The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads microbiome models and the Atlas/Compass datasets from the official repository of a well-known service (Hugging Face). These operations require an HF_TOKEN for authenticated access to gated content.
[REMOTE_CODE_EXECUTION]: The instructions specify the use of trust_remote_code=True when loading the custom TaxonomicTokenizer via the Hugging Face transformers library. This is a standard mechanism for custom model components. The documentation proactively recommends pinning a specific revision hash to ensure the integrity of the downloaded code.
[COMMAND_EXECUTION]: The skill utilizes the waypoint CLI tool provided by the waypoint-bio package to perform data processing, model training, and evaluation tasks. These are documented administrative and scientific workflows.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted microbiome abundance data from various formats (MetaPhlAn, Kraken2, QIIME 2). While this constitutes an attack surface, the ingestion logic in scripts/profiler_to_waypoint.py and scripts/vocab_coverage.py is restricted to taxonomic normalization and numerical parsing. The models themselves are specialized causal LMs for taxonomic tokens, not general-purpose instruction-following agents.
[DYNAMIC_EXECUTION]: The script scripts/vocab_coverage.py uses ast.literal_eval to parse Python list representations from CSV files. This is a safe alternative to eval() that only evaluates literal structures, preventing arbitrary code execution during data ingestion.

waypoint-bio