What problem does it solve? Solo LLM code reviews suffer from self-ratification and confident false positives, so this Skill runs two reviewers from different model families (Claude and Codex/GPT) on the same diff and has each try to refute the other's findings, reporting only cross-examined results ranked by confidence. ## Core Features & Use Cases - Cross-vendor independent review: Claude and Codex/GPT each review the same git diff in isolation, one natively and one as a subprocess CLI, then emit structured findings matching a JSON schema. - Adversarial refutation phase: Each model attempts to refute the other's findings, and confidence (high, medium, low/contested) is assigned based on whether findings survive cross-examination. - Deterministic gates: Mechanical checks like lint, typecheck, and grep-based house-rule detection run before the models and are reported separately as facts. - Use Case: Before merging a feature branch, ask for an adversarial review against main to get a ranked table of findings where each item shows which model found it and whether the other model upheld or refuted it. ## Quick Start Run an adversarial review of my current branch against main and show me the ranked findings.
