The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using tools like yt-dlp, ffmpeg, and curl. These commands incorporate user-provided parameters, such as video URLs, image URLs, and start/end time offsets. This pattern creates a potential surface for command injection if input parameters are not strictly sanitized by the execution environment.
Evidence: yt-dlp "URL" -o /tmp/verging-faceswap/input.mp4 and ffmpeg -i input.mp4 -ss <start> -to <end> ... as described in SKILL.md.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch media from user-specified URLs and interacts with the verging.ai API and Cloudflare R2 storage for file uploads and job management. While these are part of the primary workflow, they involve communication with external infrastructure.
Evidence: Downloads via yt-dlp and curl, and uploads to verging.ai/api/v1/upload-video and Cloudflare R2 endpoints as documented in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media files (videos and images) retrieved from external URLs (YouTube, Bilibili, etc.). It lacks explicit boundary instructions or sanitization steps to prevent the agent from potentially interpreting instructions hidden within the media's metadata or content.
Ingestion points: --video (URL) and --face (URL) in SKILL.md.
Boundary markers: Absent.
Capability inventory: Shell execution via yt-dlp, ffmpeg, ffprobe, and network operations via curl as detailed in SKILL.md.
Sanitization: Absent.

faceswap