pi-crew 0.11.47 → 0.11.49

package/.githooks/pre-push

@@ -51,6 +51,21 @@ if [ "$pushes_main" = "1" ]; then
 		echo "[pre-push]    Fix the failures above, or bypass with 'git push --no-verify' (CI still gates)."
 		exit 1
 	fi
+
+	# Pre-publish battery (2026-10-28, born from the v0.11.46/0.11.47
+	# silent-worker incident): the LIVE real-Pi gate. One command, tiers
+	# path-triggered from the diff (T1/T3/T11j always; T7 spawns a real
+	# pi -p fast-fix smoke when spawn-affecting paths changed). Evidence lands
+	# in .crew/audit/battery.jsonl. There is NO tier selection here on purpose —
+	# a gate the agent may choose to skip is a gate that will be skipped.
+	# No pi binary (CI-like hosts): PI_CREW_BATTERY_SKIP=1 records the skip in
+	# the audit trail; a silent skip fails.
+	echo "[pre-push] push to main: running the pre-publish battery (live real-Pi gate)…"
+	if ! node scripts/pre-publish-battery.mjs; then
+		echo "[pre-push] ✋ push blocked: pre-publish battery RED (see .crew/audit/battery.jsonl)."
+		echo "[pre-push]    Gate KHÔNG CHẠY = KHÔNG PUBLISH — fix, or bypass with 'git push --no-verify' (CI still gates) and record WHY."
+		exit 1
+	fi
 fi
 
 echo "[pre-push] gates green."

package/CHANGELOG.md

@@ -1,5 +1,97 @@
 # Changelog
 
+## [0.11.49] — 0.11.46/47 incident ROOT-CAUSED & CLOSED: fd0 handshake race → fd3 dedicated pipe; T7 nonce binding; review fix wave (2026-10-28)
+
+**Root cause (dynamically confirmed 5/5 + on-disk leak proof).** pi's `readPipedStdin` (main.js:52-65) consumes ALL
+piped stdin as PROMPT CONTENT in every mode except `--mode rpc`. The v0.11.46 stdin handshake rode fd0 and raced it
+three ways: (1) token JSON merged into the worker prompt — **secret leak, proven in a live transcript** (the model
+reasoned about the credential line); (2) stream-ended creds-less silence; (3) our reader's `pause()` before EOF
+drain deadlocking pi's boot await → **self-initiated exit 0** (not parent SIGTERM as first modeled) with
+`takeOverStdout`-buffered ~0-byte stdout — the exact incident signature (~1.2s, 0 stdout, ×3 retry).
+
+**Fix:** brokered child-pi workers now get a DEDICATED pipe as stdio entry 4 (fd3); the payload is written there
+and read by `readFirstLineFromFd`. fd0 returns to `"ignore"` (pre-0.11.46 shape), killing every race ordering. The
+detached background RUNNER keeps fd0 (plain node process, no race). Gate stays OPT-IN (`PI_CREW_BROKER_STDIN=1`);
+default route bit-for-bit unchanged. **Evidence:** pre-fix gate-ON matrix on pi@1.0.4: 5/5 silent death; post-fix
+2/2 alive (106-110 KB stdout, ~3.2 s, zero leak) — full inversion. Live-worker integration 2/2 on pi 1.4.0 with
+REPO-SRC extension proven via argv capture; security review 9/9 invariants.
+
+**Also:** `maxTasksPerRun` is now enforced at manifest admission (pre-dispatch) instead of only at finalize —
+oversized manifests were fully dispatched and >~2.5k-task handshake lines silently lose broker creds (writer now
+warns >60 KiB). Worker-side fallback gate reads the RESOLVED env (`options.env ?? process.env` via `getCrewEnv`)
+— closing the test seam that masked the earlier dead-wire. `check:env-vars` registry compliance for all 8 broker
+identity reads; fd-mechanics tests skip on win32 (reader short-circuits there by design; brokered spawn is
+auto-disabled on Windows).
+
+### Also in this release (earlier 2026-10-28) — cold-verified review fix wave (2026-10-28)
+
+**Fixed (MAJOR) — the `PI_CREW_BROKER_STDIN=1` opt-in route was dead-wired end-to-end (review + 3-agent cold-verification, dynamically proven).**
+Two compounding defects: (1) the parent never re-wrote the flag into the child's `built.env` (registry kind `host-only` strips
+the ambient value), so the worker-side mirror gate (`resolveBrokerStdinFallbackEnabled`, which reads the child's OWN
+`process.env`) could never activate — the stdin payload was written and never read, leaving every opt-in worker creds-less
+(`NOOP_HANDLE`: no ask/message/mailbox/steer, no broker audit). (2) the token was omitted from `built.env` for ALL
+branches, but only the headless-JSON branch delivers the payload — RPC (`child-pi-rpc.ts`) and surface-pane spawns compose
+without it, so their children got neither token nor payload. Unit tests masked both: they flip the gate by mutating
+`process.env` in-process, bypassing the spawn-env plumbing (structural seam: `startChildBrokerClientAsync` honors
+`options.env` for identity vars but the gate reads ambient `process.env`).
+
+Repair: `prepareSpawnContext` now writes `PI_CREW_BROKER_STDIN="1"` into `built.env` when the host gate is ON (registry
+kind → `child-write`) and puts the token in `built.env` for every brokered spawn; only the headless compose (payload
+present) strips it from the FINAL env (`buildFinalChildPiSpawnOptions`), so RPC/surface keep the env route — a token in
+their env also disarms their stdin fallback (their stdin is the JSONL protocol / a TTY). Gate OFF: bit-for-bit legacy.
+New composed-path tests assert the flag/token/payload matrix across all three call sites (no live-reader assertions — the
+0.11.46 reader exit-0 death is still open, default stays OPT-IN). Misleading docblocks + the factually-wrong registry
+rationale ("parent decision fully determines the route") corrected. `argvIndicatesRpcMode` now also matches the
+`--mode=rpc` equals-form.
+
+**Fixed (MAJOR) — T7 live-smoke judged the WRONG run (newest-mtime race → false PASS).** The battery's T7 tier picked
+`candidates[0]` (newest run dir ≥ start−60s) with zero binding to its own dispatch — a concurrent team run created in the
+window got judged instead (two real run dirs on this host measured 25 ms apart), and the `pi -p` dispatch stdout/stderr
+was piped with zero listeners, so a silent smoke death left no forensics. Now the dispatched goal carries a per-invocation
+nonce (`[t7-nonce:<uuid>]`); only run dirs whose `events.jsonl` contain that nonce are judged, zero nonce dirs = FAIL, a
+non-terminal run is a FAIL checked before the killed/hung diagnosis, and a capped tail (64 KB) of the dispatch's own
+stdout/stderr is recorded in the failure reason, audit entry, and log output.
+
+**Fixed (minor) — battery gate fail-open trigger paths.** `changedPaths()` now filters diff-base tags to release semver
+(`^v\d+\.\d+\.\d+$` — a near-HEAD non-release tag previously shrank the diff window and silently un-triggered T7) and
+parses `git status --porcelain` renames (`R old -> new`) into BOTH paths instead of one mangled string.
+`prepublish-ci.mjs` now preserves `PI_CREW_BATTERY_SKIP` through the npm `prepublishOnly` path (the documented skip
+hatch was dead there — the battery's own failure message recommended a var its caller stripped).
+
+**Changed — T7 trigger coverage gaps closed** (`src/state|utils|agents|teams|workflows|hooks|observability`, `scripts/`):
+cold-verification found the workflow-engine src subtrees and the gate's own scripts/ were invisible to the live gate; gate
+changes now live-test the gate (this commit deliberately triggers T7 on its own pre-push).
+
+**Changed — `before_agent_start` `systemPrompt` hardening (defense-in-depth).** Cold-verification refuted reachability
+(`systemPrompt` is required in every host type def and synthesized by a runner getter, unlike `prompt` which arrives via
+caller args — the GH #64 root cause), but the asymmetry was real: `autonomous-policy.ts` and `knowledge-injection.ts`
+now coerce non-strings to `""` (append-only output, safe); `prompt-runtime.ts` SKIPS the rewrite instead (coercing `""`
+there would blank a real system prompt — it strips sections). No behavior change on any current host.
+
+**Added — LOUD oversize diagnostic in `buildBrokerStdinLine`.** The reader aborts past 64 KiB without a newline; a
+~2.5k-task manifest (~131 KiB measured live — pipe chunks make the effective Linux boundary ~2× the coded cap)
+silently stripped EVERY task's broker creds. The writer now warns at 60 KiB via `logInternalError` (not PI_CREW_DEBUG-
+gated); the line is still written, never blocked.
+
+## [0.11.48] — HOTFIX: stdin broker-credential gate flipped OPT-IN — fixes silent worker death (2026-10-28)
+
+**Fixed — v0.11.46/0.11.47 regression: every headless JSON worker on the default route exited 0 with zero stdout.**
+The LOW-1 stdin handshake shipped in 0.11.46 with the gate default-ON. On real Pi hosts, the moment the worker-side
+stdin reader activated (`startChildBrokerClientAsync` → `readStdinFirstLine`: env broker identity present, env token
+absent, gate on — the default shape of every team-run worker), the whole `pi --mode json -p` process exited 0 silently
+(~1.2s, 0 stdout bytes, retry ×3, run failed). CI never caught it: unit tests inject the `readStdinLine` seam and never
+run `process.stdin` inside a real pi process; the live battery (real-test-pi-crew T7) caught it — which is also why this
+release ships with that battery re-run green (3/3 tasks + `message` round-trip through the legacy env route).
+
+`PI_CREW_BROKER_STDIN` is now **OPT-IN** (`"1"` enables the stdin route; unset/`0`/anything else keeps the legacy env
+route on BOTH sides, bit-for-bit pre-0.11.46 behavior). The LOW-1 closure code (writer, reader, payload v2) stays for
+re-enabling after the reader path is root-caused.
+
+Also adds `test/integration/broker-stdin-reader-liveworker.test.ts` — a LIVE regression test that spawns a real `pi`
+worker with the exact env shape that died and asserts non-zero stdout (skips where no `pi` binary exists, e.g. CI
+containers that run the fake-pi fixtures), plus incidental biome fixes (`useImportType` ×5, `noUselessTernary`) in
+crew-broker-delegate-request.ts / surface-worker.ts.
+
 ## [0.11.47] — live-session split: model/prompt clusters out of live-session-runtime.ts (2026-10-27)
 
 **Changed — `live-session-runtime.ts` (1971 → 1634 lines).** The file sat 29 lines under the 2000-line wc-gate LIMIT,