orchestration

name	orchestration
description	Parallel execution using multiple Claude instances in Kitty terminal
allowed-tools	["Read","Glob","Grep","Bash","Task"]
context	fork
user-invocable	true
version	1.0.0

Multi-Worker Parallel Orchestration Skill

Overview

Parallel execution of plans using Claude and/or Copilot CLI workers in Kitty terminal tabs. Supports mixed-engine mode where Claude handles complex tasks and Copilot handles mechanical ones.

When to Use

• Complex tasks with 4+ independent subtasks
• Release preparations with multiple verification steps
• Large refactoring across multiple file domains
• Any plan created by strategic-planner agent

Requirements

Requirement	Details
Terminal	Kitty only (not Warp/iTerm/Terminal.app)
Config	allow_remote_control yes in kitty.conf
Claude	wildClaude='claude --dangerously-skip-permissions'
Copilot	copilot CLI + GH_TOKEN or COPILOT_TOKEN env var
Max Workers	4 (hard limit)

Quick Start

# Claude-only (default):
orchestrate.sh <plan> 4

# Copilot-only (all workers use copilot --yolo):
orchestrate.sh <plan> 4 --engine copilot

# Mixed (Claude for complex, Copilot for codex:true tasks):
orchestrate.sh <plan> 4 --engine mixed

# Monitor (detects both Claude-N and Copilot-N tabs):
claude-monitor.sh

# Launch single worker by type:
worker-launch.sh claude "Claude-2" <task_db_id> --cwd /worktree
worker-launch.sh copilot "Copilot-3" <task_db_id> --cwd /worktree

# Standalone Copilot worker (no Kitty needed):
copilot-worker.sh <task_db_id> --model claude-sonnet-4-5 --timeout 600

Integration with Agents

strategic-planner

The strategic-planner agent can create plans with Claude assignments and execute them in parallel:

@strategic-planner Create an execution plan for [task] with parallel execution

When asked "Vuoi eseguire in parallelo?", it will:

1. Verify Kitty environment
2. Launch Claude workers
3. Send tasks to each worker
4. Monitor progress
5. Report completion

Plan Format

Plans for parallel execution must include:

## 🎭 RUOLI CLAUDE

| Claude   | Role        | Tasks      | Files           |
| -------- | ----------- | ---------- | --------------- |
| CLAUDE 1 | Coordinator | Monitor    | -               |
| CLAUDE 2 | Implementer | T-01, T-02 | src/api/        |
| CLAUDE 3 | Implementer | T-03, T-04 | src/components/ |
| CLAUDE 4 | Implementer | T-05       | src/lib/        |

Critical Rules

1. NO FILE OVERLAP - Avoid git conflicts
2. MAX 4 WORKERS - Beyond = chaos
3. VERIFY LAST - lint/typecheck/build at end
4. GIT COORDINATION - One commit at a time
5. WORKTREE MANDATORY - Every worker runs worktree-guard.sh first. NEVER on main.
6. COPILOT: --yolo - Always use copilot --yolo for full autonomy (no confirmation prompts)

Copilot CLI Flags Reference

Flag	Purpose
--yolo	Full autonomy, no confirmations (REQUIRED)
--add-dir <path>	Trust worktree directory
--model <model>	Select model (claude-sonnet-4-5, gpt-4.1, etc.)
-p "prompt"	Non-interactive mode (for scripted execution)

Scripts

Script	Purpose
orchestrate.sh	Main orchestrator (--engine claude|copilot|mixed)
worker-launch.sh	Launch single worker in Kitty tab
copilot-worker.sh	Standalone Copilot task execution
copilot-task-prompt.sh	Generate Copilot-compatible prompt from DB task
worktree-guard.sh	Block execution if not in correct worktree
claude-monitor.sh	Monitor all worker tabs (Claude + Copilot)
kitty-check.sh	Verify Kitty setup

Related

• Agent: .claude/agents/core_utility/strategic-planner.md
• Global config: ${CLAUDE_HOME:-.claude}/commands/planner.md
• Drift check: plan-db.sh drift-check <plan_id>