Get spark RAM/VRAM/GPU usage, ComfyUI version, and current queue size.

List recent ComfyUI output images on spark, newest first.

Fetch a ComfyUI output image and return it as an image block visible in the conversation.

Fetch the last N ComfyUI output images and return them as image blocks.

List ComfyUI workflow JSON files on spark, newest first.

Read a workflow JSON file from spark as raw text.

Write a workflow JSON file to spark. Validates that content parses as JSON with nodes+links arrays.

Queue a ComfyUI workflow with optional overrides (prompt, seed, steps, guidance, sampler, dims, loras). Returns prompt_id; call wait_for_image to retrieve the result.

Get current ComfyUI queue state.

Interrupt the currently running ComfyUI job.

Poll ComfyUI until the given prompt_id finishes, then return the generated image.

List installed model files in a ComfyUI models subdir on spark.

Download a model file (Civitai / HuggingFace / direct URL) onto spark via aria2c into the chosen models dir.

Curate selected ComfyUI output images into a named folder on spark and (by default) sync to ~/Pictures// on the Mac. Optionally writes a NOTES.md alongside.

Queue a workflow with N override variants (different seeds, prompts, LoRAs, etc.), wait for all to complete in parallel, and return all images in a single response. Max 8 variants per call.

Author and run a ComfyUI workflow as Python via ComfyScript — runs on spark in the comfyui venv against the live server. Use this for COMPLEX/multi-stage graphs (LTX-2.3 audio, two-stage pipelines) that are awkward to express as JSON or drive via the GUI. Provide a snippet with a with Workflow(wait=True): block ending in SaveImage(...)/SaveVideo(...); the imports + load() are auto-prepended. Every server node (incl. custom: HunyuanImage3Generate, LTXV*, RES4LYF samplers) is available as a function. Returns the new output filenames (view with view_image) plus an output tail. Heavy models (HunyuanImage-3.0 ~13min) need a high timeout_sec.