The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process data from external sources, including clinical trial CSV files and pre-computed results in Jupyter notebooks.
Ingestion points: The skill uses tu run read_executed_notebook to search and read local notebook outputs and pd.read_csv across multiple analysis scripts (e.g., prepare_ae_cohort.py, expression_anova.py, logistic_regression_or.py) to load user-provided data.
Capability inventory: The agent has capabilities for file system access, tool execution (tu run), and shell command execution (subprocess.run).
Boundary markers: There are no specific instructions or delimiters mentioned to prevent the agent from following instructions that might be embedded within the processed CSV or notebook data.
Sanitization: The scripts include basic path validation (e.g., reject_writing_inside_input in spline scripts) to prevent overwriting input data, but they do not sanitize the content of the data for potential injection patterns.
[COMMAND_EXECUTION]: Several scripts within the skill utilize the subprocess module to execute external commands.
Evidence: scripts/r_natural_spline_regression.py and scripts/spline_model_compare.py use subprocess.run to call Rscript. This is used to ensure that natural spline calculations match R's specific implementation.
Evidence: The test_skill.py script uses subprocess.run to execute the skill's own scripts with the --help flag for verification purposes.
[DYNAMIC_EXECUTION]: The skill generates and executes code at runtime to bridge between Python and R.
Evidence: scripts/r_natural_spline_regression.py contains a hardcoded R script template (R_DRIVER). At runtime, the script interpolates user-supplied parameters, such as column names and degrees of freedom, into this template, writes it to a temporary file, and executes it via Rscript.

tooluniverse-statistical-modeling