Agent Reviewer

Independent QA layer over agent-produced work. Spawns parallel review sub-agents to audit changes across four dimensions, scores findings, and optionally enters Plan Mode to produce a correction plan.

When to Use

Trigger when the user asks any of:

• "Review what the agent just did"
• "Second opinion on these commits / this PR"
• "Validate the agent's work before I merge"
• "Check if agent changes are correct"
• "Did the agent do what I asked?"

Do NOT use this skill when:

• The user wants a standard PR review without agent context → call /code-review directly.
• The user wants to execute or plan new work → use mdn-run or mdn-plan.

Inputs

Gather from context or ask one clarifying question:

1. scope (required, infer from context):
   • local — no scope given; auto-detect via git status + git diff HEAD.
   • commits:<sha1>..<sha2> — user specifies a commit range.
   • pr:<number> — user provides a Pull Request number.
2. intent (required): what the agent was supposed to do. If not stated, ask before proceeding.
3. create_plan (yes / no / ask, default ask): whether to enter Plan Mode if significant issues are found.

Step 0 — Detect Scope and Load Context

1. If the user passed pr:<N> → scope = PR.
2. If a commit range is mentioned → scope = commits.
3. Otherwise → run git status and git diff HEAD; scope = local.
4. Read the root CLAUDE.md (if it exists) for compliance checks in Step 2.
5. Restate scope and intent to the user before proceeding (cheap sanity check).

Step 1 — Collect the Diff

Depending on scope:

• PR: gh pr diff <N> + gh pr view <N> (title, description, body, changed files).
• Commits: git log <range> --oneline + git diff <range> + commit messages.
• Local: git diff HEAD (staged + unstaged) + git status.

Extract and store: list of changed files, full diff, associated commit messages.

Step 2 — Parallel Multi-Agent Review

Launch 4 Sonnet agents in parallel (single message, 4 Agent tool calls):

Each agent returns a list of findings with: description, estimated severity (low / medium / high), and justification.

False positive guidelines for all agents (identical to the built-in code-review plugin):

• Pre-existing issues not touched by this diff.
• Something that looks like a bug but is not.
• Pedantic nitpicks a senior engineer would not call out.
• Issues a linter or type-checker would catch — assume CI handles those.
• General quality concerns (test coverage, documentation) unless CLAUDE.md explicitly requires them.

Step 3 — Score Findings

For each finding from Step 2, launch a parallel Haiku agent that scores it 0–100 (give this rubric verbatim):

• 0: False positive. Does not survive light scrutiny or is a pre-existing issue.
• 25: Possible issue, unverified. May be a false positive.
• 50: Real issue, but minor or unlikely to be hit in practice.
• 75: Important issue, well-verified, likely to occur in practice.
• 100: Critical and confirmed. Will definitely occur frequently.

Filter out findings with score < 75. If no findings remain, the agent work is clean — report that and stop.

Step 4 — Optional: Invoke /code-review as Sub-Agent (PR scope only)

If scope = PR and create_plan != no:

1. Launch an agent that runs the built-in code-review command on the PR.
2. Merge its findings into the consolidated report, labelled [code-review] as source.
3. Apply the same ≥ 75 filter before merging.

This step is additive — it supplements Steps 2–3, does not replace them.

Step 5 — Consolidate and Report

Generate a Markdown report in the conversation:

## Agent Review — <scope>

### Intent
<what the agent was supposed to do>

### Verdict
✓ Aligned | ⚠ Minor issues | ✗ Significant issues

### Findings (<N> issues)

1. [HIGH] <brief description> — score: 85
   File: path/to/file.py:42
   Reason: <justification>

2. [MEDIUM] <brief description> — score: 77
   File: path/to/other.ts:10
   Reason: <justification>

### Clean Dimensions
- Intent alignment ✓  (or listed here if no issues found for that agent)
- CLAUDE.md compliance ✓

### Recommendation
<free text: merge / fix before merge / requires correction plan>

Rules:

• Keep descriptions brief. Link to file and line for every finding.
• Do not emit emojis beyond the verdict line.
• If zero findings survive the filter, report "No issues found" and stop.

Step 6 — Correction Plan (Conditional)

If findings with score ≥ 75 exist:

• create_plan=yes — Enter Plan Mode automatically (EnterPlanMode) with findings as requirements context.
• create_plan=ask — Ask the user: "Do you want me to generate a correction plan for these findings?" Proceed only on explicit confirmation.
• create_plan=no — Skip.

The correction plan uses the findings as requirements and follows the standard Claude Code plan structure. Reference the specific files and line numbers from the findings.

Hard Constraints

• Read-only against git and GitHub — never commit, push, or merge as part of this review.
• No fabricated findings — if an agent finds nothing, record "no issues" and continue. Do not invent issues to fill the report.
• No build steps — do not run tests, linters, or type-checkers. Assume CI handles those.
• Cite everything — every finding must include a file path and line reference.