fix
Processes a PR's review comments (Copilot or human), classifies them, applies the fixes that belong, replies to each thread, and pushes. Use whenever the user says "address the review", "fix what Copi
Файлы1
Открыть в GitHubSKILL.mdskills/fix/SKILL.md
GitHubAddress review comments
Turn a PR's comments into commits or into justified replies. Two automatic rounds max; the third is for a human.
Resolve the repo once up front; every gh call below uses it:
gh repo view --json nameWithOwner -q .nameWithOwner # -> {{owner}}/{{repo}}
1. Determine the round
gh pr view {{n}} --json labels,title,url
- No
ai-round-*label -> this is round 1. - Has
ai-round-1-> this is round 2. - Has
ai-round-2-> STOP. Don't touch code. Go straight to step 5 (escalate).
The needs-human label does not stop this skill: it means a human does the merge, not
that comments go unaddressed. Both Copilot rounds run anyway. The only things that stop the
rounds are ai-round-2, or a comment landing in the "Human" group (step 3).
2. Pull the unresolved comments
gh pr view {{n}} --comments
gh api repos/{{owner}}/{{repo}}/pulls/{{n}}/comments --paginate
Discard threads already resolved and comments already answered in earlier rounds. Copilot tends to repeat comments that were already dismissed: if a previous thread has a justified rejection, don't reopen it.
Make sure you're on the PR branch and up to date:
git switch {{pr-branch}} && git pull
3. Classify each comment
Before touching anything, classify all comments into three groups and show the user the classification:
| Group | Criterion | Action |
|---|---|---|
| Accept | Real bug, concrete risk, or a CLAUDE.md convention broken |
Fix it |
| Reject | False positive, generic style suggestion that contradicts CLAUDE.md, or out-of-scope refactor |
Reply to the thread with the reason, don't touch code |
| Human | Touches something on the "Changes that require human review" list, or implies a design decision the plan didn't make | Don't touch. Goes to step 5 |
Core criterion: a comment is not an order. Rejecting with an argument is a valid response and preferable to a mechanical fix that dirties the diff.
4. Apply and reply
For the "Accept" group:
- Fix it. One commit per topic, not one "fix review comments" commit with everything in it.
- Run lint and tests after each fix.
- Reply to the thread saying what changed and in which commit.
For the "Reject" group:
- Reply to the thread with the concrete reason. No bare "doesn't apply".
gh api repos/{{owner}}/{{repo}}/pulls/comments/{{comment_id}}/replies -f body="{{reply}}"
Then:
git push
gh pr edit {{n}} --add-label "ai-round-{{N}}" --remove-label "ai-round-{{N-1}}"
5. Escalate what doesn't resolve itself
If there are "Human" group comments, or this was round 3:
gh pr edit {{n}} --add-label "needs-human" --add-assignee "@me"
And leave one comment on the PR with the status:
## Pending human review
- [ ] {{comment/topic}} — {{why it needs a human decision}}
Automatic rounds completed: {{N}}/2
Auto-resolved this round: {{count}} fixes, {{count}} justified rejections.
6. Close
Report to the user in three lines: how many were fixed, how many rejected and why, and what's left for them. If nothing is pending and the PR is green, say so explicitly — that's the signal they can merge.
Never merge.
Convention: {{like-this}} marks a value to replace. <> is avoided because it breaks skill parsing.