The Agent Skills Directory

[SAFE]: The skill's instructions in SKILL.md explicitly prohibit the use of sudo or privilege escalation, defaulting to unprivileged shell access and requiring users to run probes themselves if direct access is not available.
[SAFE]: All installation and validation commands are presented to the user in discrete blocks and require explicit approval before any execution occurs.
[SAFE]: System information collection is performed via a transparent, read-only, and unprivileged shell script (scripts/collect-system-info.sh) that targets only hardware and software configuration data (OS, GPU, CUDA, and RDMA drivers) relevant to the installation process.
[SAFE]: Software artifacts and source code are fetched exclusively from official, well-known NVIDIA domains, including developer.nvidia.com and github.com/NVIDIA.
[SAFE]: The skill maintains a strict posture against silent execution, ensuring that even validation tests are approved by the user and their outcomes are recorded truthfully without substituting simple checks for actual multi-node validation.
[SAFE]: No evidence of obfuscation, malicious prompt injection, or persistence mechanisms was found in the provided instructions or scripts.

nvshmem-install