The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the Cell Ontology (CL) from the OBO Foundry mirror and NLP models from HuggingFace. These are well-known, trusted scientific and technology resources. The download logic is idempotent and provides options for local caching (local_model_dir, embeddings_path).
[CREDENTIALS_UNSAFE]: The skill correctly instructs users to provide API keys as arguments to setup_llm_expansion. While it uses a placeholder example (sk-...), it follows standard practices for secret management rather than hardcoding sensitive information. It also supports local OpenAI-compatible endpoints via base_url for privacy-conscious environments.
[COMMAND_EXECUTION]: The skill utilizes high-level Python libraries (omicverse, scanpy, sentence-transformers) for its operations. No arbitrary shell command execution or unsafe subprocess calls were detected.
[DATA_EXFILTRATION]: Outbound network activity is limited to downloading required ontology/model files and optional LLM expansion calls to OpenAI or user-defined endpoints. No patterns for exfiltrating sensitive local data (like .ssh or .aws credentials) were found.
[PROMPT_INJECTION]: The skill uses LLMs for abbreviation expansion and match selection. While this involves processing author-provided cell labels, the risk of indirect prompt injection is minimal due to the narrow scope of the task (mapping biological terms) and the use of specific context parameters (tissue_context, species).

omicverse-single-cell-cellmatch-ontology