deep-diagnostic-procedure-public

Deep Diagnostic Procedure for OpenClaw Issues

When to Use

• A tool, channel, or feature isn't working as expected
• Symptoms persist after "obvious" fixes
• You're tempted to blame an external component without evidence
• You've restarted the gateway more than twice and the problem persists

Core Principles

1. Don't conclude from first impression. "It's probably X" is a hypothesis, not a diagnosis.
2. Investigate in layers, from closest to the problem to farthest.
3. Evidence, not assumptions. Every hypothesis must be verified with logs, config, or code.
4. One real cause is more likely than three coincidental ones.
5. Policy before plumbing. Permission/config issues cause more outages than network bugs.

Investigation Layers (In Order)

Layer 1 — Policy & Permissions (MOST OVERLOOKED)

• What does tools.profile remove? Check with openclaw config get tools.profile
• Are the required tools in the allowed list or blocked?
• Channel-specific policies: dmScope, allowFrom, groupPolicy
• Cross-context restrictions (e.g., messaging across channels denied)
• Exec security, host approvals, elevated permissions

Key question: "Does the agent actually HAVE ACCESS to the tool/channel it needs?"

💡 Real-world lesson: Weeks spent debugging outbound message delivery. Root cause? tools.profile silently removed the message tool from the agent's session. The agent literally couldn't send — nothing to do with the external library, DNS, or connection.

Layer 2 — Configuration

• Is the config correct in openclaw.json? Verify with openclaw config get
• Empty values, missing keys, or defaults that should be explicit?
• Format correctness (e.g., phone numbers with/without +)
• Plugins and skills enabled correctly?

Key question: "Is it configured the way it should be, not the way I assume it is?"

Layer 3 — Runtime & State

• Is the gateway running? openclaw gateway status
• Are channels connected? openclaw channels status --probe
• Is there an active listener for the target channel?
• Are sessions active and routed correctly?
• Are tokens/credentials valid?

Key question: "Is everything actually running and connected RIGHT NOW?"

Layer 4 — Logs & Errors

• Run openclaw logs --follow while reproducing the issue
• Enable verbose logging temporarily
• Capture the EXACT error message (not a summary)
• Note the timeline: when does the error appear relative to the action?

Key question: "What is the system TELLING US is wrong?"

Layer 5 — Network & External Services

• Does DNS resolve? (Only if symptoms point to network)
• Can you reach external services (Ollama, Groq, GitHub, etc.)?
• Firewall or proxy blocking?
• Latency or timeout issues?

Key question: "Is the problem OUTSIDE of OpenClaw?"

Layer 6 — Code & Plugin Internals

• Audit the plugin code (trace the data flow)
• Map the full path: input → processing → output
• Look for silent failure points (try/catch that swallows errors)
• Check plugin version compatibility with OpenClaw version

Key question: "Where exactly in the code does the flow stop?"

Execution Procedure

1. Write down the precise symptom: "When I do X, Y happens instead of Z"
2. Layers 1-3: Quick verification (target: 10 minutes total)
3. Layer 4: Enable logs, reproduce the problem, capture errors
4. Form primary hypothesis → verify with evidence
5. If hypothesis doesn't confirm → move to next layer
6. Layers 5-6 only if 1-4 reveal nothing
7. Document: CAUSE → EVIDENCE → SOLUTION

Anti-Patterns (WHAT NOT TO DO)

• ❌ "It's probably [external component]" without evidence
• ❌ Restarting as diagnosis (hides the problem, doesn't fix it)
• ❌ "I've seen this before" without verifying it's the same cause
• ❌ Investigating only one layer (e.g., only network)
• ❌ Concluding before checking policy & permissions
• ❌ Blaming the external component before checking internal config

Report Template

## Diagnostic: [Issue title]
**Symptom:** [Precise description]
**Initial hypothesis:** [What I thought first]
**Actual cause:** [What I found]
**Evidence:** [Log/config/code that confirms]
**Layers investigated:** [Which layers, in what order]
**Solution:** [What fixed it]
**Time wasted on wrong hypotheses:** [Estimate]
**Lesson:** [What to remember for next time]

Historical Cases

Case 1: Silent tool removal by tools.profile

• Symptom: Agent couldn't send messages to an external channel from another channel
• Initial hypothesis: External library bug, DNS issues, disconnection
• Actual cause: tools.profile setting silently removed the message tool from the agent session
• Evidence: Logs showed tool policy removed N tool(s) via tools.profile: ...message...
• Layers investigated: 6 (code) → 1 (policy) — WRONG ORDER, started from the end
• Solution: Changed tools.profile to include the message tool
• Time wasted: Weeks (chasing DNS, re-linking, blaming external library)
• Lesson: ALWAYS check policy & permissions FIRST. It's the most overlooked layer and the most common cause of "silent" failures.