Red Team Review

# Adversarial review, one model attacks another's work

**Goal:** stop a model from validating its own output. The author model
produces; a *different* model is told to break it.

## Pattern
1. Author model does the work (your CLI's native model, e.g. Claude).
2. `challenge` forces a critical stance on that work.
3. `clink` hands the same artifact to a different vendor's model for a
   second, independent attack in a fresh context.
4. You read only the surviving objections.

## Template