Repository navigation
Dead-PID and zombie job reconciliation #14
Description
Summary
When a tracked job's worker process dies between SessionStart and SessionEnd (crash, OOM, SIGKILL, parent detached), the job record stays status: "running" for the rest of the session. buildStatusSnapshot and enrichJob in plugins/opencode/scripts/lib/job-control.mjs never probe process.kill(pid, 0) — they only read the stored status field.
Cleanup today only happens in plugins/opencode/scripts/session-lifecycle-hook.mjs:27-44, which runs on SessionEnd. During the session itself, /opencode:status and /opencode:result can display rows whose PIDs are long dead.
Repro
- Start a background task via
/opencode:rescue --background. kill -9the background worker PID./opencode:statusstill shows the job asrunning.
Suggested fix
In enrichJob (or buildStatusSnapshot), for any job.status === "running" with a job.pid, probe with process.kill(pid, 0). On ESRCH, downgrade to failed + errorMessage: "Worker process exited without updating job status." and persist via upsertJob.
Upstream reference
Derived from openai/codex-plugin-cc#164.
Activity
[-]Dead-PID detection missing — /opencode:status lies about crashed jobs[/-][+]Dead-PID and zombie job reconciliation[/+]JohnnyVicious commentedon Apr 12, 2026
Consolidating zombie-job follow-ups here: #15 covers stale running rows accumulating in state, and #42 covers the reference implementation for reconciling dead PIDs during status reads. This issue is now the canonical tracker for dead-PID detection plus zombie state cleanup.
JohnnyVicious commentedon Apr 12, 2026
Dead-PID reconciliation implemented in PR #51 commit a4e0097 (isProcessAlive probe + markDeadPidJobFailed + reconcileIfDead wired into buildStatusSnapshot/handleResult/handleCancel). Issue was a parent tracking stub; implementation is done.