The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system commands for package management (APT), compilation (gcc/g++/nvcc), and running official verification samples. These executions are secured via a central wrapper (run_command in setup_contract.py) that strips potentially dangerous environment variables, enforces Python isolation (-I flag), and uses absolute paths for system tools.\n- [EXTERNAL_DOWNLOADS]: Dependencies and source archives (e.g., pycuda) are fetched from trusted domains including pythonhosted.org, pytorch.org, and nvidia.com. Every download is strictly verified against a pre-defined SHA-256 hash before use.\n- [PRIVILEGE_ESCALATION]: System package operations utilize sudo -n for non-interactive execution. Access is limited to specific APT commands defined within a canonical plan. The skill employs a digest-matching mechanism to ensure the execution plan has not been altered between review and application.\n- [REMOTE_CODE_EXECUTION]: The skill downloads, builds, and installs the pycuda source package. This process is protected by an explicit SHA-256 integrity check of the source archive and the use of scoped build environments to prevent system-wide side effects.\n- [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface as it ingests system configuration data (e.g., /etc/nv_tegra_release, APT sources). This risk is mitigated through the use of strict, versioned JSON schemas (e.g., nvcodec-environment schema 1.2), canonical digest verification of artifacts, and robust sanitization of tool outputs. Mandatory Evidence Chain: 1. Ingestion points: Reads system release and APT source files. 2. Boundary markers: Uses versioned JSON artifacts and plan digests. 3. Capability inventory: Limited subprocess execution and module imports. 4. Sanitization: Hashing and strict JSON parsing hooks.

jetson-video-setup