@gajae-code/coding-agent 0.10.1 → 0.10.2

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

@@ -29,7 +29,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"gjc-dogfood-skill-template.md": "# GJC dogfood local skill template\n\nIssue #93 requested a gaebal-gajae/operator dogfood skill. The live issue has no comment approving a fifth bundled default workflow skill, so this stays a local template instead of changing the default workflow surface. Operators can copy it into a user or project override when they want GJC-first session guidance.\n\nThe installable skill body is everything from the first frontmatter marker down; the frontmatter must be the **first line** of the installed file or the skill scan silently skips it (the scan requires a parsed `description`). Install into the user-level scan location (`~/.gjc/agent/skills/`, not `~/.gjc/skills/`):\n\n```sh\nmkdir -p ~/.gjc/agent/skills/gjc-dogfood\nsed -n '/^---$/,$p' docs/gjc-dogfood-skill-template.md > ~/.gjc/agent/skills/gjc-dogfood/SKILL.md\n```\n\nFor a single project, install to `<project>/.gjc/skills/gjc-dogfood/SKILL.md` with the same extraction. Do not commit that project `.gjc` copy unless the project explicitly wants a local override.\n\nFilesystem skill discovery is off by default, so enable it once. Set `skills.enabled`, then enable **only the scan that matches where you installed** — `enablePiUser` and `enablePiProject` default to `false` in `DEFAULT_SKILL_DISCOVERY_SETTINGS`, and enabling the project scan opts every future session into repo-local `.gjc/skills` discovery, so do not enable it for a user-only install:\n\n```sh\ngjc config set skills.enabled true\n\n# for the user-level install (~/.gjc/agent/skills/):\ngjc config set skills.enablePiUser true\n\n# OR, for the project-level install (<project>/.gjc/skills/):\ngjc config set skills.enablePiProject true\n```\n\nThen verify in a new session: `/skill:gjc-dogfood` should autocomplete.\n\n---\nname: gjc-dogfood\ndescription: Use when running or reviewing work through GJC sessions, dogfooding Gajae-Code, or migrating an operator workflow from OMX to GJC.\n---\n\n# GJC Dogfood Operator Workflow\n\nUse GJC first for coding, review, planning, and follow-up sessions. Treat OMX as a fallback only when GJC is unavailable, broken, or missing a required capability.\n\n## Locate and launch GJC\n\n- Installed CLI: run `command -v gjc` and then launch with `gjc --tmux`.\n- Repository checkout: from the gajae-code repo, prefer `bun packages/coding-agent/src/cli.ts --tmux` when testing source changes before install.\n- Worktree isolation: for branch-specific work, either let GJC create a managed sibling worktree with `gjc --tmux --worktree <branch-like-name>` or `cd <existing-worktree-path>` and run `gjc --tmux` there. Do not pass filesystem paths to `--worktree`.\n- Name sessions explicitly with the project and issue, for example `gajae-code-93-dogfood-skill`, so tmux panes, logs, and exports remain traceable.\n\n## Start the session\n\n- Put git operations inside the GJC session: fetch, branch/worktree setup, focused commits, pushes, and PR creation should be visible in-session.\n- Submit the initial prompt with the issue URL, target branch, acceptance criteria, verification limits, and any existing plan/spec link.\n- Verify the prompt was accepted: the TUI should show the user prompt, an active assistant turn, or a tool/action request. If the session silently idles, resend once with a shorter prompt and capture the failure.\n- Verify working state before leaving the session unattended: confirm the target cwd/worktree, branch, and issue scope are visible in the transcript or command output.\n\n## During work\n\n- Keep session names and branch names issue-scoped.\n- Prefer GJC workflow skills only when they fit: `deep-interview` for unclear requirements, `ralplan` for planning, `ultragoal` for durable ledgers, and `team` for coordinated tmux execution.\n- Keep evidence in the session: issue reads, focused tests/checks, screenshots only when visual behavior matters, and PR URLs.\n- When GJC is weaker than OMX, finish the urgent work with the smallest safe fallback and file a gajae-code follow-up issue with the missing capability, exact command/session context, expected behavior, and evidence.\n\n## Fallback policy\n\nUse OMX or another operator path only when:\n\n- `gjc` cannot be located or launched after checking installed and repo-local commands;\n- authentication, model routing, tmux, or prompt submission is broken;\n- GJC lacks a required capability that OMX already has;\n- an urgent production/review deadline would be missed by debugging GJC first.\n\nRecord the fallback reason and create or link the gajae-code issue that would make GJC sufficient next time.\n\n## Evidence checklist\n\nReport:\n\n- project, issue, branch/worktree, and session name;\n- whether GJC was installed or repo-local;\n- prompt acceptance and working-state evidence;\n- git operations performed in-session;\n- focused verification commands and results;\n- PR/issue URLs;\n- follow-up gajae-code issues for any GJC gap or fallback.\n",
 	"gjc-plugins.md": "# GJC Plugin Bundles\n\nGJC supports two distinct plugin families. Do not confuse them:\n\n1. **Legacy marketplace / npm plugins** (`packages/coding-agent/src/extensibility/plugins`) — installed through the existing `gjc plugin install <marketplace-ref|npm-spec>` marketplace/npm flows. Unchanged by this system.\n2. **GJC plugin bundles** — directories whose root contains a **`gajae-plugin.json`** manifest (`kind: \"gajae-code-plugin\"`). These *extend* existing GJC capabilities and are the subject of this document.\n\nA GJC plugin bundle may only **extend** existing skills/agents — it can never register a new top-level skill, slash-command, command, or agent. GJC exposes exactly four default workflow skills (`deep-interview`, `ralplan`, `team`, `ultragoal`) and four role agents (`executor`, `architect`, `planner`, `critic`); bundles add sub-skills/appendices/tools/hooks/MCPs to those existing parents only.\n\n## Manifest (`gajae-plugin.json`)\n\n```json\n{\n  \"kind\": \"gajae-code-plugin\",\n  \"name\": \"example-domain-bundle\",\n  \"version\": \"1.0.0\",\n  \"subskills\": [\"subskills/ralplan-design/SKILL.md\"],\n  \"tools\": [\n    { \"name\": \"domain_note\", \"path\": \"tools/domain-note.ts\", \"description\": \"...\" }\n  ],\n  \"hooks\": [\n    { \"name\": \"audit-read\", \"event\": \"tool_call\", \"target\": \"read\", \"phase\": \"before\", \"path\": \"hooks/audit-read.ts\" }\n  ],\n  \"mcps\": [\n    { \"name\": \"domain_docs\", \"transport\": \"stdio\", \"command\": \"bun\", \"args\": [\"mcp/domain-docs.ts\"], \"cwd\": \".\" }\n  ],\n  \"system_appendix\": [{ \"name\": \"domain-policy\", \"path\": \"prompts/system-appendix.md\" }],\n  \"agent-appendix\": [{ \"agent\": \"executor\", \"name\": \"domain-executor\", \"path\": \"prompts/executor-appendix.md\" }]\n}\n```\n\n### Surfaces (the only allowed extension points)\n\n| Surface | Purpose | Additive rule |\n|---------|---------|---------------|\n| `subskills` | Inline sub-skills bound to an existing skill/agent (`binds_to`/`phase`/`activation_arg`) | Two-tier (see below) |\n| `tools` | Always-on custom tools (object entries) or legacy subskill-scoped string paths | Additive; manifest-declared name is authoritative, never overwrites an existing tool |\n| `hooks` | Constrained event hooks bound to a declared `event`/`target`/`phase` | Additive; run alongside built-ins, never replace |\n| `mcps` | MCP servers (`stdio`/`http`/`sse`) | Additive; server-name collisions are hard errors |\n| `system_appendix` | Lower-authority text appended to the default agent system prompt | Append-only, never overrides base |\n| `agent-appendix` | Lower-authority text appended to an existing role agent's prompt | Append-only per named agent |\n\n### Forbidden / unsupported keys\n\n- **Forbidden** (`forbidden_surface`): `skills`, `slash-commands`, `commands`, `agents` — bundles may not register new top-level definitions.\n- **Unsupported** (`unsupported_surface`): `mcp`, `mcpServers` (use the canonical `mcps`), and any unknown top-level key.\n\n## Installation\n\n```sh\ngjc plugin install <path|git-url|tarball> --user      # install into the user root\ngjc plugin install <path|git-url|tarball> --project   # install into the project root\n```\n\nExactly one of `--user` / `--project` is required for GJC plugin bundles (there is no default root). A source containing `gajae-plugin.json` is classified as a GJC bundle and routed to the bundle installer **before** the marketplace/npm path; non-bundle sources fall through to the legacy flow.\n\nInstall is **compile-validate-then-copy**:\n\n1. The bundle is compiled and validated **without importing any plugin code** (manifest, frontmatter, and declared files are read as bytes only).\n2. Collision and MCP security policy are enforced (the durable registry is the collision authority — never capability \"first-wins\").\n3. Only the validated, hashed files are copied into a temp sibling, then atomically renamed into place; the registry entry is written last under a per-scope lock. Nothing is mutated on failure.\n\nIdempotency: re-installing identical content is a no-op; different content requires `--force`.\n\n## Security model\n\n- **Install validation never executes plugin code.** Tool/hook names are manifest-declared; at runtime the loaded factory must return/register exactly the declared name/event or the surface is quarantined (`runtime_mismatch`).\n- **MCP policy** (install + runtime connect): HTTPS-only for `http`/`sse`; private/loopback/link-local/unique-local/multicast and the `169.254.169.254` metadata endpoint are denied across IPv4, IPv6, IPv4-mapped/compatible, zone-id and trailing-dot forms; URL credentials and CRLF headers are rejected; DNS is re-resolved before connect (rebinding defence). `stdio` servers are confined to the plugin root (allowed launchers `node`/`bun` or a root-confined executable; required bundled script argument; no eval/loader flags; no env expansion).\n- **Hooks** run through a *constrained* API: only a handler for the declared event may be registered. `registerCommand`, `sendMessage`, `appendEntry`, renderer registration, and shell `exec` are denied (`security_policy`). The broad first-party hook API is never exposed to bundle hooks.\n- **Appendices** render as lower-authority, delimited `<gjc-plugin-system-appendix>` / `<gjc-plugin-agent-appendix>` blocks appended after the base/project prompt; size-capped (8 KiB/appendix, 32 KiB total) fail-closed; content is escaped and control-char sanitized. They can never override base/developer instructions.\n- **Hash drift**: installed files are re-verified against the registry at session start; any drift quarantines the plugin (`runtime_mismatch`).\n\n## Sub-skills: Tier-1 vs Tier-2\n\n- **Tier-1 advertisement** (metadata-only): when a parent skill/agent prompt is built, installed sub-skills bound to it are advertised as a bounded list (`plugin` / `name` / `description` / `activation_arg` / `phase`; max 12 items, 200-char descriptions, 4 KiB block, with an overflow note). No body content; rendered only in the target parent prompt, never the global public-workflow surface.\n- **Tier-2 activation** (full body): on explicit activation (e.g. `deep-interview --autoresearch`) or an agent's contextual choice, the full sub-skill body is injected as a `<gjc-subskill>` block at the matching phase.\n\n## Registry, enablement, and quarantine\n\nEach scope keeps a durable `registry.json` recording per-plugin: name/version, source (`path`/`git`/`tarball` + ref/sha), manifest hash, copied files (relative path + sha256 — the uninstall ownership boundary), per-surface extension IDs, `enabled` flag, `disabledSurfaceIds`, and any `quarantine` entries.\n\nExtension IDs are stable: `tool:<name>`, `hook:<event>:<phase>:<target>:<name>`, `mcp:<name>`, `system-appendix:<plugin>:<name>`, `agent-appendix:<agent>:<plugin>:<name>`, `subskill:<parent>:<phase>:<activation_arg>`. Disabled is user-controlled (not an error); quarantine is fail-closed and visible.\n\n## Status / scope notes\n\n- Always-on **tools**, **system appendices**, **agent appendices**, and **Tier-1 advertisement** activate at session start (additive; no-op when no bundle is installed).\n- **MCP runtime connection** and the **live hook runner** integration are gated behind the same validated registry + policy; consult the ledger/run notes for their wiring status.\n- Full enable/disable/uninstall/upgrade UX is a planned follow-up; the registry already records everything required for it (per-surface IDs + copied-file ownership).\n",
 	"gjc-session-clawhip-routing.md": "# Clawhip-routed GJC sessions\n\nThis guide documents the visible tmux session pattern used by operator bots such as Clawhip, Hermes, and OpenClaw when repository work must stay observable in a routed channel.\n\nUse this pattern when a human or chatops router needs to watch the session, receive stale-session alerts, and send follow-up prompts into the same visible GJC pane.\n\nFor pure machine control, prefer the Coordinator MCP tools in [`docs/hermes-mcp-bridge.md`](./hermes-mcp-bridge.md). For a single embedded worker process, prefer [`docs/rpc.md`](./rpc.md). This visible-session pattern is the operator-facing fallback/interop lane.\n\n## Contract\n\n1. Create or verify a dedicated git worktree for the issue or PR.\n2. Register a named tmux session with the host router before launching GJC.\n3. Start interactive `gjc` inside the worktree.\n4. Wait until the GJC TUI is ready.\n5. Inject the real task prompt separately.\n6. Verify acceptance from actual work evidence, not from a visible pasted prompt.\n\nDo not launch visible routed work in the canonical repo checkout. Use a worktree so branch changes, generated files, tests, and cleanup stay scoped to the task.\n\n## Session naming\n\nUse stable names that include the project and artifact id:\n\n```text\ngajae-code-issue-905-ctrl-shift-enter-newline\ngajae-code-pr-911-ctrl-shift-enter-review\nclawhip-issue-269-lightweight-zero-receipt\n```\n\nAvoid ambiguous names such as `fix-tui`, `review`, or `issue-905` when multiple repositories route into the same chat surface.\n\n## Portable script shape\n\nThe exact router command is host-owned. A Clawhip-style wrapper usually has three small scripts:\n\n```sh\n# create.sh\n# create/register a routed tmux session and start interactive gjc in the worktree\nscripts/gjc-session/create.sh <session-name> <worktree-path> [channel-id] [mention]\n\n# prompt.sh\n# inject the real task after the TUI is ready\nscripts/gjc-session/prompt.sh <session-name> @/path/to/task.md\n\n# tail.sh\n# inspect bounded pane output before/after prompt delivery\nscripts/gjc-session/tail.sh <session-name> [lines]\n```\n\nThis repository includes a portable implementation in `scripts/gjc-session/`. It keeps private routing values outside the script body: channel ids and mentions are runtime arguments, the router binary is optional, and credentials are never embedded. Host deployments can still override router behavior with environment variables instead of editing the scripts.\n\n\n## Included helper scripts\n\nThe `scripts/gjc-session/` directory contains the public version of the operator helpers:\n\n- `create.sh` validates a dedicated git worktree, starts interactive `gjc` in tmux, preserves the pane after exit, prints and writes the session-specific durable state path, writes `metadata.json`, mirrors pane output to `pane.log`, records lifecycle events in `events.log`, bridges the inner `gjc` process to a public-safe `runtime-state.json`, writes normal-exit `final.json`, and optionally registers a Clawhip-style `tmux watch`.\n- `prompt.sh` sends a text or `@file` prompt only after the pane looks like a ready GJC TUI; if the tmux session vanished, it refuses injection and prints the durable metadata/log/final/events recovery paths plus the last pane-log excerpt.\n- `tail.sh` captures bounded pane output for readiness and acceptance checks, with durable metadata, pane-log, event-log, and final-status fallback when tmux vanished.\n- `harness-tmux-owner-start.sh` starts the GJC harness control plane with the RuntimeOwner resident inside tmux for dogfood/debug cases that need visible owner liveness.\n\nConfiguration is runtime-only:\n\n```sh\nexport GJC_BIN=/path/to/gjc                         # optional; defaults to command -v gjc\nexport GJC_SESSION_FLAGS=\"--model provider/model\"   # optional interactive gjc flags\nexport GJC_SESSION_ROUTER=clawhip                   # optional router binary\nexport GJC_SESSION_SKIP_ROUTER=1                    # skip router registration\nexport GJC_SESSION_STATE_DIR=/tmp/gjc-session-state # optional durable metadata/log root\nexport GJC_SESSION_LOG_SEARCH_ROOT=$HOME/Workspace  # optional tail/prompt fallback search root\nexport GJC_SESSION_STALE_MINUTES=60                 # router stale window\nexport GJC_SESSION_KEYWORDS=\"/skill:ralplan,Question\"\n```\n\nNo token, channel id, mention, workspace root, or private host path is hard-coded. Pass channel/mention values at invocation time when your router needs them.\n\n## Example flow\n\n```sh\n# 1. Prepare a dedicated worktree.\ngit -C /repo/gajae-code fetch origin dev\ngit -C /repo/gajae-code worktree add \\\n  /repo/worktrees/gajae-code-issue-905-ctrl-shift-enter-newline \\\n  -b issue-905-ctrl-shift-enter-newline origin/dev\n\n# 2. Start the routed visible session.\n./scripts/gjc-session/create.sh \\\n  gajae-code-issue-905-ctrl-shift-enter-newline \\\n  /repo/worktrees/gajae-code-issue-905-ctrl-shift-enter-newline \\\n  \"$CHANNEL_ID\" \\\n  \"$MENTION\"\n\n# 3. Confirm TUI readiness.\n./scripts/gjc-session/tail.sh gajae-code-issue-905-ctrl-shift-enter-newline 80\n\n# 4. Inject the task prompt.\n./scripts/gjc-session/prompt.sh \\\n  gajae-code-issue-905-ctrl-shift-enter-newline \\\n  @/tmp/issue-905-task.md\n\n# 5. Confirm real work started.\n./scripts/gjc-session/tail.sh gajae-code-issue-905-ctrl-shift-enter-newline 160\n```\n\n## Prompt shape\n\nImplementation prompt:\n\n```text\n/skill:ralplan\n\ngjc ultragoal fix issue #905 missed Ctrl+Shift+Enter newline case.\n\nRepo: Yeachan-Heo/gajae-code\nWorktree: /repo/worktrees/gajae-code-issue-905-ctrl-shift-enter-newline\nBranch: issue-905-ctrl-shift-enter-newline\nBase: dev\n\nScope:\n- inspect parser/key matching and packages/tui/src/components/editor.ts\n- add explicit ctrl+shift+enter newline handling\n- add focused tests for the reported terminal sequences\n- run targeted verification\n- commit, push, and open a PR to dev\n\nNon-goals:\n- no unrelated tmux/session/process changes\n- no synchronous filesystem, process, tmux, network, or durable writes in keystroke paths\n```\n\nReview prompt:\n\n```text\n/skill:ralplan\n\nReview PR #911 as a red-team-only merge gate.\nInspect origin/dev...HEAD, changed files, CI, and contract risks.\nLook for blockers, regressions, test gaps, and hidden user-facing drift.\nPost MERGE_READY or REQUEST_CHANGES with evidence. Do not merge.\n```\n\n## Acceptance checks\n\nAfter prompt delivery, require one of these before reporting that the session is working:\n\n- a tool call or file read in the pane,\n- an explicit plan or todo update,\n- a diff or test command,\n- a GitHub comment/review/PR URL,\n- a terminal verdict such as `MERGE_READY` or `REQUEST_CHANGES`.\n\nA prompt being visible in tmux scrollback is not acceptance by itself. If tmux disappears before terminal verdict, inspect the state path printed by `create.sh`: `metadata.json` identifies the worktree/session and links `runtime-state.json`; `runtime-state.json` contains public-safe GJC lifecycle state (`completed` / `errored`, timestamps, cwd/workdir, branch, session file, exit code/signal/error when available) without pane logs, prompts, transcripts, tokens, config, environment dumps, or raw tool output; `pane.log` contains the private mirrored transcript; `events.log` records launch/exit milestones; `final.json` is present when the wrapper observed `gjc` exit; and `vanished.json` is present when the external monitor observed the tmux session disappear. Use `tail.sh <session-name> [lines]` to surface these artifacts without a live tmux server.\n\n## Anti-patterns\n\n- Starting `gjc -p` for long-running visible repo work.\n- Launching from the canonical repo checkout instead of a task worktree.\n- Running a long GJC/tmux session under a short shell timeout that can SIGKILL the owner process.\n- Treating tmux process existence as proof that the prompt was accepted.\n- Restarting a vanished session without first checking its durable metadata, pane log, event log, and final status.\n- Hard-coding private channel ids, bot mentions, or router tokens into public GJC docs.\n- Using this visible-session pattern when Coordinator MCP turn state is available and sufficient.\n",
-	"gpt-5.6-codex-preset-benchmark.md": "# GPT-5.6 Codex preset benchmark\n\nThis report records descriptive local exact-edit evidence and the product judgments used to assign GPT-5.6 Sol, Terra, and Luna to GJC's built-in Codex-related model profiles.\n\n## Decision summary\n\nBuilt-in role assignments are product judgments. The selected TypeScript edit evidence below directly compares only bounded executor-style edits; it does not establish superiority, statistical significance, production reliability, or stability for any role.\n\n- **Luna high**: assigned to bounded execution in `codex-eco`.\n- **Terra high**: assigned to planning and lower-stakes criticism.\n- **Terra xhigh**: assigned to technically difficult execution and criticism in Medium, Pro, and combo profiles.\n- **Sol medium**: assigned to general orchestration for the `codex-medium` default agent.\n- **Sol high/xhigh/max**: assigned to architecture, high-stakes criticism, and Pro orchestration.\n\nThe edit benchmark does not measure default-agent interpretation, orchestration, explanation, or routing, and it does not measure planner, architect, or critic work. Those non-executor assignments are product judgments, not benchmark findings.\n\n## Environment\n\n- Date: 2026-07-11\n- GJC provider: local `layofflabs` OpenAI Responses-compatible endpoint\n- Models: `gpt-5.6-luna`, `gpt-5.6-terra`, `gpt-5.6-sol`\n- Benchmark: `packages/typescript-edit-benchmark`\n- Verification: exact expected-file comparison after formatting normalization\n- Required tools: at least one `read` and one `edit` call per successful sample\n- Guided edits: disabled\n- Attempts: one per sample\n\nThe local provider recorded zero cost. The amounts below are non-billing list-price estimates calculated from the listed rates; they are not provider charges or production-cost predictions.\n\n| Model | Input / 1M | Output / 1M |\n|---|---:|---:|\n| Luna | $1.00 | $6.00 |\n| Terra | $2.50 | $15.00 |\n| Sol | $5.00 | $30.00 |\n\n## Initial broad sample\n\nThe first pass used eight mutation tasks with one run per task:\n\n- multi-location identifier replacement\n- call-argument swap\n- early-return removal\n- `if`/`else` structural swap\n- named-import swap\n- duplicate-line disambiguation\n- off-by-one literal correction\n- optional-chain removal\n\n| Setup | Tasks passed | Avg time/run | Input tokens | Output tokens | Est. cost |\n|---|---:|---:|---:|---:|---:|\n| Luna high | 6/8 | 54.8s | 2.86M | 10.8K | $2.92 |\n| Luna xhigh | 7/8 | 31.2s | 784K | 6.6K | $0.82 |\n| Terra high | 7/8 | 51.1s | 1.13M | 5.9K | $2.92 |\n| Terra xhigh | 8/8 | 50.9s | 820K | 5.9K | $2.14 |\n| Sol medium | 6/8 | 30.1s | 376K | 4.3K | $2.01 |\n\nIn this eight-task, one-attempt-per-task sample, Terra xhigh recorded 8/8 verified edits. Luna xhigh recorded 7/8; one run per task does not establish stability.\n\n## Repeated selected-task sample\n\nThe selected pass ran four discriminating TypeScript edit tasks three times each, scheduling 12 samples per setup:\n\n1. Remove the intended early return from a file containing several similar returns.\n2. Swap the intended `if`/`else` branches without changing nearby equivalent logic.\n3. Correct one specific off-by-one value among several plausible candidates.\n4. Remove the intended optional chain without modifying similar occurrences.\n\nThe confirmation command shape was:\n\n```sh\nbun --cwd=packages/typescript-edit-benchmark run start \\\n  --model \"layofflabs/<model>\" \\\n  --thinking \"<effort>\" \\\n  --runs 3 \\\n  --task-concurrency 2 \\\n  --timeout 180000 \\\n  --max-turns 40 \\\n  --tasks \"structural-remove-early-return-003,structural-swap-if-else-004,literal-off-by-one-003,access-remove-optional-chain-004\" \\\n  --require-read-tool-call \\\n  --require-edit-tool-call \\\n  --format json\n```\n\n| Setup | Verified edits / recorded runs | Rate | Avg time | Input tokens | Output tokens | Est. list-price cost | Est. cost / verified edit |\n|---|---:|---:|---:|---:|---:|---:|---:|\n| Luna high | 8/12 | 66.7% | 75.2s | 3.61M | 18.9K | $3.73 | $0.47 |\n| Luna xhigh | 9/12 | 75.0% | 80.5s | 6.60M | 25.0K | $6.75 | $0.75 |\n| Terra high | 6/11 | 54.5% | 58.9s | 572K | 10.0K | $1.58 | $0.26 |\n| Terra xhigh | 9/12 | 75.0% | 57.3s | 1.86M | 14.2K | $4.86 | $0.54 |\n| Sol medium | 4/12 | 33.3% | 46.3s | 558K | 10.1K | $3.09 | $0.77 |\n\nTerra high had one transport/ghost failure, so it has 11 recorded runs rather than 12 scheduled samples; its rate and cost per verified edit use those recorded results.\n\n## Findings\n\n### Terra xhigh's selected-task executor result\n\nAcross these four selected TypeScript edit tasks under the documented local setup, Terra xhigh and Luna xhigh each recorded 9/12 verified edits. Terra xhigh's reported totals were 72% fewer input tokens, 43% fewer output tokens, 28% less estimated list-price cost, and 29% less time than Luna xhigh. These descriptive results inform, but do not prove, the Terra xhigh executor assignment.\n\n### Luna remains useful, but not as the premium executor\n\nLuna xhigh recorded 7/8 in the broad sample and 9/12 in the selected-task sample. Luna high remains the Eco executor as a product judgment for that preset's lower-priced-family-member trade-off; these local runs do not establish a capability ceiling or production behavior.\n\n### Terra high's product assignment\n\nTerra high recorded 6/11 verified edits after one transport/ghost failure in the selected-task sample. Its planning and lower-stakes critic assignments are product judgments; this edit benchmark does not measure those roles.\n\n### Sol medium's product assignment\n\nSol medium recorded 4/12 verified edits in the selected-task sample and was faster with fewer reported input tokens than the other listed xhigh setups. Its `codex-medium` default-agent assignment and the Sol-family architecture assignments are product judgments because the benchmark does not measure those broader roles.\n\n### Higher effort is not automatically cheaper\n\nThe selected-task data show that Luna xhigh used more reported tokens than Luna high in this local setup. They do not establish a general cost rule for thinking effort; effort selection remains a product decision informed by model tier and role shape.\n\n## Resulting built-in profiles\n\n| Profile | Default | Executor | Planner | Critic | Architect |\n|---|---|---|---|---|---|\n| `codex-eco` | Terra high | Luna high | Terra medium | Terra high | Sol medium |\n| `codex-medium` | Sol medium | Terra xhigh | Terra high | Terra xhigh | Sol high |\n| `codex-pro` | Sol xhigh | Terra xhigh | Sol high | Sol xhigh | Sol max |\n| `opus-codex` | Claude Opus xhigh | Terra xhigh | Terra high | Terra xhigh | Sol xhigh |\n| `codex-opencodego` | Terra xhigh | DeepSeek V4 Pro | Kimi K2.6 | MiMo 2.5 Pro | Sol xhigh |\n| `fable-opus-codex` | Claude Fable high | Terra xhigh | Claude Opus medium | Claude Opus high | Sol xhigh |\n\n## Limitations\n\n- The benchmark measures four selected precise TypeScript source mutations in the repeated sample, not full-session planning, architecture, criticism, or default-agent quality.\n- The corpus is small and intentionally adversarial; the results are descriptive, not statistically significant or a proof of general superiority, production reliability, or stability.\n- Samples used a local OpenAI-compatible provider rather than OpenAI's production endpoint.\n- Terra high has 11 recorded runs because one of 12 scheduled samples ended in a transport/ghost failure.\n- Token accounting reflects the local transport and benchmark context construction. The provider recorded zero cost; displayed costs are rounded list-price estimates, not billing predictions.\n- Model behavior can change as provider snapshots are updated.\n\nThe raw JSON reports and conversation dumps were generated under `runs/gpt-5.6-local-2026-07-11/` and `runs/gpt-5.6-confirmation-2026-07-11/`, but are not committed. The committed tables support the displayed denominators and rounded comparisons, not reconstruction of unrounded token totals or list-price estimates.\n",
+	"gpt-5.6-codex-preset-benchmark.md": "# GPT-5.6 Codex preset benchmark\n\nThis report records descriptive local exact-edit evidence and the product judgments used to assign GPT-5.6 Sol, Terra, and Luna to GJC's built-in Codex-related model profiles.\n\n## Decision summary\n\nBuilt-in role assignments are product judgments. The selected TypeScript edit evidence below directly compares only bounded executor-style edits; it does not establish superiority, statistical significance, production reliability, or stability for any role.\n\n- **Eco**: `terra:low` default, `luna:low` executor, `luna:high` planner, `terra:xhigh` critic, and `terra:high` architect.\n- **Medium**: `sol:low` default, `terra:low` executor, `terra:high` planner, `sol:xhigh` critic, and `sol:high` architect.\n- **Pro**: `sol:medium` default, `terra:medium` executor, `sol:high` planner, `sol:max` critic, and `sol:xhigh` architect.\n- **Combos**: `opus-codex` uses the Medium Codex executor, critic, and architect roles, with the durable `anthropic/claude-sonnet-5` planner override; `codex-opencodego` uses Medium Codex default and architect roles; and `fable-opus-codex` uses Pro Codex executor and architect roles with `anthropic/claude-opus-4-8:medium` as planner.\n\nThe edit benchmark does not measure default-agent interpretation, orchestration, explanation, or routing, and it does not measure planner, architect, or critic work. Those non-executor assignments are product judgments, not benchmark findings.\n\n## Environment\n\n- Date: 2026-07-11\n- GJC provider: local `layofflabs` OpenAI Responses-compatible endpoint\n- Models: `gpt-5.6-luna`, `gpt-5.6-terra`, `gpt-5.6-sol`\n- Benchmark: `packages/typescript-edit-benchmark`\n- Verification: exact expected-file comparison after formatting normalization\n- Required tools: at least one `read` and one `edit` call per successful sample\n- Guided edits: disabled\n- Attempts: one per sample\n\nThe local provider recorded zero cost. The amounts below are non-billing list-price estimates calculated from the listed rates; they are not provider charges or production-cost predictions.\n\n| Model | Input / 1M | Output / 1M |\n|---|---:|---:|\n| Luna | $1.00 | $6.00 |\n| Terra | $2.50 | $15.00 |\n| Sol | $5.00 | $30.00 |\n\n## Initial broad sample\n\nThe first pass used eight mutation tasks with one run per task:\n\n- multi-location identifier replacement\n- call-argument swap\n- early-return removal\n- `if`/`else` structural swap\n- named-import swap\n- duplicate-line disambiguation\n- off-by-one literal correction\n- optional-chain removal\n\n| Setup | Tasks passed | Avg time/run | Input tokens | Output tokens | Est. cost |\n|---|---:|---:|---:|---:|---:|\n| Luna high | 6/8 | 54.8s | 2.86M | 10.8K | $2.92 |\n| Luna xhigh | 7/8 | 31.2s | 784K | 6.6K | $0.82 |\n| Terra high | 7/8 | 51.1s | 1.13M | 5.9K | $2.92 |\n| Terra xhigh | 8/8 | 50.9s | 820K | 5.9K | $2.14 |\n| Sol medium | 6/8 | 30.1s | 376K | 4.3K | $2.01 |\n\nIn this eight-task, one-attempt-per-task sample, Terra xhigh recorded 8/8 verified edits. Luna xhigh recorded 7/8; one run per task does not establish stability.\n\n## Repeated selected-task sample\n\nThe selected pass ran four discriminating TypeScript edit tasks three times each, scheduling 12 samples per setup:\n\n1. Remove the intended early return from a file containing several similar returns.\n2. Swap the intended `if`/`else` branches without changing nearby equivalent logic.\n3. Correct one specific off-by-one value among several plausible candidates.\n4. Remove the intended optional chain without modifying similar occurrences.\n\nThe confirmation command shape was:\n\n```sh\nbun --cwd=packages/typescript-edit-benchmark run start \\\n  --model \"layofflabs/<model>\" \\\n  --thinking \"<effort>\" \\\n  --runs 3 \\\n  --task-concurrency 2 \\\n  --timeout 180000 \\\n  --max-turns 40 \\\n  --tasks \"structural-remove-early-return-003,structural-swap-if-else-004,literal-off-by-one-003,access-remove-optional-chain-004\" \\\n  --require-read-tool-call \\\n  --require-edit-tool-call \\\n  --format json\n```\n\n| Setup | Verified edits / recorded runs | Rate | Avg time | Input tokens | Output tokens | Est. list-price cost | Est. cost / verified edit |\n|---|---:|---:|---:|---:|---:|---:|---:|\n| Luna high | 8/12 | 66.7% | 75.2s | 3.61M | 18.9K | $3.73 | $0.47 |\n| Luna xhigh | 9/12 | 75.0% | 80.5s | 6.60M | 25.0K | $6.75 | $0.75 |\n| Terra high | 6/11 | 54.5% | 58.9s | 572K | 10.0K | $1.58 | $0.26 |\n| Terra xhigh | 9/12 | 75.0% | 57.3s | 1.86M | 14.2K | $4.86 | $0.54 |\n| Sol medium | 4/12 | 33.3% | 46.3s | 558K | 10.1K | $3.09 | $0.77 |\n\nTerra high had one transport/ghost failure, so it has 11 recorded runs rather than 12 scheduled samples; its rate and cost per verified edit use those recorded results.\n\n## Findings\n\n### Terra xhigh's selected-task executor result\n\nAcross these four selected TypeScript edit tasks under the documented local setup, Terra xhigh and Luna xhigh each recorded 9/12 verified edits. Terra xhigh's reported totals were 72% fewer input tokens, 43% fewer output tokens, 28% less estimated list-price cost, and 29% less time than Luna xhigh. These descriptive results inform, but do not prove, the Terra xhigh executor assignment.\n\n### Luna remains useful, but not as the premium executor\n\nLuna xhigh recorded 7/8 in the broad sample and 9/12 in the selected-task sample. Luna high remains the Eco executor as a product judgment for that preset's lower-priced-family-member trade-off; these local runs do not establish a capability ceiling or production behavior.\n\n### Terra high's product assignment\n\nTerra high recorded 6/11 verified edits after one transport/ghost failure in the selected-task sample. Its planning and lower-stakes critic assignments are product judgments; this edit benchmark does not measure those roles.\n\n### Sol medium's product assignment\n\nSol medium recorded 4/12 verified edits in the selected-task sample and was faster with fewer reported input tokens than the other listed xhigh setups. Its `codex-medium` default-agent assignment and the Sol-family architecture assignments are product judgments because the benchmark does not measure those broader roles.\n\n### Higher effort is not automatically cheaper\n\nThe selected-task data show that Luna xhigh used more reported tokens than Luna high in this local setup. They do not establish a general cost rule for thinking effort; effort selection remains a product decision informed by model tier and role shape.\n\n## Resulting built-in profiles\n\n| Profile | Default | Executor | Planner | Critic | Architect |\n|---|---|---|---|---|---|\n| `codex-eco` | `openai-codex/gpt-5.6-terra:low` | `openai-codex/gpt-5.6-luna:low` | `openai-codex/gpt-5.6-luna:high` | `openai-codex/gpt-5.6-terra:xhigh` | `openai-codex/gpt-5.6-terra:high` |\n| `codex-medium` | `openai-codex/gpt-5.6-sol:low` | `openai-codex/gpt-5.6-terra:low` | `openai-codex/gpt-5.6-terra:high` | `openai-codex/gpt-5.6-sol:xhigh` | `openai-codex/gpt-5.6-sol:high` |\n| `codex-pro` | `openai-codex/gpt-5.6-sol:medium` | `openai-codex/gpt-5.6-terra:medium` | `openai-codex/gpt-5.6-sol:high` | `openai-codex/gpt-5.6-sol:max` | `openai-codex/gpt-5.6-sol:xhigh` |\n| `opus-codex` | `anthropic/claude-opus-4-8:xhigh` | `openai-codex/gpt-5.6-terra:low` | `anthropic/claude-sonnet-5` | `openai-codex/gpt-5.6-sol:xhigh` | `openai-codex/gpt-5.6-sol:high` |\n| `codex-opencodego` | `openai-codex/gpt-5.6-sol:low` | `opencode-go/deepseek-v4-pro` | `opencode-go/kimi-k2.6` | `opencode-go/mimo-v2.5-pro` | `openai-codex/gpt-5.6-sol:high` |\n| `fable-opus-codex` | `anthropic/claude-fable-5:high` | `openai-codex/gpt-5.6-terra:medium` | `anthropic/claude-opus-4-8:medium` | `anthropic/claude-opus-4-8:high` | `openai-codex/gpt-5.6-sol:xhigh` |\n\n## Limitations\n\n- The benchmark measures four selected precise TypeScript source mutations in the repeated sample, not full-session planning, architecture, criticism, or default-agent quality.\n- The corpus is small and intentionally adversarial; the results are descriptive, not statistically significant or a proof of general superiority, production reliability, or stability.\n- Samples used a local OpenAI-compatible provider rather than OpenAI's production endpoint.\n- Terra high has 11 recorded runs because one of 12 scheduled samples ended in a transport/ghost failure.\n- Token accounting reflects the local transport and benchmark context construction. The provider recorded zero cost; displayed costs are rounded list-price estimates, not billing predictions.\n- Model behavior can change as provider snapshots are updated.\n\nThe raw JSON reports and conversation dumps were generated under `runs/gpt-5.6-local-2026-07-11/` and `runs/gpt-5.6-confirmation-2026-07-11/`, but are not committed. The committed tables support the displayed denominators and rounded comparisons, not reconstruction of unrounded token totals or list-price estimates.\n",
 	"grok-build-provider-design.md": "# Grok Build provider design\n\n## Status\n\nProposal for maintainer design review. This document intentionally does not add a bundled provider implementation. It records the product/API decisions that must be accepted before any Grok Build implementation PR should land.\n\nThis is not an authorization claim for xAI endpoints, not a final naming decision, not approval for a bundled-loading exception, and not trademark/display-name approval. Those items require explicit owner sign-off before implementation.\n\n## Required owner sign-off gates\n\nImplementation should remain blocked until the owner signs off on these gates:\n\n1. **Authorized use / ToS** — confirm that GJC may use `cli-chat-proxy.grok.com` and the xAI CLI OAuth public client from a third-party tool. A public OAuth client id is not proof that this use is authorized.\n2. **Bundled-loading trust boundary** — confirm whether a source-controlled bundled provider may load even when ordinary user extension discovery is disabled.\n3. **Public selector naming** — choose the stable provider selector prefix: `grok-cli`, `grok-build`, or another owner-selected id.\n4. **Trademark/display-name** — confirm whether GJC may present the provider/profile using `Grok Build` or should use a more neutral owner-approved label.\n\nIf gate 1 is not accepted, the Grok Build provider implementation should not ship against `cli-chat-proxy.grok.com`. The fallback direction would be a documented user-supplied xAI/API-key provider or a different officially authorized integration path.\n\n## Problem\n\nGJC can load third-party extensions, but the first-run interactive path needs a maintainer-owned decision before a bundled Grok Build provider can be accepted. The desired product flow is:\n\n```text\ngjc -> /login -> OAuth -> Grok Build -> browser xAI login -> /model -> <provider-id>/grok-composer-2.5-fast\n```\n\nThe previously proposed implementation touched bundled extension loading, OAuth registration, model profiles, vendor code, usage reporting, and tests in one PR. That is too much surface for review without first agreeing on the provider contract and the owner sign-off gates above.\n\n## Goals\n\n- Keep Grok Build, if accepted, as a bundled provider extension rather than a workflow skill.\n- Preserve the existing four bundled workflow skills and four role agents.\n- Define the `/login` OAuth contract for an owner-approved display name, with `Grok Build` only as a candidate label.\n- Define the `/model` contract for `grok-composer-2.5-fast` without committing to the final selector prefix before owner sign-off.\n- Define the guardrails for any bundled provider that loads while ordinary extension discovery is disabled.\n- Keep credentials in the existing auth storage path; no tokens or user env values are checked into the repo.\n- Keep implementation PRs small enough for independent review, rejection, or rollback.\n\n## Non-goals\n\n- No new workflow command or `/skill` surface.\n- No automatic installation from npm or remote code at runtime.\n- No direct `packages/ai/src/models.json` edits.\n- No broad model-profile reshuffle.\n- No provider-specific secrets in source.\n- No claim that xAI has authorized this endpoint/client usage without owner review.\n\n## Candidate provider contract\n\nThese are candidate values for owner review, not final commitments:\n\n| Field | Candidate value | Decision status | Notes |\n| --- | --- | --- | --- |\n| Public provider id | `grok-cli` or `grok-build` | **Owner decision required** | See naming section below. |\n| Display name | `Grok Build` or owner-selected label | **Owner decision required** | Name shown in `/login` and UI surfaces; see trademark/display-name section below. |\n| Default model id | `grok-composer-2.5-fast` | Proposed | Full selector depends on final provider id. |\n| Secondary model id | `grok-build` | Proposed | Candidate for executor/architect roles if a profile is accepted. |\n| Base URL | `https://cli-chat-proxy.grok.com/v1` | **Authorized-use sign-off required** | Undocumented/private-looking endpoint; do not ship without owner approval. |\n| OAuth issuer | `https://auth.x.ai` | **Authorized-use sign-off required** | OIDC discovery must validate xAI-owned HTTPS endpoints. |\n| OAuth callback | loopback `127.0.0.1` | Proposed | Uses PKCE + state validation. |\n| API adapter | `grok-cli-responses` | Proposed internal name | Provider-specific stream adapter; not a new generic API shape. |\n| Env bypass | `GROK_CLI_OAUTH_TOKEN` | Optional follow-up | Local bypass only; no refresh or discovery guarantees. |\n\n## Authorized-use and ToS caveat\n\n`cli-chat-proxy.grok.com` and the xAI CLI OAuth public client appear to be designed for xAI/Grok CLI traffic. Reusing them from GJC may be technically possible but still unauthorized or contrary to xAI terms.\n\nBefore implementation, the owner should explicitly decide one of:\n\n- **Accept** — proceed with this integration after reviewing the legal/product risk.\n- **Defer** — keep this design document only; no code ships until authorization is clarified.\n- **Reject** — do not integrate against `cli-chat-proxy.grok.com`; use only an official public API path.\n\nImplementation PRs must not describe the public client id as a secret, but they also must not present it as authorization. Tests should avoid real tokens and should not require an xAI account.\n\n## Trademark/display-name caveat\n\n`Grok` and `xAI` are third-party marks. `Grok Build` may also imply an official xAI/Grok product relationship even when the integration is third-party. Before implementation, the owner should explicitly choose one of:\n\n- **Use `Grok Build`** — acceptable as the user-facing provider/profile label after trademark/product-risk review.\n- **Use a neutral label** — for example `xAI Grok`, `Grok OAuth`, or another owner-selected name that avoids implying official endorsement.\n- **Avoid built-in branding** — keep any Grok-specific naming only in user-provided configuration until authorization/branding is clarified.\n\nImplementation PRs should avoid lock-in language such as \"official\" unless there is explicit authorization. UI labels, profile names, docs, tests, and screenshots must all use the owner-approved label consistently.\n\n## OAuth behavior\n\nIf authorized-use is accepted, the OAuth implementation should use the existing custom OAuth provider path:\n\n1. The chosen provider id registers an OAuth provider using the owner-approved display name.\n2. `/login` calls the existing auth storage login path for that provider.\n3. The provider opens an xAI authorization URL using OIDC discovery, PKCE, `state`, and a loopback callback.\n4. The callback exchanges the authorization code for access and refresh tokens.\n5. Credentials are stored by the existing auth storage code path.\n6. Refresh uses the stored refresh token and validates the token endpoint origin.\n\nSecurity constraints:\n\n- OIDC `authorization_endpoint` and `token_endpoint` must be HTTPS and under owner-approved xAI hosts.\n- The callback server binds to loopback by default.\n- The callback must reject state mismatches.\n- Access and refresh tokens must not be logged, rendered, committed, or included in tests.\n- Error messages may include status and provider error text, but not credential values.\n- Env overrides for base URL, scope, callback host, or client id must be treated as local developer/debug escape hatches, not default product behavior.\n\n## Bundled-loading trust boundary\n\nA bundled provider is different from ordinary user extension discovery, but loading it while `disableExtensionDiscovery: true` still expands the bootstrap trust boundary. Owner sign-off is required before implementation.\n\nMinimum guardrails if accepted:\n\n- Load only source-controlled, maintainer-reviewed bundled provider paths.\n- Use a static allowlist or exported enumerator; never scan arbitrary user directories for this path.\n- Do not install, fetch, or resolve remote package code at runtime.\n- Keep ordinary user extension discovery disabled when `disableExtensionDiscovery: true`; the exception is only for bundled provider defaults.\n- Add tests proving bundled providers load before model selection and caller-supplied `additionalExtensionPaths` still coexist.\n- Keep this bootstrap change separate from the Grok vendor implementation so it can be reviewed independently.\n\nAlternatives the owner may choose:\n\n- Do not load bundled providers when extension discovery is disabled; require explicit setup/defaults install.\n- Gate bundled provider loading behind a setting or compile-time default.\n- Allow bundled loading only in packaged builds, not arbitrary source checkouts.\n\n## Provider selector naming\n\nThe selector prefix is a stable user-facing contract and must be chosen before implementation.\n\n| Option | Example selector | Pros | Cons |\n| --- | --- | --- | --- |\n| `grok-cli` | `grok-cli/grok-composer-2.5-fast` | Matches the upstream CLI/proxy lineage and existing prototype. | User-facing name is less aligned with `Grok Build`; may expose implementation detail. |\n| `grok-build` | `grok-build/grok-composer-2.5-fast` | Matches UI label and requested product wording. | Diverges from existing prototype and env names; migration needed if prototypes used `grok-cli`. |\n| Owner-selected third id | `<id>/grok-composer-2.5-fast` | Lets maintainers align with broader provider taxonomy. | Requires updating all docs/tests before implementation. |\n\nUntil this is decided, implementation docs and PRs should use `<provider-id>` when describing the public selector. Internal adapter names may still use `grok-cli-responses` if maintainers accept that as an implementation detail.\n\n## Model/profile behavior\n\nModel registration should be provider-owned. If accepted, the provider should register at least:\n\n- `grok-composer-2.5-fast`\n- `grok-build`\n\nA built-in profile is optional and should be reviewed separately. If accepted, a candidate profile is:\n\n```text\ngrok-pro.default   -> <provider-id>/grok-composer-2.5-fast\ngrok-pro.planner   -> <provider-id>/grok-composer-2.5-fast\ngrok-pro.critic    -> <provider-id>/grok-composer-2.5-fast\ngrok-pro.executor  -> <provider-id>/grok-build\ngrok-pro.architect -> <provider-id>/grok-build\n```\n\nIf maintainers prefer not to add a built-in profile, the provider can still satisfy the core `/login` and `/model` flow through direct model selection.\n\n## Usage reporting behavior\n\nUsage reporting should be an optional follow-up after login/model support lands:\n\n- Provider id: the owner-selected `<provider-id>`.\n- Fetches usage with the effective OAuth access token.\n- Returns `null` when no token is available.\n- Does not require the usage provider for chat/model selection to work.\n- Should be skipped entirely if the authorized-use gate is not accepted.\n\n## Staged PR plan\n\n### PR 1: this design document\n\nPurpose: agree on caveats, owner sign-off gates, provider id, OAuth contract, bundled-loading trust boundary, model selector, security boundaries, and implementation split.\n\n### PR 2: bundled provider bootstrap contract\n\nSmall core change only, after owner sign-off on the bundled-loading gate:\n\n- Add a maintainer-owned way to enumerate bundled provider extension paths.\n- Load those paths during session/bootstrap only under the accepted guardrails.\n- Add tests proving bundled providers and caller-supplied extension paths coexist.\n\nNo Grok vendor implementation in this PR.\n\n### PR 3: Grok Build provider extension\n\nProvider implementation only, after owner sign-off on authorized use, public selector naming, and trademark/display-name:\n\n- Add bundled Grok Build provider source.\n- Register the chosen provider id, OAuth provider, and models.\n- Include sanitize and provider-specific stream handling.\n- Test `/login` provider registration and `grok-composer-2.5-fast` model availability.\n\n### PR 4: profile and model defaults\n\nOptional product-surface PR:\n\n- Add `grok-pro` only if maintainers accept a built-in profile.\n- Add model profile catalog tests.\n\n### PR 5: usage reporting\n\nOptional observability PR:\n\n- Add usage provider for the owner-selected provider id.\n- Add focused usage tests.\n\n## Acceptance criteria for the implementation series\n\n- Owner sign-off is recorded for authorized use, bundled loading, selector naming, and trademark/display-name before implementation lands.\n- Fresh checkout test proves `createAgentSession` registers the bundled provider under the accepted bootstrap rules.\n- `/login` includes the owner-approved display name for the owner-selected provider id.\n- `/model` includes `<provider-id>/grok-composer-2.5-fast`.\n- A real OAuth URL redirects to the owner-approved xAI account login page.\n- Third-party extension paths still load alongside bundled providers when configured.\n- Token values never appear in tests, logs, checked-in docs, or git history.\n\n## Open maintainer decisions\n\n- Is using `cli-chat-proxy.grok.com` plus the xAI CLI OAuth client from GJC authorized and acceptable for this project?\n- Should bundled provider defaults load while `disableExtensionDiscovery: true`, and under which guardrails?\n- Should the final public provider id be `grok-cli`, `grok-build`, or another id?\n- May GJC use `Grok Build` as the display/profile name, or should the integration use a neutral owner-selected label?\n- Should `grok-pro` be a built-in profile or documented as a user profile?\n- Should usage reporting be included in the initial provider PR or kept as a separate follow-up?",
 	"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/extensibility/slash-commands.ts`](../packages/coding-agent/src/extensibility/slash-commands.ts)\n\n## Trigger path\n\n1. `/handoff` is declared in builtin slash command metadata (`slash-commands.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 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 `completeSimple(...)` directly:\n\n```ts\nawait completeSimple(\n  model,\n  {\n    systemPrompt,\n    messages: requestMessages,\n    tools,\n  },\n  {\n    apiKey,\n    signal,\n    reasoning: Effort.High,\n    toolChoice: \"none\",\n    initiatorOverride,\n    metadata,\n  },\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 state 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(\"handoff\", handoffContent, true, undefined, \"agent\");\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 `completeSimple(...)` request with live system prompt, tools, message history, 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",
 	"hermes-mcp-bridge.md": "# Coordinator MCP bridge\n\nGJC exposes a native outward MCP bridge for external coordinators:\n\n```bash\ngjc mcp-serve coordinator\n```\n\n`gjc mcp-serve hermes` is accepted as a compatibility alias for the same coordinator bridge.\n\nThe bridge is intentionally separate from GJC's client-side MCP runtime. It lets an external coordinator list sessions, start worktree/tmux-oriented sessions, queue bounded follow-up prompts, read status/tail/artifacts, handle structured questions, and write coordination reports without scraping terminal scrollback.\n\n## Core contract and adapters\n\nThe coordinator bridge is intentionally a core contract with multiple adapters, not an MCP-only or Hermes-only product direction. Hermes is one compatibility preset, not a privileged integration mode:\n\n- `packages/coding-agent/src/coordinator/contract.ts` owns transport-neutral server metadata and tool names.\n- `gjc mcp-serve coordinator` is the outward MCP adapter for external agents.\n- `gjc coordinator` is the read-only CLI/debug adapter for humans and scripts that need to inspect the same contract without starting MCP transport.\n- `gjc setup hermes` is the compatibility setup adapter that renders coordinator config and operator guidance.\n\nFuture session, turn, question, artifact, and report behavior should move toward shared coordinator core services that both MCP and CLI adapters call instead of duplicating transport-specific logic.\n\n## Coordinator setup adapter\n\nUse `gjc setup hermes` to render or install a portable MCP setup package for any controller that accepts Hermes-compatible MCP config:\n\n```bash\ngjc setup hermes --root /path/to/repo --profile my-bot --repo gajae-code\n```\n\nThe default mode is render-only and writes no files. To install into a Hermes profile:\n\n```bash\ngjc setup hermes \\\n  --root /path/to/repo \\\n  --profile my-bot \\\n  --repo gajae-code \\\n  --mutation sessions,questions,reports \\\n  --profile-dir /path/to/hermes/profile \\\n  --install\n```\n\nThe generated setup is model-agnostic and worktree-isolated. By default it renders `GJC_COORDINATOR_MCP_SESSION_COMMAND` as `gjc --worktree`, so spawned sessions launch inside a GJC-managed sibling worktree while GJC still records the original repo as the project identity for tmux/session resume. Users who need a stable named branch can set `--worktree-name`; users who need a specific local wrapper, dev checkout, or provider/model can opt in explicitly:\n\n```bash\ngjc setup hermes \\\n  --root /path/to/repo \\\n  --worktree-name hermes-gajae-code\n```\n\n```bash\ngjc setup hermes \\\n  --root /path/to/repo \\\n  --session-command \"gjc --worktree hermes-custom --model <provider/model>\"\n```\n\nProvider/model examples are examples only; GJC does not hard-code GPT, Anthropic, or any other provider as the Hermes bridge default.\n\nRun a non-mutating setup smoke check with:\n\n```bash\ngjc setup hermes --root /path/to/repo --smoke\n```\n\nSmoke verifies the MCP server/tool contract. It does not call a downstream LLM and does not validate provider credentials.\n\n\n## Safety model\n\nThe bridge is read-only and fail-closed by default.\n\nRequired root allowlist:\n\n```bash\nexport GJC_COORDINATOR_MCP_WORKDIR_ROOTS=\"/path/to/repo:/path/to/worktrees\"\n```\n\nMutating tools require both startup opt-in and per-call consent:\n\n```bash\nexport GJC_COORDINATOR_MCP_MUTATIONS=\"sessions,questions,reports\"\n```\n\nEvery mutating MCP call must also include `allow_mutation: true`. Missing startup opt-in or missing per-call consent returns an error instead of falling back to shell or terminal relay.\n\nReal tmux/GJC actuation uses the configured GJC-compatible session command. `gjc setup hermes` writes this as `gjc --worktree` by default so GJC owns worktree creation and resume identity:\n\n```bash\nexport GJC_COORDINATOR_MCP_SESSION_COMMAND=\"gjc --worktree\"\n```\n\nWith that command configured, `gjc_coordinator_start_session` launches a detached tmux session, `gjc_coordinator_send_prompt` creates a durable turn and sends input to that pane, `gjc_coordinator_read_coordination_status` returns a canonical polling snapshot for sessions, session states, turns, questions, reports, and bounded event summaries, and `gjc_coordinator_read_tail` reads bounded advisory pane output. Tmux tail parsing is not the completion source of truth; turn completion comes from explicit durable turn state such as runtime session state or `gjc_coordinator_report_status`.\n\nFor resume safety, prefer the generated GJC-native worktree command over creating a git worktree in Hermes itself. GJC's launch path records the original repo as the project identity while running in the worktree, so session listing/resume can still group the session under the source project. If Hermes creates and later deletes an unmanaged worktree, a saved session may still exist but its cwd can be gone.\n\nWhen an operator needs the session to stay visible in a routed tmux pane (for example a Clawhip/Hermes/OpenClaw channel that watches stale sessions and accepts follow-up prompts), use the documented visible-session fallback instead of inventing a private terminal protocol: [`docs/gjc-session-clawhip-routing.md`](./gjc-session-clawhip-routing.md). It keeps the same worktree isolation discipline while making the router, not GJC internals, own channel ids, mentions, and notification policy.\n\nArtifact reads are canonicalized, symlink escapes are rejected, and returned content is byte-capped by `GJC_COORDINATOR_MCP_ARTIFACT_BYTE_CAP`.\n\n`gjc setup hermes` renders `GJC_COORDINATOR_MCP_WORKDIR_ROOTS` with the host platform path delimiter (`:` on POSIX, `;` on Windows). Manual configs should prefer the same encoding.\n\n## Optional namespace\n\nUse namespace variables to prevent cross-profile or cross-repo enumeration:\n\n```bash\nexport GJC_COORDINATOR_MCP_PROFILE=\"team-a\"\nexport GJC_COORDINATOR_MCP_REPO=\"gajae-code\"\n```\n\nMissing namespace never widens into global session enumeration.\n\n## Tool surface\n\nRead tools:\n\n- `gjc_coordinator_list_sessions`\n- `gjc_coordinator_read_status`\n- `gjc_coordinator_read_tail`\n- `gjc_coordinator_list_questions`\n- `gjc_coordinator_list_artifacts`\n- `gjc_coordinator_read_artifact`\n- `gjc_coordinator_read_coordination_status`\n- `gjc_coordinator_read_turn`\n- `gjc_coordinator_await_turn`\n- `gjc_coordinator_watch_events`\n\n\nMutating tools:\n\n- `gjc_coordinator_start_session`\n- `gjc_coordinator_register_session`\n- `gjc_coordinator_send_prompt`\n- `gjc_coordinator_submit_question_answer`\n- `gjc_coordinator_report_status`\n- `gjc_delegate_plan`\n- `gjc_delegate_execute`\n- `gjc_delegate_team`\n\nThe `gjc_delegate_*` tools are high-level, session-level delegation: each starts (or reuses) a session and sends one workflow-tagged turn that runs `/skill:ralplan`, `/skill:ultragoal`, or `/skill:team` to completion, returning a durable `turn_id`, status, and artifact references. They use the same `sessions` mutation class and fail-closed workdir gating as `gjc_coordinator_start_session`, and emit a `delegation.started` event. Pass `cwd` and `task`; set `allow_mutation: true` only with startup mutation opt-in plus per-call consent. Optionally pass `mpreset` (same semantics as `gjc --mpreset <profile>`) to `gjc_coordinator_start_session` or a delegate tool to authoritatively activate a GJC model profile when starting a fresh session — it is resolved through the merged built-in/custom profile registry, applied from the first turn, and surfaced in status; unknown names are rejected with the available-profile listing, and reusing a session with a conflicting `mpreset` fails with `mpreset_conflict`. This is distinct from the advisory `model` prompt hint. Prefer these over manual `start_session` + `send_prompt` when delegating a whole workflow.\n\n\n`gjc_coordinator_register_session` registers an existing visible tmux-backed GJC pane as the coordinator-authoritative session. Use it when an operator has already launched a visible terminal/tmux lane and the external coordinator must send prompts to that same pane instead of creating a hidden `gjc-coordinator-*` session. The tool validates the workdir allowlist, safe session/target tokens, and tmux target liveness before writing session state.\n## Turn orchestration flow\n\nExternal coordinators should treat turns, not terminal scrollback, as the unit of work:\n\n1. Call `gjc_coordinator_start_session` with `allow_mutation: true`.\n2. Call `gjc_coordinator_send_prompt` with `allow_mutation: true`.\n3. Store the returned `turn_id`.\n4. Poll `gjc_coordinator_read_turn`, or call bounded `gjc_coordinator_await_turn`, until the turn is terminal.\n5. If `gjc_coordinator_list_questions` shows a question for that turn, answer with `gjc_coordinator_submit_question_answer`.\n6. Use `gjc_coordinator_report_status` with `session_id` and `turn_id` to write explicit completion/failure evidence.\n   Use `status: \"cancelled\"` for coordinator-policy cancellation, and `status: \"failed\"` plus `blocker` for provider/tool/task failures.\n\n`gjc_coordinator_send_prompt` preserves the legacy `queued` and `delivered` fields and adds turn fields:\n\n```json\n{\n  \"ok\": true,\n  \"session_id\": \"gjc-coordinator-demo\",\n  \"turn_id\": \"turn-00000000-0000-0000-0000-000000000000\",\n  \"active_turn_id\": \"turn-00000000-0000-0000-0000-000000000000\",\n  \"status\": \"active\",\n  \"queued\": false,\n  \"delivered\": true\n}\n```\n\nA session may have only one active turn by default. A second prompt is rejected with `active_turn_exists` unless the caller explicitly passes `queue: true` or `force: true`. Queued turns are durable and the next queued turn is promoted when the active turn reaches a terminal `gjc_coordinator_report_status`. Force supersedes the previous active turn and audits that state in the turn journal.\nCoordinator cancellation is recorded through `gjc_coordinator_report_status` with terminal `status: \"cancelled\"`; this updates durable turn state but does not kill the underlying tmux process. If the correct policy is replacement work rather than cancellation, send the replacement prompt with `force: true` so the previous active turn is superseded and audited.\n\n`gjc_coordinator_read_turn` returns the authoritative durable turn plus advisory pane status:\n\n```json\n{\n  \"ok\": true,\n  \"turn\": {\n    \"schema_version\": 1,\n    \"turn_id\": \"turn-00000000-0000-0000-0000-000000000000\",\n    \"session_id\": \"gjc-coordinator-demo\",\n    \"status\": \"completed\",\n    \"final_response\": {\n      \"text\": \"Done\",\n      \"format\": \"markdown\",\n      \"source\": \"report_status\",\n      \"artifact_path\": null,\n      \"truncated\": false\n    },\n    \"evidence\": [{ \"path\": \"artifact.txt\" }],\n    \"error\": null\n  },\n  \"advisory_status\": {\n    \"live\": true,\n    \"state\": \"idle_or_unknown\"\n  }\n}\n```\n\nThe coordinator MCP bridge is currently a durable polling/await surface. It does not expose a push subscription stream; external coordinators should poll `gjc_coordinator_read_coordination_status`, `gjc_coordinator_read_turn`, or bounded `gjc_coordinator_await_turn` instead of waiting for server-sent push events.\n\nExternal `session_id`, `turn_id`, and `question_id` values are validated before path use, and loaded records must match the requested session/turn owner.\n\n## Coordinator event journal\n\nThe bridge persists a restart-safe event journal under the configured coordinator state namespace, for example:\n\n```text\n$GJC_COORDINATOR_MCP_STATE_ROOT/<profile>/<repo>/events/event-journal.jsonl\n```\n\nEach event is a bounded JSONL record with `schema_version`, monotonic namespace-local `seq`, stable `id`, `timestamp`, canonical `kind`, optional `session_id`/`turn_id`/`question_id`/`report_id`, short `summary`, optional `payload_ref`, and bounded scalar `metadata`. Full prompts, reports, final responses, and artifacts stay in their existing turn/report/artifact read paths; event records only point at them.\n\n`gjc_coordinator_watch_events` is a bounded long-poll MCP tool, not an unbounded stream. Inputs are `after_seq` (default `0`), optional `session_id`, optional `event_types`, `timeout_ms` capped at 30000, and `limit` capped at 100. If matching events already exist after `after_seq`, it returns immediately. Otherwise it waits for the event journal to change or for timeout. The response includes `events`, `latest_seq`, `timed_out`, and `transport: { \"mcp\": \"long_poll\", \"push_subscriptions\": false }`, so coordinators can persist `latest_seq` and resume safely after restart.\n\n`gjc_coordinator_read_coordination_status` keeps its existing report fields and now also includes `latest_event_seq` plus recent event summaries for snapshot-style consumers.\n\n## Generic controller config snippet\n\n```json\n{\n  \"mcp_servers\": {\n    \"gjc_coordinator\": {\n      \"command\": \"gjc\",\n      \"args\": [\"mcp-serve\", \"coordinator\"],\n      \"env\": {\n        \"GJC_COORDINATOR_MCP_WORKDIR_ROOTS\": \"/path/to/repo\",\n        \"GJC_COORDINATOR_MCP_PROFILE\": \"team-a\",\n        \"GJC_COORDINATOR_MCP_REPO\": \"project\",\n        \"GJC_COORDINATOR_MCP_SESSION_COMMAND\": \"gjc --worktree\"\n      },\n      \"enabled\": true\n    }\n  }\n}\n```\n\n## Smoke check\n\n```bash\ngjc mcp-serve coordinator --check --json\n```\n\nExpected result includes `ok: true`, server name `gjc-coordinator-mcp`, and the GJC-named tool list.\n",
@@ -37,7 +37,7 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"keybindings.md": "# Keybindings\n\nRun `/hotkeys` inside an `gjc` 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 `~/.gjc/agent/keybindings.json`. The file is a JSON object 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 `~/.gjc/agent/config.yml`, and there is no nested `keybindings` object.\n\n```json\n{\n  \"app.commandPalette.open\": \"Ctrl+P\",\n  \"app.model.cycleForward\": \"Alt+N\",\n  \"app.model.selectTemporary\": \"Alt+P\",\n  \"app.plan.toggle\": \"Alt+Shift+P\"\n}\n```\n\nChord names are case-insensitive and use the same notation shown in the UI, such as `Ctrl+P`, `Alt+N`, `Alt+Shift+P`, `Shift+Enter`, and `Ctrl+Backspace`.\n\nSet an action to an empty array to disable it:\n\n```json\n{\n  \"app.stt.toggle\": []\n}\n```\n\n## Common action IDs\n\n| Action ID | Default | Meaning |\n| --- | --- | --- |\n| `app.commandPalette.open` | `Ctrl+P` | Open the command palette |\n| `app.model.cycleForward` | `Alt+N` | Cycle role models forward |\n| `app.model.cycleBackward` | `Alt+Shift+N` | Cycle role models backward |\n| `app.model.selectTemporary` | `Alt+P` | Pick a model temporarily for this session |\n| `app.model.select` | `Ctrl+L` | 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` | _(none)_ | Optional remap for a follow-up message; `Ctrl+Enter` is reserved for editor newline |\n| `app.message.queue` | `Alt+Enter` (`Alt+Q` on win32) | Explicitly queue a message for the next turn |\n| `app.message.dequeue` | `Alt+Up` | Dequeue a queued message back into the editor |\n\n| `app.clipboard.copyLine` | `Alt+Shift+L` | Copy the current line |\n| `app.clipboard.copyPrompt` | `Alt+Shift+C` | Copy the whole prompt |\n| `app.stt.toggle` | `Alt+H` | Toggle speech-to-text recording |\n| `app.irc.sidebar.toggle` | `Alt+I` | Toggle IRC sidebar |\n\nOlder unqualified action names are migrated when `keybindings.json` is loaded, but new docs and new configs should use the namespaced action IDs above.\n\nOn native Windows terminals, GJC defaults `app.message.queue` to `Alt+Q` because Windows Terminal and PowerShell commonly reserve `Alt+Enter` for fullscreen before GJC can receive it. Users who prefer another chord can remap `app.message.queue` in `~/.gjc/agent/keybindings.json`.\n\nIn the main GJC composer, plain `PageUp` / `PageDown` page the visible transcript viewport instead of browsing prompt history; use `Up` / `Down` or `Ctrl+R` for prompt history. Autocomplete and selector surfaces still use `PageUp` / `PageDown` for list paging while they have focus.\n\n## Auditing default-key collisions\n\nSome default chords are intentionally reused across different UI contexts, where the focused component disambiguates them at dispatch time. For example `Enter` maps to both input submit and selection confirm, and `Ctrl+C` maps to both input copy and selection cancel. These are not conflicts — only one context is active at a time.\n\nTo audit the registry for keys whose default binding is claimed by more than one action, use `detectDefaultKeyCollisions(definitions)` from `@gajae-code/tui/keybindings`. It returns one entry per colliding key with the list of claiming action IDs, which is useful when adding new defaults or reviewing the surface. User-remap conflicts (multiple actions bound to the same chord in `keybindings.json`) continue to be reported separately by `KeybindingsManager.getConflicts()`.\n\nTwo audit clarifications for the current surface:\n\n- `app.clipboard.copyLine` is registry-backed and dispatched through the input controller's custom key handlers, not hardcoded.\n- `tui.input.copy` is declared in the registry but is not currently dispatched by `Editor.handleInput`.\n\nThe editor's configurable action defaults (including the platform-aware `app.clipboard.pasteImage` default) are derived directly from the central `KEYBINDINGS` registry, so there is a single source of truth for those defaults.\n\n## Current surface audit\n\nAuthoritative inventory of the keybinding registry, one row per action. Generated from `TUI_KEYBINDINGS` (`packages/tui/src/keybindings.ts`) and `KEYBINDINGS` (`packages/coding-agent/src/config/keybindings.ts`). Every action ID below is remappable via `~/.gjc/agent/keybindings.json` unless noted. A drift test (`packages/coding-agent/test/keybindings-audit.test.ts`) asserts every registry action ID appears in this table.\n\n### Editor context (`tui.editor.*`)\n\n| Action ID | Default | Notes |\n| --- | --- | --- |\n| `tui.editor.cursorUp` | `up` | |\n| `tui.editor.cursorDown` | `down` | |\n| `tui.editor.cursorLeft` | `left`, `ctrl+b` | `ctrl+b` also `app.tool.backgroundFold` (other context) |\n| `tui.editor.cursorRight` | `right`, `ctrl+f` | |\n| `tui.editor.cursorWordLeft` | `alt+left`, `ctrl+left`, `alt+b` | `ctrl+left` also `app.tree.foldOrUp` |\n| `tui.editor.cursorWordRight` | `alt+right`, `ctrl+right`, `alt+f` | `ctrl+right` also `app.tree.unfoldOrDown` |\n| `tui.editor.cursorLineStart` | `home`, `ctrl+a` | |\n| `tui.editor.cursorLineEnd` | `end`, `ctrl+e` | |\n| `tui.editor.jumpForward` | `ctrl+]` | |\n| `tui.editor.jumpBackward` | `ctrl+alt+]` | |\n| `tui.editor.pageUp` | `pageUp` | |\n| `tui.editor.pageDown` | `pageDown` | |\n| `tui.editor.deleteCharBackward` | `backspace` | |\n| `tui.editor.deleteCharForward` | `delete`, `ctrl+d` | `ctrl+d` also `app.exit` / `app.session.delete` |\n| `tui.editor.deleteWordBackward` | `ctrl+w`, `alt+backspace`, `ctrl+backspace` | |\n| `tui.editor.deleteWordForward` | `alt+delete`, `alt+d` | |\n| `tui.editor.deleteToLineStart` | `ctrl+u` | |\n| `tui.editor.deleteToLineEnd` | `ctrl+k` | |\n| `tui.editor.yank` | `ctrl+y` | |\n| `tui.editor.yankPop` | `alt+y` | |\n| `tui.editor.undo` | `ctrl+-`, `ctrl+_` | |\n\n### Input context (`tui.input.*`)\n\n| Action ID | Default | Notes |\n| --- | --- | --- |\n| `tui.input.newLine` | `Shift+Enter` | `Ctrl+Enter` and `Ctrl+Shift+Enter` are also accepted by the editor when the terminal encodes them distinctly |\n\n| `tui.input.submit` | `enter` | also `tui.select.confirm` (other context) |\n| `tui.input.tab` | `tab` | |\n| `tui.input.copy` | `ctrl+c` | declared but not dispatched by `Editor.handleInput` |\n\n### Selection context (`tui.select.*`)\n\n| Action ID | Default | Notes |\n| --- | --- | --- |\n| `tui.select.up` | `up` | |\n| `tui.select.down` | `down` | |\n| `tui.select.pageUp` | `pageUp` | |\n| `tui.select.pageDown` | `pageDown` | |\n| `tui.select.confirm` | `enter` | |\n| `tui.select.cancel` | `escape`, `ctrl+c` | `escape` also `app.interrupt` |\n\n### Application context (`app.*`)\n\n| Action ID | Default | Notes |\n| --- | --- | --- |\n| `app.interrupt` | `escape` | |\n| `app.clear` | `ctrl+c` | |\n| `app.exit` | `ctrl+d` | |\n| `app.suspend` | `ctrl+z` | |\n| `app.thinking.cycle` | `shift+tab` | |\n| `app.thinking.toggle` | `ctrl+t` | |\n| `app.commandPalette.open` | `ctrl+p` | Open command palette from the editor |\n| `app.model.cycleForward` | `alt+n` | |\n| `app.model.cycleBackward` | `alt+shift+n` | |\n| `app.model.select` | `ctrl+l` | |\n| `app.model.selectTemporary` | `alt+p` | |\n| `app.tools.expand` | `ctrl+o` | |\n| `app.tool.backgroundFold` | `ctrl+b` | |\n| `app.editor.external` | `ctrl+g` | |\n| `app.message.followUp` | _(none)_ | `Ctrl+Enter` remains newline unless the user explicitly remaps this action; while idle the chord still falls through to newline |\n| `app.message.queue` | `alt+enter` (`alt+q` on win32) | platform-aware; avoids the Windows Terminal fullscreen shortcut |\n| `app.message.dequeue` | `alt+up` | |\n| `app.clipboard.pasteImage` | `ctrl+v` (`alt+v` on win32) | platform-aware; single source of truth in `KEYBINDINGS` |\n| `app.clipboard.copyLine` | `alt+shift+l` | registry-backed via input-controller custom handler |\n| `app.clipboard.copyPrompt` | `alt+shift+c` | |\n| `app.session.new` | _(none)_ | command-driven; empty default |\n| `app.session.tree` | _(none)_ | command-driven; empty default |\n| `app.session.fork` | _(none)_ | command-driven; empty default |\n| `app.session.resume` | _(none)_ | command-driven; empty default |\n| `app.session.observe` | `ctrl+s` | also `app.session.toggleSort` (session list) |\n| `app.jobs.open` | `alt+j` | |\n| `app.session.togglePath` | `ctrl+p` | session-list context |\n| `app.session.toggleSort` | `ctrl+s` | session-list context |\n| `app.session.rename` | `ctrl+r` | also `app.history.search` |\n| `app.session.delete` | `ctrl+d` | session-list context |\n| `app.session.deleteNoninvasive` | `ctrl+backspace` | |\n| `app.tree.foldOrUp` | `ctrl+left`, `alt+left` | |\n| `app.tree.unfoldOrDown` | `ctrl+right`, `alt+right` | |\n| `app.plan.toggle` | `alt+shift+p` | |\n| `app.history.search` | `ctrl+r` | |\n| `app.stt.toggle` | `alt+h` | |\n| `app.irc.sidebar.toggle` | `alt+i` | |\n\n### Global engine context (`tui.global.*`)\n\n| Action ID | Default | Notes |\n| --- | --- | --- |\n| `tui.global.debug` | `shift+ctrl+d` | Toggle debug overlay; resolved through the registry in `tui.ts` |\n\nCross-context default reuse (`ctrl+s`, `ctrl+r`, `ctrl+d`, `ctrl+b`, `ctrl+left`/`ctrl+right`, `enter`, `escape`, `ctrl+c`) is intentional: each pair is active in a different focused context and is disambiguated at dispatch time. Use `detectDefaultKeyCollisions()` (above) to re-derive this list from the registry.\n\n### Not yet registry-managed\n\nA few contexts still match chords directly instead of resolving through the registry, and are tracked for a later phase:\n\n- Tree selector (`tree-selector.ts`): up/down/left/right/enter, `ctrl+c`, filter cycling (`ctrl+o` / `ctrl+shift+o`), filter modes (`alt+d/t/u/l/a`), label edit (`shift+l`).\n- Parts of the model selector.\n",
 	"lsp-config.md": "# LSP configuration in GJC\n\nThis guide explains how to configure language servers for the GJC 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, GJC 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\nGJC 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` |\n| 3 | `~/.gjc/agent/lsp.json`, `~/.gjc/agent/lsp.yaml`, `~/.gemini/lsp.*` |\n| 2 | `<project>/.gjc/lsp.json`, `<project>/.gjc/lsp.yaml`, `<project>/.gemini/lsp.*` |\n| 1 (highest) | `<project>/lsp.json`, `<project>/.lsp.json`, `<project>/lsp.yaml` |\n\nEach location accepts both `.json` and `.yaml` / `.yml` variants, as well as hidden-file versions (`.lsp.json`, `.lsp.yaml`). 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 → `~/.gjc/agent/lsp.json`\n- Project-specific overrides → `<project>/.gjc/lsp.json`\n\n> **Note:** The presence of any LSP config file disables auto-detection. When at least one file is found, GJC skips the binary-scan phase and loads all 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 GJC 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>/.gjc/lsp.json`:\n\n```json\n{\n  \"servers\": {\n    \"pylsp\": {\n      \"disabled\": true\n    }\n  }\n}\n```\n\nThe user-level config in `~/.gjc/agent/lsp.json` is unaffected; pylsp is only suppressed in this project.\n\nWhen multiple built-in primary servers support the same file, a default server can list lower-precedence servers in `supersedes`. For example, `csharp-ls` supersedes `omnisharp` only when both C# servers are installed and detected; if `csharp-ls` is unavailable, `omnisharp` remains the fallback.\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| `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| `csharp-ls` | C# | `csharp-ls` |\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",
 	"memory.md": "# Autonomous Memory\n\nWhen enabled, the agent automatically extracts durable knowledge from past sessions and injects a compact summary into each new session. 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 via `/settings` or `config.yml`:\n\n```yaml\nmemories:\n  enabled: true\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- Pair memory-influenced decisions 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### Memory artifacts\n\nGenerated local-memory artifacts are private runtime state, not a public tool or URI surface. They may be summarized into the system prompt when local memory is enabled, but users and model-facing tool docs should not rely on direct `memory://` reads. The legacy internal `memory://` resolver remains only for compatibility with existing persisted guidance and is not part of the public coding harness contract; remove it after legacy local-memory prompts no longer reference it.\n### `/memory` slash command\n\n| Subcommand            | Effect                                         |\n| --------------------- | ---------------------------------------------- |\n| `view`                | Show the current memory injection payload      |\n| `clear` / `reset`     | Delete all memory data and generated artifacts |\n| `enqueue` / `rebuild` | Force consolidation to run at next startup     |\n\n## How it works\n\nMemories are built by a background pipeline that runs at startup or when manually triggered via slash command.\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, or currently active 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 to prevent double-running when multiple processes start simultaneously. Stale skill directories from prior runs are pruned automatically.\n\nAll output is scanned for secrets before being 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.md`    | 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| `memories.enabled`                    | `false` | Master switch                                             |\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, slash command handling\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` — legacy non-public `memory://` compatibility handler\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 models for the default and agent roles\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/ai/src/models.ts` and `packages/ai/src/types.ts` — built-in providers/models and `Model`/`compat` types\n\n## Config file location and legacy behavior\n\nDefault config path:\n\n- `~/.gjc/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    cacheRetention: short  # none | short | long; model entries and modelOverrides can override this\n    discovery:\n      type: ollama\n    modelOverrides:\n      some-model-id:\n        name: Renamed model\n        cacheRetention: long\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        cacheRetention: none\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\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\nmodelBindings:\n  modelRoles:\n    default: my-provider/some-model-id:high\n  agentModelOverrides:\n    executor: my-provider/some-model-id\n```\n\n### Allowed provider/model `api` values\n\n- `openai-completions`\n- `openai-responses`\n- `openai-codex-responses`\n- `azure-openai-responses`\n- `bedrock-converse-stream`\n- `anthropic-messages`\n- `bedrock-converse-stream`\n- `google-generative-ai`\n- `google-vertex`\n- `google-gemini-cli`\n- `ollama-chat`\n- `cursor-agent`\n\n\n### First-class DeepInfra, Azure OpenAI, and Amazon Bedrock examples\n\nAzure OpenAI uses canonical OpenAI model IDs in GJC and resolves those IDs to Azure deployment names at request time. Set `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` to avoid assuming model id equals deployment name:\n\n```yaml\nproviders:\n  azure-openai:\n    baseUrl: https://my-resource.openai.azure.com/openai/v1\n    apiKeyEnv: AZURE_OPENAI_API_KEY\n    api: azure-openai-responses\n    models:\n      - id: gpt-4.1\n      - id: o3\n```\n\n```sh\nexport AZURE_OPENAI_DEPLOYMENT_NAME_MAP='gpt-4.1=gpt-41-prod,o3=o3-reasoning-prod'\n```\n\nDeepInfra is available as the first-class `deepinfra` provider. It uses DeepInfra's OpenAI-compatible Chat Completions endpoint and reads `DEEPINFRA_API_KEY` when no explicit config key is provided. Set `serviceTier: priority` in GJC config or use the runtime service-tier controls to send DeepInfra's `service_tier: \"priority\"` request field for supported models:\n\n```yaml\nproviders:\n  deepinfra:\n    baseUrl: https://api.deepinfra.com/v1/openai\n    apiKeyEnv: DEEPINFRA_API_KEY\n    api: openai-completions\n    models:\n      - id: deepseek-ai/DeepSeek-V3.2\n```\n\nAmazon Bedrock uses the native `bedrock-converse-stream` transport and AWS credential chain auth. Do not put AWS access keys in `models.yml`; configure `AWS_REGION` / `AWS_PROFILE` or standard static AWS credential environment variables instead:\n\n```yaml\nproviders:\n  amazon-bedrock:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com\n    api: bedrock-converse-stream\n    models:\n      - id: us.anthropic.claude-opus-4-6-v1\n      - id: anthropic.claude-3-5-sonnet-20241022-v2:0\n```\n\n### MiniMax and GLM custom provider examples\n\nFor common MiniMax and GLM/zAI setup, prefer the provider presets so the OpenAI-compatible API, base URL, env var, model id, and compatibility flags are written together:\n\n```sh\ngjc setup provider --preset minimax\ngjc setup provider --preset minimax-cn\ngjc setup provider --preset glm\n```\n\nThe same presets are available inside the TUI:\n\n```text\n/provider add --preset minimax\n/provider add --preset glm\n/provider add zai\n```\n\nPresets only write `models.yml` entries that reference documented environment variable names (`MINIMAX_CODE_API_KEY`, `MINIMAX_CODE_CN_API_KEY`, or `ZAI_API_KEY`); they do not store or validate real credentials. The GLM preset aliases (`glm`, `zai`, `z-ai`) write an OpenAI-compatible custom provider named `glm-proxy` and do not replace the first-class `zai` provider.\n\n## Model profiles (`--mpreset`)\n\nModel profiles are optional top-level `profiles:` entries in `~/.gjc/agent/models.yml`. A profile can require provider credentials before activation and can map one or more model roles; omitted roles inherit from the active defaults.\n\n> See also: [Cross-vendor role-based profiles](./multi-vendor-profiles.md) — a curated multi-vendor `profiles:` recipe and verified selector notes that build on the mechanism described here.\n\n```yaml\nprofiles:\n  team-standard:\n    required_providers: [openai, anthropic]\n    model_mapping:\n      default: openai/gpt-5.2\n      executor: anthropic/claude-sonnet-5:medium\n      architect: openai/o3:high\n      planner: openai/o3:high\n      critic: openai/o3:high\n```\n\n`model_mapping` keys are role names (`default`, `executor`, `architect`, `planner`, `critic`). Each role maps to exactly one model selector in the form `provider/modelId[:effort]`; comma-separated fallback chains are not supported in a single role value.\n`required_providers` is the aggregate set of providers required across the profile's mapped roles, not a per-role fallback chain.\n\nBuilt-in profiles are grouped by provider mix and tier:\n\n- `codex-{eco,medium,pro}` — GPT-5.6 Sol/Terra/Luna role mixes tuned by tier and reasoning effort\n- `opencodego` — single OpenCode Go preset (Kimi default, DeepSeek executor/architect, Qwen planner, MiMo critic)\n- `claude-opus` — Anthropic OAuth preset centered on `claude-opus-4-8`\n- Single-provider tiers: `glm-{eco,medium,pro}`, `kimi-coding-plan-{eco,medium,pro}`, `mimo-{eco,medium,pro}`, `grok-{eco,medium,pro}`, `cursor-{eco,medium,pro}`, `minimax-{eco,medium,pro}`\n- Combos: `opus-codex`, `codex-opencodego`, and `fable-opus-codex`\n\nThe `eco` tier favors cheaper models, `medium` assigns Sol to general orchestration with Terra xhigh to difficult execution, and `pro` assigns Sol's highest efforts to orchestration and architecture. The GPT-5.6 profile assignments are product judgments informed by descriptive repeated local exact-edit evidence for selected executor-style TypeScript tasks; that evidence does not evaluate or prove default, planner, architect, or critic performance. See [GPT-5.6 Codex preset benchmark](./gpt-5.6-codex-preset-benchmark.md). Effort suffixes are clamped to each model's supported thinking range at preview and activation time. Single-provider tiers pin each provider's current flagship (`zai/glm-5.2`, `kimi-code/kimi-k2.7-code`, `xiaomi/mimo-v2.5-pro`, `xai/grok-4.3`, `cursor/composer-1.5`, `minimax-code/minimax-m3`). User-defined profiles override built-ins by exact profile name.\n\n\nUse `gjc --mpreset <name>` to activate a profile for the current session only. Activation hard-blocks when any provider listed in `required_providers` lacks credentials. Add `--default` to persist the selected profile as `modelProfile.default` in `config.yml`, so it applies at startup:\n\n```sh\ngjc --mpreset codex-medium\ngjc --mpreset opencodego --default\n```\n\nThe `/model` command opens to a preset landing view: presets are grouped by provider with live auth marks (✓/✗), highlighting a group expands its tiers, and selecting a tier shows the full role→model preview before applying for the session or as default. Typing jumps straight to model search, and `Browse all models` opens the classic tabbed model selector. In `/login`, `Add custom provider` is the first option for configuring credentials needed by custom or profile-required providers; after a successful provider login, the matching preset is recommended automatically.\n\nMiniMax's OpenAI-compatible endpoint rejects multiple system messages and emits thinking in `reasoning_content`, so pin the public-safe compatibility fields when hand-authoring a custom provider:\n\n```yaml\nproviders:\n  minimax-custom:\n    baseUrl: https://api.minimax.io/v1\n    apiKeyEnv: MINIMAX_API_KEY\n    api: openai-completions\n    compat:\n      supportsStore: false\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n      reasoningContentField: reasoning_content\n    models:\n      - id: MiniMax-M2.5\n```\n\nGLM via z.ai is available as the first-class `zai` provider. For a private GLM-compatible proxy, keep secrets in an env var and disable OpenAI-only request fields as needed:\n\n```yaml\nproviders:\n  glm-proxy:\n    baseUrl: https://api.z.ai/api/paas/v4\n    apiKeyEnv: ZAI_API_KEY\n    api: openai-completions\n    compat:\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n    models:\n      - id: glm-4.6\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- `models.yml` is strict: unknown provider/model keys fail validation before provider dispatch, so stale keys such as `requestTransform` or `wireModelId` only work where this document lists them.\n- `discovery.type`: `ollama`, `llama.cpp`, or `lm-studio`\n- `cacheRetention`: `none`, `short`, or `long`; request-time options win over model/modelOverride values, then provider values, then `GJC_CACHE_RETENTION`, then the runtime default. The runtime default is `short` for most providers, but the Anthropic provider defaults to `long` (`ttl: \"1h\"`) because the ~5m default is too fragile for long-running subagent workflows. The 1h marker is only emitted on the canonical Anthropic API (`api.anthropic.com`) for models advertising `supportsLongCacheRetention`; proxies, gateways, and incapable models fall back to the default ephemeral (~5m) breakpoint. For OpenAI Responses, this controls `prompt_cache_retention` only; it does not disable `prompt_cache_key` when a stable session id exists.\n\n## OpenAI-compatible proxy configuration\n\nOpenAI-compatible proxy providers should use schema-supported provider keys first:\n\n```yaml\nproviders:\n  proxy-provider:\n    baseUrl: https://api.proxy.example/v1\n    apiKeyEnv: PROXY_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: local-gpt\n        name: Local GPT\n        reasoning: true\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n```\n\nUse provider-level `headers` for proxy-required headers. Keep the provider `api` set to `openai-completions` when the proxy exposes Chat Completions-compatible `/v1/chat/completions` semantics. `auth: apiKey` sends the resolved token as bearer auth; use `auth: none` only for trusted local/no-auth endpoints.\n\n`requestTransform` and `wireModelId` remain supported for request-body shaping, but they are not needed for ordinary OpenAI-compatible proxies whose local model id is already the upstream wire id. Unknown config keys fail validation before a provider request is sent.\n\nWhen request shaping is needed:\n\n- `requestTransform.profile: openai-proxy` strips OpenAI SDK/Stainless telemetry and beta headers at final fetch time and sets a generic GJC user agent.\n- `stripHeaders` replaces the preset strip list when provided.\n- `setHeaders` is applied after stripping; use `null` to remove a header.\n- `extraBody` is shallow-merged into the JSON request body after provider compatibility fields; core transport keys such as `model`, `messages`/`input`, `stream`, `tools`, and `tool_choice` are protected and ignored.\n- Model-level `requestTransform` overrides provider-level fields and shallow-merges `setHeaders`/`extraBody`.\n- `wireModelId` changes only the upstream request body model id; local selection still uses `provider/id`.\n\n### Layofflabs-style proxy example\n\n```yaml\nproviders:\n  layofflabs:\n    baseUrl: https://api.layofflabs.com/v1\n    apiKeyEnv: OPENAI_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: gpt-5.5\n        name: GPT 5.5 via Layofflabs\n        reasoning: true\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n\nmodelBindings:\n  modelRoles:\n    default: layofflabs/gpt-5.5:high\n  agentModelOverrides:\n    executor: layofflabs/gpt-5.5:high\n```\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- `headers`\n- `compat`\n- `requestTransform`\n- `disableStrictTools`\n- `modelOverrides`\n- `discovery`\n\n### Discovery\n\n- `discovery` requires provider-level `api`.\n\n### Model value checks\n\n- `id` required\n- `contextWindow` and `maxTokens` must be positive if provided\n- unknown provider, model, override, and request-transform keys fail schema validation; remove stale keys instead of relying on them being ignored.\n\n## Merge and override order\n\nModelRegistry pipeline (on refresh):\n\n1. Load built-in providers/models from `@gajae-code/ai`.\n2. Load `models.yml` custom config.\n3. Apply provider overrides (`baseUrl`, `headers`, `requestTransform`, `disableStrictTools`, `cacheRetention`) 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 (schema v3) with a `static_fingerprint` column that hashes the\nstatic 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 via a WeakMap keyed by the static-models array\nreference, 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- `anthropic-model-opus-4-6`\n- `anthropic-model-haiku-4-5`\n- `gpt-5.3-openai-code`\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: openai-code\n        name: Zenmux OpenAI code\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/openai-code: gpt-5.3-openai-code\n    p-openai-code/openai-code: gpt-5.3-openai-code\n  exclude:\n    - demo/openai-code-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`, `input`, `cost`, `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 `http://127.0.0.1:11434`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls Ollama endpoints and normalizes discovered OpenAI-compatible models to `openai-responses`.\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\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### 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 `GJC_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 (`GJC_AUTH_BROKER_URL`, `GJC_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` plus the agent assignment targets `executor`, `architect`, `planner`, `critic`\n\nRole aliases like `pi/default` 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 code provider 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-openai-code` 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  - anthropic-model-sonnet-4-5\n  - path: ~/work\n    models:\n      - anthropic/anthropic-model-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. It is **off by default** (`contextPromotion.enabled` is `false`); opt in to enable it.\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 code provider 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 Spark -> non-Spark on the same provider:\n\n```yaml\nproviders:\n  openai-code:\n    modelOverrides:\n      gpt-5.3-openai-code-spark:\n        contextPromotionTarget: openai-code/gpt-5.3-openai-code\n```\n\nThe built-in model generator also assigns this automatically for `*-spark` models when a same-provider base model exists.\n\n## Compatibility and routing fields\n\nThe `compat` block on a provider or model overrides the URL-based auto-detection in `packages/ai/src/providers/openai-completions-compat.ts`. It is validated by `OpenAICompatSchema` in `packages/coding-agent/src/config/model-registry.ts` and consumed by every `openai-completions` transport (`packages/ai/src/providers/openai-completions.ts`). The canonical type is `OpenAICompat` in `packages/ai/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- `sendSessionHeaders` — forward the agent session id as `session_id` and `x-session-id` request headers so OpenAI-compatible relays/proxies can do session-affinity routing and reuse a server-side prompt cache. Default: `false`. Caller-set `headers`/`requestTransform` values are never overwritten.\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- `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 and zAI).\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- `requiresAssistantContentForToolCalls` — assistant tool-call turns must include non-empty text content (Kimi). Default: `false`.\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/ai/src/types.ts`). The `models.yml` schema currently exposes only the strict-tools opt-out as a top-level provider field (see below); the remaining Anthropic-side knobs (`disableAdaptiveThinking`, `supportsEagerToolInputStreaming`, `supportsLongCacheRetention`) 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. This is enabled by default for all `anthropic-messages` providers because it guarantees schema conformance in agentic systems.\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:\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: anthropic-model-sonnet-4-20250514\n        name: Anthropic model 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.\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## 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\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: anthropic-model-sonnet-4-20250514\n        name: Anthropic model 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/anthropic-model-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 `~/.gjc/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 models for the default and agent roles\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/ai/src/models.ts` and `packages/ai/src/types.ts` — built-in providers/models and `Model`/`compat` types\n\n## Config file location and legacy behavior\n\nDefault config path:\n\n- `~/.gjc/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    cacheRetention: short  # none | short | long; model entries and modelOverrides can override this\n    discovery:\n      type: ollama\n    modelOverrides:\n      some-model-id:\n        name: Renamed model\n        cacheRetention: long\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        cacheRetention: none\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\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\nmodelBindings:\n  modelRoles:\n    default: my-provider/some-model-id:high\n  agentModelOverrides:\n    executor: my-provider/some-model-id\n```\n\n### Allowed provider/model `api` values\n\n- `openai-completions`\n- `openai-responses`\n- `openai-codex-responses`\n- `azure-openai-responses`\n- `bedrock-converse-stream`\n- `anthropic-messages`\n- `bedrock-converse-stream`\n- `google-generative-ai`\n- `google-vertex`\n- `google-gemini-cli`\n- `ollama-chat`\n- `cursor-agent`\n\n\n### First-class DeepInfra, Azure OpenAI, and Amazon Bedrock examples\n\nAzure OpenAI uses canonical OpenAI model IDs in GJC and resolves those IDs to Azure deployment names at request time. Set `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` to avoid assuming model id equals deployment name:\n\n```yaml\nproviders:\n  azure-openai:\n    baseUrl: https://my-resource.openai.azure.com/openai/v1\n    apiKeyEnv: AZURE_OPENAI_API_KEY\n    api: azure-openai-responses\n    models:\n      - id: gpt-4.1\n      - id: o3\n```\n\n```sh\nexport AZURE_OPENAI_DEPLOYMENT_NAME_MAP='gpt-4.1=gpt-41-prod,o3=o3-reasoning-prod'\n```\n\nDeepInfra is available as the first-class `deepinfra` provider. It uses DeepInfra's OpenAI-compatible Chat Completions endpoint and reads `DEEPINFRA_API_KEY` when no explicit config key is provided. Set `serviceTier: priority` in GJC config or use the runtime service-tier controls to send DeepInfra's `service_tier: \"priority\"` request field for supported models:\n\n```yaml\nproviders:\n  deepinfra:\n    baseUrl: https://api.deepinfra.com/v1/openai\n    apiKeyEnv: DEEPINFRA_API_KEY\n    api: openai-completions\n    models:\n      - id: deepseek-ai/DeepSeek-V3.2\n```\n\nAmazon Bedrock uses the native `bedrock-converse-stream` transport and AWS credential chain auth. Do not put AWS access keys in `models.yml`; configure `AWS_REGION` / `AWS_PROFILE` or standard static AWS credential environment variables instead:\n\n```yaml\nproviders:\n  amazon-bedrock:\n    baseUrl: https://bedrock-runtime.us-east-1.amazonaws.com\n    api: bedrock-converse-stream\n    models:\n      - id: us.anthropic.claude-opus-4-6-v1\n      - id: anthropic.claude-3-5-sonnet-20241022-v2:0\n```\n\n### MiniMax and GLM custom provider examples\n\nFor common MiniMax and GLM/zAI setup, prefer the provider presets so the OpenAI-compatible API, base URL, env var, model id, and compatibility flags are written together:\n\n```sh\ngjc setup provider --preset minimax\ngjc setup provider --preset minimax-cn\ngjc setup provider --preset glm\n```\n\nThe same presets are available inside the TUI:\n\n```text\n/provider add --preset minimax\n/provider add --preset glm\n/provider add zai\n```\n\nPresets only write `models.yml` entries that reference documented environment variable names (`MINIMAX_CODE_API_KEY`, `MINIMAX_CODE_CN_API_KEY`, or `ZAI_API_KEY`); they do not store or validate real credentials. The GLM preset aliases (`glm`, `zai`, `z-ai`) write an OpenAI-compatible custom provider named `glm-proxy` and do not replace the first-class `zai` provider.\n\n## Model profiles (`--mpreset`)\n\nModel profiles are optional top-level `profiles:` entries in `~/.gjc/agent/models.yml`. A profile can require provider credentials before activation and can map one or more model roles; omitted roles inherit from the active defaults.\n\n> See also: [Cross-vendor role-based profiles](./multi-vendor-profiles.md) — a curated multi-vendor `profiles:` recipe and verified selector notes that build on the mechanism described here.\n\n```yaml\nprofiles:\n  team-standard:\n    required_providers: [openai, anthropic]\n    model_mapping:\n      default: openai/gpt-5.2\n      executor: anthropic/claude-sonnet-5:medium\n      architect: openai/o3:high\n      planner: openai/o3:high\n      critic: openai/o3:high\n```\n\n`model_mapping` keys are role names (`default`, `executor`, `architect`, `planner`, `critic`). Each role maps to exactly one model selector in the form `provider/modelId[:effort]`; comma-separated fallback chains are not supported in a single role value.\n`required_providers` is the aggregate set of providers required across the profile's mapped roles, not a per-role fallback chain.\n\nBuilt-in profiles are grouped by provider mix and tier:\n\n- `codex-{eco,medium,pro}` — GPT-5.6 Sol/Terra/Luna role mixes tuned by tier and reasoning effort\n- `opencodego` — single OpenCode Go preset (Kimi default, DeepSeek executor/architect, Qwen planner, MiMo critic)\n- `claude-opus` — Anthropic OAuth preset centered on `claude-opus-4-8`\n- Single-provider tiers: `glm-{eco,medium,pro}`, `kimi-coding-plan-{eco,medium,pro}`, `mimo-{eco,medium,pro}`, `grok-{eco,medium,pro}`, `cursor-{eco,medium,pro}`, `minimax-{eco,medium,pro}`\n- Combos: `opus-codex`, `codex-opencodego`, and `fable-opus-codex`\n\nThe `eco`, `medium`, and `pro` Codex profile mappings are current product judgments: Eco assigns Terra low/Luna low/Luna high/Terra xhigh/Terra high to default/executor/planner/critic/architect; Medium assigns Sol low/Terra low/Terra high/Sol xhigh/Sol high; and Pro assigns Sol medium/Terra medium/Sol high/Sol max/Sol xhigh. `opus-codex` retains the Medium Codex executor, critic, and architect roles but uses `anthropic/claude-sonnet-5` for planner; `codex-opencodego` retains the Medium Codex default and architect roles; and `fable-opus-codex` uses the Pro Codex executor and architect roles with `anthropic/claude-opus-4-8:medium` for planner. The descriptive repeated local exact-edit evidence informs only selected executor-style TypeScript tasks; it does not evaluate or prove default, planner, architect, or critic performance. See [GPT-5.6 Codex preset benchmark](./gpt-5.6-codex-preset-benchmark.md). Effort suffixes are clamped to each model's supported thinking range at preview and activation time. Single-provider tiers pin each provider's current flagship (`zai/glm-5.2`, `kimi-code/kimi-k2.7-code`, `xiaomi/mimo-v2.5-pro`, `xai/grok-4.3`, `cursor/composer-1.5`, `minimax-code/minimax-m3`). User-defined profiles override built-ins by exact profile name.\n\n\nUse `gjc --mpreset <name>` to activate a profile for the current session only. Activation hard-blocks when any provider listed in `required_providers` lacks credentials. Add `--default` to persist the selected profile as `modelProfile.default` in `config.yml`, so it applies at startup:\n\n```sh\ngjc --mpreset codex-medium\ngjc --mpreset opencodego --default\n```\n\nThe `/model` command opens to a preset landing view: presets are grouped by provider with live auth marks (✓/✗), highlighting a group expands its tiers, and selecting a tier shows the full role→model preview before applying for the session or as default. Typing jumps straight to model search, and `Browse all models` opens the classic tabbed model selector. In `/login`, `Add custom provider` is the first option for configuring credentials needed by custom or profile-required providers; after a successful provider login, the matching preset is recommended automatically.\n\nMiniMax's OpenAI-compatible endpoint rejects multiple system messages and emits thinking in `reasoning_content`, so pin the public-safe compatibility fields when hand-authoring a custom provider:\n\n```yaml\nproviders:\n  minimax-custom:\n    baseUrl: https://api.minimax.io/v1\n    apiKeyEnv: MINIMAX_API_KEY\n    api: openai-completions\n    compat:\n      supportsStore: false\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n      reasoningContentField: reasoning_content\n    models:\n      - id: MiniMax-M2.5\n```\n\nGLM via z.ai is available as the first-class `zai` provider. For a private GLM-compatible proxy, keep secrets in an env var and disable OpenAI-only request fields as needed:\n\n```yaml\nproviders:\n  glm-proxy:\n    baseUrl: https://api.z.ai/api/paas/v4\n    apiKeyEnv: ZAI_API_KEY\n    api: openai-completions\n    compat:\n      supportsDeveloperRole: false\n      supportsReasoningEffort: false\n    models:\n      - id: glm-4.6\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- `models.yml` is strict: unknown provider/model keys fail validation before provider dispatch, so stale keys such as `requestTransform` or `wireModelId` only work where this document lists them.\n- `discovery.type`: `ollama`, `llama.cpp`, or `lm-studio`\n- `cacheRetention`: `none`, `short`, or `long`; request-time options win over model/modelOverride values, then provider values, then `GJC_CACHE_RETENTION`, then the runtime default. The runtime default is `short` for most providers, but the Anthropic provider defaults to `long` (`ttl: \"1h\"`) because the ~5m default is too fragile for long-running subagent workflows. The 1h marker is only emitted on the canonical Anthropic API (`api.anthropic.com`) for models advertising `supportsLongCacheRetention`; proxies, gateways, and incapable models fall back to the default ephemeral (~5m) breakpoint. For OpenAI Responses, this controls `prompt_cache_retention` only; it does not disable `prompt_cache_key` when a stable session id exists.\n\n## OpenAI-compatible proxy configuration\n\nOpenAI-compatible proxy providers should use schema-supported provider keys first:\n\n```yaml\nproviders:\n  proxy-provider:\n    baseUrl: https://api.proxy.example/v1\n    apiKeyEnv: PROXY_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: local-gpt\n        name: Local GPT\n        reasoning: true\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n```\n\nUse provider-level `headers` for proxy-required headers. Keep the provider `api` set to `openai-completions` when the proxy exposes Chat Completions-compatible `/v1/chat/completions` semantics. `auth: apiKey` sends the resolved token as bearer auth; use `auth: none` only for trusted local/no-auth endpoints.\n\n`requestTransform` and `wireModelId` remain supported for request-body shaping, but they are not needed for ordinary OpenAI-compatible proxies whose local model id is already the upstream wire id. Unknown config keys fail validation before a provider request is sent.\n\nWhen request shaping is needed:\n\n- `requestTransform.profile: openai-proxy` strips OpenAI SDK/Stainless telemetry and beta headers at final fetch time and sets a generic GJC user agent.\n- `stripHeaders` replaces the preset strip list when provided.\n- `setHeaders` is applied after stripping; use `null` to remove a header.\n- `extraBody` is shallow-merged into the JSON request body after provider compatibility fields; core transport keys such as `model`, `messages`/`input`, `stream`, `tools`, and `tool_choice` are protected and ignored.\n- Model-level `requestTransform` overrides provider-level fields and shallow-merges `setHeaders`/`extraBody`.\n- `wireModelId` changes only the upstream request body model id; local selection still uses `provider/id`.\n\n### Layofflabs-style proxy example\n\n```yaml\nproviders:\n  layofflabs:\n    baseUrl: https://api.layofflabs.com/v1\n    apiKeyEnv: OPENAI_API_KEY\n    api: openai-completions\n    auth: apiKey\n    headers:\n      User-Agent: curl/8.7.1\n    models:\n      - id: gpt-5.5\n        name: GPT 5.5 via Layofflabs\n        reasoning: true\n        thinking:\n          minLevel: low\n          maxLevel: xhigh\n          mode: effort\n          defaultLevel: high\n          levels: [low, medium, high, xhigh]\n        input: [text]\n        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }\n        contextWindow: 400000\n        maxTokens: 128000\n\nmodelBindings:\n  modelRoles:\n    default: layofflabs/gpt-5.5:high\n  agentModelOverrides:\n    executor: layofflabs/gpt-5.5:high\n```\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- `headers`\n- `compat`\n- `requestTransform`\n- `disableStrictTools`\n- `modelOverrides`\n- `discovery`\n\n### Discovery\n\n- `discovery` requires provider-level `api`.\n\n### Model value checks\n\n- `id` required\n- `contextWindow` and `maxTokens` must be positive if provided\n- unknown provider, model, override, and request-transform keys fail schema validation; remove stale keys instead of relying on them being ignored.\n\n## Merge and override order\n\nModelRegistry pipeline (on refresh):\n\n1. Load built-in providers/models from `@gajae-code/ai`.\n2. Load `models.yml` custom config.\n3. Apply provider overrides (`baseUrl`, `headers`, `requestTransform`, `disableStrictTools`, `cacheRetention`) 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 (schema v3) with a `static_fingerprint` column that hashes the\nstatic 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 via a WeakMap keyed by the static-models array\nreference, 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- `anthropic-model-opus-4-6`\n- `anthropic-model-haiku-4-5`\n- `gpt-5.3-openai-code`\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: openai-code\n        name: Zenmux OpenAI code\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/openai-code: gpt-5.3-openai-code\n    p-openai-code/openai-code: gpt-5.3-openai-code\n  exclude:\n    - demo/openai-code-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`, `input`, `cost`, `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 `http://127.0.0.1:11434`\n- auth mode: keyless (`auth: none` behavior)\n\nRuntime discovery calls Ollama endpoints and normalizes discovered OpenAI-compatible models to `openai-responses`.\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\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### 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 `GJC_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 (`GJC_AUTH_BROKER_URL`, `GJC_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` plus the agent assignment targets `executor`, `architect`, `planner`, `critic`\n\nRole aliases like `pi/default` 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 code provider 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-openai-code` 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  - anthropic-model-sonnet-4-5\n  - path: ~/work\n    models:\n      - anthropic/anthropic-model-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. It is **off by default** (`contextPromotion.enabled` is `false`); opt in to enable it.\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 code provider 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 Spark -> non-Spark on the same provider:\n\n```yaml\nproviders:\n  openai-code:\n    modelOverrides:\n      gpt-5.3-openai-code-spark:\n        contextPromotionTarget: openai-code/gpt-5.3-openai-code\n```\n\nThe built-in model generator also assigns this automatically for `*-spark` models when a same-provider base model exists.\n\n## Compatibility and routing fields\n\nThe `compat` block on a provider or model overrides the URL-based auto-detection in `packages/ai/src/providers/openai-completions-compat.ts`. It is validated by `OpenAICompatSchema` in `packages/coding-agent/src/config/model-registry.ts` and consumed by every `openai-completions` transport (`packages/ai/src/providers/openai-completions.ts`). The canonical type is `OpenAICompat` in `packages/ai/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- `sendSessionHeaders` — forward the agent session id as `session_id` and `x-session-id` request headers so OpenAI-compatible relays/proxies can do session-affinity routing and reuse a server-side prompt cache. Default: `false`. Caller-set `headers`/`requestTransform` values are never overwritten.\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- `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 and zAI).\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- `requiresAssistantContentForToolCalls` — assistant tool-call turns must include non-empty text content (Kimi). Default: `false`.\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/ai/src/types.ts`). The `models.yml` schema currently exposes only the strict-tools opt-out as a top-level provider field (see below); the remaining Anthropic-side knobs (`disableAdaptiveThinking`, `supportsEagerToolInputStreaming`, `supportsLongCacheRetention`) 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. This is enabled by default for all `anthropic-messages` providers because it guarantees schema conformance in agentic systems.\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:\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: anthropic-model-sonnet-4-20250514\n        name: Anthropic model 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.\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## 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\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: anthropic-model-sonnet-4-20250514\n        name: Anthropic model 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/anthropic-model-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 `~/.gjc/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",
 	"multi-vendor-profiles.md": "# Choosing models in GJC: role-based profiles\n\nA practical guide to picking models for GJC's roles, for every subscription situation — one vendor, two vendors, or the full multi-vendor set. It adds curated cross-vendor `profiles:` for `~/.gjc/agent/models.yml` and verified selector notes on top of the mechanism in [Model profiles](./models.md#model-profiles---mpreset). Everything here is **user config**; it complements the built-in `--mpreset` presets and overrides a built-in only when it shares its exact name.\n\n> Selectors, prices, and \"axis leaders\" are catalog- and time-sensitive (observed 2026-06 on the current bundled catalog). Re-verify any selector with `gjc -p --no-session --no-tools --model <selector> \"Reply OK\"`.\n\n## The five roles\n\n`default` runs the main loop and most turns; `executor` / `architect` / `planner` / `critic` are the four bundled task agents, delegated only when the work calls for it.\n\n| Role | What it optimizes for |\n| --- | --- |\n| `default` | tool-calling reliability + honesty (it routes — its quality bounds the whole system) |\n| `executor` | real coding (SWE-bench Verified) |\n| `planner` | reasoning + sequencing (GPQA / ARC-AGI-2) |\n| `architect` | large-context + multimodal review |\n| `critic` | independent adversarial review (different family from what it reviews) |\n\n## Pick by what you subscribe to\n\n| You have | Use |\n| --- | --- |\n| **One vendor** | the built-in preset for that vendor — `claude-opus` (Anthropic), `codex-{eco,medium,pro}` (OpenAI/Codex), `opencodego` (OpenCode Go), or a single-vendor flagship tier (`zai/glm-5.2`, `kimi-code/...`, `xiaomi/...`, `xai/grok-4.3`, `minimax-code/...`). These already map all five roles inside one vendor. |\n| **Claude + Codex** | the built-in `opus-codex` (Claude main loop + Codex support roles). |\n| **Three or more / all five** | the cross-vendor profiles below — each role on its axis leader, `critic` kept cross-family. |\n\nThe single guiding rule across all of these: **keep `default` on the strongest router you have** (Anthropic Opus when available). A weak `default` caps quality regardless of the delegated models.\n\n## Cross-vendor profiles (3+ vendors)\n\nNo single vendor leads every axis, so these put each role on its axis leader and keep `critic` on a different family from the `executor` it reviews.\n\n```yaml\nprofiles:\n\n  daily:                 # everyday balance\n    required_providers: [anthropic, openai-codex, google-antigravity, xai]\n    model_mapping:\n      default:   anthropic/claude-opus-4-8:medium\n      executor:  openai-codex/gpt-5.4:high\n      planner:   google-antigravity/gemini-3.1-pro-low:high\n      architect: google-antigravity/gemini-3.1-pro-low:high\n      critic:    xai/grok-4.3:medium\n\n  ultimate:              # cost-no-object, best per role\n    required_providers: [anthropic, openai-codex, google-antigravity, xai]\n    model_mapping:\n      default:   anthropic/claude-opus-4-8:high\n      executor:  anthropic/claude-opus-4-8:max\n      planner:   openai-codex/gpt-5.5:xhigh\n      architect: google-antigravity/gemini-3.1-pro-low:high\n      critic:    xai/grok-4.3:high\n\n  eco:                   # cheapest delegated work; main loop stays on Opus\n    required_providers: [anthropic, opencode-go, google-antigravity, xai]\n    model_mapping:\n      default:   anthropic/claude-opus-4-8:low\n      executor:  opencode-go/deepseek-v4-flash\n      planner:   xai/grok-4-1-fast:high\n      architect: google-antigravity/gemini-3.1-pro-low\n      critic:    google-antigravity/gemini-3.5-flash\n\n  monorepo:              # huge codebases (openai-codex excluded: 272k context cap)\n    required_providers: [anthropic, google-antigravity, opencode-go]\n    model_mapping:\n      default:   anthropic/claude-opus-4-8:medium\n      executor:  anthropic/claude-opus-4-8:high\n      planner:   google-antigravity/gemini-3.1-pro-low:high\n      architect: anthropic/claude-opus-4-8:high\n      critic:    opencode-go/glm-5.2\n\n  reviewer:              # review/audit stance — the author-mode role split, inverted\n    required_providers: [anthropic, openai-codex, google-antigravity]\n    model_mapping:\n      default:   anthropic/claude-opus-4-8:high                 # aggregator restraint: preserve raw reviewer verdicts\n      executor:  openai-codex/gpt-5.5:high                      # support — repro PoCs, failing tests, harnesses\n      planner:   google-antigravity/gemini-3.1-pro-low:high     # review checklists / audit scoping\n      architect: anthropic/claude-opus-4-8:high                 # lead 1 — primary code-review judge (effective long-context)\n      critic:    openai-codex/gpt-5.5:high                      # lead 2 — merge gate, cross-family vs Claude-authored code\n```\n\n## Reviewer stance and the external review gate\n\nThe profiles above assume an **authoring** stance: `executor` is the lead and `architect`/`critic` verify its work. In a session whose primary job is reviewing or auditing (not writing) code, the roles invert — `architect`/`critic` become the leads and `executor` is support (reproduction PoCs, failing tests). The `reviewer` profile encodes that inversion, with one generalized provenance rule: **the reviewing model family must differ from the family that authored the code under review**, not merely from the session's own executor.\n\nA verified use is the cross-session final review gate: the authoring session launches a fresh, stateless reviewer sub-session so the finished diff is judged without the authoring context:\n\n```sh\n# the one-shot gate needs only a cross-family --model; add --mpreset reviewer as an\n# optional enhancement AFTER installing this profile in ~/.gjc/agent/models.yml:\ngjc -p --no-session --model openai-codex/gpt-5.5:xhigh --tools read,search,find \"<review prompt: diff + spec paths, severity findings, final line VERDICT: APPROVE|REQUEST_CHANGES>\"\n```\n\nThe `--tools` allowlist is part of the contract: it enforces the reviewer's read-only boundary for the built-in tool surface instead of trusting the prompt (the runtime still injects the session `goal` tool unless `goal.enabled` is off — disabling it for the reviewer invocation is **mandatory**, via a dedicated gate directory outside the repo so the reviewed checkout stays clean, see the template — plus `generate_image` when an image credential exists). In this one-shot form the session's `default` model authors the verdict — a tool-restricted print session cannot delegate to the profile's `critic`/`architect` roles — so the explicit cross-family `--model` carries provenance, and the `reviewer` profile itself serves the interactive review-session case (activate it with `--mpreset reviewer` only after copying it into `models.yml`; otherwise activation fails with an unknown-profile error). Profile names in this document live in the user namespace — a user profile overrides a builtin preset only on an exact name match, and a future builtin with the same name would be silently shadowed by your copy.\n\nSee [Extragoal local skill template](./extragoal-skill-template.md) for the full gate workflow (verdict contract, findings triage, bounded re-sign loop, secret-scan and injection guards) built on this recipe.\n\n## Model cheatsheet (by need)\n\nCurrent axis leaders and the cheaper second option, with metered price ($/1M in/out; Gemini via Antigravity runs on the Google AI subscription):\n\n| Need | First pick | Cheaper option |\n| --- | --- | --- |\n| Router / tool-calling (`default`) | `anthropic/claude-opus-4-8` (5/25) | `anthropic/claude-sonnet-5` (3/15) |\n| Coding (`executor`) | `anthropic/claude-opus-4-8` — SWE-bench Verified ~88.6 (5/25) | `openai-codex/gpt-5.4` (2.5/15) · `opencode-go/deepseek-v4-flash` (0.14/0.28) |\n| Reasoning (`planner`) | `openai-codex/gpt-5.5` (ARC-AGI-2) / `google-antigravity/gemini-3.1-pro-low:high` (GPQA) | `xai/grok-4-1-fast` (0.2/0.5) |\n| Large context (`architect`) | `anthropic/claude-opus-4-8` (effective long-context) | `xai/grok-4-fast` (2M nominal, 0.2/0.5) |\n| Multimodal review (`architect`) | `google-antigravity/gemini-3.1-pro-low:high` | `google-antigravity/gemini-3.5-flash` |\n| Independent critic | `xai/grok-4.3` (1.25/2.5) | `opencode-go/glm-5.2` · `google-antigravity/gemini-3.5-flash` |\n\nOn standard tasks, all current frontier models in the catalog are accurate; **pick by cost, latency, and role fit, not by raw accuracy on easy prompts.** As an indicative GJC-routed latency reference (`gjc -p`, identical coding + reasoning prompts, all correct): `grok-4.3` and `glm-5.2` ≈ 2–3s, `deepseek-v4-pro` ≈ 3–4s, `claude-opus-4-8` / `gpt-5.5` ≈ 4–7s, `gemini-3.1-pro-low:high` ≈ 7s.\n\n## Verified selector notes (current catalog)\n\nObserved via live `gjc -p` calls; useful when wiring the profiles above:\n\n- **Antigravity Gemini, high reasoning** → use `google-antigravity/gemini-3.1-pro-low:high`. The id `gemini-3.1-pro-high` returns HTTP 400 (no matching backend model); `thinkingLevel` is a per-request parameter, so raising it on `gemini-3.1-pro-low` invokes the model's native high-reasoning mode rather than a degraded one.\n- **openai-codex on a ChatGPT account** serves base GPT only (`gpt-5.5`, `gpt-5.4`). Standalone `-codex` variants (`gpt-5.3-codex`, `gpt-5.2-codex`, `gpt-5.1-codex-max` / `-mini`) return `not supported when using Codex with a ChatGPT account`.\n- **Single-message input limit is separate from the context window.** `claude-opus-4-8` runs with a 1M window via multi-turn accumulation, but a single `@file` message above ~400k tokens returns 400 on `anthropic` / `google-antigravity`; `xai` / `opencode-go` accept larger single messages. Chunk very large inputs across turns instead of pasting one block.\n- **Some selectors come from a provider's live catalog, not the bundled snapshot.** `opencode-go/glm-5.2` and `google-antigravity/gemini-3.5-flash` resolved in `gjc -p` tests but are **not** in `packages/ai/src/models.json`; they appear only after the provider's online model discovery has populated the registry. `required_providers` verifies credentials at activation — it does **not** guarantee fresh, non-stale discovery — so activation can still fail with `selector did not resolve` until discovery runs (re-login or retry to refresh). If you hit that, substitute a bundled id: `opencode-go/deepseek-v4-pro` for the critic, or `zai/glm-5.2` (add `zai` to `required_providers`) for GLM 5.2.\n\n## Activation\n\n```bash\ngjc --mpreset daily               # this session only\ngjc --mpreset ultimate --default  # persist as the startup default (config.yml)\n```\n\nActivation hard-blocks when any provider in `required_providers` lacks credentials, so log in first: `/login anthropic`, `/login openai-codex`, `/login google-antigravity`, `/login xai` (and `opencode-go` via `OPENCODE_API_KEY`).\n",
 	"native-ffi-optimization-policy.md": "# ADR: Native FFI Optimization Policy\n\n- Status: Accepted\n- Scope: `crates/pi-natives` algorithmic ports proposed for performance reasons\n- Related: [`porting-to-natives.md`](./porting-to-natives.md), [`natives-architecture.md`](./natives-architecture.md), [`natives-binding-contract.md`](./natives-binding-contract.md), [`cpu-hotspot-map.json`](./cpu-hotspot-map.json), [`hotspot-map-successor.md`](./hotspot-map-successor.md)\n\n## Decision\n\nA new native (Rust N-API / FFI) port proposed **to optimize a leftover hot path** does not land unless **all** of the following gates pass:\n\n1. **Corpus evidence** — a profiling-corpus trace shows the path has user-visible latency or RSS impact on a representative workload (not just a static complexity argument).\n2. **Self-time attribution** — a `profilerSelfTime` artifact identifies the proposed hotspot, **or** fallback-toggle evidence proves an end-to-end benefit without byte changes. Wall-clock proxy timing alone is never sufficient.\n3. **Measured FFI overhead** — the N-API call/marshalling overhead is measured against the JS/TS baseline, not assumed away.\n4. **Representative win** — a representative p50/p95 win exists on realistic inputs, not only microbenchmark seed results.\n5. **Byte parity** — a byte-identical corpus covers rendered, persisted, and provider-visible bytes for the changed path.\n6. **Operational cost** — fallback, packaging, and rollback costs are documented.\n\nThis policy governs **speculative algorithmic ports**. It does **not** re-litigate already-native platform/system surfaces (see [Scope boundary](#scope-boundary)).\n\n## Context\n\nThe CPU/memory hotspot program (Optimization Suites v1–v3, tracked in [`cpu-hotspot-map.json`](./cpu-hotspot-map.json)) is closed out. Its prioritization was a **static structural ranking** (algorithmic complexity × trigger frequency), and the map's own `method` field records that real CPU self-time was \"to be measured by the agreed profiling corpus during optimization.\" That corpus is being built separately; until its evidence exists, new native ports for leftover hotspots would repeat the same evidence gap.\n\nThe suites already produced concrete decisions that this policy codifies so they are not re-discovered:\n\n- **v2 (#530)** measured and **rejected the five remaining Rust port candidates** per the FFI cost gates after shipping only `diffLines` (H03) natively. Native overhead did not beat the JS/TS baseline for those candidates on realistic inputs.\n- **v3 (#558) rejected a native word-diff (H04)** \"without a fresh FFI gate\" — the TS fast paths were retained instead; a native port would need to re-clear gates 1–6 above.\n- **Hunt-Szymanski LCS (H05)** was implemented as a native/algorithmic replacement, then **reverted** because it produced byte-different rendered diffs (reproduced by red-team). Byte parity is the gate, not raw speed.\n- **The custom JSON length counter (H08)** was implemented, made exact, then **deleted** — an exact JS reimplementation was not faster than native `JSON.stringify`. \"More native\" is not automatically \"faster.\"\n\nThese four precedents share a root cause: a plausible algorithmic/native win that failed a real gate (cost, byte parity, or end-to-end benefit). The policy makes those gates a precondition rather than a post-hoc discovery.\n\n## Evidence taxonomy\n\nNative-port claims must classify their evidence using the same separated classes as the profiling corpus. These classes must never be conflated:\n\n- **`wallClockPhase`** — elapsed timing around a phase or operation. Useful for perceived-latency and regression detection; **insufficient** to confirm CPU self-time or to justify a port on its own.\n- **`processCpuUsage`** — `process.cpuUsage()` user/system deltas, optionally normalized by elapsed time. Indicates process-level CPU pressure; **cannot** attribute self-time to a specific hotspot.\n- **`profilerSelfTime`** — profiler (or equivalent sampled/trace) attribution of self-time to a function, module, or native symbol. **Required** before a hotspot may be called \"CPU-self-time confirmed.\"\n\nA native-optimization proposal that cites only `wallClockPhase` or `processCpuUsage` is **not** CPU-self-time confirmed and does not clear gate 2.\n\n## Approval checklist\n\nBefore opening a native-optimization PR, confirm and attach evidence for each:\n\n- [ ] Corpus trace shows user-visible latency or RSS impact for the path (gate 1).\n- [ ] `profilerSelfTime` artifact identifies the hotspot, **or** fallback-toggle before/after evidence proves end-to-end benefit without byte changes (gate 2).\n- [ ] FFI/marshalling overhead measured vs the JS/TS baseline in the same benchmark run (gate 3).\n- [ ] Representative p50/p95 win on realistic inputs, not only seeded microbench results (gate 4).\n- [ ] Byte-identical corpus covers rendered, persisted, and provider-visible bytes (gate 5).\n- [ ] Fallback, packaging (platform variants / embedded addon), and rollback costs documented (gate 6).\n\nIf any box is unchecked, keep the work in TypeScript or hold it as a tracked candidate; do not switch callsites. This mirrors the existing **Rule of thumb** in [`porting-to-natives.md`](./porting-to-natives.md): if native is not faster *and* behavior-compatible, do not switch callsites.\n\n## Scope boundary\n\nThis policy targets **speculative algorithmic ports**, not the established native surface. The following are **already native** by design and are explicitly out of scope (see `alreadyNativeExcluded` in [`cpu-hotspot-map.json`](./cpu-hotspot-map.json)):\n\n`grep`, `fd`/`glob`, text width/wrap/truncate/slice, syntax highlighting, HTML→Markdown, token counting, AST, summary, process/PTY/shell, SIXEL, clipboard, `Bun.hash.xxHash32/64`, and `JSON.parse`/`JSON.stringify`.\n\nThese are native because they are I/O, OS/process integration, or platform primitives — the criteria in [`porting-to-natives.md`](./porting-to-natives.md#when-to-port). Distinguishing them from algorithmic ports matters: a leftover algorithmic hotspot must clear gates 1–6, whereas adding a new OS/process/native-primitive binding follows the standard porting guide.\n\n## Consequences\n\n- New native algorithmic ports require profiling-corpus evidence and a measured cost gate before review; this slows speculative optimization but prevents byte-parity regressions and dead native code.\n- The default answer for a leftover hotspot is \"keep it in TypeScript\" until the corpus proves it matters.\n- Already-native platform/system primitives and new OS/process bindings are unaffected; they follow [`porting-to-natives.md`](./porting-to-natives.md) as before.\n- Reviewers can reject a native-optimization PR purely on a missing gate, citing this ADR, without re-deriving the rationale.\n\n## Follow-ups\n\n- Held native candidates (H04 word-diff, H05 LCS, and other v2-rejected candidates) stay held unless a future PR clears gates 1–6 with fresh corpus evidence.\n- When the profiling corpus lands, link its threshold/evidence ledger here so native-port proposals can cite concrete corpus artifacts.\n",
 	"natives-addon-loader-runtime.md": "# Natives Addon Loader Runtime\n\nThis document covers the runtime loader shipped by `@gajae-code/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 the authoritative compiled-binary signal when present.\n- Optionally materialize an embedded addon into a versioned per-user cache directory.\n- Attempt candidates in deterministic order and return the first addon that `require(...)` loads.\n\nThe current loader does **not** run a separate `validateNative(...)` export-presence gate. API shape is provided by the generated N-API binding file (`native/index.d.ts`) and the loaded addon itself. A stale binary therefore normally fails as a missing property or native load error rather than as a custom \"missing exports\" validation error.\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  - `nativeDir`: package-local `packages/natives/native`.\n  - `execDir`: directory containing `process.execPath`.\n  - `versionedDir`: `<getNativesDir()>/<packageVersion>`.\n  - `userDataDir` fallback:\n    - Windows: `%LOCALAPPDATA%/gjc` or `%USERPROFILE%/AppData/Local/gjc`.\n    - Non-Windows: `~/.local/bin`.\n- **Natives cache root** (`getNativesDir()`):\n  - if `$XDG_DATA_HOME/gjc` exists, `$XDG_DATA_HOME/gjc/natives`;\n  - otherwise `~/.gjc/natives`.\n- **Compiled-binary mode** (`detectCompiledBinary`): true if any of:\n  - embedded-addon manifest is non-null,\n  - `GJC_COMPILED` env var is set,\n  - `import.meta.url` contains Bun embedded markers (`$bunfs`, `~BUN`, `%7EBUN`).\n- **Variant override**: `GJC_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-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. `GJC_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\nFor each filename, candidates are:\n\n1. `<nativeDir>/<filename>`\n2. `<execDir>/<filename>`\n\n### Compiled runtime\n\nFor each filename, candidates are:\n\n1. `<versionedDir>/<filename>`\n2. `<userDataDir>/<filename>`\n3. `<nativeDir>/<filename>`\n4. `<execDir>/<filename>`\n\nAt load time, an extracted embedded candidate, when produced, 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- `files[]` entries with `variant`, `filename`, and `filePath`\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. Reuse `<versionedDir>/<selected filename>` if it already exists.\n3. Otherwise read `selectedEmbeddedFile.filePath` and write the target path.\n4. Return the target path as the first candidate.\n\nDirectory creation 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 -> try extract to versionedDir (record errors, continue)\n       no  -> skip extraction\n  -> For each runtime candidate in order:\n       require(candidate)\n       -> success: 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(...)` 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\n### Non-compiled startup failures\n\nNormal package/runtime diagnostics include:\n\n- reinstall hint (`bun install @gajae-code/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",
@@ -115,6 +115,6 @@ export const EMBEDDED_DOCS: Readonly<Record<string, string>> = {
 	"tools/write.md": "# write\n\n> Create or overwrite a file, archive entry, or SQLite row.\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. `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. |\n| `content` | `string` | Yes | Full replacement file content, archive entry content, 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 <bytes> bytes to <relative-path>`.\n  - Archive write: `Successfully wrote <bytes> 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- If hashline prefixes were copied from `read` output and stripped first, the first text block gets an extra note.\n- Plain file writes may also return `details.diagnostics` plus `details.meta.diagnostics` when LSP diagnostics-on-write is enabled.\n- SQLite writes use `toolResult(...).sourcePath(...)`, so `details.meta.sourcePath` points at the database file.\n- Archive writes return empty `details`.\n\n## Flow\n1. `WriteTool.execute()` in `packages/coding-agent/src/tools/write.ts` strips `LINE+ID|` hashline prefixes from `content` when the session is in hashline display mode.\n2. It calls `#resolveArchiveWritePath()` first. 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.\n3. 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.\n4. 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.\n5. 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`.\n6. 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   - 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   - `invalidateFsScanAfterWrite()` runs on the file path.\n7. The tool returns a text result and optional diagnostics 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; parent-directory creation is only implemented in the archive branch.\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  - Creates parent directories for archive files only.\n  - Mutates existing SQLite databases; never creates a new SQLite DB.\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 `nonAbortable = true` and `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\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- 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 writes report byte count using `cleanContent.length`, 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/modes/controllers/input-controller.ts` (`/tree`, 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 action `tree`\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 most conversational nodes, but hides bookkeeping entry types:\n\n- `label`\n- `custom`\n- `model_change`\n- `thinking_level_change`\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 and constructs a `TtsrManager`:\n\n```ts\nconst ttsrSettings = settings.getGroup(\"ttsr\");\nconst ttsrManager = new TtsrManager(ttsrSettings);\nconst rulesResult = await loadCapability<Rule>(ruleCapability.id, { cwd });\nfor (const rule of rulesResult.items) {\n  if (rule.condition?.length && ttsrManager.addRule(rule)) continue;\n  // non-TTSR rules continue through normal rule handling\n}\n```\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- `rule.condition` is absent or all condition regexes fail to compile\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.\n\n### Setting caveat\n\n`TtsrSettings.enabled=false` makes `TtsrManager` no-op (`addRule`, `checkDelta`, `hasRules`, `restoreInjected`), so no conditional rules are registered or injected.\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- append delta into a source/tool scoped manager buffer\n- call `checkDelta(delta, matchContext)`\n\n`checkDelta()` iterates registered rules and returns all matching rules that pass scope, global-path, condition, and repeat policy checks.\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().catch(() => {})` swallows follow-up errors.\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- `contextMode: \"keep\"`: partial violating output can remain in context before reminder retry. **Cost warning:** every aborted partial turn is retained, so context (and token spend) grows each time a rule fires. Prefer the default `discard` unless the partial output is specifically needed.\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-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## 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- [`../src/modes/interactive-mode.ts`](../packages/coding-agent/src/modes/interactive-mode.ts)\n- [`../src/modes/controllers/event-controller.ts`](../packages/coding-agent/src/modes/controllers/event-controller.ts)\n- [`../src/modes/controllers/input-controller.ts`](../packages/coding-agent/src/modes/controllers/input-controller.ts)\n- [`../src/modes/components/custom-editor.ts`](../packages/coding-agent/src/modes/components/custom-editor.ts)\n- [`../../tui/src/tui.ts`](../packages/tui/src/tui.ts)\n- [`../../tui/src/terminal.ts`](../packages/tui/src/terminal.ts)\n- [`../../tui/src/editor-component.ts`](../packages/tui/src/editor-component.ts)\n- [`../../tui/src/stdin-buffer.ts`](../packages/tui/src/stdin-buffer.ts)\n- [`../../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 `settings.get(\"clearOnShrink\")`, and creates persistent containers:\n\n- `chatContainer`\n- `pendingMessagesContainer`\n- `statusContainer`\n- `todoContainer`\n- `btwContainer`\n- `statusLine`\n- `hookWidgetContainerAbove`\n- `editorContainer` (holds `CustomEditor`)\n- `hookWidgetContainerBelow`\n\n`init()` wires the tree in that order, focuses the editor, registers input handlers via `InputController`, subscribes terminal appearance changes into theme auto-detection, starts TUI, and requests a forced render.\nA forced render (`requestRender(true)`) resets previous-line caches and cursor bookkeeping before repainting.\n\n## Terminal lifecycle and stdin normalization\n\n`ProcessTerminal.start()`:\n\n1. Enables raw mode and bracketed paste.\n2. Attaches resize handler.\n3. Creates a `StdinBuffer` to split partial escape chunks into complete sequences.\n4. Queries Kitty keyboard protocol support (`CSI ? u`), then enables protocol flags if supported; otherwise enables modifyOtherKeys fallback after a short timeout.\n5. Queries OSC 11 background color and enables Mode 2031 appearance notifications for dark/light theme detection.\n6. On Windows, attempts VT input enablement via `kernel32` mode flags.\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 diffing strategy\n\n`TUI.requestRender()` is debounced to one render per tick using `process.nextTick`. Multiple state changes in the same turn coalesce.\n\n`#doRender()` pipeline:\n\n1. Render root component tree to `newLines`.\n2. Composite visible overlays (if any).\n3. Extract and strip `CURSOR_MARKER` from visible viewport lines.\n4. Append segment reset suffixes for non-image lines.\n5. Choose full repaint vs differential patch:\n   - first frame\n   - width change\n   - shrink with `clearOnShrink` enabled and no overlays\n   - edits above previous viewport\n6. For differential updates, patch only changed line range and clear stale trailing lines when needed.\n7. Reposition hardware cursor for IME support.\n\nRender writes use synchronized output mode (`CSI ? 2026 h/l`) to reduce flicker/tearing.\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 full redraw 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\n## Virtual viewport (default-on, `PI_TUI_VIRTUAL_VIEWPORT`)\n\nBy default `#doRender` reuses the previous normalized off-screen prefix and only normalizes/diffs the visible window (terminal rows + a small overscan) when the terminal width is unchanged and the off-screen raw prefix is unchanged from the previous frame. This bounds steady-state append/edit work on very long sessions / weak hardware while preserving byte-identical output.\n\nSet `PI_TUI_VIRTUAL_VIEWPORT=0` (or `false`) to opt out and restore the legacy path that normalizes/truncates and diffs the full rendered transcript every frame (`O(total lines)`). The fast path compares the off-screen raw prefix by raw value equality per line, which short-circuits to a fast reference check when components return stable string instances for unchanged lines; reused entries are deterministic normalizations of identical raw lines. Any width change, off-screen edit, forced render (`requestRender(true)`), or first frame transparently falls back to the full path. `PI_TUI_METRICS` exposes `lineCounts` gauges (`rendered`, `normalized`, `measured`, `diffed`, and `offscreenScan`) to observe the bound.\n\n## Resize handling\n\nResize events are event-driven from `ProcessTerminal` to `TUI.requestRender()`.\n\nEffects:\n\n- Width changes trigger full redraw.\n- Height changes trigger full redraw except in Termux and terminal multiplexers, where the renderer avoids scrollback-hostile full replays.\n- Viewport/top tracking (`#previousViewportTop`, `#maxLinesRendered`) avoids invalid relative cursor math when content or terminal size changes.\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 terminal completion notification if backgrounded, and runs the user-level completion command hook when configured.\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` updates every 80ms via interval and requests render each frame.\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### Background mode (`/background` or `/bg`)\n\n`handleBackgroundCommand()`:\n\n- Rejects when idle.\n- Switches tool UI context to non-interactive (`hasUI=false`) so interactive UI tools fail fast.\n- Stops loaders/status line and unsubscribes foreground event handler.\n- Subscribes background event handler (primarily waits for `agent_end`).\n- Stops TUI and sends `SIGTSTP` (POSIX job control path).\n\nOn `agent_end` in background with no queued work, controller sends the terminal completion notification and shuts down. The configured `completion.notifyCommand` hook is not limited to background mode; it runs on completed agent turns so local tools such as cmux can receive foreground and background completion events.\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 fixed-interval (80ms), each frame requesting 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-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## 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- [`../src/modes/interactive-mode.ts`](../packages/coding-agent/src/modes/interactive-mode.ts)\n- [`../src/modes/controllers/event-controller.ts`](../packages/coding-agent/src/modes/controllers/event-controller.ts)\n- [`../src/modes/controllers/input-controller.ts`](../packages/coding-agent/src/modes/controllers/input-controller.ts)\n- [`../src/modes/components/custom-editor.ts`](../packages/coding-agent/src/modes/components/custom-editor.ts)\n- [`../../tui/src/tui.ts`](../packages/tui/src/tui.ts)\n- [`../../tui/src/terminal.ts`](../packages/tui/src/terminal.ts)\n- [`../../tui/src/editor-component.ts`](../packages/tui/src/editor-component.ts)\n- [`../../tui/src/stdin-buffer.ts`](../packages/tui/src/stdin-buffer.ts)\n- [`../../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 `settings.get(\"clearOnShrink\")`, and creates persistent containers:\n\n- `chatContainer`\n- `pendingMessagesContainer`\n- `statusContainer`\n- `todoContainer`\n- `btwContainer`\n- `statusLine`\n- `hookWidgetContainerAbove`\n- `editorContainer` (holds `CustomEditor`)\n- `hookWidgetContainerBelow`\n\n`init()` wires the tree in that order, focuses the editor, registers input handlers via `InputController`, subscribes terminal appearance changes into theme auto-detection, starts TUI, and requests a forced render.\nA forced render (`requestRender(true)`) resets previous-line caches and cursor bookkeeping before repainting.\n\n## Terminal lifecycle and stdin normalization\n\n`ProcessTerminal.start()`:\n\n1. Enables raw mode and bracketed paste.\n2. Attaches resize handler.\n3. Creates a `StdinBuffer` to split partial escape chunks into complete sequences.\n4. Queries Kitty keyboard protocol support (`CSI ? u`), then enables protocol flags if supported; otherwise enables modifyOtherKeys fallback after a short timeout.\n5. Queries OSC 11 background color and enables Mode 2031 appearance notifications for dark/light theme detection.\n6. On Windows, attempts VT input enablement via `kernel32` mode flags.\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 diffing strategy\n\n`TUI.requestRender()` is debounced to one render per tick using `process.nextTick`. Multiple state changes in the same turn coalesce.\n\n`#doRender()` pipeline:\n\n1. Render root component tree to `newLines`.\n2. Composite visible overlays (if any).\n3. Extract and strip `CURSOR_MARKER` from visible viewport lines.\n4. Append segment reset suffixes for non-image lines.\n5. Choose a viewport repaint, full repaint, or differential patch:\n   - real process terminals repaint the visible viewport for width/height changes, forced renders, and edits above the live viewport so native scrollback is not cleared/replayed; host markers refine policy only after this process-terminal capability is established;\n   - virtual/headless terminals retain full clear/replay regardless of inherited terminal-host environment markers, keeping historical buffer repair deterministic;\n   - steady-state visible changes use differential patches, including viewport repaint when a contraction exposes earlier transcript rows.\n6. For differential updates, patch only changed line ranges and clear stale trailing lines when needed.\n7. Reposition hardware cursor for IME support.\n\nRender writes use synchronized output mode (`CSI ? 2026 h/l`) to reduce flicker/tearing.\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 re-render wrapped content; real process terminals limit emission to the visible viewport because their native scrollback position is not observable.\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\n## Virtual viewport (default-on, `PI_TUI_VIRTUAL_VIEWPORT`)\n\nBy default `#doRender` reuses the previous normalized off-screen prefix and only normalizes/diffs the visible window (terminal rows + a small overscan) when the terminal width is unchanged and the off-screen raw prefix is unchanged from the previous frame. This bounds steady-state append/edit work on very long sessions / weak hardware while preserving byte-identical output.\n\nSet `PI_TUI_VIRTUAL_VIEWPORT=0` (or `false`) to opt out and restore the legacy path that normalizes/truncates and diffs the full rendered transcript every frame (`O(total lines)`). The fast path compares the off-screen raw prefix by raw value equality per line, which short-circuits to a fast reference check when components return stable string instances for unchanged lines; reused entries are deterministic normalizations of identical raw lines. Any width change, off-screen edit, forced render (`requestRender(true)`), or first frame transparently falls back to the full path. `PI_TUI_METRICS` exposes `lineCounts` gauges (`rendered`, `normalized`, `measured`, `diffed`, and `offscreenScan`) to observe the bound.\n\n## Resize handling\n\nResize events are event-driven from `ProcessTerminal` to `TUI.requestResizeRender()`.\n\nEffects:\n\n- Real process terminals repaint only the visible viewport on width/height changes, avoiding scrollback-hostile clear/replay cycles; known host markers and the legacy multiplexer override refine this process-terminal policy.\n- Virtual/headless terminals retain full redraw regardless of inherited host markers for deterministic buffer repair.\n- Viewport/top tracking avoids invalid relative cursor math when content or terminal size changes.\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 terminal completion notification if backgrounded, and runs the user-level completion command hook when configured.\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` updates every 80ms via interval and requests render each frame.\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### Background mode (`/background` or `/bg`)\n\n`handleBackgroundCommand()`:\n\n- Rejects when idle.\n- Switches tool UI context to non-interactive (`hasUI=false`) so interactive UI tools fail fast.\n- Stops loaders/status line and unsubscribes foreground event handler.\n- Subscribes background event handler (primarily waits for `agent_end`).\n- Stops TUI and sends `SIGTSTP` (POSIX job control path).\n\nOn `agent_end` in background with no queued work, controller sends the terminal completion notification and shuts down. The configured `completion.notifyCommand` hook is not limited to background mode; it runs on completed agent turns so local tools such as cmux can receive foreground and background completion events.\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 fixed-interval (80ms), each frame requesting 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",
 	"ui-design-visual-qa.md": "# UI design and visual QA workflow\n\nThis is the repo-owned contract for future Gajae-Code UI, web, dashboard, terminal, and TUI visual work. It adapts the useful OMO design-reference and visual-QA workflow without vendoring any third-party design corpus.\n\nIt is not a fifth bundled workflow skill. Gajae-Code's public workflow surface remains `deep-interview`, `ralplan`, `ultragoal`, and `team`; use this document as planning/review guidance inside those workflows or direct implementation.\n\n## Required branch before implementation\n\nBefore writing broad UI code, choose and record exactly one workflow branch in the plan, issue update, PR body, or local `DESIGN.md`:\n\n1. **Existing design system** — the repository or product area already has a `DESIGN.md`, component system, theme, or comparable visual grammar. Read it first and update it when the change extends the system.\n2. **Greenfield selected references** — no usable system exists, so pick a small shortlist of design references before implementation. Record the exact references loaded and translate them into first-party project guidance; do not copy raw vendor files into the repo.\n3. **Extract existing system first** — the surface exists but its rules are implicit. The first deliverable is a minimal `DESIGN.md` that extracts the current tokens, layout grammar, component anatomy, and state rules before new product screens are built.\n\nA UI task that skips this branch selection is not ready for implementation.\n\n## `DESIGN.md` source material\n\nNew UI work must create or update the nearest product `DESIGN.md` before broad product-screen implementation. The document must be first-party source material for the implementation, not a screenshot dump or mood board. Include:\n\n- **Tokens** — colors, typography, spacing, radii, borders, iconography, density, terminal color roles, and accessibility contrast constraints.\n- **Layout grammar** — grids, page regions, navigation, hierarchy, rhythm, empty space, alignment, and composition rules.\n- **Component anatomy** — named parts, slots, content rules, affordances, constraints, and when not to use the component.\n- **States** — default, hover, active, focus, disabled, loading, empty, error, selected, expanded/collapsed, and permission/connection failure states where relevant.\n- **Motion and depth** — animation timing/easing, transitions, shadows/elevation, overlays, focus movement, and reduced-motion behavior.\n- **Responsive behavior** — mobile, tablet, desktop, narrow terminal, wide terminal, font scaling, wrapping, overflow, and high-density layouts.\n\nFor greenfield selected references, `DESIGN.md` must name the selected references and explain what was translated from each into the first-party system. It must not embed raw third-party corpus text as the system of record.\n\n## Component showcase before product screens\n\nBuild or update a component showcase/state harness before implementing broad product screens. The harness may be Storybook, a local route, a CLI/TUI fixture, a docs page with runnable examples, or a purpose-built screenshot fixture, but it must expose the component states needed by the product surface.\n\nThe harness must cover:\n\n- all states listed in `DESIGN.md` that the component supports;\n- representative realistic content, including long strings, empty content, error copy, and localization-sensitive text;\n- mobile/tablet/desktop or narrow/medium/wide terminal layouts as applicable;\n- keyboard focus and accessibility-visible states for interactive components.\n\nProduct screens can follow only after the harness proves the component vocabulary is stable enough to reuse.\n\n## Visual QA contract\n\nVisual QA is completion evidence, not decoration. A UI story is not done until it has fresh full-surface evidence from the current branch.\n\nRequired evidence:\n\n- **Fresh capture** — evidence must be generated after the current implementation, not reused from an older branch, design reference, or unrelated run.\n- **Full-surface coverage** — capture every page, route, modal, drawer, tab, breakpoint, component state, and error/empty/loading state in scope. Sampling a few representative screenshots is not enough.\n- **No hidden tails** — scrollable surfaces require evidence for top, middle, bottom, sticky regions, overflow behavior, and any virtualized content boundaries.\n- **CJK semantic line breaks block completion** — Korean, Japanese, Chinese, and mixed CJK/Latin copy must not wrap in semantically broken or visually misleading ways. Bad CJK line breaks are blocking defects, not cosmetic polish.\n- **Independent review before done** — a reviewer or review lane that did not author the implementation must inspect the full evidence set against `DESIGN.md`, the component harness, and acceptance criteria before the work is marked complete.\n- **Evidence references** — PRs should link or attach the current evidence artifacts instead of describing them only in prose.\n\n## Terminal and TUI evidence\n\nTerminal/TUI visual work must preserve terminal semantics in future helper flows. Until Gajae-Code has a dedicated helper, terminal visual-QA evidence must document the exact helper requirements and avoid flattening away the data needed for review.\n\nA terminal/TUI evidence helper must produce, at minimum:\n\n- `terminal.txt` with readable plain text;\n- `terminal-ansi.txt` preserving ANSI SGR color/style sequences and terminal control semantics needed for replay/review;\n- `terminal.html` rendering the ANSI-styled output for browser review;\n- optional `terminal.png` generated from the styled rendering when image evidence is useful;\n- `metadata.json` containing command or replay source, terminal size, font/rendering assumptions, capture timestamp, tool version, wrapping/truncation policy, and whether the artifact is from a live PTY, replay, or fixture.\n\nReviewers must reject terminal/TUI evidence that only contains flattened text when the change depends on color, emphasis, cursor state, layout, wrapping, or other ANSI/terminal behavior.\n\n## Provenance boundary\n\nDo not vendor raw third-party design corpora, screenshots, brand guides, prompt packs, or critique/reference material into this repository without an explicit materializer and provenance design.\n\nAllowed in a first PR:\n\n- first-party `DESIGN.md` guidance written for Gajae-Code;\n- a small hand-written reference index that names public sources and records why they were consulted;\n- pinned links, citations, or manifests that do not copy raw third-party corpus content.\n\nRequires explicit design before inclusion:\n\n- raw third-party corpus files;\n- copied design-system text, screenshots, or asset packs;\n- submodules or generated materialized corpora;\n- generated screenshots committed as durable source assets.\n\nA future materializer design must define source ownership, license/provenance metadata, pinning strategy, reproducible generation, update review, generated-file boundaries, and what artifacts are safe to attach to GitHub without committing.\n\n## PR checklist\n\nFor UI work, include this checklist in the PR body or equivalent review artifact:\n\n- [ ] UI workflow branch selected: existing `DESIGN.md`, greenfield selected references, or extract existing system first.\n- [ ] `DESIGN.md` created/updated with tokens, layout grammar, component anatomy, states, motion/depth, and responsive behavior.\n- [ ] Component showcase/state harness exists before broad product screens.\n- [ ] Fresh full-surface visual evidence covers every in-scope page/state/breakpoint; no sampling.\n- [ ] CJK semantic line breaks were checked and any defects were fixed before completion.\n- [ ] Terminal/TUI evidence preserves ANSI semantics or documents the exact helper requirements above.\n- [ ] Independent review inspected the evidence before the work was marked done.\n- [ ] No raw third-party corpus was vendored without an explicit materializer/provenance design.\n",
 };