The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions and reference materials describe patterns for loading graph data from external sources, including CSV files (references/custom_datasets.md) and pre-downloaded dataset files. This creates an attack surface where maliciously crafted data could potentially influence the agent's behavior during processing.
Ingestion points: SKILL.md (dataset loaders), references/custom_datasets.md (CSV and tensor dictionary loading).
Boundary markers: Absent from the generic file loading examples.
Capability inventory: File system access, network operations for dataset downloads, and tensor processing via torch.load.
Sanitization: The provided code examples utilize weights_only=True in torch.load calls, which is a recommended security measure to prevent arbitrary code execution during deserialization.
[DYNAMIC_EXECUTION]: The skill documentation includes examples of using torch.compile to optimize neural network models (references/scaling.md). This involves runtime compilation of model graphs into machine code. Additionally, custom MessagePassing layers involve dynamic computation of adjacency matrices and feature aggregations.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of binary wheels and the retrieval of research datasets from well-known technical and academic repositories.
Binary Wheels: Fetches from data.pyg.org, which is the official wheel index for PyTorch Geometric.
Datasets: References datasets hosted by the University of Stanford (stanford.edu), Princeton University (princeton.edu), Max Planck Institute (tue.mpg.de), and established open-source repositories on GitHub and AWS S3 (deepchemdata.s3-us-west-1.amazonaws.com).

torch-geometric