The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates C++ source code for a bounding box parser and a corresponding Makefile, then executes make to compile a shared object library (.so). This library is subsequently loaded into the DeepStream execution environment. While this is a core functional requirement for custom model support in DeepStream, the runtime generation and execution of compiled code is a high-capability behavior.
[COMMAND_EXECUTION]: The skill makes extensive use of shell command execution across its scripts, including subprocess.run calls in the reporting modules. It utilizes gst-launch-1.0, deepstream-app, and trtexec to perform model benchmarking and validation.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading vision models and configurations from HuggingFace and NVIDIA NGC. These are well-known and reputable services for model hosting, and the scripts implement basic input validation to prevent URL or path manipulation.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of model metadata and config.json files.
Ingestion points: Data enters the environment via model acquisition from HuggingFace and NGC APIs.
Boundary markers: The skill does not implement explicit delimiters for data processed during labeling or architecture detection.
Capability inventory: The agent possesses high capabilities including network access, file system writes, and code compilation.
Sanitization: The report generation script (md-to-html-pdf.py) includes specific security controls, such as disabling local file access for wkhtmltopdf to mitigate potential HTML injection vulnerabilities from external model descriptions.
[PROMPT_INJECTION]: The SKILL.md file contains an instruction to "Run the full pipeline autonomously without pausing for confirmation at each step." This promotes a high degree of agent autonomy and reduces the opportunity for human-in-the-loop oversight during sensitive operations like code compilation and network downloads.

deepstream-import-vision-model