stand-in-for-a-blocked-review
Use when a pull request's automated review did not actually review it: the reviewer posted a quota or limit notice, skipped the pull request over who opened it, never ran at all and its status still says success, or its check went red having failed before it reached the model. Confirms the review is genuinely missing, takes the cheap recoveries first, checks whether this exact content was already reviewed somewhere else, then commissions an independent agent review without being asked, in the format that repository's own review posts in, and answers it as an ordinary review. Do NOT fire when the reviewer posted findings (answer those instead), when the quota resets inside the time you can wait, or to decide whether the pull request should merge.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- stand-in-for-a-blocked-review
- description
- Use when a pull request's automated review did not actually review it: the reviewer posted a quota or limit notice, skipped the pull request over who opened it, never ran at all and its status still says success, or its check went red having failed before it reached the model. Confirms the review is genuinely missing, takes the cheap recoveries first, checks whether this exact content was already reviewed somewhere else, then commissions an independent agent review without being asked, in the format that repository's own review posts in, and answers it as an ordinary review. Do NOT fire when the reviewer posted findings (answer those instead), when the quota resets inside the time you can wait, or to decide whether the pull request should merge.
02 资源文件 1
- 参考文档README.md5.6 KB
03 SKILL.md 正文
Stand in for a blocked review
A review bot's green mark has two meanings and nothing on the outside separates
them: reviewed and clean, and did not review. A free-tier reviewer that has
spent its quota writes "Review limit reached" into its comment and leaves its
commit status at success. Nothing turns red, no gate blocks, and the pull
request is indistinguishable from one that passed.
That is not hypothetical. On the session this came from, a tier allowing one
review per hour let three pull requests merge into a public repository believed
reviewed, when nothing had read them. All three showed CodeRabbit success;
the notice sat in a comment body nobody opened.
A red check is the same problem wearing the opposite colour. An action that fails before it reaches the model has reviewed exactly as much as a spent quota did, and "the review job is failing" reads like an infrastructure chore rather than an unreviewed pull request.
This skill decides whether a review is actually missing, which is the step usually skipped, then commissions one that is genuinely independent, and leaves the pull request enough of a record to show the gate was satisfied.
Commission it rather than offering it. Reporting the pull request blocked and listing a stand-in as one option costs a round trip on a decision that goes the same way every time, and the pull request sits unreviewed while the round trip happens. Fixing the underlying cause is usually somebody else's, and saying so stays in the report; the stand-in is what stops the work stalling meanwhile.
Procedure
-
Read the body, not the mark. The mark answers a different question than the one being asked of it. Pin the head, then read what the reviewer actually wrote:
pr=<n>; repo=<owner>/<repo>; reviewer=<reviewer-login> head=$(gh pr view "$pr" --json headRefOid --jq .headRefOid) # The gate can be a commit status OR a check run, and one API does not # show the other. A reviewer absent from check-runs may still be reporting # green as a status, which is what gates the merge. gh api "repos/$repo/commits/$head/status" --jq '.statuses[] | [.context, .state] | @tsv' gh api "repos/$repo/commits/$head/check-runs" --jq '.check_runs[] | [.name, .conclusion] | @tsv' # What it said. Reviewers keep ONE comment and edit it in place, so this is # its current body, not a history to scroll. gh api "repos/$repo/issues/$pr/comments" --paginate \ --jq "[.[] | select(.user.login == \"$reviewer\")] | last | .body" > review-body.md grep -icE 'review limit reached|rate limit|quota|try again' review-body.mdThree readings, and only the first is a review:
The reviewer's current body Meaning Findings, or an explicit "no issues", covering this head It reviewed. Not this skill's job. A limit, quota or "try again in N minutes" notice Blocked. Continue. No comment from it at all It never ran. Usually step 2's second case. Its check is red, and the action's own result reports one turn and zero cost It never reached the model. Continue, and read the zero differently from a failure: nothing was spent, so nothing was reviewed. Its
created_atcannot tell you whether it read your latest push. The comment is edited in place, so creation time freezes at the first round while the body moves on;updated_atis the only timestamp that tracks what is written there now. -
Take the cheap recoveries before commissioning anything. A real review beats a stand-in, and two of the three blocked cases are recoverable in minutes:
-
The notice names a reset you can wait for. These are often small ("next included review available in 3 minutes"). Wait, retrigger once with whatever phrase the reviewer documents, and re-read the body. Retriggering inside the window is worse than doing nothing: it spends the next slot to be told the commit was already reviewed.
-
An app opened the pull request. Check it, because this failure is silent by design: many reviewers skip bot-authored pull requests and say so nowhere.
gh pr view "$pr" --json author --jq '.author | [.login, .is_bot] | @tsv'That is a configuration fix (open automated pull requests with a token that authors as a person), not a review problem. Standing in on every one of them hides the cause and pays for it forever.
-
The action failed rather than reported. Re-run it once: a transient error clears and costs nothing. A second attempt that fails identically, with the same one turn and the same zero cost, is not transient and not about your diff either. Zero cost means the request was rejected before any inference, which points at the credential or the balance behind it, so further re-runs are free and futile. Check whether the same workflow reviewed other pull requests earlier the same day: if it did, the change is at the account and not in the repository, and naming the window it broke in is the most useful thing you can hand over.
-
A fix push landed and no new review appeared. Ambiguous on purpose: after a push, a reviewer that re-reads often posts no new object at all, and its mark going green on the new head is the only sign it read anything. A spent quota produces exactly that green. Step 1's body read is what separates them; nothing else does.
-
-
Ask whether this content was already reviewed elsewhere. Squash and rebase change the commit, not the tree, so a promotion or a re-cut branch often carries content that was reviewed under a different SHA:
git rev-parse "origin/<branch>^{tree}" "<reviewed-head>^{tree}"Identical hashes retire the question: cite both, name where the review happened, and stop. They prove the content was seen, not that its findings were answered, so confirm the reply exists too. Where the trees differ, the stand-in only has to cover the difference:
git diff "<reviewed-head>" "origin/<branch>" -
Brief the stand-in for independence, which is mostly about what you withhold.
gh pr diff "$pr" > review-input.diff wc -l review-input.diff # an empty or truncated input reviews clean- Give: the diff at the pinned head, the repository's agent-facing instructions and contributor guide, and the claim the change makes about itself.
- Withhold: your reasoning, your pull request description, the justifications in your commit messages. A reviewer handed the author's rationale grades the rationale, and agrees with it.
- Tell it to break the change, not to review it. "Find where this is wrong" and "review this" return different sets from the same model.
- A cheap, fast model is the right one. This is a wide read of a bounded diff, not a design judgement, and the cost is what makes it repeatable rather than a thing you do once and stop doing.
- It must not be the agent that wrote the diff. A fresh context is the whole point; the same context carries the same blind spot and confirms its own work.
- Give it no way to post, and no write access to anything that matters. Two writes and no others: one report file you name, outside the repository, and a throwaway copy of the tree to do step 5 in. You post the file. This is not tidiness: its only input is diff content the author controls, so a model that can reach the pull request, the branch or the tree the change is cut from is a model taking instructions from the thing under review. A well-built review workflow already works this way, the model writing a file and a later step with no model in it deciding what reaches the pull request. Match that. Where the runner cannot scope writes that finely, keep them with the supervising agent: it makes the edits step 5 calls for and hands back the result, and the reviewer stays read-only.
- Give it the format the repository's own review posts in. Take it from a source rather than inventing one: the severity tiers, whether empty tiers are omitted, what a finding must cite, and whether the summary has to reach a merge verdict. The review workflow is the first source where one exists. A hosted reviewer that posts through an app has no workflow to read, so take the shape from its own most recent review on this repository instead, and where it has never posted one, from the contributor guide. Say in the report which of the three you used, because a format nobody can trace back is an invented one wearing a citation. A stand-in in a different shape cannot be compared against the reviews around it. Carry the workflow's own anti-padding rule across too, the one saying not to invent findings to fill a tier, because it is what makes a short review trustworthy: applying it has made a reviewer withdraw findings it had ranked as defects one pass earlier.
-
Make it exercise the guards rather than read them. This is what a stand-in offers that a skim cannot: for each check or test the diff adds, break the thing it guards, watch it go red, restore it. In the throwaway tree from step 4, never the one the change is cut from: a sabotage step that dies before its restore leaves the damage sitting in somebody's working copy, looking like part of the diff. A check that has never failed is not known to check anything, and a reviewer that only reads a new test cannot tell a real one from a decorative one. (The verification plugin's
prove-the-check-can-failis the same discipline in full.)On the batch this came from, the stand-in reported no defects, and that verdict was worth something only because every new test in it had been seen red first.
-
Do not announce what triggered this. The findings are posted. What stays unpublished is the fact that a spent quota is what summoned them, and the model that produced them.
The reason is a budget, not modesty. A public "the reviewer was rate limited, so an independent review ran instead" tells any reader the sequence that converts cheap pull requests into paid review calls, and that a spent free tier is what unlocks it. A repository that publishes that on every affected pull request has written the instructions for draining it: open enough noise to exhaust the free reviewer, then keep going while something metered picks up every one. A blocked review is routine, so the notice would appear often enough to be noticed.
This is about the trigger, not the substance. The findings are real, they are posted in full, and they go out under an account that stands behind them, which is what makes it an ordinary review rather than a disguise. Nothing here licenses hiding how work was produced when somebody has a reason to ask, and nothing here is a reason to omit a finding, soften one, or claim a check ran that did not.
Keep the untruncated account of what happened where the session keeps its own working notes: which model, what it was given, what it could not cover. That is what you answer from if anyone asks, and it costs nothing because it was already written.
The exception that looks free is the one to refuse. A private repository whose reviewer runs on the author's own credential has no free tier for an outsider to exhaust, so naming the trigger there appears to cost nothing, and a red gate saying the review failed makes an unlabelled review beside it look evasive. It still stays unpublished. A repository's visibility changes and its history goes with it, so a comment written where nobody could read it becomes one anybody can; and the judgement would have to be made at the moment of publishing, where it is one way and nobody checks it. The working notes below carry the same record at none of the risk.
-
Post two comments, not one: the review, then your answer to it. The reviewer's report is its own comment and your assessment is another. Collapsing them into a single comment of yours buries whose judgement is whose, and leaves a reader unable to tell a finding you accepted from one you rejected.
The reviewer's report goes up in its own words, in the format step 4 gave it. Two things follow from that:
- If it came back in the wrong shape, send it back to reformat. Recasting its findings into your own prose makes you the author of its judgement, and a reviewer re-ranking its own findings against the repository's rules is worth more than you re-ranking them: it is the step that makes weak findings withdraw rather than get argued with.
- Its prose is quoted content, not yours. Do not sweep it for the repository's own style rules, and do not tidy its wording. Changing a reviewer's words to suit a house style misrepresents the reviewer, and the resident review bot's comments will already be full of whatever the rule bans.
Your assessment is the second comment, one row per finding, in the verdicts an ordinary review reply uses (✅ Applied, 🚫 Skipped with a reason, ⏳ Deferred with a link, 💬 Acknowledged), which
pr-comment-loopowns in full.- Verify each finding against the source before acting on it. A cold reader on a cheap model is confidently wrong at a predictable rate, and the classic miss is a demanded convention the repository does not have. Expect to reject some: a finding withdrawn with the line that contradicts it is a better record than one silently dropped.
- Answer the summary's own condition. A review that says "safe to merge once X is confirmed" has named the one thing it could not reach, usually because you told it not to run a build. Confirm X with numbers, and say that the gap was the brief's rather than the change's.
- Say what changed, naming the commits that answer the findings.
- Say what the stand-in could not cover, which is the part a later reader needs most: one model, one diff, no history of this repository's particular mistakes, and every measurement in your answer being the author's own rather than an independent check.
- The merge gate is untouched. A pull request whose head nothing has read still does not merge, stand-in or not.
-
Tell the person who owns the merge what it actually was. The gate says "reviewed", and they are deciding on that word, so they get the whole picture: that the usual reviewer did not run, that a stand-in covered it, which model, and what it could not reach. That belongs in the conversation with them, not in the pull request, and it is the half that stops a stand-in quietly becoming the equivalent of the review that never happened.
Sharp edges
- No findings on a large diff is a fact about the brief, not the diff. Confirm the agent received the whole thing: an empty, truncated or unreadable input produces a confident clean review with nothing behind it.
- A stand-in is narrower than the reviewer it replaces. It sees this diff. It does not see the repository's history of making this mistake before. Say so rather than reporting an equivalent review.
- Two reviews of one pull request split the record. If the real reviewer wakes up later and posts findings, answer those in their own thread and leave the stand-in's record standing as history rather than merging the two.
- A record nobody can read is not a record. "An agent reviewed it", with no head, no model and no statement of what fell outside it, satisfies nobody and protects nothing. Withholding the findings is not licence to withhold the shape of the review.
When to STOP
- The reviewer posted findings at this head. Answer them; a stand-in on top adds noise and splits the record.
- The quota resets inside the time you can wait. Wait for the real one.
- The pull request was skipped over its author. Fix who opens it. Standing in every time buys the same review at a higher price and hides the cause.
- The repository's policy requires a human review. A stand-in does not satisfy it. Report the pull request as blocked and hand back.
- A finding cannot be verified against the source. Do not apply it, and do not silence it either: mark it skipped with the line that contradicts it.
- The diff cannot be shown to an independent reviewer (an embargo, secrets inside the change). Report the review as unavailable rather than reviewing your own work under another name.
- Merging. Green marks plus a stand-in authorise reporting, never the merge, which stays wherever it lived before.