neohotaru

Let a DeepSeek Harness agent actually see an animated GIF or animated WebP: decode every frame, sample the animation across time, and hand the model a labelled sequence of real frames instead of the single first frame the harness otherwise delivers.