| name | orchestrating-agents |
| description | Orchestrates task handoff and "handshaking" between Gemini and Qwen CLI agents. Facilitates specialized reviews, second opinions, and cross-validation of complex logic. Use when a task requires multi-model collaboration, a second LLM perspective, or iterative feedback between terminal agents. |
| gemini-command | orchestrate |
| gemini-args | [{"name":"workflow","description":"Orchestration workflow type","type":"choice","choices":["[Truncated]","[Truncated]","[Truncated]","[Truncated]"],"default":"collaborative"},{"name":"prompt","description":"Task description for orchestration","type":"string","required":true}] |
| gemini-prompt | You are orchestrating a multi-agent collaboration workflow.
Workflow Type: {{workflow}}
Task: {{prompt}}
Steps:
1. Detect available neighbor agents (gemini, qwen)
2. Map workflow type to execution pattern:
- collaborative → Collaborative Design (3 turns)
- adversarial → Adversarial Review (3 turns)
- troubleshoot → Troubleshoot Session (4 turns)
- handshake → Single Handshake (1 turn)
3. Execute the turn protocol defined in references/workflows.md
4. Synthesize and present results
|
Orchestrating Agents
This skill provides a structured workflow for multi-model collaboration between Gemini and Qwen, enabling deep "handshaking" sessions for complex tasks.
Usage
Command-Line Invocation (Gemini)
/orchestrate collaborative "Design authentication system with JWT"
/orchestrate adversarial "Review payment processing security"
/orchestrate troubleshoot "Debug production memory leak"
/orchestrate handshake "Quick review of rate limiting logic"
/orchestrate "Design authentication system"
Workflow Types
| Workflow | Turns | Use Case |
|---|
| collaborative | 3 | New features, architecture planning, design decisions |
| adversarial | 3 | Security audits, vulnerability assessment, red-team review |
| troubleshoot | 4 | Complex debugging, production errors, emergency fixes |
| handshake | 1 | Quick second opinion, simple validation |
Interactive Selection (Fallback)
If no workflow type is specified, the agent will prompt for selection:
For Gemini CLI (ask_user)
ask_user({
questions: [{
question: "This task requires deep collaboration. Which orchestration workflow should we use?",
header: "Orchestrate",
options: [
{ label: "Collaborative Design", description: "Agent A proposes -> Agent B critiques -> Agent A refines." },
{ label: "Adversarial Review", description: "Agent A proposes -> Agent B attacks (Red Team) -> Agent A defends." },
{ label: "Troubleshoot Session", description: "Multi-agent hypothesis testing for emergency bugs." },
{ label: "Single Handshake", description: "Standard one-turn second opinion." }
]
}]
});
For Claude Code (AskUserQuestion)
AskUserQuestion({
questions: [{
question: "This task requires deep collaboration. Which orchestration workflow should we use?",
header: "Orchestrate",
options: [
{ label: "Collaborative Design", description: "Agent A proposes -> Agent B critiques -> Agent A refines." },
{ label: "Adversarial Review", description: "Agent A proposes -> Agent B attacks (Red Team) -> Agent A defends." },
{ label: "Troubleshoot Session", description: "Multi-agent hypothesis testing for emergency bugs." },
{ label: "Single Handshake", description: "Standard one-turn second opinion." }
]
}]
});
Workflows
1. Handshake Protocol (Standard)
| Agent | Initial Command | Follow-up (Resume) |
|---|
| Gemini | gemini -p "..." | gemini -r latest -p "..." |
| Qwen | qwen "..." | qwen -c "..." |
2. Multi-Turn Loops
For complex scenarios, follow the turns defined in references/workflows.md.
Loop Progress:
- [ ] Turn 1: Proposal (Agent A)
- [ ] Turn 2: Cross-Examination (Agent B)
- [ ] Turn 3: Final Synthesis (Agent A)
Resources
Best Practices
- Concise Context: Send only relevant snippets, not entire files, to the neighboring agent.
- Specific Tasks: Ask for a "Security Audit" or "Refactoring Suggestions" rather than a general review.
- Fail-Safe: If a neighboring agent is not found, proceed with a self-review but notify the user.
