| name | add-new-provider |
| description | Use when adding or switching an STT, TTS, or LLM backend (STT_BACKEND, TTS_BACKEND, LLM_BACKEND) in the ClinicGuard server, or adding provider-specific settings to server/config.py. Follow this checklist in order — each step is cross-coupled with the next. |
Adding a new provider (STT / TTS / LLM)
Adding a backend touches ~6 files in a strict order. Missing step 5 (LLM only) produces a silent fallback bug in summaries. Run every step from server/ (settings load .env relative to cwd).
Checklist
-
server/config.py — add Settings fields: a <name>_backend-style switch only if introducing a new backend category, plus the provider's key/model/voice fields. Match the existing comment style (e.g. the # --- STT --- blocks).
-
server/.env.example — add the new vars with one-line docs. Never touch the real .env.
-
server/plugins/<name>_<kind>.py — new subclass (e.g. livekit.agents.stt.STT / tts.TTS). Use whisper_stt.py / piper_tts.py as the template: module-level model cache + threading.Lock, asyncio.to_thread for blocking inference, offline-first downloads, prewarm() for local models. Register nothing here — the factory does it.
-
server/agent.py — add a branch in the matching factory build_llm() / build_tts() / build_stt() using a lazy import inside the function; add the _prewarm() hook if the new model is local.
-
LLM only: server/triage/summarizer.py — update _client() (~line 37) and _default_model() (~line 46). This is the dupe LLM switch that is routinely forgotten; if it stays on the old backend, summaries silently fall back.
-
server/pyproject.toml — add the SDK dependency, then uv sync.
-
Verify — run the verify-stack skill (server path: uv run python agent.py console and uvicorn + curl localhost:8000/health).
Fallback interaction
If the new backend is an LLM, consider wiring it into LLM_FALLBACK_BACKEND for 429 auto-failover (see fallback_llm.py). Default llm_backend is opencode; README may still read Groq-first — trust config.py + .env.example.
