openclaw - 💡(How to fix) Fix Heartbeat fork bug still reproducing on 2026.5.26 — follow-up to #84332 (locked)

Follow-up to #84332 (locked, can't comment) — heartbeat fork bug is still firing on 2026.5.26. Commit cce0049 either didn't ship in this release or isn't catching the case.

Original bug closed as fixed via PR #83647 on 2026-05-19. I upgraded from 2026.5.3-1 to 2026.5.26 last night (2026-05-27 ~18:41 BST). Same behaviour.

Mitigation we're running

Local LaunchAgent com.openclaw.paige.fork-watchdog polls ps every 60s for claude --resume processes. When exactly two are present it SIGTERMs the younger one. Source: bin/fork-watchdog.py in our workspace.

Since the upgrade the watchdog has detected and killed 15 forks in ~15 hours of active WhatsApp use (Paige is on a WhatsApp DM channel through the gateway).

Version

OpenClaw 2026.5.26 (10ad3aa)

Frequency

• 15 forks killed, 2026-05-27T19:30 BST → 2026-05-28T09:32 BST
• Roughly one per hour of active use
• No idle gaps long enough to claim the bug was "off" at any point

First three forks post-upgrade

[2026-05-27T19:30:34] FORK DETECTED: keep PID 26708 (etime 10:13), kill PID 27029 (etime 00:18)
[2026-05-27T19:45:36] FORK DETECTED: keep PID 26708 (etime 25:15), kill PID 28004 (etime 00:12)
[2026-05-27T20:45:42] FORK DETECTED: keep PID 29055 (etime 11:59), kill PID 29408 (etime 00:18)

Most recent three

[2026-05-28T08:30:58] FORK DETECTED: keep PID 50161 (etime 00:58), kill PID 50197 (etime 00:32)
[2026-05-28T08:54:01] FORK DETECTED: keep PID 50500 (etime 13:04), kill PID 51179 (etime 00:16)
[2026-05-28T09:32:05] FORK DETECTED: keep PID 50500 (etime 51:08), kill PID 52422 (etime 00:48)

Killed PID etime is always seconds-to-a-minute. Kept PID is always the older one with real conversation context — same shape as the original bug.

Asks

1. Can you confirm whether commit cce0049 actually shipped in 2026.5.26? A quick release-notes check should answer it. If it didn't ship, that explains everything and I'll wait for the next tagged release.
2. If cce0049 IS in 2026.5.26 — there's a code path it isn't catching. Happy to share the full watchdog log, the gateway log windows around each fork event, or run a small repro script. Tell me what would be most useful.

Without the watchdog this is end-user-visible: duplicated replies, context-amnesia replies on busy threads. With it, the duplicates get killed within 60s but the underlying spawn is still happening every ~hour.

Appreciated the quick turnaround on the original — happy to help debug this round too.

— Dan