The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches pre-trained models (e.g., YuNet face detector ONNX) from the official OpenCV model zoo on GitHub as part of the setup instructions.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image, video, and model files which could be used as an ingestion surface for indirect prompt injection attacks.
Ingestion points: The scripts/cv.py script ingests data via subcommands info, capture, detect-faces, track, yolo, calibrate, and stitch, utilizing cv2.imread and cv2.VideoCapture to load external files and network streams.
Boundary markers: No specific delimiters or boundary warnings are implemented for the binary content of images or videos being processed.
Capability inventory: The skill possesses the capability to write files (cv2.imwrite, cv2.VideoWriter, cv2.FileStorage) and access network streams via RTSP (cv2.VideoCapture).
Sanitization: No sanitization is performed on the internal content of the binary data or the architecture of loaded DNN models.

cv-opencv