Autonomous agent pipeline — CI fixer, review responder, quality gate

[microsasa]

Problem

Agent-created PRs require manual babysitting: marking ready, fixing CI failures (ruff/pyright), requesting Copilot reviews, addressing review comments, and merging. This slows down the feedback loop and defeats the purpose of automated agents.

Spec

Build a fully autonomous pipeline for agent-created PRs (aw-labeled):

Implementer upgrade

• Run ruff check --fix, ruff format, pyright, and pytest before pushing
• Open regular PRs (not drafts) with auto-merge enabled
• Add aw label to created PRs
• Allow Python package network access for uv sync

CI Fixer agent (new)

• Triggered by CI workflow when check job fails on aw-labeled PRs
• Reads CI logs, identifies failures, pushes fix commit
• Label guard (ci-fix-attempted) prevents infinite loops — max 1 retry

Review Responder agent (new)

• Triggers on pull_request_review events for aw-labeled PRs
• Reads unresolved review comment threads (up to 10)
• Makes requested fixes, replies to threads, resolves them
• Label guard (review-response-attempted) — max 1 attempt

Quality Gate agent (new)

• Triggers on pull_request_review (Copilot approval) for aw-labeled PRs
• Evaluates code quality and blast radius
• LOW/MEDIUM impact → submits approval review → auto-merge fires
• HIGH impact → adds comment flagging for human review, does NOT approve

Repo settings (post-merge)

• Enable auto-merge on repo
• Configure Copilot auto-review ruleset
• Update branch protection to require 2 approving reviews (Copilot + Quality Gate)

End state

Audit/Health Agent → Issue → Implementer PR (lint-clean, auto-merge)
  → CI + Copilot review (parallel)
    → CI fails → CI Fixer (1 retry)
    → Copilot comments → Review Responder (1 attempt)
    → Copilot approves → Quality Gate evaluates
      → ≤ Medium impact → approves → auto-merge
      → High impact → flags for human