review-github-pr（审查 GitHub PR）

Read-only: Do not modify any files. Your job is to review, report, and submit feedback on GitHub. Do not to make local changes.

Before proceeding, verify prerequisites:

• Run gh auth status to confirm GitHub CLI is authenticated. If not, stop and tell the user to run gh auth login.
• Confirm you have a PR or diff to review. If unclear, ask the user what to review before proceeding.

Instructions

Review a GitHub pull request, generate prioritized inline comments, and submit them as a batch review.

Use when: reviewing, evaluating, critiquing, auditing, or assessing a GitHub PR. Triggers on requests like "review this PR", "review PR #123", or when given a PR URL to review.

All GitHub interactions must use the gh CLI tool. Do not use curl, direct API URLs, or web fetching. Always use gh pr, gh api, or other gh subcommands.

1. Fetch the PR diff

Use gh CLI to get the full diff and PR metadata:

gh pr view <PR> --json number,title,body,baseRefName,headRefName
gh pr diff <PR>

If the user provides a URL, extract the repo and PR number from it. For private repos, always use gh CLI rather than fetching URLs directly.

2. Analyze the entire diff

• Read the complete diff, not just individual chunks. Understand the full scope of changes.
• Consider how changes across multiple files relate to each other.
• Check for correctness, security, performance, readability, and maintainability.
• Do not include fluff, praise, or "LGTM"-style commentary. Only include actionable findings: critical issues (P0/P1) and nits (P2/P3). Every comment must point to a specific problem or concrete suggestion. Comments that merely note something is "a good change" or compliment the author are not actionable — omit them entirely.

3. Generate prioritized comments

Organize all findings into priority levels:

• P0 — Blocking: Bugs, security vulnerabilities, data loss risks, broken functionality. These must be fixed before merge.
• P1 — Important: Logic errors, missing edge cases, poor error handling, test gaps. Strongly recommended to fix.
• P2 — Suggestion: Code style improvements, better naming, minor refactors, documentation gaps.
• P3 — Nit: Trivial style preferences, optional improvements, minor observations.

Before finalizing, review all findings — including P3 nits — and deliberately decide whether each is worth including. Do not silently drop nits; either include them or note that they were considered and omitted. Try to frame feedback as questions.

Present the full list of comments to the user, grouped and sorted by priority, before submitting. Each comment should include:

• An index of the feedback to be used in step 4 below.
• Priority level (P0/P1/P2/P3).
• File path and line number.
• The comment text.

4. Ask for review disposition

Ask the user if they want to submit all feedback, or just some, and how they want to submit the review:

• Approve — approve the PR with the comments.
• Request changes — request changes (appropriate when there are P0 or P1 items).
• Comment — submit comments without approving or requesting changes.

The user may skip submitting certain feedback by index (provided in part 3).

5. Submit as a single batch review

Construct a JSON payload and submit all comments as a single review using gh api with --input.

Always write the JSON to a temp file rather than piping via echo — comment bodies often contain single quotes, backticks, and other characters that break shell quoting:

cat <<'PAYLOAD' > /tmp/pr_review.json
<json_payload>
PAYLOAD
gh api repos/{owner}/{repo}/pulls/{number}/reviews --method POST --input /tmp/pr_review.json

JSON payload format:

{
  "event": "APPROVE|REQUEST_CHANGES|COMMENT",
  "body": "",
  "comments": [
    {
      "path": "relative/file/path.ext",
      "line": 42,
      "side": "RIGHT",
      "body": "**[P0]** Comment text here"
    }
  ]
}

Every comment body must start with the priority as a bold prefix: **[P0]**, **[P1]**, **[P2]**, or **[P3]**.

For the review body/summary, leave it empty ("body": ""). Do not include summaries, priority counts (e.g. "1 P1, 3 P2"), fluff, filler, or praise. All feedback belongs in inline comments only.

Important — use line and side, NOT position:

• line: The actual file line number in the new version of the file on the PR branch. Do NOT try to calculate this from diff hunk headers — hunk offsets are error-prone. Instead, fetch the real file from the PR branch to confirm each line number before submitting: gh api "repos/{owner}/{repo}/contents/{path}?ref={head_branch}" --jq '.content' | base64 -d | cat -n | sed -n '{start},{end}p'
• Only fetch files whose paths appear in the diff. Do not guess or infer file paths that are not explicitly listed in the gh pr diff output — they may not exist, leading to 404 errors.
• side: Always "RIGHT" for comments on new/changed lines.
• Do not use the deprecated position field — it refers to a line's offset within the diff hunk and is error-prone across multi-hunk files.

Handling lines between diff hunks: If a comment targets a line that falls between diff hunks (i.e. it's unchanged code not shown as context in any hunk), the line-based review comment will be rejected with a 422 error. In this case, submit that comment separately as a file-level comment using the individual comment endpoint:

cat <<'PAYLOAD' > /tmp/pr_file_comment.json
{
  "body": "**[P2]** Comment text here",
  "path": "relative/file/path.ext",
  "subject_type": "file",
  "commit_id": "<head_commit_sha>"
}
PAYLOAD
gh api repos/{owner}/{repo}/pulls/{number}/comments --method POST --input /tmp/pr_file_comment.json

Get the head commit SHA from gh pr view <PR> --json headRefOid --jq '.headRefOid'. Reference the specific line number in the comment body so the author knows where to look. Never combine multiple comments into a single comment body to work around this limitation — always submit each finding as its own comment.

Return the review URL to the user when complete.