codex-watchdog

Codex Watchdog

Use scripts/codex_watchdog.py as the only state-changing interface. Resolve it from this skill's directory and invoke it by absolute path; do not rely on the current working directory.

Honor control requests

Treat these requests as executable controls, including in a brand-new conversation:

• On “关闭watchdog” or equivalent, run python <WD_SCRIPT> disable, then report the returned state.
• On “启用watchdog” or equivalent, run python <WD_SCRIPT> enable, then report the returned state.
• On a watchdog status request, run python <WD_SCRIPT> status, then report the result.
• On “卸载/彻底移除 watchdog”, run disable and then uninstall. Preserve the skill and incident files unless the user separately asks to delete them.

Do not merely acknowledge these requests. disable persists across conversations; while disabled, do not arm jobs unless the user enables monitoring again.

Monitor long work

For each tool or delegated worker expected to exceed 30 seconds:

1. Check status once before an immediately launched batch. If enabled, classify each attempt and have the main conversation choose a rolling no-progress threshold from references/timeout-policy.md. Run arm --kind KIND --turn auto --generation 1 --timeout-seconds CHOSEN-SECONDS --label SHORT-LABEL for every initial attempt. The script uses CODEX_THREAD_ID when available and otherwise keeps an unknown-thread diagnostic label; tag UUIDs still isolate jobs. Retain the exact unique tag returned by the script.
2. Keep parallel work isolated: assign one tag to one attempt, and pass that exact tag to every later command.
3. Observe the real worker or output at least every 30 seconds. The main conversation or a dedicated monitoring subagent may record a heartbeat after judging the attempt healthy from current evidence such as advancing scan counters, new stream/log/tool-call records, changing output files, active process work, worker phase changes, or other task-specific progress. A timer tick or an unchanged “thinking” label alone is not proof of progress.
4. Treat expiry of the chosen interval as a mandatory review point, not an automatic stop. Inspect the exact attempt without preempting the main task. If current evidence shows normal progress, run heartbeat TAG --note EVIDENCE and continue waiting. Absence-only evidence—including no completed command, unchanged files, no child process, post_tool_transition_unobserved, model_preparing_no_request, or a quiet model stream—never authorizes interruption. Stop only on explicit user instruction or positive terminal/failure evidence. A task may run much longer than timeout_seconds while it continues producing verified progress. Keep the separate 180-second hard limit for a single image-generation attempt when the active repository or user instructions require it.
5. Always run disarm TAG --reason REASON immediately on completion, failure, cancellation, or a confirmed abnormal stall. Use a new generation and a new tag for any retry.