The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill includes functionality in scripts/auto_fix_deps.py and scripts/comfyui_setup.sh to download and install third-party Python modules (custom nodes) and machine learning models from external registries and arbitrary user-provided URLs. Because ComfyUI custom nodes are executable Python code, this represents a significant external code execution surface.
[COMMAND_EXECUTION]: The skill utilizes subprocess.run across several scripts (e.g., hardware_check.py, auto_fix_deps.py) to execute system commands, manage the ComfyUI server lifecycle, and interface with the comfy-cli utility. Commands are constructed using lists rather than raw strings to mitigate shell injection risks.
[EXTERNAL_DOWNLOADS]: Setup and maintenance scripts automate the fetching of the comfy-cli tool, the ComfyUI framework, and various AI models from public sources like GitHub, PyPI, Hugging Face, and CivitAI.
[INDIRECT_PROMPT_INJECTION]: The skill processes complex workflow JSON files which can contain instructions for custom nodes. run_workflow.py and _common.py implement mitigations including safe_path_join to prevent path traversal during output downloads and logic to prevent parameters from overwriting established link connections within the graph.
[SAFE]: The automated scan findings for remote code execution via curl | python3 in SKILL.md and scripts/comfyui_setup.sh are confirmed as false positives. In all instances, the pipe targets the standard library json.tool module (e.g., python3 -m json.tool) for pretty-printing JSON output rather than executing the received data as a script.

comfyui