pi-crew 0.11.11 → 0.11.13

package/CHANGELOG.md

@@ -1,5 +1,107 @@
 # Changelog
 
+## [0.11.13] — Reliability: runner survives run.lock contention + 2 fixes from live wave runs (2026-10-10)
+
+Three reliability fixes discovered while shipping Waves 1-3 of the upgrade-spec
+program (each bug reproduced live against a real team run today).
+
+### Fixed
+
+- **Runner death on run.lock contention (critical)** — the background runner
+  died with an unhandled rejection when `mergeUnitResult` hit
+  `Run 'run.lock' is locked by another operation`, which stopped heartbeats
+  and let the stale-reconciler mass-cancel workers as pid_dead (killed the
+  Wave-3 run mid-flight). Merge sites now treat lock-contention as retryable
+  (backoff + event log); fatal errors still reach the deadletter path. New
+  2-writer contention tests (merge-loop-lock-contention.test.ts).
+- **Chain/foreground watch-timeout no longer cancels child runs** — a 60-min
+  watch window expiring now reports partial-watch only; the child run keeps
+  running (previously: chain steps got their run cancelled).
+- **`resume` reports real status** — resuming a run replied "Status:
+  completed" instantly while the work continued for hours; it now reports
+  the actual state and re-queued work.
+
+# Changelog
+
+## [0.11.12] — Upgrade Wave 2: U4 live-session v2, U5 mailbox dedup, U6A classifier triage, U14 rpc mode, U15 OwnedProcess (2026-10-10)
+
+Wave 2 of the upgrade-spec program. One commit per spec; test:critical 120/0,
+full npm test 0 fail (unit 8842/8845 + integration 139/143, gated skips only),
+bundle rebuilt fresh; wave-2 review 0 MAJOR with MINOR-1..4 pool/env hardening
+landed as follow-up. Waves 2+3 released together as 0.11.12.
+
+### Changed
+
+- **U4** — live-session v2: shared services bundle (one lazy build, sessions
+  from `createAgentSessionFromServices` — measured init drop ~2240ms → <50ms
+  via `session_created.elapsedMs`), two-layer recursion guard (loader-level
+  noExtensions + foreign-extension assert), `streamingBehavior` on every
+  prompt path, normalized dispose (drain pendingSteers, terminated handles
+  reject steer), D4 planning-window steer-loss fix (placeholder session
+  queues and replays to the real session), hung-tool drop-reference escape.
+- **U5** — mailbox `requestId` dedup: every send carries a unique requestId,
+  receivers dedup by (direction, requestId) durably, resend-after-resume is
+  idempotent (7/7 tests).
+- **U14** — child-pi protocol v2: `--mode rpc` transport (steer receipts,
+  isStreaming-gated steer + clear_queue before prompt, type-filtered event
+  subscription, absolute bundle cliPath), 18/18 tests; pool reuse opt-in
+  (`PI_CREW_CHILD_PI_POOL`, default off). MINOR-1..4: busy-alive pool entries
+  never evicted, drain ceiling gated on pool flag, run-scoped named listeners
+  detached in settle(), env-splice canary.
+- **U15** — OwnedProcess port: process-group ownership (detached spawn →
+  group-kill), escalating dispose with hard cap, root-exit drain-reconcile,
+  pid-recycle guard, postmortem registry; 11/11 tests.
+
+### Added
+
+- **U6A** — classifier retry-triage flag (`classifierEnabled`, default off):
+  confidence passthrough + histogram + `task.noise_gate` events, soft-fail on
+  every branch (12 tests).
+
+
+## [0.11.12] — Upgrade Wave 3 (con't): U7 events view layer, U8 sqlite state, U9 reconcile-at-open, U6B classifier pre-gate, U13 verify_gate (2026-10-10)
+
+Wave 3 of the upgrade-spec program (`pi-crew-upgrade-spec-2026-10-09.md`), branch
+`upgrade/wave-3`, one commit per spec. Final gate: unit 8934 tests (8925 pass;
+4 shard/snapshot registrations fixed and re-verified green; 2 load-flakes green in
+isolation), integration 144/0 fail, typecheck + lint clean, `dist/index.mjs` rebuilt
+fresh with every wave-3 string-literal marker verified in-bundle.
+
+### Added
+
+- **U7** — events state source view layer: `eventsStateSource(runId)`
+  CommittedStateSource adapter (snapshot = one full parse to the current seq; frames =
+  `readEventsCursor` tail-follow batches; >100-frame backlog or inode change → ONE
+  root-replacement resync). `run-snapshot-cache` opt-in `{ eventsState: { pollMs } }`
+  shares ONE unref'd poll interval per entry; browser tick CPU 6.98ms → 0.41ms.
+  `events.jsonl` is never written — pure derived view. (665b443a)
+- **U8** — opt-in sqlite run-state backend: `PI_CREW_STATE_BACKEND=sqlite` stores
+  manifest + tasks + worker-status in `<runId>/state.sqlite` via `node:sqlite` with
+  ONE transaction per batch write (all-or-nothing); JSON files stay written as
+  read-side mirrors for legacy consumers while loads read the db. The `events.jsonl`
+  audit spine is unchanged in both modes. (ae49628f)
+- **U9** — reconcile-at-open fast-path: crashed runs with NO live claim (dead async
+  pid + no live registry entry) requeue `running`→`queued` in ONE locked pass
+  (attempt / checkpoint / deps / partial state preserved) and the run is marked
+  `failed`-for-resume; the verdict tree now only runs for live-claim runs. Interrupted-
+  run detection latency 300000ms → 1-9ms. New reconcile verdict `requeued_no_claim`.
+  (456c9283)
+- **U6B** — classifier verifier pre-gate: score-question variant consults the ambient
+  classifier before a verifier-role spawn (block score ≥ 0.9 / confidence ≥ 0.8);
+  soft-fail — every classifier error path escalates to the LLM verifier. (6b6c0aca)
+- **U13** — deterministic `verify_gate` pre-gate + tool: verifier-role spawns first run
+  the package.json-script checks (typecheck / test:critical) via
+  `ctx.executeTool("bash", ...)` with per-check timeout (`PI_CREW_VERIFY_GATE`,
+  `PI_CREW_VERIFY_GATE_TIMEOUT_MS`, default 120s); PASS skips the LLM verifier spawn,
+  FAILED/INCONCLUSIVE spawns it with the gate results in the prompt. (ef67cd04)
+
+### Fixed (wave-3 gate remediation)
+
+- env-vars registry snapshot now includes the wave-3 keys (`PI_CREW_STATE_BACKEND`,
+  `PI_CREW_VERIFY_GATE`, `PI_CREW_VERIFY_GATE_TIMEOUT_MS`).
+- shard-partition table registers the new `test/unit/registration/` dir (U13 test).
+- import sort in `sqlite-run-state.test.ts` (biome organizeImports).
+
 ## [0.11.11] — Upgrade Wave 1: U1 dead-option, U2 buffered events, U3 key-free E2E, U10 measured-token handoff (2026-10-10)
 
 Wave 1 of the upgrade-spec program (`pi-crew-upgrade-spec-2026-10-09.md`, verified over