What problem does it solve? Running and observing ComfyUI workflows programmatically requires juggling the POST /prompt endpoint, the /ws event stream, and /history lookups, plus decoding two distinct error channels. This Skill wraps that entire control surface in standalone scripts so an agent can submit, monitor, interrupt, and partially re-execute prompts without hand-writing protocol code. ## Core Features & Use Cases - Prompt submission and targeted reruns: Submit prompt JSON via POST /prompt, including partial_execution_targets to re-run only specific output nodes during repair loops. - Live execution monitoring: Subscribe to the /ws WebSocket and stream typed events (executing, progress, executed, execution_error, execution_interrupted) as JSON lines until a terminal frame. - Result and schema retrieval: Fetch outputs from /history/{prompt_id} and query /object_info node schemas with a filesystem cache, plus interrupt running prompts and free VRAM. - Use Case: An agent submits a txt2img prompt, watches the WebSocket until the graph finishes, pulls the generated images from history, and—on an execution_error—feeds the node_id, exception_message, and current_inputs triple into a repair loop that resubmits only the affected output node. ## Quick Start Ask the agent to submit your ComfyUI prompt JSON, watch its execution events, and return the output images from history.
