OpenClaw and Hermes messages are now sent exactly as spoken — the injected "voice assistant request... keep it brief" instruction is gone, so responses follow your agent's own prompts and style. If TTS starts reading markdown or emojis aloud, enable the per-profile "Clean Responses for Voice Output" option.