capabilities-agent

What the agent does best 1. image_resize — the best. Works right now, no models needed. Resize to exact WB/Ozon dimensions (1080×1440, 3:4), crop, white background, JPG/PNG. It's a deterministic operation: it's either done to spec or not. One second per photo. The best because nothing can go wrong — no GPU, no networks needed. 2. image_upscale — strong, but only with a model. With WITH_TORCH=1 and Real-ESRGAN weights, it’s an honest x4 upscale, quality is noticeable. Without them — LANCZOS + UnsharpMask, which honestly looks kind of “blurry.” 3. doc_extract — fast, but not very valuable. Extracting text from PDF/TXT in fractions of a second. But it’s without a neural network, and there are plenty of free alternatives — people won’t pay much for this. 4. image_restore — the weakest. GFPGAN is good on faces (old photos, blurry ones). But the demand is narrow, and GFPGAN often “repaints” the face — for marketplace product photos, that’s more of a downside.