The released installer could not start: it carried a better-sqlite3 build for the
host Node runtime instead of Electron, so the app died immediately with a module
version mismatch when it opened its database.
Packaging now proves the Electron build of every runtime-sensitive native module
before an installer or archive exists, and installers are produced only from that
verified tree, so the mistake cannot pass silently. The release pipelines run the
same check.
The default local model also pointed at a retired model: a *.gguf name that
Ollama cannot serve, while the settings, onboarding, and guide screens
recommended an older model. All of them now use the model the service code
already preferred.
Local dictation had never produced a transcript on an installed build. The
engine itself was healthy; every connection to it was broken.
Installed builds shipped no speech engine at all: the packaging config had no
entry for the faster-whisper sidecar and no pipeline step built one, so the app
always fell back to a system Python without the runtime. Development was broken
too, because the sidecar and SoX paths were resolved against the Vite output
directory instead of the app root, which also meant recording failed with a SoX
ENOENT. On hosts where localhost resolves only to IPv6, every local request was
refused outright, which silently disabled both local transcription and the local
LLM.
The sidecar is now built and bundled (including the Silero VAD data it needs),
gated by a packaging check that fails when the engine or its data is missing.
Paths are discovered from the app root and fail loudly when the engine is
absent. Local engine URLs are normalized to the IPv4 loopback, decoding is tuned
so repeated hallucinations cannot compound (the same transcript now takes about
a fifth of the time), the engine is warmed up at startup, and holding the hotkey
now shows the text forming live in the recording tip.
- verifying/manifest 상태(바이트 진행률 없음)에서 인디터미넌트 바 + 친화적 문구
— 9.6GB 검증 ~2분간 멈춘 것처럼 보이던 문제 해소
- OnboardingModal 중복 실행 가드 (runningRef)
- LocalLLMService.pullModel: 동일 모델 동시 pull은 기존 promise 합류
- LocalSTTService.downloadModel: /download 409는 실패가 아닌 기존 진행 합류
증상:
- STT 426ms로 빠른데 LLM refine 단계가 39초 소요 후 빈 문자열 반환
- _completeSession('') → finalText.length === 0이라 paste 호출 자체 건너뜀
- 사용자에게는 '느리고 paste 안 됨'으로 보임
원인:
- 사용자가 ollama pull qwen3:4b 한 직후 첫 호출 (모델 cold start 일부 있음)
- qwen3는 reasoning model이라 응답에 <think>...</think> 블록을 길게 출력
- system prompt에 '/no_think' 토큰 없음 → reasoning mode ON
- generate()는 data.response.trim() 그대로 반환 → think 블록 + 빈 본문이면
trim 후 빈 문자열
- 빈 문자열에 대한 fallback이 없어서 그대로 _completeSession('')
수정:
- SYSTEM_PROMPTS 모두에 '/no_think' 헤더 추가
- qwen3 reasoning 비활성화 → 응답 속도 대폭 단축
- 다른 모델(llama, mistral, gemma)은 토큰 무시 → 호환성 OK
- stripReasoningBlocks() 추가
- <think>...</think> + <thinking>...</thinking> 블록 제거 (gi flag)
- /no_think를 무시하는 모델 + 응답에 think tag가 섞여 들어오는 케이스 안전망
- processText() 결과:
- stripReasoningBlocks(result.text)
- 빈 문자열이면 원본 transcript fallback + warn 로그
- npm workspaces 루트 (apps/*, packages/*) 세팅
- V1 전체를 apps/desktop/으로 git mv (src, resources, tests, sidecar,
scripts, electron.vite.config.ts, electron-builder.yml, vitest.config.ts,
tsconfig.node.json, tsconfig.web.json)
- apps/desktop/package.json 신규 (name=@d3ro/desktop)
- productName: 'd3ro-voice' 명시 — app.getName()을 고정하여 userData 경로
%APPDATA%\d3ro-voice\ 그대로 유지 (기존 DB/설정 연속성 보장)
- 루트 package.json을 workspace 루트로 재구성, 공통 devDep만 유지
(typescript, eslint, prettier)
- turbo.json, tsconfig.base.json 추가 (Turborepo 자체 설치는 별도 sub-phase)
- memory/project_status.md 생성 (규칙 13)
검증:
- npm run typecheck 통과
- npm run build 통과 (electron-vite main+preload+renderer)
- npm run dev 실제 실행 → DB/핫키/Ollama 자동 실행 모두 정상
2026-04-08 14:04:41 +09:00
Renamed from src/main/services/LocalLLMService.ts (Browse further)
