The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python scripts to inventory and repair ComfyUI workflows. These scripts are executed locally via Bash and use standard library modules. The subprocess.run calls detected in the test files (tests/test_png_workflow_extractor.py and tests/test_model_metadata_scripts.py) are used strictly for unit testing the skill's own components and do not present a security risk to the user environment.- [DATA_EXFILTRATION]: The skill implements rigorous outbound network safety controls. The inventory_workflow_models.py script enforces a strict URL policy, requiring HTTPS and restricting hosts to known providers (Hugging Face and Civitai). It specifically blocks IP-literal hosts, credentials in URLs, and non-standard ports to prevent SSRF and data leakage. Evaluation files (such as evals/gated-model.eval.md) provide clear instructions for the agent to avoid exfiltrating or persisting user tokens.- [INDIRECT_PROMPT_INJECTION]: While the skill ingests untrusted data from ComfyUI workflow files (JSON and PNG), it employs multiple layers of sanitization. The extract_png_workflow.py script enforces MAX_PNG_BYTES, MAX_TEXT_VALUE_BYTES, and MAX_JSON_DEPTH to prevent resource exhaustion attacks. Furthermore, the apply_model_metadata.py script validates all imported metadata against a strict schema before generating a repaired workflow, effectively neutralizing potential injection vectors in the processed data.- [SAFE]: The skill author ('runpod') provides high-quality, security-conscious scripts that avoid dangerous functions like eval() or unsafe deserialization. Intermediate processing is handled within temporary directories that are cleaned up after execution, ensuring no sensitive traces remain on the host filesystem.

runpod-comfyui-models