The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's setup process involves cloning an external, unverified repository (https://github.com/78/xiaozhi-esp32-server.git) to the user's local machine. It also installs a large volume of Python packages, including some with low popularity like pyyml, which may pose a supply chain risk.
[COMMAND_EXECUTION]: The provided installation scripts (scripts/setup_server.sh and scripts/setup_client.sh) utilize sudo to perform system-level operations, including installing packages like libopus-dev, portaudio19-dev, and ffmpeg via the system's package manager.
[REMOTE_CODE_EXECUTION]: Instructions for setting up local LLM providers suggest the execution of a remote shell script from a well-known service (curl -fsSL https://ollama.com/install.sh | sh), which constitutes remote code execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio data through a real-time WebSocket stream that is converted to text (ASR) and passed to an LLM for response generation.
Ingestion points: WebSocket audio data stream in templates/voice_chat_api.py and templates/voice_chat_client.py.
Boundary markers: No specific delimiters or "ignore instructions" warnings are present when passing ASR output to the LLM context.
Capability inventory: The skill possesses network operation capabilities (API calls to 50+ providers), file system write access (saving audio to tmp/ folders), and shell execution via the included setup scripts.
Sanitization: No sanitization or filtering is performed on the recognized speech text before it is interpolated into the LLM prompt.
[DYNAMIC_EXECUTION]: The templates/minimal_client.py file uses dynamic module loading (__import__('random')) to generate randomized device identifiers at runtime.

realtime-voice-chat