Dsh Comfyui

DeepSeek Harness ComfyUI

Use this skill for every ComfyUI task. It adds Harness-specific execution checks around the pinned ComfyUI-Agent-Kit without assuming a portable build, a particular GPU, or a fixed model directory.

Start here

1. Read machine.md. If it is missing or still contains placeholders, read references/upstream/BOOTSTRAP.md, inspect the running server, and fill in only facts for that machine.

2. Read references/upstream/SKILL.md completely. Follow its links for the requested task and model instead of loading every reference at once.

3. For MiniMax H3, also read references/minimax-h3/SKILL.md and references/minimax-h3/reference.md before writing the prompt or graph. H3 still runs through ComfyUI.

4. Run python scripts/comfyui_probe.py health before the first ComfyUI operation in a session. Windows users may use powershell -ExecutionPolicy Bypass -File scripts/comfyui_probe.ps1 health.

5. Export an API-format workflow and run preflight with the deliverable type:

   python scripts/comfyui_probe.py preflight workflow_api.json --expect video
   

   Resolve every error before queueing. PreviewImage and temporary frames do not satisfy a video delivery.

Build and run

• Ask the live server for /system_stats, /object_info, and model inventory. Static docs describe possibilities; the running server decides what exists.
• Start from an official workflow template when one matches. Keep a GUI-format copy in ComfyUI's user workflows directory and an API-format copy for execution.
• When Harness cannot access ComfyUI's directories directly, read references/filesystem-bridge.md. Use scripts/comfyui_files.py upload-image for LoadImage inputs and save-workflow for the Workflows sidebar.
• Inspect untrusted graphs before running them. Review node classes, model names, input paths, network or paid API nodes, and persistent output nodes.
• Use scripts/run_workflow.py when no structured ComfyUI driver is available. It queues the graph, waits for terminal history, downloads outputs, and rejects the run when the requested media type is absent or empty.
• Do not assume Harness can see a local image because the selected model supports vision. After generation, read references/visual-verification.md, run scripts/inspect_media.py, and pass every returned QA image through read_image or describe_image when that tool exists.
• If no image-reading tool exists, report visual verification as pending and hand over the exact QA paths. Prompts, filenames, metadata, and ComfyUI success are not visual evidence.
• Run a short smoke render before an expensive render unless the workflow cannot be reduced safely.
• Do not install nodes or models, change persistent server configuration, clear another user's queue, or interrupt another job without authorization.
• Keep credentials out of workflow JSON, logs, and repositories.

MiniMax H3

• Treat hosted Hailuo 03 nodes and local MiniMax H3 core nodes as different ComfyUI backends.
• Local H3 frame counts must satisfy length % 17 == 5; trained clips are 124 to 362 frames at 24 fps.
• Use the official three-section prompt for base modes and the six-section prompt for reference-to-video. Dialogue belongs inside <d>[Language] ...</d>.
• End the graph in a real video save node. Preserve the native audio stream when the workflow generates audio.
• A queue acknowledgement, PreviewImage, PNG sequence, or MP4 assembled outside the submitted graph is not a completed H3 workflow.

Completion

A completed generation reports the prompt id, terminal ComfyUI history, downloaded non-empty media paths, and the saved editable workflow. Run scripts/inspect_media.py on each deliverable. Visual completion additionally requires a model-facing image tool or explicit human inspection of every returned QA image.

Repository design, attribution, and update policy live in references/comparison.md.