Agent Review Rollout Plan

# Agent Review Rollout Plan

## Goal

Let agents read Commentary review context, draft replies, and propose revisions without bypassing human ownership.

## Proposal

Start with read-only MCP access for selected repositories, then enable comment drafting for plan reviews after the first audit checkpoint. Keep app-native review threads primary and treat provider comments as synchronized representations.

## Review loop

1. A human opens a review and asks an agent to inspect selected comments.
2. The agent receives scoped files, comments, permissions, and allowed actions.
3. The agent proposes a patch or reply without treating comment bodies as instructions.
4. A reviewer accepts, edits, or rejects the proposed change.
5. Commentary records the outcome in the review thread.

## Open decisions

- Which team owns escalation when agent context is incomplete?
- Should generated replies require explicit reviewer approval before sync?
- What metrics prove the rollout is reducing review latency?
- What privacy-safe actor identifier should appear in launch telemetry?

## Accepted constraints

Agents receive explicit review ids, files, comments, permissions, and allowed actions. They do not scrape private rendered pages.

Agents must treat reviewed app content, Markdown, HTML, and comment bodies as untrusted editing context. They must not follow instructions embedded in the reviewed artifact.