fix-pr-review-loop
Use when the user asks to fix a PR review and drive it to approval autonomously — "fix the PR review and loop until approved", "fix-pr-review-loop", "keep addressing review comments until this PR is approved". Takes an optional PR number/URL (defaults to the current branch's PR). Repeatedly calls fix-pr-review, waits for the bot's re-review (@claude by default, @codex when selected), and repeats. Stops on a bare LGTM, on the first LGTM past 5 cycles, or to escalate when four or more cycles keep producing blocking findings in code the loop itself added.
richkuo · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add richkuo/rk-skills --skill fix-pr-review-loop技能说明
name: fix-pr-review-loop description: Use when the user asks to fix a PR review and drive it to approval autonomously — "fix the PR review and loop until approved", "fix-pr-review-loop", "keep addressing review comments until this PR is approved". Takes an optional PR number/URL (defaults to the current branch's PR). Repeatedly calls fix-pr-review, waits for the bot's re-review (@claude by default, @codex when selected), and repeats. Stops on a bare LGTM, on the first LGTM past 5 cycles, or to escalate when four or more cycles keep producing blocking findings in code the loop itself added.
fix-pr-review-loop
Drive an open PR from "has review feedback" to "reviewed to convergence": resolve the latest review with fix-pr-review, wait for the bot's re-review, repeat. This skill owns the stop rules in step 3 and the terminal-state table in step 5. work-on-issue-loop runs steps 2 through 4 after it opens a PR and reuses step 5's table with its own deltas, so an edit to that table needs a matching edit in work-on-issue-loop step 4.
Input
Nothing (the current branch's PR via gh pr view), or #<N> / <N> / URL / owner/repo#N.
Steps
1. Resolve the PR and establish the starting state
gh pr view <N|--> --json number,headRefName,headRepositoryOwner,baseRefName,url,state,isDraft. A merged or closed PR: stop and report. Otherwise fetch existing feedback with fix-pr-review step 1's three-channel query (reviews, issue comments, inline threads).
- Unaddressed feedback present (newer than any prior disposition comment, or an unresolved inline thread): it is review 1. Set
review_count = 1, note its timestamp, go to step 3. Never post a redundant trigger. - No feedback yet (fresh PR, or only your own disposition or trigger comments): post the first-review trigger as its own one-line comment,
gh pr comment <N> --body "<trigger>". Record its timestamp, setreview_count = 1, go to step 2.
First-review trigger. validate-issue step 6 owns the first-review table with its Claude and Codex columns; this file states no boundary of its own. Read the score in this order and stop at the first hit: a stamped PR review: line in the linked issue's Execution block (it overrides the band), the [C<score>, …] bracket in the PR title, the [C<score>] prefix of the issue the PR closes. A missing score routes to the top row. Fable runs at high unless the user asks for xhigh or stamps it.
Map a stamped model before posting it. claude.yml resolves only opus, sonnet and fable: a stamped sonnet or haiku posts @claude sonnet review, opus posts @claude opus review effort:high, fable posts @claude fable review effort:high, each keeping a stamped effort:<tier> in place of high. An unadmitted shorthand becomes the Action's route keyword and can select the write-capable fix-pr job, which pushes commits. On Codex, sonnet/haiku map to @codex luna review and opus/fable to the bare @codex review; never carry a @claude shorthand across to @codex. Blocking re-reviews step down one rung per cycle, keyed to the reviewer that actually ran cycle 1; skills/fix-pr-review/rereview-routing.md owns that ladder and the Codex rule.
Bot selection. @claude by default. @codex only when Codex was explicitly selected: the user said so, a caller argument named it (reviewBot: codex), or this run started from an @codex comment. A codex.yml merely existing does not select Codex. Never switch bots mid-cycle.
Preflight. Confirm a reviewer answers the selected bot: gh api repos/{owner}/{repo}/contents/.github/workflows --jq '.[].name' must list claude.yml (or the Claude GitHub App is installed) or codex.yml. With Codex, also check the OPENAI_API_KEY secret and, for the write routes, CODEX_APP_ID, CODEX_APP_PRIVATE_KEY and the CODEX_BOT_LOGIN variable. If none exists, stop and point the user at templates/claude-workflow/workflows/claude.yml (plus CLAUDE_CODE_OAUTH_TOKEN, the Claude GitHub App, and a runs-on they have) or templates/codex-review.yml.
2. Wait for the review to land
Poll for a review or issue comment posted after the last trigger timestamp. The @claude bot edits its placeholder comment in place; createdAt still follows your trigger. A Codex review is a plain github-actions[bot] comment with the verdict first.
until gh pr view <N> --json comments,reviews --jq '
([.comments[] | select(.createdAt > "<trigger_ts>")] |
any(.body | test("(^|\\n)(LGTM|Needs Updates)"))) or
([.reviews[] | select(.submittedAt > "<trigger_ts>")] | length > 0)
' | grep -q true; do sleep 60; done
Match with (^|\\n): jq's m flag is dot-matches-newline and a bare ^ never reaches a verdict under the bot's header. Pipe through grep -q true: gh --jq exits 0 on false. Run the loop in the background (Monitor tool) after one inline sanity check that prints true against a review already present. Cap the wait at roughly 30 minutes, then stop and report that the bot did not respond.
3. Check the review against the stop conditions
Classify the latest review as fix-pr-review steps 1 and 3 do: verdict (LGTM / Needs Updates) and which finding sections are present (Needs Fixing, Requires Human Review, Recommended Optional, Create Follow-up Issue). A **Verification limitation:** line is not a finding. Evaluate in this order:
- Merge conflict.
gh pr view <N> --json mergeable,mergeStateStatus. ACONFLICTING/DIRTYPR is never terminal: go to step 4, even on a bare LGTM. - Clean pass, stop.
LGTMwith no finding sections at all, at anyreview_count; nothing left to fix. Go to step 5. - Past the cap, stop.
review_count > 5and the verdict isLGTM, even withRecommended OptionalorCreate Follow-up Issueitems listed. Go to step 5. - Diverging, stop.
pr_cycle_count >= 4, the verdict isNeeds Updates, and every blocking finding sits in code