Debug a run

Start with the run, not the prompt. A plausible answer can still be wrong because it read stale data, resolved the wrong entity, or skipped a tool.

1. Confirm the input and the version that executed.
2. Check the retrieved context: was the relevant record present, in scope, and fresh?
3. Inspect each tool call and policy decision.
4. Check waits, retries, and the final stop reason.
5. Reproduce with a controlled input before changing the definition.

If the same class of failure can recur, add it to an evaluation case set before promoting a fix.