From agent-teams-review
Collaborative Agent Teams code review with cross-reviewer communication, AI Slop detection, and a unified professional report. Use when the user asks to review a PR, review code changes, or run a code review. Spawns a team of independent reviewer teammates (Virtual Mariusz, Backend Solidifier, Frontend Virtuoso, Quality Purist, Security Sentinel, Devil's Advocate) that can communicate with each other and share findings. Produces a professional report with executive summary, AI Slop score, cross-reviewer findings, and verdict.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-teams-review:agent-teams-reviewThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Run a collaborative code review using Agent Teams. Unlike subagent-based reviews, teammates can communicate with each other, producing cross-reviewer findings that single-pass reviews cannot achieve.
Run a collaborative code review using Agent Teams. Unlike subagent-based reviews, teammates can communicate with each other, producing cross-reviewer findings that single-pass reviews cannot achieve.
Collect the review target from $ARGUMENTS.
Input: The argument is the base branch name to compare against (e.g. develop, main). Default: main if no argument provided.
Git diff mode:
git diff <base-branch>...HEAD --name-only # changed file list
git diff <base-branch>...HEAD # full diff
Specific files: Read each file directly.
Directory: List and read source files, excluding node_modules, dist, build, .git, vendor, storage, etc.
Also read CLAUDE.md if it exists in the project root — this contains project conventions that reviewers need.
Virtual Mariusz (VM-) — ALWAYS RUNS. No condition needed.
Conditional reviewers:
| Reviewer | Trigger Condition |
|---|---|
| Backend Solidifier (BE-) | Files match: *.php, composer.json, config/*.php, migrations/*, database/* |
| Frontend Virtuoso (FE-) | Files match: *.vue, *.ts, *.tsx, *.js, *.jsx, nuxt.config.*, *.css, *.scss |
| Quality Purist (QA-) | When Backend Solidifier OR Frontend Virtuoso is selected |
| Security Sentinel (SC-) | File path matches: **/auth/**, **/security/**, **/middleware/**, **/guard/**, **/policy/** OR file content contains: password, login, token, secret, session, csrf, sanitize, encrypt, hash |
| Devil's Advocate (DV-) | New classes/functions >30 lines, complex conditionals (>3 nesting levels), financial/payment logic detected, OR total lines changed >150 |
After auto-selection, show the user which reviewers are selected and which are skipped (with reasons).
Use AskUserQuestion with multiSelect: true to let the user override the selection. Present all 6 reviewers with their auto-selected state, allowing the user to add or remove reviewers.
Read each selected reviewer's profile from references/:
| Reviewer | Profile File |
|---|---|
| Virtual Mariusz | references/virtual-mariusz.md |
| Backend Solidifier | references/backend-solidifier.md |
| Frontend Virtuoso | references/frontend-virtuoso.md |
| Quality Purist | references/quality-purist.md |
| Security Sentinel | references/security-sentinel.md |
| Devil's Advocate | references/devils-advocate.md |
CRITICAL: This step uses the Agent Teams feature, NOT the Task tool with subagents.
Instruct Claude to create an Agent Team with the selected reviewers. Use Sonnet for each teammate to optimize token costs — code review does not require Opus-level reasoning. For each reviewer:
## Review Context
### Files Changed
[list of changed files from Step 2]
### Code Diff
[full diff or file contents from Step 2]
### Project Conventions
[CLAUDE.md content if it exists]
## Team Communication
You are part of a review team. You can message other active teammates:
[List active teammate names and their roles]
When you find something in another reviewer's domain:
- Message them directly with the file, line number, and your concern
- They will investigate and report if it's a real issue
- These cross-reviewer findings are especially valuable
When you receive a message from a teammate:
- Investigate the flagged concern within your expertise
- Report findings with the CROSS- prefix if confirmed
- Acknowledge if you've already covered it
| Task | Assigned To | Depends On |
|---|---|---|
| Review: [domain-specific files] | [Reviewer Name] | — |
| Review: [domain-specific files] | [Reviewer Name] | — |
| ... | ... | — |
| Aggregate & Format Report | Lead | All review tasks |
Each review task should describe:
Teammates claim their assigned tasks and mark them as completed when done. The aggregation task unblocks automatically when all reviews finish.
CRITICAL: Do NOT start aggregating, formatting, or doing any review work yourself. You are the lead — your job is to coordinate, not review code. Wait for ALL teammates to complete their tasks before proceeding to Step 7.
As teammates finish their reviews, collect their findings. Each teammate will produce:
If a teammate appears stuck or idle without completing their task, nudge them to continue. Only proceed to Step 7 when every review task in the shared task list is marked as completed.
CROSS and attribution (do NOT create a separate section)Format the aggregated findings using the Action-First report template from references/output-format.md.
See references/example-report.md for a complete example of the expected output.
Key sections (in order):
Save the formatted report to:
docs/reviews/{branch-name}-{YYYY-MM-DD}.md
Create the docs/reviews/ directory if it does not exist.
If the branch name contains slashes (e.g. feature/user-auth), replace them with hyphens in the filename (e.g. feature-user-auth-2026-03-02.md).
Display the executive summary to the user:
Action Loop — repeat until the user selects "Done":
Use AskUserQuestion to offer:
Invoke the triage skill to run a full prioritization analysis on the review report. Use the Skill tool:
agent-teams-review:triagedocs/reviews/feature-auth-2026-03-16.md)The triage skill handles everything: loading the report, assessing each finding across 7 dimensions, classifying into Fix Now / Fix Later / Skip groups, smart grouping, and presenting results to the user. After the triage skill completes (including its own follow-up questions and optional save), return to the action loop.
Show each Critical/High issue with its suggested fix. Let the user pick which fixes to apply using AskUserQuestion with multiSelect: true. Apply selected fixes to the code.
Display the complete report.
After all follow-up actions are complete, clean up the agent team to release shared resources.
IMPORTANT: Only the lead (this session) should run cleanup. Never let a teammate run cleanup — their team context may not resolve correctly.
git diff fails, fall back to asking the user for specific file pathsnpx claudepluginhub mrstroz/claude-code-plugins --plugin agent-teams-reviewOrchestrates team-based code review with 3 specialized reviewers (security, correctness, spec compliance) examining all files. For large diffs requiring cross-file analysis.
Code review with semantic diffs, expert routing by file type, and auto-task creation for critical issues. Works on staged changes, files, or PRs.
Deploys team of 5 specialist agents (backend, frontend, test, security, UX) to parallel-review git diff changes and generate integrated Markdown report in docs/reviews/. Use for team code reviews and quality checks.