The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes fixed, read-only system inventory commands (e.g., nvidia-smi, sysctl, system_profiler) to detect hardware capabilities. It explicitly disables shell execution (shell=False) and uses hardcoded argument tuples to prevent command injection. It also implements timeouts and output size limits for all subprocess calls.
[EXTERNAL_DOWNLOADS]: The documentation references the psutil package (version 7.2.2) from the official Python Package Index (PyPI). This is a well-known, legitimate library used for cross-platform resource monitoring, and its inclusion is standard for the skill's stated purpose.
[DYNAMIC_EXECUTION]: The skill uses importlib.import_module to optionally load the psutil dependency if available on the system. The module name is hardcoded as a static string, and no user-supplied input is used to influence the import path.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-supplied JSON snapshot files for workload planning and diffing. It mitigates injection risks by enforcing strict file size limits (1 MiB), rejecting symbolic links, validating the JSON schema, and sanitizing extracted strings using a dedicated safety helper function.
[DATA_EXFILTRATION]: While the skill reads environment variables (Slurm and accelerator visibility tags), it uses an allowlist-only approach and explicitly redacts sensitive information such as job IDs, node names, and absolute paths. Output is directed only to standard output or a local file created with restricted (0o600) permissions in the current directory.

get-available-resources