Troubleshooting¶

Metrics stop but the service is up. Check codingagent.selfobs.ingest_lag first. A stalled source reads as a quiet period, not as an error.

A record type vanished after a client upgrade. codingagent.selfobs.unhandled_record_types and codingagent.drift_findings. Run ca2probe against a fresh sample and compare to the baseline.

Tool counts doubled for Codex. Deduplication on item id regressed. response_item:custom_tool_call and event_msg:item_completed describe the same call.

A metric dimension collapsed to other. The value hit metrics.dimension_cap. Check codingagent.selfobs.attribute_rejections and decide whether the allowlist needs widening or the field belongs in AttrRecord instead.

TTFT is empty for Claude Code. Correct. Claude Code transcripts carry no measured TTFT and a derived value must never enter that instrument.

Cost is missing for most sessions. cost-state is present in only a minority of Claude Code sessions and Codex has none at all. Native claude_code.cost.usage is the live source.

Nothing exports and there is no error. Configuration validation. An unknown key is an error at startup; check the service logs for the validation failure rather than assuming the endpoint.