Resolve clone model before GPU lock to avoid wait stall.

Serialize clone background load against live CustomVoice CUDA use.

Defer Qwen3-TTS clone model load off the cold-start critical path.

Add cold-start warmup progress via early health-only bind

Rename legacy ragchat wire names to HumaticAI in the voice stack.

Rename ragchat_adapter.py to humaticai_adapter.py.

Rebrand deploy as HumaticAI Voice with Digital Peers adapter naming.

Remove unused conv_id from runtime config and LLM forwarding.

Stop sending adapter identity keys to the HF Inference Providers router.

Do not send vLLM chat_template_kwargs to the HF Inference Providers router.

Simple HumaticAI/HF routing: hf default, loud errors, no fallbacks.

Persist HumaticAI thread in browser until user starts new chat

Fix HumaticAI thread continuity: server-minted thread_id only

Require per-user HumaticAI Digital Peer API key instead of hardcoded secret.

Dual CustomVoice+Base TTS, per-user LLM switch, and HumaticAI thread continuity.

Add voice style (instruct) on top of Qwen3 voice cloning

Add Qwen3 voice cloning: base model, upload API, per-user registry.

Per-browser identity: forward user_id + conv_id to RAGchat adapter

Vanilla upstream pipeline: revert voice-guard + dep tweaks, Gemma 4 LLM, no custom flags

Speech-to-speech realtime demo for HF Spaces GPU (parakeet + qwen3, port 7860)