fix: reduce post-completion idle watchdog and add cleanup timeouts to prevent Copilot CLI hang on exit

[Copilot]

The Impeccable Skills Reviewer was completing its review successfully but the Copilot CLI process never exited, causing every run to be killed by the 15-minute hard step timeout — a false failure on otherwise successful reviews.

Root cause

Two compounding issues in copilot_sdk_session.cjs:

1. Watchdog too slow: SDK_POST_COMPLETION_IDLE_MS_DEFAULT was 5 minutes. If the agent finishes at T+10min with a 15-min step, the watchdog fires at T+15min — racing the timeout instead of beating it.

2. client.stop() can hang: After sendAndWait times out, the finally block calls await client.stop(). When the SDK server is unresponsive this never settles, blocking the process until the step timeout kills it.

Changes

actions/setup/js/copilot_sdk_session.cjs

• Reduce SDK_POST_COMPLETION_IDLE_MS_DEFAULT from 5 * 60 * 1000 → 30 * 1000. The watchdog guards against false positives via three conditions (output collected, no in-flight tool calls, not mid-inference turn), so 30 s is safe in practice.

• Add 5-second cleanup timeout to session.disconnect() and client.stop() in the finally block. Uses Promise.race so a hung server unblocks the driver within a bounded window, with a warning logged on timeout:

constwithCleanupTimeout=p=>{lettimedOut=false;constdeadline=newPromise(resolve=>setTimeout(()=>{timedOut=true;resolve(false);},CLEANUP_TIMEOUT_MS));returnPromise.race([p.then(()=>true,()=>true),deadline]).then(settled=>{if(!settled&&timedOut)log(`warning: cleanup operation timed out after ${CLEANUP_TIMEOUT_MS}ms`);returnsettled;});};

actions/setup/js/copilot_sdk_driver.test.cjs

• Adds regression test "cleanup timeout prevents hang when client.stop() never resolves" — client.stop() returns a promise that never settles; asserts exitCode: 0 and wall-clock completion under 10 s.

---

Generated by 👨‍🍳 PR Sous Chef · 14.3 AIC · ⌖ 7.84 AIC · ⊞ 7.1K · ◷
_{Comment /souschef to run again}

---

Generated by 👨‍🍳 PR Sous Chef · 5.89 AIC · ⌖ 8.02 AIC · ⊞ 7.1K · ◷
_{Comment /souschef to run again}

[github-actions]

Hey @copilot 👋 — thanks for picking up the Impeccable Skills Reviewer timeout fix! The problem statement is well scoped (post-completion hang preventing clean CLI exit) and the PR is correctly rooted in the failure-investigator issue trail.

A couple of things to address before this is ready for review:

• No diff yet — the PR currently has 0 changed files. Once the fix lands (engine exit signal, MCP/stdio handle cleanup, or idle-timeout addition), make sure the implementation is committed and the description is updated to reflect what changed and why.
• Tests are missing — the fix will touch process-lifecycle or MCP-server teardown paths. Add tests (or extend existing ones) that confirm the CLI exits cleanly after the final assistant turn, without requiring the full 15-minute hard timeout to fire.

When you're ready to fill in the implementation, here's a prompt you can use:

Fix the post-completion hang in the Copilot CLI that causes the process to stay alive after the agent emits its final assistant turn.
Steps:
1. Identify the MCP server(s) and safe-output stdio bridge(s) that remain open after the last turn — trace from the engine's shutdown path.
2. Ensure all child processes (MCP servers, safe-output bridge) receive a close/terminate signal when the engine decides it is done.
3. Add an inactivity/idle timeout (shorter than the 15-minute hard step timeout) so a finished-but-hung agent exits with success when output was produced.
4. Write a test (unit or integration) that starts a minimal agent run, asserts it completes, and verifies the process exits within a short deadline after the final turn — no manual kill required.
5. Update the PR description with: what was hanging, what was changed to fix it, and how to verify the fix locally.

Generated by ✅ Contribution Check · 160.8 AIC · ⌖ 33.9 AIC · ⊞ 6.2K · ◷

[Copilot]

Pull request overview

This PR prevents successful Copilot SDK-driven runs from being reported as failures due to the Copilot CLI hanging on exit, by tightening the post-completion watchdog and bounding cleanup operations that can otherwise block forever.

Changes:

• Reduced the default post-completion idle watchdog from 5 minutes to 30 seconds to beat typical GitHub Actions step timeouts.
• Added a per-cleanup 5s timeout wrapper around session.disconnect() and client.stop() to avoid indefinite hangs in the finally block.
• Added a regression test covering the “client.stop() never resolves” hang scenario.

Show a summary per file

File	Description
actions/setup/js/copilot_sdk_session.cjs	Shortens the post-completion watchdog and adds timeout-bounded best-effort cleanup in the finally block.
actions/setup/js/copilot_sdk_driver.test.cjs	Adds a regression test ensuring cleanup timeout prevents hangs when client.stop() never resolves.

Review details

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Low

[gh-aw-bot]

@copilot please run the pr-finisher skill, address these unresolved review threads, and rerun checks once the branch is up to date:

• Copilot: fix: reduce post-completion idle watchdog and add cleanup timeouts to prevent Copilot CLI hang on exit #44254 (comment)

Generated by 👨‍🍳 PR Sous Chef · 14.3 AIC · ⌖ 7.84 AIC · ⊞ 7.1K · ◷
_{Comment /souschef to run again}

[Copilot]

@copilot please run the pr-finisher skill, address these unresolved review threads, and rerun checks once the branch is up to date:

• Copilot: fix: reduce post-completion idle watchdog and add cleanup timeouts to prevent Copilot CLI hang on exit #44254 (comment)...

Addressed in bbe84f5: the cleanup timeout helper now clears and unreferences its deadline timer when cleanup settles, and the JS regression tests cover the prompt-cleanup path. Local make fmt, make lint, make test-unit, make test, make recompile, and make agent-report-progress passed. Checks on the new head still need a maintainer re-trigger because agent pushes do not start CI.

[pelikhan]

@copilot revert lock files

[gh-aw-bot]

\n@copilot please run the skill, address unresolved review comments, and rerun checks once the branch is up to date.

Generated by 👨‍🍳 PR Sous Chef · 5.89 AIC · ⌖ 8.02 AIC · ⊞ 7.1K · ◷
_{Comment /souschef to run again}

[github-actions]

🎉 This pull request is included in a new release.

Release: v0.82.7