Multi-agent AI system for autonomous GitHub repository management.
Note: This project is archived and no longer in active development or production use. The codebase is preserved as-is for reference.
Commit-Relay was an autonomous multi-agent platform that managed the entire GitHub repository lifecycle — from task routing to code implementation, security scanning, testing, and deployment — with zero manual intervention.
Agents communicated through structured coordination files rather than direct API calls, creating a fully auditable, transparent orchestration system. A central Coordinator routed incoming tasks to specialized master agents using a Mixture of Experts (MoE) router, which then spawned lightweight workers to execute in parallel.
| Component | Description |
|---|---|
| 5 Master Agents | Coordinator, Development, Security, Inventory, CI/CD |
| 7 Worker Types | Implementation, Fix, Test, Scan, Security Fix, Documentation, Analysis |
| 8+ Autonomous Daemons | Coordinator, Worker Manager, Process Monitor, Heartbeat, Zombie Cleanup, Worker Restart, Failure Detection, Auto-Fix |
| LLM Mesh Gateway | Multi-provider support (Anthropic, OpenAI, Ollama) with circuit breakers, cost tracking, and automatic failover |
| MoE Router v4.0 | 350+ activation keywords, learned weights, semantic routing, 100% routing confidence |
| RAG System | FAISS vector store with 5 collections (code, docs, decisions, patterns, tasks) using sentence-transformers |
| API Server | 128 REST endpoints, WebSocket streaming, rate limiting, authentication |
| MCP Server | Model Context Protocol interface exposing system capabilities as tools |
| Python SDK | Full client library with task orchestration, analytics, health monitoring, and reporting |
| Observability | Elastic APM, LangSmith tracing, 27 event types, distributed tracing, anomaly detection |
- Mixture of Experts router with 350+ keywords and continuous learning
- PyTorch neural routing with training pipeline
- Semantic routing via embeddings (94.5% coverage)
- Margin-based confidence with automatic fallback
- 12+ automated remediation strategies
- Heartbeat monitoring with 2-minute intervals
- Zombie worker detection and cleanup
- Exponential backoff restart logic
- ML-based failure pattern recognition
- Anthropic Claude (primary), OpenAI, and Ollama support
- Cost-aware model selection (simple tasks -> haiku, complex -> opus)
- Circuit breaker middleware with provider health monitoring
- Automatic failover chains with retries
- Token usage and cost analytics
- PII scanning (emails, phone numbers, SSNs, API keys)
- RBAC with permission inheritance across 7 namespaces
- SOC2, GDPR, HIPAA compliance policy checking
- Data quality monitoring with schema validation
- Complete audit trail via file-based coordination
- Elastic Cloud APM with custom spans and business metrics
- LangSmith for LLM performance tracking
- 27 event types with real-time streaming
- Distributed tracing with waterfall visualization
- 50+ system metrics with aggregation
- Pluggable vector store (Weaviate, Qdrant, file-based)
- Connectors for GitHub, Confluence, Slack
- Hybrid search (BM25 + semantic with Reciprocal Rank Fusion)
- 5 specialized collections with metadata schemas
commit-relay/
├── agents/ # Agent configs, prompts, logs, worker outputs
├── api-server/ # Express.js API server (128 endpoints)
├── config/ # System configuration
├── coordination/ # File-based coordination (task queue, worker pool, handoffs)
│ ├── masters/ # Master agent configurations and libraries
│ ├── governance/ # Governance policies and audit logs
│ ├── catalog/ # Data and AI catalog
│ └── observability/ # Event streams and metrics
├── docs/ # 40+ documentation files
├── examples/ # Usage examples
├── lib/ # Shared libraries
│ ├── cache/ # Adaptive LRU cache
│ ├── governance/ # PII scanner, access control, compliance
│ ├── orchestration/ # Workflow engine, SLA monitor, rate limiter
│ └── rag/ # Vector store, embeddings, connectors
├── llm-mesh/ # Multi-provider LLM gateway
├── mcp-server/ # Model Context Protocol server
├── python-sdk/ # Python client library
├── scripts/ # 117+ operational scripts
├── security/ # Security scanning and CVE tracking
└── testing/ # Test suites and test utilities
| Category | Technologies |
|---|---|
| Runtime | Node.js 18+, Python 3.8+, Bash |
| AI/ML | Anthropic Claude, OpenAI, Ollama, PyTorch, sentence-transformers, FAISS |
| API | Express.js 5, WebSocket, JSON-RPC 2.0 (MCP) |
| Observability | Elastic APM, LangSmith, OpenTelemetry |
| Security | Helmet, express-rate-limit, JWT, PII scanning |
| Data | FAISS, Weaviate, Qdrant, MiniSearch |
| Deployment | Docker, systemd/launchd |