Log Triage

Turn noisy logs into a ranked, human-readable incident digest.

Sources

• Bot log — the chat process log (e.g. bot.log / wherever run_chat.sh writes). Start here.
• Langfuse — if LANGFUSE_* is configured, traces give per-request timing, model calls, and errors. See .claude/sections/08_* (observability).

Triage method

1. Scope — last N hours, or since the last known-good marker.
2. Extract errors — grep for ERROR/CRITICAL/Traceback/Unhandled.
3. Cluster — group by normalized message (strip timestamps, IDs, paths) so "the same error 200 times" reads as one cluster with a count.
4. Rank — by frequency × severity. The top 1–3 clusters are the headline.
5. Diagnose — for each headline cluster: first seen, last seen, count, the representative stack frame, and a one-line hypothesis.

Output

🔧 Log triage — last <window>

1. <normalized error>  ×<count>
   first <ts> · last <ts>
   at <file:line>
   likely: <hypothesis>

2. ...

Quiet otherwise: <count> warnings, no other errors.

Keep it to the signal. Do not paste raw log spans — cite file:line and counts. This skill is read-only; it diagnoses, it does not restart or patch.