comfyui-api

This skill teaches an agent how to talk to a running ComfyUI instance over its REST API so that image and video generation workflows can be queued, monitored, and collected without leaving the assistant. It solves the practical gap between having a ComfyUI node graph and actually executing it: connecting to the server (default `http://127.0.0.1:8188`), queuing a workflow, waiting for it to finish, and pulling the resulting files back down. It also defines an offline fallback that exports workflow JSON to disk for manual drag-and-drop loading when no server is available.

The documented operations are all standard ComfyUI HTTP endpoints driven with curl: check server status via `GET /system_stats`, submit a workflow with `POST /prompt`, poll `GET /history/{prompt_id}` every five seconds for completion, download outputs from `GET /view`, upload reference images to `POST /upload/image`, interrupt the current run, and free VRAM. It includes a polling strategy with a 10-minute warning threshold (video generation can take much longer), a workflow validation step that cross-checks node class_types and model filenames against a cached inventory, and an error-handling table mapping connection refusals and HTTP errors to remediation steps.

It is aimed at people building ComfyUI automation, generative-media agents, or batch pipelines who want a reliable programmatic bridge to their own local or self-hosted ComfyUI server. It is part of a larger "comfyui-expert" skill set and defers to sibling skills (comfyui-inventory, comfyui-troubleshooter) for discovery and error diagnosis.