adversarial-review

In code review, a single reviewer is inevitably blind to their own gaps, and generic AI review often lumps style nits together with genuine defects. adversarial-review tackles this with an adversarial multi-lens strategy: review is split into independent directions such as correctness, security, and maintainability. Each lens attacks the code change in parallel without interfering with the others, then an arbitration layer merges cross-lens findings, deduplicates them, and grades them using a fully deterministic formula, producing a report that traces every issue back to a concrete checkpoint list entry. The entire process is strictly read-only and never modifies code.

The plugin ships with 16 built-in deterministic static rules (hardcoded credentials, SQL concatenation, disabled TLS verification, and more) that guarantee coverage even if the model layer fails. Suppression rules prevent the same class of problems from being reported repeatedly, a review docket provides incremental fingerprint comparison against known issues, and an optional deep-review stage challenges each candidate finding one by one, downgrading refuted items to observations rather than deleting them. Every severity score is fully deterministic: the same input always yields the same verdict.

This tool is well suited to teams and individual developers who want a structured, auditable review before merges or after refactors, and can also serve as a CI quality gate via a configurable fail-on exit code. The standalone CLI carries zero third-party runtime dependencies and works with any model service speaking the OpenAI Chat Completions protocol. A DeepSeek Harness plugin entry point is also provided, allowing the model to invoke the gavel_review tool directly within a session with no human involvement.