The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches model weights, runtime code, and Python wheels from Hugging Face and PyPI (files.pythonhosted.org). These are well-known, established services for distributing machine learning models and software packages. Downloads are explicitly guarded by SHA256 checksum verification against hardcoded pins before use, as seen in scripts/qwen3.py and the various *-files.json manifests.
[REMOTE_CODE_EXECUTION]: The skill downloads and executes Python code stored in remote model repositories. While this involves running external code, the skill implements rigorous integrity checks, verifying the hash of every downloaded file before execution. scripts/ppaso.py dynamically loads and executes the verified runtime code using importlib.util.module_from_spec.
[COMMAND_EXECUTION]: The skill executes the ffmpeg command-line tool to perform audio normalization and verification. This is implemented in scripts/normalize_audio.py using subprocess.run to call ffmpeg with controlled arguments and audio filters.
[DYNAMIC_EXECUTION]: The skill uses dynamic library loading and dynamic imports to provide advanced features and runtime isolation. scripts/qwen3.py uses ctypes.CDLL to access macOS system APIs (/usr/lib/libproc.dylib) for accurate memory footprint reporting. scripts/ppaso.py uses importlib.import_module and sys.modules manipulation to load its local runtime environment from a hash-verified snapshot.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface as it processes user-provided Korean text manuscripts to generate audio narration. While this creates a potential surface for instructions embedded in the text, the skill is designed for the specific task of speech synthesis and does not exhibit patterns of unsafe instruction execution from the ingested data.

tk-tts