The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from Parquet files provided by the user in the data/factors directory. While this represents a data ingestion surface, the processing is limited to mathematical operations (correlation, z-score, weighting) using robust libraries like pandas and scipy. No exploitable capabilities such as shell execution, dynamic code evaluation, or network operations on user-controlled data were identified.
Ingestion points: Reads .parquet files from the path specified in the --factor-dir argument (defaulting to data/factors).
Boundary markers: None explicitly implemented for data content, as the data is expected to be strictly numerical/tabular.
Capability inventory: File read/write (Parquet, JSON, Pickle), mathematical synthesis. No eval, exec, or subprocess calls on data.
Sanitization: Uses standard pandas and pyarrow loaders which perform schema validation during ingestion.
[COMMAND_EXECUTION]: The script scripts/blend.py modifies the Python path (sys.path.insert) to resolve a dependency on a local peer skill skill-pandadata-api. This is a legitimate functional requirement for the QuantSkills ecosystem to access the pandadata_runtime library.
[DATA_EXPOSURE]: The skill uses pickle.dump to save the intermediate state of the blend process. While pickle can be dangerous when loading untrusted data, this skill only performs a write operation (dump) to save session state, which is a safe practice in this context.

factor-blend