The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute user-provided evaluation commands (e.g., pytest bench.py, python evaluate.py) through a local setup script. This execution is the primary intended functionality for measuring experiment metrics, but it involves running arbitrary strings in a shell environment.
[INDIRECT_PROMPT_INJECTION]: The skill accepts input for the eval_cmd parameter which is later interpolated into a shell command. This creates a potential surface for indirect injection if the agent is influenced by malicious external data to call the setup command with a manipulated payload.
Ingestion points: Command-line arguments or interactive user prompts provided to the /ar:setup command.
Boundary markers: None are specified for the interpolation of the evaluation command.
Capability inventory: The skill executes shell commands via the {skill_path}/scripts/setup_experiment.py script.
Sanitization: No specific sanitization or validation of the evaluation command string is described in the prompt instructions.

setup