AgentSkillsCN

group-review-mr

在 GitLab 合并请求上运行多代理集成代码审查（4 个并行代理 + 综合分析）

SKILL.md
--- frontmatter
name: group-review-mr
description: Run a multi-agent ensemble code review on a GitLab merge request (4 parallel agents + synthesis)
argument-hint: <MR URL>
allowed-tools:
  - Bash
  - Write
  - Glob
  - Read
  - Task
  - mcp__gitlab-mcp-code-review__fetch_merge_request
  - mcp__gitlab-mcp-code-review__fetch_merge_request_diff

Group Review (Ensemble) - Merge Request

Review the GitLab merge request at: $ARGUMENTS

This uses a multi-agent ensemble approach with 4 specialized reviewers running in parallel, followed by a synthesis step.

Phase 1: Setup

  1. •Parse URL - Extract project name and MR IID from the URL
  2. •Create Directory - Create reviews/[PROJECT]/[MR_ID]/ if it doesn't exist
  3. •Check Existing Reviews - Find existing review-notes-*.md files and determine next increment
  4. •Fetch MR Data - Use GitLab MCP tools to get MR details and diff
  5. •Setup Local Codebase - Clone/update repo in ./codebase/[PROJECT]/ and checkout source branch

Phase 2: Launch Parallel Agents

Launch 4 Task agents IN PARALLEL (single message, multiple tool calls) with these prompts:

Agent 1 - Security

Read workflow/prompts/security-agent.md for the full prompt. Key focus areas:

  • •Injection vulnerabilities (SQL, command, LDAP, template)
  • •XSS (cross-site scripting)
  • •Authentication & authorization issues
  • •Data protection (sensitive data exposure, hardcoded secrets)
  • •Input validation
  • •Security misconfigurations
  • •CSRF & request forgery

Output to: reviews/[PROJECT]/[MR_ID]/agent-security.md

Agent 2 - Logic

Read workflow/prompts/logic-agent.md for the full prompt. Key focus areas:

  • •Logic errors (conditionals, algorithms, race conditions)
  • •Edge cases (null handling, boundaries, empty collections)
  • •Error handling (missing try/catch, silent failures)
  • •Business logic correctness
  • •Data integrity
  • •Control flow issues
  • •API contract issues

Output to: reviews/[PROJECT]/[MR_ID]/agent-logic.md

Agent 3 - Quality

Read workflow/prompts/quality-agent.md for the full prompt. Key focus areas:

  • •Code design (SOLID violations, coupling, abstraction)
  • •Maintainability (complexity, duplication, naming)
  • •Performance (N+1 queries, missing indexes, inefficient algorithms)
  • •Database concerns
  • •API design
  • •Testing concerns
  • •Framework best practices

Output to: reviews/[PROJECT]/[MR_ID]/agent-quality.md

Agent 4 - General

Read workflow/prompts/general-agent.md for the full prompt.

  • •Holistic review without specialization
  • •Catches cross-cutting issues
  • •Typical senior developer perspective

Output to: reviews/[PROJECT]/[MR_ID]/agent-general.md

Context to Provide Each Agent

Include in each agent's prompt:

markdown
# Your Role
[Content from the agent's prompt file in workflow/prompts/]

# MR Information
- **Title:** [MR title]
- **Author:** [author name]
- **Source Branch:** [source] -> **Target:** [target]
- **Description:** [MR description]

# Project-Specific Guidelines
[Content from project-instructions/[PROJECT].md if it exists]

# General Review Guidelines
[Content from reviewing.md]

# The Diff
[Full diff content from GitLab MCP]

# Your Task
Review the diff above from your specialized perspective.
Write your findings to: `reviews/[PROJECT]/[MR_ID]/agent-[type].md`
Use the output format specified in your role instructions.
Focus ONLY on the diff - don't review the entire codebase.

Phase 3: Synthesis

After all 4 agents complete, launch synthesis agent:

Read workflow/prompts/synthesis-agent.md for instructions. The synthesis agent should:

  1. •Read all 4 agent outputs
  2. •Identify consensus findings (2+ agents = high confidence)
  3. •Deduplicate overlapping issues
  4. •Validate single-agent findings
  5. •Prioritize by severity
  6. •Write final review to: reviews/[PROJECT]/[MR_ID]/review-notes-[N].md

Phase 4: Report

Present the final synthesized review to the user, including:

  • •Summary of findings
  • •Issues by severity (with consensus indicators)
  • •Agent agreement table

Output Format for Each Agent

markdown
## [Agent Type] Review

### Critical
- [ID-001] **[Title]** - file:line
  - Description: [What the issue is]
  - Impact/Risk: [What could happen]
  - Recommendation: [How to fix]

### Major
- [ID-002] ...

### Minor
- [ID-003] ...

### Notes
- [Observations]

Important Notes

  • •Run the 4 review agents IN PARALLEL (single message with multiple Task tool calls)
  • •Wait for all 4 to complete before running synthesis
  • •Focus on the DIFF. The entire file and the rest of the codebase is there to gain a better understand in relation to the diff
  • •Only flag issues outside the diff if Critical/Major severity
  • •Check for project-instructions/[PROJECT].md for project-specific guidelines
  • •Follow CLAUDE.md and reviewing.md for general review practices