The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill contains functionality to download RAG (Retrieval-Augmented Generation) assets from Hugging Face (huggingface.co/datasets/JuneYao/nihaisha-rag-assets). These downloads are not automatic and require explicit user consent and separate commands. Hugging Face is a well-known service for AI assets.
[COMMAND_EXECUTION]: Several Python scripts in the evals/scripts/ directory utilize subprocess.run() and subprocess.Popen() to execute the codex command-line tool. This is a standard development workflow used for benchmarking and evaluating the agent's performance and does not incorporate unsanitized user input into the shell commands.
[INDIRECT_PROMPT_INJECTION]: The skill has a broad ingestion surface as it processes numerous Markdown files and JSONL evidence cards containing TCM course transcripts. While this creates an inherent surface for indirect prompt injection, the skill includes explicit instructions in SKILL.md for the agent to distinguish between course-derived claims and general reasoning, and to maintain strict safety boundaries.
[SAFE_PRACTICES]: The nihaisha_kg/rerank.py utility includes a robust error-sanitization function (sanitize_rerank_error) that uses regular expressions to proactively redact API keys, tokens, and passwords from logs and error messages, preventing accidental credential exposure.
[SAFETY_REQUIREMENTS]: The skill implements a sophisticated medical safety policy. It identifies high-risk keywords (e.g., tumor, emergency, pediatric, toxic herbs) and mandates that the agent provide disclaimers and redirect users to professional medical services.

nihaisha