What should an AI code reviewer examine before changes are committed?

What should an AI code reviewer examine before changes are committed?

An AI code reviewer should examine whether the proposed change satisfies its requirements and introduces defects that matter to users. Review is most effective when the assistant has the relevant diff, surrounding implementation, and acceptance criteria.

Follow data and permission boundaries

For an event-creation endpoint, inspect how the authenticated user is identified, how wedding membership is checked, and how the new event is associated with its owner. Do not assume that a wedding identifier supplied by the client proves authorization.

Check validation, error responses, and persistence behavior. Consider whether retries can create duplicate records, whether date handling is consistent, and whether failed operations leave the interface in a misleading state.

Compare contracts across layers

Verify that the form sends the fields the API expects and handles the response the API actually returns. Review empty, loading, and failure states along with the successful path. Examine related views that may need refreshing after a mutation.

Look for unnecessary dependencies, exposed configuration values, and unrelated changes that expand the review scope. Assess maintainability in the context of existing repository conventions rather than demanding stylistic rewrites with no practical benefit.

Evaluate tests and report actionable defects

Ask whether tests exercise meaningful behavior and whether they would fail for the defect they claim to prevent. Generated tests can accidentally repeat implementation assumptions or omit permission scenarios.

Findings should identify a specific condition, location, and consequence. For example: “A collaborator from another wedding can submit this identifier because membership is never checked.” Prioritize confirmed correctness and security issues over speculative improvements. After a fix, review the updated diff and verify the affected scenario before treating the finding as resolved.