The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes standard system discovery commands to assess platform compatibility. This includes uname, lsb_release, nvidia-smi, docker, and pip. These are legitimate operations for a setup utility identifying hardware and software prerequisites.
[EXTERNAL_DOWNLOADS]: The skill fetches documentation from https://docs.nvidia.com/holoscan/sdk-user-guide/ to determine supported platforms and package names. Since this targets an official, trusted vendor domain, the operation is considered safe and standard for keeping installation logic synchronized with the latest releases.
[COMMAND_EXECUTION]: The included scripts (scripts/check_conda.sh and scripts/check_ngc_image.sh) perform localized environment checks. check_conda.sh searches common installation directories and shell configuration files (e.g., .bashrc, .zshrc) to detect Conda even when not on the active PATH. check_ngc_image.sh verifies the presence of specific Holoscan container images via docker images. Both scripts operate within the scope of environment detection.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it fetches and processes external content from docs.nvidia.com. While this is a trusted source, an attacker who compromises the documentation could theoretically inject instructions. However, given the source's reputation and the skill's specific purpose, this is a low-risk surface.
Ingestion points: Step 1 fetches content from https://docs.nvidia.com/holoscan/sdk-user-guide/ and sdk_installation.html into the agent's context.
Boundary markers: None observed; the agent is instructed to read the docs to get requirements.
Capability inventory: The skill has capabilities to execute shell commands and delegate to other installation skills (/holoscan-install-*).
Sanitization: No specific sanitization or validation of the fetched documentation content is described.

holoscan-setup