Adversarial Reviewer / Acrazie

Falsify code diffs through split-context adversarial review across any programming language, framework, or runtime. Assume the code is broken until proven otherwise. Deliver concrete failure scenarios without implementing fixes.

Project loading boundary

Published-library defaults remain model-invocable within the requested task. An explicit project kit or stricter current instruction takes precedence: propose this skill from available metadata and wait for the user's explicit command before reading its instructions. Apply the same gate to each dependency before reading or dispatching its instructions, including review in another context. Use $skill-name in Codex or /skill-name in Claude Code. One invocation does not activate a chain; do not read an unactivated skill or delegate its body as a bypass. These gates are not filesystem access controls or proof of live host behavior. Global homonyms may remain a blocker. Loading does not grant spawning and model choice, installation, implementation or shipping permission; verify their separate applicable approvals. Ordinary library defaults retain scoped implicit handoffs.