The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external geospatial data formats (GeoPackage, GeoParquet, Shapefiles) which represents a potential attack surface. The risk is effectively mitigated by a strict 'local-only' intake policy that rejects URLs, GDAL virtual file systems, and archives.
Ingestion points: Data is ingested via gpd.read_file and pyarrow using the load_geodataframe and inspect_local_vector functions in scripts/_common.py.
Boundary markers: The skill instructions mandate out-of-band hash verification and manual staging of data into isolated local directories.
Capability inventory: The skill utilizes file read/write operations and subprocess calls to Python scripts for auditing.
Sanitization: Robust sanitization is implemented via reject_nonlocal, which blocks remote protocols and archive members, and checked_input_file, which enforces root directory constraints, prohibits symbolic links, and validates file suffixes.
[EXTERNAL_DOWNLOADS]: The skill references a pinned stack of standard geospatial libraries (GeoPandas, NumPy, Pandas, etc.) available from official package registries. These dependencies are treated as safe and the skill provides a smoke-tested installation command for a reproducible environment.
[COMMAND_EXECUTION]: The skill interacts with native-code libraries (GDAL, GEOS, PROJ) through Python bindings. It correctly identifies these as a trust boundary and implements safety gates such as disabling PROJ network access (network.set_network_enabled(False)) to prevent uncontrolled grid retrieval.
[DATA_EXFILTRATION]: To protect sensitive coordinate and identity data, the skill's CLI tools are designed to emit deterministic, redacted JSON reports. These reports exclude exact coordinates, record identifiers, and precise bounding boxes, focusing instead on aggregate counts and technical metadata.

geopandas