Related Issues

Closes: https://github.com/docker/gordon/issues/423

Summary

This PR integrates the data team's PII detection guide (Marlin v2 producer code review) into the PR review agent. It updates the workflow to copy all reference documentation files and makes them available to the review agent, enabling domain-specific guidance for Marlin SDK integration and PII risk detection.

Key Changes

• Reference file handling (review-pr/action.yml): Changed from copying a single posting format file to copying all reference files (.md) from agents/refs/ into /tmp/refs/, making them accessible to the review agent.

• Review agent configuration (review-pr/agents/pr-review.yaml): Added documentation for domain-specific review guides. The agent now checks /tmp/refs/ for additional review guidance files (Marlin PII detection guide, posting format, etc.) after analyzing the diff.

• Evaluation tests (new files):

  • marlin-false-positive-1.json: Validates that the agent correctly skips Marlin guidance for non-Marlin PRs (e.g., generic analytics using Segment).
  • marlin-pii-detection-1.json: Validates that the agent detects Marlin SDK imports and identifies PII risks (unvalidated error messages, command arguments) in analytics event payloads.

✓ Marlin SDK web event from Node server with raw integer enum and non-UUID account_id (run 1) ($0.166632)
  ✓ relevance 6/6
✓ Marlin SDK PageView fired in React render body with wrong timestamp format (run 1) ($0.164642)
  ✓ relevance 5/5
✓ Marlin SDK WebClick with action and elementTag from user input — string_kind and action discriminator violations (run 1) ($0.162822)
  ✓ relevance 5/5
✓ Non-Marlin analytics library — should not trigger Marlin guide (run 1) ($0.114113)
  ✓ relevance 7/7
✓ Marlin SDK Go producer setting ingestor-owned fields, identity mismatch, and UGC truncation (run 1) ($0.183323)
  ✓ relevance 6/6
✓ Marlin SDK AppInvoke analytics with PII risk — err.Error() and command args (run 1) ($0.646032)
  ✓ relevance 9/9
[██████████] 100% (6/6) ✓6 ✗0

✅      Relevance: 38/38 passed (100.0%)