A streaming action language for computer-use agents. LLMs emit one JSON action per line; the executor runs each the moment its newline arrives — while the model is still generating.