review-bot

/review-bot Skill

Triage review bot comments on a GitHub PR. For each comment: re-review the flagged code, apply valid fixes (or a better fix), and dismiss false positives. Post a structured summary comment on the PR when done.

Step 0 — Plan mode guard

If you are currently in plan mode, stop immediately. Tell the user:

"/review-bot requires Edit and Write access to apply fixes. You're in plan mode. Please exit plan mode first, then re-run /review-bot."

Do not proceed.

Step 1 — Resolve the PR

Parse $ARGUMENTS for:

• A PR number (bare integer, e.g. 123)
• --bot <name> (optional, a GitHub username to treat as the target bot)

If no PR number was provided, auto-detect from the current branch:

gh pr view --json number,url,title,headRefName

If this fails (no PR for the current branch), tell the user and stop.

Store the PR number, URL, title, and the repo owner/name (extract from gh repo view --json nameWithOwner -q '.nameWithOwner').

Step 2 — Fetch bot review comments

Fetch line-level review comments from the PR:

gh api repos/{owner}/{repo}/pulls/<PR_NUMBER>/comments --paginate

Each comment object includes:

• id — unique comment ID
• body — the comment text
• path — file path the comment is on
• line (or original_line) — line number in the diff
• diff_hunk — surrounding diff context
• user.login — author username
• user.type — "Bot" or "User"
• in_reply_to_id — parent comment ID if this is a thread reply
• html_url — link to the comment on GitHub

Filter to bot comments

1. If --bot <name> was provided: keep only comments where user.login matches (case-insensitive).
2. Otherwise: keep comments where user.type == "Bot".
3. Always exclude comments by github-actions[bot] (CI status, not code review).

If no bot comments are found, tell the user "No bot review comments found on PR #N" and stop.

Group threaded conversations

Group comments by thread using in_reply_to_id:

• Comments with no in_reply_to_id are root comments (one finding each).
• Comments whose in_reply_to_id points to another bot comment in the set belong to that root's thread.
• For each thread, the root comment is the primary finding. Append follow-up messages as additional context.

Present a brief summary:

Found N bot comments (M threads) from @botname on PR #X.

Proceed immediately — do not wait for confirmation.

Step 3 — Evaluate each comment

Process each thread/comment sequentially.

3a. Read the code

Read the file at path. Use line and diff_hunk to locate the exact code. Read at least 30 lines above and below for context.

If the file does not exist (deleted or renamed since the review), mark the comment as outdated and skip to the next.

3b. Understand the bot's feedback

Parse the comment body to extract:

• What it flags: the specific issue or concern
• Why: the reasoning (if provided)
• Suggested fix: any code suggestion (markdown code blocks or GitHub suggestion blocks)

3c. Assess validity

Determine whether the feedback is valid:

1. Is the flagged issue real? Does the code actually have the problem described?
2. Is it already handled? Check if the concern is addressed elsewhere (middleware, parent scope, etc.).
3. Is it relevant to the project? Check CLAUDE.md / project conventions — the bot may flag something the project intentionally does differently.
4. Is the suggested fix correct? Even if the issue is real, the bot's suggestion may be wrong or suboptimal.

Classify each comment:

Verdict	Meaning
fix	Valid issue. Apply the bot's suggestion.
fix-alt	Valid issue, but apply a better fix than the bot suggested.
dismiss	False positive or not applicable.
outdated	File/code no longer exists.
flag	Uncertain — needs user decision.

Assign confidence: high (>= 90%), medium (70–89%), low (< 70%).

If confidence is low, do not auto-fix — classify as flag instead.

3d. Apply fix (if applicable)

For comments with verdict fix or fix-alt and confidence high or medium:

• Use the Edit tool to apply the change.
• If the bot provided a GitHub suggestion block, use it as a starting point but verify correctness before applying.
• For fix-alt, implement the better approach rather than the bot's literal suggestion.
• After editing, briefly note what was changed and why.

Do not apply fixes that conflict with each other. If two bot comments suggest contradictory changes to the same code region, classify both as flag and let the user decide.

Step 4 — Present results

After processing all comments, present a summary:

### Review Bot Triage — PR #N

**Bot**: @botname | **Comments**: X | **Fixed**: Y | **Dismissed**: Z | **Flagged**: W

| # | File | Bot says | Verdict | Confidence | Action taken |
|---|------|----------|---------|------------|--------------|
| 1 | src/Foo.php:42 | Missing null check | fix | high | Added null check on `$user` |
| 2 | src/Bar.php:15 | Unused import | fix | high | Removed unused import |
| 3 | src/Baz.php:88 | Possible N+1 query | dismiss | high | Already eager-loaded on line 72 |
| 4 | src/Qux.php:30 | Type mismatch | fix-alt | medium | Used stricter type instead of bot's suggestion |
| 5 | src/Old.php:10 | Missing return type | outdated | — | File was deleted |
| 6 | src/Edge.php:55 | Complex refactor | flag | low | Needs manual review |

Flagged items

If any items are marked flag, present the details and use AskUserQuestion:

• Question: "How would you like to handle the flagged items?"
• Options:
  1. "Fix all" — Apply fixes to all flagged items
  2. "Skip all" — Dismiss all flagged items
  3. "Done" — Leave as-is, proceed to PR comment

The tool adds an "Other" option automatically — use that for specific inputs like "fix 6, skip 7".

Handle the response:

• "fix all": Apply fixes to all flagged items.
• "fix 6" (specific number): Fix only that item.
• "skip all" / "skip 6": Dismiss the specified items.
• "done": Proceed without changes to flagged items.

Step 5 — Post PR summary comment

Compose a single comment on the PR:

gh pr comment <PR_NUMBER> --body "<comment body>"

Use a HEREDOC for the body to preserve formatting. The comment structure:

## Review Bot Triage

Reviewed **N** comments from @botname.

### Applied fixes (Y)

| Comment | File | Issue | Fix applied |
|---------|------|-------|-------------|
| [Link](url) | `src/Foo.php:42` | Missing null check | Added null check |

### Dismissed (Z)

| Comment | File | Issue | Reason |
|---------|------|-------|--------|
| [Link](url) | `src/Baz.php:88` | Possible N+1 | Already eager-loaded at line 72 |

### Skipped — outdated

- [Link](url) — `src/Old.php:10` — File no longer exists

Omit any section that has zero items. If nothing was dismissed, omit the "Dismissed" section entirely. Same for "Applied fixes", "Skipped", etc.

Step 6 — Final status

After posting the comment, tell the user:

Posted triage summary on PR #N. Applied X fixes, dismissed Y comments, flagged Z for review.

If fixes were applied, remind the user:

Changes have been applied but not committed. Run /mella:commit or stage and commit when ready.