Multi-Agent Review

Orchestrate the review; do not perform every specialist review yourself. Produce a single result that preserves useful disagreement and makes its evidence and confidence clear.

Frame the review

Establish the following before dispatching agents. Use the user's wording when it provides the answer, and state any material assumption.

• Context: the change's purpose, relevant constraints, and any policy or acceptance criteria.
• Scope: an exact boundary: files and changed lines, a component, a document section, or another inspectable target. Exclude adjacent work unless it is needed to demonstrate an impact within scope.
• Review dimensions: the quality or risk lenses to apply, such as correctness, security, performance, maintainability, accessibility, or operational safety.
• Review policy: the allowed severity values and the standard for require_action. When absent, use critical, important, and suggestion; reserve required action for a concrete change needed to prevent a material problem.

If the requested scope or dimensions cannot be determined safely, ask a focused question rather than silently widening the review.

Dispatch independent reviewers

Assign each reviewer an explicit subset of the scope and one or more review dimensions. Use independent agents in parallel when available. Give every reviewer the same context, its own scope and dimensions, the review policy, and the reviewer-result schema in references/reviewer-result-schema.json.