Three independent reliability fixes that all surfaced as the same bug:
live tasks losing their worktrees mid-execution and emitting
`wrong_toplevel` errors.
Fix A — executor stale-self-owned classifier:
`reconcileSelfOwnedActiveSessionForRemoval` now takes a process-active
probe (`executingTaskLock.has`) and a minimum-idle window (default 5s)
in addition to the existing in-memory `activeWorktrees` binding probe.
Recently-registered or still-running entries are refused with
`process-active-refuses` / `too-recent-refuses`, with audit-grade
log lines. Both the pre-remove path
(`reconcileSelfOwnedBeforeRemove`), the post-throw retry in
`removeOwnWorktreeWithReconcile`, and the defensive reconcile in
`removeWorktree` route through the same hardened gates.
Fix B — pause synchronously reaps the agent session:
New `awaitAbortInFlightTaskWork` mirrors the existing fire-and-forget
abort but awaits each `session.abort()` /
`stepExecutor.terminateAllSessions()` /
`workflowSession.abort()`. `parkTaskAfterWorkflowStepPause` calls it
before `moveTask("todo")`, and the `task:updated` user-pause handler
routes through it, so a fast re-dispatch can no longer race a still-
live shell.
Fix C — self-healing realpath + active-task skip:
`reconcileTaskWorktreeMetadata` now realpath-normalizes both sides of
the registry comparison (handling macOS `/private/var/...`) and
refuses to clear `worktree`/`branch` on in-progress or in-review
tasks. The skip emits a new
`task:auto-recover-worktree-metadata-skipped-active` audit event;
executor-level recovery paths remain in charge of active tasks.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The merge adds a post-completion defensive backstop that probes and removes stale same-task `activeSessionRegistry` entries on `done`/`archived` transitions, completing FN-5346 with a shared reconcile helper, a defensive ownership probe wired into paused cleanup, audit event alignment, and regressio
Fusion-Task-Id: FN-5346
Investigating FN-4814 + FN-4811 re-failures after commit 8bef30655 (which
added per-instance synchronous this.executing.add) revealed the per-instance
guard was insufficient. FN-4809 log at 02:48:17-18 UTC:
02:48:17 [-] Resuming execution after unpause
02:48:17 [-] Step 4 (Testing & Verification) -> pending
02:48:17 [-] Step 4 (Testing & Verification) -> pending
02:48:17 [6097725-y2nb] Executor detected stale merge state ...
02:48:18 [6097816-9gde] Executor detected stale merge state ...
Both runs y2nb and 9gde reached executor.ts:2661 (which is INSIDE execute(),
past the synchronous this.executing.add claim). The only viable explanation
is that there is more than one TaskExecutor instance in the same Node
process (engine restart race, multi-project hybrid runtime, or similar code
path). Each instance has its own executing Set, so the per-instance guard
doesn't help.
Fix: module-level singleton executingTaskLock in active-session-registry.ts,
shared across all TaskExecutor instances. execute() synchronously tryClaim()s
the lock; if false, bails. Every existing this.executing.delete() site also
calls executingTaskLock.release(). Per-instance this.executing kept because
many other call sites use it (this.executing.has at handler gates,
stuck-detector, resumeTaskForAgent, etc.).
Test setup (resetExecutorMocks in executor-test-helpers.ts) clears the lock
between tests so process-wide state doesn't leak (executor-pause and
executor-prompt tests would otherwise show 'expected 2 createFnAgent calls
but got 0' / 'expected not called but called 3 times' flakes).
Tests:
- executing-task-lock.test.ts: 2 cases. Key case creates TWO TaskExecutor
instances and races them on the same task ID, asserts only ONE actually
runs. Verified FAILS on prior code (8bef30655) and PASSES on fix.
Verification:
- Targeted suite (4 files, 170 tests): pass.
- pnpm --filter @fusion/engine build: clean.
- pnpm lint: clean.
Fusion-Task-Id: FN-4811
