Adversarial Review

Decide whether the proposed work can responsibly proceed. Find material gaps; do not manufacture findings or redesign the product.

Independence And Context

Use a fresh reviewer for every round. For full work, run a panel: one reviewer per entry in the plan review role, on different providers when the catalog has them, each seeded identically. Merge their findings, dedupe by root cause, and note where reviewers agreed, since cross-model agreement is high-signal. fast work uses the first entry. Pass only the brief, plan, episode state, relevant evidence pointers, repository instructions, and the review question. Do not pass the orchestrator's suspected findings or the full conversation.

Resolve the reviewer from the plan review role per the kickoff skill's references/config.md, and map the spawn through its references/harness.md. A missing role runs as a fresh child on the parent's model. Never stop to ask the user to configure a reviewer.

The review is read-only. The worker must not edit files, implement fixes, or spawn more workers. When the harness cannot spawn a child, review inline, label the result non-independent, and record the degraded gate in the episode state.

When Required

• Always for full work.
• For fast work.
• For tiny work only when route assessment finds material ambiguity, blast radius, consequence, irreversibility, novelty, or coordination—or repository policy requires it.

Assess risk in the work's own domain. Include user journeys, interaction and content states, visual-system consistency, accessibility, trust, compatibility, performance, data and security, delivery, and operations when relevant; do not default to backend concerns.