Figure Generation

It expands your description into specs, generates matplotlib code with up to 4 retry attempts on errors, then uses visual feedback to refine the output. Handles the usual suspects: bar charts, training curves, heatmaps, ablation studies. Outputs both 300 DPI PNG and vector PDF, plus the LaTeX boilerplate to drop it in your paper. The quality checks are sensible, enforcing colorblind-friendly palettes and readable font sizes at print scale. Honestly, the VLM feedback loop is the interesting bit here. Most figure tools stop at code generation, but having it actually look at the rendered output and fix labeling or scaling issues saves the manual iteration you'd normally do.