The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external media data (video, images) through vision models, which presents a surface for indirect prompt injection where adversarial content in the media could influence the agent's reactive logic or decision-making processes.
Ingestion points: Processes video and image assets (e.g., assets/footage.mp4, assets/dancer.mp4) provided as input to composition and layer methods in SKILL.md.
Boundary markers: No specific delimiters or warnings for embedded instructions in media data are mentioned.
Capability inventory: The skill can write to the local file system (~/.cache/framefields/models), perform network requests to download models, and execute shell commands via onnxruntime.
Sanitization: Implements SHA-256 checksum verification for downloaded model files to ensure integrity.
[EXTERNAL_DOWNLOADS]: Machine learning models for tasks such as object detection, pose estimation, and segmentation are fetched lazily from a remote base URL, which is configurable via the FRAMEFIELDS_MODELS_BASE_URL environment variable. This is a documented and standard feature for the framework's vision capabilities.
[COMMAND_EXECUTION]: The skill utilizes onnxruntime-node and onnxruntime-web to execute vision models on CPU or GPU. It interacts with environment variables (e.g., FRAMEFIELDS_VISION_PROVIDER) to manage hardware acceleration and execution threads, which is consistent with its stated purpose of high-performance media rendering.

framefields-effects