SIGKILLed foreground companion orphans a live app-server turn; /cancel can't reach it

[axisrow]

Problem

A foreground companion run that is SIGKILLed by Claude Code's Bash tool (~120s ceiling) leaves the underlying Codex turn running inside the detached broker, still able to modify the workspace. The job is then reconciled to failed, and /codex:cancel refuses to touch it — so the orphaned turn cannot be stopped through the plugin at all.

This is not introduced by #41. It is reachable on main today. Verified against 15f22e6: the broker's socket-close handler, the /cancel status filter, and the missing turnTimeoutMs on the adversarial-review path are all identical there. #41's idle semantics widen the window for actively-progressing turns, but every link in the chain predates it.

Failure chain

1. Foreground runs the turn in-process. handleTask (plugins/codex/scripts/codex-companion.mjs:905-922) and handleReviewCommand (:828-843) both go through runForegroundCommand (:744-755). Only task --background spawns a detached worker (spawnDetachedTaskWorker, :757-768). Note handleReviewCommand parses --background (:801) but never consults options.background — for reviews the flag only raises the idle budget via resolveTurnTimeoutMsFromOptions (:726), it does not background anything.

2. An active turn never trips the internal deadline. armIdleDeadline (plugins/codex/scripts/lib/codex.mjs:713-726) resets on every lifecycle and in-item progress notification, so a healthy long-running foreground turn sails past the host's Bash ceiling.

3. SIGKILL leaves no trace. There are no signal handlers anywhere in plugins/codex/scripts/, and runTrackedJob (plugins/codex/scripts/lib/tracked-jobs.mjs:194-226) has no finally, so the job record stays status: "running".

4. The broker never interrupts. socket.on("close") / socket.on("error") (plugins/codex/scripts/lib/broker-controller.mjs:353-370) call only sockets.delete + clearSocketOwnership (:149-157) + scheduleIdleShutdown. No turn/interrupt is issued; the app-server child keeps executing the turn.

5. Notifications are dropped and the broker reports itself idle. routeNotification (:159-175) returns early once ownership is cleared, and scheduleIdleShutdown arms the 15-minute timer — so probeBroker sees an idle broker while a turn is live, and a reuse path may even shut it down mid-turn.

6. Reconciliation marks the job failed. reconcileRunningJobs (plugins/codex/scripts/lib/state.mjs:151-196) lazily flips running → failed with UNREPORTED_PROCESS_EXIT_MESSAGE on the next listJobs.

7. /cancel can no longer reach it. resolveCancelableJob (plugins/codex/scripts/lib/job-control.mjs:326-353) calls listJobs (which reconciles first) and then filters status === "queued" || "running". The job is already failed, so handleCancel throws before ever reaching interruptAppServerTurn (plugins/codex/scripts/lib/codex.mjs:1166).

Why the fix is cheap

Everything needed for recovery is already in place:

• createJobProgressUpdater (plugins/codex/scripts/lib/tracked-jobs.mjs:89-97) persists threadId and turnId incrementally, as soon as each is observed — so a killed job's record already carries valid ids.
• interruptAppServerTurn passes allowBusyStaleBroker: true (plugins/codex/scripts/lib/codex.mjs:1186-1188), so it can attach to a busy broker.
• The broker already has a dedicated bypass for interrupting a stream owned by another socket (plugins/codex/scripts/lib/broker-controller.mjs:298-323), and it tracks activeStreamThreadIds (:99, :334).

The machinery exists; only the reachability gate is missing.

Proposed fixes (independent, both worth doing)

• Broker-side (defense in depth). In socket.on("close") / on("error"), if the closing socket was activeStreamSocket, issue turn/interrupt for the active stream before clearing ownership. This needs the turn id as well as the thread id — buildStreamThreadIds currently retains only thread ids, so it would need to capture turnId too.
• Cancel-side (user-facing). Let /cancel reach a job that reconcileRunningJobs just marked failed with UNREPORTED_PROCESS_EXIT_MESSAGE and a non-null threadId/turnId, so interruptAppServerTurn becomes reachable. Either widen the filter at job-control.mjs:328 or introduce a distinct terminal state (e.g. orphaned) that cancel accepts.
• Companion-side (partial). Install SIGTERM/SIGINT handlers that interrupt the in-flight turn. Does not help against SIGKILL, but covers ordinary termination.

Related

• adversarial-review ignores --turn-timeout-ms — turnTimeoutMs not forwarded to runAppServerTurn #38 / fix: thread turn-timeout-ms into adversarial-review's app-server turn #39 — adversarial-review drops turnTimeoutMs (codex-companion.mjs:465), so it runs with the 600000ms default even in the foreground. That is today's single worst exposure to this bug, but it is tracked separately and should not be folded in here.
• Turn timeout should measure agent-idle time, not total turn wall-clock #40 / fix: measure turn timeout as agent idle time, not total wall-clock #41 — the idle-timeout redesign. It makes actively-progressing foreground turns far more likely to outlive the Bash ceiling, which is what surfaced this issue, but it does not create any link in the chain above.

Verification

• /codex:status must not report an idle broker while a turn is live.
• After a kill that orphans a turn, /codex:cancel <id> must successfully interrupt it.