code-review-and-quality

Review the tests first to confirm intent and coverage, then walk each changed file across five axes: correctness, readability, architecture, security, and performance. Label every finding with a severity prefix and finish with an adversarial pass that tries to disprove each acceptance criterion.

A review checklist should cover correctness against the spec, edge case and error path handling, naming and simplicity, architectural fit, input validation and secret hygiene, and performance issues like N+1 queries. It should also verify that tests pass and the verification story is documented.

AI-generated code needs more scrutiny, not less, because it is confident and plausible even when wrong. Verify behavioral tests actually execute each acceptance criterion rather than asserting mocked values, and run an adversarial pass targeting boundary, persistence, and concurrency inputs.

Split a change when it approaches 1000 lines, since changes around 100 lines are reviewable in one sitting and 300 lines is acceptable only for a single logical change. Use stacking, file-group, horizontal, or vertical splitting strategies, and always separate refactoring from feature work.

An adversarial pass is a mandatory disconfirming review that tries to prove each acceptance criterion is not met, attacking domain failure modes and boundary inputs. If it returns REFUTED, the verdict is request changes regardless of how clean the constructive review looked.