Adversarial review

Adversaries try to break a change or a design. A finding survives only when a run proves it. Report the findings. Do not apply fixes.

This skill never changes the user's working tree. Adversaries write repros only inside their own isolated copy. They must never edit the user's checkout, submit a review, post a comment, change a pull request, create an issue, or message anyone. Subagents may communicate with each other, but no agent may communicate with a person. Repeat that sentence verbatim in every child prompt. Keep all findings and drafts local.

code-review remains authoritative for standards and specification review. Use this skill as an additional pass that hunts for failures a run can prove. It must not weaken that skill's scope, evidence, or authorization rules.

1. Frame

The scope is what the user names. Otherwise it is the branch diff from its merge base with the base branch, through the working tree. Record the exact commit and whether uncommitted changes are included.

State the intent. Use the in-your-own-words skill when installed. Otherwise write one paragraph from the user's message, the commits, and any local spec or PR description already in scope. Ask the user when the intent is unclear.

2. Fan out

Use arena when installed: follow its model preference and seat rules and its Generate phase, with the changes below. Otherwise apply the same rules inline: