The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Python script that executes ffprobe and ffmpeg to validate the integrity and metadata of generated video files. These executions are restricted to specific arguments and use shell=False to prevent command injection.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of AI model weights and generated video artifacts. It targets official vendor repositories on GitHub (Wan-Video) and Hugging Face (Wan-AI). The implementation includes strict hostname allowlisting and IP address validation to ensure downloads occur only from approved infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied prompts for video generation, which represents an indirect injection surface.
Ingestion points: User prompts and negative prompts ingested via CLI arguments in the provided Python utility.
Boundary markers: The script uses a hash-based approval mechanism (SHA-256) to ensure the agent only executes requests that have been explicitly reviewed and approved in a 'dry-run' phase.
Capability inventory: Capabilities include network access to Alibaba Cloud Model Studio endpoints, file system writes for artifacts and state tracking, and local execution of media validation tools.
Sanitization: The implementation performs extensive validation, including input length constraints, URL hostname verification, and media integrity checks using ffprobe and full ffmpeg decoding passes.

alibaba-wan-video