The Agent Skills Directory

[SAFE]: The skill demonstrates exceptional security practices by providing local scripts for validating reinforcement learning (RL) environment contracts and inspecting model checkpoints without performing dangerous operations like unpickling or loading arbitrary native code. For example, scripts/inspect_checkpoint.py hashes and classifies checkpoint files without importing torch or pickle.
[SAFE]: The documentation (SKILL.md and references/) provides detailed guidance on secure installation using pinned versions and SHA256 hashes, as well as explicit warnings against common RL security pitfalls such as credential leakage in loggers like Weights & Biases or Neptune.
[SAFE]: The bundled scripts implement strict data validation. scripts/_common.py includes helpers for strict JSON loading, path resolution with root-escape protection, and regex-based rejection of credential-bearing keys (e.g., api_key, token) in configuration files.
[SAFE]: The skill instructions and scripts do not involve remote code execution, prompt injection, or data exfiltration. All external references are to well-known and trusted platforms like PyPI and GitHub.

pufferlib