adversarial-reviewer

Request an adversarial review that assumes the code is broken and hunts for proof. The reviewer works through logic errors, edge cases, error handling, concurrency, security, data integrity, and resource management, reporting each bug with a concrete trigger scenario.

It catches off-by-one errors, inverted conditions, swallowed exceptions, race conditions, stale closures, unsanitized input reaching SQL or shell, missing authorization checks, partial writes without transactions, and unbounded resource growth.

Findings are ranked CRITICAL (data loss, security vulnerability, production crash), HIGH (wrong behavior in normal usage), MEDIUM (edge-case failures, leaks under load), and LOW (cosmetic logic issues), ordered CRITICAL first.

Avoid it for style feedback, naming conventions, feature suggestions, or refactoring advice — it explicitly excludes those. Use it when you need correctness auditing, not general code quality or design guidance.

No. It complements testing by constructing concrete failure scenarios a human or AI can reason about, but it does not execute code. Combine it with unit and integration tests for verification coverage.