The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Automated scans flagged shell pipes to Python (e.g., curl ... | python3). Technical analysis confirms these are benign local health checks targeting 127.0.0.1 and localhost. The scripts use python3 -m json.tool to format JSON responses from the local ComfyUI server, which is a standard diagnostic practice.
[COMMAND_EXECUTION]: The skill uses subprocess.run and shell scripts to execute lifecycle commands via the official comfy-cli and system discovery tools like nvidia-smi or rocm-smi for hardware detection. These operations are restricted to the intended functionality of setting up and managing a local ComfyUI environment.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of AI models and custom nodes from established repositories such as HuggingFace, CivitAI, and GitHub. These downloads are performed through the comfy-cli tool.
[CREDENTIALS_UNSAFE]: The skill manages the COMFY_CLOUD_API_KEY for hosted services. It follows security best practices by encouraging environment variable usage and implementing a custom HTTP session handler in _common.py that explicitly strips sensitive authentication headers when following cross-host redirects (e.g., from the Comfy Cloud API to S3/GCS storage backends).
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface for indirect prompt injection as it processes untrusted workflow JSON files and user-supplied prompts. However, the risk is mitigated by the use of safe_path_join to prevent path traversal during output downloads and the fact that prompts are passed as data to the underlying generation models rather than being executed by the shell.

comfyui