The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes subprocess.run with argument lists to call ffmpeg, ffprobe, and internal Python scripts. This implementation ensures that user-provided file paths are not interpreted by a command shell, preventing command injection vulnerabilities.\n- [EXTERNAL_DOWNLOADS]: The openai-whisper library downloads pre-trained models from official sources to a local cache directory (~/.cache/whisper) during initial execution. This is standard behavior for the library and necessary for its functionality.\n- [PROMPT_INJECTION]: The skill processes untrusted video content (visual frames and audio transcripts), creating a surface for indirect prompt injection. \n
Ingestion points: Extracted frames in extract_frames.py and generated transcripts in transcribe_audio.py.\n
Boundary markers: None explicitly present in the data output to the agent.\n
Capability inventory: The skill performs perception tasks via ffmpeg and whisper but does not expose privileged system capabilities to the processed data.\n
Sanitization: No content filtering is applied to the extracted video data.\n- [SAFE]: No malicious patterns such as obfuscation, credential harvesting, or unauthorized network activity were detected. The skill's functionality is consistent with its stated purpose of local video analysis.

video-vision