The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/setup_environment.py utility uses the subprocess module to manage Conda environments and install dependencies via pip. These operations are performed using structured arguments rather than shell strings, following security best practices for environment management tools.
[DYNAMIC_EXECUTION]: The framework performs two types of dynamic execution:
The scripts/setup_environment.py script generates a temporary Python file (test_biomni.py) and executes it using subprocess to verify the installation and API key configuration.
The core A1 agent class generates and executes Python code at runtime to process biomedical datasets, a behavior that is explicitly documented in SKILL.md with accompanying warnings for users to review code and use sandboxed environments.
[EXTERNAL_DOWNLOADS]: The skill downloads approximately 11GB of integrated biomedical databases (a 'data lake') from Stanford University's SNAP Lab repositories (including GitHub and official Stanford domains). These downloads are performed upon first use to enable offline knowledge retrieval for research tasks.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an inherent attack surface for indirect prompt injection as it is designed to ingest and analyze untrusted external research data (e.g., VCF, H5AD, and CSV files provided by the user) while possessing code execution capabilities.
Ingestion points: User-specified biomedical data files loaded via the A1.go() method.
Boundary markers: The documentation encourages human oversight and code review rather than automated delimiters.
Capability inventory: Includes file system access for data reading, network access for LLM providers, and Python code execution.
Sanitization: The skill provides clear security guidelines in SKILL.md advising users to run the agent in isolated environments (Docker/VMs) and to review generated code before execution.

biomni