Interrogate

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.

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. From the user's message, the commits, and the PR description, restate in your own words what you think the user's goals are and what problem they are trying to solve. This restatement is the intent. Ask the user when the intent is unclear.

Fan out

By default there are two adversaries, on two model families. Start each subagent with a fresh context window.

• Adversary A is a subagent on your own model. Set its effort one step below your own. If you do not know your effort, do not set it.
• Adversary B runs on a different model family, through that family's CLI, with its prompt on stdin. If you are a Claude model, run codex exec -s workspace-write -C <worktree> -o <output-file> -. Otherwise, run claude -p --permission-mode auto in its worktree, and send its output to a file. The CLI uses its own configured model. If the CLI is missing or fails, run Adversary B as a second subagent on your own model, and say in the output that the review has one model family.
• If the user names more models, add one adversary for each, labeled Adversary C, D, and so on.

Give each adversary its own git worktree, so no two adversaries write to the same file. Copy the scope's uncommitted changes into each worktree. Start all adversaries at the same time, in the background. Give each adversary the scope, the intent, the path of its worktree, and this brief:

Break this change. Prove each finding with a run against the real code, such as a failing test, a repro script, or command output that shows the wrong behavior. For a design or plan, run a prototype or probe the current tree. Keep repros in your own worktree. Report each finding with a title, the exact command, the output it printed, and the output the intent requires. Leave out anything you did not run. That covers style, naming, structure, and risks that need a future change to matter. No findings is a valid result.