feat: BYO-LLM conversational voice mode — OpenAI-compatible endpoint + voice agent loop

[andreolf]

Summary

Turns Voicebox from a TTS studio into a local voice agent: connect any OpenAI-compatible LLM (Ollama, LM Studio, OpenAI, Anthropic, etc.), speak or type to it, and hear replies in a cloned voice — running entirely on your machine.

This is the text-LLM bridge that the roadmap lists as the stepping stone before end-to-end speech LLMs (Moshi, GLM-4-Voice). It reuses the entire existing generation pipeline unchanged.

What's new

Backend

• ConversationSettings ORM singleton (same pattern as CaptureSettings) — enabled, llm_endpoint, llm_api_key, llm_model, system_prompt_prefix
• backend/backends/openai_compat_backend.py — OpenAICompatLLMBackend via httpx (already a dependency); implements both generate() and generate_with_history() for multi-turn chat
• POST /conversation/turn — validates settings → calls external LLM with the profile's personality as system prompt + conversation history → enqueues TTS via the existing background job pipeline → returns assistant_text + generation_id immediately (non-blocking, same pattern as /generate)
• GET /PUT /settings/conversation — settings CRUD, same pattern as captures/generation settings

Frontend

• ConversationPanel.tsx — chat UI with user/assistant bubbles, Enter-to-send, polls SSE endpoint and auto-plays audio via the existing audio player store
• ConversationPage.tsx — settings page: enable toggle, endpoint URL, API key (password field), model name, system-prompt textarea, provider compatibility sidebar (Ollama · LM Studio · OpenAI · any compatible API)
• "Conversation" entry added to settings nav + TanStack Router
• "Voice Chat" tab in VoiceInspector — visible only when conversation mode is enabled

Compatibility

Works out of the box with any server that speaks /v1/chat/completions:

• Ollama — http://localhost:11434, no API key
• LM Studio — http://localhost:1234, no API key
• OpenAI — https://api.openai.com, API key required
• Anthropic — via compatible proxy

Test plan

• Enable conversation mode in Settings → Conversation
• Set endpoint to a running Ollama instance (ollama run llama3)
• Open a voice profile → Voice Chat tab → send a message
• Verify reply text appears and audio plays in the cloned voice
• Disable conversation mode → Voice Chat tab disappears
• Test with missing endpoint → 400 error surfaced in UI

🤖 Generated with Claude Code

[ext-sakamoro]

Hi @andreolf — really nice work on this. I only spotted it after opening #947 / #948 today, which have some real overlap with the backend layer here:

• We both create backend/backends/openai_compat_backend.py with an OpenAICompatLLMBackend on the LLMBackend protocol using httpx. That's a merge conflict if both PRs land as-is.
• Feature scope diverges cleanly on top of it: this PR builds a new conversational voice chat UX (agent loop, ConversationSettings, ConversationPanel), while Add custom OpenAI-compatible LLM endpoint for refinement and personality #947 wires the same class into the existing refinement + personality-speak paths (extends CaptureSettings in place). Different DB tables, different frontend surfaces.

Whichever one @jamiepine wants first, happy to rebase mine onto the shared backend rather than duplicating it. If this PR lands first I'll pick up your OpenAICompatLLMBackend and drop mine; if #947 lands first, this one gets the same treatment.

Also — I noticed you're using the multi-turn generate_with_history() variant that isn't in the LLMBackend protocol yet. Probably worth adding it to the protocol so it stays available across backends, but that's a follow-up either way.

No rush from my side; happy to coordinate on whichever direction the maintainer prefers.