The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external tensor program specifications, IR graphs, and compilation manifests, which creates a potential surface for indirect prompt injection. 1. Ingestion points: SKILL.md and ir-contracts-and-semantic-identity.md detail the processing of source IR, target graphs, and artifact manifests. 2. Boundary markers: The instructions lack explicit boundary delimiters or warnings to treat ingested content as untrusted data. 3. Capability inventory: The skill leverages torch.fx for runtime code generation and torch.compile for execution, which could be misused if the agent follows instructions embedded in the processed data. 4. Sanitization: There are no sanitization or verification steps to ensure that the ingested specifications do not contain malicious natural language instructions.
[DYNAMIC_EXECUTION]: The skill relies on PyTorch's torch.fx and torch.compile systems for graph capture and transformation. As documented in torch-fx-capture-and-transformation.md, the GraphModule.recompile() function is used to dynamically generate and execute Python code at runtime from a graph representation. While this is a standard and necessary operation for tensor compilation, it represents a runtime code generation capability.

using-tensor-compiler-engineering