neural-network-diagnostics（神经网络诊断）

Neural Network Ops

Purpose

Keep OpenClaw responsive by managing model providers, routing, and fallback behavior.

Fast Triage

Run these checks first:

systemctl is-active openclaw-gateway ollama
journalctl -u openclaw-gateway -n 40 --no-pager
free -h

Focus on these log patterns:

• rate limit reached
• Model context window too small
• Unknown model
• No endpoints available
• sendMessage failed
• embedded run timeout
• Removed orphaned user message

Routing Policy

Use this default priority for production:

1. groq/llama-3.3-70b-versatile (fastest cloud path)
2. openrouter/xiaomi/mimo-v2-pro (high quality backup)
3. openrouter/meta-llama/llama-3.3-70b-instruct:free
4. ollama/qwen2.5:7b (last-resort local fallback)

Avoid 35B local models on 30GB RAM CPU servers for real-time Telegram replies.

Stable Model Constraints

For local Ollama fallbacks:

• contextWindow >= 16000
• Keep maxTokens moderate (1024-2048) for latency
• Pre-warm after restart if local fallback is expected

Example local provider entry:

{
  "id": "qwen2.5:7b",
  "name": "Qwen 2.5 7B (local)",
  "contextWindow": 32768,
  "maxTokens": 2048
}

Recovery Playbook

1) Bot silent in Telegram

journalctl -u openclaw-gateway --since '10 min ago' --no-pager

If sendMessage failed, check network/provider errors first, then restart:

systemctl restart openclaw-gateway

2) Repeated orphaned user message

systemctl stop openclaw-gateway
rm -rf /root/.openclaw/.openclaw/agents/main/sessions/*
echo '{}' > /root/.openclaw/.openclaw/agents/main/sessions/sessions.json
chmod 600 /root/.openclaw/.openclaw/agents/main/sessions/sessions.json
systemctl start openclaw-gateway

3) Groq/OpenRouter rate limits

• Keep Groq as primary, but ensure at least one non-free fallback.
• For OpenRouter 404 privacy/policy errors, adjust data-policy settings in OpenRouter dashboard.
• Do not loop retries endlessly; rely on fallback chain.

4) Local fallback too slow

• Restart Ollama cleanly and warm one small model.
• Do not keep multiple heavy runners resident.

systemctl restart ollama
curl -s -X POST http://127.0.0.1:11434/api/chat \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen2.5:7b","messages":[{"role":"user","content":"hi"}],"stream":false}'

Output Format

When reporting health, return:

## Status
- Gateway: <active/inactive>
- Telegram provider: <connected/stalled>
- Primary model: <provider/model>
- Fallback chain: <ordered list>

## Findings
- <most critical issue first>
- <secondary issues>

## Actions Applied
- <exact changes made>

## Next Step
- <single user action to verify>

Guardrails

• Never expose raw API keys in replies.
• Never execute irreversible financial actions automatically.
• Ask for explicit confirmation before account registrations or external postings.
• Prefer reversible config changes and keep backups before major edits.