The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill parses user-supplied code, scripts, or commands to classify GPU requirements and determine environment settings, creating a surface where malicious instructions could influence system probes or command generation.
Ingestion points: The skill explicitly reads the user's code or commands in 'Decision 1' and uses project/script/workspace names provided by the environment for table output and shell interpolation.
Boundary markers: No specific delimiters or instructions are used to isolate or ignore potential embedded commands within the user-provided code during the classification process.
Capability inventory: The skill executes various shell commands including nvidia-smi, conda, python, nproc, uptime, free, and df across multiple decision steps.
Sanitization: There is no evidence of sanitization or strict validation for variables like <env_name> or <workspace_name> before they are interpolated into shell command strings (e.g., in the conda activate command).
[COMMAND_EXECUTION]: The skill executes local shell commands to gather system telemetry. These include nvidia-smi for GPU metrics, uptime and free -h for system load/RAM, and df -h for disk space monitoring. These commands are consistent with the skill's stated purpose of server preflight checks.
[DYNAMIC_EXECUTION]: The skill generates and executes dynamic Python code snippets (e.g., python -c 'import torch; ...') within activated environments to verify library availability and CUDA connectivity. This follows a template-based generation pattern for environment validation.

ar-gpu-preflight