The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions involve processing user-provided operations, irreps, and code logic to generate and run validation scripts. This introduces a surface for indirect prompt injection if user-provided content is improperly handled.
Ingestion points: User requests for equivariant tensor products and GNN layer definitions in SKILL.md.
Boundary markers: No explicit input delimiters or safety warnings for embedded instructions in processed data are provided.
Capability inventory: The skill creates temporary directories, writes scripts to files (check.py), and executes them via subprocesses.
Sanitization: There is no evidence of sanitization for user-provided code strings before they are interpolated into validation scripts.
[DYNAMIC_EXECUTION]: The skill uses a template-based script generation and execution pattern to verify the correctness of the cuEquivariance implementations.
Evidence: SKILL.md contains a recipe that saves preamble and user-specified checks into a check.py file and subsequently executes it using the configured Python interpreter.
[EXTERNAL_DOWNLOADS]: The skill downloads PyTorch CPU wheels and dependencies from established and trusted official sources.
Evidence: scripts/torch-cpu-wheels.txt contains direct wheel URLs from download-r2.pytorch.org (PyTorch's official hosting), which is a well-known and trusted service.

cuequivariance-torch