feat(skill): add triage-pr-review Claude Code skill

[dsilvam]

Description

Add a new Claude Code skill (.claude/commands/review-pr.md) that fetches all GitHub PR review feedback from every available source and surfaces only Critical 🔴 and High 🟠 findings with file:line pointers ready to act on.

Context

When reviewing a PR today, review feedback was missed because it was posted as a top-level issue comment by the backend-review bot — a different GitHub endpoint than the inline code review threads. The skill solves this by querying all three GitHub sources in parallel and deduplicating results.

A second pain point: reviewers sometimes flag code that the PR didn't touch (pre-existing issues) or raise cosmetic concerns that aren't worth a fix cycle. The skill addresses this by reading the PR diff and description and evaluating each finding for relevance and necessity before presenting it.

What the skill does

• Accepts a PR number or URL as input
• Fetches from all three GitHub review sources:
  • pulls/{n}/comments — inline reviewer thread comments
  • pulls/{n}/reviews — APPROVE / REQUEST_CHANGES summaries
  • issues/{n}/comments — top-level PR comments (where review bots post structured findings)
• Reads the PR description and diff to understand what the PR actually changed
• Automatically switches to the PR branch (git stash → git fetch → git checkout) so file reads reflect actual PR code
• Classifies findings by severity using explicit emoji markers (🔴, 🟠, 🟡) from bot comments and keyword inference for plain human comments
• For each Critical/High finding, checks whether it is already fixed by scanning recent commit messages and reading the affected file
• Evaluates relevance and necessity: was the flagged code touched by this PR? is the fix aligned with the PR's goal? is it just cosmetic?
• Assigns a recommendation label per finding: Must fix / Nice to have / Out of scope — skip
• Outputs a structured action list: ❌ Needs Action / ✅ Already Fixed / Skipped count

Acceptance Criteria

• Skill file exists at .claude/commands/review-pr.md
• Skill fetches from all three GitHub endpoints (inline comments, reviews, issue comments)
• Skill fetches the PR diff and description to understand scope of changes
• Skill switches to the PR branch before reading files
• Only Critical 🔴 and High 🟠 findings are shown by default; Medium/Low are counted but skipped
• Already-fixed findings are detected by reading the current branch files and matching recent commits
• Each finding includes a Recommendation label: Must fix / Nice to have / Out of scope — skip
• Each finding includes a Relevance note explaining why that label was assigned (e.g. "line is in a hunk added by this PR" vs. "file was not modified by this PR")
• Output follows the structured format: Needs Action / Already Fixed / Skipped sections
• Skill is invokable via /review-pr <pr-number|pr-url>

Additional Context

• Skill format follows the existing .claude/commands/diagnose.md pattern used in this repo
• Severity inference for plain human comments uses security/data-loss/correctness keywords
• The --all flag (future) can be added to also show Medium/Low findings
• "Out of scope — skip" findings should be logged as follow-up issues rather than blocking the current PR

[github-actions]

PRs linked to this issue

• feat(skill): add triage-pr-review Claude Code skill (#35828) by @dsilvam