The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill converts user voice input into text that is then processed by an LLM via the OpenClaw gateway, creating an attack surface for instructions embedded in audio to manipulate the agent.
Ingestion points: Microphone audio captured by the assistant and transcribed via SenseVoice or Zipformer models as specified in SKILL.md.
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions for the processed speech text.
Capability inventory: The skill uses WebSockets to communicate with the OpenClaw gateway and provides an HTTP API for shutdown and status monitoring.
Sanitization: No explicit sanitization or filtering of transcribed speech is described before it is passed to the LLM context.
[EXTERNAL_DOWNLOADS]: Installation involves fetching source code from GitHub and downloading binary models for ASR, TTS, and wake word detection.
Evidence: git clone https://github.com/RubinTry/assistant-x-openclaw.git and instructions to download models into the models/ directory.
[COMMAND_EXECUTION]: The skill relies on shell scripts to initialize the environment and manage the assistant process.
Evidence: Usage of scripts/start.sh and scripts\start.bat for application startup and device pairing.

assistant-x-openclaw-jarvis-voice-assistant