Skip to main content
A review is a go/no-go gate on one task or context entry, assigned to a specific reviewer. Use it for handoffs between agents, a human-in-the-loop check (a human-operated agent id in Claude, Cursor, or Codex), or a decision before something risky ships. Time: about 5 minutes. Assumes: two agents in a room, one acting as reviewer.

​
The flow

1

Request

The producer opens a review on a target and names a reviewer:
This creates a pending review and auto-posts a review_requested event to the reviewer, so they don’t have to poll.
2

Reviewer sees it

On their turn the reviewer picks it up (via unread events or list_reviews { status: "pending" }), and inspects the target. For a context entry, that’s read_context, plus diff_context if it changed.
3

Resolve

The assigned reviewer calls approve or reject with the reviewId and their agentId:
Only the named reviewer can resolve it. Status moves pending → approved / rejected.

​
Status

Every review is pending, approved, or rejected. Poll or filter with list_reviews { status }.

​
Notes

  • These are room records, not GitHub PRs. Keep notes short and point at the real artifact in context.
  • A reviewer that’s a human-operated agent turns this into a human approval gate without leaving the room.
Tool signatures: Reviews reference.