The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes external data files in the .h5ad format (AnnData) via scanpy.read_h5ad. This creates a vulnerability surface for indirect prompt injection if the data files contain malicious instructions within their metadata (e.g., cell types or gene labels) that the agent might interpret as commands.
Ingestion points: sc.read_h5ad is used in SKILL.md and the validation script scripts/graphvelo_smoke.py to load input data.
Boundary markers: The skill does not use specific delimiters or instructions to prevent the agent from obeying text found inside the data files.
Capability inventory: The skill utilizes omicverse for analysis and matplotlib for file generation (rendering plots), which represent the main actions performed on the data.
Sanitization: There is no evidence of metadata sanitization or validation of the input data's content before processing.
[EXTERNAL_DOWNLOADS]: The references/compatibility.md file notes that scv.datasets.dentategyrus() may download data from the internet. This is a standard and expected behavior for the scVelo library, which fetches well-known academic tutorial datasets from official sources.

omicverse-single-cell-rna-velocity