- Implement SSE mode for MCP server (mcp/skills.py)
- Add MCP service to docker-compose.yml on port 3000
- Add uvicorn dependency to mcp/requirements.txt
- Create SETUP.md, USAGE.md, OPENCODE-MCP.md
- Update README with quick links and MCP section
- Remove semantic cache references throughout
- Add cross-platform Python MCP setup script to template repo
- semantic_cache.py: Semantic similarity matching for cache hits
- rag.py: RAG-based context selection with local embeddings
- compression.py: Conversation history summarization
- New endpoints: /cache/semantic-lookup, /cache/semantic-store, /context/rag, /compress
- Uses sentence-transformers (all-MiniLM-L6-v2) - no external API calls
- No vector DB needed - cosine similarity on small datasets is fast enough
- Expected savings: 50-70% token reduction