex/Azure Responses, Google Gemini/Gemini CLI/Vertex, Bedrock Converse, Ollama, Cursor, pi-native gateway transport, plus GitLab Duo/Kimi/Synthetic/xAI-Grok-Responses wrappers and extension-registered custom APIs.\n3. Each provider pushes events into `AssistantMessageEventStream` (`packages/ai/src/utils/event-stream.ts`), which exposes:\n   - async iteration for incremental updates\n   - `result()` for final `AssistantMessage`\n4. `agentLoop` (`packages/agent/src/agent-loop.ts`) consumes those events, mutates in-flight assistant state, and emits `message_update` events carrying the raw `assistantMessageEvent`.\n5. `AgentSession` (`packages/coding-agent/src/session/agent-session.ts`) subscribes to agent events, persists messages, drives extension hooks, and applies session behaviors (retry, compaction, TTSR, streaming-edit abort checks).\n\n## Unified stream contract in `@oh-my-pi/pi-ai`\n\nAll providers emit the same shape (`AssistantMessageEvent` in `packages/ai/src/types.ts`):\n\n- `start`\n- content block lifecycle triplets:\n  - text: `text_start` → `text_delta`\\* → `text_end`\n  - thinking: `thinking_start` → `thinking_delta`\\* → `thinking_end`\n  - tool call: `toolcall_start` → `toolcall_delta`\\* → `toolcall_end`\n- terminal event:\n  - `done` with `reason: \"stop\" | \"length\" | \"toolUse\"`\n  - or `error` with `reason: \"aborted\" | \"error\"`\n\n`AssistantMessageEventStream` guarantees:\n\n- final result is resolved by terminal event (`done` or `error`)\n- events are delivered to consumers immediately, in push order (no batching or merging)\n\n## Delta throttling behavior\n\n`AssistantMessageEventStream` itself no longer throttles or merges delta events — every provider event is delivered as pushed. The per-delta cost control moved into tool-call argument parsing: providers accumulate partial JSON and re-parse it via `parseStreamingJsonThrottled()` (`packages/ai/src/utils/json-parse.ts`), which skips the re-parse until at least `STREAMING_JSON_PARSE_MIN_GROWTH` (256) new bytes have arrived, bounding mid-stream parse cost from quadratic to linear. The final `toolcall_end` parse is always unconditional and authoritative.\n\nThere is no provider backpressure: providers still produce at full speed, while the local stream queues.\n\n## Provider normalization details\n\n## Anthropic (`anthropic-messages`)\n\nSource: `packages/ai/src/providers/anthropic.ts`\n\nNormalization points:\n\n- `message_start` initializes usage (input/output/cache tokens)\n- `content_block_start` maps to text/thinking/toolcall starts\n- `content_block_delta` maps:\n  - `text_delta` → `text_delta`\n  - `thinking_delta` → `thinking_delta`\n  - `input_json_delta` → `toolcall_delta`\n  - `signature_delta` updates `thinkingSignature` only (no event)\n- `content_block_stop` emits corresponding `*_end`\n- `message_delta.stop_reason` maps via `mapStopReason()`\n\nTool-call argument streaming:\n\n- each tool block carries internal `partialJson`\n- every JSON delta appends to `partialJson`\n- `arguments` are reparsed on appended deltas via `parseStreamingJsonThrottled()` (re-parse only after ≥256 new bytes)\n- `toolcall_end` reparses once more, then strips `partialJson`\n\n## OpenAI Responses family (`openai-responses`, `openai-codex-responses`, `azure-openai-responses`)\n\nSources: `packages/ai/src/providers/openai-responses.ts`, `openai-codex-responses.ts`, and `azure-openai-responses.ts`\n\nNormalization points:\n\n- `response.output_item.added` starts reasoning/text/function-call/custom-tool blocks\n- reasoning summary events (`response.reasoning_summary_text.delta`) and raw reasoning events (`response.reasoning_text.delta`) become `thinking_delta`\n- output/refusal deltas become `text_delta`\n- `response.function_call_arguments.delta` and `response.custom_tool_call_input.delta` become `toolcall_delta`\n- `response.output_item.done` emits `thinking_end` / `text_end` / `toolcall_end`\n- `response.completed` maps status to stop reason and usage; `response.failed` / SDK `error` events throw into the wrapper's terminal `error` path\n\nTool-call argument streaming:\n\n- same `partialJson` accumulation pattern as Anthropic for function-call JSON arguments\n- custom tools stream raw string input and expose final arguments as `{ input: <raw> }`\n- providers that send only `response.function_call_arguments.done` still populate final args\n- tool call IDs are normalized as `\"<call_id>|<item_id>\"`\n\n## Google Generative AI (`google-generative-ai`)\n\nSource: `packages/ai/src/providers/google.ts` (thin request wrapper) and `google-shared.ts` (`streamGoogleGenAI`, shared chunk-to-block translation)\n\nNormalization points:\n\n- iterates `candidate.content.parts`\n- text parts are split into thinking vs text by `isThinkingPart(part)`\n- block transitions close previous block before starting a new one\n- `part.functionCall` is treated as a complete tool call (start/delta/end emitted immediately)\n- finish reason mapped by `mapStopReason()` from `google-shared.ts`\n\nTool-call argument streaming:\n\n- function call args arrive as structured object, not incremental JSON text\n- implementation emits one synthetic `toolcall_delta` containing `JSON.stringify(arguments)`\n- no partial JSON parser needed for Google in this path\n\n## Partial tool-call JSON accumulation and recovery\n\nShared behavior for Anthropic/OpenAI Responses uses `parseStreamingJson()` / `parseStreamingJsonThrottled()` (`packages/ai/src/utils/json-parse.ts`):\n\n1. try `JSON.parse`\n2. fallback to `repairJson()` + the `partial-json` parser for incomplete fragments\n3. if both fail, return `{}`\n\nImplications:\n\n- malformed or truncated argument deltas do not crash stream processing immediately\n- in-progress `arguments` may temporarily be `{}`\n- later valid deltas can recover structured arguments because parsing is retried as the buffer grows (throttled to ≥256-byte growth steps mid-stream)\n- final `toolcall_end` performs one more parse attempt before emission\n\n## Stop reasons vs transport/runtime errors\n\nProvider stop reasons are mapped to normalized `stopReason`:\n\n- Anthropic: `end_turn`→`stop`, `max_tokens`→`length`, `tool_use`→`toolUse`, safety/refusal cases→`error`\n- OpenAI Responses: `completed`→`stop`, `incomplete`→`length`, `failed/cancelled`→`error`\n- Google: `STOP`→`stop`, `MAX_TOKENS`→`length`, safety/prohibited/malformed-function-call classes→`error`\n\nError semantics are split in two stages:\n\n1. **Model completion semantics** (provider reported finish reason/status)\n2. **Transport/runtime failure** (network/client/parser/abort exceptions)\n\nIf provider stream throws or signals failure, each provider wrapper catches and emits terminal `error` event with:\n\n- `stopReason = \"aborted\"` when abort signal is set\n- otherwise `stopReason = \"error\"`\n- `errorMessage = formatErrorMessageWithRetryAfter(error)`\n\n## Malformed chunk / SSE parse failure behavior\n\nThe OpenAI Completions/Responses paths delegate chunk/SSE framing to the `openai` SDK stream. Anthropic uses the in-repo `AnthropicMessagesClient` (`packages/ai/src/providers/anthropic-client.ts`); the Google paths and the Codex SSE fallback read SSE via `readSseJson()` directly, and websocket Codex frames are normalized through the same event handler.\n\nObserved behavior in current implementation:\n\n- malformed SDK stream parsing surfaces as an exception or stream `error` event\n- malformed Codex SSE JSON/framing throws from the local SSE reader\n- provider wrapper converts failures into unified terminal `error` events\n- no provider-specific resume/retry inside the stream function itself, except Codex websocket-to-SSE transport fallback before replay-unsafe output is emitted\n- higher-level retries are handled in `AgentSession` auto-retry logic (message-level retry, not stream-chunk replay)\n\n## Cancellation boundaries\n\nCancellation is layered:\n\n- AI provider request: `options.signal` is passed into provider client stream call.\n- Provider wrapper: after stream loop, aborted signal forces error path (`\"Request was aborted\"`).\n- Agent loop: checks `signal.aborted` before handling each provider event and can synthesize an aborted assistant message from the latest partial.\n- Session/agent controls: `AgentSession.abort()` -> `agent.abort()` -> shared abort controller cancellation.\n\nTool execution cancellation is separate from model stream cancellation:\n\n- tool runners use `AbortSignal.any([agentSignal, steeringAbortSignal])`\n- steering interrupts can abort remaining tool execution while preserving already-produced tool results\n\n## Backpressure boundaries\n\nThere is no hard backpressure mechanism between provider SDK stream and downstream consumers:\n\n- `EventStream` uses in-memory queues with no max size\n- the throttled partial-JSON re-parse reduces per-delta CPU cost but does not slow provider intake\n- if consumers lag significantly, queued events can grow until completion\n\nCurrent design favors responsiveness and simple ordering over bounded-buffer flow control.\n\n## How stream events surface as agent/session events\n\n`agentLoop.streamAssistantResponse()` bridges `AssistantMessageEvent` to `AgentEvent`:\n\n- on `start`: pushes placeholder assistant message and emits `message_start`\n- on block events (`text_*`, `thinking_*`, `toolcall_*`): updates last assistant message, emits `message_update` with raw `assistantMessageEvent`\n- on terminal (`done`/`error`): resolves final message from `response.result()`, emits `message_end`\n\n`AgentSession` then consumes those events for session-level behaviors:\n\n- TTSR watches `message_update.assistantMessageEvent` for `text_delta`, `thinking_delta`, and `toolcall_delta`\n- streaming edit guard inspects `toolcall_delta`/`toolcall_end` on `edit` calls and can abort early\n- persistence writes finalized messages at `message_end`\n- auto-retry examines assistant `stopReason === \"error\"` plus `errorMessage` heuristics\n\n## Unified vs provider-specific responsibilities\n\nUnified (common contract):\n\n- event shape (`AssistantMessageEvent`)\n- final result extraction (`done`/`error`)\n- immediate in-order event delivery\n- agent/session event propagation model\n\nProvider-specific (not fully abstracted):\n\n- upstream event taxonomies and mapping logic\n- stop-reason translation tables\n- tool-call ID conventions\n- reasoning/thinking block semantics and signatures\n- usage token semantics and availability timing\n- message conversion constraints per API\n\n## Implementation files\n\n- [`../../ai/src/stream.ts`](../packages/ai/src/stream.ts) — provider dispatch, option mapping, API key/session plumbing, custom API dispatch, and provider-specific credential handling.\n- [`../../ai/src/utils/event-stream.ts`](../packages/ai/src/utils/event-stream.ts) — generic stream queue + final-result resolution.\n- [`../../ai/src/utils/json-parse.ts`](../packages/ai/src/utils/json-parse.ts) — partial JSON parsing for streamed tool arguments.\n- [`../../ai/src/providers/anthropic.ts`](../packages/ai/src/providers/anthropic.ts) — Anthropic event translation and tool JSON delta accumulation.\n- [`../../ai/src/providers/openai-responses.ts`](../packages/ai/src/providers/openai-responses.ts), [`openai-responses-shared.ts`](../packages/ai/src/providers/openai-responses-shared.ts), [`openai-codex-responses.ts`](../packages/ai/src/providers/openai-codex-responses.ts), [`azure-openai-responses.ts`](../packages/ai/src/providers/azure-openai-responses.ts) — Responses-family event translation and status mapping.\n- [`../../ai/src/providers/google.ts`](../packages/ai/src/providers/google.ts), [`google-gemini-cli.ts`](../packages/ai/src/providers/google-gemini-cli.ts), [`google-vertex.ts`](../packages/ai/src/providers/google-vertex.ts) — Gemini stream chunk-to-block translation variants.\n- [`../../ai/src/providers/google-shared.ts`](../packages/ai/src/providers/google-shared.ts) — Gemini finish-reason mapping and shared conversion rules.\n- [`../../ai/src/providers/amazon-bedrock.ts`](../packages/ai/src/providers/amazon-bedrock.ts), [`openai-completions.ts`](../packages/ai/src/providers/openai-completions.ts), [`ollama.ts`](../packages/ai/src/providers/ollama.ts), [`cursor.ts`](../packages/ai/src/providers/cursor.ts), [`pi-native-client.ts`](../packages/ai/src/providers/pi-native-client.ts) — additional built-in stream adapters using the same event contract.\n- [`../../agent/src/agent-loop.ts`](../packages/agent/src/agent-loop.ts) — provider stream consumption and `message_update` bridging.\n- [`../src/session/agent-session.ts`](../packages/coding-agent/src/session/agent-session.ts) — session-level handling of streaming updates, abort, retry, and persistence.\n",
-	"providers.md": "# Providers\n\nProviders are the model backends `omp` can route requests to: Anthropic, OpenAI, Google Gemini, Groq, OpenRouter, Mistral, xAI, local engines like Ollama, hosted gateways, custom `models.yml` providers, and providers registered by extensions.\n\nA **provider** is the account or backend namespace, such as `anthropic`, `openai`, `google`, or `ollama`. A **model** is a concrete model under that provider, selected as `provider/model-id`, such as `anthropic/claude-opus-4-6`. Disabling a provider removes every model under it from selection; if you only want to narrow individual models, use model settings instead.\n\nThis page covers how providers become available, how credentials are resolved, the provider/environment-variable map, local engines, disabling providers, and custom providers. For model selection and the full `models.yml` schema, see [Model and Provider Configuration](./models.md). For config-file locations and merge precedence, see [Settings](./settings.md). For credential storage and login flows in depth, see [Secrets and credentials](./secrets.md). For the complete environment-variable reference, see [Environment variables](./environment-variables.md). For local engine setup, see [Local models](./local-models.md). For context-file discovery providers, see [Context files](./context-files.md).\n\n## How `omp` decides a provider is available\n\nAt startup the model registry assembles its catalog from four sources, in order:\n\n1. The bundled model catalog (every built-in provider and its known models).\n2. Custom provider and model entries from `~/.omp/agent/models.yml`.\n3. Runtime-discovered models for providers that support discovery (local engines and discovery-enabled gateways).\n4. Providers and models registered by extensions.\n\nThe registry can hold a model even when it is not currently selectable. A model becomes **available** only when both conditions hold:\n\n1. its provider ID is **not** in the effective `disabledProviders` list; **and**\n2. the provider is either **keyless** (an implicit local provider, or a custom provider with `auth: none`) **or** has resolvable credentials.\n\n`disabledProviders` is checked *before* credentials. If a provider ID is disabled, no stored key, OAuth session, environment variable, `.env` entry, or `models.yml` `apiKey` will make it selectable — the provider's models are dropped from availability regardless of credentials. Removing the ID from the effective list restores them.\n\nKeyless local engines are a special case: `ollama`, `llama.cpp`, and `lm-studio` are treated as keyless when no key is configured, so their discovered models are selectable as soon as the engine answers — no login required. See [Built-in local engines](#built-in-local-engines).\n\n## Credentials and precedence\n\nWhen a provider needs an API key, `omp` resolves it in this order (first match wins):\n\n1. **Runtime override** — a key supplied for the current process, e.g. CLI `--api-key`. Never persisted.\n2. **`models.yml` config key** — an `apiKey` pinned on a custom provider, registered as a config-sourced bearer. This deliberately beats stored OAuth, so a key supplied for a custom `baseUrl`/gateway is honored instead of forwarding an upstream OAuth token the proxy would reject.\n3. **Stored API key** — an API-key credential saved in the auth store.\n4. **Stored OAuth credential** — refreshed when needed; multiple accounts are ranked/rotated automatically.\n5. **Provider environment variable** — including values loaded from `.env` files (see [the env-var table](#environment-variables-and-env-files)).\n6. **`models.yml` fallback resolver** — keys for custom providers not otherwise registered.\n\nStored credentials live in the auth store at `~/.omp/agent/agent.db` for local auth, or in the configured auth-broker snapshot when running in broker mode. (`PI_CODING_AGENT_DIR` relocates the `~/.omp/agent` base, and the auth store moves with it.)\n\n### OAuth vs API key, and provider-scoped logins\n\nLogins are **provider-scoped**: authenticating `anthropic` does not authenticate `openai`, and each provider tracks its own credentials. A disabled provider stays disabled even with valid stored auth.\n\nUse the interactive slash commands inside a session:\n\n- `/login` — opens the OAuth/key selector. `/login <provider>` jumps straight to one provider (e.g. `/login anthropic`); for an OAuth flow that needs a pasted callback, run `/login <redirect-url>` to complete it.\n- `/logout` — opens the provider selector to remove stored credentials.\n\nFor headless or remote setups backed by a shared auth broker, the CLI exposes `omp auth-broker login <provider>` / `omp auth-broker logout` (and `status`, `list`, `import`, `migrate`). See [Secrets and credentials](./secrets.md) for the broker model.\n\nWhen a model has no credentials, `omp` tells you to run `/login` or set the provider's environment variable.\n\n### Pinning a key in `models.yml`\n\nA custom provider's `apiKey` is resolved as **environment-variable-name-or-literal**: if the value names an existing environment variable, that variable's value is used; otherwise the string itself is the key. Prefixing the value with `!` runs it as a shell command and uses the trimmed stdout (see [Model and Provider Configuration](./models.md) for the full value syntax).\n\n```yaml\n# ~/.omp/agent/models.yml\nproviders:\n  my-gateway:\n    baseUrl: https://gateway.example.com/v1\n    api: openai-completions\n    apiKey: MY_GATEWAY_API_KEY # reads this env var if set, else literal text\n    models:\n      - id: claude-sonnet\n        name: Claude Sonnet via Gateway\n        contextWindow: 200000\n        maxTokens: 8192\n```\n\nIf `authHeader: true` is set on a custom provider, the resolved key is injected as an `Authorization: Bearer <key>` header on every request to that provider.\n\n## Environment variables and `.env` files\n\nEach provider has one or more environment variables that supply a key when no stored credential exists. The table below is the verified provider → variable map; the full catalog is large, so it is split into core and additional providers. OAuth-backed providers can also accept a token variable in addition to (or instead of) an API key.\n\n### Core providers\n\n| Provider ID | Environment variable(s) |\n|---|---|\n| `anthropic` | `ANTHROPIC_OAUTH_TOKEN`, then `ANTHROPIC_API_KEY` (Foundry mode prefers `ANTHROPIC_FOUNDRY_API_KEY` when `CLAUDE_CODE_USE_FOUNDRY=true`) |\n| `openai` | `OPENAI_API_KEY` |\n| `openai-codex` | `OPENAI_CODEX_OAUTH_TOKEN` |\n| `google` | `GEMINI_API_KEY` |\n| `google-vertex` | `GOOGLE_CLOUD_API_KEY`, or Application Default Credentials (`GOOGLE_APPLICATION_CREDENTIALS` + `GOOGLE_CLOUD_PROJECT` + `GOOGLE_CLOUD_LOCATION`) |\n| `groq` | `GROQ_API_KEY` |\n| `openrouter` | `OPENROUTER_API_KEY` |\n| `mistral` | `MISTRAL_API_KEY` |\n| `xai` | `XAI_API_KEY` |\n| `xai-oauth` | `XAI_OAUTH_TOKEN`, then `XAI_API_KEY` |\n| `github-copilot` | `COPILOT_GITHUB_TOKEN` |\n| `cursor` | `CURSOR_ACCESS_TOKEN` |\n| `azure-openai-responses` | `AZURE_OPENAI_API_KEY` |\n| `amazon-bedrock` | `AWS_PROFILE`, or `AWS_ACCESS_KEY_ID` + `AWS_SECRET_ACCESS_KEY`, or an ECS/IRSA credential chain |\n\n### Additional hosted providers\n\n| Provider ID | Environment variable(s) |\n|---|---|\n| `cerebras` | `CEREBRAS_API_KEY` |\n| `deepseek` | `DEEPSEEK_API_KEY` |\n| `fireworks` | `FIREWORKS_API_KEY` |\n| `together` | `TOGETHER_API_KEY` |\n| `nvidia` | `NVIDIA_API_KEY` |\n| `huggingface` | `HUGGINGFACE_HUB_TOKEN`, then `HF_TOKEN` |\n| `moonshot` | `MOONSHOT_API_KEY` |\n| `kimi-code` | `KIMI_API_KEY` |\n| `nanogpt` | `NANO_GPT_API_KEY` |\n| `venice` | `VENICE_API_KEY` |\n| `vercel-ai-gateway` | `AI_GATEWAY_API_KEY` (also `VERCEL_AI_GATEWAY_API_KEY` for catalog discovery) |\n| `cloudflare-ai-gateway` | `CLOUDFLARE_AI_GATEWAY_API_KEY` |\n| `litellm` | `LITELLM_API_KEY` |\n| `kilo` | `KILO_API_KEY` |\n| `zai` | `ZAI_API_KEY` |\n| `zenmux` | `ZENMUX_API_KEY` |\n| `zhipu-coding-plan` | `ZHIPU_API_KEY` |\n| `qianfan` | `QIANFAN_API_KEY` |\n| `qwen-portal` | `QWEN_OAUTH_TOKEN`, then `QWEN_PORTAL_API_KEY` |\n| `synthetic` | `SYNTHETIC_API_KEY` |\n| `minimax` | `MINIMAX_API_KEY` |\n| `alibaba-coding-plan` | `ALIBABA_CODING_PLAN_API_KEY` |\n| `aimlapi` | `AIMLAPI_API_KEY` |\n| `gitlab-duo` | `GITLAB_TOKEN` |\n| `opencode-zen`, `opencode-go` | `OPENCODE_API_KEY` |\n| `firepass` | `FIREPASS_API_KEY` |\n| `wafer-pass` | `WAFER_PASS_API_KEY` |\n| `wafer-serverless` | `WAFER_SERVERLESS_API_KEY` |\n| `xiaomi` | `XIAOMI_API_KEY` |\n| `ollama-cloud` | `OLLAMA_CLOUD_API_KEY` |\n| `ollama` | `OLLAMA_API_KEY` (optional; local discovery is keyless by default) |\n| `lm-studio` | `LM_STUDIO_API_KEY` (optional; keyless by default) |\n| `llama.cpp` | `LLAMA_CPP_API_KEY` (only when the server requires auth) |\n\nOAuth-backed providers such as `anthropic`, `github-copilot`, `cursor`, `ollama-cloud`, `qwen-portal`, `xai-oauth`, `wafer-pass`, `wafer-serverless`, `google-gemini-cli`, and `google-antigravity` are normally reached through `/login` rather than an environment variable. See [Environment variables](./environment-variables.md) for search-tool and configuration variables not listed here.\n\n### `.env` discovery and precedence\n\n`omp` eagerly loads `.env` files into the process environment before any provider lookup. It reads four files and, for each variable, the **first** source that defines it wins. Effective precedence, high to low:\n\n1. The process environment inherited by `omp` (already-set variables always win).\n2. `<cwd>/.env`\n3. `~/.omp/agent/.env`\n4. `~/.omp/.env`\n5. `~/.env`\n\nA variable already present in the process environment is never overwritten by a `.env` file. Among the files, a value set in `<cwd>/.env` wins over `~/.omp/agent/.env`, which wins over `~/.omp/.env`, which wins over `~/.env`. So a shell-exported `OPENAI_API_KEY` beats every `.env` file, and a project's `<cwd>/.env` beats your home `~/.env`.\n\nProject-local `.env` is the simplest way to make one repository use a project-specific gateway, key, or local endpoint:\n\n```dotenv\n# <project>/.env\nOPENROUTER_API_KEY=sk-or-...\nOLLAMA_BASE_URL=http://127.0.0.1:11434\n```\n\n`.env` parsing is intentionally minimal:\n\n- blank lines and lines starting with `#` are ignored;\n- keys must match `[A-Za-z_][A-Za-z0-9_]*` (shell-identifier shape) — other names are dropped;\n- values may be wrapped in single or double quotes, which are stripped;\n- values containing a NUL byte are dropped;\n- an `OMP_`-prefixed key is also mirrored to the matching `PI_`-prefixed name.\n\n## Built-in local engines\n\nThree local engines are discovered automatically without needing a `models.yml` entry. Each uses a base URL that can be overridden by an environment variable:\n\n| Provider ID | Base URL (env override → default) | Notes |\n|---|---|---|\n| `ollama` | `OLLAMA_BASE_URL`, then `OLLAMA_HOST` (normalized), else `http://127.0.0.1:11434` | Keyless by default. |\n| `llama.cpp` | `LLAMA_CPP_BASE_URL`, else `http://127.0.0.1:8080` | Keyless unless a key is stored for `llama.cpp`. |\n| `lm-studio` | `LM_STUDIO_BASE_URL`, else `http://127.0.0.1:1234/v1` | Keyless by default. |\n\nThese implicit engines are **skipped** when:\n\n- a provider with the same ID is already configured in `models.yml` (your explicit config wins); or\n- the provider ID appears in the effective `disabledProviders` list.\n\nFor installing and running these engines, see [Local models](./local-models.md).\n\n## Disabling model providers\n\nUse the `disabledProviders` setting to remove a provider's models from selection:\n\n```yaml\n# ~/.omp/agent/config.yml or <project>/.omp/config.yml\ndisabledProviders:\n  - anthropic\n  - openai\n  - google\n  - groq\n```\n\nProvider IDs are matched exactly. Disable `google` to hide the Google Gemini API provider; the OAuth-backed Google providers `google-gemini-cli` and `google-antigravity` are separate IDs and must be disabled individually. Disable `ollama`, `llama.cpp`, or `lm-studio` to stop local discovery for that engine.\n\n`disabledProviders` applies uniformly to:\n\n- bundled catalog providers;\n- custom `models.yml` providers;\n- runtime-discovered provider models;\n- extension-registered providers;\n- implicit local engines.\n\nDisabling a provider does not delete its stored credentials — re-enable it by removing its ID from the effective list.\n\n## Project-specific provider control\n\nProject settings live in `<project>/.omp/config.yml`. Use them when one repository must allow or hide a different provider set than your global default:\n\n```yaml\n# <project>/.omp/config.yml\ndisabledProviders:\n  - openai\n  - openrouter\n```\n\nSettings arrays are **replaced** wholesale by the higher-precedence layer, not merged or appended. If the global file disables three providers and the project file disables one, the project sees only the project list:\n\n```yaml\n# ~/.omp/agent/config.yml\ndisabledProviders:\n  - anthropic\n  - openai\n  - google\n\n# <project>/.omp/config.yml\ndisabledProviders:\n  - groq\n```\n\nEffective result inside the project:\n\n```json\n[\"groq\"]\n```\n\nThe project array re-enables `anthropic`, `openai`, and `google` for sessions launched from that project. If you want a project to *add* to the global set, repeat the global IDs in the project file. See [Settings](./settings.md) for the full precedence chain, including `--config` overlays and runtime overrides.\n\n## Path-scoped `disabledProviders`\n\n`disabledProviders` can mix plain string entries (apply everywhere) with path-scoped entries (apply only when the current working directory matches a configured path):\n\n```yaml\ndisabledProviders:\n  - ollama\n  - path: ~/projects/sensitive\n    providers:\n      - anthropic\n      - openai\n  - paths:\n      - ~/work/client-a\n      - ~/work/client-b\n    values:\n      - openrouter\n```\n\n- Bare string entries always apply.\n- A scoped entry applies when the current working directory **is** the configured path or sits **under** it. `~` expands to the home directory.\n- Accepted path keys: `path`, `paths`, `pathPrefix`, `pathPrefixes`.\n- Accepted value keys: `providers`, `values`, `items`.\n\nFor the example above:\n\n- `ollama` is disabled everywhere.\n- `anthropic` and `openai` are additionally disabled under `~/projects/sensitive`.\n- `openrouter` is additionally disabled under `~/work/client-a` and `~/work/client-b`.\n\nPath scopes are resolved **after** the settings merge. Because a higher-precedence layer replaces the whole array, a project-level `disabledProviders` array drops any scoped entries that only existed in the global array. `enabledModels` is the only other setting that supports the same path-scoped form. See [Settings](./settings.md) for details.\n\n## Provider IDs vs discovery provider IDs\n\n`disabledProviders` uses a **single shared ID namespace** that gates two different subsystems:\n\n- **Model providers** — the backends on this page (`anthropic`, `openai`, `ollama`, a custom `models.yml` ID, …). Disabling one removes its models from selection.\n- **Discovery providers** — sources of context files, MCP servers, commands, skills, hooks, tools, prompts, and settings. Disabling one stops that source from contributing capability items.\n\n| Entry type | Examples | Effect |\n|---|---|---|\n| Model provider ID | `anthropic`, `openai`, `google`, `groq`, `openrouter`, `ollama`, `my-gateway` | Removes that provider's models from availability. |\n| Discovery provider ID | `native`, `claude`, `codex`, `gemini`, `agents`, `github` | Stops that discovery source from contributing capability items. |\n\nWatch the related names. The Google Gemini **API** models use the model provider ID `google`; `gemini` is a **discovery** provider ID (the source that reads `GEMINI.md`), not the Google model provider. Use discovery IDs only when you intend to disable an entire config source. See [Context files](./context-files.md) for the discovery-provider side.\n\n## Custom providers in `models.yml`\n\nCustom providers live in `~/.omp/agent/models.yml` under `providers:`. A provider ID defined there participates in the same selection, credential resolution, and `disabledProviders` rules as built-in providers.\n\nMinimal OpenAI-compatible provider:\n\n```yaml\nproviders:\n  my-openai-compatible:\n    baseUrl: https://api.example.com/v1\n    api: openai-completions\n    apiKey: MY_OPENAI_COMPATIBLE_KEY # env-var-name or literal\n    models:\n      - id: fast-chat\n        name: Fast Chat\n        contextWindow: 128000\n        maxTokens: 8192\n```\n\nKeyless local provider (no credentials required):\n\n```yaml\nproviders:\n  local-proxy:\n    baseUrl: http://127.0.0.1:4000/v1\n    api: openai-completions\n    auth: none\n    models:\n      - id: local-model\n        name: Local Model\n        contextWindow: 32768\n        maxTokens: 4096\n```\n\nDiscovery-enabled provider (models fetched from the endpoint at runtime):\n\n```yaml\nproviders:\n  team-proxy:\n    baseUrl: https://models.example.com/v1\n    apiKey: TEAM_PROXY_API_KEY\n    authHeader: true # send Authorization: Bearer <resolved key>\n    disableStrictTools: true\n    discovery:\n      type: proxy\n```\n\nFor the full schema, all allowed `api` values, discovery `type`s, model overrides, and equivalence settings, see [Model and Provider Configuration](./models.md).\n\nTo disable a custom provider, list its ID exactly:\n\n```yaml\ndisabledProviders:\n  - my-openai-compatible\n  - team-proxy\n```\n\n## Troubleshooting\n\n**A provider's models are not selectable.** Confirm the provider has credentials (`/login <provider>`, an exported environment variable, or a `models.yml` `apiKey`) and that its ID is not in the effective `disabledProviders` list. Remember the rule: not disabled **and** (keyless **or** has credentials). Keyless local engines only appear once the engine is actually running and responding.\n\n**The wrong key is being used (a stale key from `.env`).** Resolution favors runtime `--api-key`, then a `models.yml` config key, then stored credentials, then environment/`.env`. An already-set process environment variable also beats every `.env` file, and `<cwd>/.env` beats `~/.env`. If an unexpected key wins, check for an exported shell variable and the four `.env` files in precedence order, and clear the one that should not apply.\n\n**A provider still appears even though I disabled it.** `disabledProviders` arrays are replaced, not merged: a project `<project>/.omp/config.yml` array fully overrides the global one. Verify the *effective* list for the directory you are in (path-scoped entries only apply at or under their configured path), and confirm the ID is spelled exactly. Use `omp config get disabledProviders` to inspect the merged value (see [Settings](./settings.md)).\n\n**A discovery provider name had no effect on models (or vice-versa).** The ID namespace is shared. `gemini`, `codex`, `claude`, `native`, and `agents` are discovery-source IDs; the Google model backend is `google`. Make sure you are disabling the right kind of provider.\n\n**A custom `models.yml` provider does not load.** A YAML or schema error makes the registry skip the custom file. Validate the file with `omp --list-models`, confirm each provider has a `baseUrl`, a valid `api`, and at least one model entry, and that an implicit local engine is not silently shadowing it (an explicit `ollama`/`lm-studio`/`llama.cpp` entry replaces the built-in discovery for that ID). See [Model and Provider Configuration](./models.md).\n",
+	"providers.md": "# Providers\n\nProviders are the model backends `omp` can route requests to: Anthropic, OpenAI, Google Gemini, Groq, OpenRouter, Mistral, xAI, local engines like Ollama, hosted gateways, custom `models.yml` providers, and providers registered by extensions.\n\nA **provider** is the account or backend namespace, such as `anthropic`, `openai`, `google`, or `ollama`. A **model** is a concrete model under that provider, selected as `provider/model-id`, such as `anthropic/claude-opus-4-6`. Disabling a provider removes every model under it from selection; if you only want to narrow individual models, use model settings instead.\n\nThis page covers how providers become available, how credentials are resolved, the provider/environment-variable map, local engines, disabling providers, and custom providers. For model selection and the full `models.yml` schema, see [Model and Provider Configuration](./models.md). For config-file locations and merge precedence, see [Settings](./settings.md). For credential storage and login flows in depth, see [Secrets and credentials](./secrets.md). For the complete environment-variable reference, see [Environment variables](./environment-variables.md). For local engine setup, see [Local models](./local-models.md). For context-file discovery providers, see [Context files](./context-files.md).\n\n## How `omp` decides a provider is available\n\nAt startup the model registry assembles its catalog from four sources, in order:\n\n1. The bundled model catalog (every built-in provider and its known models).\n2. Custom provider and model entries from `~/.omp/agent/models.yml`.\n3. Runtime-discovered models for providers that support discovery (local engines and discovery-enabled gateways).\n4. Providers and models registered by extensions.\n\nThe registry can hold a model even when it is not currently selectable. A model becomes **available** only when both conditions hold:\n\n1. its provider ID is **not** in the effective `disabledProviders` list; **and**\n2. the provider is either **keyless** (an implicit local provider, or a custom provider with `auth: none`) **or** has resolvable credentials.\n\n`disabledProviders` is checked *before* credentials. If a provider ID is disabled, no stored key, OAuth session, environment variable, `.env` entry, or `models.yml` `apiKey` will make it selectable — the provider's models are dropped from availability regardless of credentials. Removing the ID from the effective list restores them.\n\nKeyless local engines are a special case: `ollama`, `llama.cpp`, and `lm-studio` are treated as keyless when no key is configured, so their discovered models are selectable as soon as the engine answers — no login required. See [Built-in local engines](#built-in-local-engines).\n\n## Credentials and precedence\n\nWhen a provider needs an API key, `omp` resolves it in this order (first match wins):\n\n1. **Runtime override** — a key supplied for the current process, e.g. CLI `--api-key`. Never persisted.\n2. **`models.yml` config key** — an `apiKey` pinned on a custom provider, registered as a config-sourced bearer. This deliberately beats stored OAuth, so a key supplied for a custom `baseUrl`/gateway is honored instead of forwarding an upstream OAuth token the proxy would reject.\n3. **Stored API key** — an API-key credential saved in the auth store.\n4. **Stored OAuth credential** — refreshed when needed; multiple accounts are ranked/rotated automatically.\n5. **Provider environment variable** — including values loaded from `.env` files (see [the env-var table](#environment-variables-and-env-files)).\n6. **`models.yml` fallback resolver** — keys for custom providers not otherwise registered.\n\nStored credentials live in the auth store at `~/.omp/agent/agent.db` for local auth, or in the configured auth-broker snapshot when running in broker mode. (`PI_CODING_AGENT_DIR` relocates the `~/.omp/agent` base, and the auth store moves with it.)\n\n### OAuth vs API key, and provider-scoped logins\n\nLogins are **provider-scoped**: authenticating `anthropic` does not authenticate `openai`, and each provider tracks its own credentials. A disabled provider stays disabled even with valid stored auth.\n\nUse the interactive slash commands inside a session:\n\n- `/login` — opens the OAuth/key selector. `/login <provider>` jumps straight to one provider (e.g. `/login anthropic`); for an OAuth flow that needs a pasted callback, run `/login <redirect-url>` to complete it.\n- `/logout` — opens the provider selector to remove stored credentials.\n\nFor headless or remote setups backed by a shared auth broker, the CLI exposes `omp auth-broker login <provider>` / `omp auth-broker logout` (and `status`, `list`, `import`, `migrate`). See [Secrets and credentials](./secrets.md) for the broker model.\n\nWhen a model has no credentials, `omp` tells you to run `/login` or set the provider's environment variable.\n\n### Pinning a key in `models.yml`\n\nA custom provider's `apiKey` is resolved as **environment-variable-name-or-literal**: if the value names an existing environment variable, that variable's value is used; otherwise the string itself is the key. Prefixing the value with `!` runs it as a shell command and uses the trimmed stdout (see [Model and Provider Configuration](./models.md) for the full value syntax).\n\n```yaml\n# ~/.omp/agent/models.yml\nproviders:\n  my-gateway:\n    baseUrl: https://gateway.example.com/v1\n    api: openai-completions\n    apiKey: MY_GATEWAY_API_KEY # reads this env var if set, else literal text\n    models:\n      - id: claude-sonnet\n        name: Claude Sonnet via Gateway\n        contextWindow: 200000\n        maxTokens: 8192\n```\n\nIf `authHeader: true` is set on a custom provider, the resolved key is injected as an `Authorization: Bearer <key>` header on every request to that provider.\n\n## Environment variables and `.env` files\n\nEach provider has one or more environment variables that supply a key when no stored credential exists. The table below is the verified provider → variable map; the full catalog is large, so it is split into core and additional providers. OAuth-backed providers can also accept a token variable in addition to (or instead of) an API key.\n\n### Core providers\n\n| Provider ID | Environment variable(s) |\n|---|---|\n| `anthropic` | `ANTHROPIC_OAUTH_TOKEN`, then `ANTHROPIC_API_KEY` (Foundry mode prefers `ANTHROPIC_FOUNDRY_API_KEY` when `CLAUDE_CODE_USE_FOUNDRY=true`) |\n| `openai` | `OPENAI_API_KEY` |\n| `openai-codex` | `OPENAI_CODEX_OAUTH_TOKEN` |\n| `google` | `GEMINI_API_KEY` |\n| `google-vertex` | `GOOGLE_CLOUD_API_KEY`, or Application Default Credentials (`GOOGLE_APPLICATION_CREDENTIALS` + `GOOGLE_CLOUD_PROJECT` + `GOOGLE_CLOUD_LOCATION`) |\n| `groq` | `GROQ_API_KEY` |\n| `openrouter` | `OPENROUTER_API_KEY` |\n| `mistral` | `MISTRAL_API_KEY` |\n| `xai` | `XAI_API_KEY` |\n| `xai-oauth` | `XAI_OAUTH_TOKEN`, then `XAI_API_KEY` |\n| `github-copilot` | `COPILOT_GITHUB_TOKEN` |\n| `cursor` | `CURSOR_ACCESS_TOKEN` |\n| `azure-openai-responses` | `AZURE_OPENAI_API_KEY` |\n| `amazon-bedrock` | `AWS_PROFILE`, or `AWS_ACCESS_KEY_ID` + `AWS_SECRET_ACCESS_KEY`, or an ECS/IRSA credential chain |\n\n### Additional hosted providers\n\n| Provider ID | Environment variable(s) |\n|---|---|\n| `cerebras` | `CEREBRAS_API_KEY` |\n| `deepseek` | `DEEPSEEK_API_KEY` |\n| `fireworks` | `FIREWORKS_API_KEY` |\n| `together` | `TOGETHER_API_KEY` |\n| `nvidia` | `NVIDIA_API_KEY` |\n| `huggingface` | `HUGGINGFACE_HUB_TOKEN`, then `HF_TOKEN` |\n| `moonshot` | `MOONSHOT_API_KEY` |\n| `kimi-code` | `KIMI_API_KEY` |\n| `nanogpt` | `NANO_GPT_API_KEY` |\n| `venice` | `VENICE_API_KEY` |\n| `vercel-ai-gateway` | `AI_GATEWAY_API_KEY` (also `VERCEL_AI_GATEWAY_API_KEY` for catalog discovery) |\n| `cloudflare-ai-gateway` | `CLOUDFLARE_AI_GATEWAY_API_KEY` |\n| `litellm` | `LITELLM_API_KEY` |\n| `kilo` | `KILO_API_KEY` |\n| `zai` | `ZAI_API_KEY` |\n| `zenmux` | `ZENMUX_API_KEY` |\n| `zhipu-coding-plan` | `ZHIPU_API_KEY` |\n| `qianfan` | `QIANFAN_API_KEY` |\n| `qwen-portal` | `QWEN_OAUTH_TOKEN`, then `QWEN_PORTAL_API_KEY` |\n| `synthetic` | `SYNTHETIC_API_KEY` |\n| `minimax` | `MINIMAX_API_KEY` |\n| `alibaba-coding-plan` | `ALIBABA_CODING_PLAN_API_KEY` |\n| `aimlapi` | `AIMLAPI_API_KEY` |\n| `gitlab-duo` | `GITLAB_TOKEN` |\n| `opencode-zen`, `opencode-go` | `OPENCODE_API_KEY` |\n| `firepass` | `FIREPASS_API_KEY` |\n| `wafer-pass` | `WAFER_PASS_API_KEY` |\n| `wafer-serverless` | `WAFER_SERVERLESS_API_KEY` |\n| `xiaomi` | `XIAOMI_API_KEY` |\n| `ollama-cloud` | `OLLAMA_CLOUD_API_KEY` |\n| `ollama` | `OLLAMA_API_KEY` (optional; local discovery is keyless by default) |\n| `lm-studio` | `LM_STUDIO_API_KEY` (optional; keyless by default) |\n| `llama.cpp` | `LLAMA_CPP_API_KEY` (only when the server requires auth) |\n\nOAuth-backed providers such as `anthropic`, `github-copilot`, `cursor`, `ollama-cloud`, `qwen-portal`, `xai-oauth`, `wafer-pass`, `wafer-serverless`, `google-gemini-cli`, and `google-antigravity` are normally reached through `/login` rather than an environment variable. See [Environment variables](./environment-variables.md) for search-tool and configuration variables not listed here.\n\n### `.env` discovery and precedence\n\n`omp` eagerly loads `.env` files into the process environment before any provider lookup. It reads four files and, for each variable, the **first** source that defines it wins. Effective precedence, high to low:\n\n1. The process environment inherited by `omp` (already-set variables always win).\n2. `<cwd>/.env`\n3. `~/.omp/agent/.env`\n4. `~/.omp/.env`\n5. `~/.env`\n\nA variable already present in the process environment is never overwritten by a `.env` file. Among the files, a value set in `<cwd>/.env` wins over `~/.omp/agent/.env`, which wins over `~/.omp/.env`, which wins over `~/.env`. So a shell-exported `OPENAI_API_KEY` beats every `.env` file, and a project's `<cwd>/.env` beats your home `~/.env`.\n\nProject-local `.env` is the simplest way to make one repository use a project-specific gateway, key, or local endpoint:\n\n```dotenv\n# <project>/.env\nOPENROUTER_API_KEY=sk-or-...\nOLLAMA_BASE_URL=http://127.0.0.1:11434\n```\n\n`.env` parsing is intentionally minimal:\n\n- blank lines and lines starting with `#` are ignored;\n- keys must match `[A-Za-z_][A-Za-z0-9_]*` (shell-identifier shape) — other names are dropped;\n- values may be wrapped in single or double quotes, which are stripped;\n- values containing a NUL byte are dropped;\n- an `OMP_`-prefixed key is also mirrored to the matching `PI_`-prefixed name.\n\n## Built-in local engines\n\nThree local engines are discovered automatically without needing a `models.yml` entry. Each uses a base URL that can be overridden by an environment variable:\n\n| Provider ID | Base URL (env override → default) | Notes |\n|---|---|---|\n| `ollama` | `OLLAMA_BASE_URL`, then `OLLAMA_HOST` (normalized), else `http://127.0.0.1:11434` | Keyless by default. |\n| `llama.cpp` | `LLAMA_CPP_BASE_URL`, else `http://127.0.0.1:8080` | Keyless unless a key is stored for `llama.cpp`. |\n| `lm-studio` | `LM_STUDIO_BASE_URL`, else `http://127.0.0.1:1234/v1` | Keyless by default. |\n\nThese implicit engines are **skipped** when:\n\n- a provider with the same ID is already configured in `models.yml` (your explicit config wins); or\n- the provider ID appears in the effective `disabledProviders` list.\n\nFor installing and running these engines, see [Local models](./local-models.md).\n\n## Disabling model providers\n\nUse the `disabledProviders` setting to remove a provider's models from selection:\n\n```yaml\n# ~/.omp/agent/config.yml or <project>/.omp/config.yml\ndisabledProviders:\n  - anthropic\n  - openai\n  - google\n  - groq\n```\n\nProvider IDs are matched exactly. Disable `google` to hide the Google Gemini API provider; the OAuth-backed Google providers `google-gemini-cli` and `google-antigravity` are separate IDs and must be disabled individually. Disable `ollama`, `llama.cpp`, or `lm-studio` to stop local discovery for that engine.\n\n`disabledProviders` applies uniformly to:\n\n- bundled catalog providers;\n- custom `models.yml` providers;\n- runtime-discovered provider models;\n- extension-registered providers;\n- implicit local engines.\n\nDisabling a provider does not delete its stored credentials — re-enable it by removing its ID from the effective list.\n\n## Project-specific provider control\n\nProject settings live in `<project>/.omp/config.yml`. Use them when one repository must allow or hide a different provider set than your global default:\n\n```yaml\n# <project>/.omp/config.yml\ndisabledProviders:\n  - openai\n  - openrouter\n```\n\nSettings arrays are **replaced** wholesale by the higher-precedence layer, not merged or appended. If the global file disables three providers and the project file disables one, the project sees only the project list:\n\n```yaml\n# ~/.omp/agent/config.yml\ndisabledProviders:\n  - anthropic\n  - openai\n  - google\n\n# <project>/.omp/config.yml\ndisabledProviders:\n  - groq\n```\n\nEffective result inside the project:\n\n```json\n[\"groq\"]\n```\n\nThe project array re-enables `anthropic`, `openai`, and `google` for sessions launched from that project. If you want a project to *add* to the global set, repeat the global IDs in the project file. See [Settings](./settings.md) for the full precedence chain, including `--config` overlays and runtime overrides.\n\n## Path-scoped `disabledProviders`\n\n`disabledProviders` can mix plain string entries (apply everywhere) with path-scoped entries (apply only when the current working directory matches a configured path):\n\n```yaml\ndisabledProviders:\n  - ollama\n  - path: ~/projects/sensitive\n    providers:\n      - anthropic\n      - openai\n  - paths:\n      - ~/work/client-a\n      - ~/work/client-b\n    values:\n      - openrouter\n```\n\n- Bare string entries always apply.\n- A scoped entry applies when the current working directory **is** the configured path or sits **under** it. `~` expands to the home directory.\n- Accepted path keys: `path`, `paths`, `pathPrefix`, `pathPrefixes`.\n- Accepted value keys: `providers`, `values`, `items`.\n\nFor the example above:\n\n- `ollama` is disabled everywhere.\n- `anthropic` and `openai` are additionally disabled under `~/projects/sensitive`.\n- `openrouter` is additionally disabled under `~/work/client-a` and `~/work/client-b`.\n\nPath scopes are resolved **after** the settings merge. Because a higher-precedence layer replaces the whole array, a project-level `disabledProviders` array drops any scoped entries that only existed in the global array. `enabledModels` is the only other setting that supports the same path-scoped form. See [Settings](./settings.md) for details.\n\n## Provider IDs vs discovery provider IDs\n\n`disabledProviders` uses a **single shared ID namespace** that gates two different subsystems:\n\n- **Model providers** — the backends on this page (`anthropic`, `openai`, `ollama`, a custom `models.yml` ID, …). Disabling one removes its models from selection.\n- **Discovery providers** — sources of context files, MCP servers, commands, skills, hooks, tools, prompts, and settings. Disabling one stops that source from contributing capability items.\n\n| Entry type | Examples | Effect |\n|---|---|---|\n| Model provider ID | `anthropic`, `openai`, `google`, `groq`, `openrouter`, `ollama`, `my-gateway` | Removes that provider's models from availability. |\n| Discovery provider ID | `native`, `claude`, `codex`, `gemini`, `agents`, `github` | Stops that discovery source from contributing capability items. |\n\nWatch the related names. The Google Gemini **API** models use the model provider ID `google`; `gemini` is a **discovery** provider ID (the source that reads `GEMINI.md`), not the Google model provider. Use discovery IDs only when you intend to disable an entire config source. See [Context files](./context-files.md) for the discovery-provider side.\n\n## Custom providers in `models.yml`\n\nCustom providers live in `~/.omp/agent/models.yml` under `providers:`. A provider ID defined there participates in the same selection, credential resolution, and `disabledProviders` rules as built-in providers.\n\nMinimal OpenAI-compatible provider:\n\n```yaml\nproviders:\n  my-openai-compatible:\n    baseUrl: https://api.example.com/v1\n    api: openai-completions\n    apiKey: MY_OPENAI_COMPATIBLE_KEY # env-var-name or literal\n    models:\n      - id: fast-chat\n        name: Fast Chat\n        contextWindow: 128000\n        maxTokens: 8192\n```\n\nKeyless local provider (no credentials required):\n\n```yaml\nproviders:\n  local-proxy:\n    baseUrl: http://127.0.0.1:4000/v1\n    api: openai-completions\n    auth: none\n    models:\n      - id: local-model\n        name: Local Model\n        contextWindow: 32768\n        maxTokens: 4096\n```\n\nDiscovery-enabled provider (models fetched from the endpoint at runtime):\n\n```yaml\nproviders:\n  team-proxy:\n    baseUrl: https://models.example.com/v1\n    apiKey: TEAM_PROXY_API_KEY\n    authHeader: true # send Authorization: Bearer <resolved key>\n    disableStrictTools: true\n    discovery:\n      type: proxy\n```\n\nFor the full schema, all allowed `api` values, discovery `type`s, model overrides, and equivalence settings, see [Model and Provider Configuration](./models.md).\n\nTo disable a custom provider, list its ID exactly:\n\n```yaml\ndisabledProviders:\n  - my-openai-compatible\n  - team-proxy\n```\n\n## Troubleshooting\n\n**A provider's models are not selectable.** Confirm the provider has credentials (`/login <provider>`, an exported environment variable, or a `models.yml` `apiKey`) and that its ID is not in the effective `disabledProviders` list. Remember the rule: not disabled **and** (keyless **or** has credentials). Keyless local engines only appear once the engine is actually running and responding.\n\n**The wrong key is being used (a stale key from `.env`).** Resolution favors runtime `--api-key`, then a `models.yml` config key, then stored credentials, then environment/`.env`. An already-set process environment variable also beats every `.env` file, and `<cwd>/.env` beats `~/.env`. If an unexpected key wins, check for an exported shell variable and the four `.env` files in precedence order, and clear the one that should not apply.\n\n**A provider still appears even though I disabled it.** `disabledProviders` arrays are replaced, not merged: a project `<project>/.omp/config.yml` array fully overrides the global one. Verify the *effective* list for the directory you are in (path-scoped entries only apply at or under their configured path), and confirm the ID is spelled exactly. Use `omp config get disabledProviders` to inspect the merged value (see [Settings](./settings.md)).\n\n**A discovery provider name had no effect on models (or vice-versa).** The ID namespace is shared. `gemini`, `codex`, `claude`, `native`, and `agents` are discovery-source IDs; the Google model backend is `google`. Make sure you are disabling the right kind of provider.\n\n**A custom `models.yml` provider does not load.** A YAML or schema error makes the registry skip the custom file. Validate the file with `omp models` (use `omp models find <substr>` to scope it to one provider), confirm each provider has a `baseUrl`, a valid `api`, and at least one model entry, and that an implicit local engine is not silently shadowing it (an explicit `ollama`/`lm-studio`/`llama.cpp` entry replaces the built-in discovery for that ID). See [Model and Provider Configuration](./models.md).\n",
 	"python-repl.md": "# Eval Tool Python Backend\n\nThis document describes the Python execution stack in `packages/coding-agent`.\nIt covers tool behavior, runner lifecycle, environment handling, execution semantics, output rendering, supported magics, and operational failure modes.\n\n## Scope and Key Files\n\n- Tool surface: `src/tools/eval.ts`\n- Session/per-call kernel orchestration: `src/eval/py/executor.ts`\n- Subprocess kernel client: `src/eval/py/kernel.ts`\n- Python wrapper / NDJSON server: `src/eval/py/runner.py`\n- Prelude helpers loaded into every kernel: `src/eval/py/prelude.py`\n- Host-side subagent helper bridge: `src/eval/agent-bridge.ts`\n- MIME bundle renderer (text + structured outputs): `src/eval/py/display.ts`\n- Interactive-mode renderer for user-triggered Python runs: `src/modes/components/eval-execution.ts`\n- Runtime/env filtering and Python resolution: `src/eval/py/runtime.ts`\n\n## What eval's Python backend is\n\nThe `eval` tool executes one or more Python cells inside a retained `python` subprocess that speaks NDJSON over stdin/stdout. No Jupyter gateway and no extra pip dependencies are required — a vanilla Python 3.8+ interpreter is enough. Rich `display()` output (PIL, pandas, plotly, matplotlib figures) keeps working because the wrapper implements MIME-bundle dispatch.\n\nTool params:\n\n```ts\n{\n  cells: Array<{\n    language: \"py\" | \"js\";\n    code: string;\n    title?: string;\n    timeout?: number; // seconds, clamped to 1..3600, default 30. Inactivity budget — see \"Cell timeout\".\n    reset?: boolean; // reset this cell's selected runtime before execution\n  }>;\n}\n```\n\nThe tool is `concurrency = \"exclusive\"` for a session, so calls do not overlap.\n\n## Kernel lifecycle\n\nEach Python kernel is a single subprocess: `<resolved-python> -u <runner.py>`. The runner is bundled with the host binary (Bun text import), written to an `omp-python-runner` cache under the OS temp directory once per script hash, and reused by subsequent spawns.\n\nKernel startup sequence:\n\n1. Availability check (`checkPythonKernelAvailability`) — verifies that a Python interpreter resolves and runs.\n2. Spawn `python -u runner.py` with filtered env and `cwd`.\n3. Send an init request that runs `os.chdir(cwd)`, injects env entries, and adds `cwd` to `sys.path`.\n4. Execute `PYTHON_PRELUDE` (idempotent — only initializes once per process).\n\nKernel shutdown:\n\n- Send `{\"type\": \"exit\"}` over stdin.\n- Wait for process exit with `SHUTDOWN_GRACE_MS` budget.\n- Escalate to `SIGTERM` and finally `SIGKILL` if the process does not exit in time.\n\n## Wire protocol (NDJSON, host ↔ runner)\n\nOne JSON object per line, UTF-8, `\\n` terminated.\n\nHost → runner:\n\n```jsonc\n{\"id\": \"<reqId>\", \"code\": \"<source>\", \"silent\": false, \"storeHistory\": true, \"cwd\": \"<optional>\", \"env\": {\"KEY\": \"VAL\"}}\n{\"type\": \"exit\"}\n```\n\nRunner → host:\n\n```jsonc\n{\"type\": \"started\",  \"id\": \"<reqId>\"}\n{\"type\": \"stdout\",   \"id\": \"<reqId>\", \"data\": \"...\"}\n{\"type\": \"stderr\",   \"id\": \"<reqId>\", \"data\": \"...\"}\n{\"type\": \"display\",  \"id\": \"<reqId>\", \"bundle\": {<mime>: <value>}}\n{\"type\": \"result\",   \"id\": \"<reqId>\", \"bundle\": {<mime>: <value>}}\n{\"type\": \"error\",    \"id\": \"<reqId>\", \"ename\": \"...\", \"evalue\": \"...\", \"traceback\": [\"...\"]}\n{\"type\": \"done\",     \"id\": \"<reqId>\", \"status\": \"ok\"|\"error\", \"executionCount\": N, \"cancelled\": false}\n```\n\nStatus events the prelude emits (e.g. `_emit_status(\"find\", count=…)`) ship inside display bundles under `application/x-omp-status` so the existing TUI status renderer keeps working.\n\n## Magics\n\nThe runner's source transformer rewrites IPython-style magics to plain Python calls before parsing. Supported set:\n\n| Magic                             | Effect                                                                                                                                                      |\n| --------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| `%pip <args>`                     | `python -m pip <args>` with live streaming output. Newly installed packages are evicted from `sys.modules` so the next `import` picks up the fresh install. |\n| `%cd <path>`                      | `os.chdir(path)` (with `~` expansion); emits status event.                                                                                                  |\n| `%pwd`                            | Returns `os.getcwd()`.                                                                                                                                      |\n| `%ls [path]`                      | Returns `sorted(os.listdir(path))`.                                                                                                                         |\n| `%env [KEY[=VAL]]`                | List, read, or set env vars (matches prelude `env()` semantics).                                                                                            |\n| `%set_env KEY VALUE`              | Set `os.environ[KEY]`.                                                                                                                                      |\n| `%time <expr>` / `%timeit <expr>` | Time the expression; emits status event with elapsed ms.                                                                                                    |\n| `%who` / `%whos`                  | List user-namespace names.                                                                                                                                  |\n| `%reset`                          | Clear user globals and re-inject prelude.                                                                                                                   |\n| `%load <path>`                    | Read a file into a fresh cell and execute.                                                                                                                  |\n| `%run <path>`                     | `runpy.run_path` and merge globals back.                                                                                                                    |\n| `%%bash` / `%%sh`                 | Run the cell body via `bash`/`sh`.                                                                                                                          |\n| `%%capture [name]`                | Run body with stdout/stderr captured into `name`.                                                                                                           |\n| `%%timeit`                        | Time the cell body.                                                                                                                                         |\n| `%%writefile <path>`              | Write body to file.                                                                                                                                         |\n| `!cmd` / `var = !cmd`             | Run command via subprocess shell; returns an SList-style result with `.n` / `.s` helpers.                                                                   |\n| `var = %name args`                | Assignment forms work for line magics and `!cmd`.                                                                                                           |\n\nUnknown magic names raise `NameError: UsageError: ...` inside the cell.\n\n## Session persistence semantics\n\n`python.kernelMode` controls retained kernel reuse:\n\n- `session` (default)\n  - Reuses kernel sessions keyed by namespaced eval session id plus normalized cwd and interpreter.\n  - Multiple owners can share the same retained kernel for that key.\n  - Calls through the tool are exclusive, so tool invocations do not overlap.\n  - A dead retained subprocess is replaced before execution.\n  - If the subprocess dies during execution, it is replaced and the cell is retried once.\n- `per-call`\n  - Spawns a fresh subprocess for each request.\n  - Shuts the subprocess down after the request.\n  - No cross-call state persistence.\n\n### Multi-cell behavior in a single tool call\n\nPython cells run sequentially in the same selected Python kernel instance for that tool call.\n\nIf an intermediate cell fails:\n\n- Earlier cell state remains in memory.\n- Tool returns a targeted error indicating which cell failed.\n- Later cells are not executed.\n\n`reset=true` is per cell and resets that language runtime before the cell executes.\n\n## Environment filtering and runtime resolution\n\nEnvironment is filtered before launching the runner:\n\n- Allowlist includes core vars like `PATH`, `HOME`, locale vars, `VIRTUAL_ENV`, `PYTHONPATH`, etc.\n- Allow-prefixes: `LC_`, `XDG_`, `PI_`\n- Denylist strips common API keys (OpenAI/Anthropic/Gemini/etc.)\n\nRuntime selection order (skipped entirely when the `python.interpreter` setting names an explicit executable):\n\n1. Active/located venv (`VIRTUAL_ENV`, then `CONDA_PREFIX`, then `<cwd>/.venv`, `<cwd>/venv`)\n2. Managed venv at `~/.omp/python-env`\n3. `python` or `python3` on PATH\n\nWhen a venv is selected, its bin/Scripts path is prepended to `PATH`.\n\nThe runner additionally receives `PYTHONUNBUFFERED=1` and `PYTHONIOENCODING=utf-8` so streamed output reaches the host promptly.\n\n## Tool availability and mode selection\n\n`eval.py` / `eval.js` (both default `true`) plus optional boolean env flags `PI_PY` / `PI_JS` control eval backend exposure:\n\n- Python backend only (`eval.py=true`, `eval.js=false`, or `PI_PY=1 PI_JS=0`)\n- JavaScript backend only (`eval.py=false`, `eval.js=true`, or `PI_PY=0 PI_JS=1`)\n- both backends (`eval.py=true`, `eval.js=true`, or `PI_PY=1 PI_JS=1`)\n\n`PI_PY` and `PI_JS` use normal boolean flag parsing. Each flag, when set, overrides only its own setting; an unset flag falls back to its setting (`eval.py` / `eval.js`, both default `true`).\n\nIf Python preflight fails and `eval.js` is enabled, `eval` remains available for `js` cells; `py` cells fail with a Python-backend availability error.\n\nPython prelude helpers include `agent(prompt, *, agent_type=\"task\", model=None, label=None, schema=None)`. It synchronously calls the host bridge, runs one subagent through the task executor, and returns the final text. When `schema` is supplied, the helper parses the subagent's JSON output and returns the object.\n\n## Execution flow and cancellation/timeout\n\n### Cell timeout\n\nEach eval cell `timeout` is in seconds, defaults to 30, and is clamped to `1..3600`. It is a **wall-clock budget on the cell's own work** that the watchdog (`IdleTimeout`, `src/eval/idle-timeout.ts`) enforces, **but it is suspended while a host-side `agent()`/`parallel()`/`completion()` bridge call is in flight**: those calls emit synthetic pause/resume timeout-control status events (`withBridgeTimeoutPause`, `src/eval/bridge-timeout.ts`) that pause the watchdog entirely and start a fresh timeout window when control returns to the runtime, so a long fanout or a slow completion runs to completion instead of being killed mid-stream. Pause is reference-counted because `parallel()` can have multiple bridge calls in flight at once.\n\nThe pause/resume events are the **sole** mechanism that suspends the budget. Everything else the cell does — compute, `stdout`/`stderr`, `log()`/`phase()`, and ordinary (non-agent) tool calls — counts against `timeout`, so a cell that is not delegating to an agent/completion is bounded by a plain wall-clock timeout. The tool combines the caller abort signal, the session abort signal, and the watchdog's signal with `AbortSignal.any(...)`; no wall-clock deadline is passed to the backend, so neither runtime arms a competing fixed timer.\n\n### Kernel execution cancellation\n\nOn abort/timeout:\n\n- The host sends `kill(\"SIGINT\")` to the runner subprocess.\n- The runner's exec-time signal handler raises `KeyboardInterrupt` inside the user code.\n- Result includes `cancelled=true`; a kernel timeout is annotated as `eval cell timed out after <n>s; kernel interrupted but remains running. Reset the kernel via { reset: true } if state appears corrupted.`\n- Between requests the runner installs `SIG_IGN` for SIGINT so a stray cancel does not tear down the kernel.\n\nIf the runner does not emit `done` within 5s of the interrupt (`INTERRUPT_ESCALATION_MS` — e.g. stuck in C code holding the GIL), the host shuts the subprocess down (escalating `exit` → `SIGTERM` → `SIGKILL`), the cell is annotated as kernel-killed, and the kernel is recreated on the next call.\n\n### stdin behavior\n\nInteractive stdin is not supported. The runner does not forward `input()` prompts; user code that calls `input()` blocks until cancellation.\n\n## Output capture and rendering\n\n### Captured output classes\n\nFrom runner frames:\n\n- `stdout` / `stderr` → plain text chunks\n- `display` / `result` → rich display handling (MIME bundle)\n- `error` → traceback text\n- `application/x-omp-status` MIME inside `display` → structured status events\n\nDisplay MIME precedence:\n\n1. `text/markdown`\n2. `text/plain`\n3. `text/html` (converted to basic markdown)\n\nAdditionally captured as structured outputs:\n\n- `application/json` → JSON tree data\n- `image/png` / `image/jpeg` → image payloads\n- `application/x-omp-status` → status events\n\n### Matplotlib\n\nThe runner sets `MPLBACKEND=Agg` as an environ default so figures render off-screen. After every cell, `pyplot.get_fignums()` is iterated; each figure is saved to PNG, emitted as an `image/png` display, and closed.\n\n### Storage and truncation\n\nOutput is streamed through `OutputSink` and may be persisted to artifact storage. Tool results can include truncation metadata and `artifact://<id>` for full output recovery.\n\n### Renderer behavior\n\n- Tool renderer (`eval.ts`):\n  - shows code-cell blocks with per-cell status\n  - collapsed preview defaults to 10 lines\n  - supports expanded mode for all output retained in the tool result\n- Interactive renderer (`eval-execution.ts`):\n  - used for user-triggered Python execution in TUI\n  - collapsed preview defaults to 20 lines\n  - clamps very long individual lines to 4000 chars for display safety\n  - shows cancellation/error/truncation notices\n\n## Operational troubleshooting\n\n- **Python backend not available** — Check `eval.py`, `PI_PY`, and that `python`/`python3` is on PATH. If preflight fails and `eval.js` is enabled, use a `js` cell.\n- **No Python on PATH** — Install a system Python 3.8+ or place a venv at `~/.omp/python-env`. `omp setup python --check` reports the resolved interpreter.\n- **Execution hangs then times out** — Increase tool `timeout` (max 3600s) if workload is legitimate. For stuck native code, cancellation triggers `SIGINT` first then escalates; the session restarts on the next request.\n- **stdin/input prompts in Python code** — `input()` is not supported; pass data programmatically.\n- **Working directory errors** — Tool validates `cwd` exists and is a directory before execution.\n\n## Relevant environment variables\n\n- `PI_PY` / `PI_JS` — eval backend exposure overrides\n- `PI_PYTHON_SKIP_CHECK=1` — bypass Python preflight/warm checks\n- `PI_PYTHON_INTEGRATION=1` — enable gated integration tests that spawn a real Python\n- `PI_PYTHON_IPC_TRACE=1` — log NDJSON frames exchanged with the runner subprocess\n",
 	"render-mermaid.md": "# RenderMermaid\n\n`RenderMermaid` is an optional built-in tool that renders Mermaid source to terminal-friendly text.\n\n## Enable it\n\nDisabled by default. Turn it on in `/settings` under **Tools → Render Mermaid**, or in `~/.omp/agent/config.yml`:\n\n```yaml\nrenderMermaid:\n  enabled: true\n```\n\n## What it does\n\n- Tool name: `render_mermaid`\n- Input: Mermaid source in the required `mermaid` field\n- Output: rendered ASCII/Unicode text, not SVG or PNG\n- Storage: when artifact storage is available, the full render is also saved as an `artifact://...`\n\nThere are no model-specific or environment-variable prerequisites. Once enabled, any model that can call built-in tools can use it.\n\n## Parameters\n\n```json\n{\n  \"mermaid\": \"graph TD\\n  A[Start] --> B[Stop]\",\n  \"config\": {\n    \"useAscii\": false,\n    \"paddingX\": 2,\n    \"paddingY\": 2,\n    \"boxBorderPadding\": 0\n  }\n}\n```\n\nAvailable `config` fields:\n\n- `useAscii` — `true` for plain ASCII, `false` for Unicode box-drawing characters (default and usually more readable)\n- `paddingX` — horizontal spacing between nodes\n- `paddingY` — vertical spacing between nodes\n- `boxBorderPadding` — inner padding inside node boxes\n\n## Current limitations\n\n`RenderMermaid` uses the `beautiful-mermaid` ASCII renderer. It works best for flowcharts and small diagrams.\n\nComplex sequence diagrams, especially with `alt` / `else` blocks, can become very wide in a terminal. That is current renderer behavior, not a provider or model configuration problem.\n\nIf a sequence diagram is hard to read:\n\n1. Keep Unicode output (`useAscii: false`)\n2. Reduce spacing with a tighter config such as `paddingX: 2`, `paddingY: 2`, `boxBorderPadding: 0`\n3. Prefer smaller sub-diagrams over one large sequence diagram\n4. Open the saved artifact if the inline preview is truncated in the TUI\n\n## Example\n\nInput:\n\n```mermaid\ngraph TD\n  A[Start] --> B{Decision}\n  B -->|Yes| C[Action]\n  B -->|No| D[End]\n```\n\nTypical result:\n\n```text\n┌─────┐\n│Start│\n└─────┘\n   │\n   ▼\n┌────────┐\n│Decision│\n└────────┘\n```\n",
 	"resolve-tool-runtime.md": "# Resolve tool runtime internals\n\nThis document explains how preview/apply workflows are modeled in coding-agent and how built-in or custom tools can participate via the tool-choice queue and `pushPendingAction`.\n\n## Scope and key files\n\n- [`src/tools/resolve.ts`](../packages/coding-agent/src/tools/resolve.ts)\n- [`src/tools/ast-edit.ts`](../packages/coding-agent/src/tools/ast-edit.ts)\n- [`src/extensibility/custom-tools/types.ts`](../packages/coding-agent/src/extensibility/custom-tools/types.ts)\n- [`src/extensibility/custom-tools/loader.ts`](../packages/coding-agent/src/extensibility/custom-tools/loader.ts)\n- [`src/sdk.ts`](../packages/coding-agent/src/sdk.ts)\n\n## What `resolve` does\n\n`resolve` is a hidden tool that finalizes a pending preview action.\n\n- `action: \"apply\"` executes the queued action's `apply(reason, extra)` callback and returns that result with resolve metadata.\n- `action: \"discard\"` invokes `reject(reason, extra)` if provided; otherwise returns `Discarded: <label>. Reason: <reason>`.\n- `extra` is optional free-form metadata. Queue handlers receive it; producers decide whether it has meaning.\n\nIf no pending action exists, `resolve(action=\"apply\")` fails with:\n\n- `No pending action to resolve. Nothing to apply or discard.`\n\n`resolve(action=\"discard\")` with no pending action succeeds instead, returning `Nothing to discard; no pending action remains.` — the desired end-state (no staged change) already holds.\n\n## Pending actions use the tool-choice queue\n\nPreview producers call `queueResolveHandler(...)`, which pushes a one-shot forced `resolve` directive onto the session tool-choice queue and adds a `resolve-reminder` steering message.\n\nRuntime behavior:\n\n- the queued handler owns the pending `apply`/`reject` callbacks,\n- `resolve` looks up the current queue invoker with `session.peekQueueInvoker()`,\n- if the model rejects the forced tool choice, the queue directive is requeued,\n- `resolve` does not maintain a separate pending-action stack.\n\n`resolve` also checks a standing resolve handler after the queue invoker; this is used by long-lived approval flows that are not ordinary preview tool calls.\n\nMultiple pending previews therefore follow the active tool-choice queue ordering, not an independent pending-action store. If an apply callback throws, the queued helper re-pushes the same resolve directive and reminder so the preview can still be discarded or retried.\n\n## Built-in producer example (`ast_edit`)\n\n`ast_edit` previews structural replacements first. When the preview has replacements and is not applied yet, it queues a resolve handler that contains:\n\n- label (human-readable summary)\n- `sourceToolName` (`ast_edit`)\n- `apply(reason: string, extra?: Record<string, unknown>)` callback that reruns AST edit with `dryRun: false`\n\n`resolve(action=\"apply\", reason=\"...\")` passes `reason` into this callback. `ast_edit` currently ignores `extra`.\n\n## Custom tools: `pushPendingAction`\n\nCustom tools can register resolve-compatible pending actions through `CustomToolAPI.pushPendingAction(...)`. The custom tool loader forwards these actions to `queueResolveHandler(...)` when that hook is available.\n\n`CustomToolPendingAction`:\n\n- `label: string` (required)\n- `apply(reason: string): Promise<AgentToolResult<unknown>>` (required) — invoked on apply; `reason` is the string passed to `resolve`\n- `reject?(reason: string): Promise<AgentToolResult<unknown> | undefined>` (optional) — invoked on discard; return value replaces the default \"Discarded\" message if provided\n- `details?: unknown` exists on the public custom-tool type but is not currently forwarded by the loader into resolve metadata\n- `sourceToolName?: string` (optional, defaults to `\"custom_tool\"`)\n\n### Minimal usage example\n\n```ts\nimport type { CustomToolFactory } from \"@oh-my-pi/pi-coding-agent\";\n\nconst factory: CustomToolFactory = (pi) => ({\n  name: \"batch_rename_preview\",\n  label: \"Batch Rename Preview\",\n  description: \"Previews renames and defers commit to resolve\",\n  parameters: pi.zod.object({\n    files: pi.zod.array(pi.zod.string()),\n  }),\n\n  async execute(_toolCallId, params) {\n    const previewSummary = `Prepared rename plan for ${params.files.length} files`;\n\n    pi.pushPendingAction({\n      label: `Batch rename: ${params.files.length} files`,\n      sourceToolName: \"batch_rename_preview\",\n      apply: async (reason) => {\n        // apply writes here\n        return {\n          content: [\n            { type: \"text\", text: `Applied batch rename. Reason: ${reason}` },\n          ],\n        };\n      },\n      reject: async (reason) => {\n        // optional: cleanup or notify on discard\n        return {\n          content: [\n            { type: \"text\", text: `Discarded batch rename. Reason: ${reason}` },\n          ],\n        };\n      },\n    });\n\n    return {\n      content: [\n        {\n          type: \"text\",\n          text: `${previewSummary}. Call resolve to apply or discard.`,\n        },\n      ],\n    };\n  },\n});\n\nexport default factory;\n```\n\n## Runtime availability and failures\n\n`pushPendingAction` is wired by the custom tool loader through the active session's resolve queue hook.\n\nIf the runtime did not provide the resolve queue hook, `pushPendingAction` throws:\n\n- `Pending action store unavailable for custom tools in this runtime.`\n\n## Tool-choice behavior\n\nWhen `queueResolveHandler(...)` registers a preview, the agent runtime forces a one-shot `resolve` tool choice so pending previews are explicitly finalized before normal tool flow continues.\n\n## Developer guidance\n\n- Use pending actions only for destructive or high-impact operations that should support explicit apply/discard.\n- Keep `label` concise and specific; it is shown in resolve renderer output.\n- Ensure `apply(reason)` is deterministic and idempotent enough for one-shot execution; `reason` is informational and should not change behavior.\n- Implement `reject(reason)` when the discard needs cleanup (temp state, locks, notifications); omit it for stateless previews where the default message suffices.\n- If your tool can stage multiple previews, remember they are mediated by the tool-choice queue rather than a separate pending-action stack.\n",
-	"rpc.md": "# RPC Protocol Reference\n\nRPC mode runs the coding agent as a newline-delimited JSON protocol over stdio.\n\n- **stdin**: commands (`RpcCommand`), extension UI responses, and host-tool updates/results\n- **stdout**: a ready frame, command responses (`RpcResponse`), session/agent events, extension UI requests, host-tool requests/cancellations\n\nPrimary implementation:\n\n- `src/modes/rpc/rpc-mode.ts`\n- `src/modes/rpc/rpc-types.ts`\n- `src/session/agent-session.ts`\n- `packages/agent/src/agent.ts`\n- `packages/agent/src/agent-loop.ts`\n\n## Startup\n\n```bash\nomp --mode rpc [regular CLI options]\n```\n\nBehavior notes:\n\n- `@file` CLI arguments are rejected in RPC mode.\n- RPC mode disables automatic session title generation by default to avoid an extra model call.\n- RPC mode resets workflow-altering `todo.*`, `task.*`, `memory.backend`/`memories.enabled`, `async.*`, and `bash.autoBackground.*` settings to their built-in defaults instead of inheriting user overrides.\n- The process reads stdin as JSONL (`readJsonl(Bun.stdin.stream())`).\n- At startup it writes `{ \"type\": \"ready\" }` before processing commands.\n- When stdin closes, pending host-tool calls and host-URI requests are rejected and the process exits with code `0`.\n- Responses/events are written as one JSON object per line.\n\n## Transport and Framing\n\nEach frame is a single JSON object followed by `\\n`.\n\nThere is no envelope beyond the object shape itself.\n\n### Outbound frame categories (stdout)\n\n1. Ready frame (`{ type: \"ready\" }`)\n2. `RpcResponse` (`{ type: \"response\", ... }`)\n3. `AgentSessionEvent` objects (`agent_start`, `message_update`, etc.)\n4. `RpcExtensionUIRequest` (`{ type: \"extension_ui_request\", ... }`)\n5. Host tool requests/cancellations (`host_tool_call`, `host_tool_cancel`)\n6. Host URI requests/cancellations (`host_uri_request`, `host_uri_cancel`)\n7. Extension errors (`{ type: \"extension_error\", extensionPath, event, error }`)\n8. Available-commands updates (`{ type: \"available_commands_update\", commands }`), emitted at startup and whenever command metadata changes\n9. Subagent frames (`subagent_lifecycle`, `subagent_progress`, `subagent_event`), gated by `set_subagent_subscription`\n10. Builtin slash-command side channels (`command_output`, `session_info_update`, `config_update`)\n\n### Inbound frame categories (stdin)\n\n1. `RpcCommand`\n2. `RpcExtensionUIResponse` (`{ type: \"extension_ui_response\", ... }`)\n3. Host tool updates/results (`host_tool_update`, `host_tool_result`)\n4. Host URI results (`host_uri_result`)\n\n## Request/Response Correlation\n\nAll commands accept optional `id?: string`.\n\n- If provided, normal command responses echo the same `id`.\n- `RpcClient` relies on this for pending-request resolution.\n\nImportant edge behavior from runtime:\n\n- Unknown command responses are emitted with `id: undefined` (even if the request had an `id`).\n- Parse/handler exceptions in the input loop emit `command: \"parse\"` with `id: undefined`.\n- `prompt` and `abort_and_prompt` return immediate success, then may emit a later error response with the **same** id if async prompt scheduling fails.\n\n## Command Schema (canonical)\n\n`RpcCommand` is defined in `src/modes/rpc/rpc-types.ts`:\n\n### Prompting\n\n- `{ id?, type: \"prompt\", message: string, images?: ImageContent[], streamingBehavior?: \"steer\" | \"followUp\" }`\n- `{ id?, type: \"steer\", message: string, images?: ImageContent[] }`\n- `{ id?, type: \"follow_up\", message: string, images?: ImageContent[] }`\n- `{ id?, type: \"abort\" }`\n- `{ id?, type: \"abort_and_prompt\", message: string, images?: ImageContent[] }`\n- `{ id?, type: \"new_session\", parentSession?: string }`\n\n### State\n\n- `{ id?, type: \"get_state\" }`\n- `{ id?, type: \"get_available_commands\" }`\n- `{ id?, type: \"set_todos\", phases: TodoPhase[] }`\n- `{ id?, type: \"set_host_tools\", tools: RpcHostToolDefinition[] }`\n- `{ id?, type: \"set_host_uri_schemes\", schemes: RpcHostUriSchemeDefinition[] }`\n- `{ id?, type: \"set_subagent_subscription\", level: \"off\" | \"progress\" | \"events\" }`\n- `{ id?, type: \"get_subagents\" }`\n- `{ id?, type: \"get_subagent_messages\", subagentId?: string, sessionFile?: string, fromByte?: number }`\n\n### Model\n\n- `{ id?, type: \"set_model\", provider: string, modelId: string }`\n- `{ id?, type: \"cycle_model\" }`\n- `{ id?, type: \"get_available_models\" }`\n\n### Thinking\n\n- `{ id?, type: \"set_thinking_level\", level: ThinkingLevel }`\n- `{ id?, type: \"cycle_thinking_level\" }`\n\n### Queue modes\n\n- `{ id?, type: \"set_steering_mode\", mode: \"all\" | \"one-at-a-time\" }`\n- `{ id?, type: \"set_follow_up_mode\", mode: \"all\" | \"one-at-a-time\" }`\n- `{ id?, type: \"set_interrupt_mode\", mode: \"immediate\" | \"wait\" }`\n\n### Compaction\n\n- `{ id?, type: \"compact\", customInstructions?: string }`\n- `{ id?, type: \"set_auto_compaction\", enabled: boolean }`\n\n### Retry\n\n- `{ id?, type: \"set_auto_retry\", enabled: boolean }`\n- `{ id?, type: \"abort_retry\" }`\n\n### Bash\n\n- `{ id?, type: \"bash\", command: string }`\n- `{ id?, type: \"abort_bash\" }`\n\n### Session\n\n- `{ id?, type: \"get_session_stats\" }`\n- `{ id?, type: \"export_html\", outputPath?: string }`\n- `{ id?, type: \"switch_session\", sessionPath: string }`\n- `{ id?, type: \"branch\", entryId: string }`\n- `{ id?, type: \"get_branch_messages\" }`\n- `{ id?, type: \"get_last_assistant_text\" }`\n- `{ id?, type: \"set_session_name\", name: string }`\n- `{ id?, type: \"handoff\", customInstructions?: string }`\n\n### Messages\n\n- `{ id?, type: \"get_messages\" }`\n\n### Login\n\n- `{ id?, type: \"get_login_providers\" }`\n- `{ id?, type: \"login\", providerId: string }`\n\n## Response Schema\n\nAll command results use `RpcResponse`:\n\n- Success: `{ id?, type: \"response\", command: <command>, success: true, data?: ... }`\n- Failure: `{ id?, type: \"response\", command: string, success: false, error: string }`\n\nData payloads are command-specific and defined in `rpc-types.ts`.\n\n### `get_state` payload\n\n```json\n{\n  \"model\": { \"provider\": \"...\", \"id\": \"...\" },\n  \"thinkingLevel\": \"off|minimal|low|medium|high|xhigh\",\n  \"isStreaming\": false,\n  \"isCompacting\": false,\n  \"steeringMode\": \"all|one-at-a-time\",\n  \"followUpMode\": \"all|one-at-a-time\",\n  \"interruptMode\": \"immediate|wait\",\n  \"sessionFile\": \"...\",\n  \"sessionId\": \"...\",\n  \"sessionName\": \"...\",\n  \"autoCompactionEnabled\": true,\n  \"messageCount\": 0,\n  \"queuedMessageCount\": 0,\n  \"todoPhases\": [\n    {\n      \"id\": \"phase-1\",\n      \"name\": \"Todos\",\n      \"tasks\": [\n        {\n          \"id\": \"task-1\",\n          \"content\": \"Map the tool surface\",\n          \"status\": \"in_progress\"\n        }\n      ]\n    }\n  ],\n  \"systemPrompt\": [\"...\"],\n  \"dumpTools\": [\n    {\n      \"name\": \"read\",\n      \"description\": \"Read files and URLs\",\n      \"parameters\": {}\n    }\n  ],\n  \"contextUsage\": {\n    \"tokens\": 1100,\n    \"contextWindow\": 200000,\n    \"percent\": 0.55\n  }\n}\n```\n\n### `set_todos` payload\n\nReplaces the in-memory todo state for the current session and returns the normalized phase list:\n\n```json\n{\n  \"id\": \"req_2\",\n  \"type\": \"set_todos\",\n  \"phases\": [\n    {\n      \"id\": \"phase-1\",\n      \"name\": \"Evaluation\",\n      \"tasks\": [\n        {\n          \"id\": \"task-1\",\n          \"content\": \"Map the read tool surface\",\n          \"status\": \"in_progress\"\n        },\n        {\n          \"id\": \"task-2\",\n          \"content\": \"Exercise edit operations\",\n          \"status\": \"pending\"\n        }\n      ]\n    }\n  ]\n}\n```\n\nThis is useful for hosts that want to pre-seed a plan before the first prompt.\n\n### `set_host_tools` payload\n\nReplaces the current set of host-owned tools that the RPC server may call back\ninto over stdio:\n\n```json\n{\n  \"id\": \"req_3\",\n  \"type\": \"set_host_tools\",\n  \"tools\": [\n    {\n      \"name\": \"echo_host\",\n      \"label\": \"Echo Host\",\n      \"description\": \"Echo a value from the embedding host\",\n      \"parameters\": {\n        \"type\": \"object\",\n        \"properties\": {\n          \"message\": { \"type\": \"string\" }\n        },\n        \"required\": [\"message\"],\n        \"additionalProperties\": false\n      }\n    }\n  ]\n}\n```\n\nThe response payload is:\n\n```json\n{\n  \"toolNames\": [\"echo_host\"]\n}\n```\n\nThese tools are added to the active session tool registry before the next model\ncall. Re-sending `set_host_tools` replaces the previous host-owned set.\n\n### `set_host_uri_schemes` payload\n\nReplaces the current set of host-owned URL schemes the RPC server should\ndispatch reads/writes through:\n\n```json\n{\n  \"id\": \"req_4\",\n  \"type\": \"set_host_uri_schemes\",\n  \"schemes\": [\n    {\n      \"scheme\": \"db\",\n      \"description\": \"Virtual db row files\",\n      \"writable\": true,\n      \"immutable\": false\n    }\n  ]\n}\n```\n\nThe response payload is:\n\n```json\n{\n  \"schemes\": [\"db\"]\n}\n```\n\nSchemes are case-insensitive on the wire and normalized to lowercase before\nthe response is sent. Re-sending `set_host_uri_schemes` replaces the entire\nprevious set — schemes missing from the new list are unregistered.\n\n## Event Stream Schema\n\nRPC mode forwards `AgentSessionEvent` objects from `AgentSession.subscribe(...)`.\n\nCommon event types:\n\n- `agent_start`, `agent_end`\n- `turn_start`, `turn_end`\n- `message_start`, `message_update`, `message_end`\n- `tool_execution_start`, `tool_execution_update`, `tool_execution_end`\n- `auto_compaction_start`, `auto_compaction_end`\n- `auto_retry_start`, `auto_retry_end`\n- `ttsr_triggered`\n- `todo_reminder`\n- `todo_auto_clear`\n\nExtension runner errors are emitted separately as:\n\n```json\n{\n  \"type\": \"extension_error\",\n  \"extensionPath\": \"...\",\n  \"event\": \"...\",\n  \"error\": \"...\"\n}\n```\n\n`message_update` includes streaming deltas in `assistantMessageEvent` (text/thinking/toolcall deltas).\n\n## Prompt/Queue Concurrency and Ordering\n\nThis is the most important operational behavior.\n\n### Immediate ack vs completion\n\n`prompt` and `abort_and_prompt` are **acknowledged immediately**:\n\n```json\n{ \"id\": \"req_1\", \"type\": \"response\", \"command\": \"prompt\", \"success\": true }\n```\n\nThat means:\n\n- command acceptance != run completion\n- final completion is observed via `agent_end`\n\n### While streaming\n\n`AgentSession.prompt()` requires `streamingBehavior` during active streaming:\n\n- `\"steer\"` => queued steering message (interrupt path)\n- `\"followUp\"` => queued follow-up message (post-turn path)\n\nIf omitted during streaming, prompt fails.\n\n### Queue defaults\n\nFrom `packages/agent/src/agent.ts` defaults:\n\n- `steeringMode`: `\"one-at-a-time\"`\n- `followUpMode`: `\"one-at-a-time\"`\n- `interruptMode`: `\"immediate\"`\n\n### Mode semantics\n\n- `set_steering_mode` / `set_follow_up_mode`\n  - `\"one-at-a-time\"`: dequeue one queued message per turn\n  - `\"all\"`: dequeue entire queue at once\n- `set_interrupt_mode`\n  - `\"immediate\"`: tool execution checks steering between tool calls; pending steering can abort remaining tool calls in the turn\n  - `\"wait\"`: defer steering until turn completion\n\n## Extension UI Sub-Protocol\n\nExtensions in RPC mode use request/response UI frames.\n\n### Outbound request\n\n`RpcExtensionUIRequest` (`type: \"extension_ui_request\"`) methods:\n\n- `select`, `confirm`, `input`, `editor`, `cancel`\n- `notify`, `setStatus`, `setWidget`, `setTitle`, `set_editor_text`\n- `open_url` (emitted by RPC login flows)\n\nRuntime note:\n\n- Automatic session title generation is disabled in RPC mode, and `setTitle` UI\n  requests are also suppressed by default because most hosts do not have a\n  meaningful terminal-title surface. Set `PI_RPC_EMIT_TITLE=1` to opt back in to\n  the UI event only.\n\nExample:\n\n```json\n{\n  \"type\": \"extension_ui_request\",\n  \"id\": \"123\",\n  \"method\": \"confirm\",\n  \"title\": \"Confirm\",\n  \"message\": \"Continue?\",\n  \"timeout\": 30000\n}\n```\n\n### Inbound response\n\n`RpcExtensionUIResponse` (`type: \"extension_ui_response\"`):\n\n- `{ type: \"extension_ui_response\", id: string, value: string }`\n- `{ type: \"extension_ui_response\", id: string, confirmed: boolean }`\n- `{ type: \"extension_ui_response\", id: string, cancelled: true, timedOut?: boolean }`\n\nIf a dialog has a timeout, RPC mode resolves to a default value when timeout/abort fires.\n\n## Host Tool Sub-Protocol\n\nRPC hosts can expose custom tools to the agent by sending `set_host_tools`, then\nserving execution requests over the same transport.\n\n### Outbound request\n\nWhen the agent wants the host to execute one of those tools, RPC mode emits:\n\n```json\n{\n  \"type\": \"host_tool_call\",\n  \"id\": \"host_1\",\n  \"toolCallId\": \"toolu_123\",\n  \"toolName\": \"echo_host\",\n  \"arguments\": { \"message\": \"hello\" }\n}\n```\n\nIf the tool execution is later aborted, RPC mode emits:\n\n```json\n{\n  \"type\": \"host_tool_cancel\",\n  \"id\": \"host_cancel_1\",\n  \"targetId\": \"host_1\"\n}\n```\n\n### Inbound updates and completion\n\nHosts can optionally stream progress:\n\n```json\n{\n  \"type\": \"host_tool_update\",\n  \"id\": \"host_1\",\n  \"partialResult\": {\n    \"content\": [{ \"type\": \"text\", \"text\": \"working\" }]\n  }\n}\n```\n\nCompletion uses:\n\n```json\n{\n  \"type\": \"host_tool_result\",\n  \"id\": \"host_1\",\n  \"result\": {\n    \"content\": [{ \"type\": \"text\", \"text\": \"done\" }]\n  }\n}\n```\n\nSet top-level `isError: true` on `host_tool_result` to reject the pending host tool call and surface the returned text content as a tool error.\n\n## Host URI Sub-Protocol\n\nRPC hosts can also own custom URL schemes (virtual files). After\n`set_host_uri_schemes`, every read of `<scheme>://…` and write of\n`<scheme>://…` (when registered as `writable`) is bounced back to the host\nover the same transport.\n\n### Outbound request\n\nWhen a session tool resolves a host-owned URL, RPC mode emits:\n\n```json\n{\n  \"type\": \"host_uri_request\",\n  \"id\": \"uri_1\",\n  \"operation\": \"read\",\n  \"url\": \"db://users/42\"\n}\n```\n\nWrites look the same with `\"operation\": \"write\"` and an additional\n`\"content\": \"...\"` field carrying the full replacement bytes.\n\nIf the request is later aborted (caller cancels, session ends), RPC mode\nemits:\n\n```json\n{\n  \"type\": \"host_uri_cancel\",\n  \"id\": \"uri_cancel_1\",\n  \"targetId\": \"uri_1\"\n}\n```\n\n### Inbound result\n\nFor successful reads:\n\n```json\n{\n  \"type\": \"host_uri_result\",\n  \"id\": \"uri_1\",\n  \"content\": \"id=42\\nname=Alice\\n\",\n  \"contentType\": \"text/plain\",\n  \"notes\": [\"fresh from cache\"],\n  \"immutable\": false\n}\n```\n\nFor successful writes, omit content:\n\n```json\n{ \"type\": \"host_uri_result\", \"id\": \"uri_1\" }\n```\n\nTo reject the request, set `isError: true` and either populate `error` with\na message or fall back to `content` for textual error surfacing:\n\n```json\n{\n  \"type\": \"host_uri_result\",\n  \"id\": \"uri_1\",\n  \"isError\": true,\n  \"error\": \"row 42 not found\"\n}\n```\n\n### Constraints\n\n- The agent's `edit` tool does not target host URIs. Hosts that want to\n  mutate virtual files expose `write` and let the model use the `write` tool\n  with replacement content.\n- Schemes are global to the process; `set_host_uri_schemes` replaces the\n  previous set, unregistering anything not in the new list.\n- Schemes are normalized to lowercase before registration.\n\n## Error Model and Recoverability\n\n### Command-level failures\n\nFailures are `success: false` with string `error`.\n\n```json\n{\n  \"id\": \"req_2\",\n  \"type\": \"response\",\n  \"command\": \"set_model\",\n  \"success\": false,\n  \"error\": \"Model not found: provider/model\"\n}\n```\n\n### Recoverability expectations\n\n- Most command failures are recoverable; process remains alive.\n- Malformed JSONL / parse-loop exceptions emit a `parse` error response and continue reading subsequent lines.\n- Empty `set_session_name` is rejected (`Session name cannot be empty`).\n- Extension UI responses with unknown `id` are ignored.\n- Process termination conditions are stdin close or explicit extension-triggered shutdown after the current command.\n\n## Compact Command Flows\n\n### 1) Prompt and stream\n\nstdin:\n\n```json\n{ \"id\": \"req_1\", \"type\": \"prompt\", \"message\": \"Summarize this repo\" }\n```\n\nstdout sequence (typical):\n\n```json\n{ \"id\": \"req_1\", \"type\": \"response\", \"command\": \"prompt\", \"success\": true }\n{ \"type\": \"agent_start\" }\n{ \"type\": \"message_update\", \"assistantMessageEvent\": { \"type\": \"text_delta\", \"delta\": \"...\" }, \"message\": { \"role\": \"assistant\", \"content\": [] } }\n{ \"type\": \"agent_end\", \"messages\": [] }\n```\n\n### 2) Prompt during streaming with explicit queue policy\n\nstdin:\n\n```json\n{\n  \"id\": \"req_2\",\n  \"type\": \"prompt\",\n  \"message\": \"Also include risks\",\n  \"streamingBehavior\": \"followUp\"\n}\n```\n\n### 3) Inspect and tune queue behavior\n\nstdin:\n\n```json\n{ \"id\": \"q1\", \"type\": \"get_state\" }\n{ \"id\": \"q2\", \"type\": \"set_steering_mode\", \"mode\": \"all\" }\n{ \"id\": \"q3\", \"type\": \"set_interrupt_mode\", \"mode\": \"wait\" }\n```\n\n### 4) Extension UI round trip\n\nstdout:\n\n```json\n{\n  \"type\": \"extension_ui_request\",\n  \"id\": \"ui_7\",\n  \"method\": \"input\",\n  \"title\": \"Branch name\",\n  \"placeholder\": \"feature/...\"\n}\n```\n\nstdin:\n\n```json\n{ \"type\": \"extension_ui_response\", \"id\": \"ui_7\", \"value\": \"feature/rpc-host\" }\n```\n\n## Notes on `RpcClient` helper\n\n`src/modes/rpc/rpc-client.ts` is a convenience wrapper, not the protocol definition.\n\nCurrent helper characteristics:\n\n- Spawns `bun <cliPath> --mode rpc`\n- Correlates responses by generated `req_<n>` ids\n- Dispatches recognized core `AgentEvent` types to listeners\n- Supports host-owned custom tools via `setCustomTools()` and automatic handling of `host_tool_call` / `host_tool_cancel`\n- Wraps common protocol commands including OAuth `getLoginProviders()` / `login(...)`; use raw protocol frames for any surface not wrapped by the helper.\n\nUse raw protocol frames if you need complete surface coverage.\n",
+	"rpc.md": "# RPC Protocol Reference\n\nRPC mode runs the coding agent as a newline-delimited JSON protocol over stdio.\n\n- **stdin**: commands (`RpcCommand`), extension UI responses, and host-tool updates/results\n- **stdout**: a ready frame, command responses (`RpcResponse`), session/agent events, extension UI requests, host-tool requests/cancellations\n\nPrimary implementation:\n\n- `src/modes/rpc/rpc-mode.ts`\n- `src/modes/rpc/rpc-types.ts`\n- `src/session/agent-session.ts`\n- `packages/agent/src/agent.ts`\n- `packages/agent/src/agent-loop.ts`\n\n## Startup\n\n```bash\nomp --mode rpc [regular CLI options]\n```\n\nBehavior notes:\n\n- `@file` CLI arguments are rejected in RPC mode.\n- RPC mode disables automatic session title generation by default to avoid an extra model call.\n- RPC mode resets workflow-altering `todo.*`, `task.*`, `memory.backend`/`memories.enabled`, `async.*`, and `bash.autoBackground.*` settings to their built-in defaults instead of inheriting user overrides.\n- The process reads stdin as JSONL (`readJsonl(Bun.stdin.stream())`).\n- At startup it writes `{ \"type\": \"ready\" }` before processing commands.\n- When stdin closes, pending host-tool calls and host-URI requests are rejected and the process exits with code `0`.\n- Responses/events are written as one JSON object per line.\n\n## Transport and Framing\n\nEach frame is a single JSON object followed by `\\n`.\n\nThere is no envelope beyond the object shape itself.\n\n### Outbound frame categories (stdout)\n\n1. Ready frame (`{ type: \"ready\" }`)\n2. `RpcResponse` (`{ type: \"response\", ... }`)\n3. `AgentSessionEvent` objects (`agent_start`, `message_update`, etc.)\n4. `RpcExtensionUIRequest` (`{ type: \"extension_ui_request\", ... }`)\n5. Host tool requests/cancellations (`host_tool_call`, `host_tool_cancel`)\n6. Host URI requests/cancellations (`host_uri_request`, `host_uri_cancel`)\n7. Extension errors (`{ type: \"extension_error\", extensionPath, event, error }`)\n8. Available-commands updates (`{ type: \"available_commands_update\", commands }`), emitted at startup and whenever command metadata changes\n9. Prompt lifecycle hints (`{ type: \"prompt_result\", id?, agentInvoked }`) for scheduled prompts that later resolve without invoking the agent\n10. Subagent frames (`subagent_lifecycle`, `subagent_progress`, `subagent_event`), gated by `set_subagent_subscription`\n11. Builtin slash-command side channels (`command_output`, `session_info_update`, `config_update`)\n\n### Inbound frame categories (stdin)\n\n1. `RpcCommand`\n2. `RpcExtensionUIResponse` (`{ type: \"extension_ui_response\", ... }`)\n3. Host tool updates/results (`host_tool_update`, `host_tool_result`)\n4. Host URI results (`host_uri_result`)\n\n## Request/Response Correlation\n\nAll commands accept optional `id?: string`.\n\n- If provided, normal command responses echo the same `id`.\n- `RpcClient` relies on this for pending-request resolution.\n\nImportant edge behavior from runtime:\n\n- Unknown command responses are emitted with `id: undefined` (even if the request had an `id`).\n- Parse/handler exceptions in the input loop emit `command: \"parse\"` with `id: undefined`.\n- `prompt` and `abort_and_prompt` return immediate success, then may emit a later error response with the **same** id if async prompt scheduling fails.\n- `prompt` success responses may include `data.agentInvoked`. `false` means the prompt completed locally without an agent turn; `true` means the prompt produced agent lifecycle events; omitted means the host must rely on session events for completion.\n- `abort_and_prompt` does not currently emit `data.agentInvoked` or `prompt_result`; hosts should treat it as the legacy abort-then-schedule path and rely on session events or same-id scheduling errors.\n\n## Command Schema (canonical)\n\n`RpcCommand` is defined in `src/modes/rpc/rpc-types.ts`:\n\n### Prompting\n\n- `{ id?, type: \"prompt\", message: string, images?: ImageContent[], streamingBehavior?: \"steer\" | \"followUp\" }`\n- `{ id?, type: \"steer\", message: string, images?: ImageContent[] }`\n- `{ id?, type: \"follow_up\", message: string, images?: ImageContent[] }`\n- `{ id?, type: \"abort\" }`\n- `{ id?, type: \"abort_and_prompt\", message: string, images?: ImageContent[] }`\n- `{ id?, type: \"new_session\", parentSession?: string }`\n\n### State\n\n- `{ id?, type: \"get_state\" }`\n- `{ id?, type: \"get_available_commands\" }`\n- `{ id?, type: \"set_todos\", phases: TodoPhase[] }`\n- `{ id?, type: \"set_host_tools\", tools: RpcHostToolDefinition[] }`\n- `{ id?, type: \"set_host_uri_schemes\", schemes: RpcHostUriSchemeDefinition[] }`\n- `{ id?, type: \"set_subagent_subscription\", level: \"off\" | \"progress\" | \"events\" }`\n- `{ id?, type: \"get_subagents\" }`\n- `{ id?, type: \"get_subagent_messages\", subagentId?: string, sessionFile?: string, fromByte?: number }`\n\n### Model\n\n- `{ id?, type: \"set_model\", provider: string, modelId: string }`\n- `{ id?, type: \"cycle_model\" }`\n- `{ id?, type: \"get_available_models\" }`\n\n### Thinking\n\n- `{ id?, type: \"set_thinking_level\", level: ThinkingLevel }`\n- `{ id?, type: \"cycle_thinking_level\" }`\n\n### Queue modes\n\n- `{ id?, type: \"set_steering_mode\", mode: \"all\" | \"one-at-a-time\" }`\n- `{ id?, type: \"set_follow_up_mode\", mode: \"all\" | \"one-at-a-time\" }`\n- `{ id?, type: \"set_interrupt_mode\", mode: \"immediate\" | \"wait\" }`\n\n### Compaction\n\n- `{ id?, type: \"compact\", customInstructions?: string }`\n- `{ id?, type: \"set_auto_compaction\", enabled: boolean }`\n\n### Retry\n\n- `{ id?, type: \"set_auto_retry\", enabled: boolean }`\n- `{ id?, type: \"abort_retry\" }`\n\n### Bash\n\n- `{ id?, type: \"bash\", command: string }`\n- `{ id?, type: \"abort_bash\" }`\n\n### Session\n\n- `{ id?, type: \"get_session_stats\" }`\n- `{ id?, type: \"export_html\", outputPath?: string }`\n- `{ id?, type: \"switch_session\", sessionPath: string }`\n- `{ id?, type: \"branch\", entryId: string }`\n- `{ id?, type: \"get_branch_messages\" }`\n- `{ id?, type: \"get_last_assistant_text\" }`\n- `{ id?, type: \"set_session_name\", name: string }`\n- `{ id?, type: \"handoff\", customInstructions?: string }`\n\n### Messages\n\n- `{ id?, type: \"get_messages\" }`\n\n### Login\n\n- `{ id?, type: \"get_login_providers\" }`\n- `{ id?, type: \"login\", providerId: string }`\n\n## Response Schema\n\nAll command results use `RpcResponse`:\n\n- Success: `{ id?, type: \"response\", command: <command>, success: true, data?: ... }`\n- Failure: `{ id?, type: \"response\", command: string, success: false, error: string }`\n\nData payloads are command-specific and defined in `rpc-types.ts`.\n\n### `prompt` payload\n\n`prompt` is acknowledged after the command is accepted, not after a model turn finishes:\n\n```json\n{\n  \"id\": \"req_1\",\n  \"type\": \"response\",\n  \"command\": \"prompt\",\n  \"success\": true,\n  \"data\": { \"agentInvoked\": false }\n}\n```\n\n`data.agentInvoked: false` is a completion signal for local-only prompts, including slash commands that produce output without starting an agent turn. `data.agentInvoked: true` means the prompt produced agent lifecycle events; those events can be emitted before or after the prompt response depending on the command path. Older runtimes may omit `data`; hosts should then rely on `agent_end`, custom message completion, or `prompt_result`.\n\n`prompt_result` is emitted when a prompt was accepted immediately but later resolves as local-only:\n\n```json\n{ \"type\": \"prompt_result\", \"id\": \"req_1\", \"agentInvoked\": false }\n```\n\nLocal-only slash commands may emit `command_output` frames before completing via `data.agentInvoked: false` or a later `prompt_result`. They do not emit `agent_end`.\n\n### `get_state` payload\n\n```json\n{\n  \"model\": { \"provider\": \"...\", \"id\": \"...\" },\n  \"thinkingLevel\": \"off|minimal|low|medium|high|xhigh\",\n  \"isStreaming\": false,\n  \"isCompacting\": false,\n  \"steeringMode\": \"all|one-at-a-time\",\n  \"followUpMode\": \"all|one-at-a-time\",\n  \"interruptMode\": \"immediate|wait\",\n  \"sessionFile\": \"...\",\n  \"sessionId\": \"...\",\n  \"sessionName\": \"...\",\n  \"autoCompactionEnabled\": true,\n  \"messageCount\": 0,\n  \"queuedMessageCount\": 0,\n  \"todoPhases\": [\n    {\n      \"id\": \"phase-1\",\n      \"name\": \"Todos\",\n      \"tasks\": [\n        {\n          \"id\": \"task-1\",\n          \"content\": \"Map the tool surface\",\n          \"status\": \"in_progress\"\n        }\n      ]\n    }\n  ],\n  \"systemPrompt\": [\"...\"],\n  \"dumpTools\": [\n    {\n      \"name\": \"read\",\n      \"description\": \"Read files and URLs\",\n      \"parameters\": {}\n    }\n  ],\n  \"contextUsage\": {\n    \"tokens\": 1100,\n    \"contextWindow\": 200000,\n    \"percent\": 0.55\n  }\n}\n```\n\n### `set_todos` payload\n\nReplaces the in-memory todo state for the current session and returns the normalized phase list:\n\n```json\n{\n  \"id\": \"req_2\",\n  \"type\": \"set_todos\",\n  \"phases\": [\n    {\n      \"id\": \"phase-1\",\n      \"name\": \"Evaluation\",\n      \"tasks\": [\n        {\n          \"id\": \"task-1\",\n          \"content\": \"Map the read tool surface\",\n          \"status\": \"in_progress\"\n        },\n        {\n          \"id\": \"task-2\",\n          \"content\": \"Exercise edit operations\",\n          \"status\": \"pending\"\n        }\n      ]\n    }\n  ]\n}\n```\n\nThis is useful for hosts that want to pre-seed a plan before the first prompt.\n\n### `set_host_tools` payload\n\nReplaces the current set of host-owned tools that the RPC server may call back\ninto over stdio:\n\n```json\n{\n  \"id\": \"req_3\",\n  \"type\": \"set_host_tools\",\n  \"tools\": [\n    {\n      \"name\": \"echo_host\",\n      \"label\": \"Echo Host\",\n      \"description\": \"Echo a value from the embedding host\",\n      \"parameters\": {\n        \"type\": \"object\",\n        \"properties\": {\n          \"message\": { \"type\": \"string\" }\n        },\n        \"required\": [\"message\"],\n        \"additionalProperties\": false\n      }\n    }\n  ]\n}\n```\n\nThe response payload is:\n\n```json\n{\n  \"toolNames\": [\"echo_host\"]\n}\n```\n\nThese tools are added to the active session tool registry before the next model\ncall. Re-sending `set_host_tools` replaces the previous host-owned set.\n\n### `set_host_uri_schemes` payload\n\nReplaces the current set of host-owned URL schemes the RPC server should\ndispatch reads/writes through:\n\n```json\n{\n  \"id\": \"req_4\",\n  \"type\": \"set_host_uri_schemes\",\n  \"schemes\": [\n    {\n      \"scheme\": \"db\",\n      \"description\": \"Virtual db row files\",\n      \"writable\": true,\n      \"immutable\": false\n    }\n  ]\n}\n```\n\nThe response payload is:\n\n```json\n{\n  \"schemes\": [\"db\"]\n}\n```\n\nSchemes are case-insensitive on the wire and normalized to lowercase before\nthe response is sent. Re-sending `set_host_uri_schemes` replaces the entire\nprevious set — schemes missing from the new list are unregistered.\n\n## Event Stream Schema\n\nRPC mode forwards `AgentSessionEvent` objects from `AgentSession.subscribe(...)`.\n\nCommon event types:\n\n- `agent_start`, `agent_end`\n- `turn_start`, `turn_end`\n- `message_start`, `message_update`, `message_end`\n- `tool_execution_start`, `tool_execution_update`, `tool_execution_end`\n- `auto_compaction_start`, `auto_compaction_end`\n- `auto_retry_start`, `auto_retry_end`\n- `ttsr_triggered`\n- `todo_reminder`\n- `todo_auto_clear`\n\nExtension runner errors are emitted separately as:\n\n```json\n{\n  \"type\": \"extension_error\",\n  \"extensionPath\": \"...\",\n  \"event\": \"...\",\n  \"error\": \"...\"\n}\n```\n\n`message_update` includes streaming deltas in `assistantMessageEvent` (text/thinking/toolcall deltas).\n\n## Prompt/Queue Concurrency and Ordering\n\nThis is the most important operational behavior.\n\n### Immediate ack vs completion\n\n`prompt` and `abort_and_prompt` are **acknowledged immediately**:\n\n```json\n{ \"id\": \"req_1\", \"type\": \"response\", \"command\": \"prompt\", \"success\": true }\n```\n\nThat means:\n\n- command acceptance != run completion\n- agent turns complete via `agent_end`\n- local-only prompts complete via `data.agentInvoked: false` on the response or via a later `prompt_result`\n\n### While streaming\n\n`AgentSession.prompt()` requires `streamingBehavior` during active streaming:\n\n- `\"steer\"` => queued steering message (interrupt path)\n- `\"followUp\"` => queued follow-up message (post-turn path)\n\nIf omitted during streaming, prompt fails.\n\n### Queue defaults\n\nFrom `packages/agent/src/agent.ts` defaults:\n\n- `steeringMode`: `\"one-at-a-time\"`\n- `followUpMode`: `\"one-at-a-time\"`\n- `interruptMode`: `\"immediate\"`\n\n### Mode semantics\n\n- `set_steering_mode` / `set_follow_up_mode`\n  - `\"one-at-a-time\"`: dequeue one queued message per turn\n  - `\"all\"`: dequeue entire queue at once\n- `set_interrupt_mode`\n  - `\"immediate\"`: tool execution checks steering between tool calls; pending steering can abort remaining tool calls in the turn\n  - `\"wait\"`: defer steering until turn completion\n\n## Extension UI Sub-Protocol\n\nExtensions in RPC mode use request/response UI frames.\n\n### Outbound request\n\n`RpcExtensionUIRequest` (`type: \"extension_ui_request\"`) methods:\n\n- `select`, `confirm`, `input`, `editor`, `cancel`\n- `notify`, `setStatus`, `setWidget`, `setTitle`, `set_editor_text`\n- `open_url` (emitted by RPC login flows)\n\nRuntime note:\n\n- Automatic session title generation is disabled in RPC mode, and `setTitle` UI\n  requests are also suppressed by default because most hosts do not have a\n  meaningful terminal-title surface. Set `PI_RPC_EMIT_TITLE=1` to opt back in to\n  the UI event only.\n\nExample:\n\n```json\n{\n  \"type\": \"extension_ui_request\",\n  \"id\": \"123\",\n  \"method\": \"confirm\",\n  \"title\": \"Confirm\",\n  \"message\": \"Continue?\",\n  \"timeout\": 30000\n}\n```\n\n### Inbound response\n\n`RpcExtensionUIResponse` (`type: \"extension_ui_response\"`):\n\n- `{ type: \"extension_ui_response\", id: string, value: string }`\n- `{ type: \"extension_ui_response\", id: string, confirmed: boolean }`\n- `{ type: \"extension_ui_response\", id: string, cancelled: true, timedOut?: boolean }`\n\nIf a dialog has a timeout, RPC mode resolves to a default value when timeout/abort fires.\n\n## Host Tool Sub-Protocol\n\nRPC hosts can expose custom tools to the agent by sending `set_host_tools`, then\nserving execution requests over the same transport.\n\n### Outbound request\n\nWhen the agent wants the host to execute one of those tools, RPC mode emits:\n\n```json\n{\n  \"type\": \"host_tool_call\",\n  \"id\": \"host_1\",\n  \"toolCallId\": \"toolu_123\",\n  \"toolName\": \"echo_host\",\n  \"arguments\": { \"message\": \"hello\" }\n}\n```\n\nIf the tool execution is later aborted, RPC mode emits:\n\n```json\n{\n  \"type\": \"host_tool_cancel\",\n  \"id\": \"host_cancel_1\",\n  \"targetId\": \"host_1\"\n}\n```\n\n### Inbound updates and completion\n\nHosts can optionally stream progress:\n\n```json\n{\n  \"type\": \"host_tool_update\",\n  \"id\": \"host_1\",\n  \"partialResult\": {\n    \"content\": [{ \"type\": \"text\", \"text\": \"working\" }]\n  }\n}\n```\n\nCompletion uses:\n\n```json\n{\n  \"type\": \"host_tool_result\",\n  \"id\": \"host_1\",\n  \"result\": {\n    \"content\": [{ \"type\": \"text\", \"text\": \"done\" }]\n  }\n}\n```\n\nSet top-level `isError: true` on `host_tool_result` to reject the pending host tool call and surface the returned text content as a tool error.\n\n## Host URI Sub-Protocol\n\nRPC hosts can also own custom URL schemes (virtual files). After\n`set_host_uri_schemes`, every read of `<scheme>://…` and write of\n`<scheme>://…` (when registered as `writable`) is bounced back to the host\nover the same transport.\n\n### Outbound request\n\nWhen a session tool resolves a host-owned URL, RPC mode emits:\n\n```json\n{\n  \"type\": \"host_uri_request\",\n  \"id\": \"uri_1\",\n  \"operation\": \"read\",\n  \"url\": \"db://users/42\"\n}\n```\n\nWrites look the same with `\"operation\": \"write\"` and an additional\n`\"content\": \"...\"` field carrying the full replacement bytes.\n\nIf the request is later aborted (caller cancels, session ends), RPC mode\nemits:\n\n```json\n{\n  \"type\": \"host_uri_cancel\",\n  \"id\": \"uri_cancel_1\",\n  \"targetId\": \"uri_1\"\n}\n```\n\n### Inbound result\n\nFor successful reads:\n\n```json\n{\n  \"type\": \"host_uri_result\",\n  \"id\": \"uri_1\",\n  \"content\": \"id=42\\nname=Alice\\n\",\n  \"contentType\": \"text/plain\",\n  \"notes\": [\"fresh from cache\"],\n  \"immutable\": false\n}\n```\n\nFor successful writes, omit content:\n\n```json\n{ \"type\": \"host_uri_result\", \"id\": \"uri_1\" }\n```\n\nTo reject the request, set `isError: true` and either populate `error` with\na message or fall back to `content` for textual error surfacing:\n\n```json\n{\n  \"type\": \"host_uri_result\",\n  \"id\": \"uri_1\",\n  \"isError\": true,\n  \"error\": \"row 42 not found\"\n}\n```\n\n### Constraints\n\n- The agent's `edit` tool does not target host URIs. Hosts that want to\n  mutate virtual files expose `write` and let the model use the `write` tool\n  with replacement content.\n- Schemes are global to the process; `set_host_uri_schemes` replaces the\n  previous set, unregistering anything not in the new list.\n- Schemes are normalized to lowercase before registration.\n\n## Error Model and Recoverability\n\n### Command-level failures\n\nFailures are `success: false` with string `error`.\n\n```json\n{\n  \"id\": \"req_2\",\n  \"type\": \"response\",\n  \"command\": \"set_model\",\n  \"success\": false,\n  \"error\": \"Model not found: provider/model\"\n}\n```\n\n### Recoverability expectations\n\n- Most command failures are recoverable; process remains alive.\n- Malformed JSONL / parse-loop exceptions emit a `parse` error response and continue reading subsequent lines.\n- Empty `set_session_name` is rejected (`Session name cannot be empty`).\n- Extension UI responses with unknown `id` are ignored.\n- Process termination conditions are stdin close or explicit extension-triggered shutdown after the current command.\n\n## Compact Command Flows\n\n### 1) Prompt and stream\n\nstdin:\n\n```json\n{ \"id\": \"req_1\", \"type\": \"prompt\", \"message\": \"Summarize this repo\" }\n```\n\nstdout sequence (typical):\n\n```json\n{ \"id\": \"req_1\", \"type\": \"response\", \"command\": \"prompt\", \"success\": true }\n{ \"type\": \"agent_start\" }\n{ \"type\": \"message_update\", \"assistantMessageEvent\": { \"type\": \"text_delta\", \"delta\": \"...\" }, \"message\": { \"role\": \"assistant\", \"content\": [] } }\n{ \"type\": \"agent_end\", \"messages\": [] }\n```\n\n### 2) Prompt during streaming with explicit queue policy\n\nstdin:\n\n```json\n{\n  \"id\": \"req_2\",\n  \"type\": \"prompt\",\n  \"message\": \"Also include risks\",\n  \"streamingBehavior\": \"followUp\"\n}\n```\n\n### 3) Inspect and tune queue behavior\n\nstdin:\n\n```json\n{ \"id\": \"q1\", \"type\": \"get_state\" }\n{ \"id\": \"q2\", \"type\": \"set_steering_mode\", \"mode\": \"all\" }\n{ \"id\": \"q3\", \"type\": \"set_interrupt_mode\", \"mode\": \"wait\" }\n```\n\n### 4) Extension UI round trip\n\nstdout:\n\n```json\n{\n  \"type\": \"extension_ui_request\",\n  \"id\": \"ui_7\",\n  \"method\": \"input\",\n  \"title\": \"Branch name\",\n  \"placeholder\": \"feature/...\"\n}\n```\n\nstdin:\n\n```json\n{ \"type\": \"extension_ui_response\", \"id\": \"ui_7\", \"value\": \"feature/rpc-host\" }\n```\n\n## Notes on `RpcClient` helper\n\n`src/modes/rpc/rpc-client.ts` is a convenience wrapper, not the protocol definition.\n\nCurrent helper characteristics:\n\n- Spawns `bun <cliPath> --mode rpc`\n- Correlates responses by generated `req_<n>` ids\n- Dispatches recognized core `AgentEvent` types to listeners\n- Supports host-owned custom tools via `setCustomTools()` and automatic handling of `host_tool_call` / `host_tool_cancel`\n- Wraps common protocol commands including OAuth `getLoginProviders()` / `login(...)`; use raw protocol frames for any surface not wrapped by the helper.\n\nUse raw protocol frames if you need complete surface coverage.\n",
 	"rulebook-matching-pipeline.md": "# Rulebook Matching Pipeline\n\nThis document describes how coding-agent discovers rules from supported config formats, normalizes them into a single `Rule` shape, resolves precedence conflicts, and splits the result into:\n\n- **Rulebook rules** (available to the model via system prompt + `rule://` URLs)\n- **TTSR rules** (time-travel stream interruption rules)\n\nIt reflects the current implementation, including partial semantics and metadata that is parsed but not enforced.\n\n## Implementation files\n\n- [`packages/coding-agent/src/capability/rule.ts`](../packages/coding-agent/src/capability/rule.ts)\n- [`packages/coding-agent/src/capability/rule-buckets.ts`](../packages/coding-agent/src/capability/rule-buckets.ts)\n- [`packages/coding-agent/src/capability/index.ts`](../packages/coding-agent/src/capability/index.ts)\n- [`packages/coding-agent/src/discovery/index.ts`](../packages/coding-agent/src/discovery/index.ts)\n- [`packages/coding-agent/src/discovery/helpers.ts`](../packages/coding-agent/src/discovery/helpers.ts)\n- [`packages/coding-agent/src/discovery/builtin.ts`](../packages/coding-agent/src/discovery/builtin.ts)\n- [`packages/coding-agent/src/discovery/builtin-defaults.ts`](../packages/coding-agent/src/discovery/builtin-defaults.ts)\n- [`packages/coding-agent/src/discovery/agents.ts`](../packages/coding-agent/src/discovery/agents.ts)\n- [`packages/coding-agent/src/discovery/cursor.ts`](../packages/coding-agent/src/discovery/cursor.ts)\n- [`packages/coding-agent/src/discovery/windsurf.ts`](../packages/coding-agent/src/discovery/windsurf.ts)\n- [`packages/coding-agent/src/discovery/cline.ts`](../packages/coding-agent/src/discovery/cline.ts)\n- [`packages/coding-agent/src/sdk.ts`](../packages/coding-agent/src/sdk.ts)\n- [`packages/coding-agent/src/system-prompt.ts`](../packages/coding-agent/src/system-prompt.ts)\n- [`packages/coding-agent/src/internal-urls/rule-protocol.ts`](../packages/coding-agent/src/internal-urls/rule-protocol.ts)\n- [`packages/utils/src/frontmatter.ts`](../packages/utils/src/frontmatter.ts)\n\n## 1. Canonical rule shape\n\nAll providers normalize source files into `Rule`:\n\n```ts\ninterface Rule {\n  name: string;\n  path: string;\n  content: string;\n  globs?: string[];\n  alwaysApply?: boolean;\n  description?: string;\n  condition?: string[];\n  astCondition?: string[];\n  scope?: string[];\n  interruptMode?: \"never\" | \"prose-only\" | \"tool-only\" | \"always\";\n  _source: SourceMeta;\n}\n```\n\nCapability identity is `rule.name` (`ruleCapability.key = rule => rule.name`).\n\nConsequence: precedence and deduplication are **name-based only**. Two different files with the same `name` are considered the same logical rule.\n\n## 2. Discovery sources and normalization\n\n`src/discovery/index.ts` auto-registers providers. For `rules`, current providers are:\n\n- `native` (priority `100`)\n- `omp-plugins` (priority `90`) — `rules/*.{md,mdc}` inside configured extension package roots, normalized via the shared `buildRuleFromMarkdown` path\n- `agents` (priority `70`)\n- `cursor` (priority `50`)\n- `windsurf` (priority `50`)\n- `cline` (priority `40`)\n- `builtin-defaults` (priority `1`)\n\n### Native provider (`builtin.ts`)\n\nLoads `.omp` rules from:\n\n- project: `<cwd>/.omp/rules/*.{md,mdc}` when the cwd `.omp` directory exists\n- user: `~/.omp/agent/rules/*.{md,mdc}`\n- sticky user rule: `~/.omp/agent/RULES.md`\n- sticky project rule: nearest ancestor `.omp/RULES.md` while walking from cwd toward the repository root\n\nNormalization:\n\n- `name` = filename without `.md`/`.mdc`\n- frontmatter parsed via `parseFrontmatter`\n- `content` = body (frontmatter stripped)\n- `globs`, `alwaysApply`, `description`, `condition`/legacy `ttsr_trigger`, `scope`, and `interruptMode` are parsed by `buildRuleFromMarkdown`\n- top-level `RULES.md` is synthesized as rule name `RULES` and forced to `alwaysApply: true`\n\nImportant caveat: `condition` values that look like file globs are converted into `tool:edit(...)` / `tool:write(...)` scope shorthands with catch-all condition `.*`.\n\n### Agents provider (`agents.ts`)\n\nLoads from both `.agent` and `.agents` directories:\n\n- project: walk upward from `cwd` to repo root, loading `<ancestor>/.agent/rules/*.{md,mdc}` and `<ancestor>/.agents/rules/*.{md,mdc}`\n- user: `~/.agent/rules/*.{md,mdc}` and `~/.agents/rules/*.{md,mdc}`\n\nNormalization uses the shared `buildRuleFromMarkdown` path: filename-derived name, stripped frontmatter body, and parsed `globs`, `alwaysApply`, `description`, `condition`/legacy `ttsr_trigger`, `scope`, and `interruptMode`.\n\n### Cursor provider (`cursor.ts`)\n\nLoads from:\n\n- user: `~/.cursor/rules/*.{mdc,md}`\n- project: `<cwd>/.cursor/rules/*.{mdc,md}`\n\nNormalization (`transformMDCRule`):\n\n- `description`: kept only if string\n- `alwaysApply`: normalized to a boolean — `true` only when frontmatter has `alwaysApply: true` (anything else becomes `false`)\n- `globs`: accepts array (string elements only) or single string\n- `condition`/legacy `ttsr_trigger`, `scope`, and `interruptMode` are parsed by shared rule helpers\n- `name` from filename without extension\n\n### Windsurf provider (`windsurf.ts`)\n\nLoads from:\n\n- user: `~/.codeium/windsurf/memories/global_rules.md` (fixed rule name `global_rules`)\n- project: `<cwd>/.windsurf/rules/*.md`\n\nNormalization:\n\n- `globs`: array-of-string or single string\n- `alwaysApply`, `description`, `condition`/legacy `ttsr_trigger`, `scope`, and `interruptMode` parsed by shared rule helpers\n- `name` is fixed to `global_rules` for the user global file and derived from filename for project rules\n\n### Cline provider (`cline.ts`)\n\nSearches upward from `cwd` for nearest `.clinerules`:\n\n- if directory: loads `*.md` inside it\n- if file: loads single file as rule named `clinerules`\n\nNormalization:\n\n- `globs`: array-of-string or single string\n- `alwaysApply`, `description`, `condition`/legacy `ttsr_trigger`, `scope`, and `interruptMode` parsed by shared rule helpers\n- `name` is fixed to `clinerules` for a `.clinerules` file and derived from filename for `.clinerules/*.md`\n\n## 3. Frontmatter parsing behavior and ambiguity\n\nAll providers use `parseFrontmatter` (`utils/frontmatter.ts`) with these semantics:\n\n1. Frontmatter is parsed only when content starts with `---` and has a closing `\\n---`.\n2. Body is trimmed after frontmatter extraction.\n3. If YAML parse fails:\n   - warning is logged,\n   - parser falls back to simple `key: value` line parsing (`^([\\w-]+):\\s*(.*)$`).\n\nAmbiguity consequences:\n\n- Fallback parser does not support arrays, nested objects, or quoting rules.\n- Fallback values become strings (for example `alwaysApply: true` becomes string `\"true\"`), so providers requiring boolean/string types may drop metadata.\n- `ttsr_trigger` works in fallback (underscore key); hyphenated keys like `thinking-level` also parse and are normalized to camelCase (`thinkingLevel`) — key normalization applies to the YAML path too.\n- Files without valid frontmatter still load as rules with empty metadata and full content body.\n\n## 4. Provider precedence and deduplication\n\n`loadCapability(\"rules\")` (`capability/index.ts`) merges provider outputs and then deduplicates by `rule.name`.\n\n### Precedence model\n\n- Providers are ordered by priority descending.\n- Equal priority keeps registration order (`cursor` before `windsurf` from `discovery/index.ts`).\n- Dedup is first-wins: first encountered rule name is kept; later same-name items are marked `_shadowed` in `all` and excluded from `items`.\n\nEffective rule provider order is currently:\n\n1. `native` (100)\n2. `omp-plugins` (90)\n3. `agents` (70)\n4. `cursor` (50)\n5. `windsurf` (50)\n6. `cline` (40)\n7. `builtin-defaults` (1)\n\n### Intra-provider ordering caveat\n\nWithin a provider, item order comes from `loadFilesFromDir` glob result ordering plus explicit push order. This is deterministic enough for normal use but not explicitly sorted in code.\n\nNotable source-order differences:\n\n- `native` appends project `.omp/rules`, user `~/.omp/agent/rules`, user `RULES.md`, then nearest project `RULES.md`.\n- `omp-plugins` appends `rules/` results per configured extension package root.\n- `agents` appends project-walk `.agent`/`.agents` rule dirs before user home dirs.\n- `cursor` appends user then project results.\n- `windsurf` appends user `global_rules` first, then project rules.\n- `cline` loads only nearest `.clinerules` source.\n- `builtin-defaults` uses the embedded rule source order.\n\n## 5. Split into Rulebook, Always-Apply, and TTSR buckets\n\nAfter rule discovery in `createAgentSession` (`sdk.ts`), `bucketRules(...)` applies session-level filtering and bucket assignment:\n\n1. Drop rules listed in `ttsr.disabledRules`.\n2. Drop rules from the `builtin-defaults` provider when `ttsr.builtinRules === false`.\n3. Register rules with a non-empty `condition` or `astCondition` into `TtsrManager`; if registration succeeds, the rule is TTSR-only.\n4. Put remaining `alwaysApply === true` rules into `alwaysApplyRules`.\n5. Put remaining rules with `description` into `rulebookRules`.\n\n### Bucket behavior\n\n- **TTSR bucket**: any enabled rule with a non-empty parsed `condition` (regex) or `astCondition` (ast-grep patterns) that `TtsrManager.addRule(...)` accepts. Takes priority over other buckets.\n- **Always-apply bucket**: `alwaysApply === true`, not TTSR. Full content injected into system prompt. Resolvable via `rule://`.\n- **Rulebook bucket**: must have description, must not be TTSR, must not be `alwaysApply`. Listed in system prompt by name+description; content read on demand via `rule://`.\n- A rule with both a trigger condition and `alwaysApply` goes to TTSR only if TTSR registration accepts it; otherwise it can fall through to always-apply.\n- A rule with both `alwaysApply` and `description` goes to always-apply only (not rulebook).\n\n## 6. How metadata affects runtime surfaces\n\n### `description`\n\n- Required for inclusion in rulebook.\n- Rendered in the system prompt rulebook block (`<domain-rules>` in the default template, `<rules>` in the custom-prompt template).\n- Missing description keeps the rule out of the rulebook listing; unless it is always-apply or an accepted TTSR rule, it is also not addressable via `rule://`.\n\n### `globs`\n\n- Carried through on `Rule`.\n- Rendered inline in the default prompt's rulebook listing (`- <name> (<glob>, ...): <description>`); the custom-prompt template renders them as `<glob>...</glob>` entries.\n- Exposed in rules UI state (`extensions` mode list).\n- Used by TTSR as a global path gate: if a TTSR rule has globs, the match context must include at least one matching file path.\n- Not used to automatically select rulebook rules for `rule://`; rulebook matching remains advisory prompt behavior.\n\n### `alwaysApply`\n\n- Parsed and preserved by providers.\n- Used in UI display (`\"always\"` trigger label in extensions state manager).\n- Used as an exclusion condition from `rulebookRules`.\n- **Full rule content is auto-injected into the system prompt** (before the rulebook rules section).\n- Rule is also addressable via `rule://<name>` for re-reading.\n\n### `condition`, `astCondition`, `scope`, and `interruptMode`\n\n- `condition` is the regex TTSR trigger field; legacy `ttsr_trigger` / `ttsrTrigger` are accepted as fallback inputs during parsing.\n- `astCondition` is the ast-grep trigger field: a string or list of structural patterns, kept verbatim (no glob inference). It only matches on edit/write tool streams, where the language is inferred from the file path. A rule may set `condition`, `astCondition`, or both.\n- `scope` narrows TTSR matching scope. A `condition` token that looks like a file glob becomes `tool:edit(<glob>)` and `tool:write(<glob>)` scope entries plus catch-all condition `.*`; `astCondition` tokens never trigger this shorthand.\n- `interruptMode` can override the global TTSR interrupt mode for the rule.\n\n## 7. System prompt inclusion path\n\n`buildSystemPromptInternal` receives both `rules` (rulebook) and `alwaysApplyRules`.\n\nAlways-apply rules are deduped against custom prompt sources (`dedupeAlwaysApplyRules` drops a rule whose content already appears in the SYSTEM/APPEND_SYSTEM customization) and rendered first, injecting their raw content directly into the prompt (inside a `<generic-rules>` block in the default template).\n\nRulebook rules are rendered in a `<domain-rules>` block as `- <name> (<globs>): <description>` lines; the URL list in the prompt documents `rule://<name>` and the workflow section tells the model to read relevant rules first. The custom-prompt template (`custom-system-prompt.md`) instead renders `<rule name=\"...\">` entries with `<glob>` children under an explicit \"You MUST read `rule://<name>`\" instruction.\n\nThis is advisory/contextual: prompt text asks the model to read applicable rules, but code does not enforce glob applicability.\n\n## 8. `rule://` internal URL behavior\n\n`RuleProtocolHandler` resolves against the process-global active-rule snapshot\ninstalled once per top-level session in `sdk.ts`:\n\n```ts\nsetActiveRules([...rulebookRules, ...alwaysApplyRules, ...ttsrManager.getRules()]);\n```\n\nImplications:\n\n- `rule://<name>` resolves against **rulebookRules**, **alwaysApplyRules**, and **registered TTSR rules**.\n- TTSR rules are bucketed out before rulebook/always, but `ttsrManager.getRules()` re-adds them to the snapshot so a triggered rule (e.g. a builtin) stays addressable for re-reading.\n- Rules with no description, no `alwaysApply`, and no accepted TTSR condition are not addressable via `rule://`.\n- Resolution is exact name match.\n- Unknown names return error listing available rule names.\n- Returned content is raw `rule.content` (frontmatter stripped), content type `text/markdown`.\n\n## 9. Known partial / non-enforced semantics\n\n1. The rule providers currently loaded for `rules` are `native`, `omp-plugins`, `agents`, `cursor`, `windsurf`, `cline`, and embedded `builtin-defaults`; provider files for other tools may parse other config formats but do not register rule loaders.\n2. `globs` metadata is surfaced to prompt/UI and is used as a global path gate for TTSR matching, but it is not used to automatically select rulebook rules for `rule://`.\n3. Rule selection for `rule://` includes rulebook, always-apply, and registered TTSR rules (so a triggered TTSR rule can be re-read), but not rules that registered no condition and carry neither a description nor `alwaysApply`.\n4. Discovery warnings (`loadCapability(\"rules\").warnings`) are produced but `createAgentSession` does not currently surface/log them in this path.\n",
 	"sdk.md": "# SDK\n\nThe SDK is the in-process integration surface for `@oh-my-pi/pi-coding-agent`.\nUse it when you want direct access to agent state, event streaming, tool wiring, and session control from your own Bun/Node process.\n\nIf you need cross-language/process isolation, use RPC mode instead.\n\n## Installation\n\n```bash\nbun add @oh-my-pi/pi-coding-agent\n```\n\n## Entry points\n\n`@oh-my-pi/pi-coding-agent` exports the SDK APIs from the package root (and also via `@oh-my-pi/pi-coding-agent/sdk`).\n\nCore exports for embedders:\n\n- `createAgentSession`\n- `SessionManager`\n- `Settings`\n- `AuthStorage`\n- `ModelRegistry`\n- `discoverAuthStorage`\n- Discovery helpers (`discoverExtensions`, `discoverSkills`, `discoverContextFiles`, `discoverPromptTemplates`, `discoverSlashCommands`, `discoverCustomTSCommands`, `discoverMCPServers`)\n- Tool factory surface (`createTools`, `BUILTIN_TOOLS`, tool classes)\n\n## Quick start (auto-discovery defaults)\n\n```ts\nimport { createAgentSession } from \"@oh-my-pi/pi-coding-agent\";\n\nconst { session, modelFallbackMessage } = await createAgentSession();\n\nif (modelFallbackMessage) {\n  process.stderr.write(`${modelFallbackMessage}\\n`);\n}\n\nconst unsubscribe = session.subscribe((event) => {\n  if (\n    event.type === \"message_update\" &&\n    event.assistantMessageEvent.type === \"text_delta\"\n  ) {\n    process.stdout.write(event.assistantMessageEvent.delta);\n  }\n});\n\nawait session.prompt(\"Summarize this repository in 3 bullets.\");\nunsubscribe();\nawait session.dispose();\n```\n\n## What `createAgentSession()` discovers by default\n\n`createAgentSession()` follows “provide to override, omit to discover”.\n\nIf omitted, it resolves:\n\n- `cwd`: `getProjectDir()`\n- `agentDir`: `~/.omp/agent` (via `getAgentDir()`)\n- `authStorage`: `discoverAuthStorage(agentDir)`\n- `modelRegistry`: `new ModelRegistry(authStorage)` + background `refreshInBackground()` when the registry is not provided\n- `settings`: `await Settings.init({ cwd, agentDir })`\n- `sessionManager`: `SessionManager.create(cwd)` (file-backed)\n- skills/context files/prompt templates/slash commands/extensions/custom TS commands\n- built-in tools via `createTools(...)`\n- MCP tools (enabled by default; Exa MCP servers are folded into native Exa integration, and browser automation MCP servers are filtered when the built-in browser tool is enabled)\n- LSP integration (enabled by default)\n- `eventBus`: new `EventBus()` unless supplied\n\n### Required vs optional inputs\n\nTypically you must provide only what you want to control:\n\n- **Must provide**: nothing for a minimal session\n- **Usually provide explicitly** in embedders:\n  - `sessionManager` (if you need in-memory or custom location)\n  - `authStorage` + `modelRegistry` (if you own credential/model lifecycle)\n  - `model` or `modelPattern` (if deterministic model selection matters)\n  - `settings` (if you need isolated/test config)\n\n## Session manager behavior (persistent vs in-memory)\n\n`AgentSession` always uses a `SessionManager`; behavior depends on which factory you use.\n\n### File-backed (default)\n\n```ts\nimport { createAgentSession, SessionManager } from \"@oh-my-pi/pi-coding-agent\";\n\nconst { session } = await createAgentSession({\n  sessionManager: SessionManager.create(process.cwd()),\n});\n\nconsole.log(session.sessionFile); // absolute .jsonl path\n```\n\n- Persists conversation/messages/state deltas to session files.\n- Supports resume/open/list/fork workflows.\n- `session.sessionFile` is defined.\n\n### In-memory\n\n```ts\nimport { createAgentSession, SessionManager } from \"@oh-my-pi/pi-coding-agent\";\n\nconst { session } = await createAgentSession({\n  sessionManager: SessionManager.inMemory(),\n});\n\nconsole.log(session.sessionFile); // undefined\n```\n\n- No filesystem persistence.\n- Useful for tests, ephemeral workers, request-scoped agents.\n- Session methods still work, but persistence-specific behaviors (file resume/fork paths) are naturally limited.\n\n### Resume/open/list helpers\n\n```ts\nimport { SessionManager } from \"@oh-my-pi/pi-coding-agent\";\n\nconst recent = await SessionManager.continueRecent(process.cwd());\nconst listed = await SessionManager.list(process.cwd());\nconst opened = listed[0] ? await SessionManager.open(listed[0].path) : null;\n```\n\n## Model and auth wiring\n\n`createAgentSession()` uses `ModelRegistry` + `AuthStorage` for model selection and API key resolution.\n\n### Explicit wiring\n\n```ts\nimport {\n  createAgentSession,\n  discoverAuthStorage,\n  ModelRegistry,\n  SessionManager,\n} from \"@oh-my-pi/pi-coding-agent\";\n\nconst authStorage = await discoverAuthStorage();\nconst modelRegistry = new ModelRegistry(authStorage);\nawait modelRegistry.refresh();\n\nconst available = modelRegistry.getAvailable();\nif (available.length === 0)\n  throw new Error(\"No authenticated models available\");\n\nconst { session } = await createAgentSession({\n  authStorage,\n  modelRegistry,\n  model: available[0],\n  thinkingLevel: \"medium\",\n  sessionManager: SessionManager.inMemory(),\n});\n```\n\n### Selection order when `model` is omitted\n\nWhen no explicit `model`/`modelPattern` is provided:\n\n1. restore model from existing session (if restorable + key available)\n2. settings default model role (`default`)\n3. first available model with valid auth\n\nIf restore fails, `modelFallbackMessage` explains fallback.\n\n### Auth priority\n\n`AuthStorage.getApiKey(...)` resolves in this order:\n\n1. runtime override (`setRuntimeApiKey`, used by CLI `--api-key`)\n2. config-sourced API key override (`models.yml` provider `apiKey`)\n3. stored API-key credential in `agent.db` / broker-backed storage\n4. stored OAuth credential, including refresh when needed\n5. provider environment variables\n6. custom-provider resolver fallback\n\n## Event subscription model\n\nSubscribe with `session.subscribe(listener)`; it returns an unsubscribe function.\n\n```ts\nconst unsubscribe = session.subscribe((event) => {\n  switch (event.type) {\n    case \"agent_start\":\n    case \"turn_start\":\n    case \"tool_execution_start\":\n      break;\n    case \"message_update\":\n      if (event.assistantMessageEvent.type === \"text_delta\") {\n        process.stdout.write(event.assistantMessageEvent.delta);\n      }\n      break;\n  }\n});\n```\n\n`AgentSessionEvent` includes core `AgentEvent` plus session-level events:\n\n- `auto_compaction_start` / `auto_compaction_end`\n- `auto_retry_start` / `auto_retry_end`\n- `retry_fallback_applied` / `retry_fallback_succeeded`\n- `ttsr_triggered`\n- `todo_reminder` / `todo_auto_clear`\n- `irc_message`\n\n## Prompt lifecycle\n\n`session.prompt(text, options?)` is the primary entry point.\n\nBehavior:\n\n1. optional command/template expansion (`/` commands, custom commands, file slash commands, prompt templates)\n2. if currently streaming:\n   - requires `streamingBehavior: \"steer\" | \"followUp\"`\n   - queues instead of throwing work away\n3. if idle:\n   - validates model + API key\n   - appends user message\n   - starts agent turn\n\nRelated APIs:\n\n- `sendUserMessage(content, { deliverAs? })`\n- `steer(text, images?)`\n- `followUp(text, images?)`\n- `sendCustomMessage({ customType, content, ... }, { deliverAs?, triggerTurn? })`\n- `abort()`\n\n## Tools and extension integration\n\n### Built-ins and filtering\n\n- Built-ins come from `createTools(...)` and `BUILTIN_TOOLS`.\n- `toolNames` acts as an allowlist for built-ins.\n- `customTools` and extension-registered tools are still included.\n- Hidden tools (for example `yield`) are opt-in unless required by options.\n\n```ts\nconst { session } = await createAgentSession({\n  toolNames: [\"read\", \"search\", \"find\", \"write\"],\n  requireYieldTool: true,\n});\n```\n\n### Extensions\n\n- `extensions`: inline `ExtensionFactory[]`\n- `additionalExtensionPaths`: load extra extension files\n- `disableExtensionDiscovery`: disable automatic extension scanning\n- `preloadedExtensions`: reuse already loaded extension set\n\n### Runtime tool set changes\n\n`AgentSession` supports runtime activation updates:\n\n- `getActiveToolNames()`\n- `getAllToolNames()`\n- `setActiveToolsByName(names)`\n- `refreshMCPTools(mcpTools)`\n\nSystem prompt is rebuilt to reflect active tool changes.\n\n## Discovery helpers\n\nUse these when you want partial control without recreating internal discovery logic:\n\n- `discoverAuthStorage(agentDir?)`\n- `discoverExtensions(cwd?)`\n- `discoverSkills(cwd?, _agentDir?, settings?)`\n- `discoverContextFiles(cwd?, _agentDir?)`\n- `discoverPromptTemplates(cwd?, agentDir?)`\n- `discoverSlashCommands(cwd?)`\n- `discoverCustomTSCommands(cwd?, agentDir?)`\n- `discoverMCPServers(cwd?)`\n- `buildSystemPrompt(options?)`\n\n## Subagent-oriented options\n\nFor SDK consumers building orchestrators (similar to task executor flow):\n\n- `outputSchema`: passes structured output expectation into tool context\n- `requireYieldTool`: forces `yield` tool inclusion\n- `taskDepth`: recursion-depth context for nested task sessions\n- `parentTaskPrefix`: artifact naming prefix for nested task outputs\n\nThese are optional for normal single-agent embedding.\n\n## `createAgentSession()` return value\n\n```ts\ntype CreateAgentSessionResult = {\n  session: AgentSession;\n  extensionsResult: LoadExtensionsResult;\n  setToolUIContext: (uiContext: ExtensionUIContext, hasUI: boolean) => void;\n  mcpManager?: MCPManager;\n  modelFallbackMessage?: string;\n  lspServers?: Array<{\n    name: string;\n    status: \"connecting\" | \"ready\" | \"error\" | \"available\";\n    fileTypes: string[];\n    error?: string;\n  }>;\n  eventBus: EventBus;\n};\n```\n\nUse `setToolUIContext(...)` only if your embedder provides UI capabilities that tools/extensions should call into.\n\n## Startup performance\n\n`createAgentSession()` runs two background optimizations to overlap I/O with the rest of session setup:\n\n- **Model-host preconnect.** As soon as the model is resolved, the SDK fires a best-effort `fetch.preconnect(model.baseUrl)` so DNS + TCP + TLS + HTTP/2 to the provider's host happens in parallel with extension/skill load, tool registry build, and system-prompt assembly. The first real `fetch(...)` then reuses the warm connection, saving 100–300 ms on transcontinental hops (e.g. residential IP → `api.anthropic.com`). Implementation lives in `preconnectModelHost()` in `packages/coding-agent/src/sdk.ts`. If `fetch.preconnect` is unavailable (non-Bun runtime) or the call throws, the optimization is silently skipped — never a hard dependency. Applies to every mode (interactive, print, RPC, ACP).\n- **Conditional LSP warmup.** Startup LSP servers (those returned by `discoverStartupLspServers(cwd)`) are only warmed when **all** of these hold:\n  - `enableLsp !== false` on the session options, **and**\n  - `options.hasUI === true` (interactive TUI), **and**\n  - the `lsp.lazy` setting is disabled (it defaults to `true`).\n\n  With `lsp.lazy` enabled — the default — no language servers are launched at startup at all; each server cold-starts on first use, i.e. when the agent invokes the `lsp` tool or an edit/write touches a file whose extension matches the server's `fileTypes`. Print / script / RPC / ACP invocations (`hasUI=false`) skip the warmup regardless of the setting: they don't render the warmup status indicator and typically finish before the language servers would stabilize, so warming them just spends CPU parsing big `initialize` responses concurrently with the LLM stream consumer and jitters perceived latency. Tools that actually need an LSP server still spin one up on demand through `getOrCreateClient()` — only the _startup_ warmup is skipped. The returned `lspServers` field in `CreateAgentSessionResult` is still populated for UI sessions in lazy mode — recognized servers are discovered (no processes spawned) and reported with status `\"available\"` so the welcome screen and `/status` can list them; it is `undefined` only when `enableLsp === false` or `hasUI === false`.\n\n## Minimal controlled embed example\n\n```ts\nimport {\n  createAgentSession,\n  discoverAuthStorage,\n  ModelRegistry,\n  SessionManager,\n  Settings,\n} from \"@oh-my-pi/pi-coding-agent\";\n\nconst authStorage = await discoverAuthStorage();\nconst modelRegistry = new ModelRegistry(authStorage);\nawait modelRegistry.refresh();\n\nconst settings = Settings.isolated({\n  \"compaction.enabled\": true,\n  \"retry.enabled\": true,\n});\n\nconst { session } = await createAgentSession({\n  authStorage,\n  modelRegistry,\n  settings,\n  sessionManager: SessionManager.inMemory(),\n  toolNames: [\"read\", \"search\", \"find\", \"edit\", \"write\"],\n  enableMCP: false,\n  enableLsp: true,\n});\n\nsession.subscribe((event) => {\n  if (\n    event.type === \"message_update\" &&\n    event.assistantMessageEvent.type === \"text_delta\"\n  ) {\n    process.stdout.write(event.assistantMessageEvent.delta);\n  }\n});\n\nawait session.prompt(\"Find all TODO comments in this repo and propose fixes.\");\nawait session.dispose();\n```\n",
 	"secrets.md": "# Secret Obfuscation\n\nPrevents sensitive values (API keys, tokens, passwords) from being sent to LLM providers. When enabled, secrets are replaced before outbound text content leaves the process. Reversible obfuscation placeholders are restored when session context is rebuilt for display or resume.\n\n## Enabling\n\nDisabled by default. Toggle via `/settings` UI or directly in `config.yml`:\n\n```yaml\nsecrets:\n  enabled: true\n```\n\n## How it works\n\n1. On session startup, secrets are collected from two sources:\n   - **Environment variables** whose names match common secret patterns (`KEY`, `SECRET`, `TOKEN`, `PASSWORD`, `PASS`, `AUTH`, `CREDENTIAL`, `PRIVATE`, `OAUTH`) with values >= 8 characters\n   - **`secrets.yml` files** (see below)\n\n2. Outbound text messages to the LLM have secret values replaced with deterministic placeholders like `#AB12#`.\n\n3. Session context is deep-walked and obfuscation placeholders are restored when building display/resume context. Replace-mode substitutions are one-way and are not restored.\n\nTwo modes control what happens to each secret:\n\n| Mode                  | Behavior                                                | Reversible                                   |\n| --------------------- | ------------------------------------------------------- | -------------------------------------------- |\n| `obfuscate` (default) | Replaced with deterministic placeholder `#[A-Z0-9]{4}#` | Yes (deobfuscated in display/resume context) |\n| `replace`             | Replaced with deterministic same-length string          | No (one-way)                                 |\n\n## secrets.yml\n\nDefine custom secret entries in YAML. Two locations are checked:\n\n| Level   | Path                       | Purpose                     |\n| ------- | -------------------------- | --------------------------- |\n| Global  | `~/.omp/agent/secrets.yml` | Secrets across all projects |\n| Project | `<cwd>/.omp/secrets.yml`   | Project-specific secrets    |\n\nProject entries override global entries with matching `content`.\n\n### Schema\n\nEach entry in the array has these fields:\n\n| Field         | Type                         | Required | Description                                       |\n| ------------- | ---------------------------- | -------- | ------------------------------------------------- |\n| `type`        | `\"plain\"` or `\"regex\"`       | Yes      | Match strategy                                    |\n| `content`     | string                       | Yes      | The secret value (plain) or regex pattern (regex) |\n| `mode`        | `\"obfuscate\"` or `\"replace\"` | No       | Default: `\"obfuscate\"`                            |\n| `replacement` | string                       | No       | Custom replacement (replace mode only)            |\n| `flags`       | string                       | No       | Regex flags (regex type only)                     |\n\n### Examples\n\n#### Plain secrets\n\n```yaml\n# Obfuscate a specific API key (default mode)\n- type: plain\n  content: sk-proj-abc123def456\n\n# Replace a database password with a fixed string\n- type: plain\n  content: hunter2\n  mode: replace\n  replacement: \"********\"\n```\n\n#### Regex secrets\n\n```yaml\n# Obfuscate any AWS-style key\n- type: regex\n  content: \"AKIA[0-9A-Z]{16}\"\n\n# Case-insensitive match with explicit flags\n- type: regex\n  content: \"api[_-]?key\\\\s*=\\\\s*\\\\w+\"\n  flags: \"i\"\n\n# Regex literal syntax (pattern and flags in one string)\n- type: regex\n  content: \"/bearer\\\\s+[a-zA-Z0-9._~+\\\\/=-]+/i\"\n```\n\nRegex entries always scan globally (the `g` flag is enforced automatically). The regex literal syntax `/pattern/flags` is supported as an alternative to separate `content` + `flags` fields. Escaped slashes within the pattern (`\\\\/`) are handled correctly.\n\n#### Replace mode with regex\n\n```yaml\n# One-way replace connection strings (not reversible)\n- type: regex\n  content: \"postgres://[^\\\\s]+\"\n  mode: replace\n  replacement: \"postgres://***\"\n```\n\n## Interaction with env var detection\n\nEnvironment variables are collected first, then file-defined entries are appended. File entries can cover secrets that don't live in env vars (config files, hardcoded values, etc.). If the same plain value appears in both env and file entries, the env entry's obfuscate-mode mapping is used first.\n\n## Key files\n\n- `packages/coding-agent/src/secrets/index.ts` -- loading, merging, env var collection\n- `packages/coding-agent/src/secrets/obfuscator.ts` -- `SecretObfuscator` class, placeholder generation, message obfuscation\n- `packages/coding-agent/src/secrets/regex.ts` -- regex literal parsing and compilation\n- `packages/coding-agent/src/config/settings-schema.ts` -- `secrets.enabled` setting definition\n\n## See also\n\n- [`auth-broker-gateway.md`](./auth-broker-gateway.md) -- remote credential vault and forward-proxy that keep provider OAuth refresh tokens and access tokens off developer hosts entirely (complementary to in-process obfuscation).\n",
@@ -61,7 +61,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"session-switching-and-recent-listing.md": "# Session switching and recent session listing\n\nThis document describes how coding-agent discovers recent sessions, resolves `--resume` targets, presents session pickers, and switches the active runtime session.\n\nIt focuses on current implementation behavior, including fallback paths and caveats.\n\n## Implementation files\n\n- [`../src/session/session-manager.ts`](../packages/coding-agent/src/session/session-manager.ts)\n- [`../src/session/agent-session.ts`](../packages/coding-agent/src/session/agent-session.ts)\n- [`../src/cli/session-picker.ts`](../packages/coding-agent/src/cli/session-picker.ts)\n- [`../src/modes/components/session-selector.ts`](../packages/coding-agent/src/modes/components/session-selector.ts)\n- [`../src/modes/controllers/selector-controller.ts`](../packages/coding-agent/src/modes/controllers/selector-controller.ts)\n- [`../src/main.ts`](../packages/coding-agent/src/main.ts)\n- [`../src/sdk.ts`](../packages/coding-agent/src/sdk.ts)\n- [`../src/modes/interactive-mode.ts`](../packages/coding-agent/src/modes/interactive-mode.ts)\n- [`../src/modes/utils/ui-helpers.ts`](../packages/coding-agent/src/modes/utils/ui-helpers.ts)\n\n## Recent-session discovery\n\n### Directory scope\n\n`SessionManager` stores sessions under a cwd-scoped directory by default:\n\n- `~/.omp/agent/sessions/<dir-encoded>/*.jsonl` (home-relative `-<rel>` names, `-tmp-<rel>` for temp paths, legacy `--<abs>--` otherwise)\n\n`SessionManager.list(cwd, sessionDir?)` reads only that directory unless an explicit `sessionDir` is provided.\n\n### Two listing paths with different payloads\n\nThere are two different listing pipelines:\n\n1. `getRecentSessions(sessionDir, limit)` (welcome/summary view)\n   - Reads only a 4KB prefix (`readTextSlices(..., 4096, 0)[0]`) from each file.\n   - Parses header + earliest user text preview.\n   - Returns lightweight `RecentSessionInfo` with lazy `name` and `timeAgo` getters.\n   - Sorts by file `mtime` descending.\n\n2. `SessionManager.list(...)` / `SessionManager.listAll()` (resume pickers and ID matching)\n   - Reads a 4KB prefix plus a bounded 32 KiB tail in one `readTextSlices(...)` call per file, not the full JSONL file.\n   - Builds `SessionInfo` objects (`id`, `cwd`, `title`, `messageCount`, `firstMessage`, `allMessagesText`, timestamps, lifecycle status).\n   - Uses prefix parsing plus marker counting for list text, and tail parsing for the final-message lifecycle status; later messages beyond the prefix may not be present in `allMessagesText`.\n   - Sorts by `modified` descending.\n\n### Metadata fallback behavior\n\nFor recent summaries (`RecentSessionInfo`):\n\n- display name preference: `header.title` -> first user prompt -> `header.id` -> filename\n- name is truncated to 40 chars for compact displays\n- control characters/newlines are stripped/sanitized from title-derived names\n\nFor `SessionInfo` list entries:\n\n- `title` is `header.title` or the last compaction `shortSummary` seen in the 4KB prefix\n- `firstMessage` is first user message text discoverable from the prefix or `\"(no messages)\"`\n\n## `--continue` resolution and terminal breadcrumb preference\n\n`SessionManager.continueRecent(cwd, sessionDir?)` resolves the target in this order:\n\n1. Read terminal-scoped breadcrumb (`~/.omp/agent/terminal-sessions/<terminal-id>`)\n2. Validate breadcrumb:\n   - current terminal can be identified\n   - referenced file still exists\n3. If the breadcrumb's cwd differs from the current cwd, that cwd no longer exists (moved/renamed dir), and the current directory has no sessions of its own, the breadcrumb session is re-rooted into the current directory (`SessionManager.open` + `moveTo`) instead of starting fresh\n4. Otherwise, if the breadcrumb cwd matches the current cwd (resolved path compare), use the breadcrumb session; else fall back to newest file by mtime in the session dir (`findMostRecentSession`)\n5. If none found, create a new session\n\nTerminal ID derivation prefers TTY path and falls back to env-based identifiers (`TMUX_PANE`, `CMUX_SURFACE_ID`, `KITTY_WINDOW_ID`, `TERM_SESSION_ID`, `WT_SESSION`).\n\nBreadcrumb writes are best-effort and non-fatal.\n\n## Startup-time resume target resolution (`main.ts`)\n\n### `--resume <value>`\n\n`createSessionManager(...)` handles string-valued `--resume` in two modes:\n\n1. Path-like value (contains `/`, `\\\\`, or ends with `.jsonl`)\n   - direct `SessionManager.open(sessionArg, parsed.sessionDir)`\n\n2. Resume key value\n   - `resolveResumableSession(...)` searches local sessions first, then all sessions when `sessionDir` is not forced\n   - matching is case-insensitive and accepts `id` prefix, full JSONL filename prefix, or the session-id suffix after the timestamp\n   - first match in modified-descending order is used (no ambiguity prompt)\n\nCross-project match behavior:\n\n- if the matched session's recorded cwd no longer exists (moved/renamed dir), CLI prompts `Move (re-root) it into the current directory? [Y/n]`; yes opens the session and `moveTo(cwd)` re-roots it (this also applies to local-scope matches whose recorded cwd is gone)\n- otherwise, if a global match's cwd differs from the current cwd, CLI prompts `Fork into current directory? [y/N]`\n- fork accepted -> `SessionManager.forkFrom(...)`\n- either prompt declined -> command cancels (`Resume cancelled: session is in another project.`)\n- non-TTY -> throws `SessionResolutionError` instead of prompting\n\nNo match -> throws error (`Session \"...\" not found.`).\n\n### `--resume` (no value)\n\nHandled after initial session-manager construction:\n\n1. list local sessions with `SessionManager.list(cwd, parsed.sessionDir)`\n2. if empty: preload `SessionManager.listAll()` and open the picker in all-projects scope; print `No sessions found` and exit early only when the global list is also empty\n3. open TUI picker (`selectSession`, with optional preloaded `allSessions`/`startInAllScope`)\n4. if canceled: print `No session selected` and exit early\n5. if selected: when the session belongs to another project, switch the process into that project's directory (`setProjectDir`, cache resets, settings reload) first; then `SessionManager.open(selected.path)`\n\n### `--continue`\n\nUses `SessionManager.continueRecent(...)` directly (breadcrumb-first behavior above).\n\n## Picker-based selection internals\n\n## CLI picker (`src/cli/session-picker.ts`)\n\n`selectSession(sessions, { allSessions?, startInAllScope? })` creates a standalone TUI with `SessionSelectorComponent` and resolves exactly once:\n\n- selection -> resolves selected `SessionInfo` (caller uses `.path` / `.cwd`)\n- cancel (Esc) -> resolves `null`\n- hard exit (Ctrl+C path) -> stops TUI and `process.exit(0)`\n- Tab toggles current-folder / all-projects scope; the all-projects list is loaded lazily via `SessionManager.listAll` (or preloaded via `allSessions`)\n- search ranking is augmented with prompt-history matches from `history.db` (`HistoryStorage.matchingSessionIds`) when available\n\n## Interactive in-session picker (`SelectorController.showSessionSelector`)\n\nFlow:\n\n1. fetch sessions from current session dir via `SessionManager.list(currentCwd, currentSessionDir)`; if empty, preload `SessionManager.listAll()` and open in all-projects scope\n2. mount `SessionSelectorComponent` in editor area using `showSelector(...)`, wired with `loadAllSessions: () => SessionManager.listAll()` and a `history.db` prompt matcher\n3. callbacks:\n   - select -> close selector and call `handleResumeSession(sessionPath)`\n   - cancel -> restore editor and rerender\n   - exit -> `ctx.shutdown()`\n\n## Session selector component behavior\n\n`SessionList` supports:\n\n- arrow/page navigation\n- Enter to select\n- Delete to delete after confirmation\n- Esc to cancel\n- Ctrl+C to exit\n- Tab to toggle current-folder / all-projects scope\n- ranked fuzzy search across session id/title/cwd/first message/all messages/path, merged with prompt-history matches from `history.db`\n\nEmpty-list render behavior:\n\n- current-folder scope renders `No sessions in current folder. Press Tab to view all.`; all-projects scope renders `No sessions found`\n- Enter/Delete on empty do nothing (no callback)\n- Esc/Ctrl+C still work\n\n## Runtime switch execution (`AgentSession.switchSession`)\n\n`switchSession(sessionPath)` is the core in-process switch path.\n\nLifecycle/state transition:\n\n1. capture `previousSessionFile`\n2. emit `session_before_switch` hook event (`reason: \"resume\"`, cancellable)\n3. if canceled -> return `false` with no switch\n4. disconnect from current agent event stream\n5. abort active generation/tool flow\n6. flush session writer (`sessionManager.flush()`) to persist pending writes, then capture rollback state\n7. clear queued steering/follow-up/next-turn message buffers\n8. `sessionManager.setSessionFile(sessionPath)`\n   - updates session file pointer\n   - writes terminal breadcrumb\n   - loads entries / migrates / blob-resolves / reindexes\n   - if missing/invalid file data: initializes a new session at that path and rewrites header\n9. update `agent.sessionId`\n10. rebuild display context via `buildDisplaySessionContext()`\n11. restore persisted/discovered MCP tool selections and rebuild active tools/system prompt when discovery is enabled\n12. emit `session_switch` hook event (`reason: \"resume\"`, `previousSessionFile`)\n13. replace agent messages with rebuilt context and sync todos\n14. close provider sessions when switching to a different session or when same-session reload changed replay messages\n15. restore model via `getRestorableSessionModels(sessionContext.models, lastModelChangeRole)` — tries the recorded models in fallback order and uses the first one present in the model registry\n16. restore thinking level and service tier:\n    - thinking uses persisted `thinking_level_change`, otherwise the configured default clamped to model capability\n    - service tier uses persisted `service_tier_change`, otherwise the configured `serviceTier` setting (`\"none\"` becomes unset)\n17. reconnect agent listeners, run the registered session-switch reconciler if any (interactive mode re-enters persisted modes; errors logged, not fatal), and return `true`\n\n## UI state rebuild after interactive switch\n\n`SelectorController.handleResumeSession` performs UI reset around `switchSession`:\n\n- stop loading animation\n- clear status container\n- clear pending-message UI and pending tool map\n- reset streaming component/message references\n- call `session.switchSession(...)`\n- if the resumed session's cwd differs from the previous one, re-point the process and cwd-derived caches at it (`applyCwdChange`)\n- clear chat container and rerender from session context (`renderInitialMessages`)\n- reload todos from new session artifacts\n- show `Resumed session` (or `Resumed session in <dir>` for a cross-project resume)\n\nSo visible conversation/todo state is rebuilt from the new session file.\n\n## Startup resume vs in-session switch\n\n### Startup resume (`--continue`, `--resume`, direct open)\n\n- Session file is chosen before `createAgentSession(...)`.\n- `sdk.ts` builds `existingSession = sessionManager.buildSessionContext()`.\n- Agent messages are restored once during session creation.\n- Model/thinking are selected during creation (including restore/fallback logic).\n- Interactive mode then runs `#reconcileModeFromSession()` to re-enter persisted mode state (e.g. plan mode).\n\n### In-session switch (`/resume`-style selector path)\n\n- Uses `AgentSession.switchSession(...)` on an already-running `AgentSession`.\n- Messages/model/thinking are rebuilt immediately in place.\n- Hook `session_before_switch`/`session_switch` events are emitted.\n- UI chat/todos are refreshed.\n- Mode re-entry is symmetric with startup: interactive mode registers `#reconcileModeFromSession()` as the session-switch reconciler (`setSessionSwitchReconciler`), and `switchSession()` invokes it after reconnecting.\n\n## Failure and edge-case behavior\n\n### Cancellation paths\n\n- CLI picker cancel -> returns `null`, caller prints `No session selected`, process exits early.\n- Interactive picker cancel -> editor restored, no session change.\n- Hook cancellation (`session_before_switch`) -> `switchSession()` returns `false`.\n\n### Empty list paths\n\n- CLI `--resume` (no value): empty list prints `No sessions found` and exits.\n- Interactive selector: empty list renders message and remains cancellable.\n\n### Missing/invalid target session file\n\nWhen opening/switching to a specific path (`setSessionFile`):\n\n- ENOENT -> treated as empty -> new session initialized at that exact path and persisted.\n- malformed/invalid header (or effectively unreadable parsed entries) -> treated as empty -> new session initialized and persisted.\n\nThis is recovery behavior, not hard failure.\n\n### Hard failures\n\nSwitch/open can still throw on true I/O failures (permission errors, rewrite failures, etc.), which propagate to callers.\n\n### ID prefix matching caveats\n\n- Matching uses `startsWith` on the lowercased session id, lowercased JSONL filename, and lowercased id suffix after the filename timestamp.\n- First match in modified-descending order wins; there is no ambiguity UI if multiple sessions share a prefix.\n- Prefix-listing metadata is intentionally lightweight, so search text may not include messages outside the first 4KB of the session file.\n",
 	"session-tree-plan.md": "# Session tree architecture (current)\n\nReference: [session.md](../docs/session.md)\n\nThis document describes how session tree navigation works today: in-memory tree model, leaf movement rules, branching behavior, and extension/event integration.\n\n## What this subsystem is\n\nThe session is stored as an append-only entry log, but runtime behavior is tree-based:\n\n- Every non-header entry has `id` and `parentId`.\n- The active position is `leafId` in `SessionManager`.\n- Appending an entry always creates a child of the current leaf.\n- Branching does **not** rewrite history; it only changes where the leaf points before the next append.\n\nKey files:\n\n- `src/session/session-manager.ts` — tree data model, traversal, leaf movement, branch/session extraction\n- `src/session/agent-session.ts` — `/tree` navigation flow, summarization, hook/event emission\n- `src/modes/components/tree-selector.ts` — interactive tree UI behavior and filtering\n- `src/modes/controllers/selector-controller.ts` — selector orchestration for `/tree` and `/branch`\n- `src/slash-commands/builtin-registry.ts` — command routing (`/tree`, `/branch`)\n- `src/modes/controllers/input-controller.ts` — double-escape behavior and `app.session.tree`/`app.session.fork` keybinding wiring\n- `src/session/messages.ts` — conversion of `branch_summary`, `compaction`, and `custom_message` entries into LLM context messages\n\n## Tree data model in `SessionManager`\n\nRuntime indices:\n\n- `#byId: Map<string, SessionEntry>` — fast lookup for any entry\n- `#leafId: string | null` — current position in the tree\n- `#labelsById: Map<string, string>` — resolved labels by target entry id\n\nTree APIs:\n\n- `getBranch(fromId?)` walks parent links to root and returns root→node path\n- `getTree()` returns `SessionTreeNode[]` (`entry`, `children`, `label`)\n  - parent links become children arrays\n  - entries with missing parents are treated as roots\n  - children are sorted oldest→newest by timestamp\n- `getChildren(parentId)` returns direct children\n- `getLabel(id)` resolves current label from `labelsById`\n\n`getTree()` is a runtime projection; persistence remains append-only JSONL entries.\n\n## Leaf movement semantics\n\nThere are three leaf movement primitives:\n\n1. `branch(entryId)`\n   - Validates entry exists\n   - Sets `leafId = entryId`\n   - No new entry is written\n\n2. `resetLeaf()`\n   - Sets `leafId = null`\n   - Next append creates a new root entry (`parentId = null`)\n\n3. `branchWithSummary(branchFromId, summary, details?, fromExtension?)`\n   - Accepts `branchFromId: string | null`\n   - Sets `leafId = branchFromId`\n   - Appends a `branch_summary` entry as child of that leaf\n   - When `branchFromId` is `null`, `fromId` is persisted as `\"root\"`\n\n## `/tree` navigation behavior (same session file)\n\n`AgentSession.navigateTree()` is navigation, not file forking.\n\nFlow:\n\n1. Validate target and compute abandoned path (`collectEntriesForBranchSummary`)\n2. Emit `session_before_tree` with `TreePreparation`\n3. Optionally summarize abandoned entries (hook-provided summary or built-in summarizer)\n4. Compute new leaf target:\n   - selecting a **user** message: leaf moves to its parent, and message text is returned for editor prefill\n   - selecting a **custom_message**: same rule as user message (leaf = parent, text prefills editor)\n   - selecting any other entry: leaf = selected entry id\n5. Apply leaf move:\n   - with summary: `branchWithSummary(newLeafId, ...)`\n   - without summary and `newLeafId === null`: `resetLeaf()`\n   - otherwise: `branch(newLeafId)`\n6. Rebuild agent context from new leaf and emit `session_tree`\n\nImportant: summary entries are attached at the **new navigation position**, not on the abandoned branch tail.\n\n## `/branch` behavior (new session file)\n\n`/branch` and `/tree` are intentionally different:\n\n- `/tree` navigates within the current session file.\n- `/branch` creates a new session branch file (or in-memory replacement for non-persistent mode).\n\nUser-facing `/branch` flow (`SelectorController.showUserMessageSelector` → `AgentSession.branch`):\n\n- Branch source must be a **user message**.\n- Selected user text is extracted for editor prefill.\n- If selected user message is root (`parentId === null`): start a new session via `newSession({ parentSession: previousSessionFile })`.\n- Otherwise: `createBranchedSession(selectedEntry.parentId)` to fork history up to the selected prompt boundary.\n\n`SessionManager.createBranchedSession(leafId)` specifics:\n\n- Builds root→leaf path via `getBranch(leafId)`; throws if missing.\n- Excludes existing `label` entries from copied path.\n- Rebuilds fresh label entries from resolved `labelsById` for entries that remain in path.\n- Persistent mode: writes new JSONL file and switches manager to it; returns new file path.\n- In-memory mode: replaces in-memory entries; returns `undefined`.\n\n## Context reconstruction and summary/custom integration\n\n`buildSessionContext()` (in `session-manager.ts`) resolves the active root→leaf path and builds effective LLM context state:\n\n- Tracks latest thinking/model/service-tier/mode/TTSR/MCP-selection state on path.\n- Handles latest compaction on path:\n  - emits compaction summary first\n  - replays kept messages from `firstKeptEntryId` to compaction point\n  - then replays post-compaction messages\n- Includes `branch_summary` and `custom_message` entries as `AgentMessage` objects.\n\n`session/messages.ts` then maps these message types for model input:\n\n- `branchSummary` and `compactionSummary` become user-role templated context messages\n- `custom`/`hookMessage` become developer-role content messages (via agent-core's `convertMessageToLlm`)\n\nSo tree movement changes context by changing the active leaf path, not by mutating old entries.\n\n## Labels and tree UI behavior\n\nLabel persistence:\n\n- `appendLabelChange(targetId, label?)` writes `label` entries on the current leaf chain.\n- `labelsById` is updated immediately (set or delete).\n- `getTree()` resolves current label onto each returned node.\n\nTree selector behavior (`tree-selector.ts`):\n\n- Flattens tree for navigation, keeps active-path highlighting, and prioritizes displaying the active branch first.\n- Supports filter modes: `default`, `no-tools`, `user-only`, `labeled-only`, `all`.\n  - `default` suppresses `label`, `custom`, `model_change`, and `thinking_level_change`; it is not a complete \"hide all internal entries\" filter.\n- Supports free-text search over rendered semantic content.\n- `Shift+L` opens inline label editing and writes via `appendLabelChange`.\n\nCommand routing:\n\n- `/tree` always opens tree selector.\n- `/branch` opens user-message selector unless `doubleEscapeAction=tree`, in which case it also uses tree selector UX.\n\n## Extension and hook touchpoints for tree operations\n\nCommand-time extension API (`ExtensionCommandContext`):\n\n- `branch(entryId)` — create branched session file\n- `navigateTree(targetId, { summarize? })` — move within current tree/file\n\nEvents around tree navigation:\n\n- `session_before_tree`\n  - receives `TreePreparation`:\n    - `targetId`\n    - `oldLeafId`\n    - `commonAncestorId`\n    - `entriesToSummarize`\n    - `userWantsSummary`\n  - may cancel navigation\n  - may provide summary payload used instead of built-in summarizer\n  - receives abort `signal` (Escape cancellation path)\n- `session_tree`\n  - emits `newLeafId`, `oldLeafId`\n  - includes `summaryEntry` when a summary was created\n  - `fromExtension` indicates summary origin\n\nAdjacent but related lifecycle hooks:\n\n- `session_before_branch` / `session_branch` for `/branch` flow\n- `session_before_compact`, `session.compacting`, `session_compact` for compaction entries that later affect tree-context reconstruction\n\n## Real constraints and edge conditions\n\n- `branch()` cannot target `null`; use `resetLeaf()` for root-before-first-entry state.\n- `branchWithSummary()` supports `null` target and records `fromId: \"root\"`.\n- Selecting current leaf in tree selector is a no-op.\n- Summarization requires an active model; if absent, summarize navigation fails fast.\n- If summarization is aborted, navigation is cancelled and leaf is unchanged.\n- In-memory sessions never return a branch file path from `createBranchedSession`.\n- Tree context reconstruction includes service-tier and MCP tool-selection state, but those entries do not become LLM messages.\n\n## Plan approval session naming\n\nWhen a user approves a plan from plan mode (`InteractiveMode.#approvePlan`), the approval handler seeds the session name from the plan's title so the resulting (fresh or compacted) session does not stay unnamed.\n\nTrigger:\n\n- Plan approval reaches `#approvePlan(...)` with `options.title` populated from the plan-approval details.\n- This runs for every approval choice (`Approve and execute`, `Approve and compact context`, plain `Approve`); the synthetic `plan-approved` prompt is what otherwise bypasses the input-controller's title-generation path.\n\nNaming source:\n\n- The normalized plan title is humanized via `humanizePlanTitle(title)` (`packages/coding-agent/src/plan-mode/approved-plan.ts`):\n  - replaces runs of `-`/`_` with a single space\n  - trims whitespace\n  - capitalizes the first character\n  - returns `\"\"` for whitespace-only / separator-only input\n- The humanized name is applied only when the current session has no name (`!sessionManager.getSessionName()`). It then calls `sessionManager.setSessionName(name, \"auto\")`, which also refuses to overwrite user-named sessions.\n- On successful apply, the terminal title (`setSessionTerminalTitle`) and the editor border color are refreshed to reflect the new name.\n\nExamples (from `humanizePlanTitle`):\n\n- `migrate-mcp-loader` → `Migrate mcp loader`\n- `fix_session_naming` → `Fix session naming`\n- `foo--bar__baz` → `Foo bar baz`\n- `RefactorRouter` → `RefactorRouter` (no separators to expand)\n- `\"\"` / `\"---\"` → `\"\"` (no name applied)\n\n## Legacy compatibility still present\n\nSession migrations still run on load:\n\n- v1→v2 adds `id`/`parentId` and converts compaction index anchor to id anchor\n- v2→v3 migrates legacy `hookMessage` role to `custom`\n\nCurrent runtime behavior is version-3 tree semantics after migration.\n",
 	"session.md": "# Session Storage and Entry Model\n\nThis document is the source of truth for how coding-agent sessions are represented, persisted, migrated, and reconstructed at runtime.\n\n## Scope\n\nCovers:\n\n- Session JSONL format and versioning\n- Entry taxonomy and tree semantics (`id`/`parentId` + leaf pointer)\n- Migration/compatibility behavior when loading old or malformed files\n- Context reconstruction (`buildSessionContext`)\n- Persistence guarantees, failure behavior, truncation/blob externalization\n- Storage abstractions (`FileSessionStorage`, `MemorySessionStorage`) and related utilities\n\nDoes not cover `/tree` UI rendering behavior beyond semantics that affect session data.\n\n## Implementation Files\n\n- [`src/session/session-manager.ts`](../packages/coding-agent/src/session/session-manager.ts)\n- [`src/session/messages.ts`](../packages/coding-agent/src/session/messages.ts)\n- [`src/session/session-storage.ts`](../packages/coding-agent/src/session/session-storage.ts)\n- [`src/session/history-storage.ts`](../packages/coding-agent/src/session/history-storage.ts)\n- [`src/session/blob-store.ts`](../packages/coding-agent/src/session/blob-store.ts)\n\n## On-Disk Layout\n\nDefault session file location:\n\n```text\n~/.omp/agent/sessions/<dir-encoded>/<timestamp>_<sessionId>.jsonl\n```\n\n`<dir-encoded>` depends on where the canonicalized cwd lives:\n\n- inside the home directory: `-<relative-path>` with `/`, `\\\\`, and `:` replaced by `-` (bare `-` for home itself)\n- inside the OS temp root: `-tmp-<relative-path>` with the same replacement\n- anywhere else: legacy absolute form `--<cwd-without-leading-slash-with-same-replacement>--`\n\nOld `--<home-encoded>-*--` directories are migrated to the new home-relative names once per sessions root on first access (best-effort).\n\nBlob store location:\n\n```text\n~/.omp/agent/blobs/<sha256>\n```\n\nTerminal breadcrumb files are written under:\n\n```text\n~/.omp/agent/terminal-sessions/<terminal-id>\n```\n\nBreadcrumb content is two lines: original cwd, then session file path. `continueRecent()` prefers this terminal-scoped pointer before scanning most-recent mtime.\n\n## File Format\n\nSession files are JSONL: one JSON object per line.\n\n- Line 1 is always the session header (`type: \"session\"`).\n- Remaining lines are `SessionEntry` values.\n- Entries are append-only at runtime; branch navigation moves a pointer (`leafId`) rather than mutating existing entries.\n\n### Header (`SessionHeader`)\n\n```json\n{\n  \"type\": \"session\",\n  \"version\": 3,\n  \"id\": \"1f9d2a6b9c0d1234\",\n  \"timestamp\": \"2026-02-16T10:20:30.000Z\",\n  \"cwd\": \"/work/pi\",\n  \"title\": \"optional session title\",\n  \"titleSource\": \"auto\",\n  \"parentSession\": \"optional lineage marker\"\n}\n```\n\nNotes:\n\n- `version` is optional in v1 files; absence means v1.\n- `parentSession` is an opaque lineage string. Current code writes either a session id or a session path depending on flow (`fork`, `forkFrom`, `createBranchedSession`, or explicit `newSession({ parentSession })`). Treat as metadata, not a typed foreign key.\n\n### Entry Base (`SessionEntryBase`)\n\nAll non-header entries include:\n\n```json\n{\n  \"type\": \"...\",\n  \"id\": \"8-char-id\",\n  \"parentId\": \"previous-or-branch-parent\",\n  \"timestamp\": \"2026-02-16T10:20:30.000Z\"\n}\n```\n\n`parentId` can be `null` for a root entry (first append, or after `resetLeaf()`).\n\n## Entry Taxonomy\n\n`SessionEntry` is the union of:\n\n- `message`\n- `thinking_level_change`\n- `model_change`\n- `service_tier_change`\n- `compaction`\n- `branch_summary`\n- `custom`\n- `custom_message`\n- `label`\n- `ttsr_injection`\n- `session_init`\n- `mode_change`\n- `mcp_tool_selection`\n\n### `message`\n\nStores an `AgentMessage` directly.\n\n```json\n{\n  \"type\": \"message\",\n  \"id\": \"a1b2c3d4\",\n  \"parentId\": null,\n  \"timestamp\": \"2026-02-16T10:21:00.000Z\",\n  \"message\": {\n    \"role\": \"assistant\",\n    \"provider\": \"anthropic\",\n    \"model\": \"claude-sonnet-4-5\",\n    \"content\": [{ \"type\": \"text\", \"text\": \"Done.\" }],\n    \"usage\": {\n      \"input\": 100,\n      \"output\": 20,\n      \"cacheRead\": 0,\n      \"cacheWrite\": 0,\n      \"cost\": {\n        \"input\": 0,\n        \"output\": 0,\n        \"cacheRead\": 0,\n        \"cacheWrite\": 0,\n        \"total\": 0\n      }\n    },\n    \"timestamp\": 1760000000000\n  }\n}\n```\n\n### `model_change`\n\n```json\n{\n  \"type\": \"model_change\",\n  \"id\": \"b1c2d3e4\",\n  \"parentId\": \"a1b2c3d4\",\n  \"timestamp\": \"2026-02-16T10:21:30.000Z\",\n  \"model\": \"openai/gpt-4o\",\n  \"role\": \"default\"\n}\n```\n\n`role` is optional; missing is treated as `default` in context reconstruction.\n\n### `service_tier_change`\n\n```json\n{\n  \"type\": \"service_tier_change\",\n  \"id\": \"c1d2e3f4\",\n  \"parentId\": \"b1c2d3e4\",\n  \"timestamp\": \"2026-02-16T10:21:45.000Z\",\n  \"serviceTier\": \"flex\"\n}\n```\n\n`serviceTier` can also be `null`.\n\n### `thinking_level_change`\n\n```json\n{\n  \"type\": \"thinking_level_change\",\n  \"id\": \"c1d2e3f4\",\n  \"parentId\": \"b1c2d3e4\",\n  \"timestamp\": \"2026-02-16T10:22:00.000Z\",\n  \"thinkingLevel\": \"high\"\n}\n```\n\n### `compaction`\n\n```json\n{\n  \"type\": \"compaction\",\n  \"id\": \"d1e2f3a4\",\n  \"parentId\": \"c1d2e3f4\",\n  \"timestamp\": \"2026-02-16T10:23:00.000Z\",\n  \"summary\": \"Conversation summary\",\n  \"shortSummary\": \"Short recap\",\n  \"firstKeptEntryId\": \"a1b2c3d4\",\n  \"tokensBefore\": 42000,\n  \"details\": { \"readFiles\": [\"src/a.ts\"] },\n  \"preserveData\": { \"hookState\": true },\n  \"fromExtension\": false\n}\n```\n\n### `branch_summary`\n\n```json\n{\n  \"type\": \"branch_summary\",\n  \"id\": \"e1f2a3b4\",\n  \"parentId\": \"a1b2c3d4\",\n  \"timestamp\": \"2026-02-16T10:24:00.000Z\",\n  \"fromId\": \"a1b2c3d4\",\n  \"summary\": \"Summary of abandoned path\",\n  \"details\": { \"note\": \"optional\" },\n  \"fromExtension\": true\n}\n```\n\nIf branching from root (`branchFromId === null`), `fromId` is the literal string `\"root\"`.\n\n### `custom`\n\nExtension state persistence; ignored by `buildSessionContext`.\n\n```json\n{\n  \"type\": \"custom\",\n  \"id\": \"f1a2b3c4\",\n  \"parentId\": \"e1f2a3b4\",\n  \"timestamp\": \"2026-02-16T10:25:00.000Z\",\n  \"customType\": \"my-extension\",\n  \"data\": { \"state\": 1 }\n}\n```\n\n### `custom_message`\n\nExtension-provided message that does participate in LLM context. `content` can be a string or text/image content blocks, and `attribution` records whether the user or agent initiated it.\n\n```json\n{\n  \"type\": \"custom_message\",\n  \"id\": \"a2b3c4d5\",\n  \"parentId\": \"f1a2b3c4\",\n  \"timestamp\": \"2026-02-16T10:26:00.000Z\",\n  \"customType\": \"my-extension\",\n  \"content\": \"Injected context\",\n  \"display\": true,\n  \"details\": { \"debug\": false },\n  \"attribution\": \"agent\"\n}\n```\n\n### `label`\n\n```json\n{\n  \"type\": \"label\",\n  \"id\": \"b2c3d4e5\",\n  \"parentId\": \"a2b3c4d5\",\n  \"timestamp\": \"2026-02-16T10:27:00.000Z\",\n  \"targetId\": \"a1b2c3d4\",\n  \"label\": \"checkpoint\"\n}\n```\n\n`label: undefined` clears a label for `targetId`.\n\n### `ttsr_injection`\n\n```json\n{\n  \"type\": \"ttsr_injection\",\n  \"id\": \"c2d3e4f5\",\n  \"parentId\": \"b2c3d4e5\",\n  \"timestamp\": \"2026-02-16T10:28:00.000Z\",\n  \"injectedRules\": [\"ruleA\", \"ruleB\"]\n}\n```\n\n### `mcp_tool_selection`\n\n```json\n{\n  \"type\": \"mcp_tool_selection\",\n  \"id\": \"d2e3f4a5\",\n  \"parentId\": \"c2d3e4f5\",\n  \"timestamp\": \"2026-02-16T10:28:30.000Z\",\n  \"selectedToolNames\": [\"server.tool\"]\n}\n```\n\n### `session_init`\n\n```json\n{\n  \"type\": \"session_init\",\n  \"id\": \"d2e3f4a5\",\n  \"parentId\": \"c2d3e4f5\",\n  \"timestamp\": \"2026-02-16T10:29:00.000Z\",\n  \"systemPrompt\": \"...\",\n  \"task\": \"...\",\n  \"tools\": [\"read\", \"edit\"],\n  \"outputSchema\": { \"type\": \"object\" }\n}\n```\n\n### `mode_change`\n\n```json\n{\n  \"type\": \"mode_change\",\n  \"id\": \"e2f3a4b5\",\n  \"parentId\": \"d2e3f4a5\",\n  \"timestamp\": \"2026-02-16T10:30:00.000Z\",\n  \"mode\": \"plan\",\n  \"data\": { \"planFile\": \"/tmp/plan.md\" }\n}\n```\n\n## Versioning and Migration\n\nCurrent session version: `3`.\n\n### v1 -> v2\n\nApplied when header `version` is missing or `< 2`:\n\n- Adds `id` and `parentId` to each non-header entry.\n- Reconstructs a linear parent chain using file order.\n- Migrates compaction field `firstKeptEntryIndex` -> `firstKeptEntryId` when present.\n- Sets header `version = 2`.\n\n### v2 -> v3\n\nApplied when header `version < 3`:\n\n- For `message` entries: rewrites legacy `message.role === \"hookMessage\"` to `\"custom\"`.\n- Sets header `version = 3`.\n\n### Migration Trigger and Persistence\n\n- Migrations run during session load (`setSessionFile`).\n- If any migration ran, the entire file is rewritten to disk immediately.\n- Migration mutates in-memory entries first, then persists rewritten JSONL.\n\n## Load and Compatibility Behavior\n\n`loadEntriesFromFile(path)` behavior:\n\n- Missing file (`ENOENT`) -> returns `[]`.\n- Non-parseable lines are handled by lenient JSONL parser (`parseJsonlLenient`).\n- If first parsed entry is not a valid session header (`type !== \"session\"` or missing string `id`) -> returns `[]`.\n\n`SessionManager.setSessionFile()` behavior:\n\n- `[]` from loader is treated as empty/nonexistent session and replaced with a new initialized session file at that path.\n- Valid files are loaded, migrated if needed, blob refs resolved, then indexed.\n\n## Tree and Leaf Semantics\n\nThe underlying model is append-only tree + mutable leaf pointer:\n\n- Every append method creates exactly one new entry whose `parentId` is current `leafId`.\n- The new entry becomes the new `leafId`.\n- `branch(entryId)` moves only `leafId`; existing entries remain unchanged.\n- `resetLeaf()` sets `leafId = null`; next append creates a new root entry (`parentId: null`).\n- `branchWithSummary()` sets leaf to branch target and appends a `branch_summary` entry.\n\n`getEntries()` returns all non-header entries in insertion order. Existing entries are not deleted in normal operation; rewrites preserve logical history while updating representation (migrations, move, targeted rewrite helpers).\n\n## Context Reconstruction (`buildSessionContext`)\n\n`buildSessionContext(entries, leafId?, byId?, options?)` resolves what is sent to the model. Passing `options.transcript: true` instead builds the full-history display transcript (compactions emitted inline at the position they fired) — display-only, never sent to a provider.\n\nAlgorithm:\n\n1. Determine leaf:\n   - `leafId === null` -> return empty context.\n   - explicit `leafId` -> use that entry if found.\n   - otherwise fallback to last entry.\n2. Walk `parentId` chain from leaf to root and reverse to root->leaf path.\n3. Derive runtime state across path:\n   - `thinkingLevel` from latest `thinking_level_change` (default `\"off\"`)\n   - `serviceTier` from latest `service_tier_change`\n   - model map from `model_change` entries (`role ?? \"default\"`)\n   - fallback `models.default` from assistant message provider/model if no explicit model change\n   - deduplicated `injectedTtsrRules` from all `ttsr_injection` entries\n   - selected MCP discovery tools from latest `mcp_tool_selection`\n   - mode/modeData from latest `mode_change` (default mode `\"none\"`)\n4. Build message list:\n   - `message` entries pass through\n   - `custom_message` entries become `custom` AgentMessages via `createCustomMessage`\n   - `branch_summary` entries become `branchSummary` AgentMessages via `createBranchSummaryMessage`\n   - if a `compaction` exists on path:\n     - emit compaction summary first (`createCompactionSummaryMessage`)\n     - emit path entries starting at `firstKeptEntryId` up to the compaction boundary\n     - emit entries after the compaction boundary\n\n`custom`, `session_init`, `service_tier_change`, `mcp_tool_selection`, and `ttsr_injection` entries do not inject model context directly.\n\n## Persistence Guarantees and Failure Model\n\n### Persist vs in-memory\n\n- `SessionManager.create/open/continueRecent/forkFrom` -> persistent mode (`persist = true`).\n- `SessionManager.inMemory` -> non-persistent mode (`persist = false`) with `MemorySessionStorage`.\n\n### Write pipeline\n\nWrites are serialized through an internal promise chain (`#persistChain`) and `NdjsonFileWriter`.\n\n- `append*` updates in-memory state immediately.\n- Persistence is deferred until at least one assistant message exists.\n  - Before first assistant: entries are retained in memory; no file append occurs.\n  - When first assistant exists: full in-memory session is flushed to file.\n  - Afterwards: new entries append incrementally.\n\nRationale in code: avoid persisting sessions that never produced an assistant response.\n\n### Durability operations\n\n- `flush()` flushes writer and calls `fsync()`.\n- Atomic full rewrites (`#rewriteFile`) write to temp file, flush+fsync, close, then rename over target.\n- Used for migrations, `setSessionName`, `rewriteEntries` (tool-output pruning/supersede passes), and move/fork operations.\n\n### Error behavior\n\n- Persistence errors are latched (`#persistError`) and rethrown on subsequent operations.\n- First error is logged once with session file context.\n- Writer close is best-effort but propagates the first meaningful error.\n\n## Data Size Controls and Blob Externalization\n\nBefore persisting entries:\n\n- Large strings are truncated to `MAX_PERSIST_CHARS` (500,000 chars) with notice:\n  - `\"[Session persistence truncated large content]\"`\n- Transient fields `partialJson` and `jsonlEvents` are removed.\n- If object has both `content` and `lineCount`, line count is recomputed after truncation.\n- Image blocks in `content` arrays with base64 length >= 1024 are externalized to blob refs:\n  - stored as `blob:sha256:<hash>`\n  - raw bytes written to blob store (`BlobStore.put`)\n\nOn load, blob refs are resolved back to base64 for message/custom_message image blocks.\n\n## Storage Abstractions\n\n`SessionStorage` interface provides all filesystem operations used by `SessionManager`:\n\n- sync: `ensureDirSync`, `existsSync`, `writeTextSync`, `statSync`, `listFilesSync`\n- async: `exists`, `readText`, `readTextSlices`, `writeText`, `rename`, `unlink`, `deleteSessionWithArtifacts`, `openWriter`\n\nImplementations:\n\n- `FileSessionStorage`: real filesystem (Bun + node fs)\n- `MemorySessionStorage`: map-backed in-memory implementation for tests/non-persistent sessions\n\n`SessionStorageWriter` exposes `writeLine`, `writeLineSync`, `flush`, `fsync`, `close`, `getError`.\n\n## Session Discovery Utilities\n\nDefined in `session-manager.ts`:\n\n- `getRecentSessions(sessionDir, limit)` -> lightweight metadata for UI/session picker, capped by `limit`\n- `findMostRecentSession(sessionDir)` -> newest by mtime\n- `list(cwd, sessionDir?)` -> sessions in one project scope\n- `listAll()` -> sessions across all project scopes under `~/.omp/agent/sessions`\n- `resolveResumableSession(sessionArg, cwd, sessionDir?)` -> local then global resume/fork target lookup\n\nMetadata extraction for `getRecentSessions` reads a prefix via `readTextSlices(..., 4096, 0)`. `list`/`listAll` read a 4KB prefix plus a bounded 32 KiB tail through one `readTextSlices(...)` call per file, using the prefix for metadata and the tail for lifecycle status. Resume matching is case-insensitive and accepts session id prefixes, full filename prefixes, or the id suffix after the timestamp in `<timestamp>_<sessionId>.jsonl`.\n\n## Related but Distinct: Prompt History Storage\n\n`HistoryStorage` (`history-storage.ts`) is a separate SQLite subsystem for prompt recall/search, not session replay.\n\n- DB: `~/.omp/agent/history.db`\n- Table: `history(id, prompt, created_at, cwd, session_id)`\n- FTS5 index: `history_fts` with trigger-maintained sync\n- Deduplicates consecutive identical prompts using in-memory last-prompt cache\n- Inserts are batched through an async drain queue (~100 ms delay) so prompt capture does not block turn execution\n\nUse session files for conversation graph/state replay; use `HistoryStorage` for prompt history UX.\n",
-	"settings.md": "# Settings\n\n`omp` resolves settings from built-in defaults, a persistent global config file, optional project-local config, one-shot CLI overlays, and in-memory runtime overrides. Reach for project settings when one repository needs a different provider set, model role, tool policy, memory backend, or UI behavior than your global defaults — without touching your machine-wide configuration.\n\nSettings are stored as plain YAML mappings. Every key, its type, default, and enum values come from the settings schema, and you can inspect or change any of them with `omp config` or the interactive `/settings` panel.\n\n- For model/provider credentials, `.env` files, and the env-var table that resolves API keys, see [Providers](./providers.md).\n- For custom model definitions in `models.yml`, see [Models](./models.md).\n- For instruction files discovered into the agent context (`AGENTS.md`, `.omp/`, etc.), see [Context files](./context-files.md).\n- For the full catalog of environment variables, see [Environment variables](./environment-variables.md).\n\n## Where settings live\n\n| Scope | Path | Read behavior | Write behavior |\n|---|---|---|---|\n| Global | `~/.omp/agent/config.yml` | The main persistent settings file. Always loaded. | `/settings`, `omp config set`, and `omp config reset` write here. |\n| Global legacy | `~/.omp/agent/settings.json` | Migrated into `config.yml` once, only when `config.yml` does not yet exist. | Not written after migration; the original is renamed to `settings.json.bak`. |\n| Project | `<cwd>/.omp/config.yml` (plus `.omp/settings.json`) | Loaded when the process working directory has a non-empty `.omp/`. | Read-only from settings commands; edit the file by hand. |\n| Project legacy | `<cwd>/.omp/settings.json` | Still read; project `config.yml` is merged on top of it. | Not written by settings commands. |\n| CLI overlay | Any file passed with `--config <file>` | Loaded after global and project settings, for that one process. Repeatable. | Never persisted. |\n| Runtime overrides | In-memory only | Set by dedicated CLI flags (`--model`, `--approval-mode`, …) and feature env vars. | Never persisted. |\n\n`PI_CODING_AGENT_DIR` relocates the `~/.omp/agent` base directory. When it is set, the global `config.yml`, the auth store (`agent.db`), and everything else under the agent directory move with it. Use `omp config path` to print the active agent directory.\n\nNative project settings are intentionally scoped to the process working directory's `.omp/` folder — settings discovery does **not** walk ancestor directories looking for the nearest `.omp/`. Other discovery providers (Claude, Codex, Gemini, Cursor, OpenCode) can also contribute project-level settings from their own files; those are read-only from `omp` settings commands and can be turned off by provider id (see [Provider and source disabling](#provider-and-source-disabling)).\n\n## Config file formats\n\nThe global `config.yml` is always YAML. The generic config loader used for other files (for example `models.yml`) accepts `.yml`, `.yaml`, `.json`, and `.jsonc`:\n\n- When a `.yml`/`.yaml` path is requested and only a sibling `.json` exists, it is migrated to YAML automatically (idempotent, once per process).\n- `.json` and `.jsonc` configs are read as-is, with no migration.\n- A file whose top level is not a mapping (a bare array or scalar) is treated as empty for persistent settings, and is a hard error for `--config` overlays.\n\n## Reading and writing settings\n\nUse the interactive `/settings` panel inside a session, or the `omp config` command from a shell. Both operate on the merged effective settings, but every persistent write lands in the **global** file only.\n\n```bash\nomp config list                 # all settings with current effective values\nomp config list --json          # same, machine-readable\nomp config get theme.dark       # one value\nomp config get theme.dark --json\nomp config set compaction.enabled false\nomp config set defaultThinkingLevel medium\nomp config reset steeringMode   # restore a key to its schema default\nomp config path                 # print the active agent directory\n```\n\n### Subcommands\n\n| Command | Effect |\n|---|---|\n| `omp config list` | Print every setting grouped by tab, with its current value and type. `--json` emits an object keyed by setting path with `{ value, type, description }`. |\n| `omp config get <key>` | Print the effective value of one key. Unknown keys exit non-zero. `--json` emits `{ key, value, type, description }`. |\n| `omp config set <key> <value>` | Parse `<value>` against the key's schema type and write it to the global `config.yml`. |\n| `omp config reset <key>` | Write the key's schema **default** back to the global config (this persists the default, it does not delete the key). |\n| `omp config path` | Print the active agent directory (honors `PI_CODING_AGENT_DIR`). |\n\n`omp config` with no subcommand, or `--help`, prints the help and lists settings. The `--json` flag is accepted by `list`, `get`, `set`, and `reset`.\n\n### Value parsing\n\n`omp config set` parses the value string according to the target key's schema type. The string is trimmed first.\n\n| Type | Accepted input | Notes |\n|---|---|---|\n| boolean | `true`, `false`, `yes`, `no`, `on`, `off`, `1`, `0` | Case-insensitive. Anything else is rejected. |\n| number | Any finite JavaScript number | `Infinity`/`NaN` are rejected. |\n| enum | One of the key's allowed values | Must match exactly; the error lists the valid values. |\n| array | A JSON array | e.g. `'[\"anthropic\",\"openai\"]'`. Must parse and be an array. |\n| record | A JSON object | e.g. `'{\"bash\":\"prompt\"}'`. Must parse and be a non-array object. |\n| string | Stored as given (trimmed) | Multi-word values are joined with spaces. |\n\nKeys must match a real schema path exactly. There is no shorthand — set `theme.dark`, not `theme`.\n\n### Where writes go\n\n`omp config set`, `omp config reset`, `/settings`, and any runtime settings change all write to the global `config.yml` under the active agent directory. They never write to `<cwd>/.omp/config.yml`. To create a project-local override, edit that file directly (see [Project-local config](#project-local-config)). Saves are debounced and re-read the file under a lock, so external edits made while a session is open are preserved.\n\n## Precedence\n\nFrom lowest to highest priority, the effective value of a setting is built as:\n\n```text\nbuilt-in defaults  <-  global config  <-  project config  <-  CLI overlays  <-  runtime overrides\n```\n\nFrom highest to lowest:\n\n1. **Runtime overrides** — dedicated CLI flags and feature env vars applied in memory for the current process: `--model`, `--smol`, `--slow`, `--plan`, `--approval-mode`, `--auto-approve`/`--yolo`, `--hide-thinking`, `--no-pty`, `--api-key`, and protocol-mode defaults. Never persisted.\n2. **CLI config overlays** — each `--config <file>`; later overlay files override earlier ones.\n3. **Project settings** — `<cwd>/.omp/settings.json` then `<cwd>/.omp/config.yml` (and contributions from other discovery providers at project level).\n4. **Global settings** — `~/.omp/agent/config.yml`.\n5. **Built-in defaults** — from the settings schema.\n\nA key that is unset at every layer resolves to its schema default at read time.\n\n### Environment overrides\n\nEnvironment variables are **not** a single settings layer. Each is read by the feature that owns the value, usually as a per-machine override or fallback, and is never written back to `config.yml`. The ones that map directly onto a setting:\n\n| Env var | Overrides setting | Notes |\n|---|---|---|\n| `PI_SMOL_MODEL` | `modelRoles.smol` | Also exposed as `--smol`. |\n| `PI_SLOW_MODEL` | `modelRoles.slow` | Also exposed as `--slow`. |\n| `PI_PLAN_MODEL` | `modelRoles.plan` | Also exposed as `--plan`. |\n| `PI_NO_PTY=1` | (disables PTY bash) | Equivalent to `--no-pty` for the process. |\n| `PI_PY` | `eval.py` | `PI_PY=0` disables the Python eval backend. |\n| `PI_JS` | `eval.js` | `PI_JS=0` disables the JavaScript eval backend. |\n| `PI_TINY_DEVICE` | `providers.tinyModelDevice` | ONNX execution provider for local tiny models. |\n| `PI_TINY_DTYPE` | `providers.tinyModelDtype` | ONNX precision for local tiny models. |\n| `OMP_AUTH_BROKER_URL` | `auth.broker.url` | Env value takes precedence over config. |\n| `OMP_AUTH_BROKER_TOKEN` | `auth.broker.token` | Env value takes precedence over config. |\n| `PI_CODING_AGENT_DIR` | (relocates agent dir) | Moves `config.yml`, `agent.db`, and the whole agent base. |\n\nProvider API keys are resolved separately (stored auth, OAuth, `models.yml`, environment, and `.env` files); see [Providers](./providers.md) and the full [Environment variables](./environment-variables.md) reference.\n\n## Merge rules\n\nLayers are combined with a deep merge:\n\n- **Objects are deep-merged** — keys present only in a lower layer are kept; keys present in a higher layer override.\n- **Scalars and arrays are replaced wholesale** by the higher-precedence layer. A higher layer's array does not append to a lower layer's array.\n\nUse nested YAML mappings for dotted setting paths:\n\n```yaml\ntheme:\n  dark: titanium\n  light: light\n\ntools:\n  approvalMode: write\n  approval:\n    bash: prompt\n    read: allow\n```\n\n### Worked example: global vs. project\n\n```yaml\n# ~/.omp/agent/config.yml\ntools:\n  approvalMode: write\n  approval:\n    bash: prompt\n    read: allow\ndisabledProviders:\n  - anthropic\n  - openai\n  - gemini\n\n# <repo>/.omp/config.yml\ntools:\n  approval:\n    bash: allow\ndisabledProviders:\n  - groq\n```\n\nEffective settings inside `<repo>`:\n\n```yaml\ntools:\n  approvalMode: write   # kept from global (object deep-merge)\n  approval:\n    bash: allow         # overridden by project\n    read: allow         # kept from global\ndisabledProviders:\n  - groq                # project array REPLACES the global array\n```\n\nArray replacement is the most common surprise: the project's `disabledProviders` does not extend the global list — it becomes the entire list for that project. The same applies to `enabledModels`, `cycleOrder`, `extensions`, and every other array-typed setting.\n\n## Project-local config\n\nCreate `<repo>/.omp/config.yml` when a repository needs its own settings:\n\n```yaml\n# <repo>/.omp/config.yml\nmodelRoles:\n  default: anthropic/claude-sonnet-4-5\n  smol: openai/gpt-4.1-mini\n  slow: anthropic/claude-opus-4-5:high\n\ntools:\n  approvalMode: write\n  approval:\n    bash: prompt\n\ncompaction:\n  strategy: context-full\n  thresholdPercent: 80\n\ntheme:\n  dark: titanium\n```\n\nKeep secrets out of committed project config unless your repository policy allows it. Prefer environment variables, stored auth, an auth broker, or an untracked `--config` overlay for credentials.\n\n### One-shot overlays\n\nUse `--config` for a temporary layer that should not persist:\n\n```bash\nomp --config ./local/ci-settings.yml \"check this failure\"\nomp --config ./base.yml --config ./experiment.yml \"try this model\"\n```\n\nOverlay paths are resolved relative to the process working directory (and `~` is expanded). Each overlay must parse as a YAML mapping; a missing file, invalid YAML, or a top-level array/scalar is a hard error — it does **not** silently fall back to lower-precedence settings.\n\n## Path-scoped arrays\n\nTwo array settings — `enabledModels` and `disabledProviders` — accept path-scoped entries in addition to bare strings, so a single global config can behave differently per directory:\n\n```yaml\nenabledModels:\n  - claude-sonnet-4-5            # applies everywhere\n  - path: ~/work/high-context\n    models:\n      - anthropic/claude-opus-4-5\n\ndisabledProviders:\n  - ollama                       # applies everywhere\n  - paths:\n      - ~/projects/sensitive\n      - ~/clients/acme\n    providers:\n      - anthropic\n      - openai\n```\n\nBare string entries apply everywhere. A scoped entry applies when the current working directory **is** the configured path or is **under** it. `~` expands to your home directory and relative paths are resolved before matching.\n\nAccepted **path** keys (any of them, combined): `path`, `paths`, `pathPrefix`, `pathPrefixes`.\n\nAccepted **value** keys:\n\n- `models` (for `enabledModels`) or `providers` (for `disabledProviders`)\n- `values` or `items` (for either setting)\n\nOnly string values are kept; malformed scoped entries are ignored. Path scoping is resolved **after** the layer merge, so it reads the final effective array.\n\n## Provider and source disabling\n\n`disabledProviders` is a single shared id namespace that gates two different subsystems, before any credential check:\n\n| Entry kind | Example ids | Effect |\n|---|---|---|\n| Model providers | `anthropic`, `openai`, `gemini`, `groq`, `ollama`, `openrouter` | Removes those backends from model selection, even when credentials are available. See [Providers](./providers.md). |\n| Discovery sources | `native`, `claude`, `codex`, `gemini`, `github`, `opencode`, `cursor`, `agents-md` | Stops that source from contributing context files, MCP servers, commands, skills, hooks, tools, prompts, or settings. See [Context files](./context-files.md). |\n\nMost provider-control use cases list model provider ids. Disabling the `claude` discovery source is different from disabling the `anthropic` model provider — one stops Claude-format config discovery, the other stops the Anthropic model backend.\n\nBecause arrays replace rather than append, a project that sets `disabledProviders` must list the complete desired set:\n\n```yaml\n# ~/.omp/agent/config.yml\ndisabledProviders:\n  - anthropic\n  - openai\n\n# <repo>/.omp/config.yml — inside this repo ONLY groq is disabled\ndisabledProviders:\n  - groq\n```\n\nThe default is an empty array (nothing disabled). For the two subsystems' provider ids and ordering, see [Providers](./providers.md) and [Context files](./context-files.md).\n\n## Settings catalog\n\nEvery key below is defined in the settings schema; `omp config list` shows the full set with current values. Defaults and enum values are taken from the schema. Settings that accept an env or flag override are noted; those overrides are process-local and not persisted.\n\n### Models\n\n`modelRoles`, `modelTags`, and `cycleOrder` work together to define the models you can switch between. Role values may carry a thinking suffix (`:minimal`, `:low`, `:medium`, `:high`, `:xhigh`).\n\n```yaml\nmodelRoles:\n  default: anthropic/claude-sonnet-4-5\n  smol: openai/gpt-4.1-mini\n  slow: anthropic/claude-opus-4-5:high\n  vision: gemini/gemini-3-pro-preview\n  plan: anthropic/claude-opus-4-5\n\ncycleOrder:\n  - smol\n  - default\n  - slow\n\nmodelProviderOrder:\n  - anthropic\n  - openai\n\nenabledModels:\n  - claude-sonnet-4-5\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `modelRoles` | record | `{}` | Map of role name -> model id. Built-in roles: `default`, `smol`, `slow`, `vision`, `plan`, `designer`, `commit`, `task`. Per-role env/flags: `--model`/`--smol`/`--slow`/`--plan`. |\n| `modelTags` | record | `{}` | Custom role/tag metadata; can introduce additional roles. |\n| `modelProviderOrder` | array | `[]` | Preferred provider order when a model id is ambiguous. |\n| `cycleOrder` | array | `[\"smol\",\"default\",\"slow\"]` | Roles cycled by the model switcher. |\n| `enabledModels` | array | `[]` | Allow-list of models; supports [path-scoped entries](#path-scoped-arrays). Empty means all available models. |\n| `disabledProviders` | array | `[]` | Disabled model/discovery providers; supports path-scoped entries. See [above](#provider-and-source-disabling). |\n| `includeModelInPrompt` | boolean | `true` | Include the active model name in the system prompt. |\n\nSee [Models](./models.md) for the `models.yml` schema and custom-provider definitions.\n\n### Thinking\n\n```yaml\ndefaultThinkingLevel: high\nhideThinkingBlock: false\nthinkingBudgets:\n  minimal: 1024\n  low: 2048\n  medium: 8192\n  high: 16384\n  xhigh: 32768\n```\n\n| Key | Type | Default | Values |\n|---|---|---|---|\n| `defaultThinkingLevel` | enum | `high` | `minimal`, `low`, `medium`, `high`, `xhigh`, `auto`. Override per run with `--thinking`. |\n| `hideThinkingBlock` | boolean | `false` | Hide thinking blocks in output. `--hide-thinking` sets it for the run (display only). |\n| `thinkingBudgets.minimal` | number | `1024` | Token budget for the `minimal` level. |\n| `thinkingBudgets.low` | number | `2048` | Token budget for `low`. |\n| `thinkingBudgets.medium` | number | `8192` | Token budget for `medium`. |\n| `thinkingBudgets.high` | number | `16384` | Token budget for `high`. |\n| `thinkingBudgets.xhigh` | number | `32768` | Token budget for `xhigh`. |\n\n### Sampling\n\nA value of `-1` means \"use the provider/model default\" — `omp` does not send that parameter.\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `temperature` | number | `-1` | Sampling temperature. |\n| `topP` | number | `-1` | Nucleus sampling. |\n| `topK` | number | `-1` | Top-K sampling. |\n| `minP` | number | `-1` | Minimum-probability cutoff. |\n| `presencePenalty` | number | `-1` | Presence penalty. |\n| `repetitionPenalty` | number | `-1` | Repetition penalty. |\n| `serviceTier` | enum | `none` | `none`, `auto`, `default`, `flex`, `scale`, `priority`, `openai-only`, `claude-only`. |\n| `personality` | enum | `default` | `default`, `friendly`, `pragmatic`, `none`. |\n\n### Retry and fallback\n\n```yaml\nretry:\n  enabled: true\n  maxRetries: 10\n  baseDelayMs: 500\n  maxDelayMs: 300000\n  modelFallback: true\n  fallbackRevertPolicy: cooldown-expiry\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `retry.enabled` | boolean | `true` | Retry transient provider errors. |\n| `retry.maxRetries` | number | `10` | Max retries per request. |\n| `retry.baseDelayMs` | number | `500` | Initial backoff. |\n| `retry.maxDelayMs` | number | `300000` | Backoff ceiling (5 min). |\n| `retry.modelFallback` | boolean | `true` | Fall back to another model when one is unavailable. |\n| `retry.fallbackChains` | record | `{}` | Per-model fallback chains. |\n| `retry.fallbackRevertPolicy` | enum | `cooldown-expiry` | `cooldown-expiry`, `never`. |\n\n### Tools and approvals\n\n```yaml\ntools:\n  approvalMode: yolo          # default\n  approval:\n    bash: prompt\n    edit: allow\n  discoveryMode: auto\n  maxTimeout: 0\n  intentTracing: true\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `tools.approvalMode` | enum | `yolo` | `always-ask` (auto-approve read-only), `write` (auto-approve read + workspace-write), `yolo` (auto-approve all tiers). `--approval-mode` and `--auto-approve`/`--yolo` override per run. |\n| `tools.approval` | record | `{}` | Per-tool policy keyed by tool name; each value is `allow`, `deny`, or `prompt`. e.g. `omp config set tools.approval '{\"bash\":\"prompt\"}'`. |\n| `tools.discoveryMode` | enum | `auto` | `auto`, `off`, `mcp-only`, `all`. Controls dynamic tool discovery. |\n| `tools.essentialOverride` | array | `[]` | Tool names kept available even when tools are narrowed. |\n| `tools.maxTimeout` | number | `0` | Max tool runtime in seconds; `0` = no cap. |\n| `tools.intentTracing` | boolean | `true` | Record per-call intent strings. |\n| `tools.outputMaxColumns` | number | `768` | Per-line byte cap for streaming output; `0` disables. |\n| `tools.artifactSpillThreshold` | number | `50` | KB of tool output above which output spills to an artifact. |\n| `tools.artifactHeadBytes` | number | `20` | KB of head kept inline on spill; `0` = tail-only. |\n| `tools.artifactTailBytes` | number | `20` | KB of tail kept inline on spill. |\n| `tools.artifactTailLines` | number | `500` | Max tail lines kept inline on spill. |\n\nIndividual built-in tools are toggled by their own keys, e.g. `bash.enabled`, `eval.py`, `eval.js`, `find.enabled`, `search.enabled`, `fetch.enabled`, `browser.enabled`, `astEdit.enabled`, `astGrep.enabled`, `web_search.enabled`, `inspect_image.enabled`, `renderMermaid.enabled`.\n\n### Shell, eval, and LSP\n\n```yaml\nbash:\n  enabled: true\n  stripTrailingHeadTail: true\n  autoBackground:\n    enabled: false\n    thresholdMs: 60000\n\neval:\n  py: true\n  js: true\n\npython:\n  kernelMode: session       # session, per-call\n  interpreter: \"\"\n\nlsp:\n  enabled: true\n  lazy: true\n  diagnosticsOnWrite: true\n  diagnosticsOnEdit: false\n  formatOnWrite: false\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `bash.enabled` | boolean | `true` | Enable the bash tool. |\n| `bash.stripTrailingHeadTail` | boolean | `true` | Strip trailing head/tail noise from output. |\n| `bash.autoBackground.enabled` | boolean | `false` | Auto-background long-running commands. |\n| `bash.autoBackground.thresholdMs` | number | `60000` | Threshold before auto-backgrounding. |\n| `eval.py` | boolean | `true` | Python eval backend. `PI_PY=0` disables for the process. |\n| `eval.js` | boolean | `true` | JavaScript eval backend. `PI_JS=0` disables for the process. |\n| `python.kernelMode` | enum | `session` | `session` (persistent kernel) or `per-call`. |\n| `python.interpreter` | string | `\"\"` | Path to a Python interpreter; empty = auto-detect. |\n| `lsp.enabled` | boolean | `true` | Language-server integration. `--no-lsp` disables for the run. |\n| `lsp.lazy` | boolean | `true` | Start servers on demand. |\n| `lsp.diagnosticsOnWrite` | boolean | `true` | Run diagnostics after a write. |\n| `lsp.diagnosticsOnEdit` | boolean | `false` | Run diagnostics after an edit. |\n| `lsp.formatOnWrite` | boolean | `false` | Format files on write. |\n| `lsp.diagnosticsDeduplicate` | boolean | `true` | Collapse duplicate diagnostics. |\n| `shellPath` | string | _(unset)_ | Override the shell binary used by bash. |\n\n### Files: editing and reading\n\n```yaml\nedit:\n  mode: hashline            # apply_patch, hashline, patch, replace\n  fuzzyMatch: true\n  fuzzyThreshold: 0.95\n  blockAutoGenerated: true\n\nread:\n  defaultLimit: 300\n  toolResultPreview: false\n  summarize:\n    enabled: true\n    prose: false\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `edit.mode` | enum | `hashline` | `apply_patch`, `hashline`, `patch`, `replace`. |\n| `edit.fuzzyMatch` | boolean | `true` | Allow fuzzy anchor matching. |\n| `edit.fuzzyThreshold` | number | `0.95` | Similarity threshold for fuzzy matching. |\n| `edit.blockAutoGenerated` | boolean | `true` | Refuse to edit generated/lockfile-like files. |\n| `edit.streamingAbort` | boolean | `false` | Abort on streaming edit mismatch. |\n| `read.defaultLimit` | number | `300` | Default line count for `read` without a selector. |\n| `read.summarize.enabled` | boolean | `true` | Structural summaries for code reads. |\n| `read.summarize.prose` | boolean | `false` | Summarize prose files too. |\n| `read.toolResultPreview` | boolean | `false` | Inline preview of tool results. |\n| `readHashLines` | boolean | `true` | Show hashline tags in read output. |\n| `readLineNumbers` | boolean | `false` | Show plain line numbers. |\n\n### Context, compaction, and memory\n\n```yaml\ncontextPromotion:\n  enabled: true\n\ncompaction:\n  enabled: true\n  strategy: context-full     # context-full, handoff, shake, snapcompact, off\n  thresholdPercent: -1       # -1 = default reserve-based behavior\n  thresholdTokens: -1        # fixed token limit when > 0\n  remoteEnabled: true\n\nmemory:\n  backend: off               # off, local, hindsight, mnemopi\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `contextPromotion.enabled` | boolean | `true` | Promote relevant earlier context. |\n| `compaction.enabled` | boolean | `true` | Automatic conversation compaction. |\n| `compaction.strategy` | enum | `context-full` | `context-full`, `handoff`, `shake`, `snapcompact`, `off`. |\n| `compaction.thresholdPercent` | number | `-1` | Percent-of-context trigger; `-1` = reserve-based default. |\n| `compaction.thresholdTokens` | number | `-1` | Fixed token trigger when `> 0`. |\n| `compaction.reserveTokens` | number | `16384` | Tokens reserved for the next turn. |\n| `compaction.keepRecentTokens` | number | `20000` | Recent tokens always preserved. |\n| `compaction.remoteEnabled` | boolean | `true` | Allow remote compaction service. |\n| `compaction.autoContinue` | boolean | `true` | Continue automatically after compaction. |\n| `memory.backend` | enum | `off` | `off`, `local`, `hindsight`, `mnemopi`. Each backend has its own `hindsight.*` / `mnemopi.*` / `memories.*` tuning keys. |\n\n`compaction` has additional tuning keys (idle compaction, supersede/drop heuristics) visible in `omp config list`. See [Compaction](./compaction.md) for the full strategy reference.\n\n### Appearance and terminal\n\n```yaml\ntheme:\n  dark: titanium\n  light: light\nsymbolPreset: unicode        # unicode, nerd, ascii\ncolorBlindMode: false\n\nstatusLine:\n  preset: default            # default, minimal, compact, full, nerd, ascii, custom\n  separator: powerline-thin\n  transparent: false\n  showHookStatus: true\n\nterminal:\n  showImages: true\nimages:\n  autoResize: true\n  blockImages: false\ntui:\n  hyperlinks: auto           # off, auto, always\n```\n\n| Key | Type | Default | Values |\n|---|---|---|---|\n| `theme.dark` | string | `titanium` | Theme used on a dark terminal background. |\n| `theme.light` | string | `light` | Theme used on a light terminal background. |\n| `symbolPreset` | enum | `unicode` | `unicode`, `nerd`, `ascii`. |\n| `colorBlindMode` | boolean | `false` | Use blue instead of green for diff additions. |\n| `showHardwareCursor` | boolean | `true` | Show the terminal hardware cursor. |\n| `statusLine.preset` | enum | `default` | `default`, `minimal`, `compact`, `full`, `nerd`, `ascii`, `custom`. |\n| `statusLine.separator` | enum | `powerline-thin` | `powerline`, `powerline-thin`, `slash`, `pipe`, `block`, `none`, `ascii`. |\n| `statusLine.sessionAccent` | boolean | `true` | Tint the editor border with the session color. |\n| `statusLine.transparent` | boolean | `false` | Use the terminal background for the status line. |\n| `statusLine.showHookStatus` | boolean | `true` | Show hook status messages. |\n| `terminal.showImages` | boolean | `true` | Render images inline (when the terminal supports it). |\n| `images.autoResize` | boolean | `true` | Resize large images for model compatibility. |\n| `images.blockImages` | boolean | `false` | Never send images to providers. |\n| `tui.hyperlinks` | enum | `auto` | `off`, `auto`, `always`. |\n\nFor a custom status line, set `statusLine.preset: custom` and configure `statusLine.leftSegments`, `statusLine.rightSegments`, and `statusLine.segmentOptions`.\n\n### Interaction\n\n| Key | Type | Default | Values |\n|---|---|---|---|\n| `steeringMode` | enum | `one-at-a-time` | `all`, `one-at-a-time`. How queued steering messages are delivered. |\n| `followUpMode` | enum | `one-at-a-time` | `all`, `one-at-a-time`. |\n| `interruptMode` | enum | `immediate` | `immediate`, `wait`. |\n| `doubleEscapeAction` | enum | `tree` | `branch`, `tree`, `none`. |\n| `autoResume` | boolean | `false` | Auto-resume the most recent session in the cwd. |\n| `ask.timeout` | number | `0` | Seconds before an `ask` prompt times out; `0` = no timeout. (Legacy ms values are migrated to seconds.) |\n| `ask.notify` | enum | `on` | `on`, `off`. |\n\n### Providers and services\n\n```yaml\nproviders:\n  webSearch: auto\n  image: auto\n  fetch: auto\n  tinyModel: online\n  tinyModelDevice: default\n  tinyModelDtype: default\n  openaiWebsockets: auto\n  openrouterVariant: default\n  kimiApiFormat: anthropic\n\nprovider:\n  appendOnlyContext: auto    # auto, on, off\n\nexa:\n  enabled: true\n  enableSearch: true\n  enableResearcher: false\n  enableWebsets: false\n\nsearxng:\n  endpoint: https://search.example.com\n  token: SEARXNG_TOKEN\n```\n\n| Key | Type | Default | Values / notes |\n|---|---|---|---|\n| `providers.webSearch` | enum | `auto` | `auto` plus the configured search providers (`tavily`, `perplexity`, `brave`, `jina`, `kimi`, `anthropic`, `gemini`, `codex`, `zai`, `exa`, `parallel`, `kagi`, `synthetic`, `searxng`). |\n| `providers.image` | enum | `auto` | `auto`, `openai`, `antigravity`, `xai`, `gemini`, `openrouter`. |\n| `providers.fetch` | enum | `auto` | `auto`, `native`, `trafilatura`, `lynx`, `parallel`, `jina`. |\n| `providers.tinyModel` | enum | `online` | `online` or a local model (`lfm2-350m`, `qwen3-0.6b`, `gemma-270m`, `qwen2.5-0.5b`, `lfm2-700m`). |\n| `providers.tinyModelDevice` | enum | `default` | ONNX execution provider for local tiny models. Overridden by `PI_TINY_DEVICE`. |\n| `providers.tinyModelDtype` | enum | `default` | ONNX precision for local tiny models. Overridden by `PI_TINY_DTYPE`. |\n| `providers.openaiWebsockets` | enum | `auto` | `auto`, `off`, `on`. |\n| `providers.openrouterVariant` | enum | `default` | `default`, `nitro`, `floor`, `online`, `exacto`. |\n| `providers.kimiApiFormat` | enum | `anthropic` | `openai`, `anthropic`. |\n| `provider.appendOnlyContext` | enum | `auto` | `auto`, `on`, `off`. |\n| `exa.enabled` | boolean | `true` | Enable Exa integration. |\n| `exa.enableSearch` | boolean | `true` | Exa search. |\n| `exa.enableResearcher` | boolean | `false` | Exa researcher. |\n| `exa.enableWebsets` | boolean | `false` | Exa websets. |\n| `searxng.endpoint` | string | _(unset)_ | SearXNG instance URL. |\n| `searxng.token` | string | _(unset)_ | SearXNG token; also `searxng.basicUsername`/`searxng.basicPassword`/`searxng.categories`/`searxng.language`. |\n| `auth.broker.url` | string | _(unset)_ | Auth-broker URL. Overridden by `OMP_AUTH_BROKER_URL`. |\n| `auth.broker.token` | string | _(unset)_ | Auth-broker token. Overridden by `OMP_AUTH_BROKER_TOKEN`. |\n\nProvider credentials and custom model definitions are configured separately — see [Providers](./providers.md) and [Models](./models.md).\n\n### Other groups\n\n`omp config list` exposes many more grouped settings, including: `task.*` (subagent concurrency, isolation, model overrides), `skills.*` and `commands.*` (discovery toggles), `mcp.*`, `github.*`, `async.*`, `goal.*`, `loop.*`, `todo.*`, `magicKeywords.*`, `ttsr.*` (sticky rules), `display.*`, `startup.*`, `share.*`, `collab.*`, `stt.*`/`tts.*`, `memories.*`/`hindsight.*`/`mnemopi.*` (memory backends), and `bashInterceptor.*`. Each follows the same type/default rules shown above.\n\n## Legacy migration\n\n`omp` migrates older config shapes automatically. None of these require action; they are listed so you know what changes you may see in `config.yml`.\n\n### Startup migration to `config.yml`\n\nWhen `~/.omp/agent/config.yml` does not exist, startup builds it once from legacy sources, then writes the result:\n\n1. `~/.omp/agent/settings.json` (renamed to `settings.json.bak` after a successful migration).\n2. Settings persisted in `agent.db`.\n\nAfter `config.yml` exists, these legacy sources are no longer consulted. The generic config loader also performs `.json` -> `.yml` migration for other config files when only the `.json` form is present.\n\n### Field-level migrations\n\nApplied whenever raw settings are loaded (global, project, overlays, and runtime overrides):\n\n| Old | New |\n|---|---|\n| `queueMode` | `steeringMode` |\n| `ask.timeout` in milliseconds (value `> 1000`) | seconds (divided by 1000) |\n| flat `theme: \"<name>\"` string | `theme.dark` / `theme.light` (slot chosen by luminance; built-in `light`/`dark` are dropped to use defaults) |\n| `task.isolation.enabled: true/false` | `task.isolation.mode: auto/none` |\n| `task.simple` | removed |\n| legacy `task.isolation.mode` (`worktree`, `fuse-overlay`, `fuse-projfs`) | `rcopy`, `overlayfs`, `projfs` |\n| `lastChangelogVersion` | moved to a marker file and stripped from `config.yml` |\n\n## Troubleshooting\n\n### A project setting is not taking effect\n\n- Start `omp` from the directory that contains `.omp/config.yml`. Settings discovery only checks the current working directory's `.omp/`, not ancestor directories.\n- Ensure `.omp/` is non-empty; empty config directories are ignored.\n- Confirm the file is valid YAML and its top level is a mapping.\n- Run `omp config get <key>` from that directory to see the effective value.\n- Remember that `--config` overlays and runtime flags override project config.\n\n### A global array disappeared in a project\n\nArrays replace; they do not append. If a project sets `disabledProviders`, `enabledModels`, `cycleOrder`, `extensions`, or any other array, include the **complete** desired value in the project layer — the global array is fully replaced.\n\n### A provider is still available after editing config\n\n- Check whether you disabled the model provider id (e.g. `anthropic`) or a discovery source id (e.g. `claude`) — they are different namespaces with different effects.\n- Check for a project (or overlay) `disabledProviders` array replacing your global one.\n- Credentials can still come from environment variables, `.env`, OAuth, stored auth, or `models.yml`; disabling a provider blocks selection regardless, but verify you edited the right layer. See [Providers](./providers.md).\n- Restart the session if the model list was already initialized.\n\n### `omp config set` changed the wrong file\n\n`omp config set` and `omp config reset` always write the global `config.yml` under the active agent directory. Run `omp config path` to print it. For project-local settings, edit `<repo>/.omp/config.yml` directly.\n\n### `omp config reset` did not remove my key\n\n`reset` writes the schema **default** value into the global config — it persists the default rather than deleting the key. To stop overriding a project value from global config, delete the key from `~/.omp/agent/config.yml` by hand.\n\n### A `--config` overlay fails at startup\n\n`--config` files are process-local YAML mappings. A missing file, invalid YAML, or a top-level array/scalar is a hard error — it does not silently fall back to lower-precedence settings. Fix the path or contents.\n\n### An environment variable beats my config\n\nSome settings (model roles, eval backends, tiny-model device/precision, auth broker, PTY) are overridable by env vars or CLI flags for per-machine convenience, and those take precedence over `config.yml`. Unset the variable or drop the flag to let the persisted value win. See [Environment overrides](#environment-overrides) and [Environment variables](./environment-variables.md).\n\n### `omp config set <key>` says \"Unknown setting\"\n\nKeys must match a schema path exactly, with no shorthand. Use `theme.dark`, not `theme`. Run `omp config list` to see every valid key.\n",
+	"settings.md": "# Settings\n\n`omp` resolves settings from built-in defaults, a persistent global config file, optional project-local config, one-shot CLI overlays, and in-memory runtime overrides. Reach for project settings when one repository needs a different provider set, model role, tool policy, memory backend, or UI behavior than your global defaults — without touching your machine-wide configuration.\n\nSettings are stored as plain YAML mappings. Every key, its type, default, and enum values come from the settings schema, and you can inspect or change any of them with `omp config` or the interactive `/settings` panel.\n\n- For model/provider credentials, `.env` files, and the env-var table that resolves API keys, see [Providers](./providers.md).\n- For custom model definitions in `models.yml`, see [Models](./models.md).\n- For instruction files discovered into the agent context (`AGENTS.md`, `.omp/`, etc.), see [Context files](./context-files.md).\n- For the full catalog of environment variables, see [Environment variables](./environment-variables.md).\n\n## Where settings live\n\n| Scope | Path | Read behavior | Write behavior |\n|---|---|---|---|\n| Global | `~/.omp/agent/config.yml` | The main persistent settings file. Always loaded. | `/settings`, `omp config set`, and `omp config reset` write here. |\n| Global legacy | `~/.omp/agent/settings.json` | Migrated into `config.yml` once, only when `config.yml` does not yet exist. | Not written after migration; the original is renamed to `settings.json.bak`. |\n| Project | `<cwd>/.omp/config.yml` (plus `.omp/settings.json`) | Loaded when the process working directory has a non-empty `.omp/`. | Read-only from settings commands; edit the file by hand. |\n| Project legacy | `<cwd>/.omp/settings.json` | Still read; project `config.yml` is merged on top of it. | Not written by settings commands. |\n| CLI overlay | Any file passed with `--config <file>` | Loaded after global and project settings, for that one process. Repeatable. | Never persisted. |\n| Runtime overrides | In-memory only | Set by dedicated CLI flags (`--model`, `--approval-mode`, …) and feature env vars. | Never persisted. |\n\n`PI_CODING_AGENT_DIR` relocates the `~/.omp/agent` base directory. When it is set, the global `config.yml`, the auth store (`agent.db`), and everything else under the agent directory move with it. Use `omp config path` to print the active agent directory.\n\nNative project settings are intentionally scoped to the process working directory's `.omp/` folder — settings discovery does **not** walk ancestor directories looking for the nearest `.omp/`. Other discovery providers (Claude, Codex, Gemini, Cursor, OpenCode) can also contribute project-level settings from their own files; those are read-only from `omp` settings commands and can be turned off by provider id (see [Provider and source disabling](#provider-and-source-disabling)).\n\n## Config file formats\n\nThe global `config.yml` is always YAML. The generic config loader used for other files (for example `models.yml`) accepts `.yml`, `.yaml`, `.json`, and `.jsonc`:\n\n- When a `.yml`/`.yaml` path is requested and only a sibling `.json` exists, it is migrated to YAML automatically (idempotent, once per process).\n- `.json` and `.jsonc` configs are read as-is, with no migration.\n- A file whose top level is not a mapping (a bare array or scalar) is treated as empty for persistent settings, and is a hard error for `--config` overlays.\n\n## Reading and writing settings\n\nUse the interactive `/settings` panel inside a session, or the `omp config` command from a shell. Both operate on the merged effective settings, but every persistent write lands in the **global** file only.\n\n```bash\nomp config list                 # all settings with current effective values\nomp config list --json          # same, machine-readable\nomp config get theme.dark       # one value\nomp config get theme.dark --json\nomp config set compaction.enabled false\nomp config set defaultThinkingLevel medium\nomp config reset steeringMode   # restore a key to its schema default\nomp config path                 # print the active agent directory\n```\n\n### Subcommands\n\n| Command | Effect |\n|---|---|\n| `omp config list` | Print every setting grouped by tab, with its current value and type. `--json` emits an object keyed by setting path with `{ value, type, description }`. |\n| `omp config get <key>` | Print the effective value of one key. Unknown keys exit non-zero. `--json` emits `{ key, value, type, description }`. |\n| `omp config set <key> <value>` | Parse `<value>` against the key's schema type and write it to the global `config.yml`. |\n| `omp config reset <key>` | Write the key's schema **default** back to the global config (this persists the default, it does not delete the key). |\n| `omp config path` | Print the active agent directory (honors `PI_CODING_AGENT_DIR`). |\n\n`omp config` with no subcommand, or `--help`, prints the help and lists settings. The `--json` flag is accepted by `list`, `get`, `set`, and `reset`.\n\n### Value parsing\n\n`omp config set` parses the value string according to the target key's schema type. The string is trimmed first.\n\n| Type | Accepted input | Notes |\n|---|---|---|\n| boolean | `true`, `false`, `yes`, `no`, `on`, `off`, `1`, `0` | Case-insensitive. Anything else is rejected. |\n| number | Any finite JavaScript number | `Infinity`/`NaN` are rejected. |\n| enum | One of the key's allowed values | Must match exactly; the error lists the valid values. |\n| array | A JSON array | e.g. `'[\"anthropic\",\"openai\"]'`. Must parse and be an array. |\n| record | A JSON object | e.g. `'{\"bash\":\"prompt\"}'`. Must parse and be a non-array object. |\n| string | Stored as given (trimmed) | Multi-word values are joined with spaces. |\n\nKeys must match a real schema path exactly. There is no shorthand — set `theme.dark`, not `theme`.\n\n### Where writes go\n\n`omp config set`, `omp config reset`, `/settings`, and any runtime settings change all write to the global `config.yml` under the active agent directory. They never write to `<cwd>/.omp/config.yml`. To create a project-local override, edit that file directly (see [Project-local config](#project-local-config)). Saves are debounced and re-read the file under a lock, so external edits made while a session is open are preserved.\n\n## Precedence\n\nFrom lowest to highest priority, the effective value of a setting is built as:\n\n```text\nbuilt-in defaults  <-  global config  <-  project config  <-  CLI overlays  <-  runtime overrides\n```\n\nFrom highest to lowest:\n\n1. **Runtime overrides** — dedicated CLI flags and feature env vars applied in memory for the current process: `--model`, `--smol`, `--slow`, `--plan`, `--approval-mode`, `--auto-approve`/`--yolo`, `--hide-thinking`, `--no-pty`, `--api-key`, and protocol-mode defaults. Never persisted.\n2. **CLI config overlays** — each `--config <file>`; later overlay files override earlier ones.\n3. **Project settings** — `<cwd>/.omp/settings.json` then `<cwd>/.omp/config.yml` (and contributions from other discovery providers at project level).\n4. **Global settings** — `~/.omp/agent/config.yml`.\n5. **Built-in defaults** — from the settings schema.\n\nA key that is unset at every layer resolves to its schema default at read time.\n\n### Environment overrides\n\nEnvironment variables are **not** a single settings layer. Each is read by the feature that owns the value, usually as a per-machine override or fallback, and is never written back to `config.yml`. The ones that map directly onto a setting:\n\n| Env var | Overrides setting | Notes |\n|---|---|---|\n| `PI_SMOL_MODEL` | `modelRoles.smol` | Also exposed as `--smol`. |\n| `PI_SLOW_MODEL` | `modelRoles.slow` | Also exposed as `--slow`. |\n| `PI_PLAN_MODEL` | `modelRoles.plan` | Also exposed as `--plan`. |\n| `PI_NO_PTY=1` | (disables PTY bash) | Equivalent to `--no-pty` for the process. |\n| `PI_PY` | `eval.py` | `PI_PY=0` disables the Python eval backend. |\n| `PI_JS` | `eval.js` | `PI_JS=0` disables the JavaScript eval backend. |\n| `PI_TINY_DEVICE` | `providers.tinyModelDevice` | ONNX execution provider for local tiny models. |\n| `PI_TINY_DTYPE` | `providers.tinyModelDtype` | ONNX precision for local tiny models. |\n| `OMP_AUTH_BROKER_URL` | `auth.broker.url` | Env value takes precedence over config. |\n| `OMP_AUTH_BROKER_TOKEN` | `auth.broker.token` | Env value takes precedence over config. |\n| `PI_CODING_AGENT_DIR` | (relocates agent dir) | Moves `config.yml`, `agent.db`, and the whole agent base. |\n\nProvider API keys are resolved separately (stored auth, OAuth, `models.yml`, environment, and `.env` files); see [Providers](./providers.md) and the full [Environment variables](./environment-variables.md) reference.\n\n## Merge rules\n\nLayers are combined with a deep merge:\n\n- **Objects are deep-merged** — keys present only in a lower layer are kept; keys present in a higher layer override.\n- **Scalars and arrays are replaced wholesale** by the higher-precedence layer. A higher layer's array does not append to a lower layer's array.\n\nUse nested YAML mappings for dotted setting paths:\n\n```yaml\ntheme:\n  dark: titanium\n  light: light\n\ntools:\n  approvalMode: write\n  approval:\n    bash: prompt\n    read: allow\n```\n\n### Worked example: global vs. project\n\n```yaml\n# ~/.omp/agent/config.yml\ntools:\n  approvalMode: write\n  approval:\n    bash: prompt\n    read: allow\ndisabledProviders:\n  - anthropic\n  - openai\n  - gemini\n\n# <repo>/.omp/config.yml\ntools:\n  approval:\n    bash: allow\ndisabledProviders:\n  - groq\n```\n\nEffective settings inside `<repo>`:\n\n```yaml\ntools:\n  approvalMode: write   # kept from global (object deep-merge)\n  approval:\n    bash: allow         # overridden by project\n    read: allow         # kept from global\ndisabledProviders:\n  - groq                # project array REPLACES the global array\n```\n\nArray replacement is the most common surprise: the project's `disabledProviders` does not extend the global list — it becomes the entire list for that project. The same applies to `enabledModels`, `cycleOrder`, `extensions`, and every other array-typed setting.\n\n## Project-local config\n\nCreate `<repo>/.omp/config.yml` when a repository needs its own settings:\n\n```yaml\n# <repo>/.omp/config.yml\nmodelRoles:\n  default: anthropic/claude-sonnet-4-5\n  smol: openai/gpt-4.1-mini\n  slow: anthropic/claude-opus-4-5:high\n\ntools:\n  approvalMode: write\n  approval:\n    bash: prompt\n\ncompaction:\n  strategy: context-full\n  thresholdPercent: 80\n\ntheme:\n  dark: titanium\n```\n\nKeep secrets out of committed project config unless your repository policy allows it. Prefer environment variables, stored auth, an auth broker, or an untracked `--config` overlay for credentials.\n\n### One-shot overlays\n\nUse `--config` for a temporary layer that should not persist:\n\n```bash\nomp --config ./local/ci-settings.yml \"check this failure\"\nomp --config ./base.yml --config ./experiment.yml \"try this model\"\n```\n\nOverlay paths are resolved relative to the process working directory (and `~` is expanded). Each overlay must parse as a YAML mapping; a missing file, invalid YAML, or a top-level array/scalar is a hard error — it does **not** silently fall back to lower-precedence settings.\n\n## Path-scoped arrays\n\nTwo array settings — `enabledModels` and `disabledProviders` — accept path-scoped entries in addition to bare strings, so a single global config can behave differently per directory:\n\n```yaml\nenabledModels:\n  - claude-sonnet-4-5            # applies everywhere\n  - path: ~/work/high-context\n    models:\n      - anthropic/claude-opus-4-5\n\ndisabledProviders:\n  - ollama                       # applies everywhere\n  - paths:\n      - ~/projects/sensitive\n      - ~/clients/acme\n    providers:\n      - anthropic\n      - openai\n```\n\nBare string entries apply everywhere. A scoped entry applies when the current working directory **is** the configured path or is **under** it. `~` expands to your home directory and relative paths are resolved before matching.\n\nAccepted **path** keys (any of them, combined): `path`, `paths`, `pathPrefix`, `pathPrefixes`.\n\nAccepted **value** keys:\n\n- `models` (for `enabledModels`) or `providers` (for `disabledProviders`)\n- `values` or `items` (for either setting)\n\nOnly string values are kept; malformed scoped entries are ignored. Path scoping is resolved **after** the layer merge, so it reads the final effective array.\n\n## Provider and source disabling\n\n`disabledProviders` is a single shared id namespace that gates two different subsystems, before any credential check:\n\n| Entry kind | Example ids | Effect |\n|---|---|---|\n| Model providers | `anthropic`, `openai`, `gemini`, `groq`, `ollama`, `openrouter` | Removes those backends from model selection, even when credentials are available. See [Providers](./providers.md). |\n| Discovery sources | `native`, `claude`, `codex`, `gemini`, `github`, `opencode`, `cursor`, `agents-md` | Stops that source from contributing context files, MCP servers, commands, skills, hooks, tools, prompts, or settings. See [Context files](./context-files.md). |\n\nMost provider-control use cases list model provider ids. Disabling the `claude` discovery source is different from disabling the `anthropic` model provider — one stops Claude-format config discovery, the other stops the Anthropic model backend.\n\nBecause arrays replace rather than append, a project that sets `disabledProviders` must list the complete desired set:\n\n```yaml\n# ~/.omp/agent/config.yml\ndisabledProviders:\n  - anthropic\n  - openai\n\n# <repo>/.omp/config.yml — inside this repo ONLY groq is disabled\ndisabledProviders:\n  - groq\n```\n\nThe default is an empty array (nothing disabled). For the two subsystems' provider ids and ordering, see [Providers](./providers.md) and [Context files](./context-files.md).\n\n## Settings catalog\n\nEvery key below is defined in the settings schema; `omp config list` shows the full set with current values. Defaults and enum values are taken from the schema. Settings that accept an env or flag override are noted; those overrides are process-local and not persisted.\n\n### Models\n\n`modelRoles`, `modelTags`, and `cycleOrder` work together to define the models you can switch between. Role values may carry a thinking suffix (`:minimal`, `:low`, `:medium`, `:high`, `:xhigh`).\n\n```yaml\nmodelRoles:\n  default: anthropic/claude-sonnet-4-5\n  smol: openai/gpt-4.1-mini\n  slow: anthropic/claude-opus-4-5:high\n  vision: gemini/gemini-3-pro-preview\n  plan: anthropic/claude-opus-4-5\n\ncycleOrder:\n  - smol\n  - default\n  - slow\n\nmodelProviderOrder:\n  - anthropic\n  - openai\n\nenabledModels:\n  - claude-sonnet-4-5\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `modelRoles` | record | `{}` | Map of role name -> model id. Built-in roles: `default`, `smol`, `slow`, `vision`, `plan`, `designer`, `commit`, `task`. Per-role env/flags: `--model`/`--smol`/`--slow`/`--plan`. |\n| `modelTags` | record | `{}` | Custom role/tag metadata; can introduce additional roles. |\n| `modelProviderOrder` | array | `[]` | Preferred provider order when a model id is ambiguous. |\n| `cycleOrder` | array | `[\"smol\",\"default\",\"slow\"]` | Roles cycled by the model switcher. |\n| `enabledModels` | array | `[]` | Allow-list of models; supports [path-scoped entries](#path-scoped-arrays). Empty means all available models. |\n| `disabledProviders` | array | `[]` | Disabled model/discovery providers; supports path-scoped entries. See [above](#provider-and-source-disabling). |\n| `includeModelInPrompt` | boolean | `true` | Include the active model name in the system prompt. |\n\nSee [Models](./models.md) for the `models.yml` schema and custom-provider definitions.\n\n### Thinking\n\n```yaml\ndefaultThinkingLevel: high\nhideThinkingBlock: false\nthinkingBudgets:\n  minimal: 1024\n  low: 2048\n  medium: 8192\n  high: 16384\n  xhigh: 32768\n```\n\n| Key | Type | Default | Values |\n|---|---|---|---|\n| `defaultThinkingLevel` | enum | `high` | `minimal`, `low`, `medium`, `high`, `xhigh`, `auto`. Override per run with `--thinking`. |\n| `hideThinkingBlock` | boolean | `false` | Hide thinking blocks in output. `--hide-thinking` sets it for the run (display only). |\n| `thinkingBudgets.minimal` | number | `1024` | Token budget for the `minimal` level. |\n| `thinkingBudgets.low` | number | `2048` | Token budget for `low`. |\n| `thinkingBudgets.medium` | number | `8192` | Token budget for `medium`. |\n| `thinkingBudgets.high` | number | `16384` | Token budget for `high`. |\n| `thinkingBudgets.xhigh` | number | `32768` | Token budget for `xhigh`. |\n\n### Sampling\n\nA value of `-1` means \"use the provider/model default\" — `omp` does not send that parameter.\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `temperature` | number | `-1` | Sampling temperature. |\n| `topP` | number | `-1` | Nucleus sampling. |\n| `topK` | number | `-1` | Top-K sampling. |\n| `minP` | number | `-1` | Minimum-probability cutoff. |\n| `presencePenalty` | number | `-1` | Presence penalty. |\n| `repetitionPenalty` | number | `-1` | Repetition penalty. |\n| `serviceTier` | enum | `none` | `none`, `auto`, `default`, `flex`, `scale`, `priority`, `openai-only`, `claude-only`. |\n| `personality` | enum | `default` | `default`, `friendly`, `pragmatic`, `none`. |\n\n### Retry and fallback\n\n```yaml\nretry:\n  enabled: true\n  maxRetries: 10\n  baseDelayMs: 500\n  maxDelayMs: 300000\n  modelFallback: true\n  fallbackRevertPolicy: cooldown-expiry\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `retry.enabled` | boolean | `true` | Retry transient provider errors. |\n| `retry.maxRetries` | number | `10` | Max retries per request. |\n| `retry.baseDelayMs` | number | `500` | Initial backoff. |\n| `retry.maxDelayMs` | number | `300000` | Backoff ceiling (5 min). |\n| `retry.modelFallback` | boolean | `true` | Fall back to another model when one is unavailable. |\n| `retry.fallbackChains` | record | `{}` | Per-model fallback chains. |\n| `retry.fallbackRevertPolicy` | enum | `cooldown-expiry` | `cooldown-expiry`, `never`. |\n\n### Tools and approvals\n\n```yaml\ntools:\n  approvalMode: yolo          # default\n  approval:\n    bash: prompt\n    edit: allow\n  discoveryMode: auto\n  maxTimeout: 0\n  intentTracing: true\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `tools.approvalMode` | enum | `yolo` | `always-ask` (auto-approve read-only), `write` (auto-approve read + workspace-write), `yolo` (auto-approve all tiers). `--approval-mode` and `--auto-approve`/`--yolo` override per run. |\n| `tools.approval` | record | `{}` | Per-tool policy keyed by tool name; each value is `allow`, `deny`, or `prompt`. e.g. `omp config set tools.approval '{\"bash\":\"prompt\"}'`. |\n| `tools.discoveryMode` | enum | `auto` | `auto`, `off`, `mcp-only`, `all`. Controls dynamic tool discovery. |\n| `tools.essentialOverride` | array | `[]` | Tool names kept available even when tools are narrowed. |\n| `tools.maxTimeout` | number | `0` | Max tool runtime in seconds; `0` = no cap. |\n| `tools.intentTracing` | boolean | `true` | Record per-call intent strings. |\n| `tools.outputMaxColumns` | number | `768` | Per-line byte cap for streaming output; `0` disables. |\n| `tools.artifactSpillThreshold` | number | `50` | KB of tool output above which output spills to an artifact. |\n| `tools.artifactHeadBytes` | number | `20` | KB of head kept inline on spill; `0` = tail-only. |\n| `tools.artifactTailBytes` | number | `20` | KB of tail kept inline on spill. |\n| `tools.artifactTailLines` | number | `500` | Max tail lines kept inline on spill. |\n\nIndividual built-in tools are toggled by their own keys, e.g. `bash.enabled`, `eval.py`, `eval.js`, `find.enabled`, `search.enabled`, `fetch.enabled`, `browser.enabled`, `astEdit.enabled`, `astGrep.enabled`, `web_search.enabled`, `inspect_image.enabled`, `renderMermaid.enabled`.\n\n### Shell, eval, and LSP\n\n```yaml\nbash:\n  enabled: true\n  stripTrailingHeadTail: true\n  autoBackground:\n    enabled: false\n    thresholdMs: 60000\n\neval:\n  py: true\n  js: true\n\npython:\n  kernelMode: session       # session, per-call\n  interpreter: \"\"\n\nlsp:\n  enabled: true\n  lazy: true\n  diagnosticsOnWrite: true\n  diagnosticsOnEdit: false\n  formatOnWrite: false\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `bash.enabled` | boolean | `true` | Enable the bash tool. |\n| `bash.stripTrailingHeadTail` | boolean | `true` | Strip trailing head/tail noise from output. |\n| `bash.autoBackground.enabled` | boolean | `false` | Auto-background long-running commands. |\n| `bash.autoBackground.thresholdMs` | number | `60000` | Threshold before auto-backgrounding. |\n| `eval.py` | boolean | `true` | Python eval backend. `PI_PY=0` disables for the process. |\n| `eval.js` | boolean | `true` | JavaScript eval backend. `PI_JS=0` disables for the process. |\n| `python.kernelMode` | enum | `session` | `session` (persistent kernel) or `per-call`. |\n| `python.interpreter` | string | `\"\"` | Path to a Python interpreter; empty = auto-detect. |\n| `lsp.enabled` | boolean | `true` | Language-server integration. `--no-lsp` disables for the run. |\n| `lsp.lazy` | boolean | `true` | Start servers on demand. |\n| `lsp.diagnosticsOnWrite` | boolean | `true` | Run diagnostics after a write. |\n| `lsp.diagnosticsOnEdit` | boolean | `false` | Run diagnostics after an edit. |\n| `lsp.formatOnWrite` | boolean | `false` | Format files on write. |\n| `lsp.diagnosticsDeduplicate` | boolean | `true` | Collapse duplicate diagnostics. |\n| `shellPath` | string | _(unset)_ | Override the shell binary used by bash. |\n\n### Files: editing and reading\n\n```yaml\nedit:\n  mode: hashline            # apply_patch, hashline, patch, replace\n  fuzzyMatch: true\n  fuzzyThreshold: 0.95\n  blockAutoGenerated: true\n\nread:\n  defaultLimit: 300\n  toolResultPreview: false\n  summarize:\n    enabled: true\n    prose: false\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `edit.mode` | enum | `hashline` | `apply_patch`, `hashline`, `patch`, `replace`. |\n| `edit.fuzzyMatch` | boolean | `true` | Allow fuzzy anchor matching. |\n| `edit.fuzzyThreshold` | number | `0.95` | Similarity threshold for fuzzy matching. |\n| `edit.blockAutoGenerated` | boolean | `true` | Refuse to edit generated/lockfile-like files. |\n| `edit.streamingAbort` | boolean | `false` | Abort on streaming edit mismatch. |\n| `read.defaultLimit` | number | `300` | Default line count for `read` without a selector. |\n| `read.summarize.enabled` | boolean | `true` | Structural summaries for code reads. |\n| `read.summarize.prose` | boolean | `false` | Summarize prose files too. |\n| `read.toolResultPreview` | boolean | `false` | Inline preview of tool results. |\n| `readHashLines` | boolean | `true` | Show hashline tags in read output. |\n| `readLineNumbers` | boolean | `false` | Show plain line numbers. |\n\n### Context, compaction, and memory\n\n```yaml\ncontextPromotion:\n  enabled: true\n\ncompaction:\n  enabled: true\n  strategy: context-full     # context-full, handoff, shake, snapcompact, off\n  thresholdPercent: -1       # -1 = default reserve-based behavior\n  thresholdTokens: -1        # fixed token limit when > 0\n  remoteEnabled: true\n\nmemory:\n  backend: off               # off, local, hindsight, mnemopi\n```\n\n| Key | Type | Default | Notes |\n|---|---|---|---|\n| `contextPromotion.enabled` | boolean | `true` | Promote relevant earlier context. |\n| `compaction.enabled` | boolean | `true` | Automatic conversation compaction. |\n| `compaction.strategy` | enum | `context-full` | `context-full`, `handoff`, `shake`, `snapcompact`, `off`. |\n| `compaction.thresholdPercent` | number | `-1` | Percent-of-context trigger; `-1` = reserve-based default. |\n| `compaction.thresholdTokens` | number | `-1` | Fixed token trigger when `> 0`. |\n| `compaction.reserveTokens` | number | `16384` | Tokens reserved for the next turn. |\n| `compaction.keepRecentTokens` | number | `20000` | Recent tokens always preserved. |\n| `compaction.remoteEnabled` | boolean | `true` | Allow remote compaction service. |\n| `compaction.autoContinue` | boolean | `true` | Continue automatically after compaction. |\n| `memory.backend` | enum | `off` | `off`, `local`, `hindsight`, `mnemopi`. Each backend has its own `hindsight.*` / `mnemopi.*` / `memories.*` tuning keys. |\n| `autolearn.enabled` | boolean | `false` | Experimental: after the agent stops, nudge it to capture lessons to memory and create/enhance isolated managed skills under `~/.omp/agent/managed-skills`. Enables the `manage_skill` tool (and `learn` when a memory backend is active). |\n| `autolearn.autoContinue` | boolean | `false` | When `autolearn.enabled`, auto-run one capture turn at stop (uses extra tokens). Off = a passive reminder rides your next turn. |\n| `autolearn.minToolCalls` | number | `5` | Only nudge after a turn that used at least this many tools. |\n\n`compaction` has additional tuning keys (idle compaction, supersede/drop heuristics) visible in `omp config list`. See [Compaction](./compaction.md) for the full strategy reference.\n\n### Appearance and terminal\n\n```yaml\ntheme:\n  dark: titanium\n  light: light\nsymbolPreset: unicode        # unicode, nerd, ascii\ncolorBlindMode: false\n\nstatusLine:\n  preset: default            # default, minimal, compact, full, nerd, ascii, custom\n  separator: powerline-thin\n  transparent: false\n  showHookStatus: true\n\nterminal:\n  showImages: true\nimages:\n  autoResize: true\n  blockImages: false\ntui:\n  hyperlinks: auto           # off, auto, always\n```\n\n| Key | Type | Default | Values |\n|---|---|---|---|\n| `theme.dark` | string | `titanium` | Theme used on a dark terminal background. |\n| `theme.light` | string | `light` | Theme used on a light terminal background. |\n| `symbolPreset` | enum | `unicode` | `unicode`, `nerd`, `ascii`. |\n| `colorBlindMode` | boolean | `false` | Use blue instead of green for diff additions. |\n| `showHardwareCursor` | boolean | `true` | Show the terminal hardware cursor. |\n| `statusLine.preset` | enum | `default` | `default`, `minimal`, `compact`, `full`, `nerd`, `ascii`, `custom`. |\n| `statusLine.separator` | enum | `powerline-thin` | `powerline`, `powerline-thin`, `slash`, `pipe`, `block`, `none`, `ascii`. |\n| `statusLine.sessionAccent` | boolean | `true` | Tint the editor border with the session color. |\n| `statusLine.transparent` | boolean | `false` | Use the terminal background for the status line. |\n| `statusLine.showHookStatus` | boolean | `true` | Show hook status messages. |\n| `terminal.showImages` | boolean | `true` | Render images inline (when the terminal supports it). |\n| `images.autoResize` | boolean | `true` | Resize large images for model compatibility. |\n| `images.blockImages` | boolean | `false` | Never send images to providers. |\n| `tui.hyperlinks` | enum | `auto` | `off`, `auto`, `always`. |\n\nFor a custom status line, set `statusLine.preset: custom` and configure `statusLine.leftSegments`, `statusLine.rightSegments`, and `statusLine.segmentOptions`.\n\n### Interaction\n\n| Key | Type | Default | Values |\n|---|---|---|---|\n| `steeringMode` | enum | `one-at-a-time` | `all`, `one-at-a-time`. How queued steering messages are delivered. |\n| `followUpMode` | enum | `one-at-a-time` | `all`, `one-at-a-time`. |\n| `interruptMode` | enum | `immediate` | `immediate`, `wait`. |\n| `doubleEscapeAction` | enum | `tree` | `branch`, `tree`, `none`. |\n| `autoResume` | boolean | `false` | Auto-resume the most recent session in the cwd. |\n| `ask.timeout` | number | `0` | Seconds before an `ask` prompt times out; `0` = no timeout. (Legacy ms values are migrated to seconds.) |\n| `ask.notify` | enum | `on` | `on`, `off`. |\n\n### Providers and services\n\n```yaml\nproviders:\n  webSearch: auto\n  image: auto\n  fetch: auto\n  tinyModel: online\n  tinyModelDevice: default\n  tinyModelDtype: default\n  openaiWebsockets: auto\n  openrouterVariant: default\n  kimiApiFormat: anthropic\n\nprovider:\n  appendOnlyContext: auto    # auto, on, off\n\nexa:\n  enabled: true\n  enableSearch: true\n  enableResearcher: false\n  enableWebsets: false\n\nsearxng:\n  endpoint: https://search.example.com\n  token: SEARXNG_TOKEN\n```\n\n| Key | Type | Default | Values / notes |\n|---|---|---|---|\n| `providers.webSearch` | enum | `auto` | `auto` plus the configured search providers (`tavily`, `perplexity`, `brave`, `jina`, `kimi`, `anthropic`, `gemini`, `codex`, `zai`, `exa`, `parallel`, `kagi`, `synthetic`, `searxng`). |\n| `providers.image` | enum | `auto` | `auto`, `openai`, `antigravity`, `xai`, `gemini`, `openrouter`. |\n| `providers.fetch` | enum | `auto` | `auto`, `native`, `trafilatura`, `lynx`, `parallel`, `jina`. |\n| `providers.tinyModel` | enum | `online` | `online` or a local model (`lfm2-350m`, `qwen3-0.6b`, `gemma-270m`, `qwen2.5-0.5b`, `lfm2-700m`). |\n| `providers.tinyModelDevice` | enum | `default` | ONNX execution provider for local tiny models. Overridden by `PI_TINY_DEVICE`. |\n| `providers.tinyModelDtype` | enum | `default` | ONNX precision for local tiny models. Overridden by `PI_TINY_DTYPE`. |\n| `providers.openaiWebsockets` | enum | `auto` | `auto`, `off`, `on`. |\n| `providers.openrouterVariant` | enum | `default` | `default`, `nitro`, `floor`, `online`, `exacto`. |\n| `providers.kimiApiFormat` | enum | `anthropic` | `openai`, `anthropic`. |\n| `provider.appendOnlyContext` | enum | `auto` | `auto`, `on`, `off`. |\n| `exa.enabled` | boolean | `true` | Enable Exa integration. |\n| `exa.enableSearch` | boolean | `true` | Exa search. |\n| `exa.enableResearcher` | boolean | `false` | Exa researcher. |\n| `exa.enableWebsets` | boolean | `false` | Exa websets. |\n| `searxng.endpoint` | string | _(unset)_ | SearXNG instance URL. |\n| `searxng.token` | string | _(unset)_ | SearXNG token; also `searxng.basicUsername`/`searxng.basicPassword`/`searxng.categories`/`searxng.language`. |\n| `auth.broker.url` | string | _(unset)_ | Auth-broker URL. Overridden by `OMP_AUTH_BROKER_URL`. |\n| `auth.broker.token` | string | _(unset)_ | Auth-broker token. Overridden by `OMP_AUTH_BROKER_TOKEN`. |\n\nProvider credentials and custom model definitions are configured separately — see [Providers](./providers.md) and [Models](./models.md).\n\n### Other groups\n\n`omp config list` exposes many more grouped settings, including: `task.*` (subagent concurrency, isolation, model overrides), `skills.*` and `commands.*` (discovery toggles), `mcp.*`, `github.*`, `async.*`, `goal.*`, `loop.*`, `todo.*`, `magicKeywords.*`, `ttsr.*` (sticky rules), `display.*`, `startup.*`, `share.*`, `collab.*`, `stt.*`/`tts.*`, `memories.*`/`hindsight.*`/`mnemopi.*` (memory backends), and `bashInterceptor.*`. Each follows the same type/default rules shown above.\n\n## Legacy migration\n\n`omp` migrates older config shapes automatically. None of these require action; they are listed so you know what changes you may see in `config.yml`.\n\n### Startup migration to `config.yml`\n\nWhen `~/.omp/agent/config.yml` does not exist, startup builds it once from legacy sources, then writes the result:\n\n1. `~/.omp/agent/settings.json` (renamed to `settings.json.bak` after a successful migration).\n2. Settings persisted in `agent.db`.\n\nAfter `config.yml` exists, these legacy sources are no longer consulted. The generic config loader also performs `.json` -> `.yml` migration for other config files when only the `.json` form is present.\n\n### Field-level migrations\n\nApplied whenever raw settings are loaded (global, project, overlays, and runtime overrides):\n\n| Old | New |\n|---|---|\n| `queueMode` | `steeringMode` |\n| `ask.timeout` in milliseconds (value `> 1000`) | seconds (divided by 1000) |\n| flat `theme: \"<name>\"` string | `theme.dark` / `theme.light` (slot chosen by luminance; built-in `light`/`dark` are dropped to use defaults) |\n| `task.isolation.enabled: true/false` | `task.isolation.mode: auto/none` |\n| `task.simple` | removed |\n| legacy `task.isolation.mode` (`worktree`, `fuse-overlay`, `fuse-projfs`) | `rcopy`, `overlayfs`, `projfs` |\n| `lastChangelogVersion` | moved to a marker file and stripped from `config.yml` |\n\n## Troubleshooting\n\n### A project setting is not taking effect\n\n- Start `omp` from the directory that contains `.omp/config.yml`. Settings discovery only checks the current working directory's `.omp/`, not ancestor directories.\n- Ensure `.omp/` is non-empty; empty config directories are ignored.\n- Confirm the file is valid YAML and its top level is a mapping.\n- Run `omp config get <key>` from that directory to see the effective value.\n- Remember that `--config` overlays and runtime flags override project config.\n\n### A global array disappeared in a project\n\nArrays replace; they do not append. If a project sets `disabledProviders`, `enabledModels`, `cycleOrder`, `extensions`, or any other array, include the **complete** desired value in the project layer — the global array is fully replaced.\n\n### A provider is still available after editing config\n\n- Check whether you disabled the model provider id (e.g. `anthropic`) or a discovery source id (e.g. `claude`) — they are different namespaces with different effects.\n- Check for a project (or overlay) `disabledProviders` array replacing your global one.\n- Credentials can still come from environment variables, `.env`, OAuth, stored auth, or `models.yml`; disabling a provider blocks selection regardless, but verify you edited the right layer. See [Providers](./providers.md).\n- Restart the session if the model list was already initialized.\n\n### `omp config set` changed the wrong file\n\n`omp config set` and `omp config reset` always write the global `config.yml` under the active agent directory. Run `omp config path` to print it. For project-local settings, edit `<repo>/.omp/config.yml` directly.\n\n### `omp config reset` did not remove my key\n\n`reset` writes the schema **default** value into the global config — it persists the default rather than deleting the key. To stop overriding a project value from global config, delete the key from `~/.omp/agent/config.yml` by hand.\n\n### A `--config` overlay fails at startup\n\n`--config` files are process-local YAML mappings. A missing file, invalid YAML, or a top-level array/scalar is a hard error — it does not silently fall back to lower-precedence settings. Fix the path or contents.\n\n### An environment variable beats my config\n\nSome settings (model roles, eval backends, tiny-model device/precision, auth broker, PTY) are overridable by env vars or CLI flags for per-machine convenience, and those take precedence over `config.yml`. Unset the variable or drop the flag to let the persisted value win. See [Environment overrides](#environment-overrides) and [Environment variables](./environment-variables.md).\n\n### `omp config set <key>` says \"Unknown setting\"\n\nKeys must match a schema path exactly, with no shorthand. Use `theme.dark`, not `theme`. Run `omp config list` to see every valid key.\n",
 	"skills.md": "# Skills\n\nSkills are file-backed capability packs discovered at startup and exposed to the model as:\n\n- lightweight metadata in the system prompt (name + description)\n- on-demand content via the `read` tool against `skill://...`\n- optional interactive `/skill:<name>` commands\n\nThis document covers current runtime behavior in `src/extensibility/skills.ts`, `src/discovery/builtin.ts`, `src/internal-urls/skill-protocol.ts`, and `src/discovery/agents-md.ts`.\n\n## What a skill is in this codebase\n\nA discovered skill is represented as:\n\n- `name`\n- `description`\n- `filePath` (the `SKILL.md` path)\n- `baseDir` (skill directory)\n- source metadata (`provider`, `level`, path)\n\nThe runtime only requires `name` and `path` for validity. In practice, matching quality depends on `description` being meaningful.\n\n## Required layout and SKILL.md expectations\n\n### Directory layout\n\nFor provider-based discovery (native/Claude/Codex/Agents/plugin providers), skills are discovered as **one level under `skills/`**:\n\n- `<skills-root>/<skill-name>/SKILL.md`\n\nNested patterns like `<skills-root>/group/<skill>/SKILL.md` are not discovered by provider loaders.\n\nFor `skills.customDirectories`, scanning uses the same non-recursive layout (`*/SKILL.md`).\n\n```text\nProvider-discovered layout (non-recursive under skills/):\n\n<root>/skills/\n  ├─ postgres/\n  │   └─ SKILL.md      ✅ discovered\n  ├─ pdf/\n  │   └─ SKILL.md      ✅ discovered\n  └─ team/\n      └─ internal/\n          └─ SKILL.md  ❌ not discovered by provider loaders\n\nCustom-directory scanning is also non-recursive, so nested paths are ignored unless you point `customDirectories` at that nested parent.\n```\n\n### `SKILL.md` frontmatter\n\nSupported frontmatter fields on the skill type:\n\n- `name?: string`\n- `description?: string`\n- `globs?: string[]`\n- `alwaysApply?: boolean`\n- `hide?: boolean`\n- `disableModelInvocation?: boolean` (Agent Skills equivalent of `hide`; normalized from kebab-case `disable-model-invocation`)\n- additional keys are preserved as unknown metadata\n\nCurrent runtime behavior:\n\n- `name` defaults to the skill directory name\n- `description` is required for:\n  - native `.omp` provider skill discovery (`requireDescription: true`)\n  - `omp-plugins` extension-package skills and the `github` provider (`.github/skills/`), which also pass `requireDescription: true`\n  - `skills.customDirectories` scans via `scanSkillsFromDir` in `src/discovery/helpers.ts` (non-recursive)\n- the claude/codex/agents/opencode/claude-plugins providers can load skills without description\n\n## Discovery pipeline\n\n`loadSkills()` in `src/extensibility/skills.ts` does two passes:\n\n1. **Capability providers** via `loadCapability(\"skills\")`\n2. **Custom directories** via `scanSkillsFromDir(..., { requireDescription: true })` (one-level directory enumeration)\n\nIf `skills.enabled` is `false`, discovery returns no skills.\n\n### Built-in skill providers and precedence\n\nProvider ordering is priority-first (higher wins), then registration order for ties.\n\nCurrent registered skill providers:\n\n1. `native` (priority 100) — `.omp` user/project skills via `src/discovery/builtin.ts`\n2. `omp-plugins` (priority 90) — `skills/` bundled next to extension packages loaded through `extensions:`, `--extension`/`-e`, or installed plugins under `~/.omp/plugins/node_modules`\n3. `claude` (priority 80)\n4. priority 70 group (in registration order):\n   - `claude-plugins`\n   - `agents`\n   - `codex`\n5. `opencode` (priority 55)\n6. `github` (priority 30) — `.github/skills/<name>/SKILL.md` (GitHub Agent Skills layout, project-only)\n\nDedup key is skill name. First item with a given name wins.\n\n### Source toggles and filtering\n\n`loadSkills()` applies these controls:\n\n- source toggles: `enableCodexUser`, `enableClaudeUser`, `enableClaudeProject`, `enablePiUser`, `enablePiProject`, `enableAgentsUser`, `enableAgentsProject`\n- `disabledExtensions` entries with `skill:<name>`\n- `ignoredSkills` (exclude; glob patterns)\n- `includeSkills` (include allowlist; glob patterns; empty means include all)\n\nFilter order is:\n\n1. not disabled by `disabledExtensions`\n2. source enabled\n3. not ignored\n4. included (if include list present)\n\nThe `agents` provider (`.agent[s]/skills`) is the canonical OMP-native location and has its own `enableAgentsUser`/`enableAgentsProject` toggles — disabling Claude/Codex/Pi does **not** turn it off. For providers without a dedicated toggle (`claude-plugins`, `opencode`, `gemini`, `github`, …), enablement falls back to: enabled if **any** named source toggle is enabled.\n\n### Collision and duplicate handling\n\n- Capability dedup already keeps first skill per name (highest-precedence provider)\n- `extensibility/skills.ts` additionally:\n  - de-duplicates identical files by `realpath` (symlink-safe)\n  - emits collision warnings when a later skill name conflicts\n  - keeps the convenience `loadSkillsFromDir({ dir, source })` API as a thin adapter over `scanSkillsFromDir`\n- Custom-directory skills are merged after provider skills and follow the same collision behavior\n\n## Runtime usage behavior\n\n### System prompt exposure\n\nSystem prompt construction (`src/system-prompt.ts`) uses discovered skills as follows:\n\n- if `read` tool is available:\n  - include discovered skills list in prompt, excluding skills with `hide: true`\n- otherwise:\n  - omit discovered list\n\n`hide: true` does not disable the skill. Hidden skills are still loaded and remain reachable through `skill://<name>` and `/skill:<name>` when skill commands are enabled.\n\nTask tool subagents receive the session's discovered/provided skills list via normal session creation; there is no per-task skill pinning override.\n\n### Interactive `/skill:<name>` commands\n\nIf `skills.enableSkillCommands` is true, interactive mode registers one slash command per discovered skill.\n\n`/skill:<name> [args]` behavior:\n\n- reads the skill file directly from `filePath`\n- strips frontmatter\n- injects skill body as a custom message\n- delivery mode follows the **submission keybinding**:\n  - **Enter** → invokes the skill on the `steer` queue while streaming (matches free-text Enter, which also steers), or as a normal idle prompt when the agent is not streaming\n  - **Ctrl+Enter** (`app.message.followUp`) → invokes the skill on the `followUp` queue while streaming, or as a normal idle prompt when the agent is not streaming\n- appends metadata (`Skill: <path>`, optional `User: <args>`)\n\nThere is no flag, mode-selector, or frontmatter knob to override this — the keybinding _is_ the choice, identical to how free text is routed during streaming (`input-controller.ts:436-442` for Enter, `input-controller.ts:770-775` for Ctrl+Enter; both dispatch through `#invokeSkillCommand`).\n\n## `skill://` URL behavior\n\n`src/internal-urls/skill-protocol.ts` supports:\n\n- `skill://<name>` → resolves to that skill's `SKILL.md`\n- `skill://<name>/<relative-path>` → resolves inside that skill directory\n\n```text\nskill:// URL resolution\n\nskill://pdf\n  -> <pdf-base>/SKILL.md\n\nskill://pdf/references/tables.md\n  -> <pdf-base>/references/tables.md\n\nGuards:\n- reject absolute paths\n- reject `..` traversal\n- reject any resolved path escaping <pdf-base>\n```\n\nResolution details:\n\n- skill name must match exactly\n- relative paths are URL-decoded\n- absolute paths are rejected\n- path traversal (`..`) is rejected\n- resolved path must remain within `baseDir`\n- missing files return an explicit `File not found` error\n\nContent type:\n\n- `.md` => `text/markdown`\n- everything else => `text/plain`\n\nNo fallback search is performed for missing assets.\n\n## Skills vs AGENTS.md, commands, tools, hooks\n\n### Skills vs AGENTS.md\n\n- **Skills**: named, optional capability packs selected by task context or explicitly requested\n- **AGENTS.md/context files**: persistent instruction files loaded as context-file capability and merged by level/depth rules\n\n`src/discovery/agents-md.ts` specifically walks ancestor directories from `cwd` to discover standalone `AGENTS.md` files (stopping at the repo root, or home when no repo root is known), skipping files whose containing directory name starts with a dot.\n\n### Skills vs slash commands\n\n- **Skills**: model-readable knowledge/workflow content\n- **Slash commands**: user-invoked command entry points\n- `/skill:<name>` is a convenience wrapper that injects skill text; it does not change skill discovery semantics\n\n### Skills vs custom tools\n\n- **Skills**: documentation/workflow content loaded through prompt context and `read`\n- **Custom tools**: executable tool APIs callable by the model with schemas and runtime side effects\n\n### Skills vs hooks\n\n- **Skills**: passive content\n- **Hooks**: event-driven runtime interceptors that can block/modify behavior during execution\n\n## Practical authoring guidance tied to discovery logic\n\n- Put each skill in its own directory: `<skills-root>/<skill-name>/SKILL.md`\n- Always include explicit `name` and `description` frontmatter\n- Keep referenced assets under the same skill directory and access with `skill://<name>/...`\n- For nested taxonomy (`team/domain/skill`), point `skills.customDirectories` to the nested parent directory; scanning itself remains non-recursive\n- Avoid duplicate skill names across sources; first match wins by provider precedence\n",
 	"skills/authoring-extensions.md": "---\nname: authoring-extensions\ndescription: Use when creating a new omp extension. Covers ExtensionAPI, factory signature, tool/command/event registration, and local-dev testing.\n---\n\n# Authoring Extensions\n\nExtensions are the primary way to add capabilities to `oh-my-pi`. A single extension module can register tools the LLM can call, slash commands users can invoke, and event handlers that run throughout the session lifecycle — all from one TypeScript file.\n\n## Minimum viable extension\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function (pi: ExtensionAPI) {\n  pi.on(\"session_start\", async (_event, ctx) => {\n    ctx.ui.notify(\"My extension loaded!\", \"info\");\n  });\n}\n```\n\nThat is a working extension. Drop it into `~/.omp/agent/extensions/hello.ts` and restart omp to see the notification.\n\n## Full example\n\nThe following extension registers a slash command, a tool, and a session-start hook:\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function myExtension(pi: ExtensionAPI) {\n  const z = pi.zod;\n\n  // Runs once when the session loads\n  pi.on(\"session_start\", async (_event, ctx) => {\n    ctx.ui.notify(`Session ready in ${ctx.cwd}`, \"info\");\n  });\n\n  // Slash command: /greet\n  pi.registerCommand(\"greet\", {\n    description: \"Send a greeting into the conversation\",\n    handler: async (args, ctx) => {\n      const name = args.trim() || \"world\";\n      pi.sendMessage(\n        {\n          customType: \"greeting\",\n          content: `Hello, ${name}!`,\n          display: true,\n          attribution: \"user\",\n        },\n        { triggerTurn: false }\n      );\n      ctx.ui.notify(`Greeted ${name}`, \"info\");\n    },\n  });\n\n  // LLM-callable tool\n  pi.registerTool({\n    name: \"word_count\",\n    label: \"Word Count\",\n    description: \"Count the words in a string\",\n    parameters: z.object({\n      text: z.string().describe(\"Text to count\"),\n    }),\n    async execute(_id, params, _signal, _onUpdate, _ctx) {\n      const count = params.text.split(/\\s+/).filter(Boolean).length;\n      return {\n        content: [{ type: \"text\", text: String(count) }],\n        details: { count },\n      };\n    },\n  });\n}\n```\n\n## Discovery paths\n\nomp loads extension modules from these sources:\n\n1. Native `.omp` locations discovered through the capability system:\n   - `<cwd>/.omp/extensions/`\n   - `~/.omp/agent/extensions/`\n   - legacy extension paths listed in `.omp/settings.json#extensions` or `~/.omp/agent/settings.json#extensions`\n2. Installed plugins under `~/.omp/plugins/node_modules` (`omp plugin install` npm/git specs, or `omp plugin link`) via their `omp.extensions`/`pi.extensions` manifests. Marketplace cache installs do not feed extension modules — they surface skills/commands/hooks/tools/MCP only.\n3. Explicit configured paths passed by the CLI (`omp --extension ./my-ext.ts`, also `-e`; `--hook` is treated as an alias) and by the `extensions:` setting in config.\n\nThe runtime de-duplicates by resolved absolute path — first seen wins.\n\nWhen a path points to a directory, omp resolves the entry point in this order:\n\n1. `package.json` with `omp.extensions` (or legacy `pi.extensions`) field\n2. `index.ts`\n3. `index.js`\n\nWhen scanning an `extensions/` directory, omp also loads direct `*.ts`/`*.js` files and one-level subdirectories that have `index.ts`, `index.js`, or a manifest.\n\nExtension packages can also bundle sibling capability directories. When a package is loaded through `extensions:` or `--extension`/`-e`, the `omp-plugins` provider discovers its `skills/`, `hooks/pre|post/`, `tools/`, `commands/`, `rules/`, `prompts/`, and `.mcp.json`.\n\n## package.json manifest\n\nTo package an extension as an installable plugin, add an `omp` field to `package.json`:\n\n```json\n{\n  \"name\": \"my-omp-extension\",\n  \"omp\": {\n    \"extensions\": [\"./src/main.ts\"]\n  }\n}\n```\n\nThe legacy `pi` key is also accepted for backwards compatibility:\n\n```json\n{\n  \"pi\": {\n    \"extensions\": [\"./index.ts\"]\n  }\n}\n```\n\nMultiple entry points are supported:\n\n```json\n{\n  \"omp\": {\n    \"extensions\": [\"./src/safety.ts\", \"./src/tools.ts\"]\n  }\n}\n```\n\n## Registering commands\n\n```ts\npi.registerCommand(\"my-cmd\", {\n  description: \"What the command does\",\n  handler: async (args, ctx) => {\n    // args: everything the user typed after /my-cmd\n    // ctx: ExtensionCommandContext — includes ctx.ui, ctx.cwd, session controls\n    ctx.ui.notify(\"Running!\", \"info\");\n    await ctx.waitForIdle();\n    await ctx.newSession();\n  },\n});\n```\n\n`ExtensionCommandContext` session-control methods (safe to call from commands only):\n\n| Method | Effect |\n|---|---|\n| `waitForIdle()` | Wait for the agent to finish streaming |\n| `newSession(opts?)` | Open a fresh session |\n| `switchSession(path)` | Switch to an existing session file |\n| `branch(entryId)` | Fork from a specific history entry |\n| `navigateTree(id, opts?)` | Jump to a different point in the session tree |\n| `reload()` | Reload the session runtime |\n| `compact(opts?)` | Compact the current context |\n\n## Registering tools\n\nTools are called by the LLM. Parameters use [Zod](https://zod.dev) schemas, available at `pi.zod`:\n\n```ts\nconst z = pi.zod;\n\npi.registerTool({\n  name: \"search_notes\",           // snake_case, unique\n  label: \"Search Notes\",          // human-readable label for TUI\n  description: \"Full-text search through project notes\",\n  parameters: z.object({\n    query: z.string().describe(\"Search query\"),\n    limit: z.number().default(10).describe(\"Max results\").optional(),\n  }),\n  async execute(toolCallId, params, signal, onUpdate, ctx) {\n    if (signal?.aborted) {\n      return { content: [{ type: \"text\", text: \"Cancelled\" }] };\n    }\n    onUpdate?.({ content: [{ type: \"text\", text: \"Searching...\" }] });\n    // ... do work ...\n    return {\n      content: [{ type: \"text\", text: `Found N results for \"${params.query}\"` }],\n      details: { query: params.query, count: 0 },\n    };\n  },\n});\n```\n\n## Subscribing to events\n\n```ts\npi.on(\"tool_call\", async (event, ctx) => {\n  // event.toolName, event.input, event.toolCallId\n  if (event.toolName !== \"bash\") return;\n\n  const command = String((event.input as { command?: unknown }).command ?? \"\");\n  if (command.includes(\"rm -rf /\")) {\n    return { block: true, reason: \"Blocked by safety policy\" };\n  }\n});\n\npi.on(\"turn_end\", async (_event, ctx) => {\n  ctx.ui.setStatus(\"tokens\", `~${ctx.getContextUsage()?.tokens ?? \"?\"} tokens`);\n});\n```\n\nFull event catalog: see [hooks authoring guide](./authoring-hooks.md).\n\n## Extension vs hook — when to use which\n\n| Need | Use |\n|---|---|\n| Tools + commands + events in one module | **Extension** (`ExtensionAPI`) |\n| Pure event interception (policy, redaction) | **Extension** or **Hook** (both work; extension is preferred) |\n| Legacy hook module already exists | **Hook** (`HookAPI` from `@oh-my-pi/pi-coding-agent/extensibility/hooks`) |\n| Registering provider / custom message renderer | **Extension only** |\n| Shipping as a marketplace plugin | **Extension** (use `package.json` manifest) |\n\nExtensions are a strict superset of hooks. New authoring should use `ExtensionAPI`.\n\n## Debugging\n\nStart omp with `--log-level debug` to see extension load diagnostics:\n\n```\nomp --log-level debug\n```\n\nFailed extension loads are logged with their path and error. Loaded extensions may also emit their own debug logs via `pi.logger`.\n\nTo temporarily disable a specific extension module by name without removing the file:\n\n```yaml\n# ~/.omp/agent/config.yml\ndisabledExtensions:\n  - extension-module:my-ext\n```\n\nThe derived name is the filename stem (or directory name for `index.ts`-style entries): `/path/to/my-ext.ts` → `my-ext`.\n\n## Important constraints\n\n- **Do not call runtime actions during load.** Methods like `pi.sendMessage()` throw `ExtensionRuntimeNotInitializedError` if called synchronously during module evaluation (before a session is active). Register handlers/tools/commands during load; perform runtime actions only from event handlers, tools, or commands.\n- **`tool_call` errors are fail-closed.** If a `tool_call` handler throws, the tool is blocked.\n- **Command names must not clash with built-ins.** Conflicts are skipped with a diagnostic log.\n- **Reserved shortcuts are ignored** (`ctrl+c`, `ctrl+d`, `ctrl+z`, `ctrl+k`, `ctrl+p`, `ctrl+l`, `ctrl+o`, `ctrl+t`, `ctrl+g`, `ctrl+q`, `alt+m`, `shift+tab`, `shift+ctrl+p`, `alt+enter`, `escape`, `enter`).\n\n## Further reading\n\n- `docs/extensions.md` — runtime internals and full API surface reference\n- `docs/extension-loading.md` — detailed path resolution rules\n- `docs/hooks.md` — hook subsystem internals\n- `docs/skills/examples/hello-extension/` — complete working example\n",
 	"skills/authoring-hooks.md": "---\nname: authoring-hooks\ndescription: Use when creating a new omp hook. Covers HookAPI, event catalog, blocking/overriding tool calls, and context modification.\n---\n\n# Authoring Hooks\n\nHooks are event-driven interceptors that run alongside the agent loop. They are best used for cross-cutting concerns: safety policy, secret redaction, context pruning, audit logging. A hook module registers handlers via `pi.on(event, handler)` and can block tool execution, override tool output, or rewrite the message context before each LLM call.\n\n> **Relationship to extensions:** The hook subsystem (`HookAPI`) is the legacy API. The extension runner now handles everything hooks can do plus more. `ExtensionAPI` supports the hook event model plus extension-only events. Use `ExtensionAPI` for new work; use `HookAPI` only if you are maintaining an existing hook module.\n\n## Factory signature\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function myHook(omp: HookAPI): void {\n  omp.on(\"tool_call\", async (event, ctx) => {\n    // intercept every tool call\n  });\n}\n```\n\nThe default export must be a plain function (not async, not a class). It receives a `HookAPI` instance and must register all handlers synchronously during execution.\n\nAlternatively, using `ExtensionAPI` (preferred):\n\n```ts\nimport type { ExtensionAPI } from \"@oh-my-pi/pi-coding-agent\";\n\nexport default function myExtension(pi: ExtensionAPI): void {\n  pi.on(\"tool_call\", async (event, ctx) => { /* ... */ });\n}\n```\n\n## Event catalog\n\n### Tool lifecycle\n\n| Event | Fires | Can return |\n|---|---|---|\n| `tool_call` | Before every tool execution | `{ block?: boolean; reason?: string }` |\n| `tool_result` | After every tool execution | `{ content?; details?; isError?: boolean }` |\n\n### Session lifecycle\n\n| Event | Fires | Can return |\n|---|---|---|\n| `session_start` | On initial session load | — |\n| `session_before_switch` | Before session switch | `{ cancel?: boolean }` |\n| `session_switch` | After session switch | — |\n| `session_before_branch` | Before session branch | `{ cancel?: boolean; skipConversationRestore?: boolean }` |\n| `session_branch` | After session branch | — |\n| `session_before_compact` | Before compaction | `{ cancel?: boolean; compaction?: CompactionResult }` |\n| `session.compacting` | During compaction (inject context) | `{ context?: string[]; prompt?: string; preserveData?: Record<string, unknown> }` |\n| `session_compact` | After compaction | — |\n| `session_before_tree` | Before tree navigation | `{ cancel?: boolean; summary?: { summary: string; details?: unknown } }` |\n| `session_tree` | After tree navigation | — |\n| `session_shutdown` | On session shutdown | — |\n\n### Agent/turn lifecycle\n\n| Event | Fires | Can return |\n|---|---|---|\n| `before_agent_start` | Before agent starts a turn | `{ message?: { customType; content; display; details; attribution? } }` |\n| `agent_start` | Agent streaming starts | — |\n| `agent_end` | Agent streaming ends | — |\n| `turn_start` | Start of a user→agent turn | — |\n| `turn_end` | End of a user→agent turn | — |\n| `context` | Before each LLM API call | `{ messages?: Message[] }` |\n| `auto_compaction_start` | Auto-compaction begins | — |\n| `auto_compaction_end` | Auto-compaction ends | — |\n| `auto_retry_start` | Auto-retry begins | — |\n| `auto_retry_end` | Auto-retry ends | — |\n| `ttsr_triggered` | TTSR (too-short response) triggered | — |\n| `todo_reminder` | Todo reminder fires | — |\n\nExtension-only events such as `tool_execution_start`, `tool_execution_update`, `tool_execution_end`, `input`, `user_bash`, and `user_python` require `ExtensionAPI`.\n\n## Pre-tool blocking contract\n\nReturn `{ block: true, reason: \"...\" }` from a `tool_call` handler to prevent execution:\n\n```ts\nomp.on(\"tool_call\", async (event, ctx) => {\n  if (event.toolName === \"bash\") {\n    const cmd = String(event.input.command ?? \"\");\n    if (/\\brm\\s+-rf\\s+\\//.test(cmd)) {\n      return { block: true, reason: \"Refusing to delete root filesystem\" };\n    }\n  }\n});\n```\n\nContract:\n\n- If **any** handler returns `{ block: true }`, execution stops immediately.\n- `reason` is returned to the LLM as the tool error text.\n- If a handler **throws**, the tool is also blocked (fail-closed).\n- Last non-blocking return wins for non-blocking results; first `block: true` short-circuits.\n\n## Post-tool override contract\n\nReturn `{ content, details, isError }` from a `tool_result` handler to patch what the LLM sees:\n\n```ts\nomp.on(\"tool_result\", async (event, ctx) => {\n  if (event.toolName === \"read\" && !event.isError) {\n    const redacted = event.content.map(chunk => {\n      if (chunk.type !== \"text\") return chunk;\n      return {\n        ...chunk,\n        text: chunk.text.replace(/(?:sk|pk)-[a-zA-Z0-9]{20,}/g, \"[REDACTED_API_KEY]\"),\n      };\n    });\n    return { content: redacted };\n  }\n});\n```\n\nContract:\n\n- Handlers run in registration order. For `HookAPI`, each handler receives the original tool result event, and the last returned override wins.\n- `content` replaces the full content array for the LLM.\n- `details` replaces the structured details object.\n- `isError` exists on the shared result type, but `HookToolWrapper` does not propagate it into a successful tool result; on a tool failure, the original error is rethrown after handlers complete.\n- On a tool failure, `tool_result` is still emitted with `isError: true`.\n\n## Context modification contract\n\nReturn `{ messages: [...] }` from a `context` handler to rewrite the message list before each LLM API call:\n\n```ts\nomp.on(\"context\", async (event, ctx) => {\n  // Remove debug-only custom messages from LLM context\n  const filtered = event.messages.filter(\n    msg => !(msg.role === \"custom\" && msg.customType === \"debug-only\")\n  );\n  return { messages: filtered };\n});\n```\n\nContract:\n\n- `event.messages` is the current accumulated list.\n- Handlers run in order; each receives the output of the previous handler.\n- Return `undefined` (or nothing) to pass messages through unmodified.\n\n## Three complete examples\n\n### 1. rm-rf blocker\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function rmRfBlocker(omp: HookAPI): void {\n  omp.on(\"tool_call\", async (event, ctx) => {\n    if (event.toolName !== \"bash\") return;\n\n    const cmd = String(event.input.command ?? \"\");\n    if (!/\\brm\\s+-rf\\s+\\//.test(cmd)) return;\n\n    // Allow if user explicitly confirms (interactive mode only)\n    if (ctx.hasUI) {\n      const allow = await ctx.ui.confirm(\n        \"Dangerous command\",\n        `This command deletes from root:\\n${cmd}\\n\\nProceed?`\n      );\n      if (allow) return;\n    }\n\n    return { block: true, reason: \"rm -rf / blocked by safety policy\" };\n  });\n}\n```\n\n### 2. API-key redactor\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\n// Common API-key shapes. Not exhaustive — providers using bespoke formats\n// (Anthropic `sk-ant-…`, JWT-style bearers, gateway-specific prefixes, etc.)\n// need their own entries.\nconst SECRET_PATTERNS = [\n  /\\b(sk|pk)-[a-zA-Z0-9]{20,}\\b/g,\n  /\\bAKIA[A-Z0-9]{16}\\b/g,\n  /\\bghp_[a-zA-Z0-9]{36}\\b/g,\n  // Zhipu / GLM Coding Plan: `<id>.<secret>` (no `sk-` prefix).\n  /\\b[a-zA-Z0-9]{16,}\\.[a-zA-Z0-9]{16,}\\b/g,\n  /\\b[a-zA-Z0-9_-]{20,}\\s*=\\s*[\"']?[a-zA-Z0-9._/+=-]{20,}[\"']?/g,\n];\n\nexport default function apiKeyRedactor(omp: HookAPI): void {\n  omp.on(\"tool_result\", async (event) => {\n    if (event.isError) return;\n\n    let changed = false;\n    const redacted = event.content.map(chunk => {\n      if (chunk.type !== \"text\") return chunk;\n      let text = chunk.text;\n      for (const pattern of SECRET_PATTERNS) {\n        const next = text.replace(pattern, \"[REDACTED]\");\n        if (next !== text) { changed = true; text = next; }\n      }\n      return { ...chunk, text };\n    });\n\n    if (changed) return { content: redacted };\n  });\n}\n```\n\n### 3. Context filter\n\n```ts\nimport type { HookAPI } from \"@oh-my-pi/pi-coding-agent/extensibility/hooks\";\n\nexport default function contextFilter(omp: HookAPI): void {\n  omp.on(\"context\", async (event) => {\n    const MAX_TOOL_OUTPUT_CHARS = 8_000;\n\n    const trimmed = event.messages.map(msg => {\n      // Truncate very large tool results to keep context manageable\n      if (msg.role !== \"tool\") return msg;\n      const content = msg.content.map(chunk => {\n        if (chunk.type !== \"text\" || chunk.text.length <= MAX_TOOL_OUTPUT_CHARS) return chunk;\n        return {\n          ...chunk,\n          text: chunk.text.slice(0, MAX_TOOL_OUTPUT_CHARS) + \"\\n[... truncated by context-filter hook]\",\n        };\n      });\n      return { ...msg, content };\n    });\n\n    return { messages: trimmed };\n  });\n}\n```\n\n## UI methods in hook context\n\n`ctx.ui` is a `HookUIContext`. Available methods:\n\n| Method | Description |\n|---|---|\n| `notify(message, type?)` | Show an in-app notification |\n| `setStatus(key, text)` | Set footer status text (keyed, sorted by key) |\n| `select(title, options)` | Show a selection dialog |\n| `confirm(title, message)` | Show a yes/no dialog |\n| `input(title, placeholder?)` | Show a text input dialog |\n| `editor(title, prefill?, { signal }?, { promptStyle }?)` | Show a multi-line editor |\n| `setEditorText(text)` | Set the input editor content |\n| `getEditorText()` | Get current input editor content |\n| `custom(factory)` | Render a custom TUI component |\n| `theme` | Current theme object |\n\nPass `{ promptStyle: true }` as the fourth argument when Enter should submit and Shift+Enter should insert a newline. The default hook editor behavior keeps Enter as newline and Ctrl+Enter as submit.\n\n`ctx.hasUI` is `false` in headless/print/subagent mode — always guard interactive calls.\n\n## Further reading\n\n- `docs/hooks.md` — hook subsystem internals, ordering rules, error propagation\n- `docs/extensions.md` — `ExtensionAPI` (superset of `HookAPI`)\n- `docs/skills/examples/safety-hook/` — complete working example\n",
@@ -81,12 +81,12 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"tools/checkpoint.md": "# checkpoint\n\n> Mark the current top-level conversation state so later `rewind` can collapse exploratory context into a report.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/checkpoint.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/checkpoint.md`\n- Key collaborators:\n  - `packages/coding-agent/src/session/agent-session.ts` — captures the active checkpoint after tool success.\n  - `packages/coding-agent/src/session/session-manager.ts` — persists the normal session entry stream; not the active checkpoint marker.\n  - `packages/coding-agent/src/tools/index.ts` — registers the tool and gates it behind `checkpoint.enabled`.\n  - `packages/coding-agent/src/config/settings-schema.ts` — defines the disabled-by-default feature flag.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `goal` | `string` | Yes | Investigation goal. Required by the schema and echoed in the tool result. |\n\n## Outputs\nThe tool returns a single text result plus structured details:\n\n- text body:\n  - `Checkpoint created.`\n  - `Goal: <goal>`\n  - `Run your investigation, then call rewind with a concise report.`\n- `details`:\n  - `goal: string`\n  - `startedAt: string` — ISO timestamp created inside `CheckpointTool.execute()`\n\nNo checkpoint ID, artifact URI, job handle, file path, or restore token is returned.\n\n## Flow\n1. `CheckpointTool.createIf()` in `packages/coding-agent/src/tools/checkpoint.ts` returns `null` for subagents by checking `session.taskDepth`; only top-level sessions can see the tool.\n2. `CheckpointTool.execute()` rejects subagent calls again with `ToolError(\"Checkpoint not available in subagents.\")`.\n3. It rejects nested checkpoints with `ToolError(\"Checkpoint already active.\")` when `session.getCheckpointState?.()` is already set.\n4. It creates `startedAt = new Date().toISOString()` and returns a normal `toolResult()` payload. The tool itself does not persist anything.\n5. On the later `tool_execution_end` event, `AgentSession` in `packages/coding-agent/src/session/agent-session.ts` detects successful `checkpoint` execution and captures three in-memory fields:\n   - `checkpointMessageCount` — current `agent.state.messages.length`, after the checkpoint tool result has already been appended\n   - `checkpointEntryId` — `sessionManager.getEntries().at(-1)?.id ?? null`, i.e. the last persisted session entry ID at checkpoint time\n   - `startedAt` — copied from tool details or regenerated\n6. `AgentSession` stores that object in its private `#checkpointState` field and clears `#pendingRewindReport`.\n\n## Side Effects\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - Sets `AgentSession.#checkpointState` in memory.\n  - Records the checkpoint boundary as a message count plus a session entry ID.\n  - Enables the later yield guard: if a checkpoint is active and no rewind report is pending, `#enforceRewindBeforeYield()` injects a developer-role warning and schedules another turn.\n- User-visible prompts / interactive UI\n  - The tool result tells the model to call `rewind` after the investigation.\n  - If the agent tries to `yield` first, `AgentSession` injects:\n\n```text\n<system-warning>\nYou are in an active checkpoint. You MUST call rewind with your investigation findings before yielding. Do NOT yield without completing the checkpoint.\n</system-warning>\n```\n\n## Limits & Caps\n- Availability is gated by `checkpoint.enabled`, default `false`, in `packages/coding-agent/src/config/settings-schema.ts`.\n- The tool is registered as discoverable in `packages/coding-agent/src/tools/index.ts`.\n- Only one active checkpoint is allowed per top-level session.\n- Checkpoint state is not persisted as a dedicated session entry. If the process exits, a resumed session can reload the conversation history, but not the live `#checkpointState` guard.\n- Session persistence still applies to the ordinary checkpoint tool call message. Global session persistence truncation is `MAX_PERSIST_CHARS = 500_000` in `packages/coding-agent/src/session/session-manager.ts`.\n\n## Errors\n- `ToolError(\"Checkpoint not available in subagents.\")` — thrown for subagent sessions.\n- `ToolError(\"Checkpoint already active.\")` — thrown when a prior checkpoint has not been rewound or cleared.\n- The tool body has no local `try/catch`; unexpected exceptions propagate.\n\n## Notes\n- Despite the summary string `Create a git-based checkpoint to save and restore session state`, the implementation does not call git and does not snapshot filesystem state.\n- Captured state is conversation/session metadata only:\n  - in-memory message count\n  - session entry ID in the session tree\n  - timestamp\n- Not captured:\n  - working tree contents\n  - staged changes\n  - artifacts\n  - blob-store contents\n  - SQLite history rows from `packages/coding-agent/src/session/history-storage.ts`\n  - auth or agent records from `packages/coding-agent/src/session/agent-storage.ts`\n",
 	"tools/debug.md": "# debug\n\n> Drive one DAP debug session; adjacent debug UI code reuses the same subsystem for logs, raw SSE capture, reports, profiling, and system diagnostics.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/debug.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/debug.md`\n- Key collaborators:\n  - `packages/coding-agent/src/dap/session.ts` — session lifecycle, breakpoint/state cache\n  - `packages/coding-agent/src/dap/client.ts` — adapter process/socket transport, DAP message loop\n  - `packages/coding-agent/src/dap/config.ts` — adapter resolution and auto-selection\n  - `packages/coding-agent/src/dap/defaults.json` — built-in adapter definitions\n  - `packages/coding-agent/src/dap/types.ts` — request/response/capability shapes\n  - `packages/coding-agent/src/tools/tool-timeouts.ts` — per-tool timeout clamp\n  - `packages/coding-agent/src/debug/index.ts` — interactive debug selector menu\n  - `packages/coding-agent/src/debug/log-viewer.ts` — recent-log TUI viewer\n  - `packages/coding-agent/src/debug/raw-sse.ts` — raw SSE TUI viewer\n  - `packages/coding-agent/src/debug/raw-sse-buffer.ts` — bounded SSE capture buffer\n  - `packages/coding-agent/src/debug/profiler.ts` — CPU/heap profiling helpers\n  - `packages/coding-agent/src/debug/report-bundle.ts` — `.tar.gz` report bundling, log source, cache cleanup\n  - `packages/coding-agent/src/debug/system-info.ts` — system snapshot collection and env redaction\n  - `packages/coding-agent/src/debug/terminal-info.ts` — terminal state collection/formatting\n  - `packages/coding-agent/src/debug/protocol-probe.ts` — terminal protocol probe panel and sample image\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `action` | `\"launch\" \\| \"attach\" \\| \"set_breakpoint\" \\| \"remove_breakpoint\" \\| \"set_instruction_breakpoint\" \\| \"remove_instruction_breakpoint\" \\| \"data_breakpoint_info\" \\| \"set_data_breakpoint\" \\| \"remove_data_breakpoint\" \\| \"continue\" \\| \"step_over\" \\| \"step_in\" \\| \"step_out\" \\| \"pause\" \\| \"evaluate\" \\| \"stack_trace\" \\| \"threads\" \\| \"scopes\" \\| \"variables\" \\| \"disassemble\" \\| \"read_memory\" \\| \"write_memory\" \\| \"modules\" \\| \"loaded_sources\" \\| \"custom_request\" \\| \"output\" \\| \"terminate\" \\| \"sessions\"` | Yes | Dispatch key for the tool switch in `packages/coding-agent/src/tools/debug.ts`. |\n| `program` | `string` | No | Launch target path. Required for `launch`. Resolved relative to `cwd` if provided, otherwise session cwd. |\n| `args` | `string[]` | No | Program argv for `launch`. |\n| `adapter` | `string` | No | Explicit adapter name. Otherwise `selectLaunchAdapter()` / `selectAttachAdapter()` auto-pick from `packages/coding-agent/src/dap/config.ts`. |\n| `cwd` | `string` | No | Launch/attach working directory. Defaults to session cwd. |\n| `file` | `string` | No | Source file path for source breakpoints. |\n| `line` | `number` | No | Source line for source breakpoints. |\n| `function` | `string` | No | Function breakpoint name. Mutually exclusive with `file`+`line` in breakpoint actions. |\n| `name` | `string` | No | Data breakpoint info target name. Required for `data_breakpoint_info`. |\n| `condition` | `string` | No | Conditional expression for source/function/instruction/data breakpoints. |\n| `hit_condition` | `string` | No | Hit-count condition for instruction/data breakpoints. |\n| `expression` | `string` | No | Expression or raw debugger command. Required for `evaluate`. |\n| `context` | `string` | No | Evaluate context. Defaults to `\"repl\"`. Passed through as DAP evaluate context. |\n| `frame_id` | `number` | No | Frame selector for `evaluate`, `scopes`, `data_breakpoint_info`. `scopes` and `evaluate` default to the current stopped frame when omitted. |\n| `scope_id` | `number` | No | Variables reference from a scope. Accepted by `variables`; also used as a fallback variables reference for `data_breakpoint_info`. |\n| `variable_ref` | `number` | No | Variables reference for `variables`; preferred over `scope_id` when both are present. |\n| `pid` | `number` | No | Local process id for `attach`. `attach` requires `pid` or `port`. |\n| `port` | `number` | No | Remote attach port. If no adapter is forced, attach prefers `debugpy` when `port` is present. |\n| `host` | `string` | No | Remote attach host for `attach`. |\n| `levels` | `number` | No | Max stack frames for `stack_trace`. |\n| `memory_reference` | `string` | No | Memory reference/address for `disassemble`, `read_memory`, `write_memory`. `disassemble` uses this when provided; otherwise it falls back to the current stopped location's instruction-pointer reference if the adapter supplied one. |\n| `instruction_reference` | `string` | No | Instruction breakpoint reference; required for instruction breakpoint actions. Not used by `disassemble`. |\n| `instruction_count` | `number` | No | Required for `disassemble`. |\n| `instruction_offset` | `number` | No | Instruction offset for `disassemble`. |\n| `count` | `number` | No | Byte count for `read_memory`. Required there. |\n| `data` | `string` | No | Base64 payload for `write_memory`. Required there. |\n| `data_id` | `string` | No | Data breakpoint id. Required for `set_data_breakpoint` / `remove_data_breakpoint`. |\n| `access_type` | `\"read\" \\| \"write\" \\| \"readWrite\"` | No | Access filter for `set_data_breakpoint`. |\n| `command` | `string` | No | Custom DAP request command. Required for `custom_request`. |\n| `arguments` | `Record<string, unknown>` | No | Custom DAP request body for `custom_request`. |\n| `offset` | `number` | No | Offset for instruction breakpoints, disassembly, memory read, memory write. |\n| `resolve_symbols` | `boolean` | No | `disassemble` symbol-resolution flag. |\n| `allow_partial` | `boolean` | No | `write_memory` partial-write allowance. |\n| `start_module` | `number` | No | Modules pagination start index for `modules`. |\n| `module_count` | `number` | No | Modules pagination count for `modules`. |\n| `timeout` | `number` | No | Per-request timeout in seconds. Default `30`, clamped to `5..300`. |\n\n### Action-specific requirements\n- `launch`: `program`\n- `attach`: `pid` or `port`\n- `set_breakpoint` / `remove_breakpoint`: `function`, or `file` + `line`\n- `set_instruction_breakpoint` / `remove_instruction_breakpoint`: `instruction_reference`\n- `data_breakpoint_info`: `name`\n- `set_data_breakpoint` / `remove_data_breakpoint`: `data_id`\n- `evaluate`: `expression`\n- `variables`: `variable_ref` or `scope_id`\n- `disassemble`: capability `supportsDisassembleRequest`, plus `instruction_count`, and either `memory_reference` or a current stopped location with `instructionPointerReference`\n- `read_memory`: capability `supportsReadMemoryRequest`, plus `memory_reference` and `count`\n- `write_memory`: capability `supportsWriteMemoryRequest`, plus `memory_reference` and `data`\n- `modules`: capability `supportsModulesRequest`\n- `loaded_sources`: capability `supportsLoadedSourcesRequest`\n- `custom_request`: `command`\n\n### Interactive selector values\n`packages/coding-agent/src/debug/index.ts` also exposes a fixed UI-only selector with values `open-artifacts`, `performance`, `work`, `dump`, `memory`, `logs`, `system`, `terminal`, `protocols`, `raw-sse`, `transcript`, `clear-cache`. These are not model-callable through `debugSchema`; they are local TUI menu routes.\n\n## Outputs\nThe agent tool returns a standard `toolResult()` payload from `packages/coding-agent/src/tools/debug.ts`:\n- `content`: one text block. Every action renders human-readable text; there is no structured JSON block in `content`.\n- `details.action`: echoed action.\n- `details.success`: always initialized `true`; failures surface by throwing before a result is returned.\n- `details.snapshot`: present for actions that operate on or create a session, using `DapSessionSummary` from `packages/coding-agent/src/dap/types.ts`.\n- Action-specific `details` fields:\n  - `launch` / `attach`: `adapter`\n  - breakpoint actions: `breakpoints`, `functionBreakpoints`, `instructionBreakpoints`, `dataBreakpoints`\n  - `data_breakpoint_info`: `dataBreakpointInfo`\n  - `continue` / `step_*`: `state`, `timedOut`\n  - `threads`: `threads`\n  - `stack_trace`: `stackFrames`\n  - `scopes`: `scopes`\n  - `variables`: `variables`\n  - `evaluate`: `evaluation`\n  - `disassemble`: `disassembly`\n  - `read_memory`: `memoryAddress`, `memoryData`, `unreadableBytes`\n  - `write_memory`: `bytesWritten`\n  - `modules`: `modules`\n  - `loaded_sources`: `sources`\n  - `custom_request`: `customBody`\n  - `output`: `output`\n  - `sessions`: `sessions`\n\nStreaming/UI behavior:\n- The tool renderer merges call and result (`mergeCallAndResult: true`) and renders inline.\n- `debug.ts` itself does not emit progress updates through `_onUpdate`; result delivery is single-shot.\n- Approval is action-sensitive: read-only actions (`output`, `threads`, `stack_trace`, `scopes`, `variables`, `disassemble`, `read_memory`, `loaded_sources`, `modules`, `sessions`) request read approval; all other actions request exec approval.\n- The interactive selector is UI-driven instead of model-driven. It swaps TUI components, appends status lines to the chat pane, opens files in external viewers, or writes archives/temp files.\n\nSide-channel artifacts outside the model tool result:\n- `createReportBundle()` writes `omp-report-<timestamp>.tar.gz` under the reports dir and returns the filesystem path to the UI handler.\n- `#handleWorkReport()` writes `/tmp/work-profile-<Date.now()>.svg` before opening it.\n- `RawSseViewerComponent` and `DebugLogViewerComponent` can copy captured text to the clipboard.\n\n## Flow\n1. Tool registration is conditional: `DebugTool.createIf()` in `packages/coding-agent/src/tools/debug.ts` returns `null` unless `session.settings.get(\"debug.enabled\")` is true. `packages/coding-agent/src/tools/index.ts` wires the factory and rechecks the same setting in tool filtering.\n2. `DebugTool.execute()` clamps `params.timeout` through `clampTimeout(\"debug\", params.timeout)` and composes the caller `AbortSignal` with `AbortSignal.timeout(...)`.\n3. `launch` and `attach` resolve cwd/program paths, select an adapter in `packages/coding-agent/src/dap/config.ts`, then delegate to `dapSessionManager.launch()` / `.attach()`.\n4. `DapSessionManager.launch()` / `.attach()` enforce the single-session rule with `#ensureLaunchSlot()`, spawn the adapter through `DapClient.spawn()`, register listeners, send `initialize`, cache capabilities, start listening for an initial stop event before sending `launch`/`attach`, then complete the `initialized` → `configurationDone` handshake in `#completeConfigurationHandshake()`.\n5. `DapClient.spawn()` starts the adapter detached with `NON_INTERACTIVE_ENV`. Most adapters use stdio; socket-mode adapters (`dlv`) use `#spawnSocketUnix()` on Linux or `#spawnSocketClientAddr()` on macOS/other.\n6. `#registerSession()` in `packages/coding-agent/src/dap/session.ts` installs reverse-request handlers:\n   - `runInTerminal`: spawns the requested debuggee command detached via `ptree.spawn()` and returns `{ processId }`\n   - `startDebugging`: logs the child-session request and returns `{}`; it does not create nested sessions\n   - events: `output`, `initialized`, `stopped`, `continued`, `exited`, `terminated` update cached session state\n7. Operational actions (`set_breakpoint`, `evaluate`, `threads`, `read_memory`, `custom_request`, and similar) call `dapSessionManager` methods. Most flow through `#sendRequestWithConfig()`, which first sends `configurationDone` when required, then sends the DAP request, then updates `lastUsedAt`.\n8. Breakpoint actions maintain local cached breakpoint sets in `DapSessionManager` and remap adapter responses back onto those cached records.\n9. `continue` and the three step actions clear cached stop state, subscribe for `stopped`/`terminated`/`exited` before sending the DAP request, then `#awaitStopOutcome()` either returns the new stopped location or reports that the program is still running after timeout.\n10. `pause` sends DAP `pause`, waits for a stopped event if needed, and reuses cached stop state if the program was already stopped.\n11. `stack_trace`, `scopes`, `variables`, and `evaluate` default to the current stopped thread/frame when the caller omits ids and cached state is available.\n12. `output` reads the in-memory output ring from `DapSessionManager.getOutput()`. `terminate` sends `terminate` when supported, always attempts `disconnect`, marks the session terminated, and disposes the client.\n13. `sessions` reads the manager’s current map and formats all summaries. Although the manager stores a map, only one active session can exist because new launch/attach calls are blocked until the active one is terminated or cleaned up.\n14. The interactive selector in `packages/coding-agent/src/debug/index.ts` builds a `SelectList` of fixed values and dispatches each to a handler:\n   - `performance`: `startCpuProfile()`, wait for Enter/Escape, stop profiling, read a 30-second work profile with `getWorkProfile(30)`, then bundle via `createReportBundle()`\n   - `work`: read `getWorkProfile(30)`, write a temp SVG, open it externally\n   - `dump`: create a report bundle immediately\n   - `memory`: force GC, call `Bun.generateHeapSnapshot(\"v8\")`, then bundle\n   - `logs`: build a `DebugLogSource` and mount `DebugLogViewerComponent`\n   - `raw-sse`: resolve a `RawSseDebugBuffer` from the session and mount `RawSseViewerComponent`\n   - `system`: call `collectSystemInfo()` and render `formatSystemInfo()` into the chat pane\n   - `terminal`: `collectTerminalState()` + `formatTerminalState()` rendered into the chat pane\n   - `protocols`: fires a test desktop notification (unless suppressed), then mounts `ProtocolProbeComponent` with a sample image\n   - `open-artifacts`: open the current session artifact directory if it exists\n   - `transcript`: delegates to `ctx.handleDebugTranscriptCommand()`\n   - `clear-cache`: show confirmation, then remove artifact directories older than 30 days with `clearArtifactCache()`\n\n## Modes / Variants\n- **Availability gate**\n  - Tool hidden when `debug.enabled` is false.\n- **Adapter selection**\n  - `launch`: explicit `adapter` wins; otherwise `selectLaunchAdapter()` ranks available adapters by extension match, root-marker match, then native-debugger preference (`gdb`, `lldb-dap`) for extensionless binaries.\n  - `attach`: explicit `adapter` wins; otherwise remote `port` prefers `debugpy`, then native debuggers, then first available adapter.\n- **Transport**\n  - stdio adapters: direct `stdin`/`stdout` framing.\n  - socket adapters: Unix domain socket on Linux; TCP callback on macOS/other.\n- **DAP agent-tool actions**\n  - `launch` — spawn adapter, initialize session, maybe stop on entry; returns formatted session snapshot and `details.adapter`.\n  - `attach` — connect to a live process or remote port; same output shape as `launch`.\n  - `set_breakpoint` — source or function breakpoint add/update; returns the current breakpoint list for that target.\n  - `remove_breakpoint` — source or function breakpoint removal; returns the remaining breakpoint list.\n  - `set_instruction_breakpoint` / `remove_instruction_breakpoint` — require `supportsInstructionBreakpoints`; return current instruction breakpoint list.\n  - `data_breakpoint_info` — require `supportsDataBreakpoints`; asks the adapter for a `dataId`, access types, and description for `name`.\n  - `set_data_breakpoint` / `remove_data_breakpoint` — require `supportsDataBreakpoints`; return the cached data-breakpoint list.\n  - `continue` / `step_over` / `step_in` / `step_out` — return text describing whether execution stopped, terminated, or kept running, plus `details.state` and `details.timedOut`.\n  - `pause` — interrupts a running target and returns a stopped snapshot.\n  - `evaluate` — adapter expression evaluation; defaults context to `repl`.\n  - `stack_trace` — fetches frames for the resolved thread.\n  - `threads` — fetches current threads.\n  - `scopes` — frame scopes for an explicit `frame_id` or the current stopped frame.\n  - `variables` — variables for `variable_ref` or `scope_id`.\n  - `disassemble` — require `supportsDisassembleRequest`; disassembles around `memory_reference`, or around the current stopped instruction pointer when no memory reference is supplied.\n  - `read_memory` — require `supportsReadMemoryRequest`; returns address, base64 data, unreadable-byte count.\n  - `write_memory` — require `supportsWriteMemoryRequest`; writes base64 data and reports bytes written.\n  - `modules` — require `supportsModulesRequest`; optional pagination via `start_module` / `module_count`.\n  - `loaded_sources` — require `supportsLoadedSourcesRequest`; returns loaded source descriptors.\n  - `custom_request` — sends any DAP request name with arbitrary arguments.\n  - `output` — dumps captured stdout/stderr/console text from the session cache.\n  - `terminate` — disconnects and disposes the active session; returns `No debug session to terminate.` when none exists.\n  - `sessions` — lists all cached session summaries.\n- **Interactive selector routes (UI-only)**\n  - `logs` — loads today’s log tail and optional older daily log files into `DebugLogViewerComponent`; supports copy, range selection, pid filtering, load-older.\n  - `raw-sse` — live view over the session’s `RawSseDebugBuffer`; supports tail-follow, scrolling, copy-all.\n  - `performance` — CPU profile + 30-second work profile + report bundle.\n  - `memory` — heap snapshot + report bundle.\n  - `dump` — report bundle without profiler artifacts.\n  - `work` — standalone work-profile flamegraph export/open.\n  - `system` — formatted OS/arch/CPU/memory/version/cwd/shell/terminal dump.\n  - `terminal` — formatted terminal subprotocol/geometry/scrollback state dump.\n  - `protocols` — terminal protocol test: desktop-notification side effect plus a probe panel sampling special protocols.\n  - `open-artifacts` / `transcript` / `clear-cache` — artifact directory open, transcript export, artifact-cache pruning.\n\n## Side Effects\n- Filesystem\n  - Resolves program/file/cwd paths against the session cwd.\n  - Report creation writes `.tar.gz` bundles and may read the session JSONL, artifact files, subagent session JSONLs, and log files.\n  - Work-profile export writes `/tmp/work-profile-<timestamp>.svg`.\n  - Log source reads daily log files from the logs dir.\n  - Artifact-cache cleanup removes session artifact directories older than the cutoff.\n  - `resolveRawSseDebugBuffer()` may attach a non-enumerable `rawSseDebugBuffer` property to the owner object.\n- Network\n  - Socket-mode adapters bind/connect local sockets.\n  - Remote attach may connect through the adapter to a remote debug port.\n- Subprocesses / native bindings\n  - Spawns debugger adapters (`gdb`, `lldb-dap`, `python -m debugpy.adapter`, `dlv`, and others from `defaults.json`) detached.\n  - Reverse DAP `runInTerminal` requests spawn the debuggee detached via `ptree.spawn()`.\n  - `getWorkProfile(30)` comes from `@oh-my-pi/pi-natives`.\n  - CPU profiling uses `node:inspector/promises`; heap snapshots use `Bun.generateHeapSnapshot(\"v8\")`; raw/log viewers sanitize text via `@oh-my-pi/pi-natives`.\n  - `openPath()` launches the OS default file/browser handler for artifact dirs and SVGs.\n  - Log/raw-SSE viewers can call `copyToClipboard()`.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - `DapSessionManager` keeps session summaries, breakpoints, threads, stack frames, stop location, output capture, capabilities, and last-used timestamps in memory.\n  - Active-session id is global to the singleton `dapSessionManager`.\n  - `RawSseDebugBuffer` stores recent SSE events per owner/session.\n  - The tool is `exclusive`; concurrent debug tool calls are blocked by the scheduler.\n- User-visible prompts / interactive UI\n  - Debug selector shows confirmation before cache deletion.\n  - Performance profiling temporarily hijacks editor Enter/Escape handlers until profiling stops.\n  - Log/raw-SSE viewers replace the editor pane with custom components.\n- Background work / cancellation\n  - Every DAP request accepts an `AbortSignal`; timeouts and caller cancellation abort the active request, not the whole session lifetime.\n  - `DapSessionManager` runs a background cleanup loop every 30 seconds.\n  - Raw SSE viewers subscribe to buffer updates until closed.\n\n## Limits & Caps\n- Tool timeout clamp: `default=30`, `min=5`, `max=300` in `packages/coding-agent/src/tools/tool-timeouts.ts`.\n- Per-request DAP default timeout: `DEFAULT_REQUEST_TIMEOUT_MS = 30_000` in `packages/coding-agent/src/dap/client.ts`.\n- Single active session: enforced by `#ensureLaunchSlot()` in `packages/coding-agent/src/dap/session.ts`.\n- Idle session cleanup: `IDLE_TIMEOUT_MS = 10 * 60 * 1000`, checked every `CLEANUP_INTERVAL_MS = 30 * 1000`.\n- Adapter liveness heartbeat: `HEARTBEAT_INTERVAL_MS = 5 * 1000`.\n- Output capture cap: `MAX_OUTPUT_BYTES = 128 * 1024`; whole chunks are dropped from the front (then the front chunk is byte-sliced so exactly the cap remains) and `outputTruncated` is recorded.\n- Initial stop capture timeout after launch/attach: `STOP_CAPTURE_TIMEOUT_MS = 5_000`.\n- Socket-mode adapter readiness timeout: `10_000` ms in `waitForCondition()` and TCP connect timeout logic in `packages/coding-agent/src/dap/client.ts`.\n- Raw SSE buffer caps in `packages/coding-agent/src/debug/raw-sse-buffer.ts`:\n  - `MAX_RAW_SSE_EVENTS = 1_000`\n  - `MAX_RAW_SSE_CHARS = 512_000`\n  - `MAX_RAW_SSE_EVENT_CHARS = 64_000` per event, with `: omp-debug-truncated ...` marker appended on trim\n- Log viewer window in `packages/coding-agent/src/debug/log-viewer.ts`:\n  - `INITIAL_LOG_CHUNK = 50`\n  - `LOAD_OLDER_CHUNK = 50`\n- Report/log ingestion caps in `packages/coding-agent/src/debug/report-bundle.ts`:\n  - `MAX_LOG_LINES = 5000` for interactive log reading\n  - `MAX_LOG_BYTES = 2 * 1024 * 1024` tail-read ceiling\n  - report bundles include only the last `1000` log lines\n  - subagent session inclusion is capped at the most recent `10` JSONL files\n- Interactive profiling windows in `packages/coding-agent/src/debug/index.ts`: both performance and work reports request `getWorkProfile(30)`.\n- Artifact cache pruning default: `30` days in `clearArtifactCache()` and the selector confirmation text.\n\n## Errors\n- Parameter validation in `packages/coding-agent/src/tools/debug.ts` throws `ToolError` with explicit messages such as:\n  - `program is required for launch`\n  - `attach requires pid or port`\n  - `set_breakpoint requires file+line or function`\n  - `variables requires variable_ref or scope_id`\n  - `instruction_count is required for disassemble`\n  - `disassemble requires memory_reference unless the current stop location has an instruction pointer reference`\n  - `memory_reference is required for read_memory`\n  - `count is required for read_memory`\n  - `data is required for write_memory`\n  - `command is required for custom_request`\n- Adapter selection failure throws `No debugger adapter available. Installed adapters: ...`.\n- Capability-gated actions throw from `requireCapability(...)`, e.g. `Active adapter does not support memory reads.`\n- No-session and state errors come from `DapSessionManager`, e.g. `No active debug session. Launch or attach first.`, `No active stack frame. Run stack_trace first or supply frame_id.`, `Debugger reported no threads.`\n- Launching a second live session throws `Debug session <id> is still active. Terminate it before launching another.`\n- DAP transport/request failures surface as thrown errors from `DapClient`:\n  - `DAP request <command> timed out after <ms>ms`\n  - `DAP event <event> timed out after <ms>ms`\n  - `DAP adapter <name> is not running`\n  - `DAP adapter exited (code N): <stderr>` or `DAP adapter exited unexpectedly (code N)`\n  - adapter response `message` when a DAP request fails\n- `continue` / `step_*` are intentionally non-fatal when the target stays running past the timeout: they return `details.timedOut = true` and `state: \"running\"` instead of throwing.\n- `terminate` suppresses adapter errors while sending `terminate`/`disconnect`; it still disposes the client and returns the last summary when possible.\n- Interactive selector handlers report UI errors instead of throwing:\n  - profiler start/stop, report bundling, log reading, system-info collection, cache clearing, and artifact opening use `ctx.showError(...)` / `ctx.showWarning(...)`\n  - empty logs and empty artifact caches are warnings/status messages, not failures\n  - copy failures in log/raw-SSE viewers become status/error text in the UI\n- Report-bundle helpers are intentionally best-effort for many file reads: missing session files, missing artifact dirs, unreadable artifact files, missing log dirs, inaccessible cache dirs, and missing subagent files are skipped silently.\n- `collectSystemInfo()` is best-effort for CPU probing; failure there falls back to `Unknown CPU`.\n\n## Notes\n- `packages/coding-agent/src/prompts/tools/debug.md` tells the model only one active session is supported; that is not advisory, it is enforced in code.\n- `configurationDone` is sent automatically both during launch/attach handshake and lazily before later requests if the adapter required it and the initial handshake did not complete.\n- `startDebugging` reverse requests are acknowledged but not implemented; child debug sessions are not spawned.\n- `output` exposes the merged `output` event stream only; the tool does not distinguish stdout, stderr, and console categories.\n- Session summaries expose `needsConfigurationDone`; this is derived from adapter capabilities and whether `configurationDone` has been sent.\n- Source breakpoint file paths are normalized with `path.resolve()` before caching and sending to the adapter.\n- `evaluate` defaults to `repl`, so the tool can forward raw debugger commands when the adapter supports them.\n- `disassemble` resolves its target from `memory_reference` first, then the current stopped session's `instructionPointerReference`; it throws if neither is present.\n- `RawSseDebugBuffer.recordEvent()` increments `totalEvents` before bounded retention. A snapshot can therefore show fewer retained records than total observed events.\n- Raw SSE buffer listener failures are swallowed so viewer bugs do not break capture.\n- `createDebugLogSource()` walks daily log files newest-first, but `loadOlderLogs()` reverses each requested slice before concatenation so older chunks prepend in chronological order.\n- `clearArtifactCache()` deletes directories by directory mtime, not per-file age.\n- `addDirectoryToArchive()` reads artifact files as text with `Bun.file(...).text()`. Binary artifact contents are not preserved byte-for-byte in the report bundle.\n- The tool renderer truncates displayed output for the TUI preview, but the underlying text result still contains the full returned string.\n",
 	"tools/edit.md": "# edit\n\n> Applies source edits; default mode is the hashline patch language consumed from a single `input` string.\n\n## Source\n- Entry: `packages/coding-agent/src/edit/index.ts`\n- Model-facing prompt: `packages/hashline/src/prompt.md`\n- Key collaborators:\n  - `packages/coding-agent/src/utils/edit-mode.ts` — selects active edit mode\n  - `packages/hashline/src/grammar.lark` — canonical constrained-decoding grammar\n  - `packages/hashline/src/format.ts` — sigils and header constants (`[`, `]`, `#`, `+`, `replace`, `delete`, `insert`)\n  - `packages/hashline/src/input.ts` — parses `[PATH#TAG]` sections\n  - `packages/hashline/src/tokenizer.ts` / `packages/hashline/src/parser.ts` — tokenizes and parses ops\n  - `packages/hashline/src/apply.ts` — applies parsed edits to file text\n  - `packages/hashline/src/mismatch.ts` — stale-anchor mismatch formatting\n  - `packages/hashline/src/recovery.ts` — snapshot-based stale-anchor recovery\n  - `packages/hashline/src/snapshots.ts` — mints and resolves per-path opaque snapshot tags\n\n## Inputs\n\n### Hashline mode (default)\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `input` | `string` | Yes | One or more file sections. Anchored sections must start with `[PATH#TAG]`; `TAG` is the four-hex snapshot tag emitted by the latest `read`/`search`/`write`/successful `edit`. Optional `*** Begin Patch` / `*** End Patch` envelope is ignored if present. |\n\nPatch language inside `input`:\n\n- **File header**: `[PATH#TAG]`. `TAG` is four uppercase-hex chars — a content-derived hash of the whole normalized file (`computeFileHash()`), recorded in the session snapshot store.\n- **Operations**:\n  - `replace N..M:` — replace original lines N..M with the body rows below.\n  - `replace block N:` — replace the whole tree-sitter block beginning on line N (its header line through its closing line) with the body rows. The line span is resolved at apply time from the file's parse tree; point N at the line that opens the construct. The resolved span is exactly the node that begins on line N — a leading decorator, attribute, or doc-comment is a separate node and is not included; point N at the first decorator line (Python wraps `@dec` + `def` as one block) or fall back to `replace N..M:` to take a leading line-comment that parses as its own node (e.g. Rust `///`). On success the result echoes the matched span (`replace block N → resolved lines A-B`). Errors (and steers to `replace N..M:`) when the language is unsupported, line N is blank or a closing delimiter, no node begins there, or the resolved block has a syntax error.\n  - `delete N..M` — delete original lines N..M. No body.\n  - `delete block N` — delete the whole tree-sitter block beginning on line N (resolved like `replace block N`, with the same decorator/comment caveat). No body. On success the result echoes the matched span (`delete block N → resolved lines A-B`). Same resolution failure modes and `delete N..M` fallback.\n  - `insert before N:` — insert body rows immediately before line N.\n  - `insert after N:` — insert body rows immediately after line N.\n  - `insert after block N:` — insert body rows after the last line of the tree-sitter block beginning on line N. Point N at the line that opens the construct, never its closing delimiter / last visible line; if you can see the last line already, use plain `insert after M:`. Same resolution failure modes and `insert after M:` fallback.\n  - `insert head:` — insert body rows at the start of the file.\n  - `insert tail:` — insert body rows at the end of the file.\n- **Body rows**:\n  - Only body-bearing headers end in `:`.\n  - Every body row is `+TEXT`; `+` alone adds a blank line.\n  - `delete` never has body rows.\n  - There is no repeat row kind. To keep a line, leave it out of every range; split edits into multiple hunks when needed.\n  - `-` rows are invalid. Literal text beginning with `-` or `+` must be written as `+-text` / `++text`.\n\nAnchors come from `read`/`search` output. `read` emits a `[PATH#TAG]` header from the session snapshot store and lines as `LINE:TEXT`; copy the header into the edit section and copy only the line number into hunk headers.\n\n### Tolerated input shapes (lenient parsing)\n\nThe canonical grammar is strict, but the hand parser accepts a few non-dangerous variants:\n\n- `replace N:` — accepted as `replace N..N:`.\n- `delete N` — accepted as single-line delete.\n- Missing trailing colon on `replace` or `insert` — accepted.\n- `replace N-M:`, `replace N…M:`, and `replace N M:` — accepted as `replace N..M:`.\n- Bare body rows with no `+` prefix are auto-prepended with `+` and a `BARE_BODY_AUTO_PIPED_WARNING` is appended.\n- `*** Begin Patch` / `*** End Patch` envelopes are silently consumed. `*** Abort` terminates parsing silently — ops parsed before the marker still apply, no warning surfaced.\n- Some malformed bracketed headers are recovered after stripping apply-patch path noise such as `Update File:` / `Add File:` and extra `***`, but the recovered header still needs a valid four-hex tag for the patcher to apply it.\n- `*** Update File:` / `*** Add File:` / `*** Delete File:` / `*** Move to:` apply_patch sentinels inside the diff body throw an `apply_patch sentinel … is not valid in hashline` error.\n- `@@`-bracketed hunk headers are rejected with guidance to write a verb header.\n- Bare `N` and bare `N M` / `N..M` headers are rejected with guidance to write `replace` or `delete`.\n- `delete N..M:` and any body rows under `delete` / `delete block` are rejected.\n- Empty `replace` / `insert` / `replace block` hunks are rejected.\n- `-` body rows are rejected with `MINUS_ROW_REJECTED`.\n- `replace block N:` / `delete block N` / `insert after block N:` require a wired tree-sitter resolver; `replace block` and `insert after block` additionally need at least one `+TEXT` body row, while `delete block` takes none. An unresolvable block (unsupported language, blank/closing-delimiter line, no node beginning on N, or a syntax error in the resolved block) is rejected on the apply/final-preview path; the streaming preview silently drops it instead. Exception: `insert after block N:` anchored on a pure closing-delimiter line is lowered to plain `insert after N:` with a warning — line N is the end of a block, and inserting after that end is exactly what the plain form does.\n\n## Outputs\n- Single-shot tool result; hashline mode does not use a `resolve` preview/apply handshake.\n- `content` contains one text block per call. For a successful single-file edit it is the post-edit `[path#TAG]` section header (a fresh snapshot tag for the written content), followed by a compact diff preview from `packages/hashline/src/diff-preview.ts` when one is emitted.\n- When the patch used `replace block`/`delete block`/`insert after block` ops (and the apply matched the tagged content), one `replace block N → resolved lines A-B (K lines)` line per block op (single-line spans render `resolved line A (1 line)`; insert-after appends `; body lands after line B`) is inserted between the `[PATH#TAG]` header and the diff preview, so the caller can confirm tree-sitter resolved the construct it intended.\n- Parse, apply, or recovery warnings are appended as:\n\n```text\nWarnings:\n...\n```\n\n- `details` is `EditToolDetails` from `packages/coding-agent/src/edit/renderer.ts`:\n  - `diff`: unified diff string\n  - `firstChangedLine`: first changed post-edit line\n  - `diagnostics`: LSP/format result if available\n  - `op`: `\"create\"` or `\"update\"` for hashline mode\n  - `meta`: output metadata\n  - `perFileResults`: present for multi-section input\n- Multi-section input returns one aggregated result with combined text and per-file details.\n\n## Worked examples\n\nReference file (the exact shape `read` returns):\n\n```text\n[a.ts#0A3B]\n1:const X = \"a\";\n2:const Y = X;\n3:\n4:console.log(X);\n5:console.log(Y);\n6:export { X, Y };\n```\n\nReplace line 1 with two lines:\n\n```text\n[a.ts#0A3B]\nreplace 1..1:\n+const X = \"b\";\n+export const Y = X;\n```\n\nInsert below line 5:\n\n```text\n[a.ts#0A3B]\ninsert after 5:\n+console.log(X + Y);\n```\n\nInsert above line 5:\n\n```text\n[a.ts#0A3B]\ninsert before 5:\n+console.log(X + Y);\n```\n\nDelete lines 4..5 entirely:\n\n```text\n[a.ts#0A3B]\ndelete 4..5\n```\n\nInsert at start and end of file:\n\n```text\n[a.ts#0A3B]\ninsert head:\n+// header\ninsert tail:\n+// trailer\n```\n\nMulti-file:\n\n```text\n[src/a.ts#0A3B]\nreplace 4..4:\n+const enabled = true;\n[src/b.ts#1F7C]\ndelete 20\n```\n\n## Limits & Caps\n- File snapshot tags are exactly four uppercase-hex chars — content-derived hashes (`computeFileHash()`) recorded in the per-session snapshot store.\n- The visible mismatch report shows 2 lines of context on each side (`MISMATCH_CONTEXT`) in `packages/hashline/src/messages.ts`.\n- Stale-anchor recovery uses `fuzzFactor: 0` in `packages/hashline/src/recovery.ts`.\n- `HL_FILE_PREFIX` is `[`, `HL_FILE_SUFFIX` is `]`, `HL_PAYLOAD_REPLACE` is `+`, `HL_RANGE_SEP` is `..`, `HL_FILE_HASH_SEP` is `#`, and hunk keyword constants are `replace` / `delete` / `insert` (`packages/hashline/src/format.ts`).\n\n## Errors\n- Missing section header:\n  - `input must begin with \"[PATH#HASH]\" on the first non-blank line for anchored edits; got: ...`\n- Missing tag for any section:\n  - `Missing hashline snapshot tag for edit to <path>; use \\`[<path>#tag]\\` from your latest read/search output. To create a new file, use the write tool.`\n- Stray payload line:\n  - `line N: payload line has no preceding hunk header. Use \\`replace N..M:\\`, \\`delete N..M\\`, or \\`insert before|after|head|tail:\\` above the body. Got \"...\".`\n- Minus row:\n  - ``line N: `-` rows are not valid; hashline ranges already name the lines being changed. To insert a literal line starting with `-`, write `+-…`.``\n- Empty body-bearing hunk:\n  - `line N: \\`replace N..M:\\` needs at least one \\`+TEXT\\` body row. To delete lines, use \\`delete N..M\\`.`\n  - `line N: \\`insert\\` needs at least one \\`+TEXT\\` body row.`\n  - `line N: \\`replace block N:\\` needs at least one \\`+TEXT\\` body row. To delete a block, use \\`delete N..M\\` with the block's line range.`\n- Unresolvable block anchor (apply / final-preview path only):\n  - `line N: \\`replace block X:\\` could not resolve a syntactic block beginning on line X. The language may be unsupported, the line may be blank or a closing delimiter, or the block may not parse. Use \\`replace X..M:\\` with the block's explicit end line instead.` — followed by a blank line and numbered `*`-marked context rows around line X (same shape as the mismatch preview).\n  - `line N: \\`insert after block X:\\` could not resolve a syntactic block beginning on line X. The language may be unsupported, the line may be blank or a closing delimiter, or the block may not parse. Use \\`insert after M:\\` with the block's explicit last line instead.` — same context preview.\n- Delete with body:\n  - `line N: \\`delete N..M\\` does not take body rows. Remove the body, or use \\`replace N..M:\\`.`\n  - `line N: \\`delete block N\\` does not take body rows. Remove the body, or use \\`replace block N:\\` to replace the block.`\n- Range out of order:\n  - `line N: range A..B ends before it starts.`\n- Overlapping hunks on the same anchor:\n  - `line N: anchor line X is already targeted by another hunk on line Y. Issue ONE hunk per range; payload is only the final desired content, never a before/after pair.`\n- apply_patch / unified-diff contamination:\n  - `line N: apply_patch sentinel \"*** …\" is not valid in hashline. File sections start with \\`[path#HASH]\\` (no \\`Update File:\\` / \\`Add File:\\` keyword). Use \\`replace N..M:\\`, \\`delete N..M\\`, or \\`insert before|after|head|tail:\\` ops.`\n  - `line N: unified-diff hunk header (\\`@@ -N,M +N,M @@\\`) is not valid in hashline. Use \\`replace N..M:\\`, \\`delete N..M\\`, or \\`insert before|after|head|tail:\\` ops.`\n  - `line N: \\`@@\\`-bracketed hunk header \"@@ …\" is not valid in hashline. Drop the \\`@@ ... @@\\` brackets and write a verb header such as \\`replace N..M:\\`.`\n  - `line N: hunk headers need a verb. Use \\`replace N..N:\\` to replace, or \\`delete N\\` to delete.`\n  - `line N: bare range hunk header \"N M\" is not valid. Hunk headers need a verb: write \\`replace N..M:\\` or \\`delete N..M\\`.`\n- Out-of-range anchor:\n  - `Line N does not exist (file has M lines)`\n- Stale snapshot tag: the `Patcher` first attempts snapshot-based recovery. When recovery cannot prove a valid result it throws `MismatchError`, which distinguishes recognized-but-drifted hashes from never-recorded hashes. The error includes the current file hash plus context around each anchor.\n- No-op edit:\n  - `Edits to <path> parsed and applied cleanly, but produced no change: your body row(s) are byte-identical to the file at the targeted lines. The bug is somewhere else — re-read the file before issuing another edit. Do NOT widen the payload or add lines; verify the anchor first.`\n  - After `NOOP_HARD_LIMIT = 3` consecutive byte-identical no-ops of the same payload on the same file, the soft text result escalates to a `ToolError` (`STOP. Edits to <path> have been a byte-identical no-op N times in a row …`) from `packages/coding-agent/src/edit/hashline/noop-loop-guard.ts`.\n- Recovery failure is silent internally: if cache-based merge cannot prove a valid result, the mismatch error is surfaced unchanged.\n\n## Warnings\n- `Auto-prefixed bare body row(s) with +. Body rows must be +TEXT literal lines …` (`BARE_BODY_AUTO_PIPED_WARNING`)\n- Recovery banners: `RECOVERY_EXTERNAL_WARNING`, `RECOVERY_SESSION_CHAIN_WARNING`, `RECOVERY_SESSION_REPLAY_WARNING` (`packages/hashline/src/messages.ts`).\n",
-	"tools/eval.md": "# eval\n\n> Execute Python or JavaScript code in persistent cell-based runtimes.\n\n> **Notice:** Do not shell out to `python -c`/`python -e`, `bun -e`, or `node -e` via the `bash` tool for ad-hoc code execution. Use this tool instead — it gives you persistent state across cells, structured `display()` output, image/JSON capture, and proper cancellation/timeout handling that one-shot `-e`/`-c` invocations cannot provide.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/eval.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/eval.md`\n- Key collaborators:\n  - `packages/coding-agent/src/eval/backend.ts` — backend execution contract\n  - `packages/coding-agent/src/eval/agent-bridge.ts` — host-side `agent()` bridge into the subagent executor\n  - `packages/coding-agent/src/eval/js/executor.ts` — JS backend adapter\n  - `packages/coding-agent/src/eval/js/worker-core.ts` — JS execution, VM context, display/log capture\n  - `packages/coding-agent/src/eval/js/shared/prelude.txt` — JS global helper installer\n  - `packages/coding-agent/src/eval/js/shared/helpers.ts` — JS filesystem/text/env helper implementations\n  - `packages/coding-agent/src/eval/py/index.ts` — Python backend adapter\n  - `packages/coding-agent/src/eval/py/executor.ts` — kernel session retention, reset, cleanup\n  - `packages/coding-agent/src/eval/py/kernel.ts` — subprocess NDJSON runner protocol, display capture\n  - `packages/coding-agent/src/eval/py/prelude.py` — Python helper functions and status events\n  - `packages/coding-agent/src/session/streaming-output.ts` — truncation, artifacts, streamed chunks\n  - `docs/python-repl.md` — Python kernel/runner internals\n\n## Inputs\n\nTool parameters are a JSON object with a single `cells` field — an ordered array of cell objects. Each cell is a structured record; there is no `*** Cell` header parsing, no language sniffing, and no implicit single-cell fallback. Cells run in array order; state persists within each language across cells and across tool calls.\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `cells` | `EvalCellInput[]` | Yes | Cells executed in order. At least one cell is required (`.min(1)`). |\n\nEach `EvalCellInput` (from `evalCellSchema` in `packages/coding-agent/src/tools/eval.ts`):\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `language` | `\"py\" \\| \"js\"` | Yes | Backend selector. `\"py\"` maps to the IPython-style subprocess kernel (`python` backend); `\"js\"` maps to the persistent JavaScript VM. |\n| `code` | `string` | Yes | Cell body, verbatim. JSON-encoded — embed newlines, quotes, and indentation directly; no fences, no headers. |\n| `title` | `string` | No | Short label rendered in the transcript (e.g. `\"imports\"`, `\"load config\"`). |\n| `timeout` | `integer` | No | Per-cell timeout in seconds, clamped to `1..3600`. Defaults to 30 when omitted. |\n| `reset` | `boolean` | No | Wipe this cell's language kernel before running. Reset is per-language: a `py` cell's reset does not touch the JS VM and vice versa. Defaults to `false`. |\n\nMinimal example matching the live schema:\n\n```json\n{\n  \"cells\": [\n    { \"language\": \"py\", \"title\": \"imports\", \"timeout\": 10, \"code\": \"import json\\nfrom pathlib import Path\" },\n    { \"language\": \"py\", \"title\": \"load config\", \"code\": \"data = json.loads(read('package.json'))\\ndisplay(data)\" },\n    { \"language\": \"js\", \"title\": \"summary\", \"reset\": true, \"code\": \"const data = JSON.parse(await read('package.json'));\\ndisplay(data);\\nreturn data.name;\" }\n  ]\n}\n```\n\n## Outputs\n\nFinal result from `EvalTool.execute()` is single-shot, but `onUpdate` streams partial text and `details` while cells run.\n\nReturned shape:\n\n- `content`: one text block containing combined cell output, `(displayed N image(s); no text output)` when only images exist, or `(no output)` when nothing visible was produced; image outputs are appended as additional image content blocks.\n- `details` (`EvalToolDetails` from `packages/coding-agent/src/eval/types.ts`):\n  - `cells`: per-cell code, status (`pending`/`running`/`complete`/`error`), output, duration, exit code, status events, markdown flag\n  - `language`: first backend used\n  - `languages`: distinct backends used, in first-use order\n  - `jsonOutputs`: structured values emitted via `display(...)`\n  - `statusEvents`: aggregated helper/tool status events\n  - `notice`: backend fallback notice (currently unused; reserved for future per-cell notices)\n  - `meta`: truncation metadata\n  - `isError`: set on cell failure or cancellation\n\nRenderer behavior in `packages/coding-agent/src/tools/eval.ts`:\n\n- call preview renders each cell's `code` with syntax highlighting based on its declared `language`\n- result view renders each cell separately, including status, duration, and output\n- markdown outputs are rendered with the Markdown component instead of plain text\n- `jsonOutputs` render as a tree, collapsed or expanded depending on UI state\n- timeout / truncation notices render as dim metadata lines\n- images are returned as content image blocks; live updates may also carry `details.images` while execution is in progress\n\nSide-channel artifacts:\n\n- `session.allocateOutputArtifact?.(\"eval\")` may allocate an `artifact://...` backing store for spilled output.\n- Truncated output metadata points at that artifact when available.\n\n## Flow\n\n1. `EvalTool.execute()` in `packages/coding-agent/src/tools/eval.ts` receives `params.cells` already validated by the Zod schema — no string parsing step.\n2. For each cell, `execute()` maps `cell.language` to an `EvalLanguage` (`\"py\"` → `\"python\"`, `\"js\"` → `\"js\"`) and calls `resolveBackend(session, language)`:\n   - `python` is gated on `eval.py !== false` and `pythonBackend.isAvailable(session)`.\n   - `js` is gated on `eval.js !== false`.\n   - A disabled or unavailable requested backend throws `ToolError`; there is no auto-fallback or sniffing.\n3. The tool allocates an `OutputSink`, a `TailBuffer`, per-cell result objects, and a `sessionAbortController`. `session.trackEvalExecution?.(...)` can wrap the whole run for external cancellation tracking.\n4. It resolves the executor session id from `session.getEvalSessionId?.()`, falling back to `defaultEvalSessionId(session)`. Subagents inherit the parent's id so both sides share the same JS VM and Python kernel for each backend.\n5. Cells execute sequentially within one eval tool call. For each cell, `execute()`:\n   - clamps `cell.timeout ?? 30` seconds through `clampTimeout(\"eval\", ...)`\n   - builds a combined abort signal from the tool signal, the timeout, and the session abort controller\n   - marks the cell `running` and emits an update\n   - calls the backend's `execute()` with `cwd`, `sessionId`, `sessionFile`, `kernelOwnerId`, `idleTimeoutMs`, `reset` (defaults to `false`), the combined signal, and chunk/status callbacks\n6. JS cells dispatch through `packages/coding-agent/src/eval/js/index.ts` into `executeJs()`; Python cells dispatch through `packages/coding-agent/src/eval/py/index.ts` into `executePython()`.\n7. Backend text chunks stream into the shared `OutputSink`; rich outputs are accumulated separately as JSON, images, markdown markers, and status events.\n8. After each cell:\n   - text output is trimmed and stored on that cell result\n   - multi-cell runs prefix text with `[i/n]` and the optional title\n   - cancellations return early with `isError: true` and a cell-specific abort message\n   - non-zero exit codes return early with `isError: true` and a message naming the failed cell\n   - later cells are skipped after the first error, but earlier cell state persists in the underlying runtime\n9. On success, the tool joins all cell outputs, synthesizes `(no text output)` or `(no output)` when needed, and attaches truncation metadata from `summarizeFinal()`.\n10. The renderer uses `details.cells`, `details.jsonOutputs`, and `details.statusEvents` to build notebook-style output. `mergeCallAndResult = true` and `inline = true`, so call and result render together in the transcript.\n\n## Modes / Variants\n\n### Backend selection\n\nBackend choice is **explicit per cell** — there is no auto-detection.\n\n- `language: \"py\"` → Python (IPython-style subprocess kernel) backend\n- `language: \"js\"` → JavaScript VM backend\n\nIf the requested backend is disabled or unavailable, the tool throws `ToolError` for that cell. The caller chooses; the tool does not silently substitute.\n\n### JavaScript runtime\n\nImplemented in `packages/coding-agent/src/eval/js/worker-core.ts`, `packages/coding-agent/src/eval/js/shared/prelude.txt`, and `packages/coding-agent/src/eval/js/shared/helpers.ts`.\n\n- Persistent worker-backed VM sessions keyed by `js:${sessionId}`\n- `reset: true` calls `resetVmContext(sessionKey)` before the cell executes; reset is destructive for all live runs on that JS session\n- Top-level `await` and bare `return` are supported by wrapping code in an async IIFE when `wrapCode()` sees `await` or `return`\n- Top-level static `import ... from ...` and dynamic `import(...)` calls are routed through `rewriteImports()`, which sends them via `__omp_import__` so the specifier resolves against the session cwd. Dynamic-import call sites are swapped for a guarded shim (`typeof __omp_import__ === \"function\" ? __omp_import__ : (s, o) => import(s, o)`) rather than the bare helper identifier: functions handed to puppeteer (`tab.evaluate`, `page.evaluate`, ...) are serialized with `Function.prototype.toString()` and re-evaluated inside the browser page, where the worker-injected helper does not exist, so the shim falls back to native dynamic import there\n- Module cache is busted for **local** imports between cells so edits to source files are picked up without restarting the runtime. `__omp_import__` deletes `require.cache[absPath]` before re-importing whenever the original specifier is a filesystem path: relative (`./x`, `../x`, `.`, `..`), POSIX-absolute (`/...`), home-prefixed (`~/...`), or Windows drive-letter (`C:\\...` / `C:/...`). Bare specifiers (`react`, `lodash/x`) and URL/scheme specifiers (`node:fs`, `file://...`, `https://...`) are left in cache so package identity stays stable across cells. The cache-bust only fires when the resolved target is an absolute path — unresolved bare-package fallbacks (`resolveImportSpecifier()` returning the original specifier) skip it.\n- The prelude installs globals:\n  - `display`, `print`\n  - `read`, `write`, `append`, `sort`, `uniq`, `counter`, `diff`, `tree`, `env`, `output`\n  - `tool.<name>(args)` proxy for arbitrary session tool calls\n  - `completion(prompt, opts?)` for oneshot, stateless model calls (see _Oneshot completion helper_ below)\n  - `agent(prompt, opts?)` for a single subagent call, plus `parallel()` / `pipeline()` bounded-pool helpers (see _Subagent helper_ below)\n- JS helpers that touch the host/runtime boundary are async and `await`able; pure text helpers (`sort`, `uniq`, `counter`) return synchronously but may still be safely awaited.\n- JS helper signatures use a trailing options object rather than Python keyword arguments:\n  - `await read(path, { offset?, limit? })`\n  - `await tree(path = \".\", { maxDepth?, hidden? })`\n  - `sort(text, { reverse?, unique? })`, `uniq(text, { count? })`, `counter(items, { limit?, reverse? })`\n  - `await agent(prompt, { agentType?, model?, label?, schema? })`\n  - `await parallel([() => agent(\"a\"), () => agent(\"b\")])`\n  - `await pipeline(items, stage1, stage2)`\n- `display(value)` behavior:\n  - plain objects/arrays become JSON outputs\n  - `{ type: \"image\", data, mimeType }` becomes an image output\n  - scalars become text\n- The VM exposes a restricted `process` subset plus `Buffer`, `fetch`, `Blob`, `File`, `Headers`, `Request`, `Response`, `fs`, `require`, and browser-style globals\n- Concurrent runs on the same VM are not queued end-to-end. Synchronous JS still runs on the single event loop; awaited regions can interleave with sibling runs.\n\n### Python runtime\n\nImplemented in `packages/coding-agent/src/eval/py/executor.ts`, `packages/coding-agent/src/eval/py/kernel.ts`, and `packages/coding-agent/src/eval/py/prelude.py`. See `docs/python-repl.md` for kernel and runner details.\n\n- Default mode is retained `session` kernels keyed by `python:${sessionId}` plus normalized cwd and interpreter\n- Optional `python.kernelMode = \"per-call\"` creates a fresh kernel for each cell and shuts it down afterward\n- `reset: true` disposes the retained kernel for that session before the cell runs; later Python cells in the same tool call reuse the fresh kernel\n- Startup path:\n  - availability check\n  - create/connect kernel\n  - initialize cwd / env / `sys.path`\n  - execute `PYTHON_PRELUDE`\n- Python cells run in the runner's persistent asyncio event loop, so top-level `await` works; the prompt warns not to use `asyncio.run(...)`\n- The Python prelude defines helpers with the same surface as JS where practical, including `tool.<name>(args)`, `completion(...)`, and `agent(...)` through a per-run loopback bridge\n- Synchronous statement blocks run in the default executor with ContextVar state copied in; the GIL still serializes bytecode execution, but awaited regions can interleave with sibling cells\n- Kernel `display` / `result` frames map to:\n  - `application/x-omp-status` → status event\n  - `image/png` → image output\n  - `application/json` → JSON output\n  - `text/markdown` → markdown output\n  - `text/plain` → text output\n  - `text/html` → HTML converted to markdown with `htmlToBasicMarkdown()`\n- Interactive stdin is rejected: a stdin-flagged result returns exit code `1` with `Kernel requested stdin; interactive input is not supported.`\n\n### Oneshot completion helper (`completion`)\n\nBoth runtimes expose `completion()` — a single stateless completion against a model tier. It is intentionally minimal: no conversation history, no agent-visible tools, pure text in / text (or object) out. Implemented host-side in `packages/coding-agent/src/eval/completion-bridge.ts` and routed through the existing tool bridge under the reserved name `__completion__`.\n\n- Signatures:\n  - JS: `await completion(prompt, { model?, system?, schema? })`\n  - Python: `completion(prompt, *, model=\"default\", system=None, schema=None)`\n- `model` selects a tier (default `\"default\"`):\n  - `\"smol\"` → `pi/smol` role (fast / cheap)\n  - `\"default\"` → the session's active model, falling back to the `pi/default` role\n  - `\"slow\"` → `pi/slow` role; requests high reasoning effort only on reasoning-capable models\n- `system` (optional) supplies a system prompt.\n- `schema` (optional) is a plain JSON-Schema object. When present, the model is forced to call a single synthetic `respond` tool with that schema (loose, non-strict), and the helper returns the parsed object. When absent, the helper returns the completion string.\n- Errors surface as exceptions: unresolved tier, missing API key, an `error`/`aborted` stop reason, or empty output each raise.\n\n### Subagent helper (`agent`)\n\nBoth runtimes expose `agent()` — a single subagent invocation routed through `packages/coding-agent/src/eval/agent-bridge.ts` into the same `runSubprocess(...)` path used by the `task` tool. It uses the current eval session's spawn policy and inherits the parent eval executor id, so parent and subagent code share JS/Python runtime state.\n\n- Signatures:\n  - JS: `await agent(prompt, { agentType?, model?, label?, schema? })`\n  - Python: `agent(prompt, *, agent_type=\"task\", model=None, label=None, schema=None)`\n- `agentType` / `agent_type` defaults to the bundled `task` agent and resolves through normal agent discovery, so project and user agents work.\n- `model` overrides the selected agent's model. Without it, normal per-agent settings and the agent frontmatter model apply.\n- Shared background is passed via files: write a `local://` file and reference it in the prompt. `label` controls the `agent://<id>` output label prefix.\n- `schema` passes a JSON Schema to the subagent structured-output path. When present, the helper parses the final JSON text and returns an object.\n- Spawn restrictions use `session.getSessionSpawns()` exactly like the `task` tool. Eval-driven subagent recursion is capped at depth 3.\n- JS and Python both expose `parallel(thunks)` and `pipeline(items, ...stages)`; both use a bounded async/threaded pool whose width tracks the `task.maxConcurrency` setting (the same ceiling the `task` tool uses; `0` = run every item at once), preserve item order, and propagate rejections. The width is fetched live from the host via the `__concurrency__` bridge, so the helpers no longer take a `concurrency` argument.\n- Errors surface as exceptions: unknown or disabled agent, disallowed spawn, recursion cap, subagent failure, or invalid structured output all fail the eval cell.\n\n### Multi-language call behavior\n\nA single tool call can mix Python and JS cells. Persistence is per language runtime:\n\n- `reset: true` on a Python cell does not touch JS state\n- `reset: true` on a JS cell does not touch Python state\n- each backend keeps its own retained session keyed from the same session-derived ID\n\n## Side Effects\n\n- Filesystem\n  - JS/Python prelude helpers can read, write, append, diff, and traverse filesystem paths under the session cwd or absolute paths.\n  - JS helper `read()` rejects protocol URIs (`://`) and directory paths; use `tool.read(...)` for internal URLs or reader-mode behavior.\n  - Output may spill to an artifact file via `OutputSink`.\n- Network\n  - Python backend speaks NDJSON to a local `python3` subprocess over stdin/stdout (no network).\n  - JS runtime exposes `fetch` and `tool.<name>()`; those tools may perform additional network I/O.\n- Subprocesses / native bindings\n  - Python availability check runs `<python> -c ...`.\n  - Python backend spawns one `python -u runner.py` subprocess per kernel; cancellation sends `SIGINT`. Details in `docs/python-repl.md`.\n  - `agent()` runs one in-process subagent via the task executor; that subagent may use its configured tools.\n- Session state\n  - `session.assertEvalExecutionAllowed?.()` can block execution.\n  - `session.trackEvalExecution?.(...)` can register cancellable eval work.\n  - `session.getSessionFile?.()`, `session.getEvalSessionId?.()`, and `session.getEvalKernelOwnerId?.()` influence VM/kernel reuse and artifact lookup.\n  - JS VM contexts persist across eval calls until reset/disposal.\n  - Python retained kernels persist until reset, owner cleanup, or process exit.\n  - `agent()` allocates `agent://<id>` output artifacts and reuses the parent's eval executor id.\n- User-visible prompts / interactive UI\n  - none; stdin requests are rejected programmatically\n- Background work / cancellation\n  - Python retained kernels have heartbeat and idle cleanup timers.\n  - Cancellation hard-kills/resets the shared executor for that backend: JS terminates the worker, Python sends SIGINT and may escalate to subprocess shutdown.\n\n## Limits & Caps\n\n- Per-cell timeout default: 30s (applied when `timeout` is omitted in `EvalTool.execute()`; clamped through `TOOL_TIMEOUTS.eval.default` in `packages/coding-agent/src/tools/tool-timeouts.ts`)\n- Schema-level `timeout` range: integer `1..3600` seconds (enforced by Zod on the cell schema)\n- Timeout clamp at runtime: 1s minimum, 3600s maximum (`TOOL_TIMEOUTS.eval` in `packages/coding-agent/src/tools/tool-timeouts.ts`)\n- Transcript code/output preview: 10 lines by default (`EVAL_DEFAULT_PREVIEW_LINES` in `packages/coding-agent/src/tools/eval-render.ts`, re-exported from `eval.ts`)\n- Output truncation window: 50KB default (`DEFAULT_MAX_BYTES` in `packages/coding-agent/src/session/streaming-output.ts`)\n- Output line cap inside truncation helpers: 3000 lines (`DEFAULT_MAX_LINES` in `packages/coding-agent/src/session/streaming-output.ts`)\n- Streaming tail buffer for live updates: `DEFAULT_MAX_BYTES * 2` = 100KB (`packages/coding-agent/src/tools/eval.ts`)\n- JS/Python `parallel()` / `pipeline()` helper pool width: the `task.maxConcurrency` setting (default 32; `0` = unbounded), resolved live via the `__concurrency__` bridge (`packages/coding-agent/src/eval/concurrency-bridge.ts`)\n- Eval-driven `agent()` recursion cap: task depth 3 (`EVAL_AGENT_MAX_DEPTH`)\n- Python kernel startup wait: 10s (`STARTUP_TIMEOUT_MS` in `packages/coding-agent/src/eval/py/kernel.ts`)\n- Python kernel shutdown grace per escalation step (`exit` request → `SIGTERM` → `SIGKILL`): 1000ms (`SHUTDOWN_GRACE_MS` in `packages/coding-agent/src/eval/py/kernel.ts`)\n- Python SIGINT escalation window: 5s without a `done` frame before the subprocess is killed (`INTERRUPT_ESCALATION_MS` in `packages/coding-agent/src/eval/py/kernel.ts`)\n- Python auto-restart budget: a dead retained kernel is replaced and the cell retried once per execution (`executeOnSession` in `packages/coding-agent/src/eval/py/executor.ts`)\n\n## Errors\n\n- Zod validation rejects malformed `cells` arrays before `execute()` runs (missing `language`/`code`, out-of-range `timeout`, empty `cells`).\n- Missing session without proxy executor throws `ToolError(\"Eval tool requires a session when not using proxy executor\")`.\n- Disabled/unavailable backends throw `ToolError` from `resolveBackend()`:\n  - `eval.py = false` and a `py` cell is requested\n  - `eval.js = false` and a `js` cell is requested\n  - Python kernel unavailable and a `py` cell is requested\n- JS runtime exceptions are converted into text output plus `exitCode: 1`; cancellations return `cancelled: true` and may append `Command timed out`.\n- Python execution errors from the kernel become text output and `exitCode: 1`; later cells are skipped.\n- Python stdin requests are treated as errors with the message `Kernel requested stdin; interactive input is not supported.`\n- Cancellation is returned, not thrown, once backend execution has started. The tool formats it as a cell failure and sets `details.isError = true`.\n- If output truncates, the tool still succeeds; truncation is surfaced through `details.meta` and artifact-backed full output when available.\n\n## Shared executor trade-offs\n\n- Parent agents and subagents share eval state bidirectionally when a subagent inherits the parent's executor id. Mutations in either direction are visible to the other participant.\n- Async regions of concurrent runs can interleave. Synchronous JS still blocks the VM event loop; synchronous Python still contends on the GIL.\n- Cancelling one run is destructive to the shared backend executor. This is intentional: JS worker termination and Python SIGINT/subprocess shutdown are the only reliable way to interrupt arbitrary user code.\n- `reset: true` is destructive for every live run on that backend session id. Concurrent Python resets coalesce — a reset already in flight is awaited rather than duplicated, and runs queued behind it proceed on the freshly-restarted kernel.\n\n## Notes\n\n- Backend selection is strictly explicit per cell: `language` must be `\"py\"` or `\"js\"`. The previous `*** Cell` header parser, the `eval.lark` constrained grammar, and the sniffer-based fallback have all been removed.\n- `EvalTool.customFormat` no longer exists. Tool calls flow through the standard JSON schema; there is no Lark-constrained sampling path.\n- `tool.<name>()` exists in both JS and Python. Python calls route through a per-run loopback bridge keyed by the current cell id.\n- JS helper paths reject protocol URIs (`://`) in `resolveRegularFile()` for `read()`, and resolve other paths against the session cwd or absolute filesystem path. Use `tool.read(...)` or another tool explicitly for internal URLs.\n- Python helper `output(...)` depends on `PI_ARTIFACTS_DIR` or `PI_SESSION_FILE`; it fails outside a session-backed run.\n- `display()` can produce text and structured outputs from the same value; the renderer prefers markdown over `text/plain` when both exist.\n- JS static imports are rewritten only at top level. Nested imports stay invalid and surface normal JS syntax/runtime errors.\n- `EvalTool` is `concurrency = \"exclusive\"` within one agent session, but parent and subagent sessions can run eval concurrently when they share an inherited executor id.\n- The tool description shown to the model is templated by backend availability (`getEvalToolDescription()`); if Python is unavailable, the prompt omits Python-specific instructions.\n",
+	"tools/eval.md": "# eval\n\n> Execute Python or JavaScript code in persistent cell-based runtimes.\n\n> **Notice:** Do not shell out to `python -c`/`python -e`, `bun -e`, or `node -e` via the `bash` tool for ad-hoc code execution. Use this tool instead — it gives you persistent state across cells, structured `display()` output, image/JSON capture, and proper cancellation/timeout handling that one-shot `-e`/`-c` invocations cannot provide.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/eval.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/eval.md`\n- Key collaborators:\n  - `packages/coding-agent/src/eval/backend.ts` — backend execution contract\n  - `packages/coding-agent/src/eval/agent-bridge.ts` — host-side `agent()` bridge into the subagent executor\n  - `packages/coding-agent/src/eval/js/executor.ts` — JS backend adapter\n  - `packages/coding-agent/src/eval/js/worker-core.ts` — JS execution, VM context, display/log capture\n  - `packages/coding-agent/src/eval/js/shared/prelude.txt` — JS global helper installer\n  - `packages/coding-agent/src/eval/js/shared/helpers.ts` — JS filesystem/text/env helper implementations\n  - `packages/coding-agent/src/eval/py/index.ts` — Python backend adapter\n  - `packages/coding-agent/src/eval/py/executor.ts` — kernel session retention, reset, cleanup\n  - `packages/coding-agent/src/eval/py/kernel.ts` — subprocess NDJSON runner protocol, display capture\n  - `packages/coding-agent/src/eval/py/prelude.py` — Python helper functions and status events\n  - `packages/coding-agent/src/session/streaming-output.ts` — truncation, artifacts, streamed chunks\n  - `docs/python-repl.md` — Python kernel/runner internals\n\n## Inputs\n\nTool parameters are a JSON object with a single `cells` field — an ordered array of cell objects. Each cell is a structured record; there is no `*** Cell` header parsing, no language sniffing, and no implicit single-cell fallback. Cells run in array order; state persists within each language across cells and across tool calls.\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `cells` | `EvalCellInput[]` | Yes | Cells executed in order. At least one cell is required (`.min(1)`). |\n\nEach `EvalCellInput` (from `evalCellSchema` in `packages/coding-agent/src/tools/eval.ts`):\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `language` | `\"py\" \\| \"js\"` | Yes | Backend selector. `\"py\"` maps to the IPython-style subprocess kernel (`python` backend); `\"js\"` maps to the persistent JavaScript VM. |\n| `code` | `string` | Yes | Cell body, verbatim. JSON-encoded — embed newlines, quotes, and indentation directly; no fences, no headers. |\n| `title` | `string` | No | Short label rendered in the transcript (e.g. `\"imports\"`, `\"load config\"`). |\n| `timeout` | `integer` | No | Per-cell timeout in seconds, clamped to `1..3600`. Defaults to 30 when omitted. |\n| `reset` | `boolean` | No | Wipe this cell's language kernel before running. Reset is per-language: a `py` cell's reset does not touch the JS VM and vice versa. Defaults to `false`. |\n\nMinimal example matching the live schema:\n\n```json\n{\n  \"cells\": [\n    { \"language\": \"py\", \"title\": \"imports\", \"timeout\": 10, \"code\": \"import json\\nfrom pathlib import Path\" },\n    { \"language\": \"py\", \"title\": \"load config\", \"code\": \"data = json.loads(read('package.json'))\\ndisplay(data)\" },\n    { \"language\": \"js\", \"title\": \"summary\", \"reset\": true, \"code\": \"const data = JSON.parse(await read('package.json'));\\ndisplay(data);\\nreturn data.name;\" }\n  ]\n}\n```\n\n## Outputs\n\nFinal result from `EvalTool.execute()` is single-shot, but `onUpdate` streams partial text and `details` while cells run.\n\nReturned shape:\n\n- `content`: one text block containing combined cell output, `(displayed N image(s); no text output)` when only images exist, or `(no output)` when nothing visible was produced; image outputs are appended as additional image content blocks.\n- `details` (`EvalToolDetails` from `packages/coding-agent/src/eval/types.ts`):\n  - `cells`: per-cell code, status (`pending`/`running`/`complete`/`error`), output, duration, exit code, status events, markdown flag\n  - `language`: first backend used\n  - `languages`: distinct backends used, in first-use order\n  - `jsonOutputs`: structured values emitted via `display(...)`\n  - `statusEvents`: aggregated helper/tool status events\n  - `notice`: backend fallback notice (currently unused; reserved for future per-cell notices)\n  - `meta`: truncation metadata\n  - `isError`: set on cell failure or cancellation\n\nRenderer behavior in `packages/coding-agent/src/tools/eval.ts`:\n\n- call preview renders each cell's `code` with syntax highlighting based on its declared `language`\n- result view renders each cell separately, including status, duration, and output\n- markdown outputs are rendered with the Markdown component instead of plain text\n- `jsonOutputs` render as a tree, collapsed or expanded depending on UI state\n- timeout / truncation notices render as dim metadata lines\n- images are returned as content image blocks; live updates may also carry `details.images` while execution is in progress\n\nSide-channel artifacts:\n\n- `session.allocateOutputArtifact?.(\"eval\")` may allocate an `artifact://...` backing store for spilled output.\n- Truncated output metadata points at that artifact when available.\n\n## Flow\n\n1. `EvalTool.execute()` in `packages/coding-agent/src/tools/eval.ts` receives `params.cells` already validated by the Zod schema — no string parsing step.\n2. For each cell, `execute()` maps `cell.language` to an `EvalLanguage` (`\"py\"` → `\"python\"`, `\"js\"` → `\"js\"`) and calls `resolveBackend(session, language)`:\n   - `python` is gated on `eval.py !== false` and `pythonBackend.isAvailable(session)`.\n   - `js` is gated on `eval.js !== false`.\n   - A disabled or unavailable requested backend throws `ToolError`; there is no auto-fallback or sniffing.\n3. The tool allocates an `OutputSink`, a `TailBuffer`, per-cell result objects, and a `sessionAbortController`. `session.trackEvalExecution?.(...)` can wrap the whole run for external cancellation tracking.\n4. It resolves the executor session id from `session.getEvalSessionId?.()`, falling back to `defaultEvalSessionId(session)`. Subagents inherit the parent's id so both sides share the same JS VM and Python kernel for each backend.\n5. Cells execute sequentially within one eval tool call. For each cell, `execute()`:\n   - clamps `cell.timeout ?? 30` seconds through `clampTimeout(\"eval\", ...)`\n   - builds a combined abort signal from the tool signal, the timeout, and the session abort controller\n   - marks the cell `running` and emits an update\n   - calls the backend's `execute()` with `cwd`, `sessionId`, `sessionFile`, `kernelOwnerId`, `idleTimeoutMs`, `reset` (defaults to `false`), the combined signal, and chunk/status callbacks\n6. JS cells dispatch through `packages/coding-agent/src/eval/js/index.ts` into `executeJs()`; Python cells dispatch through `packages/coding-agent/src/eval/py/index.ts` into `executePython()`.\n7. Backend text chunks stream into the shared `OutputSink`; rich outputs are accumulated separately as JSON, images, markdown markers, and status events.\n8. After each cell:\n   - text output is trimmed and stored on that cell result\n   - multi-cell runs prefix text with `[i/n]` and the optional title\n   - cancellations return early with `isError: true` and a cell-specific abort message\n   - non-zero exit codes return early with `isError: true` and a message naming the failed cell\n   - later cells are skipped after the first error, but earlier cell state persists in the underlying runtime\n9. On success, the tool joins all cell outputs, synthesizes `(no text output)` or `(no output)` when needed, and attaches truncation metadata from `summarizeFinal()`.\n10. The renderer uses `details.cells`, `details.jsonOutputs`, and `details.statusEvents` to build notebook-style output. `mergeCallAndResult = true` and `inline = true`, so call and result render together in the transcript.\n\n## Modes / Variants\n\n### Backend selection\n\nBackend choice is **explicit per cell** — there is no auto-detection.\n\n- `language: \"py\"` → Python (IPython-style subprocess kernel) backend\n- `language: \"js\"` → JavaScript VM backend\n\nIf the requested backend is disabled or unavailable, the tool throws `ToolError` for that cell. The caller chooses; the tool does not silently substitute.\n\n### JavaScript runtime\n\nImplemented in `packages/coding-agent/src/eval/js/worker-core.ts`, `packages/coding-agent/src/eval/js/shared/prelude.txt`, and `packages/coding-agent/src/eval/js/shared/helpers.ts`.\n\n- Persistent worker-backed VM sessions keyed by `js:${sessionId}`\n- `reset: true` calls `resetVmContext(sessionKey)` before the cell executes; reset is destructive for all live runs on that JS session\n- Top-level `await` and bare `return` are supported by wrapping code in an async IIFE when `wrapCode()` sees `await` or `return`\n- Top-level static `import ... from ...` and dynamic `import(...)` calls are routed through `rewriteImports()`, which sends them via `__omp_import__` so the specifier resolves against the session cwd. Dynamic-import call sites are swapped for a guarded shim (`typeof __omp_import__ === \"function\" ? __omp_import__ : (s, o) => import(s, o)`) rather than the bare helper identifier: functions handed to puppeteer (`tab.evaluate`, `page.evaluate`, ...) are serialized with `Function.prototype.toString()` and re-evaluated inside the browser page, where the worker-injected helper does not exist, so the shim falls back to native dynamic import there\n- Module cache is busted for **local** imports between cells so edits to source files are picked up without restarting the runtime. `__omp_import__` deletes `require.cache[absPath]` before re-importing whenever the original specifier is a filesystem path: relative (`./x`, `../x`, `.`, `..`), POSIX-absolute (`/...`), home-prefixed (`~/...`), or Windows drive-letter (`C:\\...` / `C:/...`). Bare specifiers (`react`, `lodash/x`) and URL/scheme specifiers (`node:fs`, `file://...`, `https://...`) are left in cache so package identity stays stable across cells. The cache-bust only fires when the resolved target is an absolute path — unresolved bare-package fallbacks (`resolveImportSpecifier()` returning the original specifier) skip it.\n- The prelude installs globals:\n  - `display`, `print`\n  - `read`, `write`, `append`, `sort`, `uniq`, `counter`, `diff`, `tree`, `env`, `output`\n  - `tool.<name>(args)` proxy for arbitrary session tool calls\n  - `completion(prompt, opts?)` for oneshot, stateless model calls (see _Oneshot completion helper_ below)\n  - `agent(prompt, opts?)` for a single subagent call, plus `parallel()` / `pipeline()` bounded-pool helpers (see _Subagent helper_ below)\n- JS helpers that touch the host/runtime boundary are async and `await`able; pure text helpers (`sort`, `uniq`, `counter`) return synchronously but may still be safely awaited.\n- JS helper options may be passed either positionally in the Python order or as a trailing options object. `null` and `undefined` skip positional slots:\n  - `await read(path, offset?, limit?)` or `await read(path, { offset?, limit? })`\n  - `await tree(path = \".\", maxDepth?, showHidden?)` or `await tree(path, { maxDepth?, showHidden? })`\n  - `sort(text, reverse?, unique?)`, `uniq(text, count?)`, `counter(items, limit?, reverse?)`\n  - `await agent(prompt, agentType?, model?, label?, schema?)` or `await agent(prompt, { agentType?, model?, label?, schema? })`\n  - `await parallel([() => agent(\"a\"), () => agent(\"b\")])`\n  - `await pipeline(items, stage1, stage2)`\n- `display(value)` behavior:\n  - plain objects/arrays become JSON outputs\n  - `{ type: \"image\", data, mimeType }` becomes an image output\n  - scalars become text\n- The VM exposes a restricted `process` subset plus `Buffer`, `fetch`, `Blob`, `File`, `Headers`, `Request`, `Response`, `fs`, `require`, and browser-style globals\n- Concurrent runs on the same VM are not queued end-to-end. Synchronous JS still runs on the single event loop; awaited regions can interleave with sibling runs.\n\n### Python runtime\n\nImplemented in `packages/coding-agent/src/eval/py/executor.ts`, `packages/coding-agent/src/eval/py/kernel.ts`, and `packages/coding-agent/src/eval/py/prelude.py`. See `docs/python-repl.md` for kernel and runner details.\n\n- Default mode is retained `session` kernels keyed by `python:${sessionId}` plus normalized cwd and interpreter\n- Optional `python.kernelMode = \"per-call\"` creates a fresh kernel for each cell and shuts it down afterward\n- `reset: true` disposes the retained kernel for that session before the cell runs; later Python cells in the same tool call reuse the fresh kernel\n- Startup path:\n  - availability check\n  - create/connect kernel\n  - initialize cwd / env / `sys.path`\n  - execute `PYTHON_PRELUDE`\n- Python cells run in the runner's persistent asyncio event loop, so top-level `await` works; the prompt warns not to use `asyncio.run(...)`\n- The Python prelude defines helpers with the same surface as JS where practical, including `tool.<name>(args)`, `completion(...)`, and `agent(...)` through a per-run loopback bridge\n- Synchronous statement blocks run in the default executor with ContextVar state copied in; the GIL still serializes bytecode execution, but awaited regions can interleave with sibling cells\n- Kernel `display` / `result` frames map to:\n  - `application/x-omp-status` → status event\n  - `image/png` → image output\n  - `application/json` → JSON output\n  - `text/markdown` → markdown output\n  - `text/plain` → text output\n  - `text/html` → HTML converted to markdown with `htmlToBasicMarkdown()`\n- Interactive stdin is rejected: a stdin-flagged result returns exit code `1` with `Kernel requested stdin; interactive input is not supported.`\n\n### Oneshot completion helper (`completion`)\n\nBoth runtimes expose `completion()` — a single stateless completion against a model tier. It is intentionally minimal: no conversation history, no agent-visible tools, pure text in / text (or object) out. Implemented host-side in `packages/coding-agent/src/eval/completion-bridge.ts` and routed through the existing tool bridge under the reserved name `__completion__`.\n\n- Signatures:\n  - JS: `await completion(prompt, { model?, system?, schema? })`\n  - Python: `completion(prompt, *, model=\"default\", system=None, schema=None)`\n- `model` selects a tier (default `\"default\"`):\n  - `\"smol\"` → `pi/smol` role (fast / cheap)\n  - `\"default\"` → the session's active model, falling back to the `pi/default` role\n  - `\"slow\"` → `pi/slow` role; requests high reasoning effort only on reasoning-capable models\n- `system` (optional) supplies a system prompt.\n- `schema` (optional) is a plain JSON-Schema object. When present, the model is forced to call a single synthetic `respond` tool with that schema (loose, non-strict), and the helper returns the parsed object. When absent, the helper returns the completion string.\n- Errors surface as exceptions: unresolved tier, missing API key, an `error`/`aborted` stop reason, or empty output each raise.\n\n### Subagent helper (`agent`)\n\nBoth runtimes expose `agent()` — a single subagent invocation routed through `packages/coding-agent/src/eval/agent-bridge.ts` into the same `runSubprocess(...)` path used by the `task` tool. It uses the current eval session's spawn policy and inherits the parent eval executor id, so parent and subagent code share JS/Python runtime state.\n\n- Signatures:\n  - JS: `await agent(prompt, agentType?, model?, label?, schema?)` or `await agent(prompt, { agentType?, model?, label?, schema? })`\n  - Python: `agent(prompt, *, agent_type=\"task\", model=None, label=None, schema=None)`\n- `agentType` / `agent_type` defaults to the bundled `task` agent and resolves through normal agent discovery, so project and user agents work.\n- `model` overrides the selected agent's model. Without it, normal per-agent settings and the agent frontmatter model apply.\n- Shared background is passed via files: write a `local://` file and reference it in the prompt. `label` controls the `agent://<id>` output label prefix.\n- `schema` passes a JSON Schema to the subagent structured-output path. When present, the helper parses the final JSON text and returns an object.\n- Spawn restrictions use `session.getSessionSpawns()` exactly like the `task` tool. Eval-driven subagent recursion is capped at depth 3.\n- JS and Python both expose `parallel(thunks)` and `pipeline(items, ...stages)`; both use a bounded async/threaded pool whose width tracks the `task.maxConcurrency` setting (the same ceiling the `task` tool uses; `0` = run every item at once), preserve item order, and propagate rejections. The width is fetched live from the host via the `__concurrency__` bridge, so the helpers no longer take a `concurrency` argument.\n- Errors surface as exceptions: unknown or disabled agent, disallowed spawn, recursion cap, subagent failure, or invalid structured output all fail the eval cell.\n\n### Multi-language call behavior\n\nA single tool call can mix Python and JS cells. Persistence is per language runtime:\n\n- `reset: true` on a Python cell does not touch JS state\n- `reset: true` on a JS cell does not touch Python state\n- each backend keeps its own retained session keyed from the same session-derived ID\n\n## Side Effects\n\n- Filesystem\n  - JS/Python prelude helpers can read, write, append, diff, and traverse filesystem paths under the session cwd or absolute paths.\n  - JS helper `read()` rejects protocol URIs (`://`) and directory paths; use `tool.read(...)` for internal URLs or reader-mode behavior.\n  - Output may spill to an artifact file via `OutputSink`.\n- Network\n  - Python backend speaks NDJSON to a local `python3` subprocess over stdin/stdout (no network).\n  - JS runtime exposes `fetch` and `tool.<name>()`; those tools may perform additional network I/O.\n- Subprocesses / native bindings\n  - Python availability check runs `<python> -c ...`.\n  - Python backend spawns one `python -u runner.py` subprocess per kernel; cancellation sends `SIGINT`. Details in `docs/python-repl.md`.\n  - `agent()` runs one in-process subagent via the task executor; that subagent may use its configured tools.\n- Session state\n  - `session.assertEvalExecutionAllowed?.()` can block execution.\n  - `session.trackEvalExecution?.(...)` can register cancellable eval work.\n  - `session.getSessionFile?.()`, `session.getEvalSessionId?.()`, and `session.getEvalKernelOwnerId?.()` influence VM/kernel reuse and artifact lookup.\n  - JS VM contexts persist across eval calls until reset/disposal.\n  - Python retained kernels persist until reset, owner cleanup, or process exit.\n  - `agent()` allocates `agent://<id>` output artifacts and reuses the parent's eval executor id.\n- User-visible prompts / interactive UI\n  - none; stdin requests are rejected programmatically\n- Background work / cancellation\n  - Python retained kernels have heartbeat and idle cleanup timers.\n  - Cancellation hard-kills/resets the shared executor for that backend: JS terminates the worker, Python sends SIGINT and may escalate to subprocess shutdown.\n\n## Limits & Caps\n\n- Per-cell timeout default: 30s (applied when `timeout` is omitted in `EvalTool.execute()`; clamped through `TOOL_TIMEOUTS.eval.default` in `packages/coding-agent/src/tools/tool-timeouts.ts`)\n- Schema-level `timeout` range: integer `1..3600` seconds (enforced by Zod on the cell schema)\n- Timeout clamp at runtime: 1s minimum, 3600s maximum (`TOOL_TIMEOUTS.eval` in `packages/coding-agent/src/tools/tool-timeouts.ts`)\n- Transcript code/output preview: 10 lines by default (`EVAL_DEFAULT_PREVIEW_LINES` in `packages/coding-agent/src/tools/eval-render.ts`, re-exported from `eval.ts`)\n- Output truncation window: 50KB default (`DEFAULT_MAX_BYTES` in `packages/coding-agent/src/session/streaming-output.ts`)\n- Output line cap inside truncation helpers: 3000 lines (`DEFAULT_MAX_LINES` in `packages/coding-agent/src/session/streaming-output.ts`)\n- Streaming tail buffer for live updates: `DEFAULT_MAX_BYTES * 2` = 100KB (`packages/coding-agent/src/tools/eval.ts`)\n- JS/Python `parallel()` / `pipeline()` helper pool width: the `task.maxConcurrency` setting (default 32; `0` = unbounded), resolved live via the `__concurrency__` bridge (`packages/coding-agent/src/eval/concurrency-bridge.ts`)\n- Eval-driven `agent()` recursion cap: task depth 3 (`EVAL_AGENT_MAX_DEPTH`)\n- Python kernel startup wait: 10s (`STARTUP_TIMEOUT_MS` in `packages/coding-agent/src/eval/py/kernel.ts`)\n- Python kernel shutdown grace per escalation step (`exit` request → `SIGTERM` → `SIGKILL`): 1000ms (`SHUTDOWN_GRACE_MS` in `packages/coding-agent/src/eval/py/kernel.ts`)\n- Python SIGINT escalation window: 5s without a `done` frame before the subprocess is killed (`INTERRUPT_ESCALATION_MS` in `packages/coding-agent/src/eval/py/kernel.ts`)\n- Python auto-restart budget: a dead retained kernel is replaced and the cell retried once per execution (`executeOnSession` in `packages/coding-agent/src/eval/py/executor.ts`)\n\n## Errors\n\n- Zod validation rejects malformed `cells` arrays before `execute()` runs (missing `language`/`code`, out-of-range `timeout`, empty `cells`).\n- Missing session without proxy executor throws `ToolError(\"Eval tool requires a session when not using proxy executor\")`.\n- Disabled/unavailable backends throw `ToolError` from `resolveBackend()`:\n  - `eval.py = false` and a `py` cell is requested\n  - `eval.js = false` and a `js` cell is requested\n  - Python kernel unavailable and a `py` cell is requested\n- JS runtime exceptions are converted into text output plus `exitCode: 1`; cancellations return `cancelled: true` and may append `Command timed out`.\n- Python execution errors from the kernel become text output and `exitCode: 1`; later cells are skipped.\n- Python stdin requests are treated as errors with the message `Kernel requested stdin; interactive input is not supported.`\n- Cancellation is returned, not thrown, once backend execution has started. The tool formats it as a cell failure and sets `details.isError = true`.\n- If output truncates, the tool still succeeds; truncation is surfaced through `details.meta` and artifact-backed full output when available.\n\n## Shared executor trade-offs\n\n- Parent agents and subagents share eval state bidirectionally when a subagent inherits the parent's executor id. Mutations in either direction are visible to the other participant.\n- Async regions of concurrent runs can interleave. Synchronous JS still blocks the VM event loop; synchronous Python still contends on the GIL.\n- Cancelling one run is destructive to the shared backend executor. This is intentional: JS worker termination and Python SIGINT/subprocess shutdown are the only reliable way to interrupt arbitrary user code.\n- `reset: true` is destructive for every live run on that backend session id. Concurrent Python resets coalesce — a reset already in flight is awaited rather than duplicated, and runs queued behind it proceed on the freshly-restarted kernel.\n\n## Notes\n\n- Backend selection is strictly explicit per cell: `language` must be `\"py\"` or `\"js\"`. The previous `*** Cell` header parser, the `eval.lark` constrained grammar, and the sniffer-based fallback have all been removed.\n- `EvalTool.customFormat` no longer exists. Tool calls flow through the standard JSON schema; there is no Lark-constrained sampling path.\n- `tool.<name>()` exists in both JS and Python. Python calls route through a per-run loopback bridge keyed by the current cell id.\n- JS helper paths reject protocol URIs (`://`) in `resolveRegularFile()` for `read()`, and resolve other paths against the session cwd or absolute filesystem path. Use `tool.read(...)` or another tool explicitly for internal URLs.\n- Python helper `output(...)` depends on `PI_ARTIFACTS_DIR` or `PI_SESSION_FILE`; it fails outside a session-backed run.\n- `display()` can produce text and structured outputs from the same value; the renderer prefers markdown over `text/plain` when both exist.\n- JS static imports are rewritten only at top level. Nested imports stay invalid and surface normal JS syntax/runtime errors.\n- `EvalTool` is `concurrency = \"exclusive\"` within one agent session, but parent and subagent sessions can run eval concurrently when they share an inherited executor id.\n- The tool description shown to the model is templated by backend availability (`getEvalToolDescription()`); if Python is unavailable, the prompt omits Python-specific instructions.\n",
 	"tools/find.md": "# find\n\n> Find filesystem paths by glob; use `search` when you need content matches instead of path matches.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/find.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/find.md`\n- Key collaborators:\n  - `packages/coding-agent/src/tools/path-utils.ts` — normalize inputs; split base path vs glob.\n  - `packages/coding-agent/src/tools/list-limit.ts` — apply result-count caps.\n  - `packages/coding-agent/src/session/streaming-output.ts` — truncate text output at byte cap.\n  - `packages/coding-agent/src/tools/tool-result.ts` — build `content` and `details.meta`.\n  - `packages/coding-agent/src/tools/output-meta.ts` — encode limit / truncation metadata.\n  - `packages/coding-agent/src/tools/tool-errors.ts` — map user-facing tool errors.\n  - `packages/coding-agent/src/tools/index.ts` — register the built-in local implementation.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `paths` | `string[]` | Yes | One or more globs, files, directories, or internal URLs with backing files. Empty strings are rejected. Single entries accidentally joined with comma, semicolon, or whitespace are expanded only after existence validation; existing paths containing delimiters stay intact. Each entry becomes its own walk root; multi-entry calls run those scans concurrently. |\n| `hidden` | `boolean` | No | Whether hidden files are included. Defaults to `true` (`hidden ?? true`). |\n| `gitignore` | `boolean` | No | Whether `.gitignore` is respected during local native globbing. Defaults to `true`; set `false` to include gitignored files. |\n| `limit` | `number` | No | Max returned paths. Defaults to `200`; finite positive inputs are floored then clamped to `1..200`. |\n| `timeout` | `number` | No | Timeout in seconds. Defaults to `5`; clamped to `0.5..60`. On timeout, returns partial matches collected so far with a timeout notice and `truncated: true`. |\n\n## Outputs\nThe tool returns a single text block plus structured `details`.\n\n- Success text: matching paths grouped as a multi-level, prefix-folded directory tree (`formatGroupedPaths()`): one `#` per nesting level, single-child directory chains fold into one header (`# a/b/c/`), and files are listed bare under the deepest owning header; root-level matches are listed without a header. Directory matches carry a trailing `/`. Exact file inputs return that file path as one line.\n- Empty result text: `No files found matching pattern`, optionally followed by a timeout or missing-path notice.\n- Multi-path partial miss: appends `Skipped missing paths: ...` after the result block, or after the empty-result line.\n- `details` may include:\n  - `scopePath`: display form of the searched root or merged roots.\n  - `fileCount`: number of paths returned after result limiting.\n  - `files`: returned paths as an array.\n  - `truncated`: whether result count or byte truncation occurred.\n  - `resultLimitReached`: reached result limit.\n  - `missingPaths`: skipped missing inputs in multi-path calls.\n  - `truncation` / `meta.limits`: structured truncation and limit metadata for renderers.\n- Streaming: when the runtime supplies `onUpdate`, the local implementation emits incremental newline-delimited text snapshots during globbing, throttled to 200 ms. Final output is grouped; streaming snapshots are not.\n\n## Flow\n\n1. `FindTool.execute()` expands delimiter-flattened local `paths` entries with `expandDelimitedPathEntries(..., parseFindPattern)` unless custom operations are injected. The splitter validates candidate parts by statting their parsed base paths, keeps existing delimiter-containing paths intact, accepts comma/semicolon splits when at least one part resolves, and accepts whitespace splits only when every part resolves.\n2. The tool normalizes each resulting entry with `normalizePathLikeInput()` and `/\\\\/g -> \"/\"` (`packages/coding-agent/src/tools/find.ts`). Empty normalized entries fail with `` `paths` must contain non-empty globs or paths ``.\n3. For multi-path local calls, `partitionExistingPaths(..., parseFindPattern)` (`packages/coding-agent/src/tools/path-utils.ts`) stats each base path. Missing entries are skipped; if all are missing, the tool throws `Path not found: ...`. Single missing paths still hard-fail.\n4. The tool calls `resolveExplicitFindPatterns()` for multi-entry calls; it parses each entry into its own `(basePath, globPattern, hasGlob)` target so every path is walked as its own root (collapsing to a shared ancestor would scan unrelated siblings). Single-entry calls parse with `parseFindPattern()` directly.\n5. `parseFindPattern()` determines `(basePath, globPattern, hasGlob)`:\n   - no glob chars (`*`, `?`, `[`, `{`) => search that path with implicit `**/*`.\n   - glob in the first segment => search from `.` and, unless the pattern already starts with `**/`, prefix it with `**/`.\n   - glob later in the path => split at the first glob-bearing segment.\n6. `resolveToCwd()` converts the base path to an absolute path under the session cwd. A resolved `/` is rejected with `Searching from root directory '/' is not allowed`.\n7. `limit` defaults to `DEFAULT_LIMIT` (`200`), must be positive and finite, is floored, then clamped to `MAX_LIMIT` (`200`). `hidden` and `gitignore` both default to `true`. `timeout` is converted to milliseconds and clamped to `500..60_000` before building an `AbortSignal.timeout(...)`.\n8. Execution then branches:\n   - **Custom operations branch**: if `FindToolOptions.operations.glob` exists, the tool checks existence with `operations.exists()`, short-circuits exact-file inputs via `operations.stat()` when available, then calls `operations.glob(globPattern, searchPath, { ignore: [\"**/node_modules/**\", \"**/.git/**\"], limit })`.\n   - **Built-in local branch**: the tool stats each target's `searchPath`. Exact-file inputs return immediately. Directory inputs call `natives.glob()` with `hidden`, `maxResults: effectiveLimit`, `sortByMtime: true`, `gitignore: useGitignore`, `recursive: false` (recursion comes from the `**/` prefix `parseFindPattern()` adds), and the combined abort signal; multi-target calls run their globs concurrently.\n9. In the local branch, optional `onMatch` callbacks convert each match to a cwd-relative display path and emit throttled progress updates.\n10. After native glob returns, JS merges per-target results, deduplicates repeated display paths, and sorts the merged list by `mtime` descending before formatting paths.\n11. `buildResult()` applies `applyListLimit()` to cap the array again at `effectiveLimit`, formats paths with `formatGroupedPaths()` (from `@oh-my-pi/pi-utils`), appends notices, then runs `truncateHead()` with `maxLines: Number.MAX_SAFE_INTEGER`. In practice this leaves the 50 KB byte cap in place while disabling the default 3000-line cap.\n12. `toolResult()` packages text plus `details`, and records result-limit / truncation metadata for renderers.\n\n## Modes / Variants\n- **Exact file path**: if the parsed input has no glob and the resolved path stats as a file, output is that one path.\n- **Directory path**: if the parsed input has no glob and stats as a directory, the tool searches it with implicit `**/*`.\n- **Single glob path**: one input parsed by `parseFindPattern()`.\n- **Multi-path search**: multiple inputs resolved by `resolveExplicitFindPatterns()` into per-entry targets, each walked as its own root concurrently and merged afterwards.\n- **Partial multi-path search with missing inputs**: local multi-path calls skip missing base paths and surface them as `missingPaths` / `Skipped missing paths: ...`.\n- **Internal URL input**: supported when the internal router resolves the URL to a backing file. Internal URL globs are rejected.\n- **Custom delegated search**: uses injected `FindOperations` instead of local fs + native glob.\n\n## Side Effects\n- Filesystem\n  - Stats the resolved base path, and in local multi-path mode stats every candidate base path up front.\n  - Does not write files.\n- Subprocesses / native bindings\n  - Built-in local mode calls the native `@oh-my-pi/pi-natives` glob implementation.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - Emits structured progress updates when `onUpdate` is provided.\n  - Adds truncation / limit metadata to the tool result.\n- Background work / cancellation\n  - Local globbing is cancellable through the caller abort signal plus the configured internal timeout.\n\n## Limits & Caps\n- Default result limit: `200` (`DEFAULT_LIMIT` in `packages/coding-agent/src/tools/find.ts`).\n- Maximum result limit: `200` (`MAX_LIMIT`); larger inputs are clamped.\n- Local glob timeout: default `5000` ms, clamped to `500..60_000` ms.\n- Output byte cap: `50 * 1024` bytes (`DEFAULT_MAX_BYTES` in `packages/coding-agent/src/session/streaming-output.ts`).\n- Default generic line cap in `truncateHead()` is `3000`, but `find` overrides `maxLines` to `Number.MAX_SAFE_INTEGER`, so byte size — not line count — is the practical output truncation cap.\n- Streaming update throttle: `200` ms between `onUpdate` emissions.\n- Sort order: most recent `mtime` first in the built-in local branch and promised in the prompt. The tool re-sorts in JS even though native glob receives `sortByMtime: true` so native code can still stop early at `maxResults`.\n\n## Errors\n- User-facing `ToolError`s from `FindTool.execute()` include:\n  - `` `paths` must contain non-empty globs or paths ``\n  - `Path not found: ...`\n  - `Searching from root directory '/' is not allowed`\n  - `Limit must be a positive number`\n  - `Path is not a directory: ...`\n  - timeout result text is `find timed out after <seconds>s; returning <N> partial matches — increase timeout or narrow pattern` and is returned as a successful, truncated partial result rather than an error.\n- If the caller aborts, the local branch converts `AbortError` into `ToolAbortError`.\n- Non-`ENOENT` stat failures and other unexpected errors are rethrown.\n- Empty matches are not errors; they return the no-files text result.\n\n## Notes\n- Reach for `find` for filename / path discovery. Reach for `search` when the selection criterion is file contents or regex matches; `search` takes a `pattern` and returns anchored content matches, while `find` only returns matching paths (`packages/coding-agent/src/prompts/tools/find.md`, `packages/coding-agent/src/prompts/tools/search.md`).\n- Bare top-level globs are made recursive. `*.ts` is parsed as base `.` plus glob `**/*.ts`; `src/*.ts` stays rooted at `src` with a non-recursive `*.ts` segment; `src/**/*.ts` preserves explicit recursion.\n- `.gitignore` defaults to enabled in the built-in local branch. Use `gitignore: false` to disable it for native traversal.\n- `hidden` defaults to `true`; hidden-file exclusion is opt-out, not opt-in.\n- Multi-path missing-input tolerance applies in both branches, but only the built-in local branch surfaces `missingPaths` / `Skipped missing paths: ...`. The custom-operations branch hard-fails a missing `searchPath` only for single-input calls; in multi-input calls a missing target silently contributes no results.\n- The custom `FindOperations.glob()` hook receives `ignore` and `limit`, but not the `hidden` flag or an explicit `.gitignore` toggle. A remote delegate must account for that itself if it wants parity with the local branch.\n- Built-in local globbing does not force `fileType: File`; it can return files and directories from native glob. Directory outputs also occur through exact-path passthrough or custom delegates that return them.",
 	"tools/github.md": "# github\n\n> Dispatch GitHub CLI operations for repositories, issues, pull requests, search, and Actions run watching.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/gh.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/github.md`\n- Key collaborators:\n  - `packages/coding-agent/src/tools/gh-format.ts` — shorten commit SHAs for summaries.\n  - `packages/coding-agent/src/tools/gh-renderer.ts` — TUI rendering, especially `run_watch` live/result views.\n  - `packages/coding-agent/src/utils/git.ts` — `gh`/`git` process wrappers, repo locking, branch config writes.\n  - `packages/utils/src/dirs.ts` — base directory for dedicated PR worktrees.\n  - `packages/coding-agent/src/sdk.ts` — session artifact allocation hook.\n  - `packages/coding-agent/src/session/artifacts.ts` — artifact filename format `<id>.<toolType>.log`.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `op` | `\"repo_view\" \\| \"pr_create\" \\| \"pr_checkout\" \\| \"pr_push\" \\| \"search_issues\" \\| \"search_prs\" \\| \"search_code\" \\| \"search_commits\" \\| \"search_repos\" \\| \"run_watch\"` | Yes | Dispatch selector. `GithubTool.execute()` switches only on this field. |\n| `repo` | `string` | No | `owner/repo` override. Ignored when the identifier argument is already a full GitHub URL. For `search_issues`/`search_prs`/`search_code`/`search_commits`, defaults to the current checkout's `owner/repo` when omitted (skipped when the query already contains a `repo:`/`org:`/`user:`/`owner:` qualifier or when current-repo resolution fails). Required in practice when `gh` cannot infer repo context from the current checkout. |\n| `branch` | `string` | No | Used by `repo_view`, `pr_push`, and `run_watch`. `run_watch` falls back to current git branch when `run` is omitted; `pr_push` falls back to current branch. |\n| `pr` | `string \\| string[]` | No | Used by `pr_checkout`. Each item may be a PR number, branch name, or GitHub PR URL. Array form enables batching. Omitted means current branch PR. |\n| `force` | `boolean` | No | Used only by `pr_checkout`. Defaults to `false`; allows resetting an existing `pr-<number>` local branch to the PR head commit. |\n| `forceWithLease` | `boolean` | No | Used only by `pr_push`; passed through to git push. |\n| `title` | `string` | No | Used only by `pr_create`. Required unless `fill` is `true`. |\n| `body` | `string` | No | Used only by `pr_create`. Mutually exclusive with `fill`. Empty/omitted body becomes `--body \"\"` to suppress the interactive editor. Non-empty body is written to a temp file and passed as `--body-file`. |\n| `base` | `string` | No | Used only by `pr_create`; passed as `--base`. |\n| `head` | `string` | No | Used only by `pr_create`; passed as `--head`. |\n| `draft` | `boolean` | No | Used only by `pr_create`. Defaults to `false`. |\n| `fill` | `boolean` | No | Used only by `pr_create`. Defaults to `false`. Mutually exclusive with `title` and `body`. |\n| `reviewer` | `string[]` | No | Used only by `pr_create`; each entry becomes `--reviewer`. |\n| `assignee` | `string[]` | No | Used only by `pr_create`; each entry becomes `--assignee`. |\n| `label` | `string[]` | No | Used only by `pr_create`; each entry becomes `--label`. |\n| `query` | `string` | No | Used by all `search_*` ops. Required by local validation only for `search_code`; the other search ops compose it with optional date/repo/type qualifiers and send the result to GitHub. |\n| `since` | `string` | No | Lower date bound for `search_issues`, `search_prs`, `search_commits`, and `search_repos`. Accepts relative durations (`3d`, `12h`, `2w`, `2mo`, `1y`), `YYYY-MM-DD`, or an ISO datetime. Rejected for `search_code`. |\n| `until` | `string` | No | Upper date bound for `search_issues`, `search_prs`, `search_commits`, and `search_repos`. Same formats as `since`. Rejected for `search_code`. |\n| `dateField` | `\"created\" \\| \"updated\"` | No | Date qualifier field for issue/PR/repo search. Defaults to `created`; repo search maps `updated` to GitHub's `pushed:` qualifier. Ignored for commit search, which always uses `committer-date:`. |\n| `limit` | `number` | No | Used by all `search_*` ops. Defaults to `10`, floored, clamped to `50`, and must be `> 0`. |\n| `run` | `string` | No | Used only by `run_watch`. Must be a numeric run ID or full GitHub Actions run URL. |\n| `tail` | `number` | No | Used only by `run_watch`. Defaults to `15`, floored, clamped to `200`, and must be `> 0`. |\n\n## Outputs\nThe tool returns a single text result built by `buildTextResult()` in `packages/coding-agent/src/tools/gh.ts`.\n\n- `content`: one text block. Multi-item ops join sections with blank lines and `---` separators.\n- `sourceUrl`: set for single repo/PR/run results when a canonical URL is known.\n- `details`: optional structured metadata used by the TUI renderer.\n  - Common fields: `artifactId`, `repo`, `branch`, `worktreePath`, `remote`, `remoteBranch`, `headSha`, `runId`, `runIds`, `status`, `conclusion`, `failedJobs`.\n  - `pr_checkout` adds `checkouts: GhPrCheckoutSummary[]`.\n  - `run_watch` adds `watch: GhRunWatchViewDetails`, which drives the custom live/result renderer in `packages/coding-agent/src/tools/gh-renderer.ts`.\n- Artifact trailer: when `artifactId` is present, the text body gets an appended line like `Full failed-job logs: artifact://<id>`.\n  - `run_watch` allocates artifacts with `session.allocateOutputArtifact(\"github\")`; persistent sessions therefore save failed-log bodies as `<artifact-dir>/<id>.github.log`.\n\n`run_watch` is the only streaming op. It emits `onUpdate` snapshots while polling, then returns one final text result.\n\n## Flow\n1. `GithubTool.createIf()` exposes the tool only when `git.github.available()` finds `gh` on `PATH`.\n2. `GithubTool.execute()` wraps dispatch in `untilAborted()` and switches on `params.op`.\n3. Each op normalizes optional strings, arrays, booleans, and numeric caps locally in `packages/coding-agent/src/tools/gh.ts`.\n4. CLI execution goes through `git.github.run/json/text()` in `packages/coding-agent/src/utils/git.ts`:\n   - spawns `gh ...` with `Bun.spawn()`;\n   - trims stdout/stderr unless `trimOutput: false`;\n   - maps common auth/repo-context failures into tool-facing `ToolError` messages;\n   - `json()` rejects empty or invalid JSON.\n5. Read-style ops (`repo_view`, `search_*`) fetch JSON and format Markdown-like text summaries. Single-issue and single-PR views were moved out of the tool and now resolve through the `issue://` / `pr://` internal URL schemes, which share the same SQLite cache.\n6. PR diffs moved out of the tool. `pr://<N>/diff` lists changed files, `pr://<N>/diff/<i>` slices a single file, and `pr://<N>/diff/all` returns the full unified diff — see `docs/tools/read.md`. All three variants share one `gh pr diff` invocation through the `pr-diff` cache row.\n7. `pr_checkout` resolves PR metadata first, then enters `git.withRepoLock()` before any git mutation so parallel checkout calls for the same primary repo do not race on shared `.git` state.\n8. `pr_push` reads PR head metadata back from git branch config, derives a refspec, then pushes with `git.push()`.\n9. `pr_create` shells out once, then best-effort re-reads the created PR for a richer summary.\n10. `run_watch` chooses either run mode (`run` supplied) or commit mode (`run` omitted), polls GitHub Actions APIs every 3 seconds for the first minute and every 15 seconds after that, emits streaming updates, and may save a full failed-log artifact before returning.\n11. Final text goes through `toolResult().text(...)`; if `session.allocateOutputArtifact()` returns a slot, failed-log text is persisted with `Bun.write()`.\n\n## Modes / Variants\n\n### `repo_view`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `repo`, `branch` |\n| `gh` command | `gh repo view [<repo>] [--branch <branch>] --json <GH_REPO_FIELDS>` |\n| Batching | None |\n| Output | `# <owner/repo>` header, description, URL, default branch, requested branch, visibility, permission, primary language, stars, forks, archive/fork flags, updated timestamp, homepage, topics. `sourceUrl = data.url`. |\n\nIf `repo` is omitted, `gh` repository resolution is used.\n\nSingle-issue and single-PR reads live in the `issue://<N>` / `pr://<N>` URL schemes (see `docs/tools/read.md`). They share `~/.omp/cache/github-cache.db` (override via `OMP_GITHUB_CACHE_DB`) and the `github.cache.softTtlSec` / `github.cache.hardTtlSec` / `github.cache.enabled` settings. The cache retains rendered Markdown plus the raw JSON payload returned by `gh`, including private bodies, comments, reviews, and review comments when comments are enabled; rows are scoped by the local GitHub credential fingerprint. Root and repo-scoped reads (`issue://`, `pr://owner/repo`) issue a live `gh issue list` / `gh pr list` for browsing; query params `state`, `limit`, `author`, `label` pass through to `gh` (`issue://` accepts `state=open|closed|all`; `pr://` also accepts `merged`). PR diffs ride the same cache under `pr://<N>/diff[/…]`: the listing, full diff, and per-file slices all share one `pr-diff` row keyed by repo and PR number.\n\n### `pr_create`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` plus either `fill=true` or `title` |\n| Optional fields | `repo`, `title`, `body`, `base`, `head`, `draft`, `fill`, `reviewer[]`, `assignee[]`, `label[]` |\n| `gh` command | `gh pr create ...` with flags assembled from provided fields |\n| Batching | None |\n| Output | `# Created Pull Request ...` summary with URL, state, draft flag, base/head, author, created time, labels, optional body. `sourceUrl` is the created PR URL. |\n\nBranches:\n- `fill && (title || body !== undefined)` throws.\n- Non-empty `body` is written under a temp dir `gh-pr-body-*` in `os.tmpdir()`, passed as `--body-file`, then removed in `finally`.\n- After creation, the tool parses the returned URL and best-effort runs `gh pr view <number> --repo <repo> --json <GH_PR_FIELDS_NO_COMMENTS>`; failures there are swallowed.\n\n### `pr_checkout`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `repo`, `pr`, `force` |\n| `gh` command | For each requested PR: `gh pr view [<pr>] [--repo <repo>] --json <GH_PR_CHECKOUT_FIELDS>`; cross-repo PRs may also call `gh repo view <headRepository> --json <GH_REPO_CLONE_FIELDS>`. |\n| Batching | Yes. `pr` may be `string[]`; each PR is resolved in parallel, but git mutations are serialized per primary repo by `git.withRepoLock()`. |\n| Output | Single PR: checkout/worktree summary plus `details.repo`, `details.branch`, `details.worktreePath`, `details.remote`, `details.remoteBranch`, `details.checkouts`. Batched: `# <n> Pull Request Worktrees (...)` plus one section per PR and aggregated `details.checkouts`. |\n\nWorktree and metadata behavior:\n- Local branch name is always `pr-<number>`.\n- Worktree path is `path.join(getWorktreesDir(), encodeRepoPathForFilesystem(primaryRepoRoot), localBranch)`, where `getWorktreesDir()` is `~/.omp/wt`; effective path is `~/.omp/wt/<encoded-primary-repo-root>/pr-<number>`.\n- Existing worktree detection is by branch ref `refs/heads/pr-<number>` from `git.worktree.list()`.\n- New worktree creation calls `git.worktree.add(repoRoot, finalWorktreePath, localBranch, { signal })` after verifying the path is neither already registered nor already present on disk.\n- For same-repo PRs, remote is `origin`. For cross-repo PRs, the tool resolves a clone URL for the head repo, reuses an existing remote with the same URL when possible, or creates `fork-<owner>` / `fork-<owner>-<n>`.\n- The branch push metadata is persisted with `git config` under the repository's shared `.git/config` as:\n  - `branch.pr-<number>.remote`\n  - `branch.pr-<number>.merge`\n  - `branch.pr-<number>.pushRemote`\n  - `branch.pr-<number>.ompPrHeadRef`\n  - `branch.pr-<number>.ompPrUrl`\n  - `branch.pr-<number>.ompPrIsCrossRepository`\n  - `branch.pr-<number>.ompPrMaintainerCanModify`\n- If `refs/heads/pr-<number>` already exists at a different commit, checkout fails unless `force=true`, in which case `git branch --force` resets it to the fetched PR head.\n- If a matching worktree already exists, the tool reuses it and reports `reused: true`.\n\n### `pr_push`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `branch`, `forceWithLease` |\n| `gh` command | None. This path uses git, not `gh`. |\n| Batching | None |\n| Output | `# Pushed Pull Request Branch` summary with local branch, remote, remote branch, remote URL, PR URL, and force-with-lease flag. `sourceUrl = prUrl` when known. |\n\nPush target resolution reads the `branch.<name>.ompPrHeadRef`, `pushRemote`/`remote`, `ompPrUrl`, `ompPrMaintainerCanModify`, and `ompPrIsCrossRepository` git-config keys written by `pr_checkout`. If the current checked-out branch matches the target branch, the source ref is `HEAD`; otherwise it pushes `refs/heads/<branch>`. The refspec is `HEAD:refs/heads/<headRef>` or `refs/heads/<branch>:refs/heads/<headRef>`.\n\n### `search_issues`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `repo`, `query`, `limit`, `since`, `until`, `dateField` |\n| `gh` command | `gh api -X GET /search/issues -f q=\"<query> [date qualifier] [repo:<repo>] is:issue\" -F per_page=<limit>` |\n| Batching | None |\n| Output | `# GitHub issues search`, echoed query, optional repo, result count, then one bullet per issue with repo/state/author/labels/timestamps/URL. |\n\n`repo` defaults to the current checkout's `owner/repo` via `resolveSearchRepoScope()` when omitted. The default is suppressed when the composed query already contains a leading `repo:`/`org:`/`user:`/`owner:` qualifier or when `gh repo view` fails to resolve the current checkout (e.g. outside a github remote).\n\n### `search_prs`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `repo`, `query`, `limit`, `since`, `until`, `dateField` |\n| `gh` command | `gh api -X GET /search/issues -f q=\"<query> [date qualifier] [repo:<repo>] is:pr\" -F per_page=<limit>` |\n| Batching | None |\n| Output | Same shape as `search_issues`, labeled as pull requests. |\n\n`repo` defaults to the current checkout's `owner/repo` as in `search_issues`.\n\n### `search_code`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op`, `query` |\n| Optional fields | `repo`, `limit` |\n| `gh` command | `gh api -X GET /search/code -f q=\"<query> [repo:<repo>]\" -F per_page=<limit> -H \"Accept: application/vnd.github.text-match+json\"` |\n| Batching | None |\n| Output | `# GitHub code search`, result count, then one bullet per match with path, repo, short commit SHA, URL, and first normalized text-match fragment line when present. |\n\n`repo` defaults to the current checkout's `owner/repo` as in `search_issues`. `since` and `until` are explicitly rejected for this op because GitHub code search has no supported date qualifier.\n\n### `search_commits`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `repo`, `query`, `limit`, `since`, `until`, `dateField` (accepted but ignored; commit searches use `committer-date`) |\n| `gh` command | `gh api -X GET /search/commits -f q=\"<query> [committer-date qualifier] [repo:<repo>]\" -F per_page=<limit>` |\n| Batching | None |\n| Output | `# GitHub commits search`, result count, then one bullet per commit: short SHA + first commit-message line, repo, author, date, URL. |\n\n`repo` defaults to the current checkout's `owner/repo` as in `search_issues`.\n\n### `search_repos`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `query`, `limit`, `since`, `until`, `dateField` |\n| `gh` command | `gh api -X GET /search/repositories -f q=\"<query> [date qualifier]\" -F per_page=<limit>` |\n| Batching | None |\n| Output | `# GitHub repositories search`, result count, then one bullet per repo with first description line, language, stars, forks, open issues, visibility, archive/fork flags, updated time, URL. |\n\n`repo` is intentionally not used for this op. If `query`, `since`, and `until` are all omitted, the tool sends an empty GitHub repository-search query and the GitHub API may reject it.\n\n### `run_watch`\n\n| Aspect | Value |\n| --- | --- |\n| Required fields | `op` |\n| Optional fields | `repo`, `branch`, `run`, `tail` |\n| `gh` command | Repo resolution: `gh repo view --json nameWithOwner -q .nameWithOwner` when `repo` and run URL repo are both absent. Single-run mode uses `gh api --method GET /repos/<repo>/actions/runs/<runId>` and `gh api --method GET /repos/<repo>/actions/runs/<runId>/jobs`. Commit mode uses `gh api --method GET /repos/<repo>/branches/<branch>`, `gh api --method GET /repos/<repo>/actions/runs`, `gh api --method GET /repos/<repo>/actions/runs/<runId>/jobs`, and `gh api /repos/<repo>/actions/jobs/<jobId>/logs` for failed jobs. |\n| Batching | Implicit batching only in commit mode: all workflow runs for one commit are tracked together. |\n| Output | Streaming watch snapshots via `onUpdate`, then a final text report. On failure, appends `Full failed-job logs: artifact://<id>` and sets `details.artifactId`. |\n\nWatch flow:\n- `run` parsing accepts either a decimal run ID or a full run URL. URL repo must match explicit `repo` when both are given.\n- Poll interval is `3` seconds (`RUN_WATCH_INTERVAL_DEFAULT`) for the first `60` seconds of the watch (`RUN_WATCH_FAST_WINDOW_MS`), then `15` seconds (`RUN_WATCH_INTERVAL_SLOW`). Rate-limited poll errors back off at the slow interval and are retried up to `5` consecutive failures (`RUN_WATCH_MAX_POLL_FAILURES`). Commit mode gives up with a clear message after `90` seconds if no runs ever appear (`RUN_WATCH_NO_RUNS_GIVE_UP_MS`).\n- Failure grace period is fixed at 5 seconds (`RUN_WATCH_GRACE_DEFAULT`). When any failed job appears before completion, the tool emits a note, waits once, re-fetches state, then collects logs so concurrent failures are included.\n- Failed-job logs are fetched with `gh api /repos/<repo>/actions/jobs/<jobId>/logs` via `git.github.run()`, not `json()`. Non-zero exit leaves `available: false` instead of failing the whole watch.\n- Inline result includes only the last `tail` lines per failed job. The saved artifact contains full logs (`mode: \"full\"`).\n- In commit mode, success is intentionally double-checked: once all known runs are successful, the tool waits one more poll interval and succeeds only if the set of run IDs is unchanged. This avoids returning before late workflow runs appear for the same commit.\n- `details.watch` drives a specialized renderer in `packages/coding-agent/src/tools/gh-renderer.ts`; non-watch results fall back to generic text rendering.\n\n## Side Effects\n- Filesystem\n  - `pr_create` may create a temp dir under `os.tmpdir()` named `gh-pr-body-*`, write `body.md`, then remove the dir in `finally`.\n  - `pr_checkout` may create directories under `~/.omp/wt/<encoded-primary-repo-root>/` and add git worktrees there.\n  - `run_watch` may write a session artifact with full failed-job logs.\n- Network\n  - Every op shells out to `gh`, which then talks to GitHub APIs except `pr_push`.\n  - `pr_push` uses git network transport to the configured remote.\n- Subprocesses / native bindings\n  - All `gh` calls use `Bun.spawn([\"gh\", ...args])`.\n  - `pr_checkout` and `pr_push` also invoke git helpers from `packages/coding-agent/src/utils/git.ts`.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - `run_watch` consumes `session.allocateOutputArtifact()` when failed-job logs are persisted.\n  - Returned `details` objects carry run/checkouts metadata for the renderer/UI.\n- User-visible prompts / interactive UI\n  - `gh` interactive editor fallback is suppressed for `pr_create` by forcing either `--body-file` or `--body \"\"`.\n  - `gh-renderer` provides compact headers for all ops and a custom live watch view for `run_watch`.\n- Background work / cancellation\n  - `run_watch` loops until success/failure and uses `scheduler.wait()` between polls.\n  - `GithubTool.execute()` is wrapped in `untilAborted()`; `git.github.run()` forwards the abort signal into `Bun.spawn()`.\n\n## Limits & Caps\n- Search result default: `10` (`SEARCH_LIMIT_DEFAULT` in `packages/coding-agent/src/tools/gh.ts`).\n- Search result max: `50` (`SEARCH_LIMIT_MAX`).\n- PR file preview inside the `pr://` view: first `50` files only (`FILE_PREVIEW_LIMIT` in `gh.ts`).\n- Run-watch poll interval: `3s` for the first `60s`, then `15s` (`RUN_WATCH_INTERVAL_DEFAULT`, `RUN_WATCH_FAST_WINDOW_MS`, `RUN_WATCH_INTERVAL_SLOW`); commit mode with no runs gives up after `90s` (`RUN_WATCH_NO_RUNS_GIVE_UP_MS`); up to `5` consecutive rate-limited poll failures are tolerated (`RUN_WATCH_MAX_POLL_FAILURES`).\n- Run-watch failure grace period: `5s` (`RUN_WATCH_GRACE_DEFAULT`).\n- Run-watch failed-log tail default: `15` lines (`RUN_WATCH_TAIL_DEFAULT`).\n- Run-watch failed-log tail max: `200` lines (`RUN_WATCH_TAIL_MAX`).\n- PR review comments page size: `100` (`REVIEW_COMMENTS_PAGE_SIZE`).\n- Actions jobs page size: `100` (`RUN_JOBS_PAGE_SIZE`).\n- Search and tail numeric inputs are floored with `Math.floor()`, clamped to the max, and rejected when non-finite or `<= 0`.\n- `pr_checkout` batch fan-out is unbounded in tool code; all requested PRs are launched with `Promise.all()`.\n\n## Errors\n- Tool creation is skipped entirely when `gh` is not installed.\n- `git.github.run()` throws `ToolError(\"GitHub CLI (gh) is not installed...\")` if `gh` is missing at execution time.\n- `git.github.text/json()` map common failures to model-facing messages:\n  - not authenticated → `GitHub CLI is not authenticated. Run \\`gh auth login\\`.`\n  - missing repo context without explicit `repo` → `GitHub repository context is unavailable. Pass \\`repo\\` explicitly or run the tool inside a GitHub checkout.`\n  - otherwise stderr/stdout text, or fallback `GitHub CLI command failed: gh ...`\n- `json()` also throws on empty stdout or invalid JSON.\n- Local validation errors throw `ToolError`, including:\n  - missing required per-op fields (`query` for `search_code`, `title unless fill=true`)\n  - invalid numeric `limit` / `tail`\n  - invalid `since` / `until` date bound\n  - invalid `run` format\n  - `fill` combined with `title` or `body`\n  - missing git repo / branch / HEAD context for checkout, push, or watch\n  - `pr_push` on a branch without `ompPrHeadRef` metadata\n  - conflicting existing worktree path or branch without `force`\n- `run_watch` treats failed-job log fetches specially: missing log content does not fail the watch; it marks that log `available: false` and prints `Log tail unavailable.` / `Full log unavailable.`.\n- `pr_create` swallows only the post-create best-effort `gh pr view` refresh; the create step itself still fails normally.\n\n## Notes\n- `appendRepoFlag()` intentionally skips `--repo` when the identifier argument is already a full GitHub URL; that lets `gh` derive repo/number from the URL.\n- `normalizePrIdentifierList()` accepts `reviewer`, `assignee`, and `label` arrays too; the helper name is broader than its callers.\n- `pr_push` depends on `pr_checkout` having run first for that local branch; there is no alternate metadata source.\n- `pr_checkout` stores push metadata in branch config, not in the worktree directory. Reusing the same `pr-<number>` branch reuses those config keys.\n- Worktree write serialization is keyed by the primary repo root, not the current worktree path, because git worktrees share `.git/config`, `packed-refs`, commit-graph, and worktree metadata files.\n- `search_repos` is the only search op that never forwards `repo`; repository scoping must be expressed in the query itself.\n- `run_watch` success on commit mode means “all observed runs succeeded and no additional runs appeared one poll later”, not merely “latest poll looked green”.\n- The TUI renderer collapses failed log previews unless the result view is expanded; the underlying text result still contains the same tailed lines plus any artifact reference.\n",
 	"tools/inspect_image.md": "# inspect_image\n\n> Send a local image file to a vision-capable model and return text analysis.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/inspect-image.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/inspect-image.md`\n- Key collaborators:\n  - `packages/coding-agent/src/tools/inspect-image-renderer.ts` — TUI call/result rendering.\n  - `packages/coding-agent/src/utils/image-loading.ts` — path resolution, type detection, size gate, optional resize.\n  - `packages/coding-agent/src/utils/image-resize.ts` — downscale and recompress oversized images.\n  - `packages/coding-agent/src/tools/path-utils.ts` — resolve input path relative to session cwd.\n  - `packages/utils/src/mime.ts` — detect supported image formats from file bytes.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `path` | `string` | Yes | Image path passed to `loadImageInput`; resolved relative to `session.cwd` by `resolveReadPath(...)`. |\n| `question` | `string` | Yes | User prompt sent as a text content block alongside the image. |\n\n## Outputs\nThe tool returns a single `AgentToolResult`:\n\n- `content`: one text block, `[{ type: \"text\", text }]`, where `text` is the concatenated assistant text content from the model response.\n- `details`:\n  - `model`: `<provider>/<id>` of the selected model.\n  - `imagePath`: resolved filesystem path returned by `loadImageInput(...)`.\n  - `mimeType`: MIME type actually sent to the model after optional resize/re-encode.\n\nModel-visible output is single-shot, not streamed by this tool.\n\nTUI rendering adds presentation-only truncation from `packages/coding-agent/src/tools/inspect-image-renderer.ts`:\n\n- call preview truncates `question` to 100 columns,\n- result view shows 4 lines collapsed or 16 lines expanded,\n- each rendered output line is truncated to 120 columns,\n- footer metadata shows `model · mimeType` when present.\n\n## Flow\n1. `InspectImageTool.execute(...)` rejects immediately if `images.blockImages` is enabled in session settings.\n2. It reads `session.modelRegistry`; missing registry, empty registry, missing API key, or unresolved model each raise `ToolError` from `packages/coding-agent/src/tools/inspect-image.ts`.\n3. Model selection tries, in order, `pi/vision`, `pi/default`, the active model string from the session, then `availableModels[0]`. `expandRoleAlias(...)` and `resolveModelFromString(...)` handle each lookup.\n4. The chosen model must advertise `input.includes(\"image\")`; otherwise execution fails before reading the file.\n5. `loadImageInput(...)` in `packages/coding-agent/src/utils/image-loading.ts` resolves the path with `resolveReadPath(...)`, detects MIME type with `readImageMetadata(...)`, and rejects files larger than `MAX_IMAGE_INPUT_BYTES` (`20 * 1024 * 1024`, 20 MiB) using `ImageInputTooLargeError`.\n6. `readImageMetadata(...)` in `packages/utils/src/mime.ts` inspects file headers only. Supported detected MIME types are `image/png`, `image/jpeg`, `image/gif`, and `image/webp`.\n7. If `images.autoResize` is true, `loadImageInput(...)` calls `resizeImage(...)`. Resize failures are swallowed there and the original bytes are kept.\n8. If MIME detection returned no supported image type, `execute(...)` throws `ToolError(\"inspect_image only supports PNG, JPEG, GIF, and WEBP files detected by file content.\")`.\n9. The tool calls `instrumentedCompleteSimple(...)` with one user message containing two content parts in order:\n   - `{ type: \"image\", data: imageInput.data, mimeType: imageInput.mimeType }`\n   - `{ type: \"text\", text: params.question }`\n10. `systemPrompt` is a one-element array rendered from `packages/coding-agent/src/prompts/tools/inspect-image-system.md`; telemetry is tagged with oneshot kind `inspect_image`.\n11. If the model response stop reason is `error` or `aborted`, the tool maps that to `ToolError`.\n12. `extractTextContent(...)` from `packages/coding-agent/src/commit/utils.ts` concatenates only `text` content blocks from the assistant message, trims the result, and the tool fails if nothing remains.\n13. Success returns the text plus `details`; `inspectImageToolRenderer` formats the result for the TUI.\n\n## Modes / Variants\n- **Original image path**: `images.autoResize` disabled. The original file bytes are base64-encoded and sent with the detected MIME type.\n- **Auto-resized path**: `images.autoResize` enabled. `resizeImage(...)` may downscale and re-encode the image before upload.\n- **Unsupported image path**: file exists but header sniffing does not identify PNG/JPEG/GIF/WEBP. The tool returns a `ToolError` before any model call.\n- **Oversize image path**: file size exceeds 20 MiB before upload. The tool returns a `ToolError` before any model call.\n\n## Side Effects\n- Filesystem\n  - Resolves and reads the target image from disk.\n  - Stats the file once with `Bun.file(...).stat()` and reads it fully with `fs.readFile(...)`.\n- Network\n  - Sends the final base64 image payload plus question text to the selected model through `instrumentedCompleteSimple(...)` / the configured simple completion implementation.\n- Session state\n  - Reads session settings, active model preferences, cwd, and model registry.\n- Background work / cancellation\n  - Passes the caller `AbortSignal` into `instrumentedCompleteSimple(...)` and the configured simple completion implementation.\n  - Image preprocessing is local and not cancellation-aware in these helpers.\n\n## Limits & Caps\n- Supported detected input formats: `image/png`, `image/jpeg`, `image/gif`, `image/webp` (`SUPPORTED_IMAGE_MIME_TYPES` in `packages/utils/src/mime.ts`).\n- Metadata sniff cap: `DEFAULT_IMAGE_METADATA_HEADER_BYTES = 256 * 1024` bytes. Format detection only reads up to 256 KiB from the file header.\n- Availability is gated by `inspect_image.enabled`, default `false`, in `packages/coding-agent/src/config/settings-schema.ts` / `packages/coding-agent/src/tools/index.ts`.\n- Upload input cap: `MAX_IMAGE_INPUT_BYTES = 20 * 1024 * 1024` bytes (20 MiB) in `packages/coding-agent/src/utils/image-loading.ts`.\n- Auto-resize defaults in `packages/coding-agent/src/utils/image-resize.ts`:\n  - `maxWidth: 1568`\n  - `maxHeight: 1568`\n  - `maxBytes: 500 * 1024` bytes (500 KiB target)\n  - `jpegQuality: 80`\n- Resize fast path: if the original image is already within `1568x1568` and within `maxBytes / 4` (125 KiB by default), `resizeImage(...)` returns the original bytes unchanged.\n- Resize quality ladder: after the first encode pass, lossy retries use qualities `[70, 60, 50, 40]`.\n- Resize dimension ladder: if quality reduction still misses the byte target, retries scale dimensions by `[1.0, 0.75, 0.5, 0.35, 0.25]` and stop if either dimension would fall below `100` pixels.\n- First resize pass encodes PNG, JPEG, and WebP, then keeps the smallest encoded buffer. Fallback passes encode JPEG and WebP only, again keeping the smaller output. WebP is excluded from both ladders when `OMP_NO_WEBP=1`/`true` (or `excludeWebP` is passed).\n- Renderer caps:\n  - `INSPECT_QUESTION_PREVIEW_WIDTH = 100`\n  - `INSPECT_OUTPUT_COLLAPSED_LINES = 4`\n  - `INSPECT_OUTPUT_EXPANDED_LINES = 16`\n  - `INSPECT_OUTPUT_LINE_WIDTH = 120`\n\n## Errors\n- Settings gate:\n  - `Image submission is disabled by settings (images.blockImages=true). Disable it to use inspect_image.`\n- Model resolution / capability:\n  - `Model registry is unavailable for inspect_image.`\n  - `No models available for inspect_image.`\n  - `Unable to resolve a model for inspect_image.`\n  - `Resolved model <provider>/<id> does not support image input. Configure a vision-capable model for modelRoles.vision.`\n  - `No API key available for <provider>/<id>. Configure credentials for this provider or choose another vision-capable model.`\n- Input file:\n  - `Image file too large: <size> exceeds <limit> limit.` from `ImageInputTooLargeError`, remapped to `ToolError`.\n  - `inspect_image only supports PNG, JPEG, GIF, and WEBP files detected by file content.` when header sniffing fails.\n- Model call:\n  - `inspect_image request failed.` if the response stop reason is `error` without a provider message.\n  - Provider `errorMessage` is passed through when present.\n  - `inspect_image request aborted.` on aborted responses.\n  - `inspect_image model returned no text output.` when the assistant message contains no text blocks after filtering.\n\nFailures surface as thrown `ToolError`s from `execute(...)`; the normal success return shape is not used for error reporting.\n\n## Notes\n- The tool schema is not marked strict in `InspectImageTool`; callers should still treat only `path` and `question` as supported inputs because the implementation reads no other fields.\n- The model-facing prompt path on disk is `packages/coding-agent/src/prompts/tools/inspect-image.md`; the assignment's underscore form does not exist.\n- Format support is based on file content, not filename extension. Renaming a non-image file to `.png` does not make it valid.\n- `resolveReadPath(...)` tries macOS-specific path variants: shell-unescaped spaces, AM/PM narrow no-break-space filenames, NFD normalization, and curly-quote variants.\n- `loadImageInput(...)` also computes `textNote`, `dimensionNote`, and final `bytes`, but `inspect_image` does not include those in tool output.\n- Auto-resize can change the MIME type sent to the model. A JPEG or GIF input may be uploaded as PNG, JPEG, or WebP depending on which encoder output is smallest.\n- If `resizeImage(...)` throws or cannot decode the image, `loadImageInput(...)` silently keeps the original base64 payload instead of failing.\n",
 	"tools/irc.md": "# irc\n\n> Send and receive messages between agents over a process-global mailbox bus.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/irc.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/irc.md`\n- Key collaborators:\n  - `packages/coding-agent/src/irc/bus.ts` — process-global `IrcBus`: per-agent mailboxes, delivery, waiter matching.\n  - `packages/coding-agent/src/registry/agent-registry.ts` — process-global agent directory and status.\n  - `packages/coding-agent/src/registry/agent-lifecycle.ts` — revival of parked recipients on direct send.\n  - `packages/coding-agent/src/session/agent-session.ts` — `deliverIrcMessage(...)`: recipient-side injection and wake turns.\n  - `packages/coding-agent/src/prompts/system/irc-incoming.md` — incoming-message rendering for the recipient.\n  - `packages/coding-agent/src/prompts/system/irc-autoreply.md` — prompt for the ephemeral auto-reply side turn (busy recipient, async disabled).\n  - `packages/coding-agent/src/config/settings-schema.ts` — `irc.timeoutMs`.\n  - `packages/coding-agent/src/modes/controllers/event-controller.ts` — renders IRC events into chat UI.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `op` | `\"send\" \\| \"wait\" \\| \"inbox\" \\| \"list\"` | Yes | Operation. |\n| `to` | `string` | `send` | Recipient agent id, or `\"all\"` for broadcast. Whitespace trimmed; self-send rejected. |\n| `message` | `string` | `send` | Message body. Empty-after-trim is rejected. |\n| `replyTo` | `string` | No | `send`: message id being answered. |\n| `await` | `boolean` | No | `send`: after delivery, block until the next message from that peer arrives (round-trip sugar). Invalid with `to: \"all\"`. |\n| `from` | `string` | No | `wait`: only accept a message from this agent id. |\n| `timeoutMs` | `number` | No | `wait` / `send await:true`: timeout in milliseconds; `0` waits indefinitely. Defaults to `irc.timeoutMs`. |\n| `peek` | `boolean` | No | `inbox`: list messages without consuming them. |\n\n## Outputs\n- Single-shot `AgentToolResult`; no streaming updates.\n- `content` is one text block:\n  - `list`: `No other agents.` or `<n> peer(s):` bullets — `id [displayName · kind · status]` plus unread count, parent, and last-activity age; a footer notes that parked agents are revived automatically when messaged.\n  - `send`: per-recipient delivery receipts (`injected` / `woken` / `revived` / `failed — <error>`); with `await: true`, the reply body or a clean no-reply timeout note.\n  - `wait`: the consumed message as `[<msgId>] <from>: <body>` (with a reply-to tag), or `No message within <duration>.`\n  - `inbox`: `Inbox empty.` or `<n> message(s):` bullets.\n- `details: IrcDetails`: `{ op, from?, to?, receipts?, waited?, inbox?, peers? }`. `waited` is `null` when a wait timed out; `receipts` carry `{ to, outcome, error? }`.\n\n## Flow\n1. `IrcTool.createIf` constructs the tool only when `isIrcEnabled` passes and the session has both an `AgentRegistry` and `getAgentId`. There is no `irc.enabled` setting: availability is derived — true for every subagent (`taskDepth > 0`; a parent always exists) and for any session that can still spawn subagents through the task tool. Only a top-level session with task spawning unavailable has no peers, hence no irc.\n2. `execute` resolves the registry and sender id; missing either returns a text error result instead of throwing.\n3. `op: \"list\"`: `registry.list()` minus self and minus `aborted` agents — `parked` peers ARE listed. Each row includes the unread count from `IrcBus.unreadCount(...)` and last activity.\n4. `op: \"send\"` validates `to`/`message`, rejects self-sends, and rejects `await` with `to: \"all\"`.\n5. Target resolution: broadcasts fan out to `registry.listVisibleTo(senderId)` (live peers only — `running`/`idle`; reviving every parked agent on a broadcast would be a stampede). Direct sends go through the bus unfiltered, so a parked recipient is revived.\n6. `IrcBus.send(...)` is fire-and-forget — it never blocks on the recipient generating anything. Delivery by recipient status:\n   - `running` → message enqueued and injected as a non-interrupting aside at the recipient's next step boundary (`AgentSession.deliverIrcMessage`, rendered from `irc-incoming.md`, persisted as an `irc:incoming` custom message) — receipt `injected`. If the sender awaits a reply (`expectsReply` from `await: true`) and the recipient has `async.enabled` off, the recipient also generates an ephemeral no-tools auto-reply (`runEphemeralTurn`, the `/btw` pipeline) and sends it back over the bus with `replyTo` set, recording an `irc:autoreply` aside in its own history — a recipient blocked in a synchronous task spawn can never reach a step boundary before the sender's timeout otherwise;\n   - `idle` (live session) → enqueued and a real turn is started — the message wakes the agent — receipt `woken`;\n   - `parked` → `AgentLifecycleManager.global().ensureLive(to)` revives the session first, then the wake path — receipt `revived`;\n   - resolution/revival failure → receipt `failed` with the error; other recipients still complete.\n7. `send` with `await: true` then calls `IrcBus.wait(senderId, { from: to }, timeoutMs, signal)` and appends the reply (or a no-reply note suggesting `inbox`/`wait`) to the result. Awaited sends pass `{ expectsReply: true }` to `IrcBus.send` so a busy recipient can auto-reply (see step 6).\n8. `op: \"wait\"` blocks until a message for the caller (optionally filtered by `from`) arrives, consumes it, and returns it. Timeout returns a clean \"no message\" result, not an error.\n9. `op: \"inbox\"` drains pending messages (or peeks with `peek: true`) without blocking.\n10. Timeouts resolve as `params.timeoutMs ?? irc.timeoutMs`, normalized: `0` disables the timeout, negative/non-finite values fall back to the default `120_000`, positive values are truncated and clamped to ≥ 1 ms.\n\n## Modes / Variants\n- `list`: enumerate peers with status (`running`/`idle`/`parked`), unread counts, and last activity.\n- `send` direct: one exact peer id; wakes idle peers, revives parked ones.\n- `send` broadcast: `to: \"all\"` to every live peer; parked peers are skipped.\n- `send` + `await: true`: round-trip convenience — send, then wait for the next message from that peer. Marks the send `expectsReply`, enabling the busy-recipient auto-reply path when async execution is disabled.\n- `wait`: block for an incoming message, optionally filtered by sender.\n- `inbox`: non-blocking drain or peek.\n\n## Side Effects\n- Session state\n  - Reads the process-global `AgentRegistry`; direct sends to parked agents revive their sessions through the lifecycle manager.\n  - Persists `irc:incoming` custom messages into recipient history; replies are ordinary turns in the recipient's own session.\n  - Waking an idle/parked recipient starts a real agent turn (model requests, tool use) in that recipient.\n- User-visible prompts / interactive UI\n  - IRC events render as transcript cards in the TUI; the Agent Hub shows per-agent unread counts.\n- Background work / cancellation\n  - `send` itself never blocks on reply generation; only `wait` (and `await: true`) blocks, bounded by the resolved timeout and the caller's `AbortSignal`.\n- Network\n  - No IRC server connection. Woken recipients make their own model-provider calls as part of their turn.\n- Filesystem\n  - No direct filesystem writes in the tool itself; recipient turns persist to their session JSONL as usual.\n\n## Limits & Caps\n- Availability gates: `isIrcEnabled` (running as a subagent, or task spawning available — there is no `irc.enabled` setting), an `AgentRegistry`, and a caller agent id.\n- Mailboxes are bounded at 100 messages per agent (`MAILBOX_CAP` in `packages/coding-agent/src/irc/bus.ts`); oldest messages are dropped beyond the cap.\n- `irc.timeoutMs` defaults to `120_000` and is the default `wait` / `send await:true` timeout; `0` disables the timeout, non-finite or negative values fall back to the default, positive values are truncated and clamped to at least `1` ms.\n- Broadcast scope: live peers only (`running`/`idle`) via `listVisibleTo`; direct sends address any non-aborted agent, including parked ones.\n\n## Errors\n- The tool returns text errors (with `isError: true`), not thrown exceptions, for:\n  - missing registry: `IRC is unavailable in this session.`\n  - missing sender id: `IRC is unavailable: caller has no agent id.`\n  - missing `to` / `message` on `send`\n  - self-send: `Cannot send an IRC message to yourself.`\n  - `await` with `to: \"all\"`\n  - unknown op\n- Per-recipient delivery failures surface as `failed` receipts with the error message; `send` is marked `isError` only when no recipient received the message.\n- `wait` timeout is a normal result (`waited: null`), not an error.\n\n## Notes\n- This is IRC-like naming only: no servers, sockets, channels, or join/part state. Addressing is by exact registry agent id.\n- Replies are real turns by the recipient, with one exception: an awaited send to a mid-turn recipient with `async.enabled` off triggers an ephemeral no-tools auto-reply (the old `respondAsBackground` path), because a recipient blocked in a synchronous task spawn whose batch includes the sender can never run a real turn before the sender's timeout. A recipient may otherwise keep working before answering; check `inbox` or `wait` again rather than re-sending.\n- Wake-on-message is the only resume primitive: messaging a parked agent revives it (same `ensureLive` path as the Agent Hub). The task tool has no `resume` parameter.\n- Message ids are Snowflakes; pass them as `replyTo` to thread an answer to a specific message.\n- Persistence is per recipient history: the sender gets receipts in the tool result; the recipient sees the injected `irc:incoming` message in its own transcript (visible via `history://<id>`).\n",
-	"tools/job.md": "# job\n\n> Wait for or cancel background jobs managed by the session async runtime.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/job.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/job.md`\n- Key collaborators:\n  - `packages/coding-agent/src/async/job-manager.ts` — job registry, cancellation, delivery suppression.\n  - `packages/coding-agent/src/tools/bash.ts` — explicit async bash and auto-backgrounded bash jobs.\n  - `packages/coding-agent/src/task/index.ts` — async task-job scheduling.\n  - `packages/coding-agent/src/sdk.ts` — automatic follow-up delivery for unsuppressed completions.\n  - `packages/coding-agent/src/config/settings-schema.ts` — `async.pollWaitDuration` options.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `poll` | `string[]` | No | Job ids to watch. Cannot be combined with `list`. If omitted (and `cancel` is also omitted), the tool watches all running jobs owned by the calling agent. If provided, missing ids — and ids owned by other agents — are silently filtered out before waiting. |\n| `cancel` | `string[]` | No | Job ids to cancel before any polling. Missing ids (and other agents' jobs) are reported as `not_found`; non-running ids as `already_completed`. |\n| `list` | `boolean` | No | Return an immediate snapshot of every job spawned by the calling agent (running + completed within retention) without waiting. Read-only — cannot be combined with `poll` or `cancel`. |\n\n## Outputs\nThe tool returns one text block plus `details`.\n\n- `content[0].text`: markdown-like plain text sections assembled by `#buildResult(...)`:\n  - `## Cancelled (N)` for cancel outcomes.\n  - `## Completed (N)` for non-running jobs, including stored `resultText` and `errorText`.\n  - `## Still Running (N)` for jobs still in `running`.\n- `details.jobs`: array of snapshots:\n  - `id: string`\n  - `type: \"bash\" | \"task\"`\n  - `status: \"running\" | \"completed\" | \"failed\" | \"cancelled\"`\n  - `label: string`\n  - `durationMs: number`\n  - optional `resultText`, `errorText`\n- `details.cancelled` appears only when `cancel` was passed; each item is `{ id, status }` where status is `\"cancelled\" | \"not_found\" | \"already_completed\"`.\n\nStreaming behavior:\n- During a polling wait, `execute(...)` emits `onUpdate(...)` every 500 ms with an empty text block and fresh `details.jobs` snapshots.\n- Final return is single-shot after a completion, timeout, abort, or immediate fast path.\n\nRead-only snapshot path:\n- Calling `job` with `list: true` returns a markdown summary of every job spawned by the calling agent (running + completed within retention) without waiting.\n\n## Flow\n1. `JobTool` is registered unconditionally in `packages/coding-agent/src/tools/index.ts`; there is no `async.enabled` gate (the manager may still carry bash or task jobs from before a setting change).\n2. `execute(...)` fetches `session.asyncJobManager`. If absent, it returns `Async execution is disabled; no background jobs are available.`\n3. `cancel` ids are processed first:\n   - `manager.getJob(id)` missing → `not_found`.\n   - existing job with `status !== \"running\"` → `already_completed`.\n   - running job → `manager.cancel(id)`, which sets `job.status = \"cancelled\"`, aborts the controller, and schedules eviction.\n4. Polling mode is chosen with `const shouldPoll = requestedPollIds !== undefined || cancelIds.length === 0`:\n   - only `cancel` present → return immediately, no wait.\n   - explicit `poll`, or no args at all → proceed to watch jobs.\n5. Watch set resolution:\n   - explicit `poll` → resolve ids via `#visibleJobs(...)`, dropping missing ids and jobs owned by other agents.\n   - no `poll` and no `cancel` → `manager.getRunningJobs(ownerFilter)` (jobs owned by the calling agent).\n6. Empty watch set returns immediately:\n   - if cancellations happened, return snapshots for the cancelled ids that still exist.\n   - else return either `No matching jobs found for IDs: ...` or `No running background jobs to wait for.`\n7. If every watched job is already non-running, `#buildResult(...)` returns immediately without waiting.\n8. Otherwise the tool waits on `Promise.race(...)` across:\n   - every watched running job's `job.promise`,\n   - a timeout promise for `async.pollWaitDuration`,\n   - the tool-call abort signal when present.\n9. Before waiting, it calls `manager.watchJobs(watchedJobIds)`. This suppresses automatic completion delivery for those ids while they are being watched.\n10. If `onUpdate` exists, a 500 ms interval sends progress snapshots from `#snapshotJobs(...)`; one snapshot is emitted immediately before entering the race.\n11. In `finally`, the tool always calls `manager.unwatchJobs(...)`, clears the timeout, and stops the progress interval.\n12. `#buildResult(...)` deduplicates jobs, snapshots current manager state, then calls `manager.acknowledgeDeliveries(...)` for every non-running job in the result. That suppresses later automatic follow-up delivery for the same completions and removes queued deliveries for those ids.\n13. The final text groups jobs by non-running vs still-running state. A timeout is not an error path; it simply returns the current snapshot.\n\n## Modes / Variants\n- Poll all running jobs: call with neither `poll` nor `cancel`.\n- Poll explicit ids: call with `poll` only.\n- Cancel only: call with `cancel` only; cancellations happen and the tool returns immediately.\n- Cancel then poll: call with both. Cancellations are applied first, then the tool watches the remaining resolved `poll` ids.\n- Read-only inspection: call with `list: true` for the same snapshot data without waiting on completion.\n\nSpawn paths that produce jobs:\n- `packages/coding-agent/src/tools/bash.ts`\n  - `async: true` always registers a `type: \"bash\"` job with `AsyncJobManager.register(...)` and returns a start message.\n  - auto-background mode (`bash.autoBackground.enabled`) starts the same managed job path for non-PTY commands, waits up to `min(bash.autoBackground.thresholdMs, timeoutMs - 1000)`, and if the command is still running returns a background-job start result instead of inline command output.\n- `packages/coding-agent/src/task/index.ts`\n  - every `task` call registers one `type: \"task\"` job, unless the session has no job manager or the agent definition declares `blocking: true` (sync fallback).\n\nLifecycle and exact state names:\n- Conceptual scheduling path: `pending` (only task-progress bookkeeping before work starts) → `running` → `completed` / `failed`; cancellation changes a running async job to `cancelled`.\n- Exact `AsyncJob.status` values in `packages/coding-agent/src/async/job-manager.ts`: `\"running\" | \"completed\" | \"failed\" | \"cancelled\"`.\n- Exact per-task progress values in `packages/coding-agent/src/task/types.ts`: `\"pending\" | \"running\" | \"completed\" | \"failed\" | \"aborted\"`.\n\n## Side Effects\n- Filesystem\n  - None in `job.ts` itself.\n  - Jobs being observed may already have written artifacts/results through their own tool runtimes.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - Reads and mutates `session.asyncJobManager` state.\n  - `watchJobs(...)` / `unwatchJobs(...)` toggle delivery suppression for the watched ids.\n  - `acknowledgeDeliveries(...)` marks completed ids as suppressed and removes queued deliveries for them.\n  - `cancel(...)` aborts running jobs through each job's `AbortController`.\n- User-visible prompts / interactive UI\n  - Polling emits periodic `onUpdate` snapshots every 500 ms.\n  - Automatic job completion follow-ups are generated by `packages/coding-agent/src/sdk.ts` only for unsuppressed deliveries.\n- Background work / cancellation\n  - Waiting uses a timeout plus optional tool-call abort signal.\n  - Cancelling a job does not synchronously await teardown; it flips state, aborts, and returns control to the manager/job promise.\n\n## Limits & Caps\n- Poll wait duration comes from `async.pollWaitDuration` in `packages/coding-agent/src/config/settings-schema.ts`:\n  - allowed values: `5s`, `10s`, `30s`, `1m`, `5m`\n  - default: `30s`\n- Progress update cadence while polling: `PROGRESS_INTERVAL_MS = 500` in `packages/coding-agent/src/tools/job.ts`.\n- Async job retention default: `DEFAULT_RETENTION_MS = 5 * 60 * 1000` in `packages/coding-agent/src/async/job-manager.ts`.\n- Manager fallback max-running limit: `DEFAULT_MAX_RUNNING_JOBS = 15` in `packages/coding-agent/src/async/job-manager.ts`.\n- Session wiring clamps `async.maxJobs` to `1..100` before constructing the manager in `packages/coding-agent/src/sdk.ts`; settings default is `100` in `packages/coding-agent/src/config/settings-schema.ts`.\n- Async completion delivery retry backoff in `packages/coding-agent/src/async/job-manager.ts`:\n  - base `500` ms\n  - max `30_000` ms\n  - jitter `< 200` ms\n  - exponent capped at 8 doublings\n\n## Errors\n- Tool-disabled path is returned as normal text, not thrown: `Async execution is disabled; no background jobs are available.`\n- Polling a nonexistent id is not an exception:\n  - with `poll` only, missing ids are dropped; if none remain the tool returns `No matching jobs found for IDs: ...`.\n  - with `cancel`, each missing id is reported as `not_found` in `details.cancelled` and text.\n- Cancelling a non-running job is not an exception; it reports `already_completed` even if the actual status is `completed`, `failed`, or `cancelled`.\n- Tool-call abort during polling stops waiting and returns a final snapshot through `#buildResult(...)`; it does not cancel watched jobs.\n- Failures inside the underlying async work are stored on the job (`status: \"failed\"`, `errorText`) and reported in normal tool output, not rethrown by `job`.\n- Calling `list: true` against an empty manager returns a normal empty-list result rather than throwing; missing ids passed to `poll` are silently filtered.\n- Combining `list` with `poll` or `cancel` throws a `ToolError`: `` `list` cannot be combined with `poll` or `cancel`. ``\n\n## Notes\n- `job` waits for the first watched running job to settle, not for all watched jobs. If others remain `running`, they are reported under `## Still Running`; the caller must invoke `job` again to continue waiting.\n- Delivery suppression is the key difference between snapshot and automatic delivery:\n  - snapshots (`job` calls with `poll` or `list: true`) read current manager state;\n  - follow-up delivery comes from `AsyncJobManager.#enqueueDelivery(...)` and `sdk.ts` `onJobComplete`;\n  - watched or acknowledged ids are suppressed via `isDeliverySuppressed(...)`.\n- `manager.cancel(id)` sets `status = \"cancelled\"` before the underlying promise settles. The job function may later populate `resultText` or `errorText`; `job-manager.ts` preserves that text but does not transition the status away from `cancelled`.\n- Retention eviction removes the job record, suppression flags, and watch flag together. After eviction, both `job` calls and `list: true` snapshots behave as if the id never existed.\n",
+	"tools/job.md": "# job\n\n> Wait for or cancel background jobs managed by the session async runtime.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/job.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/job.md`\n- Key collaborators:\n  - `packages/coding-agent/src/async/job-manager.ts` — job registry, cancellation, delivery suppression.\n  - `packages/coding-agent/src/tools/bash.ts` — explicit async bash and auto-backgrounded bash jobs.\n  - `packages/coding-agent/src/task/index.ts` — async task-job scheduling.\n  - `packages/coding-agent/src/sdk.ts` — automatic follow-up delivery for unsuppressed completions.\n  - `packages/coding-agent/src/config/settings-schema.ts` — `async.pollWaitDuration` options.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `poll` | `string[]` | No | Job ids to watch. Cannot be combined with `list`. If omitted (and `cancel` is also omitted), the tool watches all running jobs owned by the calling agent. If provided, missing ids — and ids owned by other agents — are silently filtered out before waiting. |\n| `cancel` | `string[]` | No | Job ids to cancel before any polling. Missing ids (and other agents' jobs) are reported as `not_found`; non-running ids as `already_completed`. |\n| `list` | `boolean` | No | Return an immediate snapshot of every job spawned by the calling agent (running + completed within retention) without waiting. Read-only — cannot be combined with `poll` or `cancel`. |\n\n## Outputs\nThe tool returns one text block plus `details`.\n\n- `content[0].text`: markdown-like plain text sections assembled by `#buildResult(...)`:\n  - `## Cancelled (N)` for cancel outcomes.\n  - `## Completed (N)` for non-running jobs, including stored `resultText` and `errorText`.\n  - `## Still Running (N)` for jobs still in `running`.\n- `details.jobs`: array of snapshots:\n  - `id: string`\n  - `type: \"bash\" | \"task\"`\n  - `status: \"running\" | \"completed\" | \"failed\" | \"cancelled\"`\n  - `label: string`\n  - `durationMs: number`\n  - optional `resultText`, `errorText`\n- `details.cancelled` appears only when `cancel` was passed; each item is `{ id, status }` where status is `\"cancelled\" | \"not_found\" | \"already_completed\"`.\n\nStreaming behavior:\n- During a polling wait, `execute(...)` emits `onUpdate(...)` every 500 ms with an empty text block and fresh `details.jobs` snapshots.\n- Final return is single-shot after a completion, timeout, abort, or immediate fast path.\n\nRead-only snapshot path:\n- Calling `job` with `list: true` returns a markdown summary of every job spawned by the calling agent (running + completed within retention) without waiting.\n\n## Flow\n1. `JobTool` is registered unconditionally in `packages/coding-agent/src/tools/index.ts`; there is no `async.enabled` gate (the manager may still carry bash or task jobs from before a setting change).\n2. `execute(...)` fetches `session.asyncJobManager`. If absent, it returns `Async execution is disabled; no background jobs are available.`\n3. `cancel` ids are processed first:\n   - `manager.getJob(id)` missing → `not_found`.\n   - existing job with `status !== \"running\"` → `already_completed`.\n   - running job → `manager.cancel(id)`, which sets `job.status = \"cancelled\"`, aborts the controller, and schedules eviction.\n4. Polling mode is chosen with `const shouldPoll = requestedPollIds !== undefined || cancelIds.length === 0`:\n   - only `cancel` present → return immediately, no wait.\n   - explicit `poll`, or no args at all → proceed to watch jobs.\n5. Watch set resolution:\n   - explicit `poll` → resolve ids via `#visibleJobs(...)`, dropping missing ids and jobs owned by other agents.\n   - no `poll` and no `cancel` → `manager.getRunningJobs(ownerFilter)` (jobs owned by the calling agent).\n6. Empty watch set returns immediately:\n   - if cancellations happened, return snapshots for the cancelled ids that still exist.\n   - else return either `No matching jobs found for IDs: ...` or `No running background jobs to wait for.`\n7. If every watched job is already non-running, `#buildResult(...)` returns immediately without waiting.\n8. Otherwise the tool waits on `Promise.race(...)` across:\n   - every watched running job's `job.promise`,\n   - a timeout promise for the poll wait window — `manager.nextPollWaitMs(ownerId)` when `async.pollWaitDuration` is `smart`, otherwise the fixed duration,\n   - the tool-call abort signal when present.\n9. Before waiting, it calls `manager.watchJobs(watchedJobIds)`. This suppresses automatic completion delivery for those ids while they are being watched.\n10. If `onUpdate` exists, a 500 ms interval sends progress snapshots from `#snapshotJobs(...)`; one snapshot is emitted immediately before entering the race.\n11. In `finally`, the tool always calls `manager.unwatchJobs(...)`, clears the timeout, and stops the progress interval.\n12. `#buildResult(...)` deduplicates jobs, snapshots current manager state, then calls `manager.acknowledgeDeliveries(...)` for every non-running job in the result. That suppresses later automatic follow-up delivery for the same completions and removes queued deliveries for those ids.\n13. The final text groups jobs by non-running vs still-running state. A timeout is not an error path; it simply returns the current snapshot.\n\n## Modes / Variants\n- Poll all running jobs: call with neither `poll` nor `cancel`.\n- Poll explicit ids: call with `poll` only.\n- Cancel only: call with `cancel` only; cancellations happen and the tool returns immediately.\n- Cancel then poll: call with both. Cancellations are applied first, then the tool watches the remaining resolved `poll` ids.\n- Read-only inspection: call with `list: true` for the same snapshot data without waiting on completion.\n\nSpawn paths that produce jobs:\n- `packages/coding-agent/src/tools/bash.ts`\n  - `async: true` always registers a `type: \"bash\"` job with `AsyncJobManager.register(...)` and returns a start message.\n  - auto-background mode (`bash.autoBackground.enabled`) starts the same managed job path for non-PTY commands, waits up to `min(bash.autoBackground.thresholdMs, timeoutMs - 1000)`, and if the command is still running returns a background-job start result instead of inline command output.\n- `packages/coding-agent/src/task/index.ts`\n  - every `task` call registers one `type: \"task\"` job, unless the session has no job manager or the agent definition declares `blocking: true` (sync fallback).\n\nLifecycle and exact state names:\n- Conceptual scheduling path: `pending` (only task-progress bookkeeping before work starts) → `running` → `completed` / `failed`; cancellation changes a running async job to `cancelled`.\n- Exact `AsyncJob.status` values in `packages/coding-agent/src/async/job-manager.ts`: `\"running\" | \"completed\" | \"failed\" | \"cancelled\"`.\n- Exact per-task progress values in `packages/coding-agent/src/task/types.ts`: `\"pending\" | \"running\" | \"completed\" | \"failed\" | \"aborted\"`.\n\n## Side Effects\n- Filesystem\n  - None in `job.ts` itself.\n  - Jobs being observed may already have written artifacts/results through their own tool runtimes.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - Reads and mutates `session.asyncJobManager` state.\n  - `watchJobs(...)` / `unwatchJobs(...)` toggle delivery suppression for the watched ids.\n  - `acknowledgeDeliveries(...)` marks completed ids as suppressed and removes queued deliveries for them.\n  - `cancel(...)` aborts running jobs through each job's `AbortController`.\n- User-visible prompts / interactive UI\n  - Polling emits periodic `onUpdate` snapshots every 500 ms.\n  - Automatic job completion follow-ups are generated by `packages/coding-agent/src/sdk.ts` only for unsuppressed deliveries.\n- Background work / cancellation\n  - Waiting uses a timeout plus optional tool-call abort signal.\n  - Cancelling a job does not synchronously await teardown; it flips state, aborts, and returns control to the manager/job promise.\n\n## Limits & Caps\n- Poll wait duration comes from `async.pollWaitDuration` (\"Max Poll Time\") in `packages/coding-agent/src/config/settings-schema.ts`:\n  - allowed values: `5s`, `10s`, `30s`, `1m`, `5m`, `smart`\n  - default: `smart`\n  - fixed values block for exactly that long; `smart` uses the adaptive ladder `POLL_WAIT_LADDER_MS = [5s, 10s, 30s, 1m, 5m]` in `packages/coding-agent/src/async/job-manager.ts`, climbing one rung per back-to-back poll and resetting to the 5s floor after `POLL_ESCALATION_RESET_MS = 60_000` ms without polling. Per-owner state is driven by `nextPollWaitMs(...)` / `recordPollWaitEnd(...)`.\n- Progress update cadence while polling: `PROGRESS_INTERVAL_MS = 500` in `packages/coding-agent/src/tools/job.ts`.\n- Async job retention default: `DEFAULT_RETENTION_MS = 5 * 60 * 1000` in `packages/coding-agent/src/async/job-manager.ts`.\n- Manager fallback max-running limit: `DEFAULT_MAX_RUNNING_JOBS = 15` in `packages/coding-agent/src/async/job-manager.ts`.\n- Session wiring clamps `async.maxJobs` to `1..100` before constructing the manager in `packages/coding-agent/src/sdk.ts`; settings default is `100` in `packages/coding-agent/src/config/settings-schema.ts`.\n- Async completion delivery retry backoff in `packages/coding-agent/src/async/job-manager.ts`:\n  - base `500` ms\n  - max `30_000` ms\n  - jitter `< 200` ms\n  - exponent capped at 8 doublings\n\n## Errors\n- Tool-disabled path is returned as normal text, not thrown: `Async execution is disabled; no background jobs are available.`\n- Polling a nonexistent id is not an exception:\n  - with `poll` only, missing ids are dropped; if none remain the tool returns `No matching jobs found for IDs: ...`.\n  - with `cancel`, each missing id is reported as `not_found` in `details.cancelled` and text.\n- Cancelling a non-running job is not an exception; it reports `already_completed` even if the actual status is `completed`, `failed`, or `cancelled`.\n- Tool-call abort during polling stops waiting and returns a final snapshot through `#buildResult(...)`; it does not cancel watched jobs.\n- Failures inside the underlying async work are stored on the job (`status: \"failed\"`, `errorText`) and reported in normal tool output, not rethrown by `job`.\n- Calling `list: true` against an empty manager returns a normal empty-list result rather than throwing; missing ids passed to `poll` are silently filtered.\n- Combining `list` with `poll` or `cancel` throws a `ToolError`: `` `list` cannot be combined with `poll` or `cancel`. ``\n\n## Notes\n- `job` waits for the first watched running job to settle, not for all watched jobs. If others remain `running`, they are reported under `## Still Running`; the caller must invoke `job` again to continue waiting.\n- Delivery suppression is the key difference between snapshot and automatic delivery:\n  - snapshots (`job` calls with `poll` or `list: true`) read current manager state;\n  - follow-up delivery comes from `AsyncJobManager.#enqueueDelivery(...)` and `sdk.ts` `onJobComplete`;\n  - watched or acknowledged ids are suppressed via `isDeliverySuppressed(...)`.\n- `manager.cancel(id)` sets `status = \"cancelled\"` before the underlying promise settles. The job function may later populate `resultText` or `errorText`; `job-manager.ts` preserves that text but does not transition the status away from `cancelled`.\n- Retention eviction removes the job record, suppression flags, and watch flag together. After eviction, both `job` calls and `list: true` snapshots behave as if the id never existed.\n",
 	"tools/lsp.md": "# lsp\n\n> Query language servers for diagnostics, navigation, symbols, renames, code actions, capabilities, and raw requests.\n\n## Source\n- Entry: `packages/coding-agent/src/lsp/index.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/lsp.md`\n- Key collaborators:\n  - `packages/coding-agent/src/lsp/client.ts` — client process lifecycle and JSON-RPC\n  - `packages/coding-agent/src/lsp/config.ts` — config loading, auto-detect, server selection\n  - `packages/coding-agent/src/lsp/lspmux.ts` — optional `lspmux` command wrapping\n  - `packages/coding-agent/src/lsp/edits.ts` — apply `WorkspaceEdit` and text edits\n  - `packages/coding-agent/src/lsp/utils.ts` — URI conversion, symbol resolution, formatting, glob expansion\n  - `packages/coding-agent/src/lsp/types.ts` — tool schema and protocol types\n  - `packages/coding-agent/src/lsp/clients/index.ts` — custom linter client cache/factory\n  - `packages/coding-agent/src/lsp/clients/lsp-linter-client.ts` — LSP-backed linter adapter\n  - `packages/coding-agent/src/lsp/clients/biome-client.ts` — Biome CLI diagnostics/formatting adapter\n  - `packages/coding-agent/src/lsp/clients/swiftlint-client.ts` — SwiftLint CLI diagnostics adapter\n  - `packages/coding-agent/src/tools/index.ts` — tool registration and `lsp.enabled` gating\n  - `packages/coding-agent/src/tools/tool-timeouts.ts` — timeout defaults and clamping\n  - `packages/coding-agent/src/lsp/defaults.json` — built-in server definitions for auto-detect\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `action` | string enum | Yes | One of `diagnostics`, `definition`, `references`, `hover`, `symbols`, `rename`, `rename_file`, `code_actions`, `type_definition`, `implementation`, `status`, `reload`, `capabilities`, `request`. |\n| `file` | string | No | File path; for `diagnostics` also a glob; for workspace forms use `\"*\"`; for `rename_file` this is the source path. |\n| `line` | number | No | 1-indexed line number for position-based actions. Defaults to `1` on the single-file action path. |\n| `symbol` | string | No | Substring used to resolve the column on `line`. Supports `name#N` occurrence selectors; `N` is 1-indexed and defaults to `1`. Required when `line` is given for `definition`/`references`/`rename` against project-aware servers. |\n| `query` | string | No | Workspace symbol query, code-action selector/filter, or LSP method name for `action=request`. |\n| `new_name` | string | No | Required for `rename` and `rename_file`. |\n| `apply` | boolean | No | For `rename`/`rename_file`, apply unless explicitly `false`. For `code_actions`, list unless explicitly `true`. |\n| `timeout` | number | No | Seconds, clamped by `clampTimeout(\"lsp\", ...)` to `5..60`, default `20`. |\n| `payload` | string | No | JSON string for `action=request`; overrides auto-built params. |\n\n## Outputs\n- Single-shot `AgentToolResult`.\n- `content` is always one text block: `[{ type: \"text\", text: string }]`.\n- `details` is `LspToolDetails`: `action`, `success`, optional `serverName`, optional original `request`.\n- No streaming updates.\n- No artifact URIs or background jobs.\n- Many validation failures are returned as ordinary text results with `details.success: false`; aborts throw `ToolAbortError` instead.\n\n## Flow\n1. `packages/coding-agent/src/tools/index.ts` registers `lsp: LspTool.createIf`; session creation also gates it behind `session.enableLsp !== false` and `settings.get(\"lsp.enabled\")`.\n2. `LspTool.execute()` in `packages/coding-agent/src/lsp/index.ts` clamps `timeout` with `clampTimeout(\"lsp\", ...)`, builds an `AbortSignal.timeout(...)`, and combines it with the caller signal.\n3. `getConfig()` loads and caches `LspConfig` per cwd, applies idle-timeout config via `setIdleTimeout()`, and reuses the cached config on later calls.\n4. Config loading in `packages/coding-agent/src/lsp/config.ts` merges `defaults.json` with JSON/YAML overrides from project, project config dirs, user config dirs, plugin roots, and home; if there are no overrides it auto-detects servers from root markers plus executable discovery.\n5. Server routing uses `getServersForFile()` / `getServerForFile()` from `config.ts`: extension or basename match, then sort primary servers before linters. `index.ts` further filters custom linter clients out of navigation/refactor paths with `getLspServersForFile()` / `getLspServerForFile()`.\n6. `getOrCreateClient()` in `client.ts` creates one process per `command:cwd`, optionally wraps supported commands with `lspmux`, spawns the server, starts the background message reader, sends `initialize`, stores server capabilities, then sends `initialized`.\n7. The message reader in `client.ts` parses LSP frames, resolves pending requests, caches `publishDiagnostics`, tracks `$/progress` tokens for project-load completion, answers `workspace/configuration`, and applies `workspace/applyEdit` requests through `applyWorkspaceEdit()`.\n8. File-scoped actions call `ensureFileOpen()` before requests. Column resolution uses `resolveSymbolColumn()` from `utils.ts`: read the target file, pick first non-whitespace when `symbol` is omitted, otherwise find the exact or case-insensitive match on the target line and honor `#N` occurrence selectors.\n9. Actions dispatch in `LspTool.execute()` through dedicated branches: workspace-only branches (`status`, some `diagnostics`, workspace `symbols`, workspace `reload`, `capabilities`, `request`) run before the single-file switch; all other single-file actions share one client lookup and `switch(action)`.\n10. Requests go through `sendRequest()` in `client.ts`, which allocates an incrementing JSON-RPC id, installs abort and timeout handling, sends `$/cancelRequest` on abort, and rejects on timeout or process exit.\n11. Actions that return edits either preview with `formatWorkspaceEdit()` or apply with `applyWorkspaceEdit()` from `edits.ts`; `rename_file` also performs the filesystem rename and then sends `workspace/didRenameFiles`.\n12. Non-abort failures inside the single-file action block are converted to `LSP error: ...`; many precondition failures return explicit text without throwing.\n\n## Modes / Variants\n### Routing and workspace scope\n- `file: \"*\"` is only special for `diagnostics`, `symbols`, and `reload`.\n- `status` ignores `file`.\n- `capabilities` with omitted `file` or `\"*\"` inspects all non-custom LSP servers; with a concrete file it scopes to matching non-custom servers.\n- `request` with omitted `file` or `\"*\"` chooses the first available non-custom LSP server; with a concrete file it chooses that file's primary non-linter server.\n- `rename_file` sends `workspace/willRenameFiles` and `workspace/didRenameFiles` to every non-custom LSP server from `getLspServers(config)` whose `fileTypes` match the source, destination, or any enumerated rename pair — not just one file-scoped server.\n- Diagnostics are the only tool action that queries both normal LSP servers and custom linter clients (`BiomeClient`, `SwiftLintClient`, or `LspLinterClient`).\n\n### `diagnostics`\n**Inputs**\n- Required: `file`, unless using workspace mode with `file: \"*\"`.\n- Optional: `timeout`.\n\n**Execution**\n- `file: \"*\"`: `runWorkspaceDiagnostics()` detects project type from root markers and runs one subprocess command: Rust `cargo check --message-format=short`, TypeScript `npx tsc --noEmit`, Python `pyright`.\n- Concrete file or glob: `resolveDiagnosticTargets()` treats non-globs as one target, otherwise expands a `Bun.Glob` up to `MAX_GLOB_DIAGNOSTIC_TARGETS`.\n- Per file, every matching server runs: custom clients call `lint(file)`; real LSP servers optionally wait for project load, capture `diagnosticsVersion`, `refreshFile()`, then `waitForDiagnostics()` for fresh `publishDiagnostics` (settles on the latest publish; exact-version match accepted immediately).\n- Results are deduplicated by range+message and severity-sorted.\n\n**Output text**\n- Single target with no issues: `OK`.\n- Single target with issues: `<summary>:\\n<grouped diagnostics>`.\n- Batch/glob target: one section per file, plus an initial truncation warning when the glob exceeds the file cap.\n- Workspace mode: `Workspace diagnostics (<detected description>):\\n<command output>`.\n\n### `definition`\n**Inputs**\n- Required: `file`.\n- Optional: `line`, `symbol`, `timeout`.\n\n**Execution**\n- Sends `textDocument/definition` with `{ textDocument, position }`.\n- Accepts `Location`, `Location[]`, `LocationLink`, or `LocationLink[]`; `normalizeLocationResult()` converts `LocationLink` to `targetSelectionRange ?? targetRange`.\n- Requires `symbol` when `line` is given on project-aware servers (the first-non-whitespace-column fallback is disabled for this action).\n- Waits for project load before the request.\n\n**Output text**\n- `No definition found` or `Found N definition(s):` followed by `file:line:col` and one context line above/below each location.\n\n### `type_definition`\nSame as `definition`, but sends `textDocument/typeDefinition` and reports `type definition(s)`.\n\n### `implementation`\nSame as `definition`, but sends `textDocument/implementation` and reports `implementation(s)`.\n\n### `references`\n**Inputs**\n- Required: `file`.\n- Optional: `line`, `symbol`, `timeout`.\n\n**Execution**\n- Sends `textDocument/references` with `includeDeclaration: true`.\n- Requires `symbol` when `line` is given on project-aware servers (the first-non-whitespace-column fallback is disabled for this action).\n- For project-aware servers, retries up to `REFERENCES_RETRY_COUNT` times when the only hit is the queried declaration; between retries it waits for project load and sleeps `REFERENCES_RETRY_DELAY_MS`.\n- First `REFERENCE_CONTEXT_LIMIT` references include surrounding context; the rest are location-only.\n\n**Output text**\n- `No references found` or `Found N reference(s):` with contextual entries first, then `... M additional reference(s) shown without context` when truncated.\n\n### `hover`\n**Inputs**\n- Required: `file`.\n- Optional: `line`, `symbol`, `timeout`.\n\n**Execution**\n- Sends `textDocument/hover`.\n- `extractHoverText()` flattens strings, markup content, marked-string objects, or arrays into plain text.\n\n**Output text**\n- `No hover information` or the extracted hover text.\n\n### `symbols`\n**Inputs**\n- Workspace mode: `file: \"*\"` or omitted file on the early workspace branch, plus required `query`.\n- Document mode: required `file`.\n- Optional: `timeout`.\n\n**Execution**\n- Workspace mode sends `workspace/symbol` to every non-custom LSP server, post-filters matches with `filterWorkspaceSymbols()`, deduplicates with `dedupeWorkspaceSymbols()`, then truncates to `WORKSPACE_SYMBOL_LIMIT`.\n- Document mode sends `textDocument/documentSymbol` to the primary server. If the first item has `selectionRange`, it formats hierarchical `DocumentSymbol`s; otherwise it formats flat `SymbolInformation`s.\n\n**Output text**\n- Workspace mode: `Found N symbol(s) matching \"query\":` plus formatted `name @ file:line:col`, with an omission line when over the limit.\n- Document mode: `Symbols in <file>:` plus hierarchical or flat symbol lines.\n\n### `rename`\n**Inputs**\n- Required: `file`, `new_name`.\n- Optional: `line`, `symbol`, `apply`, `timeout`.\n\n**Execution**\n- Requires `symbol` when `line` is given on project-aware servers, then waits for project load, sends `textDocument/rename`, receives a `WorkspaceEdit`.\n- `apply !== false` applies edits immediately with `applyWorkspaceEdit()`.\n- `apply === false` renders a preview with `formatWorkspaceEdit()`.\n\n**Output text**\n- `Rename returned no edits`, `Applied rename:` plus applied change lines, or `Rename preview:` plus summarized edits.\n\n### `rename_file`\n**Inputs**\n- Required: `file` source path, `new_name` destination path.\n- Optional: `apply`, `timeout`.\n\n**Execution**\n- Resolves absolute source and destination, rejects identical paths, missing source, existing destination, empty rename set, or directories with more than `MAX_RENAME_PAIRS` files.\n- `enumerateRenamePairs()` returns one `{oldUri,newUri}` pair for a file or walks every regular file in a directory tree.\n- Sends `workspace/willRenameFiles` with `{ files: pairs }` to every non-custom LSP server whose `fileTypes` match an affected path; collects returned `WorkspaceEdit`s and server notes.\n- Preview mode (`apply === false`) only formats those edits.\n- Apply mode coalesces the returned text edits per URI (a project-aware server's edits win on overlap; overlapping edits from other servers are discarded with a note), applies each URI once from a single snapshot, creates the destination parent directory and renames the source path on disk, sends `textDocument/didClose` for every renamed open file, deletes those `openFiles` entries, then sends `workspace/didRenameFiles`.\n\n**Output text**\n- Preview: `Rename preview: <file-count label> → <dest>` plus per-server edit summaries and optional server notes.\n- Apply: `Renamed <file-count label> → <dest>` plus applied edit summaries, filesystem rename line, and optional server notes.\n\n### `code_actions`\n**Inputs**\n- Required: `file`.\n- Optional: `line`, `symbol`, `query`, `apply`, `timeout`.\n\n**Execution**\n- Reads cached diagnostics for the open URI from `client.diagnostics` and sends `textDocument/codeAction` for a zero-width range at the resolved position.\n- When `apply !== true`, `query` is passed as `context.only: [query]`; this is a server-side kind filter.\n- When `apply === true`, `query` becomes a required client-side selector: either a zero-based numeric index or a case-insensitive substring of the action title.\n- Applying a `CodeAction` uses `applyCodeAction()`: optionally `codeAction/resolve`, then `applyWorkspaceEdit(edit)`, then optional `workspace/executeCommand`.\n- Applying a bare `Command` only runs `workspace/executeCommand`.\n\n**Output text**\n- List mode: `N code action(s):` plus `index: [kind] title` lines.\n- Apply mode success: `Applied \"title\":` plus `Workspace edit:` and/or `Executed command(s):` sections.\n- Apply mode miss: `No code action matches \"query\". Available actions:`.\n- Apply mode with no edit/command: `Action \"title\" has no workspace edit or command to apply`.\n\n### `status`\n**Inputs**\n- None.\n\n**Execution**\n- Reads configured servers from cached `LspConfig` and cross-references `getActiveClients()` so each server is labelled `(configured, not started)` or with its live client status.\n- Calls `detectLspmux()` and appends status text when `lspmux` is installed.\n\n**Output text**\n- `Language servers: <name (configured, not started) | name (<status>)>` plus an explanatory note line, or `No language servers configured for this project`, optionally followed by `lspmux: active (multiplexing enabled)` or `lspmux: installed but server not running`.\n\n### `reload`\n**Inputs**\n- Workspace mode: `file: \"*\"` or omitted `file`.\n- Single-file mode: required `file`.\n- Optional: `timeout`.\n\n**Execution**\n- Workspace mode reloads every non-custom LSP server.\n- Single-file mode reloads the primary server for that file.\n- `reloadServer()` tries `rust-analyzer/reloadWorkspace`, then `workspace/didChangeConfiguration` with `{ settings: {} }`; if neither works it kills the process so the next request cold-starts a new client.\n\n**Output text**\n- One line per server: `Reloaded <server>`, `Restarted <server>`, or `Failed to reload <server>: ...`.\n\n### `capabilities`\n**Inputs**\n- Optional: `file`, `timeout`.\n\n**Execution**\n- With a concrete `file`, inspects matching non-custom servers for that file.\n- With omitted `file` or `\"*\"`, inspects every non-custom configured server.\n- Starts servers as needed and dumps `client.serverCapabilities ?? {}` as pretty JSON.\n\n**Output text**\n- Per server: `<server>:` followed by indented `capabilities: { ... }`, or `<server>: failed to start (...)`.\n\n### `request`\n**Inputs**\n- Required: `query` method name.\n- Optional: `file`, `line`, `symbol`, `payload`, `timeout`.\n\n**Execution**\n- Chooses one non-custom server: file-scoped primary server, otherwise the first configured non-custom server.\n- Param building precedence:\n  1. If `payload` is present, parse JSON and use it verbatim.\n  2. Else if `file` is concrete and `line` is present, build `{ textDocument: { uri }, position: { line: line - 1, character } }` using `resolveSymbolColumn()`.\n  3. Else if `file` is concrete, build `{ textDocument: { uri } }`.\n  4. Else use `{}`.\n- Opens the file first when `file` is concrete.\n\n**Output text**\n- Success: `<server> ← <method>:\\n<formatted result>`, where non-string results are `JSON.stringify(..., null, 2)` and nullish values become `null`.\n- Failure: `LSP error from <server> on <method>: ...` followed by `  params: <preview>` echoing the request params (truncated to 400 chars).\n\n## Side Effects\n- Filesystem\n  - Reads config files, target files, and root markers.\n  - `rename` and `code_actions` may edit/create/delete/rename files via `applyWorkspaceEdit()`.\n  - `rename_file` always renames the source path on disk in apply mode.\n  - Server-initiated `workspace/applyEdit` requests also mutate files through `applyWorkspaceEdit()`.\n- Network\n  - None directly; communication is local stdio JSON-RPC to subprocesses.\n- Subprocesses / native bindings\n  - Spawns language servers with `ptree.spawn()`.\n  - Workspace diagnostics spawns `cargo`, `npx`, `go`, or `pyright`.\n  - `BiomeClient` and `SwiftLintClient` spawn CLI tools.\n  - Optional `lspmux` detection spawns `lspmux status`; supported servers may be wrapped through `lspmux client`.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - Caches config per cwd in `configCache`.\n  - Caches LSP clients per `command:cwd`, with `pendingRequests`, `diagnostics`, `openFiles`, `serverCapabilities`, and project-load state.\n  - Caches custom linter clients by `serverName:cwd`.\n  - Updates client `lastActivity`; optional idle-timeout cleanup is driven by `setIdleTimeout()`.\n- Background work / cancellation\n  - Every request has an abortable timeout signal.\n  - Aborting an in-flight LSP request sends `$/cancelRequest`.\n  - Background message readers persist for each live client until process exit/shutdown.\n\n## Limits & Caps\n- Tool timeout clamp: default `20`, min `5`, max `60` seconds — `TOOL_TIMEOUTS.lsp` in `packages/coding-agent/src/tools/tool-timeouts.ts`.\n- LSP request default timeout inside `sendRequest()`: `30_000ms` — `DEFAULT_REQUEST_TIMEOUT_MS` in `packages/coding-agent/src/lsp/client.ts`.\n- Warmup initialize timeout default: `5_000ms` — `WARMUP_TIMEOUT_MS` in `packages/coding-agent/src/lsp/client.ts`.\n- Project-load wait fallback: `15_000ms` — `PROJECT_LOAD_TIMEOUT_MS` in `packages/coding-agent/src/lsp/client.ts`.\n- Idle-client sweep interval when enabled: `60_000ms` — `IDLE_CHECK_INTERVAL_MS` in `packages/coding-agent/src/lsp/client.ts`.\n- Diagnostic message output cap: first `50` messages — `DIAGNOSTIC_MESSAGE_LIMIT` in `packages/coding-agent/src/lsp/index.ts`.\n- Single-file diagnostics wait: `3_000ms` — `SINGLE_DIAGNOSTICS_WAIT_TIMEOUT_MS`.\n- Batch/glob diagnostics wait per file: `400ms` — `BATCH_DIAGNOSTICS_WAIT_TIMEOUT_MS`.\n- Glob diagnostic target cap: first `20` matches — `MAX_GLOB_DIAGNOSTIC_TARGETS`.\n- Workspace symbol cap: first `200` entries — `WORKSPACE_SYMBOL_LIMIT`.\n- Reference context cap: first `50` references include source context — `REFERENCE_CONTEXT_LIMIT`.\n- References retry count: `2` retries, `250ms` backoff — `REFERENCES_RETRY_COUNT`, `REFERENCES_RETRY_DELAY_MS`.\n- Directory rename cap: `1_000` file pairs — `MAX_RENAME_PAIRS`.\n- `detectLspmux()` state cache TTL: `5 * 60 * 1000ms`; liveness check timeout: `1_000ms` — `STATE_CACHE_TTL_MS`, `LIVENESS_TIMEOUT_MS` in `packages/coding-agent/src/lsp/lspmux.ts`.\n- Workspace diagnostics output cap: first `50` lines from the subprocess.\n\n## Errors\n- Missing or invalid inputs are usually returned as text with `details.success: false`, not thrown:\n  - missing `file`/`query`/`new_name`\n  - invalid JSON in `payload`\n  - no matching server\n  - invalid `rename_file` source/destination conditions\n- `resolveSymbolColumn()` throws explicit errors for missing files, missing symbols, and out-of-bounds `#N` selectors; these surface as `LSP error: ...` or request-specific error text.\n- `sendRequest()` rejects on timeout with `LSP request <method> timed out after <ms>ms`.\n- Client process exit rejects all pending requests with an exit-code/stderr error assembled in `getOrCreateClient()`.\n- Single-file action failures inside the main `try` become `LSP error: <message>`.\n- `request` has its own error envelope: `LSP error from <server> on <method>: <message>`.\n- Some server failures are intentionally softened:\n  - diagnostics continue when one server fails\n  - `rename_file` suppresses `workspace/willRenameFiles` “method not found” errors and records other server errors as notes\n  - `code_actions` ignores `codeAction/resolve` failures and applies unresolved actions when possible\n- Caller aborts are not converted to text: `ToolAbortError` is rethrown. A wall-clock tool timeout without a caller abort instead throws `ToolError`: `LSP <action> timed out after <N>s on <server>. ...`.\n\n## Notes\n- `status` reports configured servers from `LspConfig` and labels each one via `getActiveClients()`: `(configured, not started)` means the binary resolves on PATH but no request has spawned it; a live client reports its status.\n- `getLspServerForFile()` excludes `createClient` adapters and linter-only servers; navigation/refactor actions never target Biome/SwiftLint custom clients.\n- `getServersForFile()` matches both file extensions and exact basenames from `fileTypes`; config can target names like `Dockerfile` if present.\n- `symbol` matching is exact first, then case-insensitive, and falls back to the Nth occurrence on the specified line only; it never scans other lines.\n- For `definition`, `references`, and `rename` against project-aware servers, omitting `symbol` while passing `line` is rejected with a `ToolError` instead of silently falling back to the first non-whitespace column.\n- `code_actions` uses `query` in two different ways: server-side `context.only` filter in list mode, client-side title/index selector in apply mode.\n- `rename` and `rename_file` default to apply. Preview requires `apply: false`.\n- `request` with `file: \"*\"` is treated the same as omitted `file`: it does not build workspace-specific params.\n- `reload` does not recreate a client immediately after killing it; the next request triggers reinitialization.\n- `workspace/applyEdit` can apply edits initiated by the server outside the direct tool action result path.\n- `detectLspmux()` can be disabled with `PI_DISABLE_LSPMUX=1`; only `rust-analyzer` is in `DEFAULT_SUPPORTED_SERVERS`.\n- Startup LSP discovery (`discoverStartupLspServers(cwd)` in `sdk.ts`) runs for `enableLsp && options.hasUI`; the background warmup additionally requires `!settings.get(\"lsp.lazy\")`. `lsp.lazy` defaults to `true`, so by default discovered servers are surfaced with status `\"available\"` (gray dot in the welcome screen) and cold-start through `getOrCreateClient()` on first use (lsp tool call or edit/write on a matching file type). Print/RPC/ACP/script sessions skip discovery and warmup entirely. See `docs/sdk.md` § Startup performance.\n- `configCache` is per-process and never auto-invalidated; config changes require a fresh process to be observed by `getConfig()` callers.",
 	"tools/read.md": "# read\n\n> Read files, directories, archives, SQLite databases, internal resources, images, documents, and URLs through one `path` string.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/read.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/read.md`\n- Key collaborators:\n  - `packages/coding-agent/src/tools/path-utils.ts` — split `path` from trailing selectors; normalize local paths.\n  - `packages/coding-agent/src/tools/archive-reader.ts` — detect `archive.ext:inner/path`, index archives, list/read entries.\n  - `packages/coding-agent/src/tools/sqlite-reader.ts` — detect SQLite targets, parse selectors, render tables.\n  - `packages/coding-agent/src/tools/fetch.ts` — URL parsing, fetch/render pipeline, URL cache/artifacts.\n  - `packages/coding-agent/src/internal-urls/router.ts` — resolve `agent://`, `artifact://`, `history://`, `issue://`, `local://`, `mcp://`, `memory://`, `omp://`, `pr://`, `rule://`, `skill://`, and `vault://`.\n  - `packages/coding-agent/src/edit/notebook.ts` — convert `.ipynb` to editable `# %% [...] cell:N` text.\n  - `packages/coding-agent/src/utils/file-display-mode.ts` — decide hashline vs line-number vs raw display.\n  - `packages/coding-agent/src/workspace-tree.ts` — render directory trees.\n  - `packages/coding-agent/src/edit/file-snapshot-store.ts` — stores read lines for later hashline edit verification/recovery.\n  - `packages/coding-agent/src/tools/index.ts` — registers `read: s => new ReadTool(s)`.\n\n## Inputs\n\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `path` | `string` | Yes | Filesystem path, internal URL, or web URL. May end with a trailing selector such as `:50-100` or `:raw`. |\n\n### Selector grammar\n\nFor normal file-like reads, `splitPathAndSel()` in `packages/coding-agent/src/tools/path-utils.ts` recognizes the final suffix only when it matches one of these forms:\n\n| Suffix | Meaning |\n| --- | --- |\n| `:raw` | Raw/verbatim mode. Disables structural summaries and line prefixes. |\n| `:conflicts` | Render unresolved Git merge-conflict regions for a local file. |\n| `:N` / `:LN` / `:N-` / `:N..` | Start at 1-indexed line `N`, open-ended. |\n| `:A-B` / `:LA-LB` / `:A..B` | Inclusive 1-indexed line range (`..` is a forgiving alias normalized to `-`). |\n| `:A+C` / `:LA+LC` | `C` lines starting at `A`; tool converts this to end line `A + C - 1`. |\n| `:R1,R2,...` | Multiple ranges, sorted and merged before reading (for example `:5-16,960-973`). |\n| `:range:raw` or `:raw:range` | Same line selection, but raw output. |\n\nValidation in `parseLineRangeChunk()`:\n- line numbers are 1-indexed; `:0` throws.\n- `+` counts must be `>= 1`.\n- `-` end must be `>= start`.\n\nSelector parsing intentionally falls through for unrecognized trailing `:...`; archive and SQLite paths consume their own colon syntax.\n\nURL selectors are parsed separately in `packages/coding-agent/src/tools/fetch.ts`, but use the same line-range parser for `:raw`, `:N`, `:A-B`, `:A+C`, `:5-10,20-30`, and `:range:raw` / `:raw:range`. Because URL ports also use `:`, add a trailing slash before a selector on a host/port URL, e.g. `https://example.com/:80`.\n\n## Outputs\n- Single-shot `AgentToolResult` built through `toolResult()` in `packages/coding-agent/src/tools/tool-result.ts`.\n- `content` is usually one text block. Image reads may return `[text, image]`.\n- `details` is path-dependent. `ReadToolDetails` may include:\n  - `kind: \"file\" | \"url\"` (URL path uses `kind: \"url\"`; file reads usually omit `kind`)\n  - `isDirectory`\n  - `resolvedPath`\n  - `suffixResolution`\n  - URL fields: `url`, `finalUrl`, `contentType`, `method`, `notes`\n  - `truncation`\n  - `displayContent` (unprefixed text + starting line for TUI rendering)\n  - `summary` (`lines`, `elidedSpans`, `elidedLines`) for structural summaries\n  - `meta` from `packages/coding-agent/src/tools/output-meta.ts`\n- `details.meta.source` is set to the backing path, URL, or internal URL.\n- `details.meta.truncation` carries shown range, total lines/bytes, next offset, and optional `artifactId` for cached URL output.\n- Directory/archive listings and SQLite table lists also set `details.meta.limits` when list limits trigger.\n\n## Flow\n1. `ReadTool.execute()` accepts `{ path }`. `file://...` inputs are expanded first with `expandPath()`.\n2. It tries URL handling first via `parseReadUrlTarget()` from `packages/coding-agent/src/tools/fetch.ts`.\n   - Plain URL reads call `executeReadUrl()`.\n   - URL reads with line selectors load or refresh the URL cache with `loadReadUrlCacheEntry()` and paginate the cached text locally with `#buildInMemoryTextResult()`.\n3. If not a web URL, it checks `session.internalRouter.canHandle(...)`.\n   - Internal URLs are resolved with `internalRouter.resolve()`.\n   - `agent://` query extraction (`/path` or `?q=`) bypasses pagination and returns the extracted content directly.\n   - Other internal resources are paginated in-memory by `#buildInMemoryTextResult()`.\n4. It tries archive resolution next with `#resolveArchiveReadPath()`.\n   - `parseArchivePathCandidates()` scans for `.tar`, `.tar.gz`, `.tgz`, or `.zip` anywhere before `:sub/path`.\n   - On success, `#readArchive()` either lists a directory or decodes an entry as UTF-8 text.\n5. It tries SQLite resolution with `#resolveSqliteReadPath()`.\n   - `parseSqlitePathCandidates()` scans for `.sqlite`, `.sqlite3`, `.db`, `.db3` before any `:table`, `:key`, or `?query` suffix.\n   - `#readSqlite()` dispatches on `parseSqliteSelector()`.\n6. Otherwise it treats the input as a local filesystem path.\n   - `resolveReadPath()` expands `~`, resolves relative to session cwd, treats bare `/` as session cwd, and retries macOS screenshot/NFD/curly-quote variants.\n   - If the path does not exist, `findUniqueSuffixMatch()` does a workspace glob-based unique suffix lookup (skipped for remote mounts).\n7. Directories go through `#readDirectory()`.\n8. Non-directories branch by content type:\n   - image metadata / inline image\n   - editable notebook text\n   - markit-converted document\n   - structural summary for parseable code/prose\n   - streamed text/line-range read\n9. Local text reads are streamed by `streamLinesFromFile()` rather than loading the whole file. The tool adds `1` leading and `3` trailing context lines around explicit bounded ranges (constrained sides only).\n10. Hashline-eligible local reads record a whole-file snapshot into the session snapshot store (`getFileSnapshotStore()` on `session.fileSnapshotStore`, `packages/coding-agent/src/edit/file-snapshot-store.ts`) for later hashline edit verification/recovery.\n11. If suffix resolution happened, the first text block is prefixed with `[Path '...' not found; resolved to '...' via suffix match]`.\n\n## Modes / Variants\n\n### Local text files\n- No selector: if summarization is enabled and the file is small enough, `#trySummarize()` calls `summarizeCode()`.\n  - Guards: file size `<= 2 MiB` (`MAX_SUMMARY_BYTES`), line count `<= 20_000` (`MAX_SUMMARY_LINES`).\n  - Summary output keeps selected declarations and replaces elided spans with `...` or merged brace-pair lines containing `..`. When at least one span is elided, the text content ends with a footer like `[NN lines elided; re-read needed ranges, e.g. <path>:5-16,40-80]` using concrete ranges from the actual elisions.\n  - When an elided block sits between matching brace lines, `#renderSummary()` may merge them into one anchored line rather than emitting separate opener/closer lines.\n- Explicit selector or summarization miss: streamed text read.\n  - Default open-ended limit is `min(session setting read.defaultLimit, DEFAULT_MAX_LINES)`.\n  - Explicit ranges expand by `RANGE_LEADING_CONTEXT_LINES = 1` / `RANGE_TRAILING_CONTEXT_LINES = 3` on the constrained sides only.\n  - Non-raw output uses `resolveFileDisplayMode()`:\n    - hashline numbered output when edit mode is hashline, read is not raw, source is mutable, edit tool exists, and `readHashLines !== false`\n    - otherwise optional line numbers when `readLineNumbers === true`\n    - raw mode suppresses both\n- Prefix format in hashline mode is a `[PATH#TAG]` header followed by `LINE:TEXT`, e.g. `[src/foo.ts#0A1B]` and `41:def alpha():`, from the session snapshot store plus `formatNumberedLine()` / `formatHashlineHeader()`.\n- The `edit`/hashline path consumes that header plus bare line numbers later; the four-hex tag is a content-derived hash of the whole normalized file, resolvable through the session snapshot store that recorded it. Immutable sources and `:raw` intentionally suppress hashline headers.\n\n### Directory listings\n- `#readDirectory()` calls `buildDirectoryTree()` with:\n  - `maxDepth = 2`\n  - `perDirLimit = 12`\n  - `rootLimit = null`\n  - `lineCap = limit` when a line selector was present, else unlimited at this layer\n- `buildDirectoryTree()` sorts siblings by recency, shows file sizes and relative ages, and may mark `limits.resultLimit` when the tree truncates.\n- Empty directories render as `(empty directory)`.\n\n### Archives\n- Supported archive containers: `.tar`, `.tar.gz`, `.tgz`, `.zip`.\n- Syntax: `archive.ext`, `archive.ext:path/inside`, `archive.ext:path/inside:50-60`.\n- `openArchive()` branches by format:\n  - tar/tgz reads the whole archive into memory (capped at `MAX_TAR_ARCHIVE_BYTES = 256 MiB`) and indexes it with `new Bun.Archive(bytes)`\n  - zip is indexed via ranged central-directory reads (`readZipEntries()`); entries are inflated on demand with `fflate.inflateSync()`, with declared member sizes capped at `MAX_ARCHIVE_MEMBER_BYTES = 64 MiB`\n- Archive paths normalize `/`, drop `.` segments, and reject `..`.\n- Directory reads list immediate children; files show `name` plus ` (size)` when size > 0.\n- Directory listing default limit is `500` entries in `#readArchiveDirectory()`.\n- File entries are UTF-8 decoded. Non-UTF-8 entries return `[Cannot read binary archive entry '...' (...)]` instead of bytes.\n- Text archive entries reuse the normal in-memory pagination/anchoring path.\n\n### SQLite databases\n- Database detection requires both a matching extension and a valid SQLite file header (`isSqliteFile()`).\n- Selector forms from `parseSqliteSelector()`:\n\n#### `db.sqlite`\n- `kind: \"list\"`\n- Lists non-`sqlite_%` tables with row counts.\n- `#readSqlite()` caps the rendered list to `500` tables via `applyListLimit()`.\n\n#### `db.sqlite:table`\n- `kind: \"schema\"`\n- Returns `sqlite_master.sql` plus sample rows.\n- Sample size is `DEFAULT_SCHEMA_SAMPLE_LIMIT = 5`.\n\n#### `db.sqlite:table:key`\n- `kind: \"row\"`\n- Resolves by primary key when the table has exactly one PK column; otherwise falls back to `rowid` lookup.\n- No query parameters allowed on row lookups.\n\n#### `db.sqlite:table?limit=...&offset=...&order=...&where=...`\n- `kind: \"query\"`\n- Defaults: `limit = 20`, `offset = 0`.\n- `limit` is capped at `500`.\n- `order` accepts `column` or `column:asc|desc` and must name an existing column.\n- `where` is accepted only after `validateWhereClause()` rejects comments, semicolons, and control keywords like `LIMIT`, `OFFSET`, `UNION`, `ATTACH`, `PRAGMA`.\n- Unknown query parameters throw.\n\n#### `db.sqlite?q=SELECT ...`\n- `kind: \"raw\"`\n- Cannot be combined with table selectors or any other query param.\n- Empty `q` throws.\n- `executeReadQuery()` prepares the SQL, rejects bound parameters, and collects rows from `statement.iterate()` capped at `MAX_RAW_QUERY_ROWS = 1000`; it does not verify that the SQL starts with `SELECT`.\n\n- Rendering caps in `packages/coding-agent/src/tools/sqlite-reader.ts`:\n  - ASCII table width `120` (`MAX_RENDER_WIDTH`)\n  - per-column width `40` (`MAX_COLUMN_WIDTH`)\n- `#readSqlite()` opens Bun SQLite in `{ readonly: true, strict: true }` and sets `PRAGMA busy_timeout = 3000`.\n\n### Documents\n- `CONVERTIBLE_EXTENSIONS` in `packages/coding-agent/src/tools/read.ts` covers `.pdf`, `.doc`, `.docx`, `.ppt`, `.pptx`, `.xls`, `.xlsx`, `.rtf`, `.epub`.\n- `convertFileWithMarkit()` converts the file to text/markdown.\n- Converted output is then head-truncated with normal shared limits; there is no line selector support inside the source document before conversion.\n- Conversion failures return a text block like `[Cannot read .pdf file: ...]`.\n\n### Jupyter notebooks\n- `.ipynb` goes through `readEditableNotebookText()` unless `:raw` was requested.\n- Output is editable plain text with markers like:\n\n```text\n# %% [code] cell:0\n...\n```\n\n- Raw mode bypasses that conversion and falls back to file-text reading.\n\n### Images\n- Image detection is metadata-based (`readImageMetadata()`).\n- Max accepted image size is `20 MiB` (`MAX_IMAGE_INPUT_BYTES`, re-exported as `MAX_IMAGE_SIZE`). Larger files throw.\n- If `inspect_image.enabled` is true, `read` returns metadata only (MIME, bytes, dimensions, channels, alpha) plus a suggestion to call `inspect_image`.\n- Otherwise it calls `loadImageInput()` and returns:\n  - a text note from the image loader\n  - an inline image block\n- Unsupported/undecodable image formats throw a `ToolError`.\n\n### Internal URLs\n- `read` does not resolve these itself; it delegates to `session.internalRouter.resolve()`.\n- Registered protocols are outside this file, but the router in `packages/coding-agent/src/internal-urls/router.ts` is built for `agent://`, `artifact://`, `history://`, `issue://`, `local://`, `mcp://`, `memory://`, `omp://`, `pr://`, `rule://`, `skill://`, and `vault://`.\n- `#handleInternalUrl()` behavior:\n  - parses the URL with `parseInternalUrl()` so colons inside the host segment are legal\n  - for `agent://`, treats non-root path extraction or `?q=` extraction as a special no-pagination mode\n  - otherwise paginates the resolved text in memory\n  - passes `immutable` through to `resolveFileDisplayMode()` so anchors are suppressed for immutable resources such as artifacts, skills, memory, and agent outputs\n  - sets `ignoreResultLimits: true` for `skill://` so the full skill text is paginated only by explicit selectors, not by the normal default line limit\n- `issue://<N>` / `pr://<N>` (and the long form `issue://<owner>/<repo>/<N>` / `pr://<owner>/<repo>/<N>`) route through the same SQLite cache the `github` tool writes to; `?comments=0` selects the no-comments rendering. Bare `issue://` / `pr://` (and `issue://<owner>/<repo>` / `pr://<owner>/<repo>`) issue a live `gh issue list` / `gh pr list` for browsing, accepting `?state=`, `?limit=`, `?author=`, `?label=`. PR diffs share the same cache through `pr://<N>/diff` (numbered file listing with per-file hints), `pr://<N>/diff/<i>` (single file slice; 1-indexed), and `pr://<N>/diff/all` (verbatim unified diff); the listing and per-file slices are reconstructed from the cached unified-diff payload, so all three variants share one `gh pr diff` invocation per PR. Diff content is served as `text/plain`. Soft TTL `github.cache.softTtlSec` (default 5 minutes), hard TTL `github.cache.hardTtlSec` (default 7 days). Stale-hit returns the cached row and schedules a background refresh.\n\n### Web URLs\n- `parseReadUrlTarget()` accepts `http://`, `https://`, or `www.` targets.\n- Plain URL reads call `executeReadUrl()` in `packages/coding-agent/src/tools/fetch.ts`.\n- `:raw` means raw HTML/body fallback path; plain URL reads prefer rendered/reader-friendly output.\n- `:N`, `:A-B`, `:A+C`, and comma-separated multi-ranges do not refetch when cached output is usable. They page over cached output from the prior or current URL render.\n- URL render pipeline in `renderUrl()`:\n  1. normalize scheme (`https://` added for bare `www.`)\n  2. try special handlers for known sites unless raw\n  3. fetch with `loadPage()`\n  4. if content is image/PDF/DOCX/etc., try binary fetch + markit/image handling\n  5. handle JSON directly, feeds via feed parser, plain text directly\n  6. for HTML and non-raw mode, try markdown alternates, `URL.md`, content negotiation, feed alternates, HTML-to-text renderers, extracted linked documents, then `llms.txt`\n  7. fall back to raw body text/html\n- URL output is wrapped with a small header:\n\n```text\nURL: ...\nContent-Type: ...\nMethod: ...\nNotes: ...\n\n---\n```\n\n- `method` records the winning path (`json`, `feed`, `text`, `alternate-markdown`, `md-suffix`, `content-negotiation`, `image`, `markit`, `llms.txt`, `raw`, `raw-html`, etc.).\n- URL reads may return an inline image block when the fetched resource is a supported image and survives resizing.\n\n## Side Effects\n- Filesystem\n  - Opens and streams local files.\n  - Reads tar/tgz archives fully into memory before indexing (256 MiB cap); ZIP archives are indexed via ranged central-directory reads.\n  - May read URL-cache artifact files from the session artifacts directory.\n  - Writes URL output artifacts when URL output is truncated or when line-range pagination needs a persisted cache body.\n- Network\n  - URL mode performs HTTP fetches, binary refetches, and alternate-endpoint probes.\n- Subprocesses / native bindings\n  - Uses Bun SQLite for `.db`/`.sqlite*`.\n  - Uses `Bun.Archive` for tar/tgz and `fflate` for zip.\n  - URL HTML rendering can delegate into site handlers and HTML-to-text backends from `packages/coding-agent/src/tools/fetch.ts`.\n- Session state\n  - Records whole-file snapshots of local text reads into `session.fileSnapshotStore` for later stale-anchor recovery.\n  - Uses `session.internalRouter` for internal URLs.\n  - Uses `session.allocateOutputArtifact()` for cached/truncated URL output.\n- Background work / cancellation\n  - Only the deterministic disk reads are non-abortable: plain-file line/range reads (`streamLinesFromFile`, multi-range) and directory listings (`#readDirectory`) are called with `undefined` instead of the `AbortSignal`, so an interrupt mid-read can't surface a misleading \"Operation aborted\" on a read that would have finished instantly. Every other branch keeps the signal and its helpers call `throwIfAborted(signal)` to stop promptly: URL/internal-URL reads (network), archive, sqlite, document conversion, image decode, structural summary, conflict scan, and the suffix-glob path resolution.\n\n## Limits & Caps\n- Shared text truncation defaults from `packages/coding-agent/src/session/streaming-output.ts`:\n  - `DEFAULT_MAX_LINES = 3000`\n  - `DEFAULT_MAX_BYTES = 50 * 1024`\n- Local text open-ended default line limit: `read.defaultLimit`, clamped to `[1, DEFAULT_MAX_LINES]`.\n- Explicit line ranges add `1` leading and `3` trailing context lines on the constrained sides (`RANGE_LEADING_CONTEXT_LINES` / `RANGE_TRAILING_CONTEXT_LINES`).\n- File streaming chunk size: `8 * 1024` bytes (`READ_CHUNK_SIZE`).\n- Local streamed byte budget for line reads: `max(DEFAULT_MAX_BYTES, maxLinesToCollect * 512)`.\n- Structural summaries only run when file size `<= 2 MiB` and line count `<= 20_000`.\n- Image input max: `20 MiB`.\n- Directory tree caps for local directories: depth `2`, per-directory children `12`.\n- Archive directory default list cap: `500` entries.\n- SQLite:\n  - default row query limit `20`\n  - schema sample limit `5`\n  - max query limit `500`\n  - raw `?q=` row cap `1000` (`MAX_RAW_QUERY_ROWS`)\n  - table list cap `500`\n  - render width `120`, column width `40`\n  - busy timeout `3000` ms\n- URL read result shown to the model is truncated to `300` lines and `50 KiB` in `executeReadUrl()`; full cached output can be attached as an artifact.\n- Inline fetched URL images:\n  - source bytes cap `20 MiB`\n  - post-resize inline output cap `300 KiB`\n- Unique suffix auto-resolution glob timeout: `5000` ms.\n- File snapshot store holds `30` paths with up to `4` versions each (`DEFAULT_MAX_PATHS` / `DEFAULT_MAX_VERSIONS_PER_PATH` in `packages/hashline/src/snapshots.ts`); files over `4 MiB` (`SNAPSHOT_MAX_BYTES`) are not snapshotted.\n\n## Errors\n- Validation and operational failures surface as `ToolError`.\n- Selector errors include:\n  - `Line selector 0 is invalid; lines are 1-indexed. Use :1.`\n  - invalid `A+B` / `A-B` shapes\n  - `Cannot combine query extraction with line selectors` for `agent://.../path:50`\n- Missing local/archive/sqlite paths first attempt unique suffix resolution; if no unique match exists they error.\n- Out-of-bounds line reads do not throw. They return explanatory text with a suggestion such as `Use :1 ...` or `Use :<last line> ...`.\n- Binary archive entries do not throw; they return a text notice.\n- Document conversion failure returns a text notice.\n- Image oversize/unsupported/invalid cases throw.\n- SQLite parser rejects unsupported parameter combinations early; DB/runtime errors are caught and rethrown as `ToolError(message)`.\n- URL fetch failure does not throw when HTTP fetch succeeds but `response.ok === false`; it returns a failed URL read with `method: \"failed\"` and explanatory notes.\n\n## Notes\n- Hashline anchors are suppressed for raw reads and immutable internal resources because there is no editable backing target for later `edit` consumption.\n- `splitPathAndSel()` intentionally treats unknown trailing `:...` as part of the path so `archive.zip:inner/file` and `db.sqlite:table:key` still work.\n- `resolveReadPath()` contains macOS-specific filename fallbacks for screenshot timestamps, NFD Unicode normalization, and curly apostrophes.\n- A bare `/` resolves to the session cwd, not the filesystem root.\n- URL cache keys are session-scoped and normalized by requested URL + raw/rendered mode; both requested URL and final redirected URL are cached.\n- URL line-range reads request `ensureArtifact: true, preferCached: true` so a later paginated read can reopen the same rendered body from artifact storage.\n- Raw SQLite `q=` execution is not keyword-restricted beyond “no bound parameters”; the read tool relies on the surrounding contract to keep it read-only.\n- The file snapshot store is not a read acceleration cache. It exists to verify and recover hashline edits when the file changed after the read.",
 	"tools/recall.md": "# recall\n\n> Search the active long-term memory backend and return matching memories.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/memory-recall.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/recall.md`\n- Hindsight collaborators:\n  - `packages/coding-agent/src/hindsight/state.ts` — session state, recall query defaults, prompt-side auto-recall.\n  - `packages/coding-agent/src/hindsight/content.ts` — result formatting and UTC timestamp formatting.\n  - `packages/coding-agent/src/hindsight/client.ts` — HTTP `recall` call and error mapping.\n  - `packages/coding-agent/src/hindsight/bank.ts` — bank id and tag-filter scoping.\n- Mnemopi collaborators:\n  - `packages/coding-agent/src/mnemopi/state.ts` — scoped local recall and result formatting with ids.\n  - `packages/coding-agent/src/mnemopi/config.ts` — local bank scoping and recall limits.\n  - `docs/tools/retain.md` — shared backend, storage, scoping, and retention behavior.\n\n## Inputs\n\n| Field | Type | Required | Description |\n|---|---|---:|---|\n| `query` | `string` | Yes | Natural-language search query. The tool passes it through unchanged except Mnemopi `per-project-tagged` may run an internal shared-bank fallback query. |\n\n## Outputs\nReturns a single-shot tool result.\n\nWhen matches exist:\n- `content[0].type = \"text\"`\n- `content[0].text = \"Found <n> relevant memory/memories (as of YYYY-MM-DD HH:MM UTC):\\n\\n<bullet list>\"`\n- `details = {}`\n\nHindsight bullet format comes from `formatMemories(...)`:\n- each bullet is `- <text> [<type>] (<mentioned_at>)`; the type and timestamp suffixes appear only when those fields are present.\n\nMnemopi bullet format comes from `formatScopedRecallWithIds(...)`:\n- each bullet is `- <content> (id: <id>|id unavailable) [<source>] (<YYYY-MM-DD>) c:<score>`; optional source, date, and score suffixes appear only when present.\n\nWhen no matches exist:\n- `content[0].text = \"No relevant memories found.\"`\n- `details = {}`\n\n## Flow\n1. `MemoryRecallTool.createIf(...)` exposes the tool when `memory.backend` is either `\"hindsight\"` or `\"mnemopi\"`.\n2. `execute(...)` wraps the operation in `untilAborted(...)`.\n3. If the backend is `mnemopi`:\n   - it reads `session.getMnemopiSessionState()` and throws if the backend was not started;\n   - it calls `state.recallResultsScoped(params.query)`;\n   - scoped recall queries each configured recall bank with `recallEnhanced(query, recallLimit, { includeFacts: true, channelId: bank })`, merges/deduplicates results by id/content, sorts them, and truncates to `recallLimit`;\n   - in `per-project-tagged`, the shared bank may receive one extra fallback query with project-bank literal tokens stripped so broad global memories still match;\n   - results are formatted with ids for later `memory_edit` use.\n4. If the backend is `hindsight`:\n   - it reads `session.getHindsightSessionState()` and throws if the backend was not started;\n   - it calls `state.client.recall(...)` with `bankId`, query, configured `budget`, `maxTokens`, `types`, and bank-scope tag filters;\n   - `HindsightApi.recall(...)` POSTs `/v1/default/banks/{bank_id}/memories/recall`;\n   - results are formatted into a plain-text list with `formatMemories(...)`.\n5. Backend failures are logged with `logger.warn(\"recall failed\", ...)` and rethrown as `Error` instances when needed.\n\n## Modes / Variants\n- Tool path: explicit query-only recall. It does not compose context from recent turns.\n- Backend auto-recall has a richer query-composition path in `HindsightSessionState.beforeAgentStartPrompt(...)` / `maybeRecallOnAgentStart(...)` and `MnemopiSessionState.beforeAgentStartPrompt(...)` / `maybeRecallOnAgentStart(...)`.\n- Hindsight bank scoping:\n  - `global` — no tag filter.\n  - `per-project` — separate bank id per project label (git primary checkout root basename; cwd basename outside a repo).\n  - `per-project-tagged` — shared bank id plus `project:<project label>` filter with `tagsMatch = \"any\"`, so project-tagged and untagged global memories can both surface.\n- Mnemopi bank scoping:\n  - `global` — recall reads the shared bank.\n  - `per-project` — recall reads the project bank.\n  - `per-project-tagged` — recall reads the project bank and shared bank, then merges results.\n- Session scope: reads cross-session memory data, using the active session's cached config and scope.\n\n## Side Effects\n- Network\n  - Hindsight: `POST /v1/default/banks/{bank_id}/memories/recall`.\n  - Mnemopi: none unless configured local runtime providers perform embedding/LLM work during recall.\n- Session state\n  - None on success for the explicit tool path. Unlike backend auto-recall, this tool does not update `lastRecallSnippet` or refresh the system prompt.\n- Background work / cancellation\n  - Aborts through `untilAborted(...)` if the tool call signal is cancelled.\n\n## Limits & Caps\n- Tool availability requires `memory.backend` to be `\"hindsight\"` or `\"mnemopi\"`; default `memory.backend` is `\"off\"`.\n- Hindsight client default budget for raw `HindsightApi.recall(...)` is `\"mid\"`; this tool overrides from config.\n- Hindsight recall settings:\n  - `hindsight.recallBudget = \"mid\"`\n  - `hindsight.recallMaxTokens = 1024`\n  - `hindsight.recallTypes = [\"world\", \"experience\"]`\n- Mnemopi recall settings:\n  - `mnemopi.recallLimit = 8`\n  - `mnemopi.scoping` selects which local bank(s) are searched\n- The explicit tool path does not apply `hindsight.recallContextTurns`, `hindsight.recallMaxQueryChars`, `mnemopi.recallContextTurns`, or `mnemopi.recallMaxQueryChars`; those caps only affect backend auto-recall query composition.\n\n## Errors\n- Throws `Mnemopi backend is not initialised for this session.` when `memory.backend == \"mnemopi\"` but no state exists.\n- Throws `Hindsight backend is not initialised for this session.` when `memory.backend == \"hindsight\"` but no state exists.\n- Hindsight HTTP and fetch failures become `HindsightError` with `statusCode` and parsed `details` when available.\n- Mnemopi recall target failures inside `collectScopedRecallResults(...)` are caught per bank and logged only when `mnemopi.debug` is enabled; if all targets fail, the tool can return `No relevant memories found.`\n- Non-`Error` failures caught by the tool are normalized to `new Error(String(err))` before rethrow.\n\n## Notes\n- Shared backend details are in `docs/tools/retain.md`: storage, subagent aliasing, bank scoping, mission setup, and mental-model behavior.\n- Hindsight mental models are not fetched by this tool. They may already be present in the agent's developer instructions because the backend caches a `<mental_models>` block separately from recall results.\n- Mnemopi developer instructions may include a `<memories>` block from auto-recall; this explicit tool does not update that block.\n- The tool returns memory hits; it does not synthesize across them. Use `reflect` for that path.\n",
@@ -104,7 +104,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"tools/write.md": "# write\n\n> Create or overwrite a file, writable internal resource, archive entry, SQLite row, or merge-conflict resolution.\n\n## Source\n- Entry: `packages/coding-agent/src/tools/write.ts`\n- Model-facing prompt: `packages/coding-agent/src/prompts/tools/write.md`\n- Key collaborators:\n  - `packages/coding-agent/src/tools/archive-reader.ts` — parse `archive.ext:entry` selectors.\n  - `packages/coding-agent/src/tools/sqlite-reader.ts` — detect SQLite paths and perform row insert/update/delete.\n  - `packages/coding-agent/src/lsp/index.ts` — format-on-write and diagnostics writethrough.\n  - `packages/coding-agent/src/tools/auto-generated-guard.ts` — block overwriting generated files.\n  - `packages/coding-agent/src/tools/fs-cache-invalidation.ts` — invalidate shared FS scan caches after writes.\n  - `packages/coding-agent/src/tools/plan-mode-guard.ts` — resolve paths and enforce plan-mode write policy.\n\n## Inputs\n| Field | Type | Required | Description |\n| --- | --- | --- | --- |\n| `path` | `string` | Yes | Target path. Plain file path writes a filesystem file. Writable internal URLs are delegated to their handler. `archive.ext:inner/path` writes an archive entry for `.tar`, `.tar.gz`, `.tgz`, or `.zip`. `db.sqlite:table` inserts a row. `db.sqlite:table:key` updates or deletes a row. `conflict://<id>` resolves a recorded merge conflict; `conflict://*` bulk-resolves every registered conflict. |\n| `content` | `string` | Yes | Full replacement file content, archive entry content, internal-resource content, conflict replacement, or SQLite row payload. SQLite non-delete writes must parse as a JSON5 object. Empty or whitespace-only content deletes a SQLite row when `path` includes a row key. |\n\nWorked examples:\n\n```text\npath: \"src/generated/config.json\"\ncontent: \"{\\n  \\\"enabled\\\": true\\n}\\n\"\n```\n\n```text\npath: \"fixtures/archive.zip:templates/email.txt\"\ncontent: \"hello\\n\"\n```\n\n```text\npath: \"data/app.sqlite:users:42\"\ncontent: \"{name: 'Ada', active: true}\"\n```\n\n## Outputs\nSingle-shot result.\n\n- Success always returns a text block.\n  - Plain file write: `Successfully wrote <chars> bytes to <relative-path>` (the count is `cleanContent.length`, not encoded byte length).\n  - Internal URL write: `Successfully wrote <chars> bytes to <url>`.\n  - Archive write: `Successfully wrote <chars> bytes to <relative-archive-path>:<entry-path>`.\n  - SQLite write: one of `Inserted row into <table>`, `Updated row '<key>' in <table>`, `No row updated ...`, `Deleted row ...`, `No row deleted ...`.\n  - Conflict resolution: conflict-specific success text, with fresh hashline snapshot headers when applicable.\n- If hashline prefixes were copied from `read` output and stripped first, the first text block gets an extra note.\n- In hashline display mode, plain file writes (including ACP bridge writes) and conflict resolutions prepend a fresh `[<relative-path>#TAG]` header so the next `edit` has a current snapshot tag without an extra `read`. Bulk conflict resolutions append a `Snapshots:` block listing one header per successfully written file.\n- Plain file writes may also return `details.diagnostics` plus `details.meta.diagnostics` when LSP diagnostics-on-write is enabled, and `details.madeExecutable` when a newly written shebang file is chmodded executable.\n- SQLite writes use `toolResult(...).sourcePath(...)`, so `details.meta.sourcePath` points at the database file.\n- Archive writes set `details.resolvedPath` to the archive's absolute path; internal URL writes return empty `details`.\n\n## Flow\n1. `WriteTool.execute()` in `packages/coding-agent/src/tools/write.ts` strips pasted `[PATH#HASH]` headers and `LINE:` hashline prefixes from `content` when the session is in hashline display mode.\n2. If `path` is an internal URL whose handler exposes `write`, the tool delegates directly to `handler.write(...)` and returns.\n3. `conflict://...` paths are handled next by the merge-conflict resolver. Scope reads such as `conflict://<id>/ours` are rejected as read-only; writable conflict URIs must omit the scope.\n4. It calls `#resolveArchiveWritePath()` next. That uses `parseArchivePathCandidates()` from `packages/coding-agent/src/tools/archive-reader.ts`, checks candidate archive files on disk, and falls back to the longest matching archive suffix even when the archive file does not exist yet.\n5. Archive writes call `enforcePlanModeWrite(..., { op: exists ? \"update\" : \"create\" })`, then `#writeArchiveEntry()`.\n   - The parent directory of the archive file is created with `fs.mkdir(..., { recursive: true })`.\n   - `.zip` archives are read with `fflate.unzipSync()`, the target entry is replaced in an in-memory map, and the archive is rewritten with `fflate.zipSync()` + `Bun.write()`.\n   - `.tar`, `.tar.gz`, and `.tgz` archives are read with `Bun.Archive`, existing entries are copied into an object map, the target entry is replaced, and `Bun.Archive.write()` rewrites the archive.\n   - `invalidateFsScanAfterWrite()` runs on the archive file path.\n6. If the path is not treated as an archive, `execute()` calls `#resolveSqliteWritePath()`. That uses `parseSqlitePathCandidates()` and `isSqliteFile()` from `packages/coding-agent/src/tools/sqlite-reader.ts`. Existing non-SQLite files suppress the SQLite path interpretation.\n7. SQLite writes call `enforcePlanModeWrite(..., { op: \"update\" })`, then `#writeSqliteRow()`.\n   - The database must already exist; missing DBs throw `SQLite database '<path>' not found`.\n   - The tool opens `new Database(..., { create: false, strict: true })` and sets `PRAGMA busy_timeout = 3000`.\n   - Whitespace-only `content` with a row key deletes a row.\n   - Non-empty `content` is parsed with `Bun.JSON5.parse()`, must be a JSON object, and is routed to insert/update helpers from `packages/coding-agent/src/tools/sqlite-reader.ts`.\n   - `invalidateFsScanAfterWrite()` runs on the DB path and the connection is closed in `finally`.\n8. Otherwise the tool treats `path` as a plain filesystem file.\n   - `enforcePlanModeWrite(..., { op: \"create\" })` runs before path resolution.\n   - Existing files are checked by `assertEditableFile()` to block overwriting detected generated files.\n   - ACP bridge writeTextFile is tried first when available; otherwise the session’s writethrough callback writes content. With LSP enabled and `lsp.formatOnWrite` / `lsp.diagnosticsOnWrite` settings on, `createLspWritethrough()` may format content, sync it through LSP servers, save it, and collect diagnostics. Otherwise `writethroughNoop()` writes directly with `Bun.write()` or `file.write()`.\n   - `maybeMarkExecutableForShebang()` may chmod the file executable when content starts with `#!`.\n   - `invalidateFsScanAfterWrite()` runs on the file path.\n9. The tool returns a text result and optional diagnostics / executable metadata.\n\n## Modes / Variants\n### Plain file path\n- Target is any path that does not resolve as an archive selector and does not resolve as an existing-or-new SQLite selector.\n- Existing files are overwritten.\n- `write.ts` does not call `fs.mkdir()` on this path; explicit parent-directory creation only exists in the archive branch, but `Bun.write()` itself creates missing parent directories for plain file writes.\n\nExample:\n\n```text\npath: \"tmp/output.txt\"\ncontent: \"hello\\n\"\n```\n\n### Archive entry write\n- Selector syntax: `archive.ext:inner/path`.\n- Supported archive suffixes come from `parseArchivePathCandidates()`: `.tar`, `.tar.gz`, `.tgz`, `.zip`.\n- The inner path is normalized to `/`, strips empty and `.` segments, rejects `..`, and rejects directory targets ending in `/`.\n- Rewrites the whole archive file after replacing one entry.\n- Creates the parent directory for the archive file if needed.\n\nExample:\n\n```text\npath: \"build/assets.tar.gz:css/app.css\"\ncontent: \"body { color: black; }\\n\"\n```\n\n### SQLite table insert\n- Selector syntax: `db.sqlite:table`.\n- `content` must parse as a JSON5 object.\n- Empty object is allowed and becomes `INSERT INTO <table> DEFAULT VALUES`.\n- Query parameters are rejected for SQLite writes.\n\nExample:\n\n```text\npath: \"data/app.db:users\"\ncontent: \"{name: 'Ada', active: true}\"\n```\n\n### SQLite row update / delete\n- Selector syntax: `db.sqlite:table:key`.\n- Non-empty `content` updates the row.\n- Empty or whitespace-only `content` deletes the row.\n- Row lookup uses the single-column primary key if present; otherwise it falls back to `rowid`. Composite primary keys and `WITHOUT ROWID` tables are rejected for key-based writes.\n\nExample update:\n\n```text\npath: \"data/app.sqlite:users:42\"\ncontent: \"{email: 'ada@example.com'}\"\n```\n\nExample delete:\n\n```text\npath: \"data/app.sqlite:users:42\"\ncontent: \"\"\n```\n\n## Side Effects\n- Filesystem\n  - Creates or overwrites plain files.\n  - Rewrites entire archive files when writing an archive entry.\n  - Explicitly creates parent directories (via `fs.mkdir`) for archive files only; plain file writes get parent directories from `Bun.write()`.\n  - Mutates existing SQLite databases; never creates a new SQLite DB.\n  - Resolves conflict markers in files for `conflict://...` writes.\n  - May chmod a shebang file executable after a successful plain-file write.\n- Subprocesses / native bindings\n  - Uses Bun SQLite bindings via `bun:sqlite`.\n  - Uses Bun archive APIs and lazily imports `fflate` for ZIP reads/writes.\n  - May talk to configured LSP servers through `packages/coding-agent/src/lsp/index.ts`.\n- Session state (transcript, memory, jobs, checkpoints, registries)\n  - Invalidates shared filesystem scan cache entries through `invalidateFsScanAfterWrite()`.\n  - Enforces plan-mode write restrictions before mutating the target.\n- Background work / cancellation\n  - Marks the tool `concurrency = \"exclusive\"` in `WriteTool`.\n  - LSP writethrough can schedule deferred diagnostics fetches after a timeout, but plain `write.ts` only consumes the immediate return value.\n\n## Limits & Caps\n- `WriteTool` itself exposes no byte cap beyond storing `content` in memory and, for archives, rebuilding the archive in memory.\n- Generated-file detection reads at most `CHECK_BYTE_COUNT = 1024` bytes and `HEADER_LINE_LIMIT = 40` header lines from an existing file in `packages/coding-agent/src/tools/auto-generated-guard.ts`.\n- SQLite writes set `PRAGMA busy_timeout = 3000`.\n- LSP writethrough uses a `5_000` ms operation timeout in `runLspWritethrough()` and may schedule a deferred diagnostics fetch with `AbortSignal.timeout(25_000)` in `scheduleDeferredDiagnosticsFetch()`.\n- Shebang executable handling depends on host filesystem chmod support.\n\n## Errors\n- Invalid archive subpaths throw `ToolError` with messages such as:\n  - `Archive write path must target a file inside the archive`\n  - `Archive write path must target a file, not a directory`\n  - `Archive path cannot contain '..'`\n- SQLite path parsing throws on unsupported forms:\n  - `SQLite write paths do not support query parameters`\n  - `SQLite write path must target a table`\n  - `SQLite row writes require a non-empty row key`\n- Missing SQLite DBs surface as `SQLite database '<path>' not found`.\n- SQLite content errors are model-visible `ToolError`s, including invalid JSON5, non-object payloads, unknown columns, non-scalar values, empty update objects, composite primary keys, and `WITHOUT ROWID` tables.\n- Existing plain files may be rejected by `assertEditableFile()` when they look generated.\n- Conflict scope writes such as `conflict://<id>/ours` are rejected as read-only; invalid conflict IDs or missing conflict history surface as `ToolError`s from the conflict resolver.\n- Archive read/write failures and unexpected SQLite exceptions are wrapped in `ToolError(error.message)`.\n- If no LSP server matches or LSP formatting/diagnostics times out, file writes still fall back to writing content; diagnostics may be omitted.\n\n## Notes\n- Archive path detection runs before SQLite detection. A path that matches an archive selector is never treated as SQLite.\n- SQLite detection declines when an existing file with a `.sqlite` / `.db` suffix is present but does not have SQLite magic bytes; then the path falls back to a plain file write.\n- ZIP entry content is encoded with `new TextEncoder().encode(content)` in `#writeArchiveEntry()`. Non-ZIP archive writes pass the string directly to `Bun.Archive.write()`.\n- The prompt forbids two common anti-patterns: using `write` for routine edits that should use `edit`, and creating `*.md` / `README` files unless explicitly requested. It also forbids emojis unless requested.\n- Plain file and internal URL writes report `cleanContent.length` as “bytes”, which is UTF-16 code units in JS, not an on-disk byte measurement.\n- `stripWriteContent()` only removes hashline prefixes when the session’s file display mode has `hashLines` enabled; otherwise content is written unchanged.\n",
 	"tree.md": "# `/tree` Command Reference\n\n`/tree` opens the interactive **Session Tree** navigator. It lets you jump to any entry in the current session file and continue from that point.\n\nThis is an in-file leaf move, not a new session export.\n\n## What `/tree` does\n\n- Builds a tree from current session entries (`SessionManager.getTree()`)\n- Opens `TreeSelectorComponent` with keyboard navigation, filters, and search\n- On selection, calls `AgentSession.navigateTree(targetId, { summarize, customInstructions })`\n- Rebuilds visible chat from the new leaf path\n- Optionally prefills editor text when selecting a user/custom message\n\nPrimary implementation:\n\n- `src/slash-commands/builtin-registry.ts` (`/tree`, `/branch` command routing)\n- `src/modes/controllers/input-controller.ts` (keybinding wiring, double-escape behavior)\n- `src/modes/controllers/selector-controller.ts` (tree UI launch + summary prompt flow)\n- `src/modes/components/tree-selector.ts` (navigation, filters, search, labels, rendering)\n- `src/session/agent-session.ts` (`navigateTree` leaf switching + optional summary)\n- `src/session/session-manager.ts` (`getTree`, `branch`, `branchWithSummary`, `resetLeaf`, label persistence)\n\n## How to open it\n\nAny of the following opens the same selector:\n\n- `/tree`\n- configured keybinding for the `app.session.tree` action\n- double-escape on empty editor when `doubleEscapeAction = \"tree\"` (default)\n- `/branch` when `doubleEscapeAction = \"tree\"` (routes to tree selector instead of user-only branch picker)\n\n## Tree UI model\n\nThe tree is rendered from session entry parent pointers (`id` / `parentId`).\n\n- Children are sorted by timestamp ascending (older first, newer lower)\n- Active branch (path from root to current leaf) is marked with a bullet\n- Labels (if present) render as `[label]` before node text\n- If multiple roots exist (orphaned/broken parent chains), they are shown under a virtual branching root\n\n```text\nExample tree view (active path marked with •):\n\n├─ user: \"Start task\"\n│  └─ assistant: \"Plan\"\n│     ├─ • user: \"Try approach A\"\n│     │  └─ • assistant: \"A result\"\n│     │     └─ • [milestone] user: \"Continue A\"\n│     └─ user: \"Try approach B\"\n│        └─ assistant: \"B result\"\n```\n\nThe selector recenters around current selection and shows up to:\n\n- `max(5, floor(terminalHeight / 2))` rows\n\n## Keybindings inside tree selector\n\n- `Up` / `Down`: move selection (wraps)\n- `Left` / `Right`: page up / page down\n- `Enter`: select node\n- `Esc`: clear search if active; otherwise close selector\n- `Ctrl+C`: close selector\n- `Type`: append to search query\n- `Backspace`: delete search character\n- `Shift+L`: edit/clear label on selected entry\n- `Ctrl+O`: cycle filter forward\n- `Shift+Ctrl+O`: cycle filter backward\n- `Alt+D/T/U/L/A`: jump directly to specific filter mode\n\n## Filters and search semantics\n\nFilter modes (`TreeList`):\n\n1. `default`\n2. `no-tools`\n3. `user-only`\n4. `labeled-only`\n5. `all`\n\n### `default`\n\nShows conversational nodes plus any entry types not explicitly suppressed. It hides these setting/bookkeeping entry types:\n\n- `label`\n- `custom`\n- `model_change`\n- `thinking_level_change`\n\nOther internal entry types that are not rendered specially may appear as blank rows in current code.\n\n### `no-tools`\n\nSame as `default`, plus hides `toolResult` messages.\n\n### `user-only`\n\nOnly `message` entries where role is `user`.\n\n### `labeled-only`\n\nOnly entries that currently resolve to a label.\n\n### `all`\n\nEverything in the session tree, including bookkeeping/custom entries.\n\n### Tool-only assistant node behavior\n\nAssistant messages that contain **only tool calls** (no text) are hidden by default in all filtered views unless:\n\n- message is error/aborted (`stopReason` not `stop`/`toolUse`), or\n- it is the current leaf (always kept visible)\n\n### Search behavior\n\n- Query is tokenized by spaces\n- Matching is case-insensitive\n- All tokens must match (AND semantics)\n- Searchable text includes label, role, and type-specific content (message text, branch summary text, custom type, tool command snippets, etc.)\n\n## Selection outcomes (important)\n\n`navigateTree` computes new leaf behavior from selected entry type:\n\n### Selecting `user` message\n\n- New leaf becomes selected entry’s `parentId`\n- If parent is `null` (root user message), leaf resets to root (`resetLeaf()`)\n- Selected message text is copied to editor for editing/resubmit\n\n### Selecting `custom_message`\n\n- Same leaf rule as user messages (`parentId`)\n- Text content is extracted and copied to editor\n\n### Selecting non-user node (assistant/tool/summary/compaction/custom bookkeeping/etc.)\n\n- New leaf becomes selected node id\n- Editor is not prefilled\n\n### Selecting current leaf\n\n- No-op; selector closes with “Already at this point”\n\n```text\nSelection decision (simplified):\n\nselected node\n   │\n   ├─ is current leaf? ── yes ──> close selector (no-op)\n   │\n   ├─ is user/custom_message? ── yes ──> leaf := parentId (or resetLeaf for root)\n   │                                     + prefill editor text\n   │\n   └─ otherwise ──> leaf := selected node id\n                    + no editor prefill\n```\n\n## Summary-on-switch flow\n\nSummary prompt is controlled by `branchSummary.enabled` (default: `false`).\n\nWhen enabled, after picking a node the UI asks:\n\n- `No summary`\n- `Summarize`\n- `Summarize with custom prompt`\n\nFlow details:\n\n- Escape in summary prompt reopens tree selector\n- Custom prompt cancellation returns to summary choice loop\n- During summarization, UI shows loader and binds `Esc` to `abortBranchSummary()`\n- If summarization aborts, tree selector reopens and no move is applied\n\n`navigateTree` internals:\n\n- Collects abandoned-branch entries from old leaf to common ancestor\n- Emits `session_before_tree` (extensions can cancel or inject summary)\n- Uses default summarizer only if requested and needed\n- Applies move with:\n  - `branchWithSummary(...)` when summary exists\n  - `branch(newLeafId)` for non-root move without summary\n  - `resetLeaf()` for root move without summary\n- Replaces agent conversation with rebuilt session context\n- Emits `session_tree`\n\nNote: if user requests summary but there is nothing to summarize, navigation proceeds without creating a summary entry.\n\n## Labels\n\nLabel edits in tree UI call `appendLabelChange(targetId, label)`.\n\n- non-empty label sets/updates resolved label\n- empty label clears it\n- labels are stored as append-only `label` entries\n- tree nodes display resolved label state, not raw label-entry history\n\n## `/tree` vs adjacent operations\n\n| Operation | Scope                                            | Result                                                                                                                                                   |\n| --------- | ------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| `/tree`   | Current session file                             | Moves leaf to selected point (same file)                                                                                                                 |\n| `/branch` | Usually current session file -> new session file | By default branches from selected **user** message into a new session file; if `doubleEscapeAction = \"tree\"`, `/branch` opens tree navigation UI instead |\n| `/fork`   | Whole current session                            | Duplicates session into a new persisted session file                                                                                                     |\n| `/resume` | Session list                                     | Switches to another session file                                                                                                                         |\n\nKey distinction: `/tree` is a navigation/repositioning tool inside one session file. `/branch`, `/fork`, and `/resume` all change session-file context.\n\n## Operator workflows\n\n### Re-run from an earlier user prompt without losing current branch\n\n1. `/tree`\n2. search/select earlier user message\n3. choose `No summary` (or summarize if needed)\n4. edit prefilled text in editor\n5. submit\n\nEffect: new branch grows from selected point within same session file.\n\n### Leave current branch with context breadcrumb\n\n1. enable `branchSummary.enabled`\n2. `/tree` and select target node\n3. choose `Summarize` (or custom prompt)\n\nEffect: a `branch_summary` entry is appended at the target position before continuing.\n\n### Investigate hidden bookkeeping entries\n\n1. `/tree`\n2. press `Alt+A` (all)\n3. search for `model`, `thinking`, `custom`, or labels\n\nEffect: inspect full internal timeline, not just conversational nodes.\n\n### Bookmark pivot points for later jumps\n\n1. `/tree`\n2. move to entry\n3. `Shift+L` and set label\n4. later use `Alt+L` (`labeled-only`) to jump quickly\n\nEffect: fast navigation among durable branch landmarks.\n",
 	"ttsr-injection-lifecycle.md": "# TTSR Injection Lifecycle\n\nThis document covers the current Time Traveling Stream Rules (TTSR) runtime path from rule discovery to stream interruption, retry injection, extension notifications, and session-state handling.\n\n## Implementation files\n\n- [`../src/sdk.ts`](../packages/coding-agent/src/sdk.ts)\n- [`../src/export/ttsr.ts`](../packages/coding-agent/src/export/ttsr.ts)\n- [`../src/session/agent-session.ts`](../packages/coding-agent/src/session/agent-session.ts)\n- [`../src/session/session-manager.ts`](../packages/coding-agent/src/session/session-manager.ts)\n- [`../src/prompts/system/ttsr-interrupt.md`](../packages/coding-agent/src/prompts/system/ttsr-interrupt.md)\n- [`../src/capability/index.ts`](../packages/coding-agent/src/capability/index.ts)\n- [`../src/extensibility/extensions/types.ts`](../packages/coding-agent/src/extensibility/extensions/types.ts)\n- [`../src/extensibility/hooks/types.ts`](../packages/coding-agent/src/extensibility/hooks/types.ts)\n- [`../src/extensibility/custom-tools/types.ts`](../packages/coding-agent/src/extensibility/custom-tools/types.ts)\n- [`../src/modes/controllers/event-controller.ts`](../packages/coding-agent/src/modes/controllers/event-controller.ts)\n\n## 1. Discovery feed and rule registration\n\nAt session creation, `createAgentSession()` loads discovered rules, constructs a `TtsrManager`, and buckets rules through `bucketRules(...)`:\n\n```ts\nconst ttsrSettings = settings.getGroup(\"ttsr\");\nconst ttsrManager = new TtsrManager(ttsrSettings);\nconst rulesResult = await loadCapability<Rule>(ruleCapability.id, { cwd });\nconst { rulebookRules, alwaysApplyRules } = bucketRules(\n  rulesResult.items,\n  ttsrManager,\n  {\n    builtinRules: ttsrSettings.builtinRules,\n    disabledRules: ttsrSettings.disabledRules,\n  },\n);\n```\n\n`bucketRules(...)` drops names listed in `ttsr.disabledRules`, drops embedded `builtin-defaults` rules when `ttsr.builtinRules === false`, registers accepted TTSR rules, and then routes the remaining rules to always-apply/rulebook buckets.\n\n### Pre-registration dedupe behavior\n\n`loadCapability(\"rules\")` deduplicates by `rule.name` with first-wins semantics (higher provider priority first). Shadowed duplicates are removed before TTSR registration.\n\n### `TtsrManager.addRule()` behavior\n\nRegistration is skipped when:\n\n- TTSR is disabled (`ttsr.enabled === false`)\n- both `rule.condition` (regex) and `rule.astCondition` (ast-grep patterns) are absent, or every regex condition fails to compile and there are no AST conditions\n- a rule with the same `rule.name` was already registered in this manager\n- the rule scope excludes all monitored streams\n\nInvalid regex conditions and unreachable scopes are logged as warnings and ignored; session startup continues. If a TTSR rule defines `globs`, those globs are compiled as a global file-path gate for matching.\n\n### AST conditions (`astCondition`)\n\nA rule may carry `astCondition`: a list of [ast-grep](https://ast-grep.github.io/) patterns (OR'd, same as regex `condition`), matched structurally instead of textually. A repeated metavariable inside one pattern requires both occurrences to be equal (`if ($X) clearTimeout($X)` matches but `if ($X) clearTimeout($Y)` does not).\n\nAST conditions only evaluate on **edit/write tool-argument streams** — they need a language, which is inferred from the file extension on the tool's path argument, and they match against the tool's reconstructed source snapshot (`matcherDigest`), not the raw wire delta. Matching is performed in memory by the native `astMatch` engine (no temp files) with Smart strictness. Streams without a usable file path (prose, thinking, path-less tool calls) skip AST conditions entirely. A rule may mix `condition` and `astCondition`; the regex paths keep working on every scope while AST paths apply only to those tool streams.\n\n### Setting gating\n\n`TtsrSettings.enabled` gates the manager: when `ttsr.enabled === false`, `addRule()` refuses registration and `checkDelta()`/`checkSnapshot()`/`checkAstSnapshot()`/`hasRules()`/`hasAstRules()` all return empty/false, so no matching runs.\n\n## 2. Streaming monitor lifecycle\n\nTTSR detection runs inside `AgentSession.#handleAgentEvent`.\n\n### Turn start\n\nOn `turn_start`, the stream buffer is reset:\n\n- `ttsrManager.resetBuffer()`\n\n### During stream (`message_update`)\n\nWhen assistant updates arrive and rules exist:\n\n- monitor `text_delta`, `thinking_delta`, and `toolcall_delta`\n- for tools exposing `matcherDigest` (edit/write), replace the scoped buffer with the reconstructed source snapshot and call `checkSnapshot(snapshot, matchContext)`; otherwise append the delta into a source/tool scoped manager buffer and call `checkDelta(delta, matchContext)` (synchronous regex matching either way)\n- for edit/write tool streams, when `hasAstRules()` is true, `await checkAstSnapshot(snapshot, matchContext)` (asynchronous AST matching)\n\n`checkDelta()`/`checkSnapshot()` iterate registered rules and return all matching rules that pass scope, global path-glob, regex condition, and repeat policy checks. `checkAstSnapshot()` applies the same scope/path/repeat gates, then runs each candidate rule's `astCondition` patterns against the snapshot via the native `astMatch` engine. It is throttled per stream key: an identical consecutive snapshot (common when only non-source arguments change between deltas) is skipped without re-running the matcher. Both paths feed their matches through the same trigger-decision handler.\n\n## 3. Trigger decision and immediate abort path\n\nWhen one or more rules match and at least one matched rule allows interruption:\n\n1. Matched rules are deduplicated into `#pendingTtsrInjections`.\n2. `#ttsrAbortPending = true` and a TTSR resume gate is created.\n3. `agent.abort()` is called immediately.\n4. `ttsr_triggered` event is emitted asynchronously (fire-and-forget).\n5. retry work is scheduled via the post-prompt task scheduler with a 50ms delay.\n\nAbort is not blocked on extension callbacks.\n\n## 4. Retry scheduling, context mode, and reminder injection\n\nAfter the 50ms timeout:\n\n1. `#ttsrAbortPending = false`\n2. read `ttsrManager.getSettings().contextMode`\n3. if `contextMode === \"discard\"`, drop the targeted partial assistant output with `agent.replaceMessages(...slice(0, targetAssistantIndex))`\n4. build injection content from pending rules using `ttsr-interrupt.md` template\n5. append and persist a hidden `custom_message`/runtime custom message with `customType: \"ttsr-injection\"` and `details.rules`\n6. mark those rule names injected, persist a `ttsr_injection` entry, and call `agent.continue()` to retry generation\n\nTemplate payload is:\n\n```xml\n<system-interrupt reason=\"rule_violation\" rule=\"{{name}}\" path=\"{{path}}\">\n...\n{{content}}\n</system-interrupt>\n```\n\nPending injections are cleared after content generation.\n\n### `contextMode` behavior on partial output\n\n- `discard`: partial/aborted assistant message is removed before retry.\n- `keep`: partial assistant output remains in conversation state; reminder is appended after it.\n\n### Non-interrupting matches\n\nNon-interrupting matches split by `matchContext.source`:\n\n- **`source === \"tool\"` (tool-source match).** The rule is bucketed into `#perToolTtsrInjections`, keyed by the matched tool call's `id`. There is **no** deferred follow-up turn and the stream is not aborted. When the tool actually produces a result, the `afterToolCall` hook prepends a rendered `ttsr-tool-reminder.md` block to `ctx.result.content` (a single `text` block inserted ahead of the tool's own content), and persists a `ttsr_injection` entry with the consumed rule names. The template payload is:\n\n  ```xml\n  <system-reminder reason=\"rule_violation\" rule=\"{{name}}\" path=\"{{path}}\">\n  ...\n  {{content}}\n  </system-reminder>\n  ```\n\n- **`source === \"text\"` / `\"thinking\"` (prose-source match).** Behavior is unchanged: the rule is queued in `#pendingTtsrInjections` and, after a successful non-error, non-aborted assistant message, `AgentSession` injects the hidden `ttsr-injection` custom message as a follow-up and schedules continuation.\n\nWithin a single matching batch, each rule is attached to exactly one sibling tool call — if multiple sibling tool calls would satisfy the same rule, deduplication picks one and the others are left untouched. Multiple distinct rules can still fold onto the same tool call.\n\n#### Implications for tool authors and transcript readers\n\n- The tool's own `toolResult` content is preserved verbatim; the reminder is **prepended** as an additional leading text block. Renderers that assume `content[0]` is the tool's primary output must scan past any block whose text begins with `<system-reminder reason=\"rule_violation\"` (or filter on the wrapper tag) to find the real payload.\n- The reminder is in-band on the tool result, not a separate `custom_message`/`ttsr-injection` entry. Transcript readers looking for non-interrupting TTSR activity on tool-source rules MUST inspect tool results (and the persisted `ttsr_injection` entry list), not just synthetic injection entries.\n- A single tool result may carry reminders for several rules concatenated with a blank line between rendered templates.\n- If the assistant message ends with `stopReason === \"aborted\"` or `\"error\"` before the matched tools run, the pending per-tool buckets are cleared — those rules are **not** persisted as injected and remain eligible to re-trigger on a future turn (subject to repeat policy).\n\n## 5. Repeat policy and gap logic\n\n`TtsrManager` tracks `#messageCount` and per-rule `lastInjectedAt`.\n\n### `repeatMode: \"once\"`\n\nA rule can trigger only once after it has an injection record.\n\n### `repeatMode: \"after-gap\"`\n\nA rule can re-trigger only when:\n\n- `messageCount - lastInjectedAt >= repeatGap`\n\n`messageCount` increments on `turn_end`, so gap is measured in completed turns, not stream chunks.\n\n## 6. Event emission and extension/hook surfaces\n\n### Session event\n\n`AgentSessionEvent` includes:\n\n```ts\n{ type: \"ttsr_triggered\"; rules: Rule[] }\n```\n\n### Extension runner\n\n`#emitSessionEvent()` routes the event to:\n\n- extension listeners (`ExtensionRunner.emit({ type: \"ttsr_triggered\", rules })`)\n- local session subscribers\n\n### Hook and custom-tool typing\n\n- extension API exposes `on(\"ttsr_triggered\", ...)`\n- hook API exposes `on(\"ttsr_triggered\", ...)`\n- custom tools receive `onSession({ reason: \"ttsr_triggered\", rules })`\n\n### Interactive-mode rendering difference\n\nInteractive mode uses `session.isTtsrAbortPending` to suppress showing the aborted assistant stop reason as a visible failure during TTSR interruption, and renders a `TtsrNotificationComponent` when the event arrives.\n\n## 7. Persistence and resume state (current implementation)\n\n`SessionManager` persists injected-rule state:\n\n- entry type: `ttsr_injection`\n- append API: `appendTtsrInjection(ruleNames)`\n- query API: `getInjectedTtsrRules()`\n- context reconstruction includes `SessionContext.injectedTtsrRules`\n\n`TtsrManager` supports restoration via `restoreInjected(ruleNames)`.\n\n### Current wiring status\n\nIn the current runtime path:\n\n- interrupted injections append a hidden `custom_message` with `customType: \"ttsr-injection\"` and append a `ttsr_injection` entry via `appendTtsrInjection(...)`\n- deferred non-interrupting prose-source injections are marked/persisted when their queued custom message reaches `message_end`\n- non-interrupting tool-source injections are marked at match time and persisted via `appendTtsrInjection(...)` from the `afterToolCall` hook when the matched tool's result is produced\n- `createAgentSession()` restores `existingSession.injectedTtsrRules` into `ttsrManager`\n\nNet effect: injected-rule suppression is persisted/restored across session reload/resume for the current branch path.\n\n## 8. Race boundaries and ordering guarantees\n\n### Abort vs retry callback\n\n- abort is synchronous from TTSR handler perspective (`agent.abort()` called immediately)\n- retry is deferred by timer (`50ms`)\n- extension notification is asynchronous and intentionally not awaited before abort/retry scheduling\n\n### Multiple matches in same stream window\n\n`checkDelta()` returns all currently matching eligible rules for that scoped buffer. Pending injections are deduplicated by rule name before injection.\n\n### Between abort and continue\n\nDuring the timer window, state can change (user interruption, mode actions, additional events). The retry call is best-effort: `agent.continue()` is awaited in a try/catch; on failure the error is swallowed and the TTSR resume gate is resolved.\n\n## 9. Edge cases summary\n\n- Invalid `condition` regex: skipped with warning; other conditions/rules continue.\n- Duplicate rule names at capability layer: lower-priority duplicates are shadowed before registration.\n- Duplicate names at manager layer: second registration is ignored.\n- `ttsr.disabledRules`: listed names are dropped before TTSR registration and are not surfaced through always-apply/rulebook buckets.\n- `ttsr.builtinRules: false`: embedded `builtin-defaults` rules are dropped before TTSR registration; user/project rules still load.\n- `globs` on a TTSR rule require the stream match context to include at least one matching file path.\n- `contextMode: \"keep\"`: partial violating output can remain in context before reminder retry.\n- `interruptMode: \"never\"`: prose-source matches queue a deferred hidden injection after a successful assistant message; tool-source matches fold an in-band `<system-reminder>` into the matched tool call's `toolResult` content via the `afterToolCall` hook (no mid-stream abort, no separate follow-up turn).\n- Tool-source non-interrupting buckets are cleared when the parent assistant message ends with `stopReason === \"aborted\"` or `\"error\"`, so rules whose target tool never produced a result remain eligible to re-trigger.\n- Repeat-after-gap depends on turn count increments at `turn_end`; mid-turn chunks do not advance gap counters.\n",
-	"tui-core-renderer.md": "# TUI core renderer — the append-only contract\n\nWhat you are dealing with before you touch the rendering engine. This is the\ncompanion to [`tui-runtime-internals.md`](./tui-runtime-internals.md): that doc\nmaps the *flow* (input → component tree → render); this doc explains the\n**render contract, why it is shaped this way, and the invariants you must not\nviolate**. Scope is the core engine only:\n\n- [`packages/tui/src/tui.ts`](../packages/tui/src/tui.ts) — frame pipeline, commit ledger, window math, emitters, cursor placement.\n- [`packages/tui/src/terminal.ts`](../packages/tui/src/terminal.ts) — `ProcessTerminal`, capability probes, private-CSI reassembly.\n- [`packages/tui/src/terminal-capabilities.ts`](../packages/tui/src/terminal-capabilities.ts) — `TERMINAL` profile, sync-output / DECCARA / image detection.\n- [`packages/tui/src/stdin-buffer.ts`](../packages/tui/src/stdin-buffer.ts) — escape-sequence reassembly.\n- [`packages/tui/src/utils.ts`](../packages/tui/src/utils.ts) — width/slice/wrap (the width model).\n- [`packages/tui/src/kitty-graphics.ts`](../packages/tui/src/kitty-graphics.ts) + [`components/image.ts`](../packages/tui/src/components/image.ts) — inline images.\n- [`packages/tui/src/deccara.ts`](../packages/tui/src/deccara.ts) — rectangular-fill optimizer.\n\nApplication-layer renderers (transcript, tool calls, session tree, editor,\nwidgets) are **out of scope** — they live in `packages/coding-agent`. The one\napp-layer file that is load-bearing for this contract is\n[`transcript-container.ts`](../packages/coding-agent/src/modes/components/transcript-container.ts),\nwhich implements the commit-boundary seam described below.\n\n---\n\n## 1. The one thing to understand first\n\n> **The renderer cannot observe the terminal's scroll position** (ConPTY's\n> probe lies; POSIX has no API at all). The previous engine tried to *guess*\n> when it was safe to rewrite native scrollback, and every policy choice over\n> that unobservable variable traded one failure family for another (yank ↔\n> flash ↔ corruption ↔ invisible-until-resize — see the git history of this\n> file for the full war journal). The current engine removes the guess\n> entirely: **native scrollback is append-only.**\n\nWe keep the transcript on the **normal screen** (native scrollback, native\nselection, transcript persists after exit). The engine maintains one ledger:\n\n- **`committedRows` (C)** — frame rows `[0, C)` have been physically scrolled\n  into terminal history. They are **immutable**: the engine never rewrites\n  them, and components must never change them.\n- **`windowTopRow` (W)** — the frame row mapped to grid row 0. The visible\n  window is frame rows `[W, W + height)`, repainted in place with relative\n  cursor moves.\n- **commit boundary (B)** — reported by the component tree per frame\n  (`NativeScrollbackLiveRegion`): `B = commitSafeEnd ?? liveRegionStart ??\n  frame.length`. Rows below B may still re-layout and must not enter history.\n\nPer ordinary frame: `W = max(C, L − height)`, `C' = max(C, min(B, W))`, and the\nonly bytes that ever touch history are the **chunk** `frame[C, C')` written at\nthe scrollback seam. Scrollback therefore equals `frame[0..C)` — every row\nexactly once, in order, with its content at commit time. There is nothing to\nguess, nothing to defer, and nothing to reconcile: the scroll position is\nirrelevant because ordinary updates never rewrite anything a scrolled reader\ncould be looking at.\n\n### What this costs (the accepted tradeoffs)\n\n- A block that has scrolled past the window top cannot reflow in place. Blocks\n  stay in the live region (below B) until they are final; a late mutation of\n  committed content is ignored (the stale committed copy stays in history).\n- A component tree that reports **no seam** gets shell semantics: whatever\n  scrolls off is final. Shrinking such a frame into its committed prefix\n  re-anchors the window and leaves the stale copy in history (§3).\n- Inside multiplexers, a resize leaves the pane history wrapped at the old\n  width (same as any shell output).\n\n---\n\n## 2. The frame pipeline (what you are editing)\n\n`#doRender` per frame:\n\n1. Compose the frame (`render(width)`), collecting `liveRegionStart` /\n   `commitSafeEnd` from the root children (absolute row indices).\n2. **Audit the committed prefix** (`findCommittedPrefixResync`, skipped on\n   geometry frames). Components must never re-layout rows below C, but real\n   flows violate it (a TTSR rewind truncating a streamed block, an image-cap\n   demotion shrinking a committed image) and the violation must not become\n   content loss. The detector samples the prefix *tail* (up to 8 non-blank\n   rows in the last 24, SGR-stripped): an in-place edit or restyle disturbs\n   only the touched rows (≤1 mismatch ⇒ aligned ⇒ ignored — stale styling in\n   history is the accepted artifact), while any insertion/deletion shifts\n   every row below it including the tail (⇒ re-anchor C at the first changed\n   row and recommit from there: history keeps the stale copy and gains a\n   fresh one — **duplication, never loss**).\n3. Classify: **fullPaint** (first paint, `clearScrollback` session replace, or\n   geometry change outside a multiplexer — all user gestures) or **update**.\n4. Window math as in §1. Two special rules:\n   - **Overlays freeze commits** (`C' = C`): composited rows must never enter\n     history; the hidden gap backfills via the chunk after the overlay closes.\n   - **Shrink into the committed prefix** (`L ≤ C`): re-anchor\n     `W = max(0, L − height)`, reset `C = min(B, W)`, keep the stale history\n     above (no gesture, no erase).\n5. Extract the cursor marker (strip-first: markers never reach the terminal,\n   the prefix ledger, or the audit), prepare lines (width fitting), slice the\n   window, composite overlays **into the window slice only** (screen\n   coordinates — an overlay never touches the frame or the ledger).\n6. Emit:\n\n| Emitter | Bytes | When |\n|---|---|---|\n| `#emitFullPaint` | clears + `frame[0, C')` + window rows | gestures only. `clearScrollback` ⇒ `\\x1b[2J\\x1b[H\\x1b[3J`; otherwise ED22 (when supported) + `\\x1b[2J\\x1b[H` |\n| `#emitUpdate` scroll-append | `\\r\\n` + new bottom rows + changed-row range | the rows leaving the screen are exactly the chunk, content untouched since painted |\n| `#emitUpdate` in-window diff | relative move + changed-row range rewrite | nothing scrolls, nothing commits (cursor-only when nothing changed) |\n| `#emitUpdate` seam rewrite | chunk rows + full window rewrite | commit advance, window re-anchor, hidden-gap backfill, mux resize |\n\n**ED3 (`CSI 3 J`) is emitted in exactly one place** — `#emitFullPaint` with\n`clearScrollback: true` — and is reached only by user gestures: session\nreplace/branch/resume (`requestRender(true, { clearScrollback: true })`),\nresize outside a multiplexer, `resetDisplay()` (Ctrl+L). A gesture pins the\nuser to the tail, so the snap is acceptable; multiplexers never get ED3 (it is\na no-op there and a replay would duplicate pane history).\n\nThe ordinary update path never emits ED2/ED3 or an absolute cursor home —\nseveral terminal families snap a scrolled reader to the bottom on those.\n\n### The commit-boundary seam (the load-bearing app contract)\n\n`NativeScrollbackLiveRegion` (tui.ts) is how a component keeps mutable rows out\nof history:\n\n- `getNativeScrollbackLiveRegionStart()` — first row that may still mutate\n  (everything below it, including root chrome rendered after it, stays in the\n  window).\n- `getNativeScrollbackCommitSafeEnd()` — optional deeper boundary: the\n  append-only prefix of the live region (a streaming assistant message's\n  settled rows). Without it, a single live block taller than the window would\n  hold its head out of history until it finalizes.\n\n`TranscriptContainer` implements this for the coding agent: finalized blocks\nfreeze (their render is snapshotted, so their content can never drift after\nthe engine may have committed it), still-mutating blocks\n(`isTranscriptBlockFinalized?.() === false`) anchor the live region, and\n`deriveLiveCommitState` derives the commit-safe end of the first live block\nfrom two independent signals:\n\n- **append-only detection** — a block observed growing without visibly\n  rewriting an interior row commits its full body; a rewrite suspends this\n  for `VOLATILE_REARM_FRAMES` clean frames.\n- **stable-prefix ratchet** — rows that stayed visibly identical for a full\n  `STABLE_PREFIX_COMMIT_FRAMES` window commit even while the block's tail\n  keeps rewriting (a task tool's static prompt above a ticking progress\n  tree). Without it, one perpetually animating row holds the whole block out\n  of history, so a block taller than the window reads as cut off (head\n  neither committed nor on screen) for the entire run. The ratchet tracks the\n  window-minimum common prefix; a rewrite above the promoted run retreats it\n  to the divergence, and rows that already committed are the engine audit's\n  problem (recommit → duplication, never loss). That retreat also arms a\n  permanent **rewrite floor** at the divergence: a row that mutates *after*\n  surviving a full promotion window is a slow ticker (an agent row's tool/cost\n  counter updating every few seconds), not settling content — without the\n  floor, every quiet stretch re-promoted it and every later tick forced an\n  audit recommit, spraying stale snapshots of the block into scrollback for\n  the whole run. Rows at/after the floor never re-promote while the block\n  lives (the floor index travels with append-shaped insertions above it);\n  one-off re-layouts before any promotion never arm it, and the append-only\n  path commits the full block regardless.\n\nFreezing is unconditional — it is the engine's required guarantee, not a\nper-terminal optimization.\n\n---\n\n## 3. Invariants — MUST / NEVER\n\n1. **NEVER add a new `CSI 3 J` (ED3) callsite.** ED3 flows only through\n   `#emitFullPaint({ clearScrollback: true })`, only for gestures, never inside\n   multiplexers.\n2. **NEVER rewrite a committed row.** No emitter may touch frame rows `< C`,\n   and `W ≥ C` always (re-showing a committed row on the grid duplicates it\n   for a scrolling reader — the historical corruption family). When a\n   *component* violates immutability, the audit (§2) degrades to duplication —\n   never silently skip rows, never erase history.\n3. **Commits are exactly the chunk.** Any byte shape that scrolls the screen\n   must scroll *only* rows accounted for by `C' − C` — that is what makes\n   scrollback provably `frame[0..C)`.\n4. **NEVER probe the viewport position or fork on platform in the update\n   path.** win32 behaves like POSIX. The probe APIs are gone; do not\n   reintroduce them.\n5. **Mutable content stays below the commit boundary.** App-layer renderers\n   must finalize-before-commit; the engine trusts B and clamps, it does not\n   verify content.\n6. **Park the hardware cursor at real content bottom**, not the padded window\n   bottom, or height shrinks scroll live rows into history and duplicate them\n   per resize step.\n7. **Cursor writes live inside the synchronized-output frame**, before ESU —\n   never as a second frame after it.\n8. **NEVER throw in the render hot path.** Clamp over-wide lines\n   (`truncateToWidth`); a width mismatch is cosmetic, not fatal.\n9. **Multiplexers get no destructive clear and no history rewrap on resize** —\n   repaint the window in place; pane history keeps its old wrap.\n10. **Any change to the ledger math, the emitters, or the seam must be\n    validated by the stress harness (§6)** across its full scenario matrix,\n    not by a single-terminal smoke test.\n\n---\n\n## 4. Terminal capability detection\n\n`TERMINAL` (`terminal-capabilities.ts`) is resolved once at import from\n`TERMINAL_ID` plus environment sniffing; detection helpers are pure over\n`(env, platform)` and unit-testable.\n\n- `shouldEnableSynchronizedOutputByDefault(env, id)` → DEC 2026 default.\n  Precedence: user opt-out (`PI_NO_SYNC_OUTPUT`/`PI_TUI_SYNC_OUTPUT=0`) → user\n  force-on (`PI_FORCE_SYNC_OUTPUT=1`/`PI_TUI_SYNC_OUTPUT=1`) → `TERM_FEATURES`\n  advertises `Sy` → `WT_SESSION` → known direct terminals → off for risky\n  multiplexers and unknowns. Reconciled at runtime by the DECRQM mode-2026\n  report; a user override still wins.\n- `detectRectangularSgrSupport(id, env)` → DECCARA fills: kitty only, off in\n  multiplexers and under `PI_NO_DECCARA`.\n- `supportsScreenToScrollback` → kitty's ED22 (used once, on the initial\n  paint, to preserve the pre-existing shell screen).\n\nThe old ED3-risk classifier (`eagerEraseScrollbackRisk`, `PI_TUI_ED3_SAFE`,\n`submitPinsViewportToTail`) is gone: behavior no longer depends on which\nterminal is rendering, so there is no risk class to detect. Env sniffing now\nonly selects *optimizations* (sync output, DECCARA, images), where a miss is\ncosmetic, not corrupting.\n\n---\n\n## 5. Width model\n\n`visibleWidth` / `truncateToWidth` / `sliceByColumn` / `wrapTextWithAnsi`\n(`utils.ts`) all route through **one native UAX#11 engine**\n(`@oh-my-pi/pi-natives`, Rust `unicode-width`). `Bun.stringWidth` was dropped\ndeliberately — mixing two width models in measure-vs-slice produced crashes.\n\n- Fast path: printable ASCII is one cell per code unit.\n- ZWJ pictographic emoji take the `visibleWidthByGrapheme` override.\n- OSC 66 sized text takes the native path.\n\n**Rule:** any new measuring code routes through these helpers, and the hot\npath clamps instead of throwing. Known residual: combining-heavy scripts\n(Arabic harakat) survive painting verbatim, but ghostty-web's cell readback can\nmigrate non-spacing marks across cells — the stress harness compares those rows\nwith marks stripped (`sameLinesAllowingMarkDrift`).\n\n---\n\n## 6. The fidelity gate (use it)\n\n`packages/tui/test/render-stress-harness.ts` drives the renderer's **real\nemitted ANSI** into a ghostty-web `VirtualTerminal` across randomized op\nsequences and parameterized terminal shapes, and validates the contract with a\n**shadow commit ledger**: an independent reimplementation of §1's math, fed\nonly by observed frames (a `render` wrap) and observed bytes (a `write` wrap).\nPer op it asserts:\n\n- the whole tape (scrollback + grid) equals `shadowTape + window slice`, row\n  for row, including across resizes;\n- scrolled readers stay pinned and visible history rows are never rewritten;\n- multiplexer pane history grows by exactly the committed chunk;\n- sync-output/autowrap bracket discipline, cursor parking, background columns,\n  duplicate accounting.\n\nRun it — plus `render-regressions.test.ts`,\n`streaming-scrollback-defer.test.ts`, and the `issue-*-repro.test.ts` files —\nbefore changing ledger math, emitters, or the seam. A change that passes one\nterminal and one seed is not verified.\n\n---\n\n## 7. Capability probes & stdin reassembly\n\n`ProcessTerminal` fuses capability queries with a bare DA1 (`CSI c`) sentinel so\na non-answering terminal is detected when DA1 returns first. Replies can arrive\n**split across a stdin flush**, so:\n\n- `#privateCsiResponseBuffer` accumulates `\\x1b[?…` partials while a sentinel is\n  outstanding, rejoins on the terminator byte, then runs the handlers on the\n  **complete** reply. A new `\\x1b` mid-reassembly or >256 bytes abandons the\n  partial so real keys still reach input.\n- `#da1SentinelOwners` is a **typed FIFO** discriminated by `kind` so a\n  keyboard DA1 cannot be mistaken for an OSC 11 / DECRQM / graphics-probe\n  sentinel.\n- DECRQM probes (2026/2048/2031) drive runtime feature gating.\n\n**Rule:** any new probe must own a typed sentinel and survive a split reply\n(feed the reply byte-by-byte in a test and assert nothing leaks to input).\n\n---\n\n## 8. Inline images & memory\n\nKitty images are **transmit-once, place-many** (`kitty-graphics.ts`).\n`ImageBudget` keeps only the most-recent N images live; when the cap is\nexceeded the demoted image's pixels are deleted by id (`a=d,d=I`) and its\nvisible rows re-render as the text fallback through the ordinary window diff —\n**no destructive replay**. A demoted placement already committed to history\nsimply loses its pixels (committed rows are immutable), and the text fallback\nis **height-preserving** once a graphic has rendered (reserved rows + fallback\nline), so demotion never shrinks the block and never shifts committed content\nbelow it.\n\n**Rule:** never re-emit full base64 per frame. Kitty Unicode placeholders are\ndefault-on only for kitty/ghostty (`PI_NO_KITTY_PLACEHOLDERS` /\n`PI_KITTY_PLACEHOLDERS`).\n\n---\n\n## 9. Escape hatches (env vars)\n\n| Var | Effect |\n|---|---|\n| `PI_NO_SYNC_OUTPUT=1` | Disable DEC 2026 BSU/ESU wrappers (autowrap discipline stays on). |\n| `PI_TUI_SYNC_OUTPUT=0\\|1` / `PI_FORCE_SYNC_OUTPUT=1` | Force sync output off / on. |\n| `PI_NO_DECCARA` | Disable Kitty DECCARA rectangular-fill optimization. |\n| `PI_FORCE_IMAGE_PROTOCOL=kitty\\|iterm2\\|sixel\\|off` | Override image protocol detection. |\n| `PI_NO_KITTY_PLACEHOLDERS=1` / `PI_KITTY_PLACEHOLDERS=1` | Force Kitty Unicode placeholders off / on. |\n| `PI_HARDWARE_CURSOR=1` | Show the real hardware cursor instead of a rendered one. |\n| `PI_NOTIFICATIONS=off\\|0\\|false` | Suppress terminal notifications. |\n| `PI_DEBUG_REDRAW=1` | Log the chosen render intent + ledger state per frame to the debug log. |\n\nRemoved with the old engine: `PI_TUI_ED3_SAFE` (no ED3-risk lever exists),\n`PI_CLEAR_ON_SHRINK` (shrinks always clear exactly), `PI_TUI_DEBUG` (per-render\ndump superseded by `PI_DEBUG_REDRAW` ledger logging and the stress harness\nreplay/reduce tooling).\n\n---\n\n## 10. Before you touch the render core — checklist\n\n- [ ] Are you about to emit `CSI 3 J` anywhere other than the gesture-driven\n      `clearScrollback` full paint? **Stop.**\n- [ ] Could any code path rewrite, or re-show on the grid, a frame row below\n      `committedRows`? **Stop.**\n- [ ] Does your byte shape scroll rows that are not the commit chunk? That\n      breaks `scrollback == frame[0..C)`.\n- [ ] Are you adding a viewport probe, a platform fork, or a terminal-brand\n      branch to the update path? The contract exists so none are needed.\n- [ ] New mutable UI above the editor? It must report (or live inside) the\n      live-region seam, or it will freeze at first commit.\n- [ ] Did you run the stress harness and the repro suite across the full\n      scenario matrix — not just one terminal and one seed?\n- [ ] New probe? Typed sentinel owner + split-reply test.\n- [ ] New width path? Routed through the shared native engine, clamped (never\n      thrown) in the hot path.\n",
+	"tui-core-renderer.md": "# TUI core renderer — the append-only contract\n\nWhat you are dealing with before you touch the rendering engine. This is the\ncompanion to [`tui-runtime-internals.md`](./tui-runtime-internals.md): that doc\nmaps the *flow* (input → component tree → render); this doc explains the\n**render contract, why it is shaped this way, and the invariants you must not\nviolate**. Scope is the core engine only:\n\n- [`packages/tui/src/tui.ts`](../packages/tui/src/tui.ts) — frame pipeline, commit ledger, window math, emitters, cursor placement.\n- [`packages/tui/src/terminal.ts`](../packages/tui/src/terminal.ts) — `ProcessTerminal`, capability probes, private-CSI reassembly.\n- [`packages/tui/src/terminal-capabilities.ts`](../packages/tui/src/terminal-capabilities.ts) — `TERMINAL` profile, sync-output / DECCARA / image detection.\n- [`packages/tui/src/stdin-buffer.ts`](../packages/tui/src/stdin-buffer.ts) — escape-sequence reassembly.\n- [`packages/tui/src/utils.ts`](../packages/tui/src/utils.ts) — width/slice/wrap (the width model).\n- [`packages/tui/src/kitty-graphics.ts`](../packages/tui/src/kitty-graphics.ts) + [`components/image.ts`](../packages/tui/src/components/image.ts) — inline images.\n- [`packages/tui/src/deccara.ts`](../packages/tui/src/deccara.ts) — rectangular-fill optimizer.\n\nApplication-layer renderers (transcript, tool calls, session tree, editor,\nwidgets) are **out of scope** — they live in `packages/coding-agent`. The one\napp-layer file that is load-bearing for this contract is\n[`transcript-container.ts`](../packages/coding-agent/src/modes/components/transcript-container.ts),\nwhich implements the commit-boundary seam described below.\n\n---\n\n## 1. The one thing to understand first\n\n> **The renderer cannot observe the terminal's scroll position** (ConPTY's\n> probe lies; POSIX has no API at all). The previous engine tried to *guess*\n> when it was safe to rewrite native scrollback, and every policy choice over\n> that unobservable variable traded one failure family for another (yank ↔\n> flash ↔ corruption ↔ invisible-until-resize — see the git history of this\n> file for the full war journal). The current engine removes the guess\n> entirely: **native scrollback is append-only.**\n\nWe keep the transcript on the **normal screen** (native scrollback, native\nselection, transcript persists after exit). The engine maintains one ledger:\n\n- **`committedRows` (C)** — frame rows `[0, C)` have been physically scrolled\n  into terminal history. They are **immutable**: the engine never rewrites\n  them, and components must never change them.\n- **`windowTopRow` (W)** — the frame row mapped to grid row 0. The visible\n  window is frame rows `[W, W + height)`, repainted in place with relative\n  cursor moves.\n- **commit boundary** — reported by the component tree per frame\n  (`NativeScrollbackLiveRegion`) as two nested ends:\n  - **byte-stable end (B)** — `commitSafeEnd ?? liveRegionStart ?? frame.length`.\n    Rows below B are asserted never to re-layout and stay under the\n    committed-prefix audit.\n  - **durable end (D)** — `max(B, snapshotSafeEnd ?? B)`. Rows in `[B, D)` may\n    still drift bytes later (a streaming markdown table re-aligning columns) but\n    are *durable* — their current snapshot is permanent content, so dropping them\n    when they scroll off is forbidden. They commit **audit-exempt**: later drift\n    becomes a frozen stale row in history, never a re-anchor.\n\nPer ordinary frame: `W = max(C, L − height)`, `C' = max(C, min(D, W))`, and the\nonly bytes that ever touch history are the **chunk** `frame[C, C')` written at\nthe scrollback seam. The engine also tracks **`auditRows` (A ≤ C)** — the\nbyte-stable leading prefix `[0, A)`; the committed-prefix audit (§2) samples only\nthat prefix, so the durable suffix `[A, C)` drifting never triggers a re-anchor.\nScrollback therefore equals `frame[0..C)` — every row exactly once, in order,\nwith its content at commit time. There is nothing to guess, nothing to defer,\nand nothing to reconcile: the scroll position is irrelevant because ordinary\nupdates never rewrite anything a scrolled reader could be looking at.\n\n### What this costs (the accepted tradeoffs)\n\n- A block that has scrolled past the window top cannot reflow in place. A\n  byte-stable block stays in the live region (below B) until final; a durable\n  block (below D) commits its scroll-off snapshot, so a late layout change of an\n  already-committed row is a frozen stale row in history (duplication never loss),\n  not a dropped row.\n- A component tree that reports **no seam** gets shell semantics: whatever\n  scrolls off is final. Shrinking such a frame into its committed prefix\n  re-anchors the window and leaves the stale copy in history (§3).\n- Inside multiplexers, a resize leaves the pane history wrapped at the old\n  width (same as any shell output).\n\n---\n\n## 2. The frame pipeline (what you are editing)\n\n`#doRender` per frame:\n\n1. Compose the frame (`render(width)`), collecting `liveRegionStart` /\n   `commitSafeEnd` from the root children (absolute row indices).\n2. **Audit the committed prefix** (`findCommittedPrefixResync`, skipped on\n   geometry frames). Components must never re-layout rows below C, but real\n   flows violate it (a TTSR rewind truncating a streamed block, an image-cap\n   demotion shrinking a committed image) and the violation must not become\n   content loss. The detector samples the prefix *tail* (up to 8 non-blank\n   rows in the last 24, SGR-stripped): an in-place edit or restyle disturbs\n   only the touched rows (≤1 mismatch ⇒ aligned ⇒ ignored — stale styling in\n   history is the accepted artifact), while any insertion/deletion shifts\n   every row below it including the tail (⇒ re-anchor C at the first changed\n   row and recommit from there: history keeps the stale copy and gains a\n   fresh one — **duplication, never loss**).\n3. Classify: **fullPaint** (first paint, `clearScrollback` session replace, or\n   geometry change outside a multiplexer — all user gestures) or **update**.\n4. Window math as in §1. Two special rules:\n   - **Overlays freeze commits** (`C' = C`): composited rows must never enter\n     history; the hidden gap backfills via the chunk after the overlay closes.\n   - **Shrink into the committed prefix** (`L ≤ C`): re-anchor\n     `W = max(0, L − height)`, reset `C = min(B, W)`, keep the stale history\n     above (no gesture, no erase).\n5. Extract the cursor marker (strip-first: markers never reach the terminal,\n   the prefix ledger, or the audit), prepare lines (width fitting), slice the\n   window, composite overlays **into the window slice only** (screen\n   coordinates — an overlay never touches the frame or the ledger).\n6. Emit:\n\n| Emitter | Bytes | When |\n|---|---|---|\n| `#emitFullPaint` | clears + `frame[0, C')` + window rows | gestures only. `clearScrollback` ⇒ `\\x1b[2J\\x1b[H\\x1b[3J`; otherwise ED22 (when supported) + `\\x1b[2J\\x1b[H` |\n| `#emitUpdate` scroll-append | `\\r\\n` + new bottom rows + changed-row range | the rows leaving the screen are exactly the chunk, content untouched since painted |\n| `#emitUpdate` in-window diff | relative move + changed-row range rewrite | nothing scrolls, nothing commits (cursor-only when nothing changed) |\n| `#emitUpdate` seam rewrite | chunk rows + full window rewrite | commit advance, window re-anchor, hidden-gap backfill, mux resize |\n\n**ED3 (`CSI 3 J`) is emitted in exactly one place** — `#emitFullPaint` with\n`clearScrollback: true` — and is reached only by user gestures: session\nreplace/branch/resume (`requestRender(true, { clearScrollback: true })`),\nresize outside a multiplexer, `resetDisplay()` (Ctrl+L). A gesture pins the\nuser to the tail, so the snap is acceptable; multiplexers never get ED3 (it is\na no-op there and a replay would duplicate pane history).\n\nThe ordinary update path never emits ED2/ED3 or an absolute cursor home —\nseveral terminal families snap a scrolled reader to the bottom on those.\n\n### The commit-boundary seam (the load-bearing app contract)\n\n`NativeScrollbackLiveRegion` (tui.ts) is how a component keeps mutable rows out\nof history:\n\n- `getNativeScrollbackLiveRegionStart()` — first row that may still mutate\n  (everything below it, including root chrome rendered after it, stays in the\n  window).\n- `getNativeScrollbackCommitSafeEnd()` — optional **byte-stable** deeper boundary\n  (B): the append-only prefix of the live region (a streaming assistant message's\n  settled rows), asserted never to re-layout, so it stays under the audit.\n- `getNativeScrollbackSnapshotSafeEnd()` — optional **durable** deeper boundary\n  (D ≥ B): rows whose current snapshot is permanent but may still drift bytes\n  (a streaming markdown table whose columns keep re-aligning). They commit on\n  scroll-off (never dropped) but **audit-exempt** — drift after commit freezes a\n  stale row in history rather than re-anchoring the audit and spraying duplicate\n  snapshots. Without it, a commit-stable block that perpetually re-lays-out an\n  interior row (a table taller than the window) had no byte-stable prefix past\n  the table head, so its scrolled-off rows were committed nowhere and repainted\n  nowhere — silent content loss as the reply streamed.\n\n`TranscriptContainer` implements this for the coding agent: finalized blocks\nfreeze (their render is snapshotted, so their content can never drift after\nthe engine may have committed it), still-mutating blocks\n(`isTranscriptBlockFinalized?.() === false`) anchor the live region, and\n`deriveLiveCommitState` derives the byte-stable commit-safe end of the first\nlive block from two independent signals:\n\n- **append-only detection** — a block observed growing without visibly\n  rewriting an interior row commits its full body; a rewrite suspends this\n  for `VOLATILE_REARM_FRAMES` clean frames.\n- **stable-prefix ratchet** — rows that stayed visibly identical for a full\n  `STABLE_PREFIX_COMMIT_FRAMES` window commit even while the block's tail\n  keeps rewriting (a task tool's static prompt above a ticking progress\n  tree). Without it, one perpetually animating row holds the whole block out\n  of history, so a block taller than the window reads as cut off (head\n  neither committed nor on screen) for the entire run. The ratchet tracks the\n  window-minimum common prefix; a rewrite above the promoted run retreats it\n  to the divergence, and rows that already committed are the engine audit's\n  problem (recommit → duplication, never loss). That retreat also arms a\n  permanent **rewrite floor** at the divergence: a row that mutates *after*\n  surviving a full promotion window is a slow ticker (an agent row's tool/cost\n  counter updating every few seconds), not settling content — without the\n  floor, every quiet stretch re-promoted it and every later tick forced an\n  audit recommit, spraying stale snapshots of the block into scrollback for\n  the whole run. Rows at/after the floor never re-promote while the block\n  lives (the floor index travels with append-shaped insertions above it);\n  one-off re-layouts before any promotion never arm it, and the append-only\n  path commits the full block regardless.\n\nThe byte-stable end gates audited commits; the **durable snapshot end** is the\nseparate floor that guarantees no loss. `TranscriptContainer` reports the whole\nbody of a still-live **commit-stable** block (`isTranscriptBlockCommitStable?.()\n!== false`) as the snapshot-safe end, so its scrolled-off rows always reach\nhistory even while its interior re-lays-out. Provisional blocks\n(`isTranscriptBlockCommitStable?.() === false`: a collapsing tool/edit preview\nwhose head is a throwaway tail window) report no snapshot-safe end, so their\nhead is correctly dropped rather than stranded as stale history.\n\nFreezing is unconditional — it is the engine's required guarantee, not a\nper-terminal optimization.\n\n---\n\n## 3. Invariants — MUST / NEVER\n\n1. **NEVER add a new `CSI 3 J` (ED3) callsite.** ED3 flows only through\n   `#emitFullPaint({ clearScrollback: true })`, only for gestures, never inside\n   multiplexers.\n2. **NEVER rewrite a committed row.** No emitter may touch frame rows `< C`,\n   and `W ≥ C` always (re-showing a committed row on the grid duplicates it\n   for a scrolling reader — the historical corruption family). When a\n   *component* violates immutability, the audit (§2) degrades to duplication —\n   never silently skip rows, never erase history.\n3. **Commits are exactly the chunk.** Any byte shape that scrolls the screen\n   must scroll *only* rows accounted for by `C' − C` — that is what makes\n   scrollback provably `frame[0..C)`.\n4. **NEVER probe the viewport position or fork on platform in the update\n   path.** win32 behaves like POSIX. The probe APIs are gone; do not\n   reintroduce them.\n5. **Mutable content stays below the commit boundary.** App-layer renderers\n   must finalize-before-commit; the engine trusts B and clamps, it does not\n   verify content.\n6. **Park the hardware cursor at real content bottom**, not the padded window\n   bottom, or height shrinks scroll live rows into history and duplicate them\n   per resize step.\n7. **Cursor writes live inside the synchronized-output frame**, before ESU —\n   never as a second frame after it.\n8. **NEVER throw in the render hot path.** Clamp over-wide lines\n   (`truncateToWidth`); a width mismatch is cosmetic, not fatal.\n9. **Multiplexers get no destructive clear and no history rewrap on resize** —\n   repaint the window in place; pane history keeps its old wrap.\n10. **Any change to the ledger math, the emitters, or the seam must be\n    validated by the stress harness (§6)** across its full scenario matrix,\n    not by a single-terminal smoke test.\n\n---\n\n## 4. Terminal capability detection\n\n`TERMINAL` (`terminal-capabilities.ts`) is resolved once at import from\n`TERMINAL_ID` plus environment sniffing; detection helpers are pure over\n`(env, platform)` and unit-testable.\n\n- `shouldEnableSynchronizedOutputByDefault(env, id)` → DEC 2026 default.\n  Precedence: user opt-out (`PI_NO_SYNC_OUTPUT`/`PI_TUI_SYNC_OUTPUT=0`) → user\n  force-on (`PI_FORCE_SYNC_OUTPUT=1`/`PI_TUI_SYNC_OUTPUT=1`) → `TERM_FEATURES`\n  advertises `Sy` → `WT_SESSION` → known direct terminals → off for risky\n  multiplexers and unknowns. Reconciled at runtime by the DECRQM mode-2026\n  report; a user override still wins.\n- `detectRectangularSgrSupport(id, env)` → DECCARA fills: kitty only, off in\n  multiplexers and under `PI_NO_DECCARA`.\n- `supportsScreenToScrollback` → kitty's ED22 (used once, on the initial\n  paint, to preserve the pre-existing shell screen).\n\nThe old ED3-risk classifier (`eagerEraseScrollbackRisk`, `PI_TUI_ED3_SAFE`,\n`submitPinsViewportToTail`) is gone: behavior no longer depends on which\nterminal is rendering, so there is no risk class to detect. Env sniffing now\nonly selects *optimizations* (sync output, DECCARA, images), where a miss is\ncosmetic, not corrupting.\n\n---\n\n## 5. Width model\n\n`visibleWidth` / `truncateToWidth` / `sliceByColumn` / `wrapTextWithAnsi`\n(`utils.ts`) all route through **one native UAX#11 engine**\n(`@oh-my-pi/pi-natives`, Rust `unicode-width`). `Bun.stringWidth` was dropped\ndeliberately — mixing two width models in measure-vs-slice produced crashes.\n\n- Fast path: printable ASCII is one cell per code unit.\n- ZWJ pictographic emoji take the `visibleWidthByGrapheme` override.\n- OSC 66 sized text takes the native path.\n\n**Rule:** any new measuring code routes through these helpers, and the hot\npath clamps instead of throwing. Known residual: combining-heavy scripts\n(Arabic harakat) survive painting verbatim, but ghostty-web's cell readback can\nmigrate non-spacing marks across cells — the stress harness compares those rows\nwith marks stripped (`sameLinesAllowingMarkDrift`).\n\n---\n\n## 6. The fidelity gate (use it)\n\n`packages/tui/test/render-stress-harness.ts` drives the renderer's **real\nemitted ANSI** into a ghostty-web `VirtualTerminal` across randomized op\nsequences and parameterized terminal shapes, and validates the contract with a\n**shadow commit ledger**: an independent reimplementation of §1's math, fed\nonly by observed frames (a `render` wrap) and observed bytes (a `write` wrap).\nPer op it asserts:\n\n- the whole tape (scrollback + grid) equals `shadowTape + window slice`, row\n  for row, including across resizes;\n- scrolled readers stay pinned and visible history rows are never rewritten;\n- multiplexer pane history grows by exactly the committed chunk;\n- sync-output/autowrap bracket discipline, cursor parking, background columns,\n  duplicate accounting.\n\nRun it — plus `render-regressions.test.ts`,\n`streaming-scrollback-defer.test.ts`, and the `issue-*-repro.test.ts` files —\nbefore changing ledger math, emitters, or the seam. A change that passes one\nterminal and one seed is not verified.\n\n---\n\n## 7. Capability probes & stdin reassembly\n\n`ProcessTerminal` fuses capability queries with a bare DA1 (`CSI c`) sentinel so\na non-answering terminal is detected when DA1 returns first. Replies can arrive\n**split across a stdin flush**, so:\n\n- `#privateCsiResponseBuffer` accumulates `\\x1b[?…` partials while a sentinel is\n  outstanding, rejoins on the terminator byte, then runs the handlers on the\n  **complete** reply. A new `\\x1b` mid-reassembly or >256 bytes abandons the\n  partial so real keys still reach input.\n- `#da1SentinelOwners` is a **typed FIFO** discriminated by `kind` so a\n  keyboard DA1 cannot be mistaken for an OSC 11 / DECRQM / graphics-probe\n  sentinel.\n- DECRQM probes (2026/2048/2031) drive runtime feature gating.\n\n**Rule:** any new probe must own a typed sentinel and survive a split reply\n(feed the reply byte-by-byte in a test and assert nothing leaks to input).\n\n---\n\n## 8. Inline images & memory\n\nKitty images are **transmit-once, place-many** (`kitty-graphics.ts`).\n`ImageBudget` keeps only the most-recent N images live; when the cap is\nexceeded the demoted image's pixels are deleted by id (`a=d,d=I`) and its\nvisible rows re-render as the text fallback through the ordinary window diff —\n**no destructive replay**. A demoted placement already committed to history\nsimply loses its pixels (committed rows are immutable), and the text fallback\nis **height-preserving** once a graphic has rendered (reserved rows + fallback\nline), so demotion never shrinks the block and never shifts committed content\nbelow it.\n\n**Rule:** never re-emit full base64 per frame. Kitty Unicode placeholders are\ndefault-on only for kitty/ghostty (`PI_NO_KITTY_PLACEHOLDERS` /\n`PI_KITTY_PLACEHOLDERS`).\n\n---\n\n## 9. Escape hatches (env vars)\n\n| Var | Effect |\n|---|---|\n| `PI_NO_SYNC_OUTPUT=1` | Disable DEC 2026 BSU/ESU wrappers (autowrap discipline stays on). |\n| `PI_TUI_SYNC_OUTPUT=0\\|1` / `PI_FORCE_SYNC_OUTPUT=1` | Force sync output off / on. |\n| `PI_NO_DECCARA` | Disable Kitty DECCARA rectangular-fill optimization. |\n| `PI_FORCE_IMAGE_PROTOCOL=kitty\\|iterm2\\|sixel\\|off` | Override image protocol detection. |\n| `PI_NO_KITTY_PLACEHOLDERS=1` / `PI_KITTY_PLACEHOLDERS=1` | Force Kitty Unicode placeholders off / on. |\n| `PI_HARDWARE_CURSOR=1` | Show the real hardware cursor instead of a rendered one. |\n| `PI_NOTIFICATIONS=off\\|0\\|false` | Suppress terminal notifications. |\n| `PI_DEBUG_REDRAW=1` | Log the chosen render intent + ledger state per frame to the debug log. |\n\nRemoved with the old engine: `PI_TUI_ED3_SAFE` (no ED3-risk lever exists),\n`PI_CLEAR_ON_SHRINK` (shrinks always clear exactly), `PI_TUI_DEBUG` (per-render\ndump superseded by `PI_DEBUG_REDRAW` ledger logging and the stress harness\nreplay/reduce tooling).\n\n---\n\n## 10. Before you touch the render core — checklist\n\n- [ ] Are you about to emit `CSI 3 J` anywhere other than the gesture-driven\n      `clearScrollback` full paint? **Stop.**\n- [ ] Could any code path rewrite, or re-show on the grid, a frame row below\n      `committedRows`? **Stop.**\n- [ ] Does your byte shape scroll rows that are not the commit chunk? That\n      breaks `scrollback == frame[0..C)`.\n- [ ] Are you adding a viewport probe, a platform fork, or a terminal-brand\n      branch to the update path? The contract exists so none are needed.\n- [ ] New mutable UI above the editor? It must report (or live inside) the\n      live-region seam, or it will freeze at first commit.\n- [ ] Did you run the stress harness and the repro suite across the full\n      scenario matrix — not just one terminal and one seed?\n- [ ] New probe? Typed sentinel owner + split-reply test.\n- [ ] New width path? Routed through the shared native engine, clamped (never\n      thrown) in the hot path.\n",
 	"tui-runtime-internals.md": "# TUI runtime internals\n\nThis document maps the non-theme runtime path from terminal input to rendered output in interactive mode. It focuses on behavior in `packages/tui` and its integration from `packages/coding-agent` controllers.\n\n> **Editing the rendering engine itself?** Read\n> [`tui-core-renderer.md`](./tui-core-renderer.md) first — it documents the\n> failure modes (yank / corruption / flash / width crashes) and the invariants\n> the render planner, native-scrollback bookkeeping, and capability detection\n> must not violate.\n\n## Runtime layers and ownership\n\n- **`packages/tui` engine**: terminal lifecycle, stdin normalization, focus routing, render scheduling, differential painting, overlay composition, hardware cursor placement.\n- **`packages/coding-agent` interactive mode**: builds component tree, binds editor callbacks and keymaps, reacts to agent/session events, and translates domain state (streaming, tool execution, retries, plan mode) into UI components.\n\nBoundary rule: the TUI engine is message-agnostic. It only knows `Component.render(width)`, `handleInput(data)`, focus, and overlays. Agent semantics stay in interactive controllers.\n\n## Implementation files\n\n- [`packages/coding-agent/src/modes/interactive-mode.ts`](../packages/coding-agent/src/modes/interactive-mode.ts)\n- [`packages/coding-agent/src/modes/controllers/event-controller.ts`](../packages/coding-agent/src/modes/controllers/event-controller.ts)\n- [`packages/coding-agent/src/modes/controllers/input-controller.ts`](../packages/coding-agent/src/modes/controllers/input-controller.ts)\n- [`packages/coding-agent/src/modes/components/custom-editor.ts`](../packages/coding-agent/src/modes/components/custom-editor.ts)\n- [`packages/tui/src/tui.ts`](../packages/tui/src/tui.ts)\n- [`packages/tui/src/terminal.ts`](../packages/tui/src/terminal.ts)\n- [`packages/tui/src/editor-component.ts`](../packages/tui/src/editor-component.ts)\n- [`packages/tui/src/stdin-buffer.ts`](../packages/tui/src/stdin-buffer.ts)\n- [`packages/tui/src/components/loader.ts`](../packages/tui/src/components/loader.ts)\n\n## Boot and component tree assembly\n\n`InteractiveMode` constructs `TUI(new ProcessTerminal(), settings.get(\"showHardwareCursor\"))`, applies `tui.maxInlineImages` and Kitty text-sizing settings, then creates persistent containers:\n\n- `chatContainer`\n- `pendingMessagesContainer`\n- `statusContainer`\n- `todoContainer`\n- `subagentContainer`\n- `btwContainer`\n- `omfgContainer`\n- `errorBannerContainer`\n- `statusLine`\n- `hookWidgetContainerAbove`\n- `editorContainer` (holds `CustomEditor`)\n- `hookWidgetContainerBelow`\n\n`init()` wires the tree in that order after any startup warnings/welcome/changelog, focuses the editor, registers input handlers via `InputController`, starts TUI, pushes terminal title state, updates the editor border, and requests a forced render.\nA forced render (`requestRender(true)`) queues a viewport repaint or explicit session replacement; it does **not** throw away previous-line history by default.\n\n## Terminal lifecycle and stdin normalization\n\n`ProcessTerminal.start()`:\n\n1. Enables raw mode and bracketed paste.\n2. Attaches resize handler and refreshes dimensions.\n3. Enables Windows VT input mode when running on win32.\n4. Creates a `StdinBuffer` to split partial escape chunks into complete sequences.\n5. Queries Kitty keyboard protocol support (`CSI ? u`), then enables protocol flags if supported; otherwise enables modifyOtherKeys fallback after a short timeout.\n6. Queries OSC 11 background color and Mode 2031 appearance notifications for dark/light theme detection.\n7. Queries OSC 99 notification capabilities.\n8. Starts periodic OSC 11 polling only where safe, then probes DEC private modes 2026/2048/2031 via DECRQM.\n\n`StdinBuffer` behavior:\n\n- Buffers fragmented escape sequences (CSI/OSC/DCS/APC/SS3).\n- Emits `data` only when a sequence is complete or timeout-flushed.\n- Detects bracketed paste and emits a `paste` event with raw pasted text.\n\nThis prevents partial escape chunks from being misinterpreted as normal keypresses.\n\n## Input routing and focus model\n\nInput path:\n\n`stdin -> ProcessTerminal -> StdinBuffer -> TUI.#handleInput -> focusedComponent.handleInput`\n\nRouting details:\n\n1. TUI runs registered input listeners first (`addInputListener`), allowing consume/transform behavior.\n2. TUI handles global debug shortcut (`shift+ctrl+d`) before component dispatch.\n3. If focused component belongs to an overlay that is now hidden/invisible, TUI reassigns focus to next visible overlay or saved pre-overlay focus.\n4. Key release events are filtered unless focused component sets `wantsKeyRelease = true`.\n5. After dispatch, TUI schedules render.\n\n`setFocus()` also toggles `Focusable.focused`, which controls whether components emit `CURSOR_MARKER` for hardware cursor placement.\n\n## Key handling split: editor vs controller\n\n`CustomEditor` intercepts high-priority combos first (escape, ctrl-c/d/z, ctrl-v, ctrl-p variants, ctrl-t, alt-up, extension custom keys) and delegates the rest to base `Editor` behavior (text editing, history, autocomplete, cursor movement).\n\n`InputController.setupKeyHandlers()` then binds editor callbacks to mode actions:\n\n- cancellation / mode exits on `Escape`\n- shutdown on double `Ctrl+C` or empty-editor `Ctrl+D`\n- suspend/resume on `Ctrl+Z`\n- slash-command and selector hotkeys\n- follow-up/dequeue toggles and expansion toggles\n\nThis keeps key parsing/editor mechanics in `packages/tui` and mode semantics in coding-agent controllers.\n\n## Render loop and the append-only contract\n\n`TUI.requestRender()` coalesces render requests and rate-limits ordinary frames:\n\n- forced renders (`requestRender(true, ...)`) schedule an immediate frame and force a full window rewrite; with `clearScrollback`, they trigger a destructive full paint (ED3 outside multiplexers)\n- ordinary renders schedule through `#scheduleRender()` and respect `TUI.#MIN_RENDER_INTERVAL_MS`\n- repeated requests while a render is pending collapse into the same scheduled frame\n- `requestComponentRender(component)` requests on behalf of a single self-contained change (spinner frame, blink): when every request in the coalesced frame is component-scoped and the frame is quiet (no resize, overlays, inline images, forced repaint, or root-list change), compose re-renders only the root subtrees containing the requesting components and reuses every other root child's previous rows and seam report; any unsafe condition or concurrent full request downgrades to a full compose\n\n`#doRender()` pipeline:\n\n1. Render root component tree, collecting the commit-boundary seam (`NativeScrollbackLiveRegion`) from the children.\n2. Advance the append-only ledger: `windowTop = max(committedRows, frame.length - height)`, commit chunk = settled rows crossing the window top (never past the seam).\n3. Extract and strip `CURSOR_MARKER`, normalize lines, slice the visible window, composite overlays into the window slice (screen coordinates; overlays freeze commits).\n4. Emit one of: gesture-driven full paint (initial / session replace / resize), scroll-append (chunk rows only), in-window row diff, or seam rewrite (chunk + full window).\n\nNative scrollback always equals the committed frame prefix — rows enter history exactly once, in order, when the seam says they are final. There are no viewport probes and no deferred reconciliation; see [`tui-core-renderer.md`](./tui-core-renderer.md).\n\nRender writes use synchronized output mode (`CSI ? 2026 h/l`) when enabled; capability detection, DECRQM, or `PI_NO_SYNC_OUTPUT` can disable the wrappers while leaving autowrap discipline on.\n\n## Render safety constraints\n\nCritical safety checks in `TUI`:\n\n- Non-image rendered lines are expected to fit terminal width; the differential path truncates overwide lines as a last-resort guard and can write debug diagnostics when redraw debugging is enabled.\n- Overlay compositing includes defensive truncation and post-composite width guarding.\n- Width changes force repaint/rebuild planning because wrapping semantics change.\n- Cursor position is clamped before movement.\n\nThese constraints are runtime guards plus component conventions; renderers should still return width-safe lines rather than rely on truncation.\n\nThe deeper reasons these guards exist — why the renderer cannot observe scroll\nposition, why ED3 (`CSI 3 J`) is confined to one path, and why the hot path\nclamps instead of throwing — are documented in\n[`tui-core-renderer.md`](./tui-core-renderer.md).\n\n## Resize handling\n\nResize events are event-driven from `ProcessTerminal` to `TUI.requestRender()`.\n\nEffects:\n\n- A resize is an explicit user gesture: outside multiplexers the engine erases and replays (`ED3` + full paint) so history rewraps at the new geometry; the commit ledger restarts from the replayed frame.\n- Inside terminal multiplexers, resize repaints the visible window in place after a settle debounce (issue #2088); pane history keeps its old wrap, like any shell output, because pane scrollback cannot be erased safely.\n- Overlay visibility can depend on terminal dimensions (`OverlayOptions.visible`); focus is corrected when overlays become non-visible after resize.\n\n## Streaming and incremental UI updates\n\n`EventController` subscribes to `AgentSessionEvent` and updates UI incrementally:\n\n- `agent_start`: starts loader in `statusContainer`.\n- `message_start` assistant: creates `streamingComponent` and mounts it.\n- `message_update`: updates streaming assistant content; creates/updates tool execution components as tool calls appear.\n- `tool_execution_update/end`: updates tool result components and completion state.\n- `message_end`: finalizes assistant stream, handles aborted/error annotations, marks pending tool args complete on normal stop.\n- `agent_end`: stops loaders, clears transient stream state, flushes deferred model switch, issues completion notification if backgrounded.\n\nRead-tool grouping is intentionally stateful (`#lastReadGroup`) to coalesce consecutive read tool calls into one visual block until a non-read break occurs.\n\n## Status and loader orchestration\n\nStatus lane ownership:\n\n- `statusContainer` holds transient loaders (`loadingAnimation`, `autoCompactionLoader`, `retryLoader`).\n- `statusLine` renders persistent status/hooks/plan indicators and drives editor top border updates.\n\nLoader behavior:\n\n- `Loader` advances its spinner every 80ms (animated message colorizers redraw at ~30fps) and requests a component-scoped render each frame (`requestComponentRender`), so idle spinner ticks repaint without re-walking the transcript.\n- Escape handlers are temporarily overridden during auto-compaction and auto-retry to cancel those operations.\n- On end/cancel paths, controllers restore prior escape handlers and stop/clear loader components.\n\n## Mode transitions and backgrounding\n\n### Bash/Python input modes\n\nInput text prefixes toggle editor border mode flags:\n\n- `!` -> bash mode\n- `$` (non-template literal prefix) -> python mode\n\nEscape exits inactive mode by clearing editor text and restoring border color; when execution is active, escape aborts the running task instead.\n\n### Plan mode\n\n`InteractiveMode` tracks plan mode flags, status-line state, active tools, and model switching. Enter/exit updates session mode entries and status/UI state, including deferred model switch if streaming is active.\n\n### Suspend/resume (`Ctrl+Z`)\n\n`InputController.handleCtrlZ()`:\n\n1. Registers one-shot `SIGCONT` handler to restart TUI and force render.\n2. Stops TUI before suspend.\n3. Sends `SIGTSTP` to process group.\n\n## Cancellation paths\n\nPrimary cancellation inputs:\n\n- `Escape` during active stream loader: restores queued messages to editor and aborts agent.\n- `Escape` during bash/python execution: aborts running command.\n- `Escape` during auto-compaction/retry: invokes dedicated abort methods through temporary escape handlers.\n- `Ctrl+C` single press: clear editor; double press within 500ms: shutdown.\n\nCancellation is state-conditional; same key can mean abort, mode-exit, selector trigger, or no-op depending on runtime state.\n\n## Event-driven vs throttled behavior\n\nEvent-driven updates:\n\n- Agent session events (`EventController`)\n- Key input callbacks (`InputController`)\n- terminal resize callback\n- terminal appearance callbacks, SIGWINCH theme reevaluation, and git branch watchers in `InteractiveMode`\n\nThrottled/debounced paths:\n\n- TUI rendering is tick-debounced (`requestRender` coalescing).\n- Loader animation is interval-driven (80ms spinner advance; ~30fps when the message colorizer is animated), each frame requesting a component-scoped render.\n- Editor autocomplete updates (inside `Editor`) use debounce timers, reducing recompute churn during typing.\n\nThe runtime therefore mixes event-driven state transitions with bounded render cadence to keep interactivity responsive without repaint storms.\n",
 	"tui.md": "# TUI integration for extensions and custom tools\n\nThis document covers the **current** TUI contract used by `packages/coding-agent` and `packages/tui` for extension UI, custom tool UI, and custom renderers.\n\n## What this subsystem is\n\nThe runtime has two layers:\n\n- **Rendering engine (`packages/tui`)**: differential terminal renderer, input dispatch, focus, overlays, cursor placement.\n- **Integration layer (`packages/coding-agent`)**: mounts extension/custom-tool components, wires keybindings/theme, and restores editor state.\n\n## Runtime behavior by mode\n\n| Mode                | `ctx.ui.custom(...)` availability | Notes                                                                                                                          |\n| ------------------- | --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ |\n| Interactive TUI     | Supported                         | Component is mounted in the editor area or overlay, focused, and must call `done(result)` to resolve.                          |\n| Background/headless | Not interactive                   | UI context is no-op (`hasUI === false`).                                                                                       |\n| RPC mode            | Not mounted                       | `custom()` is implemented as unsupported UI and returns `undefined as never`; do not depend on interactive UI in RPC handlers. |\n\nIf your extension/tool can run in non-interactive mode, guard with `ctx.hasUI` / `pi.hasUI`.\n\n## Core component contract (`@oh-my-pi/pi-tui`)\n\n`packages/tui/src/tui.ts` defines:\n\n```ts\nexport interface Component {\n  render(width: number): readonly string[];\n  handleInput?(data: string): void;\n  wantsKeyRelease?: boolean;\n  invalidate?(): void;\n  dispose?(): void;\n}\n```\n\nRender results are component-owned and immutable to callers; a component that did not change should return the **same array reference** it returned last time (reference equality is what enables the renderer's memoization and row virtualization), and must return a new array whenever its content changed.\n\n`Focusable` is separate:\n\n```ts\nexport interface Focusable {\n  focused: boolean;\n}\n```\n\nCursor behavior uses `CURSOR_MARKER` (not `getCursorPosition`). Focused components emit the marker in rendered text; `TUI` extracts it and positions the hardware cursor.\n\n## Rendering constraints (terminal safety)\n\nYour `render(width)` output must be terminal-safe:\n\n1. **Do not intentionally exceed `width` on any line**. The renderer truncates overwide non-image lines as a last-resort guard, but components should still return width-safe output.\n2. **Measure visual width**, not string length: use `visibleWidth()`.\n3. **Truncate/wrap ANSI-aware text** with `truncateToWidth()` / `wrapTextWithAnsi()`.\n4. **Sanitize tabs/content** from external sources using `replaceTabs()` (and higher-level sanitizers in coding-agent render paths).\n\nMinimal pattern:\n\n```ts\nimport { replaceTabs, truncateToWidth } from \"@oh-my-pi/pi-tui\";\n\nrender(width: number): readonly string[] {\n  return this.lines.map(line => truncateToWidth(replaceTabs(line), width));\n}\n```\n\n## Input handling and keybindings\n\n### Raw key matching\n\nUse `matchesKey(data, \"...\")` for navigation keys and combos.\n\n### Match app keybinding actions\n\nExtension UI factories receive a `KeybindingsManager` (interactive mode; an in-memory instance carrying the default bindings, not the user's `keybindings.yml`) so you can match action ids instead of hardcoding keys:\n\n```ts\nif (keybindings.matches(data, \"app.interrupt\")) {\n  done(undefined);\n  return;\n}\n```\n\n### Key release/repeat events\n\nKey release events are filtered unless your component sets:\n\n```ts\nwantsKeyRelease = true;\n```\n\nThen use `isKeyRelease()` / `isKeyRepeat()` if needed.\n\n## Focus, overlays, and cursor\n\n- `TUI.setFocus(component)` routes input to that component.\n- Overlay APIs exist in `TUI` (`showOverlay`, `OverlayHandle`). In interactive extension/custom UI, `custom(..., { overlay: true })` mounts your component through `TUI.showOverlay(...)`; without `overlay`, it replaces the editor component area directly.\n- Overlay custom UI is anchored at `bottom-center` with full terminal width/max height and is removed through the returned overlay handle when `done(...)` closes the flow.\n\n## Mount points and return contracts\n\n## 1) Extension UI (`ExtensionUIContext`)\n\nCurrent signature (`extensibility/extensions/types.ts`):\n\n```ts\ncustom<T>(\n  factory: (\n    tui: TUI,\n    theme: Theme,\n    keybindings: KeybindingsManager,\n    done: (result: T) => void,\n  ) => (Component & { dispose?(): void }) | Promise<Component & { dispose?(): void }>,\n  options?: { overlay?: boolean },\n): Promise<T>\n```\n\nBehavior in interactive mode (`extension-ui-controller.ts`):\n\n- Saves editor text.\n- Without `options.overlay`, replaces the editor component with your component.\n- With `options.overlay`, mounts your component as a bottom-centered overlay instead of replacing the editor.\n- Focuses your component.\n- On `done(result)`: calls `component.dispose?.()`, hides the overlay if present, restores editor + text for non-overlay flows, focuses editor, resolves promise.\n  So `done(...)` is mandatory for completion.\n\n## 2) Hook/custom-tool UI context (legacy typing)\n\n`HookUIContext.custom` is typed as `(tui, theme, done)` in hook/custom-tool types.\nUnderlying interactive implementation calls factories with `(tui, theme, keybindings, done)`. JS consumers can use the extra arg; type-level compatibility still reflects the 3-arg legacy signature.\n\nCustom tools typically use the same UI entrypoint via the factory-scoped `pi.ui` object, then return the selected value in normal tool content:\n\n```ts\nasync execute(toolCallId, params, onUpdate, ctx, signal) {\n  if (!pi.hasUI) {\n    return { content: [{ type: \"text\", text: \"UI unavailable\" }] };\n  }\n\n  const picked = await pi.ui.custom<string | undefined>((tui, theme, done) => {\n    const component = new MyPickerComponent(done, signal);\n    return component;\n  });\n\n  return { content: [{ type: \"text\", text: picked ? `Picked: ${picked}` : \"Cancelled\" }] };\n}\n```\n\n## 3) Custom tool call/result renderers\n\nCustom tools and extension tools can return components from:\n\n- `renderCall(args, options, theme)`\n- `renderResult(result, options, theme, args?)`\n\n`options` currently includes:\n\n- `expanded: boolean`\n- `isPartial: boolean`\n- `spinnerFrame?: number`\n\nThese renderers are mounted by `ToolExecutionComponent`.\n\n## Lifecycle and cancellation\n\n- `dispose()` is optional at type level but should be implemented when you own timers, subprocesses, watchers, sockets, or overlays.\n- `done(...)` should be called exactly once from your component flow.\n- For cancellable long-running UI, pair `CancellableLoader` with `AbortSignal` and call `done(...)` from `onAbort`.\n\nExample cancellation pattern:\n\n```ts\nconst loader = new CancellableLoader(\n  tui,\n  theme.fg(\"accent\"),\n  theme.fg(\"muted\"),\n  \"Working...\",\n);\nloader.onAbort = () => done(undefined);\nvoid doWork(loader.signal).then((result) => done(result));\nreturn loader;\n```\n\n## Realistic custom component example (extension command)\n\n```ts\nimport type { Component } from \"@oh-my-pi/pi-tui\";\nimport {\n  SelectList,\n  matchesKey,\n  replaceTabs,\n  truncateToWidth,\n} from \"@oh-my-pi/pi-tui\";\nimport {\n  getSelectListTheme,\n  type ExtensionAPI,\n} from \"@oh-my-pi/pi-coding-agent\";\n\nclass Picker implements Component {\n  list: SelectList;\n  keybindings: any;\n  done: (value: string | undefined) => void;\n\n  constructor(\n    items: Array<{ value: string; label: string }>,\n    keybindings: any,\n    done: (value: string | undefined) => void,\n  ) {\n    this.list = new SelectList(items, 8, getSelectListTheme());\n    this.keybindings = keybindings;\n    this.done = done;\n    this.list.onSelect = (item) => this.done(item.value);\n    this.list.onCancel = () => this.done(undefined);\n  }\n\n  handleInput(data: string): void {\n    if (this.keybindings.matches(data, \"app.interrupt\")) {\n      this.done(undefined);\n      return;\n    }\n    this.list.handleInput(data);\n  }\n\n  render(width: number): readonly string[] {\n    return this.list\n      .render(width)\n      .map((line) => truncateToWidth(replaceTabs(line), width));\n  }\n\n  invalidate(): void {\n    this.list.invalidate();\n  }\n}\n\nexport default function extension(pi: ExtensionAPI): void {\n  pi.registerCommand(\"pick-model\", {\n    description: \"Pick a model profile\",\n    handler: async (_args, ctx) => {\n      if (!ctx.hasUI) return;\n\n      const selected = await ctx.ui.custom<string | undefined>(\n        (tui, theme, keybindings, done) => {\n          const items = [\n            { value: \"fast\", label: theme.fg(\"accent\", \"Fast\") },\n            { value: \"balanced\", label: \"Balanced\" },\n            { value: \"quality\", label: \"Quality\" },\n          ];\n          return new Picker(items, keybindings, done);\n        },\n      );\n\n      if (selected) ctx.ui.notify(`Selected profile: ${selected}`, \"info\");\n    },\n  });\n}\n```\n\n## Key implementation files\n\n- `packages/tui/src/tui.ts` — `Component`, `Focusable`, cursor marker, focus, overlay, input dispatch.\n- `packages/tui/src/utils.ts` — width/truncation/sanitization primitives.\n- `packages/tui/src/keys.ts` / `keybindings.ts` — key parsing and configurable action mapping.\n- `packages/coding-agent/src/modes/controllers/extension-ui-controller.ts` — interactive mounting/unmounting for extension/hook/custom-tool UI.\n- `packages/coding-agent/src/extensibility/extensions/types.ts` — extension UI and renderer contracts.\n- `packages/coding-agent/src/extensibility/hooks/types.ts` — hook UI contract (legacy custom signature).\n- `packages/coding-agent/src/extensibility/custom-tools/types.ts` — custom tool execute/render contracts.\n- `packages/coding-agent/src/modes/components/tool-execution.ts` — mounting `renderCall`/`renderResult` components and partial-state options.\n- `packages/coding-agent/src/tools/context.ts` — tool UI context propagation (`hasUI`, `ui`).\n",
 };