Multi-Agent Code Review

Pragmatic Quality framework — orchestrate parallel review agents for high-signal findings.

Parse Arguments

• If $ARGUMENTS contains --post-to-pr: enable GitHub PR posting (Phase 5)
• Remaining non-flag arguments: use as output directory (default: ./reviews/code/)

Example usage:

• /review:code-review — local report to ./reviews/code/
• /review:code-review custom/dir — local report to custom/dir/
• /review:code-review --post-to-pr — local report + post to GitHub PR
• /review:code-review custom/dir --post-to-pr — both

Git Analysis

Analyze these outputs to understand the scope and content of the changes.

GIT STATUS:

!`git status`

FILES MODIFIED:

!`git diff --name-only origin/HEAD...`

COMMITS:

!`git log --no-decorate origin/HEAD...`

DIFF CONTENT:

!`git diff --merge-base origin/HEAD`

Phase 1 — Context Gathering

Launch 2 parallel Haiku agents:

Agent A — CLAUDE.md Discovery: Find all CLAUDE.md files in the repo (root + directories modified by the changes). Return file paths and brief content summaries of each.

Agent B — Change Summary: Analyze the diff above. Return: files changed count, primary areas affected, change type (feature/bugfix/refactor/config/test/docs), estimated risk level (Low/Medium/High/Critical).

Phase 2 — Parallel Review

Launch 4-5 Sonnet agents simultaneously. Provide each with: the full diff content, the CLAUDE.md summaries from Phase 1, and the change summary. Each agent returns findings in this format:

Finding: {description}
File: {path}:{line}
Category: {CLAUDE.md | Bug | History | Security | Comments}
Reason: {why this was flagged}
Suggested fix: {code snippet, if applicable}

Reviewer 1 — CLAUDE.md Compliance

Audit changes against all discovered CLAUDE.md rules. Only flag items specifically called out in a CLAUDE.md. Double-check that the CLAUDE.md actually requires what is being flagged. Ignore silenced rules (lint-ignore comments).

Reviewer 2 — Bug Scanner

Shallow scan for obvious bugs in the diff only. Focus on large bugs — avoid nitpicks. Do NOT read extra context beyond the changes. Ignore issues linters/typecheckers would catch.

Reviewer 3 — Git Blame/History Analyzer

Read git blame and history of modified files. Identify issues in light of historical context: reverted changes being re-modified, recently-fixed areas, breaking established conventions, patterns from previous PR comments.

Reviewer 4 — Security Reviewer

Security-focused scan: injection (SQLi, XSS, command), auth/access control, secrets/credentials, data exposure in logs/responses, crypto misuse. Only report HIGH confidence exploitable findings.

Reviewer 5 — Code Comments Compliance (conditional)

Only launch if modified files contain substantive code comments (// NOTE:, // IMPORTANT:, // INVARIANT:, // SAFETY:, // TODO:). Ensure changes comply with guidance in those comments.

False Positive Awareness

All reviewers must skip these false positive categories:

• Pre-existing issues not introduced in the changes
• Issues that linters, typecheckers, or compilers would catch
• Pedantic nitpicks a senior engineer wouldn't flag
• Framework-handled concerns (e.g., XSS in React unless using unsafe HTML injection APIs)
• General quality issues unless explicitly required in CLAUDE.md
• Style preferences matching existing codebase conventions
• Real issues on lines the author did not modify

See WORKFLOW.md for detailed false positive filtering rules.

Phase 3 — Confidence Scoring

For each finding from Phase 2, launch a parallel Haiku agent that:

1. Receives: the finding description, the relevant diff section, and the CLAUDE.md files list
2. Scores 0-100 using this rubric:
   • 0: False positive — doesn't hold up to scrutiny, or pre-existing issue
   • 25: Might be real, but may also be false positive. Unable to verify.
   • 50: Verified issue, but minor or nitpick. Not very important relative to the rest.
   • 75: Very likely real. Existing approach is insufficient. Important and impactful.
   • 100: Confirmed. Will happen in practice. Evidence directly confirms this.
3. For CLAUDE.md findings: double-check the CLAUDE.md actually calls out the issue
4. Returns: score + brief justification

Filter: Remove all findings scoring below 80. If no findings survive, generate a clean report.

Phase 4 — Report Generation

Generate the report using the template in WORKFLOW.md.

Map confidence scores to triage levels:

• 90-100 → Blocker (if severity warrants) or Improvement
• 80-89 → Improvement or Question

Include for each finding: the Source category (CLAUDE.md, Bug Scan, Git History, Security, Comments).

1. Create output directory: mkdir -p {output-directory}
2. Save report to: {output-directory}/{YYYY-MM-DD}_{HH-MM-SS}_code-review.md
3. Display the full report to the user
4. Confirm the save path

Phase 5 — Optional GitHub PR Posting

Only execute if --post-to-pr flag was passed.

1. Check if an open PR exists for the current branch via gh pr view
2. If no PR exists, inform the user: "No open PR found for this branch. Skipping GitHub posting."
3. If a PR exists, check eligibility via a Haiku agent:
   • Is the PR closed? → skip
   • Is the PR a draft? → skip
   • Has Claude already commented on this PR? → skip
4. If eligible, format findings as a concise PR comment and post via gh pr comment
5. Use the GitHub comment format from WORKFLOW.md

Phase 6 — Automatic Verification

After saving the report, invoke the false-positive verifier:

1. Use the Skill tool to invoke review:verify-findings with the saved report path
2. The verifier runs in an isolated forked context and produces a .verified.md report
3. After verification completes, inform the user of both report locations

If the Skill tool is not available (e.g., running inside a subagent):

Run verification manually: /review:verify-findings {report-path}

Resources

• WORKFLOW.md — Detailed review checklists, agent prompt templates, scoring rubric, report template, GitHub comment format
• EXAMPLES.md — Sample reports and orchestration flow examples
• TROUBLESHOOTING.md — Common issues with pipeline, scoring, and output