adversarial-review

Adversarial code review challenges recent diffs and large refactors using multiple critical lenses from the opposite model. It executes via `codex exec` or `claude -p` to stress-test intent alignment and structural correctness, producing a severity-ordered synthesized verdict with deduplicated findings.

Model switching ensures critique is not produced by the same underlying reasoning path by spawning 1–3 reviewers on the opposite model. This principle-guided challenge uses dedicated reviewer lenses to identify real defects and verify planning artifacts before changes ship.

Yes, you can ask the system to run an adversarial review on your latest work and produce a synthesized verdict without making any changes. The review yields a final lead judgment that may accept or reject individual findings based on severity.

Opposite-model adversarial reviewing requires executing the opposite model via its required CLI, specifically `codex exec` or `claude -p`. This ensures the critique is generated independently from the original model's reasoning path.

The severity-structured verdict synthesis produces a single PASS, CONTESTED, or REJECT verdict. It consolidates deduplicated findings ordered by severity, plus a final lead judgment that may accept or reject individual findings to reduce the risk of shipping flawed changes.

Adversarial review targets recent diffs, large refactors, and planning artifacts where intent alignment and structural correctness must be stress-tested. It is designed for pre-shipment risk assessment rather than routine syntax correction or formatting tasks.