| name | github-pr-finding-triage |
| description | Use when turning the github-pr-review-comments JSONL index, including CodeRabbit PR review comments captured there, into an action queue with must-fix, verify-first, duplicate, informational, and human-decision buckets. |
GitHub PR Finding Triage
Use after github-pr-review-comments has produced a commentIndexJsonl file.
This is classification of reviewer claims, not proof that the claims are valid. Verify code before changing behavior.
Do not downgrade refactoring, cleanup, or maintainability findings merely because they are inconvenient or behavior still works; apply code-review-release/subskills/thermo-nuclear-code-quality-review/SKILL.md before dismissing them.
Command
tsc --project $HOME/.agents/skills/github/tools/tsconfig.json
node $HOME/.agents/skills/github/tools/dist/pr-finding-triage.js --index /tmp/...comments.jsonl --repo OWNER/REPO --pr PR_NUMBER
Outputs:
- JSON summary on stdout
- private Markdown report in
/tmp
- private full JSON report in
/tmp
Categories
must_fix: reviewer text claims critical/major/security/failing/broken behavior.
verify_first: inline or uncertain findings that need local code verification before action.
probably_duplicate: same normalized path/line/severity/body preview as an earlier finding.
informational: nitpick/trivial/info/style-only comments. Not for structural maintainability findings unless verified against the thermo-nuclear quality bar.
needs_human_decision: API/product/UX/tradeoff or low-confidence comments.
Each finding includes:
derivedSeverity
confidence
classificationBasis
rationale
- original GitHub node ID, body, URL, path/line metadata
Recommended workflow
- Run
pr-review-comments and confirm it is complete (ok: true, no pagination gaps).
- Run this triage tool on the JSONL index.
- Start with
must_fix, then verify_first, then needs_human_decision.
- Do not skip
probably_duplicate without checking duplicateOfId and context.
- Feed the triage JSON into
pr-review-verification before implementing fixes.
Caveats
The classifier is heuristic. It prioritizes actionability and safety over cleverness. A must_fix means “reviewer appears to claim this is important,” not “the code is definitely wrong.”