ccswarm: AI Multi-Agent Orchestration System

February 3, 2026 · View on GitHub

Version 0.4.3 - Rust-Native Multi-Agent Orchestration with ai-session Integration

CI Rust License: MIT

ccswarm is a workflow automation framework for coordinating specialized AI agents using Claude Code CLI. It provides task delegation infrastructure, template-based scaffolding, and Git worktree isolation for parallel development.

AI Integration: ccswarm uses native PTY sessions via ai-session crate. Multi-provider system exists but orchestrator coordination loop is not fully implemented. See docs/UPCOMING_FEATURES.md for roadmap.

Implementation Status

CategoryStatusDescription
CLI Infrastructure✅ WorkingAll commands parse and route correctly
Session Management✅ WorkingNative PTY-based sessions
TUI Monitoring✅ WorkingReal-time terminal UI with ratatui
Configuration✅ WorkingProject configs, agent settings
Git Worktrees✅ WorkingCreate, list, remove, prune
Template System✅ WorkingPredefined templates for app scaffolding
Task Queue✅ WorkingTask queuing and tracking
AI Execution⚠️ SimulatedReturns keyword-based responses
Provider System🚧 PlannedCode exists, not integrated
ACP Integration🚧 StubCLI wrapper only, no WebSocket
Parallel Executor⚠️ PartialStructure exists, not integrated with orchestrator
start Command⚠️ PartialInitializes but coordination loop not implemented
Auto-Create⚠️ PartialTemplates work, full AI generation incomplete
Sangha (Voting)🚧 PlannedData structures only
Extensions🚧 PlannedStub implementation

Note: See docs/analysis/ for detailed capability gap analysis.

Documentation

DocumentDescription
Getting StartedInstallation, tutorial, first project
ArchitectureSystem design, components, data flow
ConfigurationAll configuration options with examples
Commands ReferenceComplete CLI command reference
Claude ACP GuideClaude Code integration setup
TroubleshootingCommon issues and solutions
Upcoming Featuresv0.4.0 implementation patterns
Application SpecDetailed feature specifications
ContributingHow to contribute

Key Features

Status Legend: ✅ Working | ⚠️ Partial | 🚧 Planned

Multi-Provider AI Integration (🚧 Planned)

  • Provider Architecture: 5 providers implemented (ClaudeCode, Aider, ClaudeAPI, Codex, Custom)
  • Current Execution: Simulated responses (keyword-based matching)
  • ACP Module: Stub implementation (CLI wrapper, no WebSocket)
  • Status: Code exists but not integrated with orchestrator
  • See docs/UPCOMING_FEATURES.md for integration roadmap

Developer Experience (✅ Working)

  • Interactive Setup Wizard: Guided configuration for new users
  • Built-in Tutorial: Learn by doing with hands-on chapters
  • Smart Error Messages: Helpful suggestions and solutions
  • Progress Indicators: Real-time feedback for operations
  • Contextual Help: ccswarm help <topic> for detailed guidance
  • System Doctor: ccswarm doctor --fix diagnoses and fixes issues

Session Management (✅ Working)

  • Native PTY Sessions: Cross-platform terminal sessions (no tmux)
  • Session Lifecycle: Create, list, attach, detach, pause, resume
  • Auto-Recovery: Automatic reconnection and state persistence
  • Resource Monitoring: Track session usage and statistics

Template System (✅ Working)

  • Predefined Templates: Rust CLI, security review, performance optimization
  • Variable Substitution: Dynamic content generation with context
  • Category Organization: Application, utility, review, optimization
  • Custom Templates: Create and store project-specific templates
  • Validation System: Type-safe template validation

Git Worktree Isolation (✅ Working)

  • Parallel Development: Multiple branches simultaneously without conflicts
  • Per-Agent Workspaces: Isolated environments for each agent role
  • Automatic Management: Create, list, remove, prune operations
  • Safe Isolation: No cross-agent file access or conflicts

Task Queue & Tracking (✅ Working)

  • Task Management: Create, queue, and track tasks
  • Priority Levels: High, medium, low task prioritization
  • Status Tracking: Pending, in-progress, completed states
  • Task Metadata: Tags, descriptions, dependencies

Orchestrator & Delegation (⚠️ Partial)

  • ProactiveMaster: Intelligent task analysis (structure exists)
  • Agent Assignment: Optimal agent selection (basic routing works)
  • Coordination Loop: Continuous monitoring (not implemented)
  • Dependency Resolution: Auto task ordering (planned)

Parallel Execution (⚠️ Partial)

  • ParallelExecutor: Multi-agent concurrency structure exists
  • Message Bus: Inter-agent communication framework ready
  • Resource Management: File locking and conflict prevention designed
  • Integration: Not wired to orchestrator yet

Auto-Create System (⚠️ Partial)

  • Template Scaffolding: Generate project structure (working)
  • Variable Substitution: Dynamic content generation (working)
  • AI-Driven Generation: Natural language to app (incomplete)
  • Custom Templates: User-defined project types (working)

Observability & Tracing (✅ v0.4.3)

  • Span Tracking: Trace agent execution across workflows
  • Token Usage Metrics: Monitor and optimize LLM API costs
  • Trace Visualization: Hierarchical span trees for debugging
  • OpenTelemetry Export: Jaeger, Zipkin integration (file-export ready)
  • Langfuse Integration: LLM-specific observability (file-export ready)

Human-in-the-Loop (✅ v0.4.3)

  • Approval Workflows: Gate critical actions with human oversight
  • Policy-Based Rules: Define approval requirements by risk level
  • CLI Notifications: Interactive approval prompts
  • Timeout Handling: Automatic escalation on timeout
  • Audit Trail: Complete history of approval decisions
  • Multi-Channel: Slack/Email integration (planned)

Long-term Memory & RAG (✅ v0.4.3)

  • Short/Long-term Memory: Session-aware memory consolidation
  • Retrieval Augmented Generation: Context-aware responses
  • Importance Retention: Smart memory decay and prioritization
  • In-Memory Backend: Fast access for current session
  • Vector Embeddings: Semantic search (planned)
  • Persistent Storage: File/vector DB backends (planned)

Graph Workflow Engine (✅ v0.4.3)

  • DAG Workflows: Define task dependencies as graphs
  • Conditional Execution: Dynamic workflow paths
  • Parallel Tasks: Run independent tasks concurrently (structure ready)
  • Approval Gates: HITL integration at checkpoints (planned)
  • Sub-workflows: Compose complex workflows (planned)

Benchmark Integration (✅ v0.4.3)

  • SWE-Bench Style: Standardized agent performance testing
  • Predefined Suites: Basic coding, bug fixes, refactoring
  • Metrics Collection: Track pass rates, scores, improvements
  • Leaderboard System: Compare agent performance over time
  • Custom Benchmarks: Create project-specific evaluation suites

Collective Intelligence (🚧 Planned)

  • Sangha System: Democratic decision-making (data structures only)
  • Self-Extension: Agents propose improvements (stub implementation)
  • Experience Learning: Continuous introspective analysis (planned)
  • Consensus Algorithms: Smart proposal voting (planned)
  • Safe Implementation: Risk assessment and rollback (planned)

In Development (v0.4.0)

See docs/UPCOMING_FEATURES.md for detailed implementation patterns and integration guides.

Features with code already in codebase but not fully integrated:

  • 🔧 Hook System Integration - Extensible execution hooks with priority registry
  • 🔧 Verification Agent - Auto-verify applications with 6-check workflow
  • 🔧 DynamicSpawner - Intelligent workload balancing for agent selection
  • 🔧 Parallel Execution - True multi-agent parallelism (command & PTY-based)
  • 🔧 ai-session MessageBus - Inter-agent coordination with 93% token savings
  • 🔧 Session Persistence - Resume/fork/checkpoint with crash recovery

Roadmap:

PhaseTaskPriorityStatus
1Fix start command coordination loopCriticalNot Started
1Wire ParallelExecutor to orchestratorCriticalNot Started
1Implement inter-process communication (Unix socket/SQLite)CriticalNot Started
2Integrate ai-session MessageBus for agent coordinationHighNot Started
2Enable context compression (93% token savings)HighNot Started
2Make ACP optional (--enable-acp flag)MediumNot Started
3Multi-vendor agent support (Gemini CLI, Codex CLI, OpenCode)MediumPlanned
3Sangha collective decision-makingLowPlanned

Note: ACP (Agent Client Protocol) will be optional. PTY/CLI remains the primary execution method. See Issue #67 for discussion.

Quick Start

New to ccswarm? See the Getting Started Guide for a complete walkthrough.

Installation

git clone https://github.com/nwiizo/ccswarm.git
cd ccswarm
cargo build --release
cargo install --path crates/ccswarm

First Project

# Interactive setup
ccswarm setup

# Or quick init
ccswarm init --name "MyProject" --agents frontend,backend

# Start system
ccswarm start

# Launch TUI (in another terminal)
ccswarm tui

# Create a task
ccswarm task "Create login form [high] [feature]"

# Auto-create complete app
ccswarm auto-create "Create TODO app with React" --output ./todo-app

Sample Demos

cd sample/
./setup.sh
./claude_acp_demo.sh      # Claude Code integration
./task_management_demo.sh  # Task management
./multi_agent_demo.sh     # Multi-agent collaboration

Architecture

┌─────────────────────────────────────────┐
│         ProactiveMaster                 │ ← Type-State Pattern
│     ├─ Channel-Based Orchestration     │   Zero shared state
│     ├─ Task Analysis & Delegation      │   Pattern matching
│     └─ Quality Review Integration      │   Async/await
├─────────────────────────────────────────┤
│     Multi-Provider Integration          │ ← Planned (v0.4.0)
│     ├─ Provider System (5 impl.)       │   Code exists, not wired
│     └─ Current: Simulated Execution    │   Keyword-based responses
├─────────────────────────────────────────┤
│     Specialized Agent Pool              │ ← Actor Model
│     ├─ Frontend Agent (React/Vue/UI)   │
│     ├─ Backend Agent (APIs/Database)   │
│     ├─ DevOps Agent (Docker/CI/CD)     │
│     └─ QA Agent (Testing/Quality)      │
├─────────────────────────────────────────┤
│     Template System                     │ ← Variable substitution
├─────────────────────────────────────────┤
│     Git Worktree Manager                │ ← Isolated development
├─────────────────────────────────────────┤
│     Real-time Monitoring (TUI)          │ ← Crossterm-based
└─────────────────────────────────────────┘

Rust-Native Patterns

  • Type-State Pattern: Compile-time state validation with zero runtime cost
  • Channel-Based Orchestration: Message-passing without shared state or locks
  • Iterator Pipelines: Zero-cost abstractions for task processing
  • Minimal Testing: Only 8 essential tests covering core functionality
  • No Arc: Replaced with actor model and channels

Project Structure

ccswarm/
├── crates/
│   └── ccswarm/           # Main orchestration crate
│       ├── src/
│       │   ├── acp_claude/  # Claude ACP integration
│       │   ├── cli/         # CLI commands
│       │   ├── orchestrator/ # ProactiveMaster
│       │   ├── agent/       # Agent types
│       │   ├── session/     # Session management
│       │   ├── template/    # Template system
│       │   ├── subagent/    # Parallel execution
│       │   ├── tui/         # Terminal UI
│       │   └── resource/    # Resource monitoring
│       └── tests/
├── docs/                  # Documentation
└── sample/                # Demo scripts

Known Limitations

LimitationImpactWorkaround
start command exits immediatelyNo continuous orchestrationUse TUI or individual commands
ParallelExecutor not wiredNo true parallel agent executionTasks run sequentially
ACP WebSocket not functionalNo real-time bidirectional commsUses CLI wrapper instead
ai-session MessageBus unusedNo inter-agent coordinationAgents work independently
No IPC between processesCommands are isolatedEach command is standalone
macOS/Linux onlyNo Windows supportUse WSL on Windows

See docs/analysis/00-capability-gap-analysis.md for detailed analysis.

Contributing

cargo test --workspace
cargo fmt --all
cargo clippy --workspace -- -D warnings

See CONTRIBUTING.md for guidelines.

Standalone Deployment

ccswarm can run without Claude Code! See STANDALONE_DEPLOYMENT.md for:

  • Simulation mode for testing
  • Built-in templates without AI
  • Docker deployment
  • Custom providers

License

MIT License - see LICENSE for details.

Acknowledgments

Experience Claude Code integration with autonomous AI orchestration in ccswarm v0.4.3