Review PR

Review incoming contributions for factual correctness and whether they make the documentation better as a whole. Treat a technically true addition as insufficient when it is misplaced, overemphasized, redundant, or unhelpful to the page's intended reader.

Preserve the write boundary

Perform the review in two phases:

1. Research the PR, decide a verdict, and present the exact proposed comment or review text.
2. Wait for explicit user confirmation, then post only the confirmed text.

Before confirmation, do not post comments, submit a GitHub review, approve or request changes, resolve threads, push commits, edit labels, or otherwise mutate GitHub. A request to review or draft feedback is not confirmation to post it. Ask Post these comments? and stop. Treat revisions to a draft as unconfirmed until the user explicitly asks to post them.

1. Gather the full context

For each PR, inspect its metadata, body, commits, changed files, checks, conversation, reviews, and linked issues. Always fetch inline comments separately because gh pr view --json reviews omits them.

gh pr view <PR> --repo docker/docs \
  --json number,title,url,state,author,body,baseRefName,headRefName,headRefOid,commits,files,comments,reviews,reviewDecision,statusCheckRollup
gh api repos/docker/docs/pulls/<PR>/comments \
  --jq '[.[] | {id, author: .user.login, body, path, line, side, commit_id}]'
gh pr diff <PR> --repo docker/docs

Read linked issues and relevant discussion. An issue is evidence that a reader was confused, but it does not establish the reporter's diagnosis or justify a new highlighted note by itself. Green CI establishes only that automated checks passed, not that the content is correct.

Fetch the PR head when local inspection is useful. Compare it with the canonical upstream base rather than assuming the local branch is fresh. Read each changed file in full, not only its diff.

2. Research independently

Verify every material claim against authoritative sources such as product source code, upstream documentation, specifications, release notes, or safe local reproduction. Do not accept the PR description, issue diagnosis, or existing review feedback as fact.

Search the documentation for related explanations and canonical pages. Read STYLE.md, COMPONENTS.md, and applicable repository instructions. Check whether a changed file is generated or maintained upstream and identify the correct upstream repository instead of proposing a local edit.

Distinguish among:

• a wrong fact
• a correct fact expressed inaccurately
• a correct fact placed on the wrong page
• content already explained elsewhere
• a real discovery problem better addressed with a short signpost and link
• a request that needs no documentation change.

If an external claim or replacement URL cannot be verified, report that limitation instead of guessing.

3. Assess editorial fit

Apply these questions to each addition:

• Does it change a reader's decision or next action on this page?
• Is this the canonical page for the concept?
• Is the fact general, or specific to this page, feature, or component?
• Is the information already documented elsewhere?
• Would a concise local signpost to canonical coverage solve the discovery problem better than duplicating the explanation?
• Is the visual and textual weight proportional to the information's value?
• Does it preserve the page's scope, flow, and character?

Prefer one coherent explanation in the canonical location. Add local context only when it helps the reader complete the task at hand. Avoid stray notes, callouts, and exhaustive edge cases whose prominence exceeds their value.

4. Choose a decisive verdict

Lead with one of these outcomes:

• Approve: correct, useful, well placed, and ready to merge.
• Approve with optional polish: ready to merge; suggestions are genuinely non-blocking.
• Focused rewrite: the underlying need is valid, but wording, scope, placement, or structure should change before merge.
• Close / no docs change: incorrect, redundant, out of scope, or not a documentation problem.

Explain the verdict with evidence. When wording is the issue, provide exact replacement text rather than a vague request to improve it.

5. Place feedback deliberately

Use an inline comment when the finding is anchored to a narrow changed line or range and acting on it is local. Examples include an inaccurate sentence, an ambiguous option description, a broken link, or a precise wording replacement.

Use a PR-wide comment for scope, information architecture, overall approach, multiple intertwined edits, or a proposed replacement section. Do not attach holistic feedback to an arbitrary line.

Use both when appropriate: put the overall direction in the PR-wide comment and line-specific corrections inline. Do not repeat the same point in both. Consolidate related feedback so the author receives the fewest comments that remain clear and actionable.

For every proposed inline comment, resolve and display the current changed file path and right-side diff line. If the target line is not part of the current diff or cannot be identified reliably, use a PR-wide comment that quotes the target text instead. Never guess a line number.

End comments posted on the user's behalf with an accurate agent-disclosure footer, such as Generated by Codex.

6. Present drafts and stop

Before any GitHub write, show the review in this form, omitting empty sections:

## Verdict

Focused rewrite

## Findings

- <finding and evidence>

## Proposed inline comments

1. `path/to/file.md:42`
   > Exact comment text

## Proposed PR-wide comment

> Exact comment text

Post these comments?

For multiple PRs, give each PR its own verdict and comment set. Make the confirmation scope unambiguous. Do not interpret approval of one PR's drafts as approval to post comments on the others.

7. Post only confirmed feedback

Immediately before posting, re-fetch the PR head SHA and diff. If either the head or an inline target changed, stop and show the updated draft or placement for confirmation.

Post confirmed inline comments as a single comment-only review when practical. Use the current head SHA and right-side diff lines:

gh api repos/docker/docs/pulls/<PR>/reviews --method POST --input <payload>

The JSON payload contains commit_id, event: "COMMENT", and a comments array whose entries contain path, line, side: "RIGHT", and body. Submitting a review with APPROVE or REQUEST_CHANGES requires separate, explicit user authorization; a verdict alone does not grant it.

Post confirmed holistic feedback separately:

gh pr comment <PR> --repo docker/docs --body-file <file>

Use a safely created temporary file or API input so Markdown, backticks, and shell substitutions are preserved literally. Post exactly the confirmed text. Verify the resulting review/comments and report their URLs and placements. If GitHub rejects an inline location, do not silently fall back to a PR-wide comment; report the failure and prepare a revised placement for confirmation.

Definition of done

• Verify technical claims with authoritative evidence.
• Evaluate usefulness, placement, duplication, and proportionality.
• Give a decisive verdict and exact actionable wording.
• Choose inline and PR-wide placement based on the feedback's scope.
• Show every exact draft and target before any GitHub mutation.
• Post only after explicit confirmation and verify what was posted.