The Agent Skills Directory

[COMMAND_EXECUTION]: The SKILL.md file mentions the use of ffmpeg and ffprobe for audio/video processing and provides a Python script to launch a local SAM3 server. These are standard operational requirements for media processing and model serving.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to external APIs (OpenAI-compatible endpoints, DashScope) and a user-configured SAM3 server. These are the primary purpose of the skill for media analysis. The frontmatter does not restrict these tools, which is expected for this use case.
[INDIRECT_PROMPT_INJECTION]: The skill processes external media (images, video, audio) through model prompts.
Ingestion points: Media file paths and URLs are passed to various tools (e.g., vision_chat, omni_asr) in SKILL.md.
Boundary markers: The skill documentation advises using dry_run=true to inspect payloads, which serves as a manual verification step for the agent.
Capability inventory: Subprocess calls for ffmpeg and network POST requests to model endpoints are present in SKILL.md and references/launch_sam3_server.py.
Sanitization: Standard Base64 encoding is used for image data transmission in the SAM3 server script, reducing the risk of binary-to-text injection issues.

qwen-mm-plugins-api