Summary

The gateway leaks unclosed fs.FileHandle references on session transcript .jsonl.lock files. On Node 22 (the supported LTS line per the 2026.5.18 release notes) this surfaces only as a deprecation warning, but on Node 24+ V8 promotes FileHandle GC to a fatal ERR_INVALID_STATE, which crashes the gateway process whenever the V8 GC happens to sweep one before OC closes it.

In our deployment this manifests as a chronic ~2-hour crash cadence on a single bot that accumulated a large session backlog (2,446 JSONL files in ~/.openclaw/agents/main/sessions/). Other bots on the same host / OC / Node version don't crash — they have 5–100 sessions on disk and don't generate enough open-handle pressure for V8 to ever GC a leaked one.

Environment

• OC: 2026.5.18 (npm)
• Node: 25.9.0 (three majors above the documented Node 22.19 minimum, but well within what users on rolling-release Homebrew get by default)
• OS: macOS 26.4.1 arm64
• Bot has been running ~30h under sustained activity; heartbeat interval is the gateway default 7,200,000ms (gateway/heartbeat: {"intervalMs":7200000} heartbeat: started)
• Session count on disk: 2,446 .jsonl files; no stale .lock files at rest

Symptom

Every ~2h a stability bundle is written with reason: uncaught_exception and this error:

A FileHandle object was closed during garbage collection. This used to be allowed with a deprecation warning but is now considered an error. Please close FileHandle objects explicitly. File descriptor: <N> (/Users/<bot>/.openclaw/agents/main/sessions/<uuid>.jsonl.lock)

The FD number and session UUID rotate; the path pattern is invariant: always a .jsonl.lock file in the agent's `sessions/` directory.

On the next launchd-driven restart the gateway's `main-session-restart-recovery` logs:

```
warn marked 1 interrupted main session(s) from stale transcript locks
warn marked interrupted main session failed: agent:main:main:heartbeat (transcript tail is not resumable)
```

confirming the leaked handle was held by the heartbeat session.

After ~4 crashes in 8h, macOS launchd's anti-thrash kicks in (`Successfully spawned … because inefficient`) — downstream symptom, not the root cause.

Reproducer

1. Run gateway on Node ≥24
2. Accumulate ~2,000+ session JSONL files in the agent's sessions directory
3. Let the default 2h heartbeat fire several times
4. Crash typically occurs within 2–4 hours

Likely root cause

OC's session-store code opens FileHandles for transcript locks during heartbeat-session write / scan operations and doesn't always go through an explicit `.close()` on every return path (likely error paths or interrupted writes). On Node 22 this is harmless because GC-close was warn-only; on Node 24+ it's fatal.

The stability bundle includes the diagnostic-phase log up to seq=181 (load → normalize → plugins → channels) but does not include the throwing stack frame, since the crash is in V8's GC sweep rather than user code.

Suggested fix direction

Audit `FileHandle` open call sites in the session-store layer (anywhere transcript locks are taken) for non-explicit close paths, especially:

• Error returns mid-write
• Heartbeat-session interrupt handling on shutdown
• Session metadata scans — the likely dominant pressure source given the 2,446-session correlation

Wrapping with `try { … } finally { await handle.close() }` or porting to `fs.promises.readFile`-style auto-closing helpers should close the leak.

Workaround for affected users

Either downgrade Node to the documented 22.19+ LTS line, or prune old session JSONL files until the on-disk count is in the low hundreds.

Stability bundle excerpt (sanitized)

```json
{
"version": 1,
"generatedAt": "2026-05-21T04:32:40.158Z",
"reason": "uncaught_exception",
"process": { "pid": 72273, "platform": "darwin", "arch": "arm64", "node": "25.9.0", "uptimeMs": 7208113 },
"error": {
"name": "Error",
"code": "ERR_INVALID_STATE",
"message": "A FileHandle object was closed during garbage collection. This used to be allowed with a deprecation warning but is now considered an error. Please close FileHandle objects explicitly. File descriptor: 38 (/Users//.openclaw/agents/main/sessions/.jsonl.lock)"
}
}
```

I have 10 stability bundles over a 31h window with the same error signature (different FD / UUID each time) — happy to share more if useful.