Orchestra CLI

October 2, 2025 Β· View on GitHub

3-Depth Agent Orchestration System powered by Claude Code

Orchestra CLI is a hierarchical agent orchestration system with three distinct layers (L1/L2/L3) designed for safe, observable, and efficient task automation.

πŸš€ Quick Start

# Install dependencies
pnpm install

# Build
pnpm build

# Run (development)
pnpm dev

# Execute a task
orchestra run "Create README and lint" --plan-only

πŸ“š Documentation

Full documentation is available in multiple languages:

Key Documents

πŸ—οΈ Project Structure

orchestra-cli/
β”œβ”€β”€ src/
β”‚   β”œβ”€β”€ cli/          # CLI entry point
β”‚   β”œβ”€β”€ config/       # Configuration loader & schema
β”‚   β”œβ”€β”€ logger/       # JSONL logger
β”‚   β”œβ”€β”€ l1/           # L1 Orchestrator
β”‚   β”œβ”€β”€ l2/           # L2 Coordinator
β”‚   β”œβ”€β”€ l3/           # L3 Worker
β”‚   └── types/        # TypeScript type definitions
β”œβ”€β”€ tests/
β”‚   β”œβ”€β”€ unit/         # Unit tests
β”‚   β”œβ”€β”€ integration/  # Integration tests
β”‚   └── fixtures/     # Test scenarios
β”œβ”€β”€ config/           # Default configuration
└── docs/             # Documentation (EN/KO)

πŸ› οΈ Tech Stack

  • Language: TypeScript
  • Runtime: Node.js 20+
  • Package Manager: pnpm
  • LLM Backend: Claude Code (Sonnet 4.5)
  • Testing: Vitest
  • Linting: ESLint + Prettier

πŸ“‹ Features

  • 3-Layer Architecture: L1 (Orchestrator), L2 (Coordinator), L3 (Workers)
  • Security-First: Whitelist-based tool access, filesystem/network sandboxing
  • Observable: JSONL logs, trace ID propagation, markdown reports
  • Safe Execution: --plan-only and --dry-run modes
  • LLM-Powered Planning: Claude Code generates intelligent task decomposition
  • Retry & Recovery: Exponential backoff, partial failure handling

πŸ§ͺ Development

# Install dependencies
pnpm install

# Run type checking
pnpm typecheck

# Run linter
pnpm lint

# Format code
pnpm format

# Run tests
pnpm test

# Run tests with coverage
pnpm test:coverage

πŸ“¦ Available Commands

orchestra run <task> [options]     # Execute a task
orchestra agent ls                 # List agents
orchestra agent inspect <level>    # Inspect agent details
orchestra eval run <scenario>      # Run evaluation

Options

  • --plan-only - Generate plan only, do not execute
  • --dry-run - Simulate execution without running tools
  • --config <path> - Configuration file (default: ./orchestra.config.yaml)
  • --concurrency <n> - Number of parallel workers (default: 1)
  • --retries <n> - Retry count override

πŸ—ΊοΈ Roadmap

  • Day 1-2: Project scaffolding and basic infrastructure
  • Day 3-4: Logger, CLI parser, run directory management
  • Day 5-7: L2 Coordinator (LLM adapter, DAG generation)
  • Day 8-9: L3 Worker (tool execution, retry)
  • Day 10: L1 policy engine, report generation
  • Day 11-12: Testing (unit + integration)
  • Day 13-14: Debugging, polish, MVP release

See Next Steps for detailed roadmap.

🀝 Contributing

We welcome contributions! Please see our Contributing Guidelines for details.

πŸ“„ License

MIT License - see LICENSE for details.


Built with ❀️ using TypeScript + Claude Code