Every failure the system hit, whether or not it recovered. Most of these should recover — a tool error goes back to the model as data, a batch timeout falls back to the serial path. What must not happen is recovery that leaves no trace, so recovery is recorded rather than done quietly.
| Reference | When | Kind | Where | What happened | What we did about it |
|---|---|---|---|---|---|
| err_e9dc2e | 2026-08-23 18:43:40 | unhandled invalid_output | agent.loop | RuntimeError: unexpected content block type from provider | nothing — this one stopped the run |
| err_09da39 | 2026-08-23 18:43:40 | recovered timeout | triage.batch | TimeoutError: batch still in_progress after 1800s | fell back to serial triage for 61 documents, at full price |
Quote a reference to see the full traceback and the context it happened in: uv run regwatch errors err_e9dc2e