The Agent Skills Directory

[SAFE]: The skill implements significant safety guardrails, including mandatory user confirmation for resource-intensive operations and strict verification of local process state before reuse or modification.
[COMMAND_EXECUTION]: The skill uses ps, ss, and nvidia-smi to inspect the local environment. These tools are used legitimately to verify if a GR00T server is running, identify port ownership, and confirm GPU availability for the server, ensuring correct operation without interfering with unrelated processes.
[EXTERNAL_DOWNLOADS]: The skill may trigger the download of AI models (e.g., nvidia/GR00T-N1.6-DROID) and update git submodules (Isaac-GR00T). These operations are documented as standard first-run steps for the project, and the instructions explicitly require the agent to obtain user confirmation before proceeding with any downloads or repository changes.
[INDIRECT_PROMPT_INJECTION]: The skill reads local project documentation (gr00t.rst) and server scripts to resolve configuration details.
Ingestion points: Reads local files from docs/ and project submodules to determine server arguments and readiness message formats.
Boundary markers: The skill establishes a rigid contract for server resolution and explicitly instructs the agent to treat the current checkout as the source of truth, reporting any mismatches.
Capability inventory: The skill is limited to specific bash commands for project management (git, uv, ps, ss, nvidia-smi, test) and does not possess broad system modification or exfiltration capabilities.
Sanitization: While no explicit data sanitization is mentioned, the skill restricts actions to a specific, predefined server management workflow.

serve-gr00t-policy