Follow-up to #75598 after testing the current package requested in the maintainer comment.

Original issue: https://github.com/openclaw/openclaw/issues/75598

Summary

On OpenClaw 2026.5.2 (8b2a6e5), the gateway still pins one Node.js CPU thread even after reducing the runtime to a minimal CPU test profile:

• no external chat channels
• no memory plugins
• no browser
• no Discord / Telegram
• no extra skills
• minimal tools profile
• LM Studio kept as the model provider target

The gateway starts and reports ready, but idle CPU remains around 100% of one core. A simple agent turn fails through the gateway websocket and falls back to embedded mode.

Environment

• OpenClaw: 2026.5.2 (8b2a6e5)
• Node: 22.22.2
• OS: Ubuntu 24.04 VM under Proxmox
• Hardware: Intel N355 mini PC, low-power VM environment
• Service: user systemd service, openclaw-gateway.service
• Model/provider: LM Studio local provider target
• Related regression window from original issue: 2026.4.23 works; 2026.4.24/4.25+ through 2026.5.2 regress

Minimal config tested

Effective minimal test state after validation:

channels.enabled=
plugins.allow=lmstudio
plugins.slots={"memory":"none","contextEngine":"none"}
browser.enabled=False
tools.profile=minimal
tools.deny=*
tools.exec.security=deny
enabled.skills=

Gateway log after restart:

http server listening (0 plugins, 6.2s)
gateway ready
ActiveState=active
SubState=running
NRestarts=0

CPU measurements

Idle samples from the gateway PID:

105 105 105 104 104 104 104 104 104 104

During one simple chat turn:

104 104 104 104 104 104 104 104 103 103 103 103 103 103 103 103 103 103 103 103 103 103 103 103 103

After chat turn:

102 102 102 102 102 102 102 102 103 103

/ready still reports degraded event loop:

{
  "ready": true,
  "failing": [],
  "eventLoop": {
    "degraded": true,
    "reasons": ["event_loop_delay", "event_loop_utilization", "cpu"]
  }
}

Simple chat turn result

Command:

openclaw agent --agent main --message 'Reply with exactly: OK' --timeout 120 --json

Result:

gateway connect failed: Error: gateway closed (1000):
EMBEDDED FALLBACK: Gateway agent failed; running embedded agent:
GatewayTransportError: gateway closed (1000 normal closure): no close reason

Fresh trace line from fallback:

[trace:embedded-run] startup stages:
totalMs=9174
runtime-plugins=8668ms
model-resolution=165ms
auth=333ms
context-engine=1ms
attempt-dispatch=5ms

Additional profiling signal

A 15s Node inspector CPU profile taken during the degraded state showed repeated synchronous filesystem/config/plugin discovery work in the hot path, including:

node:fs                         12.1%
discovery-B19Xdk1_.js            8.0%
boundary-file-read-*.js          6.8%
parse-json-compat-*.js           6.3%
manifest-registry-*.js           1.3%
installed-plugin-index-store     0.9%

Top sampled functions included repeated:

lstat
readFileUtf8
readFileSync
realpathSync
parseJsonWithJson5Fallback

Expected

With 0 loaded plugins, no external channels, browser disabled, memory/context disabled, and minimal tools, the gateway should idle near 0% CPU and should be able to handle a trivial local agent turn through the gateway.

Actual

The gateway remains alive but consumes around 100% of one CPU core at idle. The trivial agent turn does not complete through the gateway websocket and falls back to embedded mode.

Notes

• openclaw config validate passed for the final minimal config.
• openclaw doctor --fix was run during the test process.
• The issue is still reproducible without Discord, Telegram, browser, memory-core, memory-wiki, or lossless-claw enabled.
• This suggests the remaining issue is not only channel/network related; there appears to be a core gateway/runtime plugin discovery or reload loop that persists even under a 0-plugin gateway startup.