multi-ai-collab（多 AI 协作）

Multi-AI Collaboration Skill

Overview

This skill enables the invoking AI agent to act as an Orchestrator, coordinating multiple AI agents (Codex CLI, Gemini CLI, Claude sub-agents) with assigned Personas (specialized expert roles) for collaborative software development tasks.

The primary use case is Cross-Review: having multiple AI agents independently analyze code from different expert perspectives, then synthesizing their findings to provide comprehensive, bias-reduced results.

Prerequisites

Before using this skill, ensure the required CLI tools are installed:

# Check available agents
which codex && codex --version
which gemini && gemini --version
which claude && claude --version

CLI Installation

• Codex CLI: See OpenAI Codex Documentation
• Gemini CLI: npm install -g @google/gemini-cli or brew install gemini-cli
• Claude Code: See Claude Code Documentation

Workflow

┌─────────────────────────────────────────────────────────────────────┐
│                    ORCHESTRATOR (Invoking AI)                       │
│                                                                     │
│  The AI agent that invokes this skill becomes the orchestrator.     │
│  It coordinates all sub-agents and synthesizes results.             │
└─────────────────────────────────────────────────────────────────────┘
                                │
        ┌───────────────────────┼───────────────────────┐
        ▼                       ▼                       ▼
┌───────────────┐       ┌───────────────┐       ┌───────────────┐
│   Codex CLI   │       │  Gemini CLI   │       │ Claude (sub)  │
│  latest       │       │  latest       │       │   latest      │
│               │       │               │       │               │
│   Persona:    │       │   Persona:    │       │   Persona:    │
│   Architect   │       │   Security    │       │  QA Engineer  │
└───────────────┘       └───────────────┘       └───────────────┘

Phase 1: Task Analysis (Silent)

The orchestrator performs initial analysis using a Parallel Fan-Out pattern for efficiency:

┌─────────────────────────────────────────────────────────────────────┐
│                     Phase 1: Parallel Fan-Out                       │
│                                                                     │
│   ┌─────────────────────┐     ┌─────────────────────┐              │
│   │  Identify target    │     │  Detect available   │              │
│   │  files/code         │     │  AI agents          │   PARALLEL   │
│   └─────────┬───────────┘     └──────────┬──────────┘              │
│             │                            │                          │
│             └────────────┬───────────────┘                          │
│                          ▼                                          │
│             ┌─────────────────────┐                                 │
│             │  Analyze task       │                                 │
│             │  nature             │                     SEQUENTIAL  │
│             └──────────┬──────────┘                                 │
│                        ▼                                            │
│             ┌─────────────────────┐                                 │
│             │  Recommend          │                                 │
│             │  personas           │                                 │
│             └─────────────────────┘                                 │
└─────────────────────────────────────────────────────────────────────┘

Step 1 (Parallel): Execute these tasks concurrently as they have no dependencies:

• Identify target files/code - Use Glob, Grep, Read tools to understand scope
• Detect available AI agents - Check which CLIs are installed (which codex gemini claude)

Step 2 (Sequential): After parallel tasks complete, execute in order:

• Analyze task nature - Determine if it's implementation, review, refactoring, investigation (requires file context from Step 1)
• Recommend personas - Suggest appropriate expert roles based on task nature and available agents

Phase 2: Team Assembly (Interactive)

Use the environment-appropriate user input tool to configure the team:

Q1: Select Personas (Multiple Choice)

Which expert personas should participate in this task?

1. 🏗️ Architect - System design, modularity, dependencies
2. 🔒 Security Researcher - Vulnerabilities, OWASP, auth/authz
3. 🧪 QA Engineer - Test design, edge cases, coverage
4. 👁️ Code Reviewer - Code quality, readability, best practices
5. ⚡ Performance Engineer - Complexity, memory, caching
6. 🔍 Analyzer - Static analysis, bug patterns, type safety
7. 📝 Documentarian - API docs, comments, README
8. 🧠 Domain Expert - Business logic, requirements fit

Recommended based on task analysis: 1, 2, 4

Q2: Assign AI Agents to Personas

Assign an AI agent to each selected persona:

Architect:
  1. Codex CLI (latest default) - Recommended: deep reasoning
  2. Gemini CLI (latest default)
  3. Claude (sub-agent)

Security Researcher:
  1. Codex CLI (latest default)
  2. Gemini CLI (latest default) - Recommended: can search latest CVEs
  3. Claude (sub-agent)

Code Reviewer:
  1. Codex CLI (latest default)
  2. Gemini CLI (latest default)
  3. Claude (sub-agent) - Recommended: fast iteration

Q3: Select Workflow Mode

Select workflow mode:

1. Parallel - All agents work independently, synthesize at end (Recommended for cross-review)
2. Sequential - Each agent builds on previous results
3. Pipeline - Implementation → Test → Review flow
4. Adversarial - Agents critically challenge each other's findings

Phase 3: Execution

The orchestrator executes the configured workflow.

Parallel Mode (Cross-Review)

┌─────────────┐   ┌─────────────┐   ┌─────────────┐
│   Agent A   │   │   Agent B   │   │   Agent C   │
│  (Codex)    │   │  (Gemini)   │   │  (Claude)   │
└──────┬──────┘   └──────┬──────┘   └──────┬──────┘
       │                 │                 │
       ▼                 ▼                 ▼
   Result A          Result B          Result C
       │                 │                 │
       └────────────┬────┴────────────────┘
                    ▼
             ┌─────────────┐
             │  Synthesis  │
             └─────────────┘

Execution Commands:

# Codex CLI (Architect persona) - omit --model to use latest default
codex exec "You are a Senior Software Architect.
Analyze the following code for:
- Modularity and separation of concerns
- Dependency management
- Extensibility and maintainability
- Design pattern usage

[CODE_CONTENT]

Provide findings with severity (Critical/High/Medium/Low) and recommendations."

# Gemini CLI (Security persona) - omit -m to use latest default
gemini -p "You are a Security Researcher.
Analyze the following code for:
- OWASP Top 10 vulnerabilities
- Authentication/authorization issues
- Input validation and sanitization
- Data protection and encryption

[CODE_CONTENT]

Provide vulnerabilities with CVSS scores and remediation steps."

# Claude sub-agent (QA persona)
# Use the environment-appropriate subagent tool (Claude Code Task / Codex spawn_agent)
"You are a QA Engineer.
Based on the code, design:
- Required test cases (unit, integration, e2e)
- Edge cases and boundary conditions
- Security test scenarios
- Performance test considerations

[CODE_CONTENT]"

Sequential Mode

Agent A → Agent B → Agent C → Synthesis
   │          │          │
   └── Pass results to next agent

Each agent receives the previous agent's findings and builds upon them.

Pipeline Mode

Implementer → Tester → Reviewer
     │           │          │
   Code      Tests      Review
     │           │          │
     └───────────┴──────────┴──→ Quality-assured output

Adversarial Mode (Generator/Critic Pattern)

This mode implements the Generator and Critic pattern from Google ADK for iterative refinement:

┌─────────────────────────────────────────────────────────────────────┐
│              Generator/Critic Iteration Cycle                       │
│                                                                     │
│   ┌──────────────────────────────────────────────────────────────┐ │
│   │                    Iteration Loop                             │ │
│   │                                                               │ │
│   │   ┌─────────────┐                                             │ │
│   │   │  Generator  │──────────────┐                              │ │
│   │   │  (Agent A)  │   Proposal   │                              │ │
│   │   └─────────────┘              ▼                              │ │
│   │         ▲              ┌─────────────┐                        │ │
│   │         │              │   Critic    │                        │ │
│   │         │              │  (Agent B)  │                        │ │
│   │         │              └──────┬──────┘                        │ │
│   │         │                     │                               │ │
│   │         │   Feedback          ▼                               │ │
│   │         │              ┌─────────────┐                        │ │
│   │         └──────────────│  Evaluate   │                        │ │
│   │                        │  Quality    │                        │ │
│   │                        └──────┬──────┘                        │ │
│   │                               │                               │ │
│   │                 ┌─────────────┴─────────────┐                 │ │
│   │                 ▼                           ▼                 │ │
│   │         [Quality OK?]              [Max iterations?]          │ │
│   │              │ No                        │ Yes                │ │
│   │              └───── Continue ────────────┴── Exit ──────────▶ │ │
│   └──────────────────────────────────────────────────────────────┘ │
│                                                                     │
│                               ▼                                     │
│                    ┌─────────────────────┐                         │
│                    │   Final Decision    │                         │
│                    │  (User Input Tool)  │                         │
│                    └─────────────────────┘                         │
└─────────────────────────────────────────────────────────────────────┘

Configuration Parameters:

• max_iterations: Maximum number of generate-critique cycles (default: 3)
• quality_threshold: Criteria for acceptable output (e.g., no Critical issues)
• escalate_on_deadlock: Whether to involve user when agents cannot converge

Iteration Cycle:

1. Generate Phase: Generator agent produces a proposal/analysis
2. Critique Phase: Critic agent evaluates and challenges the proposal
3. Evaluate Phase: Check termination conditions:

- Quality threshold met (no Critical/High severity issues remain) - Maximum iterations reached - Agents have converged on consensus

1. Refine or Exit: Either continue with refined proposal or exit to final decision

Example Adversarial Flow:

Iteration 1:
  Generator (Codex/Architect): "Propose microservices architecture"
  Critic (Gemini/Security): "Challenges: Service-to-service auth gaps, data consistency risks"
  Quality: Critical issues found → Continue

Iteration 2:
  Generator: "Refined proposal with OAuth2 service mesh, saga pattern for consistency"
  Critic: "Medium concerns: Observability gaps, no circuit breaker"
  Quality: No Critical issues → Continue (optional refinement)

Iteration 3:
  Generator: "Added distributed tracing, circuit breaker with fallbacks"
  Critic: "Low concerns: Consider rate limiting for external APIs"
  Quality: Acceptable → Exit

Final: Present converged proposal to user for approval

Termination Conditions:

Phase 4: Synthesis

The orchestrator consolidates all results:

1. Collect results from all agents
2. Identify consensus - Points all agents agree on
3. Identify divergence - Points where agents disagree
4. Prioritize actions - Create actionable items with priority
5. Handle conflicts - Use the environment-appropriate user input tool for unresolved disagreements

Personas Reference

🏗️ Architect

Focus Areas:

• Modularity and separation of concerns
• Dependency direction and management
• Extensibility for future changes
• Design pattern appropriateness
• Public API/interface design

Output Format:

• Architecture assessment summary
• Issues (Critical/High/Medium/Low)
• Improvement recommendations
• Diagrams if needed

🔒 Security Researcher

Focus Areas:

• OWASP Top 10 compliance
• Authentication and authorization
• Input validation and sanitization
• Cryptography and data protection
• Error handling information leakage

Output Format:

• Vulnerability summary
• Findings with CVSS scores
• Attack scenarios
• Remediation steps

🧪 QA Engineer

Focus Areas:

• Test case design (unit/integration/e2e)
• Edge cases and boundary conditions
• Regression test needs
• Test coverage gaps
• Security testing requirements

Output Format:

• Test strategy overview
• Required test cases
• Edge cases identified
• Coverage recommendations

👁️ Code Reviewer

Focus Areas:

• Code readability and clarity
• Naming conventions
• Error handling patterns
• Code duplication
• Best practices adherence

Output Format:

• Review summary
• Issues by category
• Specific line-level feedback
• Improvement suggestions

⚡ Performance Engineer

Focus Areas:

• Time complexity analysis
• Memory usage patterns
• N+1 query problems
• Caching opportunities
• Resource management

Output Format:

• Performance assessment
• Bottleneck identification
• Optimization recommendations
• Benchmarking suggestions

🔍 Analyzer

Focus Areas:

• Bug patterns and anti-patterns
• Dead code detection
• Type safety issues
• Null/undefined handling
• Race conditions

Output Format:

• Static analysis results
• Bug risk assessment
• Code smell identification
• Refactoring suggestions

📝 Documentarian

Focus Areas:

• API documentation completeness
• Code comment quality
• README accuracy
• Type definitions
• Usage examples

Output Format:

• Documentation gaps
• Improvement areas
• Template suggestions
• Priority updates

🧠 Domain Expert

Focus Areas:

• Business logic correctness
• Requirements alignment
• Use case coverage
• Domain terminology
• Edge case handling

Output Format:

• Requirements fit analysis
• Business rule verification
• Missing functionality
• Domain-specific recommendations

CLI Command Reference

Codex CLI

# Basic invocation (latest default)
codex exec "prompt"

# With explicit model (if you must pin it)
codex exec --config model='"<latest-codex-model>"' "prompt"

# Reading from file (latest default)
codex exec "Review this code: $(cat src/file.ts)"

Gemini CLI

# Basic invocation (latest default)
gemini -p "prompt"

# With JSON output (latest default)
gemini -p "prompt" --output-format json

# Non-interactive mode (required for scripting)
gemini -p "prompt"

Claude Code (Sub-agent)

For Claude Code, use the Task tool with subagent_type: general-purpose (default/latest model):

Task tool parameters:
  subagent_type: general-purpose
  prompt: "[Persona prompt with task]"
  model: omit to use latest default, or specify if you must pin

For other AI agent CLIs invoking Claude:

# Non-interactive mode
claude -p "prompt" --output-format json

# With tool restrictions
claude -p "prompt" --allowedTools Read,Grep,Glob

# With turn limit
claude -p "prompt" --max-turns 5

Output Template

# 🎭 Multi-AI Collaboration Report

## Executive Summary
[1-2 sentence summary of findings]

## Team Configuration

| Persona | AI Agent | Model | Focus |
|---------|----------|-------|-------|
| 🏗️ Architect | Codex CLI | latest default | Design & Structure |
| 🔒 Security | Gemini CLI | latest default | Vulnerabilities |
| 🧪 QA | Claude (sub) | latest default | Test Design |

**Workflow:** Parallel (Cross-Review)
**Target:** [files/directories]

---

## Agent Results

### 🏗️ Architect (Codex CLI)

**Assessment:** [Overall status]

**Findings:**
1. [Finding] - Severity: [Level]
2. [Finding] - Severity: [Level]

**Recommendations:**
- [Recommendation]

---

### 🔒 Security Researcher (Gemini CLI)

**Assessment:** [Overall status]

**Vulnerabilities:**
| ID | Type | Severity | Location |
|----|------|----------|----------|
| SEC-001 | [Type] | [Severity] | [Location] |

**Remediation:**
- [Steps]

---

### 🧪 QA Engineer (Claude)

**Test Strategy:**
- [Strategy overview]

**Required Tests:**
- [ ] [Test case]

**Edge Cases:**
- [Edge case]

---

## Synthesis

### ✅ Consensus
- [Points all agents agree on]

### ⚠️ Divergence

| Topic | Architect | Security | QA | Resolution |
|-------|-----------|----------|-----|------------|
| [Topic] | [View] | [View] | [View] | [Status] |

### ❓ User Decisions Required

1. **[Decision item]**
   - Agent A recommends: [X]
   - Agent B recommends: [Y]

---

## Priority Actions

### 🔴 Critical (P0)
- [ ] [Action]

### 🟠 High (P1)
- [ ] [Action]

### 🟡 Medium (P2)
- [ ] [Action]

### 🟢 Low (P3)
- [ ] [Action]

---

## Next Steps
1. [Step]
2. [Step]

Usage Examples

Example 1: Cross-Review a Pull Request

User: Review the authentication module changes in this PR

Orchestrator:
1. Identifies target files (src/auth/*)
2. Detects available agents (codex, gemini, claude)
3. Recommends personas: Architect, Security, Code Reviewer

User Input Tool: "Select personas for this review"
User: 1, 2, 4 (Architect, Security, Code Reviewer)

User Input Tool: "Assign agents to personas"
User: Codex→Architect, Gemini→Security, Claude→Reviewer

User Input Tool: "Select workflow mode"
User: 1 (Parallel)

Execution:
- Codex analyzes architecture
- Gemini checks security
- Claude reviews code quality
- Orchestrator synthesizes results

Example 2: Implementation with QA Split

User: Implement user profile feature with tests

Orchestrator:
1. Analyzes requirements
2. Recommends Pipeline mode: Implementer → QA → Reviewer

Execution:
- Orchestrator (Claude) implements feature
- Codex creates comprehensive tests
- Gemini reviews implementation and tests

Example 3: Security Audit

User: Perform security audit on payment module

Orchestrator:
1. Identifies payment-related files
2. Recommends personas: Security, Analyzer, Performance

Execution:
- Gemini (Security): OWASP analysis, CVE search
- Codex (Analyzer): Static analysis, bug patterns
- Claude (Performance): DoS vulnerability, resource limits

Best Practices

1. Start with Parallel mode for unbiased cross-review
2. Use Codex for deep reasoning tasks (architecture, complex bugs)
3. Use Gemini for research tasks (latest vulnerabilities, best practices)
4. Use Claude sub-agents for speed (quick iterations, implementation)
5. Always synthesize divergent opinions - don't just merge results
6. Escalate to user when agents fundamentally disagree
7. Limit personas to 3-4 per task to avoid information overload
8. If 4 personas are selected but only 3 distinct CLIs are available (codex / gemini / claude), assign two personas to Claude sub-agents (spawn the Task tool twice with different persona prompts) — do not drop a persona

Troubleshooting

Agent CLI not found

# Check installation
which codex gemini claude

# Install missing CLIs
# Codex: Follow OpenAI instructions
# Gemini: npm install -g @google/gemini-cli
# Claude: Download from anthropic.com

Agent timeout

• Reduce scope of analysis
• Split into smaller tasks
• Use simpler prompts

Conflicting results

• Use Adversarial mode for deeper analysis
• Escalate to user via the environment-appropriate user input tool
• Document disagreement in report

Environment-Specific Notes

Codex CLI Environment

• Use request_user_input for persona selection and workflow mode
• Use spawn_agent for subagents (latest default model)
• Use exec_command to invoke external CLIs (gemini, claude)

Claude Code Environment

• Use AskUserQuestion tool for user interactions (not "AskUserTool")
• Use Task tool with subagent_type: general-purpose for Claude sub-agents (latest default model)
• Use Bash tool to invoke external CLIs (codex, gemini)
• When invoking external CLIs that need file content, expand with $(cat path/to/file) or read the file via the Read tool first and inline the result; the [CODE_CONTENT] placeholder in Phase 3 examples must be replaced this way

Gemini CLI Environment

• Use numbered prompt options for user selection (no tool calls)
• Use gemini -p directly for execution (latest default model)
• For subagents, invoke other CLIs directly (codex/claude) with latest defaults

Other AI Agent Environments

• Use the platform's equivalent of: user input, subagent, and shell execution tools
• Default to each CLI's latest model unless explicitly pinned