Slice 10 — Reviewer Agent posts a review on every newly opened PR

[safayavatsal]

What to build

The Reviewer Agent — every PR, human-authored or Agent-authored, gets an automated review by default.

This is HITL because review tone, format, and what counts as "worth flagging" are product-defining. A Reviewer Agent that's noisy or annoying actively damages the wedge.

• agent-orchestrator consumes pr.opened events. For each event, the Reviewer Agent runs against the PR diff and posts comments back to platform-api.
• ReviewerAgent is implemented as a deep module: (diff, context) → []ReviewComment. Pure over its inputs once a ModelClient is injected.
• Comments support inline anchoring ({file, line, side}) when the model can identify a precise location, otherwise a top-level comment.
• All Reviewer Agent comments are clearly attributed (author = forge-reviewer, special UI badge).
• Review focus at MVP: correctness regressions, obvious security issues (e.g., shell injection, leaking secrets), and ambiguity that suggests the change is incomplete. Not style nitpicks.

Decisions to nail in this slice

• Review prompt strategy and what the Reviewer is told not to comment on.
• Comment volume cap (e.g., max N comments per PR) to avoid spam.
• Behavior when the diff is too large to fit in context.

Acceptance criteria

• Opening a PR (in slices 5 or 7's flow) triggers a review within bounded time.
• The review either posts inline comments anchored to specific lines or a top-level comment with a summary.
• All Reviewer Agent comments are visually distinct from human comments in the web UI.
• On a trivial no-op PR, the review correctly says "looks fine" rather than inventing issues.
• On a PR that introduces a deliberate bug from the eval set, the review flags the bug.
• ReviewerAgent has tests against fixed diffs with a recorded-response ModelClient, asserting comment shape (not exact wording).

Blocked by

• Slice 5 — User can open a PR from a branch and merge it from the web UI #5