The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches model code and dependencies from the official Biohub GitHub organization and Hugging Face repositories. This includes a specialized fork of the transformers library and the esm SDK required for model execution.
[COMMAND_EXECUTION]: Provides instructions for setting up a Python 3.12 environment using pip and uv to install necessary bioinformatics, machine learning, and hardware acceleration packages.
[DYNAMIC_EXECUTION]: Includes a Python implementation to monkeypatch torch.linalg.svd at runtime. This is designed to improve CUDA stability during diffusion sampling by redirecting specific operations to the CPU to avoid cusolver workspace corruption.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted biological data formats that act as potential attack surfaces.
Ingestion points: Processes protein sequences, Multiple Sequence Alignments (MSA) from A3M files, and SMILES strings via StructurePredictionInput and MSA utility classes.
Boundary markers: No specific delimiters or instructions are used to separate external sequence data from the agent's logic.
Capability inventory: The skill allows writing prediction files to the local file system, network access for fetching weights from Hugging Face, and execution of shell commands for installation.
Sanitization: No explicit validation or sanitization is applied to the input biological sequences or structural data files before they are processed by the folding model.

esmfold2