Better retries when a sub-agent fails

[derekmisler]

Summary

When a sub-agent fails (e.g., due to API overload), the root agent can recover and still post a review — but the pipeline previously treated any non-zero exit code as a hard failure, posting a redundant "Review Failed" comment and potentially retrying the entire pipeline, causing duplicate reviews. This PR disables pipeline-level retries and adds smarter exit-code handling to distinguish a true failure from a partial success.

Changes

• review-pr/action.yml — disable pipeline retries: Sets max-retries: "0" on the review step with an explanatory comment, since the root agent already recovers internally when sub-agents fail and retrying the pipeline produces duplicate reviews.
• review-pr/action.yml — smarter failure detection: Exposes verbose-log-file from the review step output and, on a non-zero exit code, checks whether a pull request review was actually posted (by grepping for pullrequestreview-[0-9]+ in the log). If a review was found, the pipeline reports ⚠️ Review completed with warnings instead of ❌ Review failed and skips posting the fallback failure comment.

How to Test

• Simulate a sub-agent failure (e.g., trigger an API overload mid-review) and confirm the pipeline exits with a warning status rather than posting a "Review Failed" comment when the root agent successfully posted a review.
• Confirm that a genuine total failure (no review posted, non-zero exit) still posts the ❌ PR Review Failed comment to the PR as before.
• Verify no duplicate reviews appear on a PR when a sub-agent fails and the root agent recovers.

[derekmisler]

/describe

[docker-agent]

✅ PR description has been generated and updated!

[docker-agent]

Review Summary

Assessment: 🟢 APPROVE

This PR adds intelligent error recovery to handle cases where a sub-agent fails but the root agent successfully posts a review anyway. The changes are well-reasoned:

Key improvements:

• ✅ Disables retries (max-retries: "0") at the pipeline level since the review agent has internal recovery logic
• ✅ Adds detection for reviews posted despite non-zero exit codes by checking for pullrequestreview-[0-9]+ in verbose logs
• ✅ Treats this scenario as a "partial success" with a warning rather than a full failure
• ✅ Only posts error comments when no review was actually posted

Review notes:

• The grep pattern pullrequestreview-[0-9]+ is specific enough to avoid false positives in practice — it matches GitHub's review ID format which only appears when the API successfully returns a review object
• The error handling logic correctly distinguishes between "failed with recovery" and "failed without recovery" scenarios

No bugs found in the changed code. This improves the robustness of the review workflow.