fix(session-end): remove .unref() so process force-exits after 1500ms

[dnraitzyk]

Fixes #1069.

What

Remove .unref() from the setTimeout in session-end.mjs / session-end.ts.

Why

The session-end hook fires fire-and-forget fetch() calls to the local daemon, then does:

setTimeout(() => process.exit(0), 1500).unref();

AGENTS.md documents the intent: the process should force-exit after 1500ms so it does not block Claude Code's shutdown. But .unref() makes the timer passive — Node.js does not keep the event loop alive for an unref'd timer, but it also will not fire it while other handles (the pending fetch() calls) are keeping the loop alive.

The fetches use AbortSignal.timeout() values of 30s, 60s, and 120s. Those pending requests keep the event loop alive the entire time. Claude Code's hook deadline fires first and kills the process with:

SessionEnd hook [node "${CLAUDE_PLUGIN_ROOT}/scripts/session-end.mjs"] failed: Hook cancelled

This is intermittent because when the daemon responds fast the event loop drains before the deadline; when it is slow the race is lost.

Fix

Remove .unref(). An active timer is the only handle that keeps the event loop alive after the fetches are dispatched, so the process exits after exactly 1500ms regardless of whether the fetches have settled. This matches the documented intent and the fire-and-forget pattern AGENTS.md prescribes.

- setTimeout(() => process.exit(0), 1500).unref();
+ setTimeout(() => process.exit(0), 1500);

Summary by CodeRabbit

• Bug Fixes
  • Improved session-end handling so processes terminate reliably after the session ends.
  • Prevented lingering network requests from delaying shutdown beyond the expected timeout.

[vercel]

@dnraitzyk is attempting to deploy a commit to the rohitg00's projects Team on Vercel.

A member of the Team first needs to authorize it.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 45de2006-8654-4e63-8036-b4218b030b35

📥 Commits

Reviewing files that changed from the base of the PR and between 93ae9bc and 1ba4419.

📒 Files selected for processing (2)

• plugin/scripts/session-end.mjs
• src/hooks/session-end.ts

---

📝 Walkthrough

Walkthrough

Updated both session-end implementations to use referenced 1500ms exit timers, ensuring the process terminates after the timeout regardless of pending fetch requests. Inline comments document the event-loop behavior.

Changes

Session-end process termination

Layer / File(s)	Summary
Force session-end termination src/hooks/session-end.ts, plugin/scripts/session-end.mjs	Replaces unref’d exit timers with active 1500ms timers and documents the resulting event-loop behavior.

Estimated code review effort: 1 (Trivial) | ~5 minutes

Possibly related issues

• rohitg00/agentmemory#991 — Addresses the same session-end deferred exit timer behavior.

Possibly related PRs

• rohitg00/agentmemory#688 — Implements the earlier fire-and-forget and unref’d timeout pattern in the same session-end exit logic.

Suggested reviewers: rohitg00

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: removing .unref() so session-end exits after 1500ms.
Linked Issues check	✅ Passed	The PR implements the linked bug fix by removing .unref() from the 1500ms session-end timer in both files.
Out of Scope Changes check	✅ Passed	The changes stay scoped to the session-end timeout behavior and related comments, with no unrelated code changes.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[dnraitzyk]

How do approvals work?

[sla-te]

I hit the same "Hook cancelled" message and went looking for this PR, but the measurements do not support the diagnosis. This change makes the healthy case 27x slower and leaves the failing case unchanged.

.unref() does not make a timer passive in the sense assumed here. It only stops the timer from keeping the event loop alive by itself. The timer still fires normally as long as something else holds the loop open, and the pending fetch() is exactly that. So the 1500 ms guard already fires and already bounds the hook.

Measured on the shipped session-end.mjs (0.9.29 dist, node v24.19.0), same input on stdin, wall clock around the process:

daemon state	current (.unref())	this PR (no .unref())
responding normally	56 ms	1549 ms
accepts the connection, never answers	1548 ms	1548 ms
SYN dropped, connect never completes	1546 ms	1548 ms

Two things follow. The hook never waits for the 30s/60s/120s AbortSignal timeouts, so that part of the root cause in #1069 does not happen. And removing .unref() cannot fix the report, because the failing case is already exiting at ~1.5 s. All the PR does is drag the fast path down to the same 1.5 s.

What actually produces the message is that ~1.5 s is longer than the shutdown grace Claude Code gives a SessionEnd hook, which is how #991 frames it. On my box the hook only crossed that line while the daemon was wedged (accept queue full, Recv-Q 129 > backlog 128, engine stuck in futex_do_wait at 12.9 GB RSS). Restarting the daemon put it back to 56 ms and the message stopped.

So the fix has to shorten the worst case, not force the best case to it. Lowering the guard on session-end and stop to something safely under the grace window would do it and keeps .unref() doing its job. The 1500 ms in d626b4e was chosen to let several fetches get initiated, and initiating them is fast, so there looks to be plenty of room below it.

Reproduction, if useful: point AGENTMEMORY_URL at nc -l 127.0.0.1 39111 for the accepts-but-silent case and at an address that drops SYNs for the other, then time echo '{"session_id":"x","reason":"other"}' | node scripts/session-end.mjs.

Is there a documented number for the SessionEnd grace window, or does the cap have to be found empirically?