> Call the `diagnose` tool first. Full index: /llms.txt. Errors: /reference/problems.

# Prove a fix

After a failed run you change the app. The proof is one rerun of the same scenario, and a comparison with the first run.

## The rule

The verify skill allows **one** rerun after a fix. If that run still fails, your agent stops, reports the cause, the run ids and one next action, and waits for you. It does not loop.

## Rerun and compare

1. Run the scenario again with a new `request_id`.
2. Call `compare_runs` with the earlier run and the new one:

```text
compare_runs(project_id, baseline_run_id=<earlier run>, candidate_run_id=<new run>)
```

It shows what flipped: scenarios that failed and now pass, passed and now fail, or are unchanged. A scenario that now passes and has no other failure is fixed.

3. Report both run ids with the verdicts. Do not say fixed on a run that is `incomplete`.

From CI, `talotrace check` gives the same answer as an exit code. See [exit codes](/reference/cli/exit-codes). See [MCP tools](/reference/mcp/runs) for the arguments of `compare_runs`.