The Agent Skills Directory

[SAFE]: The skill adopts a defensive 'planner' architecture where bundled Python scripts do not execute complex ML logic or external binaries directly; instead, they generate bounded JSON execution plans for the agent to follow with human oversight.\n- [SAFE]: The shared helper module _common.py implements a highly restrictive local_path validation function that explicitly prohibits parent directory traversal ('..'), home expansion ('~'), symlink components, and URL-like schemes (e.g., http, s3, hf) for all script operations.\n- [SAFE]: All file-reading utilities in the bundled scripts enforce strict resource limits (defined in _common.py), including hard caps on file sizes (8 GiB), total records (10M), and line lengths (1 MiB), which effectively mitigates resource exhaustion attacks during data validation.\n- [SAFE]: The model_artifact_inspector.py and tokenizer_compatibility.py scripts facilitate the inspection of model metadata and universe alignment using bounded text-based reading and a custom safe YAML parser, intentionally avoiding the use of unsafe deserialization methods such as pickle, joblib, or torch.load.\n- [SAFE]: The documentation and instructions consistently prioritize data privacy and security, providing explicit warnings against loading unverified checkpoints and requiring manual user approval for any network-dependent actions like Hugging Face Hub or BEDbase downloads.\n- [SAFE]: The corpus_auditor.py script provides a mechanism to detect patient/donor data leakage across training splits by checking grouping columns in manifests without ever disclosing or printing the actual sensitive metadata values.\n- [SAFE]: The build instructions for the legacy StarSpace dependency target a specific immutable Git commit hash from a recognized organization (Facebook Research) and guide the user through a transparent local compilation process rather than suggesting the execution of opaque third-party binaries.

geniml