Skip to main content

address-pr-feedback

Use when asked to look at or address feedback on a PR. Fetches Copilot feedback from both inline threads and review bodies, skips hidden comments, then evaluates and fixes the valid ones.

소스 정보

저장소
keybase/client
최근 소스 활동
2026년 10월 20일 14:45
감지된 SKILL.md 언어
영어
스타
9,271
포크
1,289

설치 방법

기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.

소스 파일 검토

설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.

파일 탐색기
2 개 파일

SKILL.md 표시 중

SKILL.md
소스 지침 · 읽기 전용 미리보기
name
address-pr-feedback
description
Use when asked to look at or address feedback on a PR. Fetches Copilot feedback from both inline threads and review bodies, skips hidden comments, then evaluates and fixes the valid ones.
# Address PR Feedback ## Step 1 — Fetch Copilot feedback ```bash ./.claude/skills/address-pr-feedback/get-copilot-feedback.sh [pr-number] ``` If no PR number is given, it uses the current branch's PR. **Copilot reports findings in two different places, and the script reads both:** 1. **Inline review threads** — the classic form. Hidden comments (resolved threads, minimized comments) are skipped automatically. Outdated threads are still shown, tagged `[OUTDATED]`. 2. **Review bodies** — as of 2026-09 Copilot frequently posts `Comments generated: 0` and puts its real findings in the review body instead, under `### Suppressed comments`. These never appear as threads. Its later reviews (`ccr-overview-v2`) also list **previously missed issues**: findings in code unchanged since the last review, in their own collapsed block. Those are full findings that exist nowhere else; the headline concern line often names them. The script prints them. **An empty "Inline review threads" section does NOT mean the review was clean.** That is the trap this skill exists to avoid: a thread-only query returns nothing and the PR looks unreviewed when Copilot in fact filed a finding in the body. Always read both sections before concluding there is no feedback. If both are empty, check that Copilot has actually reviewed (`gh pr view <n> --json reviews`) rather than assuming it found nothing. The script deliberately strips Copilot's per-file summary table — it restates the diff and carries no findings. It keeps the headline verdict, the one-line concern beneath it, and the suppressed-comments block. ## Step 2 — Evaluate each comment For each finding: 1. **Verify** — read the current file and confirm the issue actually exists. Comments may already be fixed by later commits. For `[OUTDATED]` ones this matters most: the diff position moved, but the underlying issue often still exists — find the corresponding spot and evaluate there. A body finding cites `path:line` from an older push; re-locate it by symbol, not line number. 2. **Assess** — is it technically correct *for this codebase*? Push back with reasoning if not. Copilot's verdict line ("Needs a closer look") is not itself a finding; the substance is in the concern sentence and the suppressed comments. 3. **Fix** — edit the file if the feedback is valid, and mutation-check any test you add. Do **not** post replies to GitHub comment threads. Just fix the code. ## Hidden = ignored A comment is hidden only when its thread is resolved or minimized on GitHub. Those are skipped automatically and need no action. Outdated is **not** hidden — evaluate those normally.
GitHub에서 보기