What problem does it solve? When a ComfyUI prompt fails validation or crashes mid-execution, re-running the entire workflow wastes time and compute. This Skill classifies the failure into a concrete error class and computes the smallest set of nodes that must be re-executed, so only the broken subgraph runs again. ## Core Features & Use Cases - Error classification: Reads node_errors from POST /prompt responses or execution_error WebSocket frames and assigns one of four classes: link, parameter, default, or missing_resource. - Deterministic repair recipes: Reference docs define JSON patch strategies per class, such as fixing link tuples, clamping parameter values to /object_info ranges, or substituting missing model filenames. - Partial rerun planning: Computes partial_execution_targets as the descendants of the failed node intersected with ancestors of the output nodes, letting ComfyUI reuse cached upstream results. - Use Case: A KSampler node fails with cfg must be <= 30. The classifier returns class parameter, the prompt is patched with a clamped value, and only the sampler, decoder, and save nodes are re-submitted instead of the full graph. ## Quick Start Ask the agent to diagnose the failed ComfyUI run, classify the error, patch the prompt, and re-run only the affected nodes.
