@loompad/cli 0.1.0 → 1.1.2

package/docs/proposals/t3code-port.md

@@ -0,0 +1,81 @@
+# Porting t3code's core into Loom
+
+Status: plan, agreed 2026-09-28. Functionality first; UI after.
+
+Loom takes t3code's provider and turn machinery (MIT, attributed per file) and keeps
+its own app, orchestra, teams, phone client and Brain. The port is plain TypeScript;
+t3code's Effect layers become classes and async functions. The differentiator is
+what t3code refuses: switching provider mid-chat
+([`ProviderCommandReactor`](../../../t3code/apps/server/src/orchestration/Layers/ProviderCommandReactor.ts):
+"is bound to driver X and cannot switch to Y"). Loom's Brain makes that switch.
+
+## Decisions
+
+- **Port, not fork.** Loom's daemon, clients and features stay; the provider layer
+  and turn lifecycle are replaced with t3code's design.
+- **Warm sessions.** One live provider session per chat, reused across turns
+  (app-server process, Agent SDK streaming query), with an idle reaper. Brain's
+  "turn owns a process group" rule becomes turn-level settlement: a turn is settled
+  when the harness reports it finished and no commands it started are running.
+- **Normalized events at the adapter boundary.** Adapters emit one canonical runtime
+  event stream; nothing above them branches on provider.
+
+## What t3code has that Loom lacks
+
+| Area | t3code | Loom today |
+| --- | --- | --- |
+| Adapter contract | `startSession`, `sendTurn`, `interruptTurn`, `respondToRequest`, `respondToUserInput`, `readThread`, `rollbackThread`, `compaction`, `stopSession`, capabilities | `send`, `interrupt`, `available` |
+| Runtime events | ~50 normalized types: session/turn/item lifecycle, `content.delta` streaming, requests, user input, plans, diffs, tasks, hooks, account, MCP | ~10 coarse kinds, completed messages only |
+| Sessions | warm, directory + reaper, per instance | one process per turn |
+| Questions | structured user input, blocking and async (Codex) | `needs_input` heuristic on a trailing `?` |
+| Plans | plan items, proposed plan, plan-mode exit | plan-mode briefing text |
+| Revert | hidden-ref checkpoints + conversation rollback, refused when the provider can't roll back | checkpoints restore files only |
+| Models | catalog per provider (`model/list`, Claude manifest), effort/thinking | free-text model override |
+| Compaction | automatic + manual start | automatic only (reported) |
+| Providers | instances (several accounts per driver), auth status, install/update ownership | one agent per config entry |
+
+## Phases
+
+Each phase ends with fakes-only tests passing and one authorized live check per
+harness.
+
+**1. Provider contract and service.** `src/providers/` in t3code's shape: a
+`ProviderAdapter` interface with capabilities, the canonical `ProviderRuntimeEvent`
+union (the subset Loom uses first, extended as phases need it), a `ProviderService`
+that routes by instance and owns a session directory and reaper, and an ingestion
+layer that projects runtime events into Loom's event log. Existing adapters are
+wrapped, so nothing user-facing changes yet.
+
+**2. Warm sessions for Codex and Claude.** Codex: one `app-server` per chat, thread
+kept open, turns started and interrupted over the same connection. Claude: one
+streaming-input `query()` per chat, turns pushed as user messages, `setModel` for
+in-provider model switches. Turn-level settlement replaces process-group settlement
+in Brain. Token streaming via `content.delta`, throttled on the websocket.
+
+**3. Interaction parity.** Tool items with start/progress/complete and streamed
+output; approvals keeping native option ids; structured user input (Claude
+`AskUserQuestion`, Codex `requestUserInput`, Codex async questions); plan items and
+plan-mode exit; per-turn diffs; manual compaction; model catalogs with effort and
+thinking settings.
+
+**4. Checkpoints and revert.** Hidden-ref workspace checkpoints per turn; revert
+coordinated with native rollback (Codex thread rollback, Claude resume-at);
+refused before touching files when the provider can't roll back.
+
+**5. Cross-provider switching (the differentiator).** Switching provider in a chat
+parks the current warm session and starts or resumes the target's, with Brain
+building what the target lacks. Budget by the target's context window; offer a
+switch when a usage limit is reached; revert across a provider boundary rolls back
+each side it touched. Brain continuity becomes the default path, not opt-in.
+
+**6. Provider management.** Instances (several accounts per driver), auth status,
+install/update detection by owning installer, importing native session history.
+
+**7. UI overhaul.** Built on the settled event model.
+
+## Not ported
+
+Relay/T3 Connect, mobile app, desktop shell, devices, voice, analytics and
+marketing: Loom has its own equivalents or doesn't need them now. Other providers
+(Cursor, Grok, OpenCode, Antigravity) move to the new contract after Codex and
+Claude.

package/docs/proposals/t3code-reference.md

@@ -0,0 +1,95 @@
+# T3 Code reference for Loom
+
+Inspected 2026-09-27. Read-only checkout: `/Volumes/Programming Vault/t3code`.
+Branch: `main`. Commit: `de251fc2971a884cb5b1305ba4daf309dc8cccb0`
+(`fix(web): continue onboarding after incomplete history imports (#13935)`).
+Read contributor and agent instructions. No dependencies installed, processes
+started, benchmarks run, or T3 source changed. This is a targeted architecture
+review, not an audit of the entire repository.
+
+## Findings and implications
+
+### Native sessions still have provider boundaries
+
+The [command reactor](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/orchestration/Layers/ProviderCommandReactor.ts#L688)
+rejects an existing thread switching to another driver. Even instances of the
+same driver must have matching `continuationIdentity.continuationKey` values.
+Model selection and native session compatibility are separate concerns.
+
+**Loom:** keep the app conversation independent of native sessions. Validate
+resume compatibility per instance. Cross-harness switching needs a fresh or
+compatible target session and our own source-backed handoff. This checkout does
+not solve shared context for us. PR #3799 was inspected separately; do not
+describe its unmerged code as behavior present here.
+
+### Cheap helper calls through harnesses are concrete
+
+[Model defaults](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/packages/contracts/src/model.ts#L164)
+select `gpt-6-luna` for Codex text generation and `claude-haiku-4-5` for Claude.
+The [TextGeneration interface](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/textGeneration/TextGeneration.ts#L87)
+provides commit messages, PR content, branch names and thread titles.
+
+- [Codex](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/textGeneration/CodexTextGeneration.ts#L197)
+  uses `exec --ephemeral`, a read-only sandbox and an output schema.
+- [Claude](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/textGeneration/ClaudeTextGeneration.ts#L199)
+  requests structured JSON, disables tools and slash commands, and specifies
+  strict MCP configuration. Thread title calls use a temporary working directory.
+
+**Loom:** reuse the architectural pattern for a small `ContextHelper` interface.
+Summary extraction and source validation are new responsibilities. Discover
+available models, bound usage, isolate helper sessions and preserve deterministic
+fallbacks. A read-only sandbox alone does not mean a helper has no tools or no
+access to unrelated readable files. These defaults do not prove account access,
+subscription savings or identical restrictions across harnesses.
+
+### Streaming efficiency is an explicit design concern
+
+The [WebSocket server](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/ws.ts)
+batches shell/sidebar projection events, coalesces updates by aggregate and keeps
+survivors ordered by sequence. Synchronization markers cannot overtake pending
+events. [Client shell state](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/packages/client-runtime/src/state/shell.ts)
+applies a batch with a single state write. Chat rendering uses a virtualized list.
+
+**Loom:** persist canonical events independently of UI update frequency. Coalesce
+replaceable projections only; preserve messages, decisions, tool outcomes and
+delivery acknowledgements. Test ordering and reconnect boundaries. These are
+useful performance mechanisms, not measured proof that every screen is fast.
+
+### Electron supervises a backend process
+
+[DesktopBackendManager](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/desktop/src/backend/DesktopBackendManager.ts)
+spawns backend processes, tracks lifecycle/readiness and schedules recovery after
+unexpected exits. Desktop, server and web are separate applications in the repo.
+
+**Loom:** keep harness execution, SQLite, indexing and helper scheduling outside
+the renderer. Specify startup, cancellation, shutdown and crash recovery before
+moving the UI to Electron. Electron packaging alone does not provide performance.
+Adopting T3's full framework stack is not required to adopt these boundaries.
+
+### SQLite and native compaction have distinct roles
+
+The [SQLite layer](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/persistence/Layers/Sqlite.ts)
+enables foreign keys, WAL, a busy timeout and migrations. Its journal-size setting
+controls retained WAL size after reset, not a hard cap on all in-flight WAL growth.
+
+The [Claude adapter](https://github.com/pingdotgg/t3code/blob/de251fc2971a884cb5b1305ba4daf309dc8cccb0/apps/server/src/provider/Layers/ClaudeAdapter.ts#L3628)
+observes native compaction boundaries and updates usage/thread state. Observing a
+native compact event does not create portable cross-provider memory.
+
+**Loom:** SQLite remains the canonical app-owned event and continuity store;
+JSONL is an export format. Preserve originals across native compaction and track
+what was actually delivered to each session separately from helper summaries.
+
+## Next work, when implementation starts
+
+- [x] Inspect session compatibility, helper isolation, persistence and desktop boundaries.
+- [x] Feed verified constraints into the [continuity plan](brain-session-continuity.md).
+- [ ] Verify installed harness capabilities and helper model availability.
+- [ ] Prototype sequential switching with exact user history and delivery receipts.
+- [ ] Test corrections, small decisions, return-to-old-session deltas and interruption.
+- [ ] Measure token overhead and continuation quality before adding vector retrieval.
+- [ ] Benchmark UI streaming and backend recovery before choosing optimizations.
+
+Pin any future T3 comparison to its commit. Read the local MIT license before
+copying source and retain required attribution if code is reused. No source was
+copied as part of this review.

package/docs/refactoring/AGENT-TEAMS-PROGRESS.md

@@ -0,0 +1,88 @@
+# Agent Teams implementation and Windows audit
+
+User scope confirmed on 2026-10-15: finish Windows reliability, then implement
+Agent Teams P2–P5 in separate reviewable PRs. Humans merge the PRs. Hosted
+acceptance and trusted signing issues remain open when external access is missing.
+The feature contract is [agent-teams.md](../proposals/agent-teams.md).
+
+## Windows foundation — PR #218
+
+Native Job Object containment, OpenCode 1.19 protocol parity, contained spec runs
+and dev servers, IPv6 localhost readiness, and source installation checks are
+implemented. Real Codex/OpenCode sequential native resume across an OS daemon
+restart, parallel Orchestra, and the sequential Ship crew passed on Windows.
+
+The audit found a native race in node-pty 1.1.0 during simultaneous terminal
+exits. The dependency is pinned to 1.2.0-beta.15, containing upstream #922. The
+DLL backend also leaves its output worker alive after public `onExit` (#887);
+Loom now drains that connection with a compatibility bridge for the pinned
+version, awaits shutdown, and ignores stale terminal callbacks after id reuse.
+45 ConPTY checks and 42 pipe checks passed; a standalone terminal process exits
+normally; production staging verifies Job Objects and ConPTY execution/shutdown.
+All ten CI jobs passed on Windows PR head bd64d26.
+
+Automatic approval review rejected launching the packaged desktop UI. Do not
+retry that launch. The packaged daemon has CLI acceptance evidence. Installers
+are unsigned, and current desktop UI acceptance remains incomplete.
+
+## P2 — persistent teammates
+
+Branch: `feat/crew-persistent-teammates`, stacked on the Windows PR.
+
+- [x] Native binding per seat, delta on the second card.
+- [x] Stable checkout across goals, distinct goal branches.
+- [x] Graceful mid-goal restart resumes the binding; uncertain actions require
+  explicit receipt reconciliation before replay.
+- [x] Explicit teammate refresh starts a new epoch.
+- [x] Crew memory scope, retrieval isolation and human promotion to project memory.
+- [x] Native fake acceptance, HTTP route checks, regression/DOM checks and CI (all ten jobs on 2cb501b).
+
+Local validation: build/typecheck passed; native restart/delta/recovery tests
+passed (3), crew/memory HTTP and the Crew DOM acceptance passed, and the existing
+crew/brain regressions passed. Full CI remains a separate release gate.
+
+## P3 — parallel and safe
+
+- [x] Builder worktrees and in-crew leases, with overlap holds and resolution.
+- [x] Race and Fan-out plays reuse Orchestra/worktree mechanisms.
+- [x] Merge-tree prediction before integration; conflicts return to the owner.
+- [x] Explicit reviewer/tester results; absent protocol output cannot approve.
+- [x] Stop and test timeout quiesce all tool descendants on Windows.
+- [x] Overlap, conflict, stop, race and fan-out integration tests.
+
+Local validation: build/typecheck passed. Nine parallel-flow checks and the
+existing crew/HTTP suites passed (24 checks). Orchestra, native continuity and
+Crew DOM acceptance passed (30 more checks). Native fan-out holds two distinct
+workspace receipts and integrates both results. All ten CI jobs passed on P3 head 7e1b326; the corrected POSIX fixture retains detached-child coverage on Windows.
+
+## P4 — several people
+
+- [x] Crew/teammate metadata in presence and Fleet, without transcript sharing.
+- [x] Cross-member requests with owner acceptance/decline.
+- [x] Runner-hosted crews, policy limits and role permission ceilings.
+- [x] Per-crew/member/goal budget accounting and visible budget holds.
+- [x] Two-member overlapping work acceptance using the existing hub harness.
+- [x] No hosted migration or landing/merge-policy changes without owner approval.
+
+## P5 — usable from desktop and phone
+
+- [x] Templates, parallel seat activity, native epochs, per-seat/goal/daily spend, and crew metadata on the shared board.
+- [x] Phone Crew view: start/approve/watch, pause/resume/stop, race selection, refresh, budget and team/runner controls.
+- [x] Crew deep links and notifications for budget/lease holds, race choices and incoming owner requests; docs.
+- [x] Build/typecheck, HTTP, four Crew DOM flows and 76 phone model checks.
+- [ ] Complete exact-head CI and physical native phone acceptance. Browser mobile layout inspected at 390x844; it does not establish native-device acceptance.
+
+P5 local evidence: 96 existing board/shell/team regressions and 24 cost/crew/parallel checks passed. Four Crew
+DOM flows cover planning/approval/apply, Stop/Resume, Race/pick/play/pause, and
+parallel-seat/owner-request/runner-control UI. Both app and daemon compile.
+Mechanical UI scan is clear; desktop and phone browser layouts inspected.
+
+## Separate open issues
+
+- #200: PR #224 aligns phone GitHub sign-in and validates PKCE callbacks.
+  Hosted redirect allow-list and physical phone acceptance still need
+  hosted/admin and device access. The issue stays open.
+- #199: public read-only links require dedicated authorization, viewer,
+  expiry/revocation and hosted acceptance; paired encrypted relay is not public sharing.
+- #70: trusted signing certificates and installed release-to-release acceptance.
+- #215 stays open until every phase's acceptance criteria are met.

package/docs/refactoring/BOUNDARIES.md

@@ -0,0 +1,120 @@
+# Brain-ready module boundaries
+
+## Native continuity implementation
+
+The prerequisite refactor below is complete. The opt-in native path now lives in
+`src/core/continuity/` and is integrated through RuntimeTurns. See
+[current architecture](../../ARCHITECTURE.md), [native continuity](../brain-continuity.md)
+and [remaining gates](BRAIN-TODO.md). SQLite mutations stay on EventLog's existing
+connection. Packet assembly and dispatch do not create another database writer;
+client diagnostics remain downstream of persistence. Legacy behavior is available
+when continuity is disabled.
+
+## Objective
+
+Prepare Loom for the Brain continuity rewrite without changing its framework,
+client protocols, persisted history, or native harness behavior. This is the
+prerequisite work in `docs/proposals/brain-session-continuity.md`, not the new
+continuity implementation. Work stays on `dev/Harsha`.
+
+## Work checklist
+
+- [x] Read contributor instructions and trace current ownership.
+- [x] Separate event contracts from SQLite/JSONL implementations; retain the
+  public EventLog entrypoint and test both backends.
+- [x] Give conversation metadata its own storage module; preserve implicit Main,
+  chat bindings, deleted-chat history and existing state.json compatibility.
+- [x] Isolate native adapter state behind per-instance storage; keep native IDs
+  separate from app conversation IDs and preserve existing resume data.
+- [x] Give adapter subscriptions/startup/retirement explicit lifecycle ownership;
+  reject events from replaced instances and prevent duplicate startup.
+- [x] Isolate Brain helper dependencies and handoff preparation from execution;
+  prevent background extraction from writing after shutdown.
+- [x] Centralize client delivery and scope filtering; a failed client must not
+  break persistence or delivery to another client.
+- [x] Add meaningful lifecycle/compatibility regressions, run focused tests,
+  typecheck, build and broader tests; record results and remaining limitations.
+
+## Invariants
+
+- Persist original events before notifying consumers. UI delivery is an observer.
+- Brain depends on an event interface, not a storage implementation or UI.
+- Native adapters translate harness behavior; they do not manage conversations.
+- Runtime coordinates execution, permissions, baton and working-tree ownership.
+- Shutdown/retirement invalidates subscriptions and late helper results.
+- No paid verification, framework migration, new summarizer or vector model.
+- Existing direct-model and bridge features remain compatible; removing them is
+  a separate product migration, not a prerequisite for these seams.
+
+## Findings
+
+- The previous extraction left chat storage and handoff preparation in runtime.
+- Five harness adapters duplicate state.json reads/writes for native sessions.
+- Replaced adapters retain callbacks and model replacement retains started state.
+- Brain extraction and semantic startup can outlive the project log.
+- WebSocket fanout repeats scope logic and allows one send failure to escape.
+
+## Verification
+
+- Initial focused run was blocked by sandbox loopback restrictions. Re-ran with
+  local socket access: 112 tests passed; subsequent expanded run: 129 passed.
+- `npm run build`: passed, including browser assets and TypeScript compilation.
+- Full `npm test`: backend passed 1,357 tests, 17 skipped; all 125 DOM tests
+  passed. Total: **1,482 passed, 17 skipped**, with no failed suites.
+- Final `npm run typecheck`, `git diff --check` and whitespace checks on new
+  files: passed. No paid provider checks or Electron GUI runs were performed.
+- Temporary logs: `/tmp/loom-boundary-focused.log`,
+  `/tmp/loom-boundary-build.log`, `/tmp/loom-boundary-full.log`.
+
+Completed on 2026-09-27. The prerequisite checklist above is complete; the next
+work is the sequential Brain continuity flow described in the proposal.
+
+## Resulting ownership
+
+| Responsibility | Owner | Callers depend on |
+| --- | --- | --- |
+| Canonical event persistence | `core/eventlog.ts`, `core/events/*` | EventJournal/EventReader; existing EventLog facade remains public |
+| App conversation metadata | `core/conversations.ts` | Chat operations; runtime validates bindings |
+| Native resume/process metadata | `core/agent-state.ts` | Per-instance read/patch via AdapterBase |
+| Configured live adapter instances | `daemon/runtime/agents.ts` | Install/start/retire/close; read-only roster |
+| Context selection and preparation | `daemon/runtime/briefings.ts` | Prepared memory/briefing and injected helper implementations |
+| Execution and workspace policy | ProjectRuntime and RuntimeTurns | Existing runtime commands |
+| Client frame delivery | `daemon/delivery.ts` | Project/log/admin audiences and transport connections |
+
+## Correctness fixes included
+
+- A failed event observer cannot make a durable append appear to fail or stop
+  later observers. Caller/subscriber mutations cannot alter canonical history;
+  SQLite and JSONL both return detached data. Closing is idempotent.
+- Retired instances are unsubscribed immediately. A replacement waits for the
+  predecessor's startup/stop and has independent startup state. Concurrent starts
+  share one attempt; failed attempts can retry.
+- Turn preparation counts as active work for replacement and hot reload. A
+  startup/send failure releases the reservation and records its error in the
+  originating chat. Dispatch refuses an instance retired during asynchronous work.
+- Brain extraction discards cancelled results before applying memory operations.
+  Semantic load completion cannot re-enable a disabled or closed channel.
+- OpenCode's asynchronous orphan check updates only its own latest state slot;
+  it no longer writes an old snapshot over intervening chat/baton changes.
+- Socket send failures are isolated for project, terminal, log and team frames.
+
+## Deliberately retained / next phase
+
+- State.json remains the metadata format; SQLite/JSONL event history is unchanged.
+  A migration to the proposed Brain SQLite schema needs its own migration tests.
+- Native sessions are still scoped as before (configured instance + working
+  directory). Mapping multiple native sessions to one app chat, provider resume
+  compatibility, handoff receipts and coverage are next-phase work.
+- Context selection still uses the legacy windows/extraction policy. These
+  changes establish ownership, not a promise that switching preserves every detail.
+- Helper cancellation discards results; it does not kill the existing helper
+  process. Its configured timeout still applies. Already running native work
+  depends on each adapter's stop implementation; no universal hard-kill policy
+  has been introduced.
+- No Electron migration, UI batching/virtualization change, model download,
+  benchmark, paid agent verification, or removal of existing integrations.
+- The T3 Code checkout is unchanged. No source was copied from it.
+- The boundary work is now committed at
+  `584a3b6a86bfa9af2a0a8ecbeac57e6aa4578e7a`. The next Brain plan is in
+  [brain-continuity-implementation.md](../proposals/brain-continuity-implementation.md);
+  implementation awaits user review. Planning work does not publish changes.

package/docs/refactoring/BRAIN-AUDIT.md

@@ -0,0 +1,176 @@
+# Brain cleanup and plan audit — 2026-09-28
+
+Baseline: `584a3b6a86bfa9af2a0a8ecbeac57e6aa4578e7a` plus the uncommitted
+implementation. Scope: native Brain, event storage, adapters, foreground turns,
+queues, runtime entrypoints, HTTP and browser controls. No commit/push or paid
+provider call was performed. Applied `code-review` (independent standards/spec
+reviews) and `codebase-design` (keep ownership behind the existing interfaces).
+
+## Standards review
+
+Five initial correctness findings were fixed: cancellation during submission,
+queue whitespace/idempotency, incomplete mode-change guards, late orchestra
+rejection, and adapter preparation outside cleanup ownership. Follow-up fixes
+cover inherited-pipe cleanup at parent exit, uncertain process-group errors and
+reassembly overflow. Each has a focused regression or is exercised by the native
+lifecycle fixtures. Heuristic duplication between legacy and native dispatch remains
+an intentional compatibility seam; neither uses the other's briefing/extractor.
+
+Additional cleanup removes four unused runtime forwarding methods, native legacy
+agent-decision mining/text accumulation, duplicated semantic startup on mode changes,
+and legacy extractor/embedding/projection controls while native mode is enabled.
+Legacy SDK/team-memory interfaces remain supported when native continuity is off;
+deleting those before protocol/route/OpenCode parity would break existing users.
+
+## Spec review
+
+Six initial gaps were corrected or bounded explicitly: current-request budget,
+checkpoint revision coverage, new user evidence invalidating unsent snapshots,
+atomic/addressable reconciliation, optional delta holes, and inherited descendant
+containment. Follow-up review caught mixed packet/receipt links, repeat reconciliation,
+failed Git observation treated as stable state, and evidence-count bounds; these were
+fixed. Review did not certify the remaining release gates below.
+
+Other fixes: distinguish agent-generated legacy decisions from user authority;
+prevent tentative/helper supersession; fence uncertain writers against rewind/handoff;
+wait for app-owned diff/commit writes before the next native turn; publish artifacts
+without overwriting existing hashes, read without following final symlinks, verify
+reference size/hash/path, and remove failed temporary artifacts. Native JSONL records
+are bounded before readline can accumulate unlimited tool output. Configuration changes validate before activation and roll back ownership on failed
+atomic config replacement; unresolved native leases also fence legacy journal writes.
+FTS has update/delete
+maintenance plus explicit transactional rebuild. Packet validation uses sets to avoid
+quadratic mandatory-coverage scans; event attribution avoids loading a whole packet
+for each native event.
+
+Native queue action edits and autonomous routes now reject before execution until
+durable revisions and root-task continuity are implemented. Queued target/model/text
+remain frozen; timing changes/reordering are supported. Removing an unsent queue item
+prevents dispatch but retains its original user evidence. This is not data erasure.
+
+## Edge-case traceability
+
+**Covered** means an implementation plus focused local fixtures, not live-provider or
+all-platform certification. **Partial** records a working safeguard plus missing
+requirements. **Gated** means explicitly unsupported/disabled or unimplemented.
+Tests are in `test/continuity.test.ts`, `test/continuity-storage.test.ts`,
+`test/continuity-http.test.ts`, native adapter and existing runtime/browser suites.
+The original research matrix remains authoritative; this table reports implementation.
+
+| Case | Status | Evidence / remaining requirement |
+| --- | --- | --- |
+| E01 | Covered | Buried user source fixture and 10,001-event benchmark; mandatory intent bypasses ranking. |
+| E02 | Covered | Exact tentative discussion, explicit item status; no automatic acceptance. |
+| E03 | Covered | Accepted user supersession, scoped revision checks; helper/tentative supersession rejected. |
+| E04 | Partial | Chat/source scope and originals preserved; explicit task-level scopes not implemented. |
+| E05 | Partial | Conflicting original statements remain; conflict classification/clarification is not automated. |
+| E06 | Partial | Topic/deferred/rejected statuses retained; no native comprehension evaluation. |
+| E07 | Gated | Message deletion/revision dependencies absent; native queue action edits explicitly rejected. |
+| E08 | Covered | New governing user evidence invalidates unsent packets; queued/unsent requests are neither history nor invalidation (validation follow-up fix). |
+| E09 | Covered | Same instance, distinct chat bindings; explicit native IDs. |
+| E10 | Covered | Provider-specific bindings; Claude → Codex → Claude fixture. |
+| E11 | Partial | Model/options/protocol/workspace fingerprint; external auth/entitlement discovery unverified. |
+| E12 | Partial | Missing native acknowledgement stays unknown; manual reconciliation resets epoch. No automatic expiry recovery. |
+| E13 | Covered | Both harnesses report compaction (Claude `compact_boundary`, Codex app-server `contextCompaction`); the binding is marked compacted and the next packet rebuilds state into the resumed session. |
+| E14 | Partial | Retention unknown; protected user evidence repeated. Native reconstruction triggers pending. |
+| E15 | Partial | Ordinary turn-input fixture avoids append-system prompt; live continuation quality unverified. |
+| E16 | Partial | Canonical checkout, HEAD/dirty content and instruction fingerprints; non-Git file applicability and moved-store migration pending. |
+| E17 | Covered | Picker changes only local selection; dispatch freezes target later. |
+| E18 | Covered | Queued captured target/model stays unchanged after picker changes. |
+| E19 | Covered | Foreground reservation, queueing and process-group interruption. |
+| E20 | Partial | Inherited tool group and output-pipe fixture; escaped/detached descendants and Windows containment remain gates. |
+| E21 | Partial | Failed lifecycle stop blocks successors; uncertain termination retains lease. Real approval dialogs unverified. |
+| E22 | Partial | Run/binding/epoch correlation and original chat lookup; broader reordered/late-provider fixtures pending. |
+| E23 | Covered | Request idempotency, one foreground writer across pinned chats/queue; unsent retries re-run instead of silently no-oping. |
+| E24 | Partial | Orchestra/subagents/bridges/routes gated before mutation; full worktree/team parity pending. |
+| E25 | Covered | Immutable prepared packet/receipt; revalidation before any native call. |
+| E26 | Covered | Restart submitting intent becomes unknown with lease retained. |
+| E27 | Covered | No automatic replay; explicit reconciliation. Native message discovery unavailable. |
+| E28 | Covered | Spawn/init separate from acceptance; proven no-launch rejection releases safely. |
+| E29 | Covered | Acceptance and execution separate; quiescence uncertainty retains lease. |
+| E30 | Partial | Stable client request IDs; stale run correlation. No universal native event-ID deduplication. |
+| E31 | Partial | Active/unknown lease retained across restart; genuine orphan OS fault matrix pending. |
+| E32 | Partial | Live owner/host/PID fence and conservative takeover; no heartbeat/remote-filesystem certification. |
+| E33 | Covered | Strict bounded Zod contracts, unknown versions, extra fields, nesting and boolean fixtures. |
+| E34 | Covered | Project/chat/hash and UTF-8 source checks; helpers disabled. |
+| E35 | Partial | Originals retained; removing checkpoint sources restores mandatory evidence; user must review summary fidelity. |
+| E36 | Covered | Exact whitespace/Unicode capture, byte-boundary checks, quoted rendering; heuristic labelled. |
+| E37 | Partial | Codex final shell exit and full large-output artifact; Claude final tool-result normalization pending. |
+| E38 | Partial | Missing completion/failed streams are not success; complete partial-message/tool lifecycle schema pending. |
+| E39 | Gated | Current snapshot checked; per-verification historical workspace applicability not implemented. |
+| E40 | Covered | Immutable packet/receipt links, source/outcome/coverage checks, epoch/state/workspace revalidation. |
+| E41 | Covered | Current request included in heuristic budget; durable overflow/reassembly and explicit resume. Native capacity unknown. |
+| E42 | Covered | Deltas cover every observation since the basis packet plus prior omissions; user sources a session lacks are resent exact or referenced. |
+| E43 | Partial | FTS backfill/update/delete/rebuild and lexical fallback; actual index/DB corruption fault matrix pending. |
+| E44 | Partial | Escaped bounded FTS, scoped reads; aliases/entity expansion and broad paginated search pending. |
+| E45 | Covered | Mandatory text inline, verified local artifact reference, attachments/capabilities fail explicitly. |
+| E46 | Gated | Helpers disabled; no automatic expensive substitution. |
+| E47 | Gated | No helper jobs; helper result invalidation must precede enablement. |
+| E48 | Gated | No helper scheduling; account/global concurrency gate remains. |
+| E49 | Partial | Transaction/foreign-key/busy and reconciliation rollback fixtures; real disk-full/OS permission matrix pending. |
+| E50 | Partial | Busy timeout and bounded indexed work; long-reader/WAL/backpressure corpora pending. |
+| E51 | Partial | SQLite required, unknown protocol/platform explicit; network filesystem detection/certification pending. |
+| E52 | Partial | Flush/exclusive artifact publication before reference; temporary cleanup; orphan GC/grace period pending. |
+| E53 | Partial | Hash/path/size/no-follow checks and corruption fixtures; media/binary decoding unsupported. |
+| E54 | Partial | Consistent activation backup and transactional import; post-cutover downgrade/export pending. |
+| E55 | Covered | Torn records/ID collisions/distinct histories rejected; originals/backups retained; idempotent import. |
+| E56 | Covered | Legacy native cursor never adopted into a chat binding. |
+| E57 | Partial | Chat/project source scope before search/read/artifact export; team/private parity remains gated. |
+| E58 | Partial | Origin-labelled JSON-quoted data; no application replay. Model resistance to adversarial evidence unmeasured. |
+| E59 | Covered | Instruction fingerprint, no user-file overwrite, legacy imported memory not injected in native mode. |
+| E60 | Gated | Complete deletion/derived artifact/backup/native-copy policy absent; hiding is not deletion. |
+| E61 | Partial | No fallback execution; pending capture preserved. Live offline/auth/quota/account matrix pending. |
+| E62 | Covered | Unsupported attachment representations explicitly rejected; no visual-understanding claim. |
+| E63 | Partial | Canonical stable request IDs and bounded packet summaries; request-ID persistence across browser reload pending. |
+| E64 | Partial | Bounded contract diagnostics and project-authenticated routes; sensitive native output redaction/export policy pending. |
+
+## Verification receipt
+
+No live harness/account, Windows Job Object, escaped-child or network-filesystem
+certification is implied by local fixtures. Native mode remains opt-in.
+
+Verified after cleanup:
+
+- `npm run build` and `npm run typecheck`: passed.
+- Final focused core/native group: 87 passed across five suites.
+- `npm test`: 1,408 non-DOM tests and all 125 browser tests passed; 17 existing skips.
+  The full suite ran outside the filesystem/network sandbox because localhost
+  listeners are required. No paid provider calls were made.
+- `git diff --check` and documentation links: passed; all 64 unique matrix IDs have audit rows.
+- Packaging dry run with already-built output: new compiled modules, audit and
+  benchmark artifacts included. No package published.
+- [Audit benchmark](brain-benchmark-audit-2026-09-28.json): same M4/Node24.3 corpus,
+  all 20 buried user points retained; preparation p95 13.85 ms, retrieval p95
+  0.224 ms, event-loop p95 18.12 ms; 3,047 heuristic turn-input tokens including
+  the current request. Native token usage and understanding remain unmeasured.
+
+Standards: eight initial/follow-up correctness findings addressed; the original
+worst issue was Stop racing with launch. Spec: ten initial/follow-up findings
+addressed in the supported path; full descendant/platform containment remains
+partial. Neither verdict is release certification for every matrix row.
+
+## Validation follow-up — 2026-09-28
+
+A later review reproduced four defects (delta coverage window, unsent requests
+rendered as history, capture before harness preflight with a silent retry no-op,
+and superseded checkpoints hiding originals). All are fixed with regressions; see
+[BRAIN-TODO](BRAIN-TODO.md#validation-follow-up--2026-09-28). The version allow-list was
+replaced by a 20 s reachability monitor, and assembly now sends a resumed native
+session only what it lacks. Full suite after the fixes: 1,419 non-DOM and 125
+browser tests passed; 17 existing skips.
+
+## Transport change (app-server / Agent SDK)
+
+Codex moved from `codex exec --json` to `codex app-server` JSON-RPC and Claude Code
+from `claude -p` to the Claude Agent SDK, following t3code. Brain-relevant effects:
+
+- Acceptance evidence is explicit on both: Codex's `turn/start` response or
+  `turn/started`; Claude's `requesting` status or first output.
+- A lost native session is reported before any turn starts (`NativeSessionMissing`);
+  `settled()` fails the receipt and moves the binding to a new epoch, so the next
+  packet reconstructs instead of sending a delta to an empty session.
+- Prompts travel on stdin, which removes the Linux argv limit.
+- A stream failure (an oversized record) now stops RPC parsing before anything else
+  is imported, and the failure's own message is kept.
+- Codex `run_complete` no longer adds cached input to input (cached is a subset).
+