release: v0.29.0 — TP-187 supervisor recovery + TP-189 polish + #559/#560 hotfixes

[HenryLach]

Minor release bundling supervisor recovery flows + accumulated polish + two critical post-merge hotfixes.

Why minor (not patch)

• New supervisor_takeover(reason) tool surface (TP-187, Post-failure event queue replays zombie 'wants to exit' alerts to supervisor; no clean drain mechanism #538) — additive but observable
• New batch-meta.json runtime artifact (TP-187, orch_resume(force=true) cannot reattach to a stopped batch after orch_abort() #539) — additive on-disk surface, written at every batch start
• CLI behavior change: taskplane doctor now actually shows the pi version (was empty parens — TP-189 / chore(release): v0.10.2 #185 follow-up)
• Refined worker / supervisor prompt templates — behavior change for new tasks
• New lane-terminated / lane-respawned IPC messages between engine-worker and supervisor

Per semver, these warrant a minor bump (0.28.x → 0.29.0). Pure bug-fix entries also shipping here as part of the same release.

Highlights

TP-187 — Supervisor recovery flows (PR #556 main batch, all reviews APPROVE)

• Post-failure event queue replays zombie 'wants to exit' alerts to supervisor; no clean drain mechanism #538 zombie alert drain: per-agent on-disk outbox drain at lane termination + supervisor-side terminated-lane filter that drops in-transit zombie alerts before they reach pi.sendUserMessage. New supervisor_takeover(reason) tool for non-destructive batch parking.
• orch_resume(force=true) cannot reattach to a stopped batch after orch_abort() #539 force-resume reconstruction: batch-meta.json artifact written at batch-start lets orch_resume(force=true) rebuild a validator-compliant PersistedBatchState from runtime artifacts after orch_abort deletes .pi/batch-state.json.
• 'Worker said:' empty in early no-progress exit alerts — supervisor lacks signal for early intervention #540 worker exit-reason surfacing: Worker said: "..." field in supervisor alerts now falls back to walking the worker's events.jsonl when the current turn produced no visible output.

TP-189 — Polish bundle (PR #556 main batch, all reviews APPROVE)

• A defensive tests for spawn-site wiring, review_step REFUSED runtime, fenced-code-block filter, removeWorktree() Windows fallback decision branches.
• B constants module migration (tool-allowlist-constants.ts single source of truth).
• C taskplane doctor shows pi version (extracted to testable bin/get-version.mjs).
• D ci.yml on Node 24 (already shipped in v0.28.8).
• E worker prompt + skill reconciliation with TP-186's Order of Operations rule.

Sage post-integration folds (PR #556)

• batchId IPC gating (sage chore(deps-dev): bump vitest from 4.0.18 to 4.1.0 in /extensions #3): lane-terminated / lane-respawned handlers gate on the supervisor's tracked orchBatchState.batchId so stale messages from a prior batch can't taint the live batch's suppression map.
• Multi-repo reconstruction guard (sage's blocker reframed): refuse force-resume reconstruction when worker manifests indicate >1 distinct repoId. Prevents silent segments: [] state loss for polyrepo batches.

Critical post-merge hotfixes

• ReferenceError: batchState is not defined — orchestrator crashes on first child IPC (regression in 0.28.x) #559 orchestrator parent crashes on first IPC frame (ReferenceError: batchState is not defined). My TP-187 sage-fold commit referenced an undefined identifier in the supervisor IPC closure. Affected every batch. Fixed in commits 4a8aea6 + ff02265 with sage post-mortem switching to orchBatchState.batchId (the canonical live runtime batch ID, not supervisorState.batchId which is only populated when the supervisor activates).
• Hardcoded @mariozechner/pi-coding-agent path breaks Runtime V2 after rename to @earendil-works #560 Pi CLI path resolution broken after @mariozechner → @earendil-works rename in Pi v0.74.0. resolvePiCliPath() now searches both scopes (new first, legacy fallback). Backward compat preserved for users still on the legacy scope.

Validation

• Tests: 3587 passing / 1 skipped / 0 failed on Node 24.15.0 (Windows local). +91 new tests over the v0.28.8 baseline (3496).
• CI: green on Linux Node 24
• CLI smoke: taskplane help, taskplane doctor, taskplane init --dry-run, taskplane uninstall --dry-run all clean. taskplane doctor correctly displays ✅ pi installed (0.74.0) (the rename target).
• Polyrepo end-to-end test: ✅ verified by operator on a real 6-wave multi-repo batch (3 repos, 6 tasks). All waves completed; ReferenceError: batchState is not defined — orchestrator crashes on first child IPC (regression in 0.28.x) #559 and Hardcoded @mariozechner/pi-coding-agent path breaks Runtime V2 after rename to @earendil-works #560 fixes confirmed working in the wild.
• Package contents: npm pack --dry-run clean — 660 kB, 77 files
• No publish-blocking external dependencies changed — peerDependencies still on @mariozechner/* (intentional; Pi's runtime aliasing handles both scopes, and updating peer deps risks breaking compat for users on Pi <v0.74.0).

Issues closed

• Post-failure event queue replays zombie 'wants to exit' alerts to supervisor; no clean drain mechanism #538, orch_resume(force=true) cannot reattach to a stopped batch after orch_abort() #539, 'Worker said:' empty in early no-progress exit alerts — supervisor lacks signal for early intervention #540 (TP-187)
• ReferenceError: batchState is not defined — orchestrator crashes on first child IPC (regression in 0.28.x) #559 (post-merge hotfix)
• Hardcoded @mariozechner/pi-coding-agent path breaks Runtime V2 after rename to @earendil-works #560 (post-merge hotfix)

Issues queued (next release)

• TP-190 / Runtime V2 lane spawn errors are silent — lanes stay 'running' forever, no alert, no failure transition #561 — Runtime V2 spawn errors are silent; lanes stay "running" forever. Task packet staged at taskplane-tasks/TP-190-runtime-v2-spawn-failure-visibility/. Review Level 2, Size M. Will ship in v0.29.x or v0.30.0 after the orchestrator runs the task.
• Dashboard: merge agent telemetry missing for some waves (snapshot not captured) #509 update — operator commented with refined hypothesis (off-by-one wave attribution in dashboard merge telemetry). Awaiting triage.

Tag

v0.29.0 will be pushed AFTER this PR merges. Use --merge to preserve the version-bump commit in main's history.