v0.3.0 — run every model locally

An interview now runs with no LLM, speech-to-text or text-to-speech key at all.

LLM_PROVIDER=ollama  STT_PROVIDER=whisper  TTS_PROVIDER=kokoro

No new dependency was added — this reuses the already-installed
livekit-plugins-openai with a base_url override. Provider values name a
contract ("an OpenAI-compatible server at a base URL"), not a vendor, so the
stack is genuinely swappable:

Stage	Endpoint	Accepted values
LLM	/v1/chat/completions	ollama · vllm · llamacpp · lmstudio · local
STT	/v1/audio/transcriptions	whisper · faster-whisper · qwen3-asr · speaches · local
TTS	/v1/audio/speech	kokoro · local

Also ships deepinterview init → "100% local models", a
docker compose --profile local stack, and docs/LOCAL_MODELS.md.

Verified

Apple M5 Pro (24 GB), qwen3:8b + Speaches + kokoro-fastapi:

• A live microphone interview, end to end — the agent spoke in Kokoro's
  voice, the local Whisper server transcribed real speech, and the report
  rendered complete, with zero recognition failures and no stage falling back
  to a cloud provider.
• Prep → question plan in ~2 minutes, CV-grounded with difficulty ramping
  1→5 and no degraded stages.
• 216 pytest + 26 vitest, lint, schema parity, compose (all profiles) and the
  agent image build, all green.

Not verified, so not claimed

• Turn latency is not benchmarked. Optimizing it is follow-up work.
• Kokoro has no Vietnamese voice — vi sessions honestly fall back to a
  cloud voice rather than mispronouncing it.
• LiveKit remains the transport. The claim is "no cloud model keys", never
  "no cloud"; livekit-server --dev is documented for a fully offline stack.
• CI has no models, so the local path is maintainer-verified only.
• Small local models lean on the shutdown-time answer recovery more than cloud
  models do — qwen3:8b did not reliably call save_answer during the live run.

Two defaults worth knowing

• WHISPER_MODEL=Systran/faster-whisper-base, not small. The OpenAI plugin
  hard-codes a 30s per-request ceiling no setting can raise, so an oversized
  model doesn't degrade — the turn dies. small needs ~3.4 GB resident and, in a
  memory-tight Docker VM, took 32s on a 3.6s clip; base needs ~220 MB and was
  just as accurate.
• KOKORO_MODEL=tts-1 is load-bearing: any other id routes synthesis down an
  SSE branch that yields no audio and no error. A wrong value is coerced with a
  warning.

Closes #55, #56, #57.

🤖 Generated with Claude Code