Require a compact evidence handoff

Use the agent summary for navigation. Because it comes from the same system that produced the code, check its claims against evidence that another process or person can reproduce.

Require a compact handoff:

• the task and starting commit;
• the changed files and reason for each;
• the test added or modified to express the requirement;
• commands run and their actual outcomes;
• manual checks that remain;
• assumptions, scope changes and unresolved concerns;
• migration or rollback notes where relevant.

A failing test before the fix is strong evidence for a defect repair. Preserve the failure through a commit, test output or reviewer-reproducible step. For work where test-first implementation is impractical, require another baseline such as a recorded query plan, an API contract fixture or a deterministic reproduction script.

Separate verification from implementation where risk justifies it. Continuous integration should run in a clean environment. A second agent can inspect the diff against the task, search for missing tests or run specialised analysis, but that review remains advisory. It can catch omissions quickly; it cannot accept the business consequence on behalf of the human owner.

Do not ask reviewers to read full agent transcripts by default. Transcripts contain exploration, abandoned theories and repeated tool output. Keep them available for investigation, while the handoff identifies the decisions that survived. If a significant design choice occurred during implementation, write it into the pull request or decision record.

Check the provenance of generated artefacts. A snapshot, fixture or migration containing hundreds of changed lines may be correct, yet it can hide an unintended update. Include the command used to generate it and a focused explanation of the meaningful difference.