The Agent Skills Directory

[DYNAMIC_EXECUTION]: The provided scripts, such as benchmark_evaluation.py and load_and_split_data.py, use importlib.import_module to lazily load components of the Therapeutics Data Commons package. This dynamic loading is secure as the module names are derived from hardcoded internal maps and are strictly validated against an allowed set of tasks and groups before invocation, preventing the loading of arbitrary or malicious modules.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of scientific datasets from external sources (e.g., Harvard Dataverse). To protect against potential injection via external data, the skill enforces boundary controls: it mandates a 'plan-then-execute' workflow, provides a dedicated cache_audit.py tool for read-only inspection of downloaded files, and utilizes a truncate_value helper to ensure that data returned to the agent context is bounded and sanitized.
[EXTERNAL_DOWNLOADS]: Dataset loading and model scoring workflows involve fetching content from the Harvard Dataverse. These network operations are implemented safely, targeting well-known academic infrastructure. Furthermore, the skill's scripts require explicit flags (--execute and --download) to trigger these network side effects, ensuring the agent cannot perform downloads without clear intent and oversight.

pytdc