Multi-Agent Orchestrator

Intelligently coordinate multiple AI agents in parallel for complex tasks. Auto-analyzes requirements, selects optimal agent combinations, and synthesizes results.

Features

✅ Smart Routing — Analyzes task keywords to select 1-3 best agents ✅ Parallel Execution — Runs agents simultaneously, not sequentially ✅ Silent Failure Handling — Continues with available results if an agent fails ✅ 4 Operating Modes — Auto, Review, Challenge, Consult ✅ Synthesized Reports — Combines multiple perspectives into decisive recommendations ✅ Multi-Language Support — Understands English, 中文, 日本語, 한국어

Quick Start

Default (Auto Mode)

User describes task → Claude auto-routes to best agents:

User: "審查這個認證流程，有沒有安全漏洞"
Claude: 分析 → 選擇 codex + cline (並行)
Output: Cross-review report

Explicit Invocations

Review Mode — Code cross-review:

/agents review

Challenge Mode — Adversarial debate:

/agents challenge

Consult Mode — Specify agents:

/agents consult kilo,gemini

Available Agents

| Agent | Specialty | |-------|-----------| | codex | Code review, refactoring, security, tests | | gemini | Architecture, design patterns, broad knowledge | | kilo | Logic, system design, complex reasoning | | cline | Implementation details, structured analysis | | opencode | Rapid prototyping, code completion, automation | | qoder | Information retrieval, documentation, knowledge base |

Task Classification

Automatically detects and routes:

• Code Review → codex + gemini
• Feature Implementation → cline + opencode
• Architecture Design → kilo + gemini
• Analysis → gemini + kilo
• Prototyping → opencode + cline
• Knowledge Lookup → qoder + gemini
• Security Audit → codex + cline
• Debugging → codex + kilo

Installation

Project Level

# Already installed at:
# ~/YD 2026/.claude/skills/agents/

Global Installation (via triple-publish)

cd ~/YD\ 2026/.claude/skills/agents/
/triple-publish  # Deploy to GitHub + npm + PyPI

Or manually:

# Copy to global skills
cp -r ~/YD\ 2026/.claude/skills/agents/ ~/.claude/skills/agents/

Output Format

All modes produce structured reports:

## 🤖 Multi-Agent Report

**Task**: [description]
**Mode**: Auto | Review | Challenge | Consult
**Agents Called**: codex, gemini
**Duration**: ~45s

### Agent 1 [PASS|CONCERNS|BLOCK]
...

### Agent 2 [PASS|CONCERNS|BLOCK]
...

### Synthesis
**Consensus**: ...
**Disagreement**: ...
**Blind Spots**: ...
**Conclusion**: [PASS | APPROVE_WITH_CHANGES | BLOCK]
**Claude's Take**: ...

Auto-Trigger Scenarios

The orchestrator automatically uses multiple agents when:

| Scenario | Trigger | Agents | |----------|---------|--------| | Code review / PR | Large changes (>100 lines) | codex + gemini | | Architecture decision | Major refactor, new pattern | gemini + kilo | | Security-sensitive changes | auth, crypto, secrets | codex + cline | | Debugging | Stuck for 3+ turns | codex or kilo | | User requests "second opinion" | Explicit mention | Best 1-2 agents |

Reliability

Constraints Respected

• ✅ Rate limit awareness (skip qoder on frequent runs)
• ✅ Context efficiency (keep prompts <2000 chars)
• ✅ Timeout protection (60s per agent, 90s total)
• ✅ Budget management (qoder: 1-2 runs/week)

Failure Handling

• 1 agent fails → Use available agent's output + note limitation
• All agents fail → Report reason (timeout, rate limit, network)
• Never silently drop results

Configuration

Enable in Claude Code

The skill is available by default. To verify:

# Check installation
cat ~/.claude/skills/agents/SKILL.md | head -20

# Or in Claude Code
/agents consult gemini "what's your favorite algorithm?"

Per-Project Settings

Add to .claude/settings.json:

{
  "enabledPlugins": {
    "agents-skill@builtin": true
  }
}

Examples

Example 1: Auto Code Review

User: "Help me review this authentication module"
→ Auto-detects: code review task
→ Routes to: codex (code quality) + gemini (design)
→ Output: Combined report with security & architecture insights

Example 2: Challenge Mode

User: "/agents challenge Should we use microservices?"
→ Auto-selects: gemini (pro) vs kilo (con)
→ Output: Debate exposing trade-offs and blind spots

Example 3: Consult Specific Agents

User: "/agents consult kilo,cline Design this database schema"
→ Routes only to: kilo + cline
→ Output: Focused analysis on both architectural & implementation aspects

Troubleshooting

Agent not responding

• Check if agent is installed: <agent> --version
• Check authentication: run agent login script
• Check rate limits: some agents have weekly quotas
• Wait a moment and retry

Slow execution

• Parallel execution normally takes 30-90s total
• If >120s, one agent likely timed out (check /tmp logs)
• Reduce task complexity for faster response

Missing results

• Check if required agent is available: /agents consult qoder test
• Some agents have budget limits (qoder: 1-2/week)
• Skip unavailable agents and use others' results

Architecture

SKILL.md
├── Mode 1: Auto (keyword matching → agent selection)
├── Mode 2: Review (codex + gemini parallel)
├── Mode 3: Challenge (adversarial debate)
└── Mode 4: Consult (manual agent selection)
    └── Parallel Bash Execution
        ├── Silent failure handling
        ├── Result aggregation
        └── Synthesis & reporting

Future Enhancements

• [ ] Learning from past runs (which agents usually agree/disagree)
• [ ] Dynamic budget allocation (adjust agents based on available quota)
• [ ] Streaming output (show results as agents finish)
• [ ] Custom agent combinations via YAML config
• [ ] Integration with CI/CD for automatic code review

Version History

v1.1.1 (2026-03-27)

• Renamed GitHub repo to orchestrator-agents-skill
• Updated all URLs and references to new repo name
• All 6 agent CLIs upgraded to latest versions

v1.1.0 (2026-03-27)

• Independent repo release with full triple-publish (GitHub + npm + PyPI)
• Synced local skill directory with all project files
• Documentation and version consistency improvements

v1.0.0 (2026-03-26)

• Initial release
• 4 modes: Auto, Review, Challenge, Consult
• Support for 6 agents (codex, cline, gemini, kilo, opencode, qoder)
• Parallel execution with silent failure handling
• Multi-language keyword matching

License

MIT — See LICENSE file

Contributing

This skill is maintained as part of YD 2026 Project. To contribute:

1. Fork from ~/.claude/skills/agents/
2. Test thoroughly (all 4 modes)
3. Submit changes via standard PR workflow
4. Deploy via /triple-publish when approved

Ready to use: Just describe your task and Claude will automatically coordinate the right agents!