The Agent Skills Directory

[SAFE]: No security issues detected. All analyzed code and instructions align with the primary purpose of the skill, and the Docker image runs as a non-privileged user by default.
[EXTERNAL_DOWNLOADS]: Fetches vendor SDKs and reference models from established sources.
Downloads Basler Pylon SDK from Basler's official blob storage during image build.
Clones the DDM-Net research repository from GitHub to provide the temporal action detection model.
[COMMAND_EXECUTION]: Performs localized shell command execution for model compilation.
Invokes an internal Python script via subprocess.run to convert models to TensorRT engines on the fly, which is a standard optimization procedure for the intended use case.
[PROMPT_INJECTION]: Contains an indirect prompt injection surface through the Vision-Language Model processing untrusted video frames and user prompts.
The risk is mitigated by using explicit system prompts and task boundaries to constrain model output to specific SOP action labels defined in a configuration file.

deepstream-sop