Apply Pr Feedback

/apply-pr-feedback

Apply every actionable review comment on a PR. Scoped to the comments that exist when the skill runs; nothing else. Used as a step inside pr-auto-merge and standalone for any PR you want to clear by hand.

When to use

• The pipeline calls this from pr-auto-merge after the bot idle window closes and at least one actionable comment is still open.
• You want to clear bot review feedback on a PR without running the full pipeline.

Inputs

• <pr-number> (optional). If omitted, the skill resolves the current branch's PR via gh pr status.

Context the skill collects before doing anything

gh pr view <pr> --json number,headRefName,baseRefName,state,mergeable
git branch --show-current
git status --short
git log --oneline -3

If the PR is not open or the working tree is dirty, abort.

Workflow

1. Fetch comments

# Review-level (bot summaries, CHANGES_REQUESTED, approvals)
gh api repos/{owner}/{repo}/pulls/<pr>/reviews   --jq '.[]|{id,user:.user.login,state,body,submitted_at}'

# Inline code comments (the ones tied to file:line)
gh api repos/{owner}/{repo}/pulls/<pr>/comments  --jq '.[]|{id,user:.user.login,path,line,body,in_reply_to_id}'

# Issue-style PR comments (CodeRabbit summaries land here)
gh api repos/{owner}/{repo}/issues/<pr>/comments --jq '.[]|{id,user:.user.login,body,updated_at}'

Capture all three. Filter out comments already resolved or already replied to with a fix commit reference.

2. Classify

For each comment, decide one of:

• Actionable — concrete code change at a specific file:line.
• Question — needs a reply, not a code change.
• Out of scope — refactor / opinion / unrelated to this PR's diff. Do NOT touch.
• Already addressed — earlier commit fixed it; mark resolved.

Build a checklist:

[ ] reviewers/coderabbitai · src/cart/checkout.ts:42 · "guard against null total"
[ ] reviewers/coderabbitai · src/cart/checkout.ts:67 · "extract magic number 4.99"
[?] human/alice           · DESIGN.md · "why not a queue?"   (question, will reply)
[~] coderabbitai          · refactor entire module           (out of scope, skip)

3. Apply fixes (scope-strict)

Per actionable comment:

1. Read the target file first. Always.
2. Group same-file changes through MultiEdit.
3. Make exactly the change requested. No bonus rewrites, no style passes, no related cleanups.
4. Check off the comment in your in-memory checklist.

The cardinal rule: never expand scope. If a reviewer suggests "this loop could be a .map" and you also notice the variable name is bad, leave the variable alone. Renames create review churn and risk regressions outside the original change.

4. Reply to non-actionable items

For each Question and each Out of scope item, post a reply on the original comment thread:

gh api -X POST repos/{owner}/{repo}/pulls/<pr>/comments/<id>/replies \
  -f body="<reply>"

For Out of scope, the reply should be specific: "Out of scope for this PR — tracking in #" or "Out of scope — this would belong in the refactor #142 we deferred." Don't dismiss without a paper trail.

5. Commit and push

git add -A
git commit -m "fix: apply PR review feedback"
git push

No co-author tags. No automated Co-authored-by: lines. The commit should look like a normal fix commit.

6. Verify

gh pr view <pr> --json reviews,statusCheckRollup

The expectation after this skill runs: every comment in the Actionable bucket is addressed in the new commit, every comment in the Question and Out of scope buckets has a reply. If a CHANGES_REQUESTED review is still active, ask the reviewer to re-review explicitly via gh pr comment <pr> --body "Addressed. Ready for re-review." — bots typically re-check on push automatically; humans need the ping.

Return contract

pr-auto-merge (or any caller) decides what to do next based on what this skill reports back, so report it explicitly. After finishing, return these four values:

• pushed_commit — true if at least one fix commit was pushed, else false.
• actionable_remaining — count of actionable comments still open after this run. 0 on a clean pass.
• all_out_of_scope — true if every comment was classified out-of-scope (nothing to fix; replies posted, no commit pushed).
• replies_posted — count of question / out-of-scope replies posted.

The disambiguation that matters: "pushed no commit" is NOT the same as "done." Three distinct outcomes share pushed_commit == false and must not be conflated:

Callers MUST branch on actionable_remaining, never on "did it push." When invoked through the workflow runtime, return this as a schema-validated object; in a prose / Codex session, state these four values explicitly in the hand-off message.

Anti-patterns

• Do NOT touch files outside the reviewer's file:line references. The blast radius of this skill is what the reviewer asked for, nothing more.
• Do NOT skip a comment because it looks trivial. Reply or fix; never silently ignore.
• Do NOT batch this skill across multiple unrelated PRs. One PR per invocation.
• Do NOT use this skill to suppress a linter / typechecker error the reviewer flagged. Fix the underlying issue.

Failure modes

• Bot disagreement (two bots flag opposite things on the same line). Reply to both explaining the chosen direction; do not commit a fix that pleases neither.
• Comment references a file no longer in the diff (e.g., file was renamed mid-review). Reply explaining the rename, link the new path.
• All comments are out of scope. Push a single comment on the PR explaining the scope boundary, do not commit anything.

Composition

• pr-auto-merge invokes this skill between bot-idle-watch and re-running CI. It branches on the Return contract above — specifically actionable_remaining, never on whether a commit was pushed: actionable_remaining == 0 means the feedback is cleared (fixes pushed OR everything was out of scope) and the loop proceeds to merge; > 0 continues only when pushed_commit == true, otherwise escalates.
• Standalone use: pair with gh pr view <pr> --comments to inspect what you're about to clear, then run this skill.

User: $ARGUMENTS