The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill automatically installs Python packages from PyPI at runtime, including transformers, torch, soundfile, numpy, google-genai, and python-dotenv. It also downloads pre-trained ML models for Whisper, Kokoro-82M, and MusicGen to local cache directories (~/.cache/hyperframes/).
[REMOTE_CODE_EXECUTION]: The scripts/lib/bgm.mjs file contains logic to programmatically run pip install for required dependencies and immediately execute the resulting libraries via spawned Python processes. This creates a dependency on external package registries at execution time.
[COMMAND_EXECUTION]: The engine frequently spawns external processes using node, python3, ffmpeg, ffprobe, and npx. These commands are used to process audio files, transcode formats, and run the hyperframes CLI tools. Input parameters from audio_request.json are passed to these subprocesses.
[DYNAMIC_EXECUTION]: In scripts/lib/bgm.mjs, the skill assembles a complete Python script as a string literal, interpolating user-provided prompts and configuration, and executes it using the python3 -c flag. While it uses JSON.stringify for basic escaping, this remains a dynamic code generation pattern.
[INDIRECT_PROMPT_INJECTION]: The skill represents a significant attack surface for indirect injection by ingesting untrusted text and mood queries from audio_request.json and external transcripts. This data is interpolated directly into prompts for media generation models and command-line arguments without explicit sanitization beyond standard JSON encoding.
[CREDENTIALS_UNSAFE]: The skill contains logic in scripts/lib/heygen.mjs to search for and load sensitive API keys from environment variables, local .env files (searching up five parent directories), and a shared global credentials file at ~/.heygen/credentials.

hyperframes-media