code-fix-pr

Evaluate and address PR review comments.

Prerequisites

Reference these skills first:

• Read jujutsu skill to understand change context.
• Read github-cli skill to fetch PR comments.

Process

Step 1 - Gather PR Data

• Get PR number: ask the user or infer from the current branch.
• Determine repository: gh repo view or jj git remote list.
• Fetch PR comments: gh api repos/owner/repo/pulls/<number>/comments -X GET --paginate.
• Fetch reviews: gh api repos/owner/repo/pulls/<number>/reviews -X GET --paginate.
• Get PR diff: gh pr diff <number> -R owner/repo.
• Check PR status: gh pr checks <number> -R owner/repo.
• Fetch failure logs: gh run view <run-id> -R owner/repo --log-failed.
• Capture CI failure context: summarize failed checks, error messages, and stack traces for later analysis.
• View working copy changes: jj diff -s.

Important: Always use -X GET to be explicit about read-only access.

Step 2 - Analyze Comments and CI

Apply a reviewer lens to review comments across all relevant perspectives:

• Ground findings in file paths and line numbers.
• Prioritize the requested lens.
• Distinguish confirmed findings from speculative risks.
• Explain why each issue matters.

Review across these perspectives:

• Correctness and quality: Classify each comment as already-addressed, valid-fix-needed, invalid, or needs-discussion.
• Security: Identify valid security concerns vs false positives; assess severity.
• Simplicity and convention: Prefer the simplest change satisfying valid feedback; distinguish valid simplifications from over-engineering.
• CI correlation: Check whether each comment relates to an observed CI failure and whether addressing it would resolve the failure; incorporate available CI failure context.

Apply a researcher lens to validate APIs and research best practices for the fixes suggested in the PR comments:

• Prefer official documentation over blog posts.

• Cite sources with URLs.

• Separate confirmed facts from plausible interpretations.

• Note version requirements.

• Lead with the single most actionable recommendation.

• If CI failure logs are included, apply a researcher lens to research the specific error messages, stack traces, or test failures to identify root causes and known solutions:

  • Prefer official documentation over blog posts; cite sources with URLs.
  • Separate confirmed facts from plausible interpretations; note version requirements.
  • Lead with the single most actionable recommendation.

• If checks failed, apply a reviewer lens to analyze the GitHub Actions CI failure logs:

  • Ground findings in file paths and line numbers.
  • Distinguish confirmed findings from speculative risks.
  • Explain why each issue matters.

For each failure:

• Failure classification: test flake, code regression, env/infra issue, dependency problem, or timeout.
• Root cause: most likely root cause with supporting evidence from logs.
• Diff relationship: whether failures occur in code touched by this PR's diff or unrelated code.
• Fix suggestion: minimal fixes for code-regression failures; mitigation steps (retry, env fixes) for flakes/infra.

Step 3 - Adjudicate Disputes

• For disputes or conflicting recommendations, apply an oracle lens to decide which feedback should be fixed, discussed, or rejected:
  • Identify the decision.
  • State assumptions and constraints.
  • Pick the smallest safe path that preserves future options.
  • Note what evidence would change the recommendation.
  • Explain the reasoning clearly.

Step 4 - Synthesize Findings

Synthesize into a unified report:

1. PR Summary: PR number, title, state; total comments analyzed.
2. Comment Classification:
   • Already Addressed: list with resolution evidence.
   • Valid - Should Fix: list with severity and location.
   • Not Valid - Won't Fix: list with reasoning.
   • Needs Discussion: list with questions.
3. Fix Plan: numbered steps with file:line targets; best-practice references; estimated effort.
4. CI Failure Analysis (only if checks failed):
   • Failing Checks: each failed check with run ID and error summary.
   • Failure Classification: test flake / code regression / env-infra / dependency / timeout.
   • Root Causes: per-failure analysis with log evidence.
   • PR Relationship: in PR-touched code or unrelated.
   • Recommendations: fix suggestions for regressions; mitigation steps for flakes/infra.
5. Next Steps: immediate actions; questions for reviewer.