openclaw - 💡(How to fix) Fix [Bug] Event loop freeze (9+ min) + Discord heartbeat timeout + local embeddings regression [2 comments, 3 participants]

Bug Description

Since upgrading from v2026.4.26 (which was stable for weeks) to v2026.4.29, the gateway experiences extreme event loop blocking that causes cascading channel failures.

Environment

• WSL2 on Windows, Linux 6.6, node 22.22.2, 11GB RAM
• Gateway v2026.4.29 running as systemd service
• 116 bundled plugins loaded
• Local node-llama-cpp embeddings with GGUF model for memory search
• Two Discord accounts, two Telegram accounts

Symptoms

1. Event loop freeze Diagnostic endpoint /readyz shows:

{
  "eventLoop": {
    "degraded": true,
    "reasons": ["event_loop_delay", "event_loop_utilization", "cpu"],
    "delayMaxMs": 566398,
    "utilization": 0.973
  }
}

eventLoopDelayMaxMs of 566,398ms (9.4 minutes) — event loop completely frozen.

2. Discord failures

• Gateway heartbeat ACK timeout — both bots drop and reconnect in loops
• Bots appear online in Discord but do not respond to any messages
• Discord API calls timing out: /users/@me at 21+ seconds
• Error: Gateway heartbeat ACK timeout

3. Telegram failures

• getUpdates polling stalls for 590+ seconds
• Multiple restart cycles with [diag] polling cycle finished reason=polling stall detected

4. Per-turn plugin re-evaluation Embedded run traces show every message triggers full plugin reload:

first run:  core-plugin-tools: 11965ms, system-prompt: 143488ms, total: 170551ms
second run: core-plugin-tools: 22732ms (grows!), system-prompt: 51513ms, total: 92050ms

system-prompt taking 143 seconds on first run — context assembly blocked by plugin re-evaluation.

Regression Timeline

• v2026.4.26: stable — no issues with CPU, memory, or gateway
• v2026.4.27: runtime-deps packaging bugs (node-llama-cpp missing, sqlite-vec missing) — fixed with manual copy workaround
• v2026.4.29: fresh upgrade — regression to event loop blocking

Comparison to Existing Issues

• Issue #75512 (per-turn plugin re-evaluation) matches our traces exactly, but our case involves local embeddings making it worse
• Issue #75516 (stuck session CPU regression) reported 22s max — our evidence shows 9+ minute freeze is possible
• Issue #74922 (ackReaction/TLS) — our Discord failures may share the same underlying cause

What might be unique to our setup

The combination of:

1. WSL2 without GPU (node-llama-cpp falls back to no-GPU mode, CPU-saturated)
2. Local memory embeddings always active
3. 116 bundled plugins

May amplify the per-turn plugin re-evaluation issue beyond what other reporters are seeing.

Workaround

Downgrading to v2026.4.26 confirmed as working. OPENCLAW_DIRTY_DISCOVERY_CACHE=1 workaround not yet tested.