adversarial-review

An adversarial code review prevents defective code by orchestrating a two-model cycle where one generates code and another critically reviews it. This process treats critical findings as blocking, requiring fixes before completion.

To validate Python code changes before merging, run an adversarial review on changed files, address all critical findings, and verify the module by running py_compile and import checks to ensure a safe completion state.

If the secondary model router is unavailable, you can still run an adversarial review using a safety-aware execution fallback. The workflow notes the instability and provides a single-model adversarial pass alternative.

Pre-merge validation supports gating developer agent implementation steps by enforcing quality over speed. It requires orchestrating the adversarial review logic and passing compilation checks before marking an implementation complete.

The best way to use LLM orchestration for critical code flaw detection is a two-model adversarial cycle. One model generates the code, another reviews it adversarially, and the first synthesizes a stronger output to catch critical flaws.

You should not use an adversarial review workflow when development speed matters more than code quality. The process enforces strict critical-before-done gating and requires compilation verification, intentionally slowing down completion.