The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides examples that utilize elevated privileges to configure the host and containers for hardware access. Documentation includes host-level setup commands using sudo for installing the NVIDIA Container Toolkit and configuring udev rules. Docker Compose templates suggest the use of privileged: true and mounting the host's /dev directory for dynamic device attachment, though the skill explicitly notes this should only be used when necessary and provides more secure cgroup-based alternatives.
[EXTERNAL_DOWNLOADS]: The skill instructions fetch configuration and security keys from established and official sources. Fetches GPG keys and repository lists from NVIDIA's official GitHub-hosted domain (nvidia.github.io). Downloads the official ROS archive keyring from the rosdistro repository on GitHub (raw.githubusercontent.com/ros/rosdistro).
[INDIRECT_PROMPT_INJECTION]: The skill describes processes that ingest external configuration files, creating a potential surface for indirect injection if those files are untrusted. Ingestion points: The rosdep install command in the Dockerfile templates processes package.xml files within the workspace (SKILL.md). Boundary markers: None present in the provided templates to delimit external content from instructions. Capability inventory: The templates include capabilities for shell execution, system package installation via apt-get, and workspace builds via colcon (SKILL.md). Sanitization: No explicit sanitization or validation of the ingested XML manifests is demonstrated.
[COMMAND_EXECUTION]: The provided Docker entrypoint script (ros_entrypoint.sh) uses exec "$@" to execute the container's command. This pattern allows the container to run arbitrary commands passed as arguments, which is standard practice for Docker images but requires user awareness of the commands being executed.

docker-ros2-development