Local model plus cloud supervisor doubles speed

A developer replaced Claude Code with a split workflow using GPT-5.6 Terra as a cloud supervisor and Qwen3-Coder-30B-A3B as a local model. The setup completes fixes and smaller features in roughly half the time while using far fewer cloud tokens.