openclaw - 💡(How to fix) Fix [Bug]: “EventLoopDelayMaxMs spikes 6–24s + stuck sessions on low-power N355 Proxmox after 4.23” [4 comments, 5 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

On OpenClaw 2026.4.23 the gateway is fast and responsive. After upgrading to any version 4.25+, periodic multi-second eventLoopDelayMaxMs spikes (6–24+ seconds) and stuck-session warnings appear even in clean sessions.

Steps to reproduce

1. Run OpenClaw 2026.4.25+ (or 4.26/4.27/4.29) on Intel N355 Proxmox VM (8 vCPU ~1.88 GHz cap, 11 GB RAM).
2. Use local LM Studio (Qwen 27B/35B) or cloud models (OpenAI/GPT-5.5).
3. Open a fresh session and send a message (or wait ~30s idle).
4. Observe spike after model response arrives.

Expected behavior

Gateway remains responsive with eventLoopDelayMaxMs staying under 1s (as observed on 2026.4.23).

Actual behavior

Repeated eventLoopDelayMaxMs spikes of 6–24+ seconds, stuck session warnings (age=122s → 440s+), and gateway feels unresponsive every ~30s. Spikes occur after model responses and in completely clean sessions.

OpenClaw version

Tested: 2026.4.23 (working) vs 2026.4.25 / 4.26 / 4.27 / 4.29 (broken). Daily driver: 2026.4.23 (a979721)

Operating system

Ubuntu VM in Proxmox (host: Intel N355 mini PC)

Install method

Official release, started via openclaw command / systemd service, auto-starts on boot.

Model

lmstudio/qwen3.6-27b-uncensored-hauhaucs-balanced (and 35B variants); also reproduced with OpenAI/GPT-5.5 cloud models.

Provider / routing chain

Local LM Studio endpoint (remote Win11 machine via Tailscale) and direct OpenAI. Issue occurs on both.

Additional provider/model setup details

N/A (reproduced on multiple providers)

Logs, screenshots, and evidence

-eventLoopDelayMaxMs values consistently 6.2s – 24.8s in logs.
-Stuck session warnings every 30–60s.
-Clean sessions: LCM assemble/compaction 10–117 ms, yet spikes continue.
-Extensive plugin disabling (lossless-claw, memory-core, memory-wiki, telegram, discord, browser, etc.) did not resolve it.

Impact and severity

Affected: All sessions on this low-power N355 VM setup. Severity: Blocks daily use (high latency, unresponsive gateway). Frequency: Every ~30 seconds or after model responses. Consequence: Unusable for real-time chat; forces staying on 4.23.

Additional information

Regression introduced in 4.25 overhaul (new plugin registry, heavier memory/context features). Not network or model-provider specific. Hardware: Intel N355 VM (capped ~1.88 GHz), UV_THREADPOOL_SIZE=8, bonjour disabled. Last known good: 2026.4.23. First known bad: 2026.4.24+.