Voice recognition panel (Windows only). Communicates with the VoiceAgent.Win.exe subprocess via JSON Lines over stdin/stdout. Protocol: Client → Agent: {"cmd":"start"} | {"cmd":"speak","text":"...","lang":"<iso2>"} | {"cmd":"stop"} Agent → Client: {"type":"ready"|"transcript"|"done"|"error", ...} The lang field is a two-letter ISO language code (e.g. "it", "en", "de") derived from . The VoiceAgent maps it to the appropriate TTS voice: supported languages → Kokoro neural TTS, unsupported → Windows SAPI fallback. == Streaming TTS buffering strategy == LLM responses are streamed token by token via . Tokens are accumulated into a buffer; a flush is triggered when sentence-ending punctuation (. ! ?) is found and there is text AFTER it (a new sentence starting). This produces the largest possible natural phrase instead of tiny fragments. Key design choices: - Punctuation at the LAST position is NOT flushed — we WAIT for more text to come. This avoids splitting "?" when the LLM might still be generating the next sentence. - Only . ! ? trigger splits — commas and semicolons are pauses the TTS handles naturally. - No timeout flush: waiting for punctuation (even 1-2s) is less annoying than chopped mid-sentence audio. The pause after . ! ? sounds natural to human listeners. Built by the BuildVoiceAgentPlugin MSBuild target. Not available on Linux.