The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for installing native dependencies (OpenSlide, JDK, build tools) and Python packages using platform-standard package managers such as apt-get, Homebrew, and uv pip.
[EXTERNAL_DOWNLOADS]: Documents the capability to fetch pre-trained models from Hugging Face (e.g., Mesmer and HoVer-Net ONNX models) and research datasets from Warwick University (PanNuke) and Zenodo (DeepFocus). These operations are part of the intended research workflow and are governed by explicit documentation requiring user opt-in and destination disclosure.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface by processing external pathology images, CSV manifests, and JSON spatial data.
Ingestion points: Data entry occurs via manifest validation in scripts/slide_manifest.py, image inspection in scripts/image_qc.py, and schema validation in scripts/validate_spatial_schema.py.
Boundary markers: The SKILL.md file includes a dedicated 'Scope and safety boundary' section outlining protocols for data isolation and the handling of pseudonymous identifiers.
Capability inventory: The agent uses allowed tools (Bash, Read, Write, Edit) to perform local processing, tiling, and inference on these files.
Sanitization: The provided helper scripts in the scripts/ directory implement robust defensive checks, including NUL byte rejection, strict path resolution, and the rejection of symlinks and URLs to ensure all I/O is local and bounded.

pathml