ensemble-claude 0.5.3

AI Orchestration Tool for Claude Code - Multi-agent orchestration for complex development tasks

Ensemble

"One task. Many minds. One result."

AI Orchestration Tool for Claude Code.

Overview

Ensemble is an AI orchestration system that combines the best practices from:

• shogun - Autonomous AI collaboration with tmux parallel execution
• takt - Workflow enforcement with quality gates
• Boris's practices - Effective use of skills, subagents, CLAUDE.md, and hooks

Features

• Autonomous AI Coordination: One instruction triggers multiple AI agents working together
• Flexible Execution Patterns:
  • Pattern A: Simple tasks via subagent
  • Pattern B: Medium tasks via tmux parallel panes (2-4 workers, auto-scaled)
  • Pattern C: Large tasks via git worktree separation
  • Mode T: Research/review tasks via Agent Teams (parallel investigation)
• Parallel Execution Enhancements:
  • Dynamic worker count (auto-scales based on task count)
  • Worker-level subagent parallelization (for 3+ files)
• Parallel Review: Architecture + Security reviews run in parallel
• Self-Improvement: Automatic learning and CLAUDE.md updates
  • Categorized learning (communication, workflow, code quality, tools)
  • Duplicate detection and consolidation
  • Subagent execution result collection
• Compaction Recovery: Built-in protocol to prevent role amnesia
• Extensibility:
  • /create-skill - Generate project-specific skill templates
  • /create-agent - Auto-generate specialized agents from tech stack
• RPI Workflow: Research → Plan → Implement staged workflow for large features
• Hooks Notification: Terminal bell on agent completion (Stop) and errors (PostToolUseFailure)
• Status Line: Real-time display of git branch, session state, worker count
• CLAUDE.md 150-line Limit Check: Pre-commit hook to prevent instruction bloat
• Event-Driven Communication: inbox_watcher.sh + inotifywait for 0ms detection
• Exclusive Locking: flock + atomic write for concurrent YAML protection
• 3-Phase Auto-Escalation: Automatic worker recovery (nudge → Escape+C-c → /clear)
• Loop Detection: LoopDetector + CycleDetector for infinite loop prevention
• Task Dependencies: blocked_by field with DependencyResolver and cycle detection
• NDJSON Session Logging: Structured logging for full session traceability
• Faceted Prompting: 5-concern separation (WHO/RULES/WHAT/CONTEXT/OUTPUT)
• Progressive Disclosure Skills: Dynamic skill injection based on task type
• CI/CD Pipeline Mode: Non-interactive execution with ensemble pipeline
• Bloom's Taxonomy Classification: Cognitive-level-based model selection (L1-L3: sonnet, L4-L6: opus)
• Bottom-Up Skill Discovery: Auto-detect repeated patterns and suggest skill creation

Installation

Using uv (recommended)

# Install globally
uv tool install ensemble-claude

# Or add to your project
uv add ensemble-claude

Using pip

pip install ensemble-claude

From source

git clone https://github.com/ChikaKakazu/ensemble.git
cd ensemble

# Using uv
uv pip install -e .

# Or using pip
pip install -e .

Quick Start

# 1. Initialize Ensemble in your project
ensemble init

# 2. Launch the tmux sessions (2 separate sessions)
ensemble launch

# 3. Open another terminal to view workers session
tmux attach -t ensemble-workers

# 4. Run a task in the Conductor session
/go implement user authentication

# Light workflow (minimal cost)
/go-light fix typo in README

# Create project-specific tools
/create-skill my-feature "Description of the skill"
/create-agent  # Interactive tech stack analysis

CLI Commands

Command	Description
ensemble init	Initialize Ensemble in current project
ensemble init --full	Also copy agent/command definitions locally
ensemble launch	Start 2 tmux sessions (conductor + workers)
ensemble launch --no-attach	Start sessions without attaching
ensemble upgrade	Sync template updates (agents, commands, scripts)
ensemble --version	Show version

In-Session Commands (Conductor)

Command	Description
/go <task>	Full workflow with auto-pattern detection
/go --teams <task>	Force Mode T: Research/review via Agent Teams
/go-light <task>	Lightweight workflow for simple changes
/go-issue [number]	Start implementation from GitHub Issue
/rpi-research <task>	Research phase: requirement analysis, technical investigation, feasibility assessment
/rpi-plan	Plan phase: detailed planning, architecture design, task breakdown
/rpi-implement	Implement phase: execute implementation based on plan (delegates to /go)
/create-skill <name> <desc>	Generate project-specific skill template
/create-agent	Auto-generate specialized agent from tech stack
/review	Run architecture + security review
/improve	Manual self-improvement analysis
/status	View current progress
/deploy	Version bump, merge, and publish to PyPI

Requirements

• Python 3.11+
• Claude Code CLI (claude command available)
• tmux
• git 2.20+ (for worktree support)
• Claude Max plan recommended (for parallel execution)

Agent Architecture

┌─────────────┐
│  Conductor  │ ← Orchestrator (planning, judgment, delegation)
└──────┬──────┘
       │
  ┌────┴────┐
  ▼         ▼
┌────────┐ ┌──────────┐
│Dispatch│ │ Learner  │
└───┬────┘ └──────────┘
    │        ↑ Learning records
    ▼
┌─────────────────────────────┐
│  Reviewer / Security-Reviewer│ ← Parallel reviews
└─────────────────────────────┘
    │
    ▼ (worktree mode)
┌──────────┐
│Integrator│ ← Merge & integrate
└──────────┘

License

MIT License - see LICENSE for details.