Post Review

Post review

A code review leaves comments on the code. Findings delivered only as a review body — or worse, only in chat — force the author to map prose back to lines themselves. Every finding that concerns a specific place in the diff becomes an inline comment on that line; the review body carries only what has no single anchor: the overall verdict, why the review blocks (or doesn't), and what was verified.

Posting the review

gh pr review cannot create inline comments — use the reviews API with a JSON payload:

gh api repos/<owner>/<repo>/pulls/<number>/reviews --input review.json

{
  "event": "REQUEST_CHANGES",
  "body": "Summary: what checks out, why this is blocked, pointer to the inline comments.",
  "comments": [
    { "path": "app/example.py", "line": 44, "side": "RIGHT", "body": "**Blocking.** ..." }
  ]
}

• event is REQUEST_CHANGES, APPROVE, or COMMENT — pick what the user asked for; never approve or request changes on your own initiative.
• line is a line number in the file as shown in the diff (side: "RIGHT" for the new version, "LEFT" for deleted lines). It must fall inside a diff hunk — context lines count. Compute it from the hunk headers in gh pr diff; an off-hunk line rejects the whole review.
• Build the payload in a file and pass --input — inline -f flags cannot express the comments array.
• When follow-up PRs will implement the findings, say so in the review body — the author must learn it in the same read as the findings. Each inline comment's "unblock" path reads as a work request, and the PRs only land minutes to hours later; an author who reads the review in that window starts implementing fixes you are about to hand them.

Structuring the findings

• Mark each blocking comment Blocking and state what unblocks it. When the finding is an undocumented behavior change, the unblock path includes "if intentional, document it in the PR description" — the author decides intent, the review only demands the decision be visible.
• Label everything else explicitly as non-blocking so the author can triage at a glance.
• The summary opens with what is good and what was verified before listing why the review blocks. Credit claims you checked and found true — a review that only lists faults reads as unverified suspicion.
• Verify before asserting: check the PR's claims against the base branch (call sites, conventions, test counts) and anchor each finding to evidence. A finding you could not verify is phrased as a question, not a defect.
• A dead-code or unused claim stands only on a complete enumeration of possible consumers — every name the module registers, every template and call site — never on a sampled read. A deadness claim built from a truncated listing once shipped a review comment that, applied, would have broken every page rendering the template.
• When any claim in a submitted review is disproven — by the author, by later verification, or by your own follow-up work — post a correction reply on that claim's own thread as soon as you know, stating exactly what was wrong and what holds instead. Never build follow-up work on the corrected premise before the retraction is up.

Implementing the suggestions

Unless the review was an APPROVE, submitting it is not the end of the job: follow up immediately with PRs that implement every suggestion the review made, targeted at the reviewed PR.

• Base each follow-up PR on the reviewed PR's head branch (gh pr create --base <head-branch>), so accepting one merges into the author's branch — never into the repo's default branch.
• Split by independent context: one PR per suggestion (or cluster of suggestions) the author can accept or reject on its own. Changes that only make sense together stay in one PR — the split exists so the author can take only what they agree with, not to scatter a coherent change.
• Each PR body links the review comment(s) it implements and tells the author to feel free to simply close the PR if they don't want the change — closing is a complete resolution, with no explanation or reply owed to the reviewer.
• After opening the PRs, reply on each originating review comment thread with a link to the PR that implements it, so every thread carries its own resolution path.
• Follow-up PRs meet the same bar as any other work: the repo's own gates pass on each one.

After the follow-ups

A review delivered through this skill is a complete task: once the review, the follow-up PRs, and the thread replies have landed, run the self-improvement skill over the whole effort before calling it done.

The founder's PRs: take over instead of posting a review

When the PR's author is the founder, nothing above applies as written. The founder contributes PRs in place of feature requests and does not return to work review feedback; by agreement, the user takes their PRs over. Never post a review asking them to change anything, and never open follow-up PRs for them to accept — the resolutions land on their PR directly:

• Put every open decision to the user as a question — with a recommendation — and build only the chosen option. Batch decisions of the same shape into one question set.
• Work lands as commits on the PR's own branch: merge its base branch in (never rebase — the branch backs an open PR), fix every confirmed finding, and finish the feature to the full Definition of Done. That includes live-browser verification of any behavior the automated suites never execute — a static read of frontend code validates its APIs but not its dynamics, and the one real bug of the first takeover was exactly there.
• When the branch is merge-ready: rewrite the PR description to current truth for an external reader, preserving the founder's original spec block (it is the intent record); reply on every open thread with its resolution; resolve the threads; and dismiss any now-stale blocking review, crediting what got fixed.
• Merging the PR remains the user's act. Report it merge-ready and stop — do not ask whether to merge: the user asks for the merge, never the other way around.

Replacing a review

To supersede an earlier review of yours, submit the new one first, then dismiss the old:

gh api -X PUT repos/<owner>/<repo>/pulls/<number>/reviews/<review-id>/dismissals \
  -f message="Superseded by the inline review below." -f event="DISMISS"

Find the review id with gh api repos/<owner>/<repo>/pulls/<number>/reviews.

Resolving your own threads

When a concern of yours has been addressed and you approve (or supersede a blocking review with an approval), resolve every thread you opened — a reply is not a resolution, and an unresolved thread keeps blocking the merge on repos that require conversation resolution. Only GraphQL can do it:

gh api graphql -f query='mutation { resolveReviewThread(input: {threadId: "<thread-id>"}) { thread { isResolved } } }'

Find thread ids via the reviewThreads connection on the pull request. Resolve only threads whose concern you have verified as addressed — resolution is the reviewer's verdict on the thread, not cleanup.