Router: fast tasks (classify, tag) run local-first; deep shaping runs cloud-first; both fall back to the other, then to built-in patterns. Every shaped plan is stamped with its true source — local 4B, cloud, or template. Keys are never written to disk. Any OpenAI-compatible endpoint works (Ollama, LM Studio, vLLM, gateways). If a local engine refuses this file, start it with OLLAMA_ORIGINS=* .