The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell scripts (doctor.sh, smoke.sh, train.sh, export_publish.sh) and uses uv run to execute Python code. These operations are within the scope of its defined purpose for managing reinforcement learning pipelines.
[PRIVILEGE_ESCALATION]: The deployment documentation in references/DEPLOY.md and the export_publish.sh script suggest running sudo robotctl commands. The instructions correctly stipulate that these should be printed for the user and only executed if the user explicitly requests installation on hardware they own.
[EXTERNAL_DOWNLOADS]: The scripts/doctor.sh script performs a git clone from https://github.com/pollen-robotics/microduck_rl. As this is a well-known repository related to the skill's primary function and the skill was developed by a known entity, this is considered a functional requirement.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data such as reward configurations, task IDs, and Hugging Face repository descriptions. While it includes validation via scripts/gate_check.py for model shapes, it lacks explicit sanitization or boundary markers for textual metadata (e.g., repository descriptions or custom task configurations) that could influence the agent's context during the training or publishing phases.

microduck-skill