Use AI-powered review

This guide shows you how to use AI-powered PR review.

What is AI-powered review?

AI-powered review dispatches one or more reviewer agents to analyze a PR. Reviewer agents produce:

Review nodes: structured findings with severity, file/line anchors, and suggested fixes
Review verdict: an aggregate assessment (ship, hold, or block)

Start an AI review

Open a PR in Control Center
Click Start AI review (or use the start_ai_review MCP tool)
Control Center:
- Selects reviewer agents based on the PR’s language and domain
- Creates a review channel for each reviewer
- Dispatches the agents in review mode with the PR diff as context

You can also start a review from a channel by mentioning a reviewer agent.

Review findings

Each reviewer produces findings as review nodes:

Field	Description
Finding kind	bug, security, performance, style, suggestion
Priority	P0 (critical) → P3 (nit)
File and line	Where the finding is located
Original code	The problematic code
Suggested code	The proposed fix
Confidence	How confident the reviewer is
Resolved	Whether the finding has been addressed

Findings appear in the review channel as structured messages.

Review verdict

After analyzing all files, the reviewer produces a verdict:

Verdict	Meaning
Ship	No significant issues found
Hold	Issues that should be addressed before merge
Block	Critical issues that must be fixed

The verdict includes an overall confidence score and per-priority finding counts.

Disagreement detection

If multiple reviewers disagree on the same finding (e.g. one says P0, another says P3), Control Center flags the disagreement. This helps you focus attention where opinions diverge.

Finalize a review

When you’re satisfied with the review:

Click Finalize review
The review summary is posted to GitHub as a PR review comment
Findings are posted as inline comments (optional)

Dispatch reviewers manually

For more control, use the dispatch_reviewers MCP tool:

dispatch_reviewers(
  workspace_id: "...",
  pr_url: "https://github.com/owner/repo/pull/123",
  reviewer_roles: ["security", "reviewer"]
)

This lets you specify exactly which roles should review the PR.