The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents how to interact with the ComfyUI API, which processes JSON-based workflows. This creates an attack surface where instructions hidden within a workflow could influence the agent or the server.
Ingestion points: Python scripts queue_prompt.py and get_results.py demonstrate how to send and receive data via the /prompt and /history endpoints.
Boundary markers: No delimiters or safety instructions are provided to ensure the agent ignores potentially malicious logic embedded within external workflow files.
Capability inventory: The skill enables the creation of a server (python main.py) that can execute complex Python-based nodes and save files to the local filesystem.
Sanitization: The provided examples do not demonstrate validation or sanitization of the JSON workflow structure before it is passed to the ComfyUI backend.
[EXTERNAL_DOWNLOADS]: The skill fetches software and pre-trained models from established public platforms.
Clones the ComfyUI and ComfyUI-Manager source code from official GitHub repositories.
Downloads Stable Diffusion XL model weights and VAE components from HuggingFace.
Note: The installation instructions do not include checksum verification for the downloaded binary model assets, which is a common security best practice for large-scale weights.
[COMMAND_EXECUTION]: The server configuration includes a command to listen on all network interfaces (--listen 0.0.0.0), which could lead to unintended exposure of the local API if not properly restricted by a firewall or container environment.

comfyui