The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is authorized to execute diagnostic and lifecycle management commands including docker ps, docker logs, ss, and a local shell script ./isaaclab_arena_openpi/docker/run_openpi_server.sh. While these tools are essential for the skill's primary function, the use of wildcards in the environment configuration for docker logs and ss permits the agent to inspect any container's logs or view all network listeners on the host system, which extends beyond the specific needs of the OpenPI server management.
[INDIRECT_PROMPT_INJECTION]: The skill determines server configuration by ingesting data from external Experiment Definition files and local documentation, creating a surface where malicious configuration data could attempt to influence agent behavior.
Ingestion points: Reads repository documentation (docs/pages/quickstart/running_a_real_policy/openpi.rst) and YAML Experiment Definition files (e.g., droid_pnp_openpi_experiment.yaml) to extract server variants and ports.
Boundary markers: The instructions do not define explicit delimiters or instructions to ignore potential commands embedded within these configuration files.
Capability inventory: Access to local shell script execution, Docker container management (status and logs), and network socket inspection.
Sanitization: The skill mitigates risks by requiring the agent to validate variants against a whitelist (pi05, pi0) and ensuring that ports are valid integers, though it relies on natural language instructions rather than programmatic enforcement.

serve-openpi-policy