Browse documentation
Review an agent change
Check requirements, boundaries and evidence against the changed source.
Review an agent change against the requirement, its affected contracts and applicable evidence. Code and tests can agree while both implement the wrong rule.
1. Fix the expected behaviour
Write success and refusal cases before evaluating the patch. A reservation change might require sufficient stock, positive quantity, account scope and duplicate-request handling.
Map each criterion to an operation or test. Record missing coverage explicitly.
2. Inspect the affected code
For a graph application at ./app:
effectful code find ./app --query reserve --format json
Use the returned definition and snapshot IDs to retrieve checked context. Search hits identify candidates; dependency relationships identify the structural context.
Check changes to state ownership, message types, effects, target actors, grants and saved-state compatibility. Include native adapters and external delivery contracts where affected.
3. Check the evidence
Using the application’s actual receipt path:
effectful code evidence ./app --receipt ./run-001.json --format json
This query runs no tests. Read outcome, origin and applicability together. A stale pass does not cover current inputs; an applicable failure remains a failure. Request fresh execution where results are missing or cannot be checked, using a new receipt path.
4. Test the changed boundary
Scripted graph tests cover their supplied inputs and effects. Adapter, UI, provider and deployment changes need corresponding integration tests. Durable schema changes need a reviewed migration and rollback procedure.
Include refused and ambiguous outcomes, not just success.
5. Retain the decision
Keep the accepted criteria, code identities and evidence with the review. State which behaviour was checked and any remaining limits. The evidence reference defines the result fields.