review

The task isn't done until it passes a reviewer panel. Run the gate before declaring done.

Panel

$ARGUMENTS = reviewer count, default 2:

• 1 → fork reviewer only.
• 2 → 1 independent + 1 fork.
• N → N−1 independents with distinct lenses (correctness, functionality, simplification, security, tests) + 1 fork.

Reviewer types:

• Fork — an agent forked from this session, carrying the full conversation context. Always on the session model.
• Independent — fresh context: give it only the task description and the branch diff vs merge-base.

Cross-model: If you are Claude, and have access to Open AI Model (like in Cursor), then prefer cross-model reviewers as long as they're not a weaker model. For example, Opus 4.8 <-> GPT 5.5