The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to fetch additional instructions and project components from official NVIDIA GitHub repositories, including nurec-skills, ncore, asset-harvester, and harmonizer.
[COMMAND_EXECUTION]: Routine use of git for version control and docker for running reconstruction engines is documented. It also includes standard cleanup commands for managing container-generated files.
[CREDENTIALS_UNSAFE]: The skill demonstrates best practices for secret management by providing safe patterns to verify the presence and length of HF_TOKEN and NGC_API_KEY using length checks and official CLI tools like hf auth whoami instead of echoing their actual values.
[PRIVILEGE_ESCALATION]: The documentation includes the use of sudo chown specifically for recovering file ownership of outputs created by Docker containers, which is a standard administrative procedure for these workflows.
[INDIRECT_PROMPT_INJECTION]: The skill functions by reading other local skill files to determine how to route user requests, representing a data ingestion surface.
Ingestion points: Files like SKILL.md are read from the nurec-skills clone as described in references/upstream-fetch.md.
Boundary markers: No explicit boundary markers are defined for the read content within the router itself.
Capability inventory: The skill execution environment includes git, docker, and sudo capabilities for administrative tasks.
Sanitization: Content is read directly via the shell without additional filtering, but the source is limited to official, vendor-controlled repositories.

physical-ai-neural-reconstruction