ucsandman / troubleshoot

Troubleshoot DashClaw

Systematic diagnostics for common DashClaw errors, signal anomalies, and configuration issues.

Error Code Diagnostics

401 Unauthorized

Symptom: API calls return 401.

Checklist:

1. Verify x-api-key header is set (not Authorization: Bearer)
2. Check DASHCLAW_API_KEY environment variable is set
3. Confirm the key hasn't been rotated — API keys are shown once at creation

4. Test with curl:
   `bash
   curl -H "x-api-key: $DASHCLAWAPIKEY" $DASHCLAWBASEURL/api/health
   `

Root cause: DashClaw uses x-api-key header, not Bearer tokens. The middleware does timing-safe comparison first, then falls back to hash lookup.

403 Forbidden

Symptom: API calls return 403.

Checklist:

1. Demo mode? Demo mode blocks all write operations. Check DASHCLAW_MODE env var.
2. Readonly key? Some keys are read-only. Check key permissions in dashboard.
3. Guard blocking? If calling /api/guard and getting 403, a policy is blocking the action — this is working as intended.
4. org_default trap? The org_default org blocks API access except onboarding routes. Create a real org first.

429 Rate Limited

Symptom: API calls return 429 Too Many Requests.

Defaults:

• Production: 100 requests/minute per IP
• Development: 1000 requests/minute per IP

Fixes:

• Set DASHCLAW_DISABLE_RATE_LIMIT=true for local development
• For production: use UPSTASH_REDIS_REST_URL for distributed rate limiting
• Batch operations where possible

503 Server Misconfigured

Symptom: API calls return 503.

Checklist:

1. Is DASHCLAW_API_KEY set? Missing key → 503 on protected routes
2. Is DATABASE_URL valid? Check connection string
3. Run health check: curl $DASHCLAW_BASE_URL/api/health
4. Check the /setup page for readiness verification

Common Gotchas

Signal Debugging

DashClaw computes 17 signal types (computeSignals in app/lib/signals.ts). The eight most common in agent integrations are below; if signals are firing unexpectedly:

ungoverned_scope ("Governance scope narrowed") is the one to check first when the ledger looks too clean. It fires when DASHCLAW_GOVERNED_CATEGORIES drops a category below the default set, which means the hook exits before the guard call for that category — those tool calls produce no row at all, so an empty ledger looks the same as a well-behaved agent. Remove the variable or set it to all, and check it for typos; a misspelled category is simply never governed. search and system are ungoverned by default and never raise it.

1. Autonomy Spikes

Trigger: >10 ungoverned actions/hour
Fix: Add guard checks before actions. Use claw.guard() before claw.createAction().

2. High Impact, Low Oversight

Trigger: Irreversible decisions with risk ≥70 and no approval
Fix: Add approval gate policy for high-risk irreversible actions.

3. Repeated Failures

Trigger: >3 failures in 24 hours
Fix: Check agent logic. Review failed actions in dashboard for patterns.

4. Stale Open Loops

Trigger: Unresolved dependencies >48 hours old
Fix: Resolve or cancel open loops: claw.resolveOpenLoop(loopId, 'resolved', 'Fixed').

5. Assumption Drift

Trigger: ≥2 invalidated assumptions in 7 days
Fix: Review assumptions. Agent may be operating on stale beliefs.

6. Stale Assumptions

Trigger: Unvalidated assumptions >14 days old
Fix: Validate or invalidate old assumptions: claw.validateAssumption(id, true/false, reason).

7. Stale Running Actions

Trigger: Actions with status running for >1 hour
Fix: Two layers of cleanup:

1. One-shot bulk repair for any historical backlog:
   `bash
   # Preview what would change
   node scripts/_run-with-env.mjs scripts/repair-stale-running-actions.mjs --dry-run --older-than-hours 1
   # Apply (status='completed' with "Auto-closed" summary; preserves error_message → 'failed')
   node scripts/_run-with-env.mjs scripts/repair-stale-running-actions.mjs --older-than-hours 1
   `

2. Prevent re-accumulation — install dashclaw_stop.py (Stop hook). At turn end it queries each open action and PATCHes status='completed' only when the action is still running. Terminal statuses written by PostToolUse are preserved. See setup-dashclaw skill.

For ad-hoc closure of a single action: claw.updateOutcome(actionId, { status: 'failed', output_summary: 'Timed out' }).

8. Agent Silent

Trigger: Agent heartbeat lost >10 minutes
Fix: Ensure agent sends heartbeats: claw.heartbeat({ status: 'online' }).

Diagnostic Tools

Health probe

curl -sf "$DASHCLAW_BASE_URL/api/health" | jq '.'

Returns {status, version, checks: {database, runtime, realtime, ...}}. Anything other than status: "healthy" points at the failing check.

Doctor (config audit)

node scripts/doctor.mjs

Audits the local config — env vars, DB connectivity, schema state, generated artifact freshness.

Live SDK round-trip

# Node — guard → createAction → updateOutcome against the live instance
node scripts/_run-with-env.mjs scripts/test-sdk-live.mjs

# Python equivalent
node scripts/_run-with-env.mjs scripts/run-sdk-live-python.mjs

Both scripts emit a real action that you should see appear on /decisions within seconds. If the round-trip fails, the error pinpoints the layer (HTTP, validation, DB).

Startup smoke test

npm run startup:smoke

Boots Next.js, hits the critical routes, and reports which (if any) fail to respond.

Hook Troubleshooting

Pretool not firing

• Verify .claude/settings.json has PreToolUse hook configured
• Check matcher pattern: Bash|Edit|Write|MultiEdit
• Verify Python is available: python --version
• Check hook script path is correct relative to project root

Pretool allows everything

• Check DASHCLAW_HOOK_MODE — if set to observe, it logs but never blocks
• Check DASHCLAW_RISK_THRESHOLD — default is 60, lower it to catch more
• Verify policies exist: curl -H "x-api-key: $KEY" $URL/api/policies

Pretool blocks everything

• Check guard policies — a too-broad policy may be catching all actions
• Try DASHCLAW_HOOK_MODE=observe first to understand what's being caught
• Check risk scoring — are file operations being scored too high?

Posttool not recording outcomes

• Check temp file bridge: pretool writes to {tempdir}/dashclaw_last_action_{tool_use_id}
• Verify DASHCLAW_BASE_URL and DASHCLAW_API_KEY are set for posttool
• Posttool never blocks — failures are silent. Check DashClaw server logs.

Stop hook not capturing tokens (Analytics shows $0 / 0 tokens)

• Verify .claude/settings.json has the Stop block (one entry, no matcher needed). If missing, re-run node /path/to/DashClaw/scripts/install-hooks.mjs --target=..
• During a turn, confirm pretool is appending actionids: `ls /tmp/dashclawturn<sessionid>(Linux/macOS) ordir %TEMP%\dashclawturn<sessionid>(Windows). After Stop fires, that file is deleted and/tmp/dashclawstopcursor<session_id>` appears.

• Manually fire the Stop hook to test:

  echo '{"session_id":"<your-session-id>","transcript_path":"<path/to/.jsonl>"}' \
    | python .claude/hooks/dashclaw_stop.py; echo "exit=$?"

  Then query the DB for any actionid from that session — `tokensin,tokensout,model,costestimate` should be populated.

• Confirm PATCH /api/actions/:id accepts token fields by checking the deployed server version: curl -sf $DASHCLAW_BASE_URL/api/health | jq '.version' should be ≥ 2.13.1. Older deploys silently drop tokens_in/tokens_out.

Stale running actions piling up

• Verify the Stop hook is installed (above). It auto-closes any action still in status='running' at turn end. Without it, every interrupted/abandoned tool stays open forever.
• For a one-shot cleanup: node scripts/_run-with-env.mjs scripts/repair-stale-running-actions.mjs --older-than-hours 1 (preview with --dry-run first).