openclaw - 💡(How to fix) Fix [Bug]: Gateway main thread CPU-bound at ~100% on v2026.4.26 / current main; clean on v2026.4.22 (fs.stat storm in microtask queue) [5 comments, 4 participants]

Bug type

Regression (worked before, now fails)

Summary

After upgrading from v2026.4.22 to v2026.4.26 (also reproduces on current main at 9bb1e59, which package.json reports as 2026.4.27), the gateway sits at ~100% CPU on its single main thread and stops responding to local probes. Same host, same ~/.openclaw, no config changes — only git checkout differs.

I've seen #74209 and the regression range overlaps, but on my machine the dominant signal in a CPU sample is a fs.stat storm in the JS microtask queue rather than bonjour. Filing separately in case the maintainer wants to triage as the same root cause or as a sibling regression.

Versions

• macOS 26.4 (Mac Studio, ARM64), Node 22 via Homebrew
• Bad: v2026.4.26 (be8c246), and current main at 9bb1e59 (reports as 2026.4.27)
• Good: v2026.4.22 (00bd2cf)

Steps to reproduce

Same ~/.openclaw, just switch versions:

git checkout v2026.4.26 && pnpm install && pnpm build
openclaw gateway restart
sleep 30
ps -o stat,%cpu,etime $(pgrep -f 'dist/index.js gateway')
# → R  95-100  sustained

git checkout v2026.4.22 && pnpm install && pnpm build
openclaw gateway restart
sleep 30
ps -o stat,%cpu,etime $(pgrep -f 'dist/index.js gateway')
# → S  2.6

Side-by-side, same host

A representative liveness warning on main:

[diagnostic] liveness warning: reasons=event_loop_delay,event_loop_utilization
  interval=320s eventLoopDelayP99Ms=314874.8 eventLoopDelayMaxMs=314874.8
  eventLoopUtilization=0.999 cpuCoreRatio=0.585 active=0 waiting=0 queued=0

i.e. the loop blocked for over 5 minutes with nothing in the active queue.

CPU sample on main (5 s, idle, no incoming traffic)

/usr/bin/sample <gateway-pid> 5 — all 4074 stacks collapse to this single path:

4074 uv__io_poll
+ 4074 uv__async_io
+   4074 uv__work_done
+     4074 MakeLibuvRequestCallback<uv_fs_s>::Wrapper
+       4074 node::fs::AfterStat
+         4074 MicrotaskQueue::PerformCheckpointInternal
+           4074 MicrotaskQueue::RunMicrotasks
+             4074 Builtins_PromiseFulfillReactionJob
+               4074 AsyncFunctionAwaitResolveClosure
+                 4074 <JIT JS, unsymbolicated>

Every sample is in node::fs::AfterStat. The main thread is consumed resolving promises from a flood of fs.stat calls. The same sample on 4.22 with the same config and data spends almost the whole 5 s in kqueue waiting.

Effect on real usage

Channel messages arrive (Discord WS frames are received), the session enters state=processing and ages indefinitely without a reply — I saw agent:main:discord:direct:oliver reach age=1376s queueDepth=1 before the gateway watchdog killed and restarted the process. No subprocess is involved (no docker, no acpx wrapper, no model API call) — the wedge is purely in-JS work on the main thread.

Workaround

Pin to v2026.4.22. Disabling the OpenClaw Auto-Update cron is recommended so the upgrade doesn't reapply.

Related

• #74209 — overlapping regression range, same restart churn, framed around bundled-plugins / bonjour. Possibly the same root cause from a different angle.