The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input to define experiment goals and reads user-provided scripts, which are then used to drive environment configuration and code execution.
Ingestion points: User input through the <experiment_goal> placeholder and reading existing script files under workspace/projects/.
Boundary markers: Absent; user-provided intent and scripts are processed without explicit delimitation or warnings to ignore embedded instructions.
Capability inventory: The skill utilizes shell execution for mkdir, conda create, pip install, python execution, and background process management via nohup/bash (found in SKILL.md).
Sanitization: No explicit sanitization, validation, or escaping of the user-provided goals or script content is described before execution.
[COMMAND_EXECUTION]: The skill is designed to execute shell commands for workspace management, environment setup, and running experiments. While it enforces use of local conda environments and specific workspace paths, it provides a direct interface for executing shell commands and scripts on the host system.
[DYNAMIC_EXECUTION]: The skill dynamically creates and executes Python code based on the user's inferred goals or executes existing scripts found in the project directory. This involves runtime environment activation and subprocess spawning.
[EXTERNAL_DOWNLOADS]: The skill allows for the downloading of installers to the /tmp/ directory. While it specifies that these should be transient, the source URLs for these downloads are not strictly defined or restricted to a trusted list within the instructions.

ar-experiment-runner