The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted user input that is subsequently used to generate executable Python code and construct shell commands.
Ingestion points: Robot names and USD file paths are extracted from the user's initial prompt or requested via the AskUserQuestion tool.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to treat the user-supplied identifiers as data rather than instructions.
Capability inventory: The skill utilizes Write, Edit, and Bash tools to modify the project's source code and execute training/test scripts.
Sanitization: While the skill recommends using a 'lowercase identifier', it does not implement or instruct the use of programmatic validation, regex checking, or escaping for the user-provided strings before they are used in sensitive operations.
[COMMAND_EXECUTION]: The skill executes a smoke test command (python run.py) using parameters derived from user input for the --embodiment flag. This command is executed via the Bash tool with run_in_background: true.

compass-newembodiment