The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/measure.py uses subprocess.run to execute system commands like nvidia-smi and git. These are used neutrally to collect environment data (GPU model, driver version, and commit hash) for performance reports.
[COMMAND_EXECUTION]: The run_isolated function in scripts/measure.py executes Python driver scripts in a separate process using sys.executable. This pattern is used for accurate peak memory measurement (VmHWM) by ensuring each test case starts with a fresh process. The skill includes security filters to prevent the injection of dangerous environment variables (like LD_PRELOAD) or execution of scripts outside the workspace.
[PROMPT_INJECTION]: The SKILL.md file contains a 'Hard rules' section with instructions such as 'These override any local reasoning.' While this uses language typical of instruction-override attacks, in this context it serves as a strict procedural guardrail to ensure the agent follows a factual evaluation methodology and avoids making unauthorized performance predictions.
[DATA_EXPOSURE]: The skill reads /proc/self/status on Linux systems to measure process memory high-water marks. This is a standard practice for performance profiling and does not expose sensitive user data.

warp-eval