The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/capacity_report.py script uses subprocess.run to execute nvidia-smi for hardware discovery. The command is called with a list of arguments and does not use a shell, which minimizes the risk of command injection. The variable gpu_id is derived from user-provided command-line arguments.
[EXTERNAL_DOWNLOADS]: The skill references the official NVIDIA container registry (nvcr.io) and technical documentation (developer.nvidia.com). These are well-known and trusted sources for the intended domain.
[PRIVILEGE_ESCALATION]: The references/nsys-cli-recipes.md document mentions using sudo sysctl to modify kernel parameters or using --privileged containers as a fallback if the profiling tool fails due to system-level restrictions. While these are high-privilege operations, they are standard requirements for deep performance profiling tools and are presented as troubleshooting steps rather than malicious instructions.
[INDIRECT_PROMPT_INJECTION]: The skill processes external CSV and log files (microbench_csv, dmon_csv) within scripts/capacity_report.py.
Ingestion points: The script reads these files via the paths provided in command-line arguments.
Boundary markers: The script treats the ingested data as structured CSV or log text rather than natural language instructions.
Capability inventory: The script can execute the nvidia-smi utility via subprocess.run.
Sanitization: The script explicitly casts parsed values to int or float within try-except blocks, which prevents non-numeric data from influencing the report logic.

deepstream-profile-pipeline