From Code to Action: Hierarchical Learning of Diffusion-VLM Policies