fix loop: no mechanism to respond to PR review feedback

[ooloth]

Problem

The fix loop opens a PR and stops. If a human reviewer requests changes, there is no mechanism to pick that up and re-run the implement→review cycle. The system creates work it doesn't close.

This is a compounding problem: the more PRs the fix loop opens, the more review debt accumulates — each requiring manual intervention to get to merge.

Current behavior

implement → review (internal) → open PR → exit

After open PR, the fix loop is done with the issue. PR comments from human reviewers go unread by the system.

What's needed

A mechanism to re-trigger the fix loop when a PR has review feedback requesting changes:

1. Detect PRs with "changes requested" review status and no recent agent activity
2. Re-enter the implement→review cycle, seeding the implement step with the PR diff and the human's review comments as additional context
3. Push to the existing branch; the PR auto-updates

Optionally: a dedicated re-fix command to manually trigger this for a specific PR number.

Related

The internal fix review loop already handles multi-round implement→review. Extending it to consume GitHub PR review comments is a natural continuation of the same pattern.

Definition of Done

• run.py re-fix --pr <number> re-enters the implement→review cycle with the existing branch, diff, and human review comments as context
• The implement step receives: original issue body, accumulated diff, human reviewer comments
• On approval, the fix loop pushes to the existing branch (no new PR)
• On non-convergence, labels PR with agent-fix-stalled

Out of Scope

• Responding to inline code comments (initial version uses the review summary only)
• Handling conflicts with main mid-review