pi-crew 0.11.2 → 0.11.4

package/CHANGELOG.md

@@ -1,6 +1,49 @@
 # Changelog
 
-## [Unreleased]
+## [0.11.4] — test-gate hardening + RR-023 battery fixes (2026-09-29)
+
+### fix: grandchild relay empty — inherit the parent task's model on model-less delegate (RR-023, 2026-09-29 battery finding #3)
+
+- **Root cause (run `team_20260929041427_5c40c4fd7e950443`)**: the relay/capture chain (`delegate-spawn.ts` capture → broker fence/mailbox → prompt-runtime poll) was FAITHFUL — all 3 delegate calls omitted `model`, so the broker spawned the grandchild model-less, the child pi booted on the pi GLOBAL default (`minimax/MiniMax-M3` while the parents ran `zai/glm-5.3`), and that model returned empty completions (assistant `content: []` ×3, per the gc session jsonl) — `rawFinalText` undefined + stdout `""` → the relay delivered an empty fenced body as `ok:true`. Same family as finding #1 (75502b68). **Fix**: the `delegate.request` admission now threads the parent task's `model` out of the locked record read and the spawner inherits it when the request omits `model` (explicit `model` still wins; no-model-everywhere passthrough unchanged) — the grandchild can no longer silently fall through to the global default. **Diagnosability guard**: the capture expression is extracted as pure `grandchildResultText()` and an exit-0-with-zero-assistant-text result now renders `[grandchild exited 0 with no assistant output]` instead of an empty mailbox body indistinguishable from a relay bug. Pinned RED-first by `test/unit/runtime/broker/delegate-grandchild-model.test.ts` (fake-spawner harness, 3 ACs) + `test/unit/runtime/delegate-grandchild-result-text.test.ts` (5 unit cases on the helper).
+
+### fix: broker run.lock contention — typed busy frame instead of connection kill (RR-023, 2026-09-29 battery finding #4)
+
+- **`wait.request`/`delegate.request` busy degradation**: both handlers ran their state RMW inside `withRunLockSync`; a live cross-process run.lock holder (e.g. the detached runner persisting task state) made `acquireLockWithRetry` throw immediately and the throw escaped `handleData` → `closeConnection` — direct-async workers saw an untyped `code=close` socket death within ~5ms (evidence: `team_20260929041005` ask-probe / `team_20260929041020` delegate-probe; the sync foreground run was the zero-contention control). New `protocol/lock-busy.ts` classifies the lock-busy throw by its exported message identity, retries on a bounded async-back-off schedule (new `CrewBrokerOptions.lockBusyRetryDelaysMs`, default 50/100/200/400/800ms), then answers a typed `busy` error frame with the connection SURVIVING (ping still answers); the delegate path additionally records `delegate.rejected(reason=run-lock-busy)` in events.jsonl — never silent. `locks.ts` steal semantics untouched (lock family v0.9.26 invariant). Pinned by `test/unit/runtime/broker/crew-broker-lock-contention.test.ts` (3 tests incl. the default-schedule retry-absorbs-transient-hold case).
+
+### fix: test-gate integrity (RR-021 review rounds 2-3)
+
+- **Test runner cancelled-test guard** (`scripts/test-runner.mjs`): a dual reporter keeps stdout UX unchanged (spec under TTY, TAP otherwise) while a full TAP stream is written to a temp file; `# cancelled N > 0` now fails the run loudly in BOTH the single-run and batch paths — independent of node:test's exit-code mapping. Belt-and-suspenders against the round-1 incident class (unref'd-race-timer tests cancelled while the harness reported green).
+- **Test runner missing-file pre-check** (round-3 B1): node:test v22 silently skips an explicitly-listed file that does not exist (when other files still run) and exits 0 — a renamed/deleted test file would silently drop out of `test:critical`-style gates forever. The runner now refuses to spawn and names the missing paths; glob-shaped args are left to node:test's own expansion.
+- **`test:bundle` now routes through the guarded runner** (was a bare `node --test` bypass — the only CI gate without the cancelled/missing-file/fail-closed wrappers).
+- **`check:lazy-imports` hardened**: composite `import()` forms that escaped every earlier rule are now caught (`[(,\[\{]\s*import(`, `>\s*import(`, `=\s*import(`); a zero-match `git grep` (fully-clean `src/`) no longer crashes the gate; the type-annotation skip now requires a member-access suffix `import("...").T` so compound runtime forms are never false-skipped; dist-side regex is whitespace-proof. Remaining gaps (ternary-else, object-property, case/label) are documented in the script.
+
+## [0.11.3] — optimization audit fixes: correctness, hang-risks, bundle, DX, dead code (2026-09-27)
+
+### fix: correctness + hang-risks (RR-021 M1)
+
+- **`classifyReviewOutcome` word-boundary bug**: the `"\baccept\b"`/`"\breject\b"` signals were TS string literals — `\b` is a BACKSPACE char there, so the two strongest verdict signals never matched anything. All 49 signals are now precompiled module-scope regex literals, and a negation guard stops "I cannot accept this" counting as approval. Reject-first precedence unchanged.
+- **Git precondition probes** (`rev-parse`/`status` in the verification worktree sandbox) gained a 5s timeout — a wedged git no longer blocks the extension forever; the existing catch degrades to `available:false`.
+- **`detectRipgrep` hardening**: a spawned-but-wedged `rg --version` is SIGKILLed by an unref'd 3s timer (defaults to unavailable); concurrent first calls share ONE in-flight detection promise instead of spawning one rg per caller.
+- **`markDeadAsyncRunIfNeeded`** reroutes to the async `withRunLock` (same v0.9.26 lock family — live-token interop); the 5s notifier tick no longer holds the event loop on a sync acquisition. The notifier tick keeps no-overlap semantics. `async-runner.ts`'s in-process test seam keeps its one-shot sync lock by design.
+- **`reconcileAllStaleRuns`** is async and reconciles runs with `mapConcurrent` bound 4 (per-run lock files are independent); results keep snapshot order, plan-approval re-check and re-read-inside-lock semantics unchanged.
+
+### perf: bundle + startup hygiene (RR-021 M2)
+
+- esbuild/acorn/diff load lazily through sync `createRequire` shims (model: `syntax-highlight.ts`) — three eager module-scope imports no longer paid on every extension load; `diff` wildcard import narrowed to `diffWords` (+ type-only `Change`).
+- `ajv` marked external (it is only reached via a lazy `await import`) — the bundle shrank 1.65 MB → 1.53 MB and skips parse cost for a rarely-taken path.
+- `check:lazy-imports` catches the common runtime `import()` positions (the old gate matched only `await import(` — audit blind-spot #2) and asserts the committed `dist/index.mjs` contains no module-scope hoisted import of esbuild/acorn/diff/jiti/cli-highlight/yaml/ajv. 23 runtime import sites across 9 files gained `// LAZY:` markers. (Wording corrected post-0.11.3: exotic positions — ternary-else, object-property, case/label — remain documented gaps; see Unreleased.)
+
+### perf(dx): incremental typecheck + flake-resistant waits (RR-021 M3)
+
+- `tsconfig.json` enables `incremental` + a cached `tsBuildInfoFile` (gitignored) — measured with TS 7.0.2: cold 4.10s → warm 1.41s.
+- Three flake-prone suites (`post-exit-stdio-guard`, `preload-idle-render`, `prompt-runtime-steer-boot-window`) replace fixed sleeps with poll-based `waitFor`/`waitForQuiet` (25ms poll, deadlines sized ≥2× the worst-case timer, quiet windows derived from a measured render timeline). Each file verified ×3 consecutive runs with zero flakes.
+
+### perf: scale + hygiene (RR-021 M4)
+
+- The notifier tick and run picker read a bounded `listRecentRuns` window (40→20 after the session filter) instead of scanning the full run index every debounce window; the startup seed keeps its full scan by design.
+- Dead code deleted (−1.6k lines, all verified 0-production-consumer): `src/benchmark/` + its two test files, `ui/card-colors.ts`, `utils/completion-dedupe.ts`, `utils/resolve-shell.ts` + their tests; the affected suites were trimmed to their live checks.
+- Eleven drive-by fixes: Set/Map lookups for hot `includes`/`find` loops (settled batch filter, task-list dependency labels, agents-pane first-wins live-handle map, group-join task resolution), allocation-free reverse scans replacing `[...arr].reverse().find()` copies (×3), a 256-entry cap for `stableComponentCache`, `TypeCompiler.Compile(TeamToolParams)` compiled once (STATIC) instead of per-call `Value.Check` schema walks, a 30-min WALL-CLOCK poll deadline (the old 1800-poll cap silently assumed a 1s configurable interval), a bounded `waitForRecord` (300s deadline racing wedged promises; expiry returns the current record — never undefined), collision-free `msg_<randomUUID>` mailbox ids, and module-scope precompiled regexes for the crew-input router + implicit-tag detector.
+- Spinner isolation (`run-dashboard` `spinnerBucket`/`metricsSig`) deferred to a follow-up — needs a live TUI profile first.
 
 ## [0.11.2] — runtime hygiene, in-process async test seam, slash-command fixes (2026-09-26)