Design Review Feedback

Use the shared review actions for design feedback:

1. Call view-screen first so the active design, screen, inspector tab, and review queue are current.
2. Call get-review-feedback for the open queue and work on one root thread at a time. Keep the thread id, target screen id, anchor node id, and nearby context together while editing.
3. For element-anchored feedback, prefer the stable data-agent-native-node-id from the anchor. Use the stored percentage point only as a visual fallback when the node cannot be resolved.
4. Read the affected file with get-design-snapshot, make the smallest persisted edit with edit-design or the appropriate design action, then verify the saved result with get-design-snapshot and a screenshot or audit when the change is visual.
5. Resolve the thread with resolve-review-thread only after the edit is persisted and verified, and pass resolutionNote with a one-line description of the persisted change. If the user needs to decide something, reply with reply-review-comment and set resolutionTarget: "human"; do not silently resolve it.
6. After applying a thread, call consume-review-feedback for the applied agent-targeted root comment so the queue does not repeatedly resurface the same work.

Reviewers can leave a human-only comment or choose Send to agent when composing. Editors can selectively route an existing open thread with send-review-thread-to-agent; that action only changes the selected thread. When a thread is sent from the Design UI, the agent chat handoff includes its thread id and instructs the agent not to apply other queued feedback.

Use create-review-comment for agent-authored follow-up notes when a durable record is more useful than a chat-only explanation. Keep review comments scoped to the design resource and the target screen; never copy large HTML, screenshots, or provider payloads into a comment.