Greptile Summary

This PR adds an early-return to triggerOpenClawRestart() that checks for registered SIGUSR1 listeners and, when found, delegates to scheduleGatewaySigusr1Restart() instead of spawning external processes (launchctl, systemctl, etc.). It also extends the RestartAttempt.method union with "sigusr1". The fix addresses the structural problem of external child processes dying when their parent gateway exits before they complete.

Confidence Score: 5/5

Safe to merge; logic is sound and the only finding is missing test coverage for the new branch.

The SIGUSR1 listener-count guard correctly identifies in-gateway context (Node's --inspect handler is OS-level and invisible to listenerCount, so there's no false-positive risk). The RestartAttempt type union is properly extended. cleanStaleGatewayProcessesSync is correctly skipped on the in-process path. scheduleGatewaySigusr1Restart always returns ok:true so dropping its return value is harmless. The single P2 finding (no unit test for the new branch) does not affect runtime correctness.

src/infra/restart.ts — new SIGUSR1 branch in triggerOpenClawRestart lacks a unit test.

This is a comment left during a code review.
Path: src/infra/restart.ts
Line: 301-305

Comment:
**No test coverage for the new SIGUSR1 early-return branch**

`restart.test.ts` only exercises the stale-PID helpers, and `process-respawn.test.ts` mocks `triggerOpenClawRestart` entirely, so the new branch is never exercised by the test suite. A minimal unit test could register a temporary `SIGUSR1` listener on `process`, call `triggerOpenClawRestart()`, assert that the result has `method === "sigusr1"` and `ok === true`, and verify `scheduleGatewaySigusr1Restart` was called (mocked). This matches the pattern recommended by CLAUDE.md for narrowly scoped changes.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix(gateway): try SIGUSR1 restart before..." | Re-trigger Greptile}

Closing this PR — Codex's P1 is correct: on Windows, the gateway run-loop always registers a SIGUSR1 listener, so the early-return fires when triggerOpenClawRestart() is re-entered from restartGatewayProcessWithFreshPid → schtasks branch, which means relaunchGatewayScheduledTask never runs and the scheduled task isn't told to relaunch the process after exit.

After re-reading the call graph, the root cause is structural, not platform-specific: triggerOpenClawRestart has two caller classes with contradictory needs, and the schtasks branch of restartGatewayProcessWithFreshPid should bypass it and go directly to relaunchGatewayScheduledTask (the same pattern launchd already uses via scheduleDetachedLaunchdRestartHandoff). Opening a revised PR with that fix + the unit test Greptile asked for.

Thanks to @codex and @greptile-bot for the review.

To use Codex here, create a Codex account and connect to github.

Thanks for the thorough write-up — the call graph analysis makes the structural issue clear. The launchd parallel (scheduleDetachedLaunchdRestartHandoff bypassing triggerOpenClawRestart) is a clean model for the schtasks fix. Looking forward to reviewing the revised PR.

Tip: You can customize Greptile's behavior for this repo with .greptile/rules and .greptile/config.