Reduce default TTS playback rate

.env.example

@@ -1,28 +1,42 @@
-DISCORD_BOT_TOKEN=
-DISCORD_APPLICATION_ID=
-DISCORD_COMMAND_GUILD_ID=
-
-OLLAMA_BASE_URL=http://localhost:11434
-OLLAMA_MODEL=qwen3:0.6b
-OLLAMA_KEEP_ALIVE=5m
-OLLAMA_NUM_CTX=4096
-
 LOCAL_AI_VENV_PATH=.local-ai/.venv
-LOCAL_AI_CACHE_DIR=.local-ai/cache
-LOCAL_AI_PYTHON=
-LOCAL_STT_MODEL=tiny
-LOCAL_STT_DEVICE=auto
-LOCAL_STT_COMPUTE_TYPE=auto
-LOCAL_STT_BEAM_SIZE=1
-LOCAL_TTS_LANGUAGE=KR
-LOCAL_TTS_SPEAKER=KR
-LOCAL_TTS_DEVICE=auto
-LOCAL_TTS_SPEED=1.12
+# Windows면 보통 python 또는 py -3
+LOCAL_AI_PYTHON=python
 
-BOT_DEFAULT_LANGUAGE=ko
-MAX_CONVERSATION_TURNS=12
-LOCAL_AUDIO_SOURCE=
-LOCAL_AUDIO_SINK=
-LOCAL_SPEAKER_NAME=local-user
-DEBUG_TEXT_EVENTS=false
+# Windows: ffmpeg dshow 장치 이름
+# Linux: pactl list sources short 에서 monitor/source 이름
+AUDIO_SOURCE=
+DOCKER_BIN=
+
+DEBUG=false
+TTS_ENABLED=true
+TTS_IMAGE=realtime-voice-bot-melotts:v0.1.2
+TTS_LANGUAGE=KR
+TTS_SPEAKER=KR
+TTS_DEVICE=cpu
+TTS_SPEED=1.18
+TTS_PLAYBACK_RATE=2.2
+TTS_SDP_RATIO=0.22
+TTS_NOISE_SCALE=0.55
+TTS_NOISE_SCALE_W=0.75
+TTS_CACHE_DIR=.local-ai/tts-cache
+TTS_OUTPUT_DIR=.local-ai/tts-output
+OLLAMA_BASE_URL=http://127.0.0.1:11434
+OLLAMA_MODEL=qwen3:8b
+OLLAMA_KEEP_ALIVE=5m
+MAX_CONVERSATION_TURNS=6
+WHISPER_MODEL=large-v3-turbo
+WHISPER_LANGUAGE=ko
+WHISPER_DEVICE=auto
+WHISPER_COMPUTE_TYPE=auto
+WHISPER_BEAM_SIZE=2
+
+SEGMENT_START_THRESHOLD=900
+SEGMENT_CONTINUE_THRESHOLD=450
+SEGMENT_START_FRAMES=2
+SEGMENT_END_FRAMES=24
+SEGMENT_PREROLL_SAMPLES=3200
+SEGMENT_MIN_SPEECH_SAMPLES=7200
+SEGMENT_MAX_SPEECH_SAMPLES=160000
+
+DEBUG_TRANSCRIPTS=true
 LOG_LEVEL=info

.gitignore

@@ -2,5 +2,5 @@ node_modules
 dist
 .env
 .local-ai
-__pycache__
 *.pyc
+__pycache__

.vscode/settings.json

@@ -0,0 +1,5 @@
+{
+  "terminal.integrated.env.windows": {
+    "PATH": "${env:PATH};C:\\Program Files\\Docker\\Docker\\resources\\bin"
+  }
+}

README.md

@@ -1,143 +1,233 @@
 # realtime_voice_bot
 
-디스코드 음성 채널 또는 로컬 PC 마이크에서 한국어 음성을 인식하고, 완전 로컬 스택으로 답변을 생성한 뒤 다시 음성으로 읽어주는 최소 프로토타입입니다.
+출력장치로 재생되는 소리를 파일 저장 없이 바로 받아서 `faster-whisper`로 STT 테스트를 하고, 필요하면 `Ollama` LLM과 연결된 통합 테스트와 LLM CLI 테스트를 할 수 있는 최소 프로토타입입니다.
 
-## 현재 스택
+현재 문서는 **Windows PC에서 실행하는 기준**으로 적었습니다.
 
-- STT: `faster-whisper` + Whisper multilingual
-- LLM: `Ollama` + `qwen3:0.6b`
-- TTS: `MeloTTS` Korean
-- VAD: `avr-vad`
+## 현재 범위
 
-외부 유료 API나 무료 한도형 API는 쓰지 않습니다.
-
-## 현재 구현 범위
-
-- Discord slash command 기반 제어: `/join`, `/leave`, `/status`, `/reset`, `/say`
-- 로컬 테스트 모드: PC 마이크로 직접 말하고 바로 응답 확인
-- `@discordjs/voice` 기반 음성 채널 입장 및 유저별 오디오 수신
-- 48k stereo PCM을 16k mono로 내려서 유저별 VAD 처리
-- 화자 발화 시작 시 현재 재생과 대기열 즉시 중단
-- Python 로컬 워커를 한 번 띄워 STT/TTS 모델을 메모리에 유지
-
-## 필수 준비물
-
-- Bun `1.3+`
-- Node.js `22.12+`
-- Python `3.11+`
-- `ffmpeg`
-- Ollama
-
-Discord 모드까지 쓸 거면 추가로:
-
-- Discord bot token
-- Discord application id
+- Node.js + TypeScript 메인 프로세스
+- 출력 오디오 실시간 캡처
+- 메모리 버퍼 기반 간단한 저지연 발화 분리
+- 미리 로드한 `faster-whisper` 워커에 PCM 직접 전달
+- 디스크에 WAV 저장 없이 바로 전사
+- STT 전용 테스트
+- STT 결과에 대해 답변 가치 판단 후 필요할 때만 LLM 답변하는 통합 테스트
+- 로컬 `Ollama` LLM 에이전트 CLI 테스트
+- 무료 로컬 `MeloTTS` 기반 음성 출력 테스트
 
 ## 빠른 시작
 
-```bash
+```bat
 bun install
-ollama pull qwen3:0.6b
-bun run setup:local-ai
+bun run setup
+copy .env.example .env
 ```
 
-그다음 로컬 장치 확인:
+또는 전체 준비를 명시적으로:
 
-```bash
+```bat
+bun run setup:all
+copy .env.example .env
+```
+
+장치 목록 확인:
+
+```bat
 bun run devices
 ```
 
 실행:
 
-```bash
-bun run start:local
+```bat
+bun run test:stt
 ```
 
-Discord 모드:
+STT + LLM 통합 테스트:
 
-```bash
-bun run start:discord
+```bat
+bun run test:sttllm
+```
+
+STT + LLM + TTS 전체 연결 테스트:
+
+```bat
+bun run test:all
+```
+
+LLM 단독 테스트:
+
+```bat
+bun run test:llm
+```
+
+TTS 단독 테스트:
+
+```bat
+bun run test:tts -- "안녕하세요. 로컬 티티에스 테스트입니다."
 ```
 
 ## 환경 변수
 
-`.env.example`를 복사해서 `.env`를 채우면 됩니다.
-
-Discord 모드에서만 필수:
-
-- `DISCORD_BOT_TOKEN`
-- `DISCORD_APPLICATION_ID`
-
-기본값이 이미 들어있는 로컬 AI 설정:
-
+- `AUDIO_SOURCE`
+  - `bun run devices` 에서 보이는 `ffmpeg dshow` 오디오 장치 이름
+  - 보통 `Stereo Mix`, 오디오 인터페이스 loopback 채널, 가상 케이블 입력 같은 이름을 넣습니다
+- `DOCKER_BIN`
+  - 비워두면 자동 탐색
+  - VSCode가 오래 떠 있어서 `docker` PATH를 못 잡을 때만 설정
+  - 예: `C:\Program Files\Docker\Docker\resources\bin\docker.exe`
+- `DEBUG`
+  - `true`면 상세 로그 출력
+  - `false`면 전사 결과만 출력
+- `WHISPER_MODEL`
+  - 기본값 `large-v3-turbo`
 - `OLLAMA_BASE_URL`
+  - 기본값 `http://127.0.0.1:11434`
 - `OLLAMA_MODEL`
+  - 기본값 `qwen3:8b`
+- `TTS_ENABLED`
+  - 기본값 `true`
+- `TTS_IMAGE`
+  - 기본값 `realtime-voice-bot-melotts:v0.1.2`
+- `TTS_LANGUAGE`
+  - 기본값 `KR`
+- `TTS_SPEAKER`
+  - 기본값 `KR`
+- `TTS_DEVICE`
+  - 기본값 `cpu`
+  - Docker GPU passthrough를 쓸 때만 `cuda`로 바꿉니다
+- `TTS_SPEED`
+  - 기본값 `1.18`
+- `TTS_PLAYBACK_RATE`
+  - 기본값 `2.2`
+  - 생성된 WAV를 `ffmpeg`로 더 빠르게 재생합니다
+- `TTS_SDP_RATIO`
+  - 기본값 `0.22`
+- `TTS_NOISE_SCALE`
+  - 기본값 `0.55`
+- `TTS_NOISE_SCALE_W`
+  - 기본값 `0.75`
 - `OLLAMA_KEEP_ALIVE`
-- `OLLAMA_NUM_CTX`
-- `LOCAL_AI_VENV_PATH`
-- `LOCAL_AI_CACHE_DIR`
-- `LOCAL_STT_MODEL`
-- `LOCAL_STT_DEVICE`
-- `LOCAL_STT_COMPUTE_TYPE`
-- `LOCAL_STT_BEAM_SIZE`
-- `LOCAL_TTS_LANGUAGE`
-- `LOCAL_TTS_SPEAKER`
-- `LOCAL_TTS_DEVICE`
-- `LOCAL_TTS_SPEED`
-
-선택:
-
-- `DISCORD_COMMAND_GUILD_ID`
-  - 테스트 서버에만 slash command를 즉시 반영하려면 설정
-- `LOCAL_AI_PYTHON`
-  - Python 경로 자동 탐지가 안 되면 설정
-  - 예시: `python`
-  - Windows 예시: `py -3`
-- `LOCAL_AUDIO_SOURCE`
-  - 로컬 입력 장치
-  - Linux는 `pw-record --target`, Windows는 `ffmpeg dshow` 장치 이름
-- `LOCAL_AUDIO_SINK`
-  - Linux 로컬 출력 장치
-  - Windows는 현재 시스템 기본 출력 장치 사용
-- `LOCAL_SPEAKER_NAME`
-  - 로컬 테스트에서 프롬프트에 넣을 화자 이름
-- `BOT_DEFAULT_LANGUAGE`
+  - 기본값 `5m`
+- `MAX_CONVERSATION_TURNS`
+  - 기본값 `6`
+  - 최근 대화 몇 턴까지 LLM 문맥으로 넘길지 정합니다
+- `WHISPER_LANGUAGE`
   - 기본값 `ko`
-- `DEBUG_TEXT_EVENTS`
-  - `true`면 transcript/reply를 콘솔에 같이 출력
+- `WHISPER_DEVICE`
+  - `auto`, `cuda`, `cpu`
+- `WHISPER_COMPUTE_TYPE`
+  - `auto`, `float16`, `int8_float16`, `int8`, `float32`
+- `WHISPER_BEAM_SIZE`
+  - 기본값 `2`
+- `SEGMENT_END_FRAMES`
+  - 기본값 `24`
+  - 끝을 조금 더 늦게 잘라서 문장이 잘게 끊기는 현상을 줄입니다
+- `SEGMENT_MAX_SPEECH_SAMPLES`
+  - 기본값 `160000`
+  - 너무 긴 구간은 강제로 끊어서 지연이 과하게 커지는 걸 막습니다
 
-## 속도 우선 기본값
+## 메모
 
-- STT 기본 모델은 `tiny`
-- LLM 기본 모델은 `qwen3:0.6b`
-- TTS 기본 속도는 `1.12`
+- 이 버전은 `STT`, `STT+LLM`, `LLM` 테스트를 따로 제공합니다.
+- `test:sttllm`은 STT와 LLM만 연결합니다.
+- `test:all`은 STT, LLM, TTS를 모두 연결합니다.
+- `test:all`에서는 자기 음성을 다시 전사하지 않도록 TTS 재생 중에는 캡처를 잠시 멈춥니다.
+- LLM 프롬프트는 `prompts/*.md` 에 분리되어 있습니다.
+- 최소 지연을 위해 파일 저장은 하지 않습니다.
+- VAD는 현재 모델 기반이 아니라 진폭 기반 단순 분리입니다.
+- Windows에서는 보통 출력 루프백이 가능한 장치나 `Stereo Mix`, 오디오 인터페이스 loopback 채널을 `AUDIO_SOURCE`로 잡아야 합니다.
+- 단순히 스피커 이름을 넣는 구조가 아니라, **루프백/캡처 가능한 입력 장치 이름**을 넣어야 합니다.
+- `ffmpeg`가 PATH에 잡혀 있어야 합니다.
+- TTS는 Windows에서 Docker Desktop이 필요합니다. MeloTTS 공식 문서도 Windows/macOS에서는 Docker 실행을 권장합니다.
+- `cmd` 기준으로 `.env`는 `copy .env.example .env`로 만들면 됩니다.
 
-정확도가 아쉬우면:
-
-```env
-LOCAL_STT_MODEL=small
-OLLAMA_MODEL=qwen3:1.7b
-```
-
-## 로컬 테스트 순서
+## Windows 테스트 순서
 
 1. `bun install`
-2. `ollama pull qwen3:0.6b`
-3. `bun run setup:local-ai`
+2. `bun run setup:stt`
+3. `copy .env.example .env`
 4. `bun run devices`
-5. 필요하면 `.env` 에 `LOCAL_AUDIO_SOURCE` 설정
-6. `bun run start:local`
+5. `.env`에서 `AUDIO_SOURCE=`에 루프백 장치 이름 입력
+6. `bun run test:stt`
+7. 유튜브, 디스코드 통화, 동영상 같은 소리를 재생해서 전사만 확인
 
-## Windows 메모
+## Windows STT+LLM 통합 테스트 순서
 
-- `bun run devices` 와 Windows 로컬 녹음은 `ffmpeg`가 필요합니다.
-- 출력 장치 직접 선택은 아직 미구현이라 시스템 기본 출력 장치로 재생됩니다.
-- Python 탐지가 안 되면 `.env` 에 `LOCAL_AI_PYTHON=py -3` 또는 `LOCAL_AI_PYTHON=python` 을 넣으면 됩니다.
+1. `bun run setup:llm`
+2. `bun run setup:tts`
+3. `bun run test:sttllm`
+4. 유튜브, 디스코드 통화, 동영상 같은 소리를 재생해서 전사와 답변 확인
 
-## 설계 메모
+## Windows 전체 연결 테스트 순서
 
-- 입력은 유저별 병렬 처리
-- 출력은 길드 세션당 단일 큐
-- 로컬 모드는 단일 화자 입력 기준
-- 화자 구분은 `speaker_id`, `speaker_name`을 LLM 프롬프트에 항상 포함
-- 모델 다운로드 캐시는 기본적으로 `.local-ai/cache` 아래에 저장
+1. `bun run setup:llm`
+2. `bun run setup:tts`
+3. `bun run test:all`
+4. 유튜브, 디스코드 통화, 동영상 같은 소리를 재생해서 전사, 답변, 음성 출력 확인
+
+## Windows LLM 테스트 순서
+
+1. `bun run setup:llm`
+2. `bun run test:llm`
+3. 콘솔에 직접 문장을 입력하고 답변 확인
+4. `/reset` 으로 문맥 초기화, `/exit` 로 종료
+
+## Windows TTS 테스트 순서
+
+1. Docker Desktop 실행
+2. `bun run setup:tts`
+3. `bun run test:tts -- "안녕하세요. 로컬 티티에스 테스트입니다."`
+
+현재 `test:llm` 에이전트 도구:
+- 현재 시간 조회
+- 현재 런타임 설정 조회
+- 주요 bun 명령 목록 조회
+- 간단한 산술식 계산
+- 웹 검색
+- URL 본문 읽기
+
+동작 원칙:
+- 일반 대화는 로컬 LLM만 답변
+- 최신 정보, 뉴스, 사실 확인, 검색 요청일 때만 웹 도구 사용
+- `test:sttllm` 에서는 먼저 "대답할 가치가 있는 텍스트인지" 판정한 뒤 필요할 때만 답변
+- 웹 검색이 실제로 시작되면 결과 전에 `검색해볼게요.` 같은 진행 메시지를 먼저 출력
+- 답변에 한글 외 다른 문자군이 섞이면 한국어만으로 한 번 더 교정
+
+현재 프롬프트 파일:
+- `prompts/assistant.md`
+- `prompts/reply-gate.md`
+- `prompts/rewrite-korean.md`
+
+## Windows용 .env 예시
+
+```env
+LOCAL_AI_PYTHON=python
+AUDIO_SOURCE=
+DEBUG=false
+TTS_ENABLED=true
+TTS_IMAGE=realtime-voice-bot-melotts:v0.1.2
+TTS_LANGUAGE=KR
+TTS_SPEAKER=KR
+TTS_DEVICE=cpu
+TTS_SPEED=1.18
+TTS_PLAYBACK_RATE=2.2
+TTS_SDP_RATIO=0.22
+TTS_NOISE_SCALE=0.55
+TTS_NOISE_SCALE_W=0.75
+TTS_CACHE_DIR=.local-ai/tts-cache
+TTS_OUTPUT_DIR=.local-ai/tts-output
+OLLAMA_BASE_URL=http://127.0.0.1:11434
+OLLAMA_MODEL=qwen3:8b
+OLLAMA_KEEP_ALIVE=5m
+MAX_CONVERSATION_TURNS=6
+WHISPER_MODEL=large-v3-turbo
+WHISPER_LANGUAGE=ko
+WHISPER_DEVICE=auto
+WHISPER_COMPUTE_TYPE=auto
+WHISPER_BEAM_SIZE=2
+SEGMENT_END_FRAMES=24
+SEGMENT_MAX_SPEECH_SAMPLES=160000
+DEBUG_TRANSCRIPTS=true
+LOG_LEVEL=info
+```

bun.lock

@@ -5,13 +5,7 @@
     "": {
       "name": "realtime_voice_bot",
       "dependencies": {
-        "@discordjs/voice": "^0.19.2",
-        "avr-vad": "^1.0.10",
-        "discord.js": "^14.26.3",
         "dotenv": "^17.4.2",
-        "ffmpeg-static": "^5.3.0",
-        "opusscript": "^0.1.1",
-        "prism-media": "^1.3.5",
         "zod": "^4.3.6",
       },
       "devDependencies": {
@@ -20,203 +14,15 @@
       },
     },
   },
-  "trustedDependencies": [
-    "ffmpeg-static",
-    "onnxruntime-node",
-  ],
   "packages": {
-    "@derhuerst/http-basic": ["@derhuerst/http-basic@8.2.4", "", { "dependencies": { "caseless": "^0.12.0", "concat-stream": "^2.0.0", "http-response-object": "^3.0.1", "parse-cache-control": "^1.0.1" } }, "sha512-F9rL9k9Xjf5blCz8HsJRO4diy111cayL2vkY2XE4r4t3n0yPXVYy3KD3nJ1qbrSn9743UWSXH4IwuCa/HWlGFw=="],
-
-    "@discordjs/builders": ["@discordjs/builders@1.14.1", "", { "dependencies": { "@discordjs/formatters": "^0.6.2", "@discordjs/util": "^1.2.0", "@sapphire/shapeshift": "^4.0.0", "discord-api-types": "^0.38.40", "fast-deep-equal": "^3.1.3", "ts-mixer": "^6.0.4", "tslib": "^2.6.3" } }, "sha512-gSKkhXLqs96TCzk66VZuHHl8z2bQMJFGwrXC0f33ngK+FLNau4hU1PYny3DNJfNdSH+gVMzE85/d5FQ2BpcNwQ=="],
-
-    "@discordjs/collection": ["@discordjs/collection@1.5.3", "", {}, "sha512-SVb428OMd3WO1paV3rm6tSjM4wC+Kecaa1EUGX7vc6/fddvw/6lg90z4QtCqm21zvVe92vMMDt9+DkIvjXImQQ=="],
-
-    "@discordjs/formatters": ["@discordjs/formatters@0.6.2", "", { "dependencies": { "discord-api-types": "^0.38.33" } }, "sha512-y4UPwWhH6vChKRkGdMB4odasUbHOUwy7KL+OVwF86PvT6QVOwElx+TiI1/6kcmcEe+g5YRXJFiXSXUdabqZOvQ=="],
-
-    "@discordjs/rest": ["@discordjs/rest@2.6.1", "", { "dependencies": { "@discordjs/collection": "^2.1.1", "@discordjs/util": "^1.2.0", "@sapphire/async-queue": "^1.5.3", "@sapphire/snowflake": "^3.5.5", "@vladfrangu/async_event_emitter": "^2.4.6", "discord-api-types": "^0.38.40", "magic-bytes.js": "^1.13.0", "tslib": "^2.6.3", "undici": "6.24.1" } }, "sha512-wwQdgjeaoYFiaG+atbqx6aJDpqW7JHAo0HrQkBTbYzM3/PJ3GweQIpgElNcGZ26DCUOXMyawYd0YF7vtr+fZXg=="],
-
-    "@discordjs/util": ["@discordjs/util@1.2.0", "", { "dependencies": { "discord-api-types": "^0.38.33" } }, "sha512-3LKP7F2+atl9vJFhaBjn4nOaSWahZ/yWjOvA4e5pnXkt2qyXRCHLxoBQy81GFtLGCq7K9lPm9R517M1U+/90Qg=="],
-
-    "@discordjs/voice": ["@discordjs/voice@0.19.2", "", { "dependencies": { "@snazzah/davey": "^0.1.9", "@types/ws": "^8.18.1", "discord-api-types": "^0.38.41", "prism-media": "^1.3.5", "tslib": "^2.8.1", "ws": "^8.19.0" } }, "sha512-3yJ255e4ag3wfZu/DSxeOZK1UtnqNxnspmLaQetGT0pDkThNZoHs+Zg6dgZZ19JEVomXygvfHn9lNpICZuYtEA=="],
-
-    "@discordjs/ws": ["@discordjs/ws@1.2.3", "", { "dependencies": { "@discordjs/collection": "^2.1.0", "@discordjs/rest": "^2.5.1", "@discordjs/util": "^1.1.0", "@sapphire/async-queue": "^1.5.2", "@types/ws": "^8.5.10", "@vladfrangu/async_event_emitter": "^2.2.4", "discord-api-types": "^0.38.1", "tslib": "^2.6.2", "ws": "^8.17.0" } }, "sha512-wPlQDxEmlDg5IxhJPuxXr3Vy9AjYq5xCvFWGJyD7w7Np8ZGu+Mc+97LCoEc/+AYCo2IDpKioiH0/c/mj5ZR9Uw=="],
-
-    "@emnapi/core": ["@emnapi/core@1.10.0", "", { "dependencies": { "@emnapi/wasi-threads": "1.2.1", "tslib": "^2.4.0" } }, "sha512-yq6OkJ4p82CAfPl0u9mQebQHKPJkY7WrIuk205cTYnYe+k2Z8YBh11FrbRG/H6ihirqcacOgl2BIO8oyMQLeXw=="],
-
-    "@emnapi/runtime": ["@emnapi/runtime@1.10.0", "", { "dependencies": { "tslib": "^2.4.0" } }, "sha512-ewvYlk86xUoGI0zQRNq/mC+16R1QeDlKQy21Ki3oSYXNgLb45GV1P6A0M+/s6nyCuNDqe5VpaY84BzXGwVbwFA=="],
-
-    "@emnapi/wasi-threads": ["@emnapi/wasi-threads@1.2.1", "", { "dependencies": { "tslib": "^2.4.0" } }, "sha512-uTII7OYF+/Mes/MrcIOYp5yOtSMLBWSIoLPpcgwipoiKbli6k322tcoFsxoIIxPDqW01SQGAgko4EzZi2BNv2w=="],
-
-    "@napi-rs/wasm-runtime": ["@napi-rs/wasm-runtime@1.1.4", "", { "dependencies": { "@tybys/wasm-util": "^0.10.1" }, "peerDependencies": { "@emnapi/core": "^1.7.1", "@emnapi/runtime": "^1.7.1" } }, "sha512-3NQNNgA1YSlJb/kMH1ildASP9HW7/7kYnRI2szWJaofaS1hWmbGI4H+d3+22aGzXXN9IJ+n+GiFVcGipJP18ow=="],
-
-    "@sapphire/async-queue": ["@sapphire/async-queue@1.5.5", "", {}, "sha512-cvGzxbba6sav2zZkH8GPf2oGk9yYoD5qrNWdu9fRehifgnFZJMV+nuy2nON2roRO4yQQ+v7MK/Pktl/HgfsUXg=="],
-
-    "@sapphire/shapeshift": ["@sapphire/shapeshift@4.0.0", "", { "dependencies": { "fast-deep-equal": "^3.1.3", "lodash": "^4.17.21" } }, "sha512-d9dUmWVA7MMiKobL3VpLF8P2aeanRTu6ypG2OIaEv/ZHH/SUQ2iHOVyi5wAPjQ+HmnMuL0whK9ez8I/raWbtIg=="],
-
-    "@sapphire/snowflake": ["@sapphire/snowflake@3.5.3", "", {}, "sha512-jjmJywLAFoWeBi1W7994zZyiNWPIiqRRNAmSERxyg93xRGzNYvGjlZ0gR6x0F4gPRi2+0O6S71kOZYyr3cxaIQ=="],
-
-    "@snazzah/davey": ["@snazzah/davey@0.1.11", "", { "optionalDependencies": { "@snazzah/davey-android-arm-eabi": "0.1.11", "@snazzah/davey-android-arm64": "0.1.11", "@snazzah/davey-darwin-arm64": "0.1.11", "@snazzah/davey-darwin-x64": "0.1.11", "@snazzah/davey-freebsd-x64": "0.1.11", "@snazzah/davey-linux-arm-gnueabihf": "0.1.11", "@snazzah/davey-linux-arm64-gnu": "0.1.11", "@snazzah/davey-linux-arm64-musl": "0.1.11", "@snazzah/davey-linux-x64-gnu": "0.1.11", "@snazzah/davey-linux-x64-musl": "0.1.11", "@snazzah/davey-wasm32-wasi": "0.1.11", "@snazzah/davey-win32-arm64-msvc": "0.1.11", "@snazzah/davey-win32-ia32-msvc": "0.1.11", "@snazzah/davey-win32-x64-msvc": "0.1.11" } }, "sha512-oBN+msHzPnm1M5DDx3wVD7iBwpNXFUtkh2MrAbUJu0OhKjliLChi28hq++mu1+qdMpAVQO5JKAvQQxYVbyneiw=="],
-
-    "@snazzah/davey-android-arm-eabi": ["@snazzah/davey-android-arm-eabi@0.1.11", "", { "os": "android", "cpu": "arm" }, "sha512-T1RYbNYKN6tLOcGIDKJd8OI6FBSEemwL7DOYdTMmhqfhhMr3YVN8WOhfoxGg63OcnpTN2e2c5tdY2bAx25RmQQ=="],
-
-    "@snazzah/davey-android-arm64": ["@snazzah/davey-android-arm64@0.1.11", "", { "os": "android", "cpu": "arm64" }, "sha512-ksJn/x2VU8h6w9eku1HT96ugSRZ7lKVkKNKbFleaFN+U99DJaPM+gMu2YvnFU4V54HR06ZBnRihnVG6VLXQpDw=="],
-
-    "@snazzah/davey-darwin-arm64": ["@snazzah/davey-darwin-arm64@0.1.11", "", { "os": "darwin", "cpu": "arm64" }, "sha512-E1d7PbaaVMO3Lj9EiAPqOVbuV0xg5+PsHzHH097DDXiD1+zUDXvJaTnUWsnm5z50pJniHpi4GtaYmk+ieB/guA=="],
-
-    "@snazzah/davey-darwin-x64": ["@snazzah/davey-darwin-x64@0.1.11", "", { "os": "darwin", "cpu": "x64" }, "sha512-Tl4TI/LTmgJZepgbgVMYDi8RqlAkPtPg1OEBPl7a9Tn3AwR36Vs6lyIT1cs/lGy/ds/+B+mKI4rPObN1cyILTw=="],
-
-    "@snazzah/davey-freebsd-x64": ["@snazzah/davey-freebsd-x64@0.1.11", "", { "os": "freebsd", "cpu": "x64" }, "sha512-T8Iw9FXkuI1T+YBAFzh9v/TXf9IOTOSqnd/BFpTRTrlW72PR2lhIidzSmg027VxO7r5pX47iFwiOkb9I/NU/EA=="],
-
-    "@snazzah/davey-linux-arm-gnueabihf": ["@snazzah/davey-linux-arm-gnueabihf@0.1.11", "", { "os": "linux", "cpu": "arm" }, "sha512-1Txj+8pqA8uq/OGtaUaBFWAPnNMQzFgIywj0iA7EI4xZl+mab48/pv+YZ1pNb/suC6ynsW44oB9efiXSdcUAgA=="],
-
-    "@snazzah/davey-linux-arm64-gnu": ["@snazzah/davey-linux-arm64-gnu@0.1.11", "", { "os": "linux", "cpu": "arm64" }, "sha512-ERzF5nM/IYW1BcN3wLXpEwBCGLFf0kGJUVhaV6yfiInz0tkU8UmvrrgpaMaACfMjIhfWdq5CcX+aTkXo/saNcg=="],
-
-    "@snazzah/davey-linux-arm64-musl": ["@snazzah/davey-linux-arm64-musl@0.1.11", "", { "os": "linux", "cpu": "arm64" }, "sha512-e6pX6Hiabtz99q+H/YHNkm9JVlpqN8HGh0qPib8G2+UY4/SSH8WvqWipk3v581dMy2oyCHt7MOoY1aU1P1N/xA=="],
-
-    "@snazzah/davey-linux-x64-gnu": ["@snazzah/davey-linux-x64-gnu@0.1.11", "", { "os": "linux", "cpu": "x64" }, "sha512-TW5bSoqChOJMbvsDb4wAATYrxmAXuNnse7wFNVSAJUaZKSeRfZbu3UAiPWSNn7GwLwSfU6hg322KZUn8IWCuvg=="],
-
-    "@snazzah/davey-linux-x64-musl": ["@snazzah/davey-linux-x64-musl@0.1.11", "", { "os": "linux", "cpu": "x64" }, "sha512-5j6Pmc+Wzv5lSxVP6quA7teYRJXibkZqQyYGfTDnTsUOO5dPpcojpqlXlkhyvsA1OAQTj4uxbOCciN3cVWwzug=="],
-
-    "@snazzah/davey-wasm32-wasi": ["@snazzah/davey-wasm32-wasi@0.1.11", "", { "dependencies": { "@napi-rs/wasm-runtime": "^1.1.2" }, "cpu": "none" }, "sha512-rKOwZ/0J8lp+4VEyOdMDBRP9KR+PksZpa9V1Qn0veMzy4FqTVKthkxwGqewheFe0SFg9fdvt798l/PBFrfDeZw=="],
-
-    "@snazzah/davey-win32-arm64-msvc": ["@snazzah/davey-win32-arm64-msvc@0.1.11", "", { "os": "win32", "cpu": "arm64" }, "sha512-5fptJU4tX901m3mj0SHiBljMrPT4ZEsynbBhR7bK1yn9TY1jjyhN8EFi7QF5IWtUEni+0mia2BCMHZ5ZkmFZqQ=="],
-
-    "@snazzah/davey-win32-ia32-msvc": ["@snazzah/davey-win32-ia32-msvc@0.1.11", "", { "os": "win32", "cpu": "ia32" }, "sha512-ualexn8SeLsiMHhWfzVrzRcjHgcBapg++FPaVgJJxoh2S/jCRiklXOu3luqIZdJdNKvhe2V9SwO/cImPeIIBKw=="],
-
-    "@snazzah/davey-win32-x64-msvc": ["@snazzah/davey-win32-x64-msvc@0.1.11", "", { "os": "win32", "cpu": "x64" }, "sha512-muNhc8UKXtknzsH/w4AIkbPR2I8BuvApn0pDXar0IEvY8PCjqU/M8MPbOOEYwQVvQRMwVTgExtxzrkBPSXB4nA=="],
-
-    "@tybys/wasm-util": ["@tybys/wasm-util@0.10.1", "", { "dependencies": { "tslib": "^2.4.0" } }, "sha512-9tTaPJLSiejZKx+Bmog4uSubteqTvFrVrURwkmHixBo0G4seD0zUxp98E1DzUBJxLQ3NPwXrGKDiVjwx/DpPsg=="],
-
     "@types/node": ["@types/node@25.6.0", "", { "dependencies": { "undici-types": "~7.19.0" } }, "sha512-+qIYRKdNYJwY3vRCZMdJbPLJAtGjQBudzZzdzwQYkEPQd+PJGixUL5QfvCLDaULoLv+RhT3LDkwEfKaAkgSmNQ=="],
 
-    "@types/ws": ["@types/ws@8.18.1", "", { "dependencies": { "@types/node": "*" } }, "sha512-ThVF6DCVhA8kUGy+aazFQ4kXQ7E1Ty7A3ypFOe0IcJV8O/M511G99AW24irKrW56Wt44yG9+ij8FaqoBGkuBXg=="],
-
-    "@vladfrangu/async_event_emitter": ["@vladfrangu/async_event_emitter@2.4.7", "", {}, "sha512-Xfe6rpCTxSxfbswi/W/Pz7zp1WWSNn4A0eW4mLkQUewCrXXtMj31lCg+iQyTkh/CkusZSq9eDflu7tjEDXUY6g=="],
-
-    "adm-zip": ["adm-zip@0.5.17", "", {}, "sha512-+Ut8d9LLqwEvHHJl1+PIHqoyDxFgVN847JTVM3Izi3xHDWPE4UtzzXysMZQs64DMcrJfBeS/uoEP4AD3HQHnQQ=="],
-
-    "agent-base": ["agent-base@6.0.2", "", { "dependencies": { "debug": "4" } }, "sha512-RZNwNclF7+MS/8bDg70amg32dyeZGZxiDuQmZxKLAlQjr3jGyLx+4Kkk58UO7D2QdgFIQCovuSuZESne6RG6XQ=="],
-
-    "avr-vad": ["avr-vad@1.0.10", "", { "dependencies": { "onnxruntime-node": "^1.22.0-rev" } }, "sha512-gM8SiQIebujfKMfy5w74tRPH+Fg78CMrBoDkMhCN3TmYVmmD8fmuVag7Q7ZCBITpFvYkOZnWEdGWuCb3YukBJw=="],
-
-    "boolean": ["boolean@3.2.0", "", {}, "sha512-d0II/GO9uf9lfUHH2BQsjxzRJZBdsjgsBiW4BvhWk/3qoKwQFjIDVN19PfX8F2D/r9PCMTtLWjYVCFrpeYUzsw=="],
-
-    "buffer-from": ["buffer-from@1.1.2", "", {}, "sha512-E+XQCRwSbaaiChtv6k6Dwgc+bx+Bs6vuKJHHl5kox/BaKbhiXzqQOwK4cO22yElGp2OCmjwVhT3HmxgyPGnJfQ=="],
-
-    "caseless": ["caseless@0.12.0", "", {}, "sha512-4tYFyifaFfGacoiObjJegolkwSU4xQNGbVgUiNYVUxbQ2x2lUsFvY4hVgVzGiIe6WLOPqycWXA40l+PWsxthUw=="],
-
-    "concat-stream": ["concat-stream@2.0.0", "", { "dependencies": { "buffer-from": "^1.0.0", "inherits": "^2.0.3", "readable-stream": "^3.0.2", "typedarray": "^0.0.6" } }, "sha512-MWufYdFw53ccGjCA+Ol7XJYpAlW6/prSMzuPOTRnJGcGzuhLn4Scrz7qf6o8bROZ514ltazcIFJZevcfbo0x7A=="],
-
-    "debug": ["debug@4.4.3", "", { "dependencies": { "ms": "^2.1.3" } }, "sha512-RGwwWnwQvkVfavKVt22FGLw+xYSdzARwm0ru6DhTVA3umU5hZc28V3kO4stgYryrTlLpuvgI9GiijltAjNbcqA=="],
-
-    "define-data-property": ["define-data-property@1.1.4", "", { "dependencies": { "es-define-property": "^1.0.0", "es-errors": "^1.3.0", "gopd": "^1.0.1" } }, "sha512-rBMvIzlpA8v6E+SJZoo++HAYqsLrkg7MSfIinMPFhmkorw7X+dOXVJQs+QT69zGkzMyfDnIMN2Wid1+NbL3T+A=="],
-
-    "define-properties": ["define-properties@1.2.1", "", { "dependencies": { "define-data-property": "^1.0.1", "has-property-descriptors": "^1.0.0", "object-keys": "^1.1.1" } }, "sha512-8QmQKqEASLd5nx0U1B1okLElbUuuttJ/AnYmRXbbbGDWh6uS208EjD4Xqq/I9wK7u0v6O08XhTWnt5XtEbR6Dg=="],
-
-    "detect-node": ["detect-node@2.1.0", "", {}, "sha512-T0NIuQpnTvFDATNuHN5roPwSBG83rFsuO+MXXH9/3N1eFbn4wcPjttvjMLEPWJ0RGUYgQE7cGgS3tNxbqCGM7g=="],
-
-    "discord-api-types": ["discord-api-types@0.38.47", "", {}, "sha512-XgXQodHQBAE6kfD7kMvVo30863iHX1LHSqNq6MGUTDwIFCCvHva13+rwxyxVXDqudyApMNAd32PGjgVETi5rjA=="],
-
-    "discord.js": ["discord.js@14.26.3", "", { "dependencies": { "@discordjs/builders": "^1.14.1", "@discordjs/collection": "1.5.3", "@discordjs/formatters": "^0.6.2", "@discordjs/rest": "^2.6.1", "@discordjs/util": "^1.2.0", "@discordjs/ws": "^1.2.3", "@sapphire/snowflake": "3.5.3", "discord-api-types": "^0.38.40", "fast-deep-equal": "3.1.3", "lodash.snakecase": "4.1.1", "magic-bytes.js": "^1.13.0", "tslib": "^2.6.3", "undici": "6.24.1" } }, "sha512-XEKtYn28YFsiJ5l4fLRyikdbo6RD5oFyqfVHQlvXz2104JhH/E8slN28dbky05w3DCrJcNVWvhVvcJCTSl/KIg=="],
-
     "dotenv": ["dotenv@17.4.2", "", {}, "sha512-nI4U3TottKAcAD9LLud4Cb7b2QztQMUEfHbvhTH09bqXTxnSie8WnjPALV/WMCrJZ6UV/qHJ6L03OqO3LcdYZw=="],
 
-    "env-paths": ["env-paths@2.2.1", "", {}, "sha512-+h1lkLKhZMTYjog1VEpJNG7NZJWcuc2DDk/qsqSTRRCOXiLjeQ1d1/udrUGhqMxUgAlwKNZ0cf2uqan5GLuS2A=="],
-
-    "es-define-property": ["es-define-property@1.0.1", "", {}, "sha512-e3nRfgfUZ4rNGL232gUgX06QNyyez04KdjFrF+LTRoOXmrOgFKDg4BCdsjW8EnT69eqdYGmRpJwiPVYNrCaW3g=="],
-
-    "es-errors": ["es-errors@1.3.0", "", {}, "sha512-Zf5H2Kxt2xjTvbJvP2ZWLEICxA6j+hAmMzIlypy4xcBg1vKVnx89Wy0GbS+kf5cwCVFFzdCFh2XSCFNULS6csw=="],
-
-    "es6-error": ["es6-error@4.1.1", "", {}, "sha512-Um/+FxMr9CISWh0bi5Zv0iOD+4cFh5qLeks1qhAopKVAJw3drgKbKySikp7wGhDL0HPeaja0P5ULZrxLkniUVg=="],
-
-    "escape-string-regexp": ["escape-string-regexp@4.0.0", "", {}, "sha512-TtpcNJ3XAzx3Gq8sWRzJaVajRs0uVxA2YAkdb1jm2YkPz4G6egUFAyA3n5vtEIZefPk5Wa4UXbKuS5fKkJWdgA=="],
-
-    "fast-deep-equal": ["fast-deep-equal@3.1.3", "", {}, "sha512-f3qQ9oQy9j2AhBe/H9VC91wLmKBCCU/gDOnKNAYG5hswO7BLKj09Hc5HYNz9cGI++xlpDCIgDaitVs03ATR84Q=="],
-
-    "ffmpeg-static": ["ffmpeg-static@5.3.0", "", { "dependencies": { "@derhuerst/http-basic": "^8.2.0", "env-paths": "^2.2.0", "https-proxy-agent": "^5.0.0", "progress": "^2.0.3" } }, "sha512-H+K6sW6TiIX6VGend0KQwthe+kaceeH/luE8dIZyOP35ik7ahYojDuqlTV1bOrtEwl01sy2HFNGQfi5IDJvotg=="],
-
-    "global-agent": ["global-agent@3.0.0", "", { "dependencies": { "boolean": "^3.0.1", "es6-error": "^4.1.1", "matcher": "^3.0.0", "roarr": "^2.15.3", "semver": "^7.3.2", "serialize-error": "^7.0.1" } }, "sha512-PT6XReJ+D07JvGoxQMkT6qji/jVNfX/h364XHZOWeRzy64sSFr+xJ5OX7LI3b4MPQzdL4H8Y8M0xzPpsVMwA8Q=="],
-
-    "globalthis": ["globalthis@1.0.4", "", { "dependencies": { "define-properties": "^1.2.1", "gopd": "^1.0.1" } }, "sha512-DpLKbNU4WylpxJykQujfCcwYWiV/Jhm50Goo0wrVILAv5jOr9d+H+UR3PhSCD2rCCEIg0uc+G+muBTwD54JhDQ=="],
-
-    "gopd": ["gopd@1.2.0", "", {}, "sha512-ZUKRh6/kUFoAiTAtTYPZJ3hw9wNxx+BIBOijnlG9PnrJsCcSjs1wyyD6vJpaYtgnzDrKYRSqf3OO6Rfa93xsRg=="],
-
-    "has-property-descriptors": ["has-property-descriptors@1.0.2", "", { "dependencies": { "es-define-property": "^1.0.0" } }, "sha512-55JNKuIW+vq4Ke1BjOTjM2YctQIvCT7GFzHwmfZPGo5wnrgkid0YQtnAleFSqumZm4az3n2BS+erby5ipJdgrg=="],
-
-    "http-response-object": ["http-response-object@3.0.2", "", { "dependencies": { "@types/node": "^10.0.3" } }, "sha512-bqX0XTF6fnXSQcEJ2Iuyr75yVakyjIDCqroJQ/aHfSdlM743Cwqoi2nDYMzLGWUcuTWGWy8AAvOKXTfiv6q9RA=="],
-
-    "https-proxy-agent": ["https-proxy-agent@5.0.1", "", { "dependencies": { "agent-base": "6", "debug": "4" } }, "sha512-dFcAjpTQFgoLMzC2VwU+C/CbS7uRL0lWmxDITmqm7C+7F0Odmj6s9l6alZc6AELXhrnggM2CeWSXHGOdX2YtwA=="],
-
-    "inherits": ["inherits@2.0.4", "", {}, "sha512-k/vGaX4/Yla3WzyMCvTQOXYeIHvqOKtnqBduzTHpzpQZzAskKMhZ2K+EnBiSM9zGSoIFeMpXKxa4dYeZIQqewQ=="],
-
-    "json-stringify-safe": ["json-stringify-safe@5.0.1", "", {}, "sha512-ZClg6AaYvamvYEE82d3Iyd3vSSIjQ+odgjaTzRuO3s7toCdFKczob2i0zCh7JE8kWn17yvAWhUVxvqGwUalsRA=="],
-
-    "lodash": ["lodash@4.18.1", "", {}, "sha512-dMInicTPVE8d1e5otfwmmjlxkZoUpiVLwyeTdUsi/Caj/gfzzblBcCE5sRHV/AsjuCmxWrte2TNGSYuCeCq+0Q=="],
-
-    "lodash.snakecase": ["lodash.snakecase@4.1.1", "", {}, "sha512-QZ1d4xoBHYUeuouhEq3lk3Uq7ldgyFXGBhg04+oRLnIz8o9T65Eh+8YdroUwn846zchkA9yDsDl5CVVaV2nqYw=="],
-
-    "magic-bytes.js": ["magic-bytes.js@1.13.0", "", {}, "sha512-afO2mnxW7GDTXMm5/AoN1WuOcdoKhtgXjIvHmobqTD1grNplhGdv3PFOyjCVmrnOZBIT/gD/koDKpYG+0mvHcg=="],
-
-    "matcher": ["matcher@3.0.0", "", { "dependencies": { "escape-string-regexp": "^4.0.0" } }, "sha512-OkeDaAZ/bQCxeFAozM55PKcKU0yJMPGifLwV4Qgjitu+5MoAfSQN4lsLJeXZ1b8w0x+/Emda6MZgXS1jvsapng=="],
-
-    "ms": ["ms@2.1.3", "", {}, "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA=="],
-
-    "object-keys": ["object-keys@1.1.1", "", {}, "sha512-NuAESUOUMrlIXOfHKzD6bpPu3tYt3xvjNdRIQ+FeT0lNb4K8WR70CaDxhuNguS2XG+GjkyMwOzsN5ZktImfhLA=="],
-
-    "onnxruntime-common": ["onnxruntime-common@1.24.3", "", {}, "sha512-GeuPZO6U/LBJXvwdaqHbuUmoXiEdeCjWi/EG7Y1HNnDwJYuk6WUbNXpF6luSUY8yASul3cmUlLGrCCL1ZgVXqA=="],
-
-    "onnxruntime-node": ["onnxruntime-node@1.24.3", "", { "dependencies": { "adm-zip": "^0.5.16", "global-agent": "^3.0.0", "onnxruntime-common": "1.24.3" }, "os": [ "linux", "win32", "darwin", ] }, "sha512-JH7+czbc8ALA819vlTgcV+Q214/+VjGeBHDjX81+ZCD0PCVCIFGFNtT0V4sXG/1JXypKPgScQcB3ij/hk3YnTg=="],
-
-    "opusscript": ["opusscript@0.1.1", "", {}, "sha512-mL0fZZOUnXdZ78woRXp18lApwpp0lF5tozJOD1Wut0dgrA9WuQTgSels/CSmFleaAZrJi/nci5KOVtbuxeWoQA=="],
-
-    "parse-cache-control": ["parse-cache-control@1.0.1", "", {}, "sha512-60zvsJReQPX5/QP0Kzfd/VrpjScIQ7SHBW6bFCYfEP+fp0Eppr1SHhIO5nd1PjZtvclzSzES9D/p5nFJurwfWg=="],
-
-    "prism-media": ["prism-media@1.3.5", "", { "peerDependencies": { "@discordjs/opus": ">=0.8.0 <1.0.0", "ffmpeg-static": "^5.0.2 || ^4.2.7 || ^3.0.0 || ^2.4.0", "node-opus": "^0.3.3", "opusscript": "^0.0.8" }, "optionalPeers": ["@discordjs/opus", "ffmpeg-static", "node-opus", "opusscript"] }, "sha512-IQdl0Q01m4LrkN1EGIE9lphov5Hy7WWlH6ulf5QdGePLlPas9p2mhgddTEHrlaXYjjFToM1/rWuwF37VF4taaA=="],
-
-    "progress": ["progress@2.0.3", "", {}, "sha512-7PiHtLll5LdnKIMw100I+8xJXR5gW2QwWYkT6iJva0bXitZKa/XMrSbdmg3r2Xnaidz9Qumd0VPaMrZlF9V9sA=="],
-
-    "readable-stream": ["readable-stream@3.6.2", "", { "dependencies": { "inherits": "^2.0.3", "string_decoder": "^1.1.1", "util-deprecate": "^1.0.1" } }, "sha512-9u/sniCrY3D5WdsERHzHE4G2YCXqoG5FTHUiCC4SIbr6XcLZBY05ya9EKjYek9O5xOAwjGq+1JdGBAS7Q9ScoA=="],
-
-    "roarr": ["roarr@2.15.4", "", { "dependencies": { "boolean": "^3.0.1", "detect-node": "^2.0.4", "globalthis": "^1.0.1", "json-stringify-safe": "^5.0.1", "semver-compare": "^1.0.0", "sprintf-js": "^1.1.2" } }, "sha512-CHhPh+UNHD2GTXNYhPWLnU8ONHdI+5DI+4EYIAOaiD63rHeYlZvyh8P+in5999TTSFgUYuKUAjzRI4mdh/p+2A=="],
-
-    "safe-buffer": ["safe-buffer@5.2.1", "", {}, "sha512-rp3So07KcdmmKbGvgaNxQSJr7bGVSVk5S9Eq1F+ppbRo70+YeaDxkw5Dd8NPN+GD6bjnYm2VuPuCXmpuYvmCXQ=="],
-
-    "semver": ["semver@7.7.4", "", { "bin": { "semver": "bin/semver.js" } }, "sha512-vFKC2IEtQnVhpT78h1Yp8wzwrf8CM+MzKMHGJZfBtzhZNycRFnXsHk6E5TxIkkMsgNS7mdX3AGB7x2QM2di4lA=="],
-
-    "semver-compare": ["semver-compare@1.0.0", "", {}, "sha512-YM3/ITh2MJ5MtzaM429anh+x2jiLVjqILF4m4oyQB18W7Ggea7BfqdH/wGMK7dDiMghv/6WG7znWMwUDzJiXow=="],
-
-    "serialize-error": ["serialize-error@7.0.1", "", { "dependencies": { "type-fest": "^0.13.1" } }, "sha512-8I8TjW5KMOKsZQTvoxjuSIa7foAwPWGOts+6o7sgjz41/qMD9VQHEDxi6PBvK2l0MXUmqZyNpUK+T2tQaaElvw=="],
-
-    "sprintf-js": ["sprintf-js@1.1.3", "", {}, "sha512-Oo+0REFV59/rz3gfJNKQiBlwfHaSESl1pcGyABQsnnIfWOFt6JNj5gCog2U6MLZ//IGYD+nA8nI+mTShREReaA=="],
-
-    "string_decoder": ["string_decoder@1.3.0", "", { "dependencies": { "safe-buffer": "~5.2.0" } }, "sha512-hkRX8U1WjJFd8LsDJ2yQ/wWWxaopEsABU1XfkM8A+j0+85JAGppt16cr1Whg6KIbb4okU6Mql6BOj+uup/wKeA=="],
-
-    "ts-mixer": ["ts-mixer@6.0.4", "", {}, "sha512-ufKpbmrugz5Aou4wcr5Wc1UUFWOLhq+Fm6qa6P0w0K5Qw2yhaUoiWszhCVuNQyNwrlGiscHOmqYoAox1PtvgjA=="],
-
-    "tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
-
-    "type-fest": ["type-fest@0.13.1", "", {}, "sha512-34R7HTnG0XIJcBSn5XhDd7nNFPRcXYRZrBB2O2jdKqYODldSzBAqzsWoZYYvduky73toYS/ESqxPvkDf/F0XMg=="],
-
-    "typedarray": ["typedarray@0.0.6", "", {}, "sha512-/aCDEGatGvZ2BIk+HmLf4ifCJFwvKFNb9/JeZPMulfgFracn9QFcAf5GO8B/mweUjSoblS5In0cWhqpfs/5PQA=="],
-
     "typescript": ["typescript@6.0.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw=="],
 
-    "undici": ["undici@6.24.1", "", {}, "sha512-sC+b0tB1whOCzbtlx20fx3WgCXwkW627p4EA9uM+/tNNPkSS+eSEld6pAs9nDv7WbY1UUljBMYPtu9BCOrCWKA=="],
-
     "undici-types": ["undici-types@7.19.2", "", {}, "sha512-qYVnV5OEm2AW8cJMCpdV20CDyaN3g0AjDlOGf1OW4iaDEx8MwdtChUp4zu4H0VP3nDRF/8RKWH+IPp9uW0YGZg=="],
 
-    "util-deprecate": ["util-deprecate@1.0.2", "", {}, "sha512-EPD5q1uXyFxJpCrLnCc1nHnq3gOa6DZBocAIiI2TaSCA7VCJ1UJDMagCzIkXNsUYfD1daK//LTEQ8xiIbrHtcw=="],
-
-    "ws": ["ws@8.20.0", "", { "peerDependencies": { "bufferutil": "^4.0.1", "utf-8-validate": ">=5.0.2" }, "optionalPeers": ["bufferutil", "utf-8-validate"] }, "sha512-sAt8BhgNbzCtgGbt2OxmpuryO63ZoDk/sqaB/znQm94T4fCEsy/yV+7CdC1kJhOU9lboAEU7R3kquuycDoibVA=="],
-
-    "zod": ["zod@4.3.6", "", {}, "sha512-rftlrkhHZOcjDwkGlnUtZZkvaPHCsDATp4pGpuOOMDaTdDDXF91wuVDJoWoPsKX/3YPQ5fHuF3STjcYyKr+Qhg=="],
-
-    "@discordjs/rest/@discordjs/collection": ["@discordjs/collection@2.1.1", "", {}, "sha512-LiSusze9Tc7qF03sLCujF5iZp7K+vRNEDBZ86FT9aQAv3vxMLihUvKvpsCWiQ2DJq1tVckopKm1rxomgNUc9hg=="],
-
-    "@discordjs/rest/@sapphire/snowflake": ["@sapphire/snowflake@3.5.5", "", {}, "sha512-xzvBr1Q1c4lCe7i6sRnrofxeO1QTP/LKQ6A6qy0iB4x5yfiSfARMEQEghojzTNALDTcv8En04qYNIco9/K9eZQ=="],
-
-    "@discordjs/ws/@discordjs/collection": ["@discordjs/collection@2.1.1", "", {}, "sha512-LiSusze9Tc7qF03sLCujF5iZp7K+vRNEDBZ86FT9aQAv3vxMLihUvKvpsCWiQ2DJq1tVckopKm1rxomgNUc9hg=="],
-
-    "http-response-object/@types/node": ["@types/node@10.17.60", "", {}, "sha512-F0KIgDJfy2nA3zMLmWGKxcH2ZVEtCZXHHdOQs2gSaQ27+lNeEfGxzkIw90aXswATX7AZ33tahPbzy6KAfUreVw=="],
+    "zod": ["zod@4.4.2", "", {}, "sha512-IynmDyxsEsb9RKzO3J9+4SxXnl2FTFSzNBaKKaMV6tsSk0rw9gYw9gs+JFCq/qk2LCZ78KDwyj+Z289TijSkUw=="],
   }
 }

docker.cmd

@@ -0,0 +1,13 @@
+@echo off
+setlocal
+
+set "DOCKER_EXE=C:\Program Files\Docker\Docker\resources\bin\docker.exe"
+
+if exist "%DOCKER_EXE%" (
+  "%DOCKER_EXE%" %*
+  exit /b %ERRORLEVEL%
+)
+
+echo Docker executable not found: "%DOCKER_EXE%"
+echo Install Docker Desktop or update this shim path.
+exit /b 1

docker/melotts/Dockerfile

@@ -0,0 +1,22 @@
+FROM python:3.11-slim
+
+ENV PYTHONUNBUFFERED=1
+
+WORKDIR /opt/realtime-voice-bot
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+  build-essential \
+  git \
+  libsndfile1 \
+  && rm -rf /var/lib/apt/lists/*
+
+RUN git clone https://github.com/myshell-ai/MeloTTS.git /opt/MeloTTS
+RUN pip install --no-cache-dir -e /opt/MeloTTS
+RUN pip install --no-cache-dir python-mecab-ko python-mecab-ko-dic
+RUN python -m unidic download
+RUN python /opt/MeloTTS/melo/init_downloads.py
+
+COPY melo_tts_cli.py /opt/realtime-voice-bot/melo_tts_cli.py
+COPY melo_tts_worker.py /opt/realtime-voice-bot/melo_tts_worker.py
+
+ENTRYPOINT ["python", "/opt/realtime-voice-bot/melo_tts_cli.py"]

docker/melotts/melo_tts_cli.py

@@ -0,0 +1,42 @@
+import argparse
+from pathlib import Path
+
+from melo.api import TTS
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--text", required=True)
+    parser.add_argument("--output", required=True)
+    parser.add_argument("--language", default="KR")
+    parser.add_argument("--speaker", default="KR")
+    parser.add_argument("--speed", type=float, default=1.0)
+    parser.add_argument("--sdp-ratio", type=float, default=0.2)
+    parser.add_argument("--noise-scale", type=float, default=0.6)
+    parser.add_argument("--noise-scale-w", type=float, default=0.8)
+    parser.add_argument("--device", default="cpu")
+    args = parser.parse_args()
+
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    model = TTS(language=args.language, device=args.device)
+    speaker_ids = model.hps.data.spk2id
+
+    if args.speaker not in speaker_ids:
+        supported = ", ".join(sorted(speaker_ids.keys()))
+        raise SystemExit(f"지원하지 않는 speaker 입니다: {args.speaker}. 사용 가능: {supported}")
+
+    model.tts_to_file(
+        args.text,
+        speaker_ids[args.speaker],
+        str(output_path),
+        speed=args.speed,
+        sdp_ratio=args.sdp_ratio,
+        noise_scale=args.noise_scale,
+        noise_scale_w=args.noise_scale_w,
+    )
+
+
+if __name__ == "__main__":
+    main()

docker/melotts/melo_tts_worker.py

@@ -0,0 +1,122 @@
+import contextlib
+import io
+import json
+import os
+import sys
+from pathlib import Path
+
+from melo.api import TTS
+
+
+LANGUAGE = os.getenv("TTS_LANGUAGE", "KR")
+SPEAKER = os.getenv("TTS_SPEAKER", "KR")
+DEVICE = os.getenv("TTS_DEVICE", "cpu")
+SPEED = float(os.getenv("TTS_SPEED", "1.18"))
+SDP_RATIO = float(os.getenv("TTS_SDP_RATIO", "0.22"))
+NOISE_SCALE = float(os.getenv("TTS_NOISE_SCALE", "0.55"))
+NOISE_SCALE_W = float(os.getenv("TTS_NOISE_SCALE_W", "0.75"))
+
+_MODEL = None
+_SPEAKER_ID = None
+
+
+def silence_stdout(func):
+    def wrapper(*args, **kwargs):
+        with contextlib.redirect_stdout(io.StringIO()):
+            return func(*args, **kwargs)
+
+    return wrapper
+
+
+@silence_stdout
+def load_model():
+    global _MODEL
+    global _SPEAKER_ID
+
+    if _MODEL is not None and _SPEAKER_ID is not None:
+        return _MODEL, _SPEAKER_ID
+
+    model = TTS(language=LANGUAGE, device=DEVICE)
+    speaker_ids = model.hps.data.spk2id
+
+    if SPEAKER not in speaker_ids:
+        supported = ", ".join(sorted(speaker_ids.keys()))
+        raise RuntimeError(f"지원하지 않는 speaker 입니다: {SPEAKER}. 사용 가능: {supported}")
+
+    _MODEL = model
+    _SPEAKER_ID = speaker_ids[SPEAKER]
+    return _MODEL, _SPEAKER_ID
+
+
+def handle_ping():
+    model, speaker_id = load_model()
+    return {
+        "language": LANGUAGE,
+        "speaker": SPEAKER,
+        "speaker_id": speaker_id,
+        "device": DEVICE,
+        "speed": SPEED,
+        "sdp_ratio": SDP_RATIO,
+        "noise_scale": NOISE_SCALE,
+        "noise_scale_w": NOISE_SCALE_W,
+        "speaker_count": len(model.hps.data.spk2id),
+    }
+
+
+@silence_stdout
+def handle_synthesize(params):
+    text = str(params["text"]).strip()
+    output_path = Path(str(params["output_path"]))
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    model, speaker_id = load_model()
+    model.tts_to_file(
+        text,
+        speaker_id,
+        str(output_path),
+        speed=SPEED,
+        sdp_ratio=SDP_RATIO,
+        noise_scale=NOISE_SCALE,
+        noise_scale_w=NOISE_SCALE_W,
+    )
+
+    return {
+        "output_path": str(output_path),
+        "text_length": len(text),
+    }
+
+
+def main():
+    for raw_line in sys.stdin:
+        line = raw_line.strip()
+        if not line:
+            continue
+
+        try:
+            payload = json.loads(line)
+            request_id = str(payload["id"])
+            method = payload["method"]
+            params = payload.get("params", {})
+
+            if method == "ping":
+                result = handle_ping()
+            elif method == "synthesize":
+                result = handle_synthesize(params)
+            else:
+                raise RuntimeError(f"알 수 없는 method 입니다: {method}")
+
+            sys.stdout.write(json.dumps({"id": request_id, "result": result}, ensure_ascii=False) + "\n")
+            sys.stdout.flush()
+        except Exception as error:
+            request_id = "unknown"
+            try:
+                request_id = str(payload.get("id", "unknown"))
+            except Exception:
+                pass
+
+            sys.stdout.write(json.dumps({"id": request_id, "error": str(error)}, ensure_ascii=False) + "\n")
+            sys.stdout.flush()
+
+
+if __name__ == "__main__":
+    main()

package.json

@@ -4,13 +4,18 @@
   "private": true,
   "type": "module",
   "scripts": {
-    "dev": "bun --watch src/index.ts",
-    "start": "bun src/index.ts discord",
-    "start:discord": "bun src/index.ts discord",
-    "start:local": "bun src/index.ts local",
-    "setup:local-ai": "bun src/setup-local-ai.ts",
-    "devices": "bun src/index.ts local-devices",
-    "audio:devices": "bun src/index.ts local-devices",
+    "setup": "bun src/setup.ts",
+    "setup:all": "bun run setup",
+    "setup:stt": "bun src/setup-python.ts",
+    "setup:llm": "bun src/setup-llm.ts",
+    "setup:tts": "bun src/setup-tts.ts",
+    "setup:python": "bun run setup:stt",
+    "test:stt": "bun src/index.ts test-stt",
+    "test:sttllm": "bun src/index.ts test-sttllm",
+    "test:all": "bun src/index.ts test-all",
+    "test:llm": "bun src/index.ts test-llm",
+    "test:tts": "bun src/index.ts test-tts",
+    "devices": "bun src/index.ts devices",
     "check": "tsc --noEmit",
     "build": "tsc -p tsconfig.json"
   },
@@ -19,21 +24,11 @@
     "node": ">=22.12.0"
   },
   "dependencies": {
-    "@discordjs/voice": "^0.19.2",
-    "avr-vad": "^1.0.10",
-    "discord.js": "^14.26.3",
     "dotenv": "^17.4.2",
-    "ffmpeg-static": "^5.3.0",
-    "opusscript": "^0.1.1",
-    "prism-media": "^1.3.5",
     "zod": "^4.3.6"
   },
   "devDependencies": {
     "@types/node": "^25.6.0",
     "typescript": "^6.0.3"
-  },
-  "trustedDependencies": [
-    "onnxruntime-node",
-    "ffmpeg-static"
-  ]
+  }
 }

prompts/assistant.md

@@ -0,0 +1,14 @@
+너는 한국어로 짧고 자연스럽게 답하는 로컬 음성 비서다.
+
+규칙:
+- 반드시 한국어로만 답한다.
+- 한자, 중국어, 일본어, 아랍어, 키릴 문자, 기타 외국 문자, 이모지 사용 금지.
+- 영어 단어는 꼭 필요한 기술명 외에는 피하고 자연스러운 한국어 표현으로 바꾼다.
+- 답변은 TTS가 읽기 쉽도록 짧고 단순한 문장으로 만든다.
+- 기본적으로 1~3문장으로 답한다.
+- 불필요한 장식, 불릿, 번호 목록, 괄호 남용, 과한 감탄 표현은 피한다.
+- 사용자의 말에 바로 답하고, 군더더기 없이 핵심만 말한다.
+- 정확한 시간, 설정 확인, 계산이 필요하면 도구를 우선 사용한다.
+- 최신 정보, 오늘/최근 정보, 뉴스, 검색 요청, 사실 확인, 외부 웹페이지 내용이 필요한 경우에만 `web_search` 와 `fetch_url` 을 사용한다.
+- 내부 지식만으로 충분한 일반 대화에는 웹 도구를 쓰지 않는다.
+- 도구가 필요한 작업이 시작되면 결과 전에 짧은 진행 메시지를 출력할 수 있다.

prompts/reply-gate.md

@@ -0,0 +1,10 @@
+다음 텍스트에 로컬 비서가 실제로 대답해야 하는지 판정한다.
+
+판정 기준:
+- 의미 없는 감탄사, 중얼거림, 문맥 없는 짧은 파편, 노래 가사 조각, 잡음성 문장은 `should_reply=false`
+- 질문, 요청, 확인, 명령, 대화 시도는 `should_reply=true`
+- 최신 정보나 사실 확인, 검색이 필요해 보이면 `likely_needs_lookup=true`
+- reason 은 아주 짧게 쓴다
+
+반드시 JSON만 출력:
+{"should_reply":true,"likely_needs_lookup":false,"reason":"짧게"}

prompts/rewrite-korean.md

@@ -0,0 +1,8 @@
+다음 답변을 의미를 유지한 채 자연스러운 한국어로만 다시 쓴다.
+
+규칙:
+- 한글, 숫자, 기본 문장부호 외 다른 문자 사용 금지
+- 이모지 사용 금지
+- 짧고 읽기 쉬운 문장으로 만든다
+- TTS가 읽기 쉽도록 불필요한 기호와 장식을 줄인다
+- 설명하지 말고 최종 답변 문장만 출력한다

python/local_stt_worker.py

@@ -1,145 +0,0 @@
-import base64
-import json
-import os
-import sys
-import tempfile
-import traceback
-import wave
-
-
-os.environ.setdefault("PYTHONIOENCODING", "utf-8")
-
-
-def log(message: str) -> None:
-    print(message, file=sys.stderr, flush=True)
-
-
-def write_response(request_id: int, ok: bool, result=None, error: str | None = None) -> None:
-    payload = {
-        "id": request_id,
-        "ok": ok,
-    }
-    if ok:
-        payload["result"] = result
-    else:
-        payload["error"] = error or "unknown error"
-
-    sys.stdout.write(json.dumps(payload, ensure_ascii=False) + "\n")
-    sys.stdout.flush()
-
-
-def resolve_device() -> str:
-    raw = os.environ.get("LOCAL_STT_DEVICE", "auto").strip().lower()
-    if raw and raw != "auto":
-        return raw
-
-    try:
-        import ctranslate2
-
-        if ctranslate2.get_cuda_device_count() > 0:
-            return "cuda"
-    except Exception:
-        pass
-
-    return "cpu"
-
-
-def resolve_compute_type(device: str) -> str:
-    raw = os.environ.get("LOCAL_STT_COMPUTE_TYPE", "auto").strip().lower()
-    if raw and raw != "auto":
-        return raw
-    if device == "cuda":
-        return "int8_float16"
-    return "int8"
-
-
-class SttWorker:
-    def __init__(self) -> None:
-        from faster_whisper import WhisperModel
-
-        self.model_name = os.environ.get("LOCAL_STT_MODEL", "tiny").strip() or "tiny"
-        self.device = resolve_device()
-        self.compute_type = resolve_compute_type(self.device)
-        self.beam_size = int(os.environ.get("LOCAL_STT_BEAM_SIZE", "1"))
-        self.model = WhisperModel(
-            self.model_name,
-            device=self.device,
-            compute_type=self.compute_type,
-        )
-        log(
-            f"local-stt ready model={self.model_name} device={self.device} compute={self.compute_type} beam={self.beam_size}"
-        )
-
-    def transcribe(self, audio_base64: str, language: str | None) -> str:
-        pcm_bytes = base64.b64decode(audio_base64)
-        temp_path = ""
-
-        try:
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as handle:
-                temp_path = handle.name
-
-            with wave.open(temp_path, "wb") as wav_file:
-                wav_file.setnchannels(1)
-                wav_file.setsampwidth(2)
-                wav_file.setframerate(16000)
-                wav_file.writeframes(pcm_bytes)
-
-            segments, _info = self.model.transcribe(
-                temp_path,
-                language=language,
-                beam_size=self.beam_size,
-                best_of=1,
-                condition_on_previous_text=False,
-                vad_filter=False,
-                without_timestamps=True,
-                temperature=0.0,
-            )
-            return " ".join(segment.text.strip() for segment in segments if segment.text.strip()).strip()
-        finally:
-            if temp_path:
-                try:
-                    os.unlink(temp_path)
-                except OSError:
-                    pass
-
-
-def main() -> int:
-    try:
-        worker = SttWorker()
-    except Exception as exc:
-        log("failed to initialize local STT worker")
-        log("run `bun run setup:local-ai` first if dependencies are missing")
-        log("".join(traceback.format_exception(exc)))
-        return 1
-
-    for line in sys.stdin:
-        line = line.strip()
-        if not line:
-            continue
-
-        try:
-            request = json.loads(line)
-            request_id = int(request["id"])
-            method = request["method"]
-            params = request.get("params", {})
-
-            if method == "ping":
-                write_response(request_id, True, {"ready": True})
-                continue
-            if method != "transcribe":
-                raise ValueError(f"unsupported method: {method}")
-
-            text = worker.transcribe(
-                audio_base64=str(params.get("audio_base64", "")),
-                language=str(params.get("language") or "").strip() or None,
-            )
-            write_response(request_id, True, {"text": text})
-        except Exception as exc:
-            error_text = "".join(traceback.format_exception_only(type(exc), exc)).strip()
-            write_response(request_id, False, error=error_text)
-
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

python/local_tts_worker.py

@@ -1,125 +0,0 @@
-import base64
-import json
-import os
-import sys
-import tempfile
-import traceback
-
-
-os.environ.setdefault("PYTHONIOENCODING", "utf-8")
-
-
-def log(message: str) -> None:
-    print(message, file=sys.stderr, flush=True)
-
-
-def write_response(request_id: int, ok: bool, result=None, error: str | None = None) -> None:
-    payload = {
-        "id": request_id,
-        "ok": ok,
-    }
-    if ok:
-        payload["result"] = result
-    else:
-        payload["error"] = error or "unknown error"
-
-    sys.stdout.write(json.dumps(payload, ensure_ascii=False) + "\n")
-    sys.stdout.flush()
-
-
-class TtsWorker:
-    def __init__(self) -> None:
-        from melo.api import TTS
-
-        self.language = os.environ.get("LOCAL_TTS_LANGUAGE", "KR").strip() or "KR"
-        self.speaker_key = os.environ.get("LOCAL_TTS_SPEAKER", "KR").strip() or "KR"
-        self.device = os.environ.get("LOCAL_TTS_DEVICE", "auto").strip() or "auto"
-        self.speed = float(os.environ.get("LOCAL_TTS_SPEED", "1.12"))
-
-        self.model = TTS(language=self.language, device=self.device)
-        speaker_ids = self.model.hps.data.spk2id
-        self.speaker_id = speaker_ids.get(self.speaker_key)
-
-        if self.speaker_id is None:
-            normalized = self.speaker_key.upper()
-            self.speaker_id = speaker_ids.get(normalized)
-
-        if self.speaker_id is None:
-            self.speaker_id = next(iter(speaker_ids.values()))
-
-        log(
-            f"local-tts ready language={self.language} speaker={self.speaker_key} device={self.device} speed={self.speed}"
-        )
-
-    def synthesize(self, text: str) -> bytes:
-        temp_path = ""
-
-        try:
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as handle:
-                temp_path = handle.name
-
-            self.model.tts_to_file(
-                text,
-                self.speaker_id,
-                temp_path,
-                speed=self.speed,
-                quiet=True,
-            )
-
-            with open(temp_path, "rb") as handle:
-                return handle.read()
-        finally:
-            if temp_path:
-                try:
-                    os.unlink(temp_path)
-                except OSError:
-                    pass
-
-
-def main() -> int:
-    try:
-        worker = TtsWorker()
-    except Exception as exc:
-        log("failed to initialize local TTS worker")
-        log("run `bun run setup:local-ai` first if dependencies are missing")
-        log("".join(traceback.format_exception(exc)))
-        return 1
-
-    for line in sys.stdin:
-        line = line.strip()
-        if not line:
-            continue
-
-        try:
-            request = json.loads(line)
-            request_id = int(request["id"])
-            method = request["method"]
-            params = request.get("params", {})
-
-            if method == "ping":
-                write_response(request_id, True, {"ready": True})
-                continue
-            if method != "synthesize":
-                raise ValueError(f"unsupported method: {method}")
-
-            text = str(params.get("text", "")).strip()
-            if not text:
-                raise ValueError("text is empty")
-
-            audio = worker.synthesize(text)
-            write_response(
-                request_id,
-                True,
-                {
-                    "wav_base64": base64.b64encode(audio).decode("ascii"),
-                },
-            )
-        except Exception as exc:
-            error_text = "".join(traceback.format_exception_only(type(exc), exc)).strip()
-            write_response(request_id, False, error=error_text)
-
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

python/loopback_stt_worker.py

@@ -0,0 +1,186 @@
+import base64
+import json
+import os
+import sys
+import site
+import traceback
+from typing import Any
+from pathlib import Path
+
+import numpy as np
+from faster_whisper import WhisperModel
+
+
+def configure_windows_dll_search_paths() -> list[str]:
+    if sys.platform != "win32":
+        return []
+
+    candidates: list[Path] = []
+    executable_dir = Path(sys.executable).resolve().parent
+    venv_root = executable_dir.parent
+    candidates.extend(
+        [
+            venv_root / "Lib" / "site-packages" / "nvidia" / "cublas" / "bin",
+            venv_root / "Lib" / "site-packages" / "nvidia" / "cudnn" / "bin",
+        ]
+    )
+
+    for package_path in site.getsitepackages():
+        base = Path(package_path)
+        candidates.extend(
+            [
+                base / "nvidia" / "cublas" / "bin",
+                base / "nvidia" / "cudnn" / "bin",
+            ]
+        )
+
+    added: list[str] = []
+    seen: set[str] = set()
+    for candidate in candidates:
+        normalized = str(candidate.resolve())
+        if normalized in seen:
+            continue
+        seen.add(normalized)
+        if not candidate.exists():
+            continue
+
+        os.add_dll_directory(normalized)
+        if normalized not in os.environ.get("PATH", ""):
+            os.environ["PATH"] = normalized + os.pathsep + os.environ.get("PATH", "")
+        added.append(normalized)
+
+    return added
+
+
+os.environ.setdefault("HF_HUB_DISABLE_SYMLINKS_WARNING", "1")
+CONFIGURED_DLL_PATHS = configure_windows_dll_search_paths()
+if CONFIGURED_DLL_PATHS:
+    print(
+        f"configured CUDA DLL search paths: {', '.join(CONFIGURED_DLL_PATHS)}",
+        file=sys.stderr,
+        flush=True,
+    )
+
+
+def resolve_model() -> WhisperModel:
+    model_name = os.environ.get("WHISPER_MODEL", "large-v3-turbo")
+    requested_device = os.environ.get("WHISPER_DEVICE", "auto")
+    requested_compute = os.environ.get("WHISPER_COMPUTE_TYPE", "auto")
+
+    attempts: list[tuple[str, str]] = []
+    if requested_device == "auto":
+        if requested_compute == "auto":
+            attempts.extend(
+                [
+                    ("cuda", "float16"),
+                    ("cuda", "int8_float16"),
+                    ("cpu", "int8"),
+                    ("cpu", "float32"),
+                ]
+            )
+        else:
+            attempts.extend(
+                [
+                    ("cuda", requested_compute),
+                    ("cpu", requested_compute),
+                ]
+            )
+    else:
+        if requested_compute == "auto":
+            compute = "float16" if requested_device == "cuda" else "int8"
+        else:
+            compute = requested_compute
+        attempts.append((requested_device, compute))
+
+    last_error: Exception | None = None
+    for device, compute_type in attempts:
+        try:
+            model = WhisperModel(model_name, device=device, compute_type=compute_type)
+            setattr(model, "_resolved_device", device)
+            setattr(model, "_resolved_compute_type", compute_type)
+            return model
+        except Exception as error:  # noqa: BLE001
+            last_error = error
+
+    assert last_error is not None
+    raise last_error
+
+
+MODEL = resolve_model()
+LANGUAGE = os.environ.get("WHISPER_LANGUAGE", "ko")
+BEAM_SIZE = int(os.environ.get("WHISPER_BEAM_SIZE", "1"))
+
+
+def write(payload: dict[str, Any]) -> None:
+    sys.stdout.write(json.dumps(payload, ensure_ascii=False) + "\n")
+    sys.stdout.flush()
+
+
+def transcribe_pcm16_base64(pcm16_base64: str) -> str:
+    audio_bytes = base64.b64decode(pcm16_base64)
+    audio = np.frombuffer(audio_bytes, dtype=np.int16).astype(np.float32) / 32768.0
+
+    segments, _info = MODEL.transcribe(
+        audio,
+        language=LANGUAGE,
+        task="transcribe",
+        beam_size=BEAM_SIZE,
+        condition_on_previous_text=False,
+        vad_filter=False,
+        without_timestamps=True,
+        word_timestamps=False,
+        temperature=0.0,
+    )
+
+    text_parts: list[str] = []
+    for segment in segments:
+        if segment.text:
+            text_parts.append(segment.text.strip())
+    return " ".join(part for part in text_parts if part).strip()
+
+
+for raw_line in sys.stdin:
+    line = raw_line.strip()
+    if not line:
+        continue
+
+    request = json.loads(line)
+    request_id = request["id"]
+    method = request["method"]
+    params = request.get("params", {})
+
+    try:
+        if method == "ping":
+            write(
+                {
+                    "id": request_id,
+                    "result": {
+                        "model": os.environ.get("WHISPER_MODEL", "large-v3-turbo"),
+                        "device": getattr(MODEL, "_resolved_device", "unknown"),
+                        "compute_type": getattr(MODEL, "_resolved_compute_type", "unknown"),
+                    },
+                }
+            )
+            continue
+
+        if method == "transcribe":
+            text = transcribe_pcm16_base64(params["pcm16_base64"])
+            write(
+                {
+                    "id": request_id,
+                    "result": {
+                        "text": text,
+                    },
+                }
+            )
+            continue
+
+        raise RuntimeError(f"unknown method: {method}")
+    except Exception as error:  # noqa: BLE001
+        traceback.print_exc(file=sys.stderr)
+        write(
+            {
+                "id": request_id,
+                "error": f"{type(error).__name__}: {error}",
+            }
+        )

python/requirements-windows.txt

@@ -0,0 +1,3 @@
+faster-whisper==1.2.1
+nvidia-cublas-cu12
+nvidia-cudnn-cu12>=9,<10

python/requirements.txt

@@ -1,2 +1 @@
 faster-whisper==1.2.1
-git+https://github.com/myshell-ai/MeloTTS.git@v0.1.2

src/audio/capture.ts

@@ -0,0 +1,138 @@
+import { spawn, type ChildProcessByStdio } from "node:child_process";
+import process from "node:process";
+import type { Readable } from "node:stream";
+
+import type { AppConfig } from "../config.js";
+import type { Logger } from "../logger.js";
+
+export function printAudioDevices(): Promise<void> {
+  if (process.platform === "win32") {
+    return new Promise<void>((resolve, reject) => {
+      const child = spawn("ffmpeg", ["-hide_banner", "-list_devices", "true", "-f", "dshow", "-i", "dummy"], {
+        stdio: ["ignore", "ignore", "inherit"],
+      });
+      child.on("exit", (code) => {
+        if (code === 0 || code === 1) {
+          resolve();
+          return;
+        }
+        reject(new Error(`ffmpeg exited with code ${code ?? "null"}`));
+      });
+      child.on("error", reject);
+    });
+  }
+
+  return new Promise<void>((resolve, reject) => {
+    const pactl = spawn("pactl", ["list", "sources", "short"], {
+      stdio: ["ignore", "inherit", "inherit"],
+    });
+
+    pactl.on("exit", (code) => {
+      if (code === 0) {
+        resolve();
+        return;
+      }
+
+      const wpctl = spawn("wpctl", ["status", "-n"], {
+        stdio: ["ignore", "inherit", "inherit"],
+      });
+      wpctl.on("exit", (wpctlCode) => {
+        if (wpctlCode === 0) {
+          resolve();
+          return;
+        }
+        reject(new Error(`pactl exited with code ${code ?? "null"} and wpctl exited with code ${wpctlCode ?? "null"}`));
+      });
+      wpctl.on("error", reject);
+    });
+
+    pactl.on("error", () => {
+      const wpctl = spawn("wpctl", ["status", "-n"], {
+        stdio: ["ignore", "inherit", "inherit"],
+      });
+      wpctl.on("exit", (code) => {
+        if (code === 0) {
+          resolve();
+          return;
+        }
+        reject(new Error(`pactl, wpctl 둘 다 실행할 수 없습니다. code=${code ?? "null"}`));
+      });
+      wpctl.on("error", reject);
+    });
+  });
+}
+
+export function spawnLoopbackCapture(
+  config: AppConfig,
+  logger: Logger,
+): ChildProcessByStdio<null, Readable, Readable> {
+  if (!config.AUDIO_SOURCE) {
+    throw new Error("AUDIO_SOURCE 설정이 필요합니다. 먼저 `bun run devices` 로 장치 이름을 확인하세요.");
+  }
+
+  if (process.platform === "win32") {
+    const args = [
+      "-hide_banner",
+      "-loglevel",
+      "warning",
+      "-fflags",
+      "nobuffer",
+      "-flags",
+      "low_delay",
+      "-f",
+      "dshow",
+      "-i",
+      `audio=${config.AUDIO_SOURCE}`,
+      "-ac",
+      "1",
+      "-ar",
+      "16000",
+      "-f",
+      "s16le",
+      "pipe:1",
+    ];
+
+    logger.info("Starting Windows loopback capture", {
+      source: config.AUDIO_SOURCE,
+      backend: "ffmpeg-dshow",
+    });
+
+    return spawn("ffmpeg", args, {
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+  }
+
+  if (process.platform === "linux") {
+    const args = [
+      "-hide_banner",
+      "-loglevel",
+      "warning",
+      "-fflags",
+      "nobuffer",
+      "-flags",
+      "low_delay",
+      "-f",
+      "pulse",
+      "-i",
+      config.AUDIO_SOURCE,
+      "-ac",
+      "1",
+      "-ar",
+      "16000",
+      "-f",
+      "s16le",
+      "pipe:1",
+    ];
+
+    logger.info("Starting Linux loopback capture", {
+      source: config.AUDIO_SOURCE,
+      backend: "ffmpeg-pulse",
+    });
+
+    return spawn("ffmpeg", args, {
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+  }
+
+  throw new Error(`지원하지 않는 플랫폼입니다: ${process.platform}`);
+}

src/audio/ffmpeg-path.ts

@@ -1,59 +0,0 @@
-import { existsSync } from "node:fs";
-import { spawnSync } from "node:child_process";
-import process from "node:process";
-
-import ffmpegStatic from "ffmpeg-static";
-
-function firstExisting(paths: Array<string | null | undefined>): string | null {
-  for (const candidate of paths) {
-    if (candidate && existsSync(candidate)) {
-      return candidate;
-    }
-  }
-  return null;
-}
-
-function findOnPath(): string | null {
-  const locator = process.platform === "win32" ? "where" : "which";
-  const binaryName = process.platform === "win32" ? "ffmpeg.exe" : "ffmpeg";
-  const result = spawnSync(locator, [binaryName], {
-    encoding: "utf8",
-  });
-
-  if (result.status !== 0) {
-    return null;
-  }
-
-  const match = result.stdout
-    .split(/\r?\n/)
-    .map((line) => line.trim())
-    .find((line) => line.length > 0 && existsSync(line));
-
-  return match ?? null;
-}
-
-export function resolveFfmpegPath(): string | null {
-  const staticPath = ffmpegStatic as unknown as string | null;
-  return firstExisting([
-    process.env.FFMPEG_PATH,
-    process.env.FFMPEG_BIN,
-    staticPath,
-    findOnPath(),
-  ]);
-}
-
-export function requireFfmpegPath(): string {
-  const resolved = resolveFfmpegPath();
-  if (resolved) {
-    return resolved;
-  }
-
-  throw new Error(
-    [
-      "ffmpeg를 찾지 못했습니다.",
-      "1. `bun install` 재실행",
-      "2. 안 되면 `bun pm trust ffmpeg-static` 후 다시 `bun install`",
-      "3. 또는 시스템 ffmpeg를 설치해서 PATH에 추가",
-    ].join("\n"),
-  );
-}

src/audio/guild-voice-session.ts

@@ -1,456 +0,0 @@
-import { EventEmitter } from "node:events";
-
-import prism from "prism-media";
-import { RealTimeVAD } from "avr-vad";
-import {
-  AudioPlayerStatus,
-  EndBehaviorType,
-  NoSubscriberBehavior,
-  VoiceConnectionStatus,
-  createAudioPlayer,
-  createAudioResource,
-  entersState,
-  joinVoiceChannel,
-  StreamType,
-  type AudioPlayer,
-  type AudioReceiveStream,
-  type VoiceConnection,
-} from "@discordjs/voice";
-import type { Client, Guild, VoiceBasedChannel } from "discord.js";
-
-import type { AppConfig } from "../config.js";
-import { Logger } from "../logger.js";
-import { float32ToPcm16Buffer, int16ArrayToFloat32, Stereo48kToMono16kDownsampler, takeFrame } from "./pcm.js";
-import { ConversationMemory, type UserUtterance } from "../services/conversation.js";
-import type { LlmService } from "../services/llm.js";
-import type { SttService } from "../services/stt.js";
-import type { PreparedSpeechAudio, TtsService } from "../services/tts.js";
-
-interface GuildVoiceSessionOptions {
-  client: Client;
-  config: AppConfig;
-  logger: Logger;
-  guild: Guild;
-  voiceChannel: VoiceBasedChannel;
-  textChannelId?: string;
-  stt: SttService;
-  tts: TtsService;
-  llm: LlmService;
-}
-
-interface SpeechJob {
-  text: string;
-  source: "assistant" | "manual";
-}
-
-class UserAudioSession {
-  private readonly downsampler = new Stereo48kToMono16kDownsampler();
-  private readonly pendingSamples: number[] = [];
-  private readonly vad: RealTimeVAD;
-  private processing = Promise.resolve();
-
-  private constructor(
-    private readonly logger: Logger,
-    private readonly speakerId: string,
-    private readonly speakerName: string,
-    private readonly receiveStream: AudioReceiveStream,
-    private readonly decoder: NodeJS.ReadWriteStream & { destroy: () => void },
-    vad: RealTimeVAD,
-    private readonly onSpeechEnd: (utterance: UserUtterance, audio: Float32Array) => void,
-  ) {
-    this.vad = vad;
-  }
-
-  static async create(options: {
-      logger: Logger;
-      speakerId: string;
-      speakerName: string;
-      receiveStream: AudioReceiveStream;
-      decoder: NodeJS.ReadWriteStream & { destroy: () => void };
-      onSpeechStart: () => void;
-      onSpeechEnd: (utterance: UserUtterance, audio: Float32Array) => void;
-    }): Promise<UserAudioSession> {
-    const vadInstance = await RealTimeVAD.new({
-      model: "v5",
-      sampleRate: 16000,
-      frameSamples: 1536,
-      positiveSpeechThreshold: 0.55,
-      negativeSpeechThreshold: 0.35,
-      redemptionFrames: 8,
-      preSpeechPadFrames: 2,
-      minSpeechFrames: 3,
-      onFrameProcessed: () => undefined,
-      onVADMisfire: () => undefined,
-      onSpeechStart: () => {
-        options.onSpeechStart();
-      },
-      onSpeechRealStart: () => undefined,
-      onSpeechEnd: (audio: Float32Array) => {
-        options.onSpeechEnd(
-          {
-            speakerId: options.speakerId,
-            speakerName: options.speakerName,
-            text: "",
-          },
-          audio,
-        );
-      },
-    });
-
-    const session = new UserAudioSession(
-      options.logger,
-      options.speakerId,
-      options.speakerName,
-      options.receiveStream,
-      options.decoder,
-      vadInstance,
-      options.onSpeechEnd,
-    );
-
-    session.decoder.on("data", (chunk: Buffer) => {
-      session.pushPcmChunk(chunk);
-    });
-
-    session.decoder.on("error", (error) => {
-      options.logger.warn("PCM decoder error", options.speakerId, error);
-    });
-
-    session.receiveStream.on("error", (error) => {
-      options.logger.warn("Audio receive stream error", options.speakerId, error);
-    });
-
-    return session;
-  }
-
-  private pushPcmChunk(chunk: Buffer): void {
-    const mono16k = this.downsampler.pushStereo48kChunk(chunk);
-    if (mono16k.length === 0) {
-      return;
-    }
-
-    for (const sample of mono16k) {
-      this.pendingSamples.push(sample);
-    }
-
-    while (true) {
-      const frame = takeFrame(this.pendingSamples, 1536);
-      if (!frame) {
-        return;
-      }
-
-      const floatFrame = int16ArrayToFloat32(frame);
-      this.processing = this.processing
-        .then(() => this.vad.processAudio(floatFrame))
-        .catch((error) => {
-          this.logger.warn("VAD frame processing failed", this.speakerId, this.speakerName, error);
-        });
-    }
-  }
-
-  destroy(): void {
-    this.receiveStream.destroy();
-    this.decoder.destroy();
-    void this.vad.destroy().catch((error) => {
-      this.logger.warn("VAD destroy failed", this.speakerId, this.speakerName, error);
-    });
-  }
-}
-
-export class GuildVoiceSession extends EventEmitter {
-  readonly guildId: string;
-  readonly voiceChannelId: string;
-
-  private readonly connection: VoiceConnection;
-  private readonly player: AudioPlayer;
-  private readonly memory: ConversationMemory;
-  private readonly trackedUsers = new Map<string, UserAudioSession>();
-  private readonly pendingUsers = new Map<string, Promise<void>>();
-  private readonly queue: SpeechJob[] = [];
-
-  private draining = false;
-  private currentAbortController: AbortController | null = null;
-  private currentPlayback: PreparedSpeechAudio | null = null;
-  private textChannelId?: string;
-
-  private constructor(private readonly options: GuildVoiceSessionOptions) {
-    super();
-
-    this.guildId = options.guild.id;
-    this.voiceChannelId = options.voiceChannel.id;
-    this.textChannelId = options.textChannelId;
-    this.memory = new ConversationMemory(options.config.MAX_CONVERSATION_TURNS);
-    this.player = createAudioPlayer({
-      behaviors: {
-        noSubscriber: NoSubscriberBehavior.Pause,
-      },
-    });
-    this.connection = joinVoiceChannel({
-      guildId: options.guild.id,
-      channelId: options.voiceChannel.id,
-      adapterCreator: options.guild.voiceAdapterCreator,
-      selfDeaf: false,
-      selfMute: false,
-    });
-  }
-
-  static async create(options: GuildVoiceSessionOptions): Promise<GuildVoiceSession> {
-    const session = new GuildVoiceSession(options);
-    await session.initialize();
-    return session;
-  }
-
-  private async initialize(): Promise<void> {
-    this.player.on("error", (error) => {
-      this.options.logger.warn("Audio player error", this.guildId, error);
-    });
-
-    this.connection.on("stateChange", (_oldState, newState) => {
-      if (newState.status === VoiceConnectionStatus.Destroyed) {
-        this.options.logger.info("Voice connection destroyed", this.guildId);
-      }
-    });
-
-    this.connection.subscribe(this.player);
-    await entersState(this.connection, VoiceConnectionStatus.Ready, 30_000);
-
-    this.connection.receiver.speaking.on("start", (userId: string) => {
-      if (userId === this.options.client.user?.id) {
-        return;
-      }
-
-      void this.ensureTrackedUser(userId);
-    });
-  }
-
-  setTextChannel(textChannelId?: string): void {
-    this.textChannelId = textChannelId;
-  }
-
-  clearConversation(): void {
-    this.memory.clear();
-    this.interruptPlayback("conversation-reset");
-  }
-
-  statusSummary(): string {
-    const playbackState = this.player.state.status;
-    return [
-      `세션 활성: 예`,
-      `음성 채널: ${this.options.voiceChannel.name}`,
-      `추적 유저 수: ${this.trackedUsers.size}`,
-      `재생 상태: ${playbackState}`,
-      `대기열: ${this.queue.length}`,
-      `최근 대화 턴: ${this.memory.recentTurns().length}`,
-    ].join("\n");
-  }
-
-  async speakText(text: string): Promise<void> {
-    this.queue.push({
-      text,
-      source: "manual",
-    });
-    await this.drainQueue();
-  }
-
-  interruptPlayback(reason: string): void {
-    if (this.queue.length > 0 || this.player.state.status !== AudioPlayerStatus.Idle) {
-      this.options.logger.info("Interrupting playback", this.guildId, reason);
-    }
-
-    this.queue.splice(0, this.queue.length);
-    this.currentAbortController?.abort();
-    this.currentAbortController = null;
-    this.currentPlayback?.dispose();
-    this.currentPlayback = null;
-    this.player.stop(true);
-  }
-
-  async destroy(): Promise<void> {
-    this.interruptPlayback("session-destroy");
-    for (const session of this.trackedUsers.values()) {
-      session.destroy();
-    }
-    this.trackedUsers.clear();
-    this.pendingUsers.clear();
-    this.connection.destroy();
-  }
-
-  private async ensureTrackedUser(userId: string): Promise<void> {
-    if (this.trackedUsers.has(userId)) {
-      return;
-    }
-
-    const existing = this.pendingUsers.get(userId);
-    if (existing) {
-      await existing;
-      return;
-    }
-
-    const pending = this.createTrackedUser(userId).finally(() => {
-      this.pendingUsers.delete(userId);
-    });
-    this.pendingUsers.set(userId, pending);
-    await pending;
-  }
-
-  private async createTrackedUser(userId: string): Promise<void> {
-    const speakerName = await this.resolveSpeakerName(userId);
-    const receiveStream = this.connection.receiver.subscribe(userId, {
-      end: {
-        behavior: EndBehaviorType.Manual,
-      },
-    });
-
-    const decoder = new prism.opus.Decoder({
-      rate: 48000,
-      channels: 2,
-      frameSize: 960,
-    }) as NodeJS.ReadWriteStream & { destroy: () => void };
-
-    receiveStream.pipe(decoder);
-
-    const session = await UserAudioSession.create({
-      logger: this.options.logger,
-      speakerId: userId,
-      speakerName,
-      receiveStream,
-      decoder,
-      onSpeechStart: () => {
-        this.interruptPlayback(`barge-in:${speakerName}`);
-      },
-      onSpeechEnd: (utterance, audio) => {
-        void this.handleSpeechEnd(utterance, audio);
-      },
-    });
-
-    this.trackedUsers.set(userId, session);
-    this.options.logger.info("Tracking speaker", this.guildId, userId, speakerName);
-  }
-
-  private async resolveSpeakerName(userId: string): Promise<string> {
-    try {
-      const user = await this.options.client.users.fetch(userId);
-      return user.globalName ?? user.username;
-    } catch {
-      return `user-${userId.slice(-6)}`;
-    }
-  }
-
-  private async handleSpeechEnd(utterance: UserUtterance, audio: Float32Array): Promise<void> {
-    if (audio.length < 16000 * 0.25) {
-      return;
-    }
-
-    const pcmBuffer = float32ToPcm16Buffer(audio);
-    let transcript: string | null = null;
-
-    try {
-      transcript = await this.options.stt.transcribePcm16(pcmBuffer);
-    } catch (error) {
-      this.options.logger.warn("STT failed", this.guildId, utterance.speakerId, error);
-      await this.announce(`음성 인식 실패: ${utterance.speakerName}`);
-      return;
-    }
-
-    if (!transcript || transcript.trim().length === 0) {
-      return;
-    }
-
-    const hydratedUtterance: UserUtterance = {
-      ...utterance,
-      text: transcript.trim(),
-    };
-
-    this.options.logger.info("Transcript committed", this.guildId, hydratedUtterance.speakerName, hydratedUtterance.text);
-    this.memory.addUserTurn(hydratedUtterance);
-
-    if (this.options.config.DEBUG_TEXT_EVENTS) {
-      await this.announce(`🗣️ ${hydratedUtterance.speakerName}: ${hydratedUtterance.text}`);
-    }
-
-    let reply: string;
-    try {
-      reply = await this.options.llm.generateReply(this.memory, hydratedUtterance);
-    } catch (error) {
-      this.options.logger.warn("LLM failed", this.guildId, utterance.speakerId, error);
-      reply = "지금은 답변 생성에 실패했습니다. 잠시 후 다시 말씀해 주세요.";
-    }
-
-    this.memory.addAssistantTurn(reply);
-    if (this.options.config.DEBUG_TEXT_EVENTS) {
-      await this.announce(`🤖 ${reply}`);
-    }
-
-    this.queue.push({
-      text: reply,
-      source: "assistant",
-    });
-    await this.drainQueue();
-  }
-
-  private async drainQueue(): Promise<void> {
-    if (this.draining) {
-      return;
-    }
-
-    this.draining = true;
-
-    try {
-      while (this.queue.length > 0) {
-        const job = this.queue.shift();
-        if (!job) {
-          continue;
-        }
-
-        const abortController = new AbortController();
-        this.currentAbortController = abortController;
-
-        try {
-          this.currentPlayback = await this.options.tts.preparePlayback(job.text, abortController.signal);
-        } catch (error) {
-          if (abortController.signal.aborted) {
-            continue;
-          }
-
-          this.options.logger.warn("TTS synthesis failed", this.guildId, job.source, error);
-          await this.announce("음성 출력 생성에 실패했습니다.");
-          continue;
-        }
-
-        try {
-          const resource = createAudioResource(this.currentPlayback.stream, {
-            inputType: StreamType.Raw,
-          });
-          this.player.play(resource);
-
-          await entersState(this.player, AudioPlayerStatus.Playing, 20_000).catch(() => null);
-          await entersState(this.player, AudioPlayerStatus.Idle, 300_000);
-        } catch (error) {
-          if (!abortController.signal.aborted) {
-            this.options.logger.warn("Audio playback failed", this.guildId, error);
-          }
-        } finally {
-          this.currentPlayback?.dispose();
-          this.currentPlayback = null;
-          if (this.currentAbortController === abortController) {
-            this.currentAbortController = null;
-          }
-        }
-      }
-    } finally {
-      this.draining = false;
-    }
-  }
-
-  private async announce(message: string): Promise<void> {
-    if (!this.textChannelId) {
-      return;
-    }
-
-    const channel = await this.options.client.channels.fetch(this.textChannelId).catch(() => null);
-    if (!channel?.isTextBased() || !("send" in channel) || typeof channel.send !== "function") {
-      return;
-    }
-
-    await channel.send(message).catch(() => null);
-  }
-}

src/audio/local-voice-session.ts

@@ -1,499 +0,0 @@
-import { spawn, type ChildProcess, type ChildProcessByStdio } from "node:child_process";
-import { once } from "node:events";
-import { promises as fs } from "node:fs";
-import os from "node:os";
-import path from "node:path";
-import type { Readable, Writable } from "node:stream";
-
-import { RealTimeVAD } from "avr-vad";
-
-import type { AssistantRuntimeConfig } from "../config.js";
-import { Logger } from "../logger.js";
-import { requireFfmpegPath } from "./ffmpeg-path.js";
-import { takeFrame, int16ArrayToFloat32, float32ToPcm16Buffer } from "./pcm.js";
-import { ConversationMemory, type UserUtterance } from "../services/conversation.js";
-import type { LlmService } from "../services/llm.js";
-import type { SttService } from "../services/stt.js";
-import type { PreparedSpeechAudio, TtsService } from "../services/tts.js";
-
-interface LocalVoiceSessionOptions {
-  config: AssistantRuntimeConfig;
-  logger: Logger;
-  stt: SttService;
-  tts: TtsService;
-  llm: LlmService;
-}
-
-interface SpeechJob {
-  text: string;
-  source: "assistant" | "manual";
-}
-
-export class LocalVoiceSession {
-  private readonly memory: ConversationMemory;
-  private readonly queue: SpeechJob[] = [];
-  private readonly pendingSamples: number[] = [];
-
-  private vad: RealTimeVAD | null = null;
-  private recorder: ChildProcessByStdio<null, Readable, Readable> | null = null;
-  private currentPlayer: ChildProcess | null = null;
-  private currentAbortController: AbortController | null = null;
-  private currentPlayback: PreparedSpeechAudio | null = null;
-  private processing = Promise.resolve();
-  private draining = false;
-  private destroyed = false;
-
-  constructor(private readonly options: LocalVoiceSessionOptions) {
-    this.memory = new ConversationMemory(options.config.MAX_CONVERSATION_TURNS);
-  }
-
-  async start(): Promise<void> {
-    this.vad = await RealTimeVAD.new({
-      model: "v5",
-      sampleRate: 16000,
-      frameSamples: 1536,
-      positiveSpeechThreshold: 0.55,
-      negativeSpeechThreshold: 0.35,
-      redemptionFrames: 8,
-      preSpeechPadFrames: 2,
-      minSpeechFrames: 3,
-      onFrameProcessed: () => undefined,
-      onVADMisfire: () => undefined,
-      onSpeechStart: () => {
-        this.interruptPlayback("local-barge-in");
-      },
-      onSpeechRealStart: () => undefined,
-      onSpeechEnd: (audio: Float32Array) => {
-        void this.handleSpeechEnd(audio);
-      },
-    });
-
-    this.recorder = this.spawnRecorder();
-    this.recorder.stdout.on("data", (chunk: Buffer) => {
-      this.pushPcm16Chunk(chunk);
-    });
-    this.recorder.stderr.on("data", (chunk: Buffer) => {
-      const text = chunk.toString().trim();
-      if (text.length > 0) {
-        this.options.logger.debug("[pw-record]", text);
-      }
-    });
-    this.recorder.on("exit", (code, signal) => {
-      if (!this.destroyed) {
-        this.options.logger.warn("pw-record exited unexpectedly", { code, signal });
-      }
-    });
-  }
-
-  async destroy(): Promise<void> {
-    this.destroyed = true;
-    this.interruptPlayback("local-shutdown");
-
-    if (this.recorder && !this.recorder.killed) {
-      this.recorder.kill("SIGTERM");
-      await once(this.recorder, "exit").catch(() => null);
-    }
-
-    if (this.vad) {
-      await this.vad.destroy().catch((error) => {
-        this.options.logger.warn("Local VAD destroy failed", error);
-      });
-      this.vad = null;
-    }
-  }
-
-  clearConversation(): void {
-    this.memory.clear();
-    this.interruptPlayback("local-reset");
-  }
-
-  async speakText(text: string): Promise<void> {
-    this.queue.push({
-      text,
-      source: "manual",
-    });
-    await this.drainQueue();
-  }
-
-  statusSummary(): string {
-    return [
-      "모드: local",
-      `플랫폼: ${process.platform}`,
-      `입력 source: ${this.options.config.LOCAL_AUDIO_SOURCE ?? "default"}`,
-      `출력 sink: ${this.describeSink()}`,
-      `대기열: ${this.queue.length}`,
-      `최근 대화 턴: ${this.memory.recentTurns().length}`,
-    ].join("\n");
-  }
-
-  private spawnRecorder(): ChildProcessByStdio<null, Readable, Readable> {
-    if (process.platform === "win32") {
-      return this.spawnWindowsRecorder();
-    }
-
-    const args = [
-      "--rate",
-      "16000",
-      "--channels",
-      "1",
-      "--format",
-      "s16",
-      "--raw",
-    ];
-
-    if (this.options.config.LOCAL_AUDIO_SOURCE) {
-      args.push("--target", this.options.config.LOCAL_AUDIO_SOURCE);
-    }
-
-    args.push("-");
-
-    this.options.logger.info("Starting local recorder", {
-      source: this.options.config.LOCAL_AUDIO_SOURCE ?? "default",
-    });
-
-    return spawn("pw-record", args, {
-      stdio: ["ignore", "pipe", "pipe"],
-    });
-  }
-
-  private spawnWindowsRecorder(): ChildProcessByStdio<null, Readable, Readable> {
-    const ffmpegPath = this.getFfmpegPath();
-    const sourceName = this.options.config.LOCAL_AUDIO_SOURCE;
-    if (!sourceName) {
-      throw new Error("Windows 로컬 모드는 LOCAL_AUDIO_SOURCE 설정이 필요합니다. `bun run audio:devices` 로 이름을 확인해 주세요.");
-    }
-
-    const args = [
-      "-hide_banner",
-      "-loglevel",
-      "warning",
-      "-f",
-      "dshow",
-      "-i",
-      `audio=${sourceName}`,
-      "-ac",
-      "1",
-      "-ar",
-      "16000",
-      "-f",
-      "s16le",
-      "pipe:1",
-    ];
-
-    this.options.logger.info("Starting local recorder", {
-      source: sourceName,
-      backend: "ffmpeg-dshow",
-    });
-
-    return spawn(ffmpegPath, args, {
-      stdio: ["ignore", "pipe", "pipe"],
-    });
-  }
-
-  private pushPcm16Chunk(chunk: Buffer): void {
-    if (this.destroyed || !this.vad) {
-      return;
-    }
-
-    for (let offset = 0; offset + 1 < chunk.length; offset += 2) {
-      this.pendingSamples.push(chunk.readInt16LE(offset));
-    }
-
-    while (true) {
-      const frame = takeFrame(this.pendingSamples, 1536);
-      if (!frame) {
-        return;
-      }
-
-      const floatFrame = int16ArrayToFloat32(frame);
-      this.processing = this.processing
-        .then(() => this.vad?.processAudio(floatFrame))
-        .catch((error) => {
-          this.options.logger.warn("Local VAD processing failed", error);
-        });
-    }
-  }
-
-  private async handleSpeechEnd(audio: Float32Array): Promise<void> {
-    if (audio.length < 16000 * 0.25) {
-      return;
-    }
-
-    const utterance: UserUtterance = {
-      speakerId: "local-user",
-      speakerName: this.options.config.LOCAL_SPEAKER_NAME,
-      text: "",
-    };
-
-    let transcript: string | null = null;
-    try {
-      transcript = await this.options.stt.transcribePcm16(float32ToPcm16Buffer(audio));
-    } catch (error) {
-      this.options.logger.warn("Local STT failed", error);
-      return;
-    }
-
-    if (!transcript || transcript.trim().length === 0) {
-      return;
-    }
-
-    utterance.text = transcript.trim();
-    this.memory.addUserTurn(utterance);
-    this.options.logger.info("Local transcript", utterance.text);
-    if (this.options.config.DEBUG_TEXT_EVENTS) {
-      console.log(`\n[you] ${utterance.text}`);
-    }
-
-    let reply: string;
-    try {
-      reply = await this.options.llm.generateReply(this.memory, utterance);
-    } catch (error) {
-      this.options.logger.warn("Local LLM failed", error);
-      reply = "지금은 답변 생성에 실패했습니다. 잠시 후 다시 말씀해 주세요.";
-    }
-
-    this.memory.addAssistantTurn(reply);
-    if (this.options.config.DEBUG_TEXT_EVENTS) {
-      console.log(`[bot] ${reply}\n`);
-    }
-
-    this.queue.push({
-      text: reply,
-      source: "assistant",
-    });
-    await this.drainQueue();
-  }
-
-  private interruptPlayback(reason: string): void {
-    if (this.queue.length > 0 || this.currentPlayer) {
-      this.options.logger.info("Interrupting local playback", reason);
-    }
-
-    this.queue.splice(0, this.queue.length);
-    this.currentAbortController?.abort();
-    this.currentAbortController = null;
-    this.currentPlayback?.dispose();
-    this.currentPlayback = null;
-
-    if (this.currentPlayer && !this.currentPlayer.killed) {
-      this.currentPlayer.kill("SIGKILL");
-    }
-    this.currentPlayer = null;
-  }
-
-  private async drainQueue(): Promise<void> {
-    if (this.draining || this.destroyed) {
-      return;
-    }
-
-    this.draining = true;
-
-    try {
-      while (this.queue.length > 0 && !this.destroyed) {
-        const job = this.queue.shift();
-        if (!job) {
-          continue;
-        }
-
-        const abortController = new AbortController();
-        this.currentAbortController = abortController;
-
-        try {
-          this.currentPlayback = await this.options.tts.preparePlayback(job.text, abortController.signal);
-        } catch (error) {
-          if (!abortController.signal.aborted) {
-            this.options.logger.warn("Local TTS synthesis failed", error);
-          }
-          continue;
-        }
-
-        try {
-          await this.playToSink(this.currentPlayback, abortController.signal);
-        } catch (error) {
-          if (!abortController.signal.aborted) {
-            this.options.logger.warn("Local playback failed", error);
-          }
-        } finally {
-          this.currentPlayback?.dispose();
-          this.currentPlayback = null;
-          if (this.currentAbortController === abortController) {
-            this.currentAbortController = null;
-          }
-        }
-      }
-    } finally {
-      this.draining = false;
-    }
-  }
-
-  private async playToSink(playback: PreparedSpeechAudio, signal: AbortSignal): Promise<void> {
-    if (process.platform === "win32") {
-      await this.playToWindowsDefaultSink(playback, signal);
-      return;
-    }
-
-    const args = [
-      "--rate",
-      "48000",
-      "--channels",
-      "2",
-      "--format",
-      "s16",
-      "--raw",
-    ];
-
-    if (this.options.config.LOCAL_AUDIO_SINK) {
-      args.push("--target", this.options.config.LOCAL_AUDIO_SINK);
-    }
-
-    args.push("-");
-
-    const player = spawn("pw-play", args, {
-      stdio: ["pipe", "ignore", "pipe"],
-    });
-    this.currentPlayer = player;
-
-    player.stderr.on("data", (chunk: Buffer) => {
-      const text = chunk.toString().trim();
-      if (text.length > 0) {
-        this.options.logger.debug("[pw-play]", text);
-      }
-    });
-
-    signal.addEventListener(
-      "abort",
-      () => {
-        playback.stream.destroy();
-        if (!player.killed) {
-          player.kill("SIGKILL");
-        }
-      },
-      { once: true },
-    );
-
-    playback.stream.pipe(player.stdin);
-
-    const [code, playSignal] = (await once(player, "exit")) as [number | null, NodeJS.Signals | null];
-    this.currentPlayer = null;
-
-    if (signal.aborted) {
-      return;
-    }
-
-    if (code !== 0) {
-      throw new Error(`pw-play exited with code=${code ?? "null"} signal=${playSignal ?? "null"}`);
-    }
-  }
-
-  private async playToWindowsDefaultSink(playback: PreparedSpeechAudio, signal: AbortSignal): Promise<void> {
-    const chunks: Buffer[] = [];
-
-    await new Promise<void>((resolve, reject) => {
-      playback.stream.on("data", (chunk: Buffer) => {
-        chunks.push(Buffer.from(chunk));
-      });
-      playback.stream.once("end", resolve);
-      playback.stream.once("error", reject);
-      signal.addEventListener(
-        "abort",
-        () => {
-          playback.stream.destroy();
-          reject(new Error("playback aborted"));
-        },
-        { once: true },
-      );
-    }).catch((error) => {
-      if (signal.aborted) {
-        return;
-      }
-      throw error;
-    });
-
-    if (signal.aborted) {
-      return;
-    }
-
-    const pcm = Buffer.concat(chunks);
-    const wav = createWaveFileBuffer(pcm, 48000, 2, 16);
-    const tempPath = path.join(os.tmpdir(), `realtime-voice-bot-${Date.now()}.wav`);
-    await fs.writeFile(tempPath, wav);
-
-    const psScript = [
-      "Add-Type -AssemblyName System;",
-      `$player = New-Object System.Media.SoundPlayer('${tempPath.replace(/'/g, "''")}');`,
-      "$player.PlaySync();",
-    ].join(" ");
-
-    const player = spawn("powershell", ["-NoProfile", "-Command", psScript], {
-      stdio: ["ignore", "ignore", "pipe"],
-    });
-    this.currentPlayer = player;
-
-    player.stderr.on("data", (chunk: Buffer) => {
-      const text = chunk.toString().trim();
-      if (text.length > 0) {
-        this.options.logger.debug("[powershell-player]", text);
-      }
-    });
-
-    signal.addEventListener(
-      "abort",
-      () => {
-        if (!player.killed) {
-          player.kill("SIGKILL");
-        }
-      },
-      { once: true },
-    );
-
-    const [code, playSignal] = (await once(player, "exit")) as [number | null, NodeJS.Signals | null];
-    this.currentPlayer = null;
-    await fs.unlink(tempPath).catch(() => null);
-
-    if (signal.aborted) {
-      return;
-    }
-
-    if (code !== 0) {
-      throw new Error(`powershell playback exited with code=${code ?? "null"} signal=${playSignal ?? "null"}`);
-    }
-  }
-
-  private getFfmpegPath(): string {
-    return requireFfmpegPath();
-  }
-
-  private describeSink(): string {
-    if (process.platform === "win32") {
-      return this.options.config.LOCAL_AUDIO_SINK ?? "system-default";
-    }
-    return this.options.config.LOCAL_AUDIO_SINK ?? "default";
-  }
-}
-
-function createWaveFileBuffer(
-  pcm: Buffer,
-  sampleRate: number,
-  channels: number,
-  bitsPerSample: number,
-): Buffer {
-  const header = Buffer.alloc(44);
-  const byteRate = sampleRate * channels * (bitsPerSample / 8);
-  const blockAlign = channels * (bitsPerSample / 8);
-
-  header.write("RIFF", 0, 4, "ascii");
-  header.writeUInt32LE(36 + pcm.length, 4);
-  header.write("WAVE", 8, 4, "ascii");
-  header.write("fmt ", 12, 4, "ascii");
-  header.writeUInt32LE(16, 16);
-  header.writeUInt16LE(1, 20);
-  header.writeUInt16LE(channels, 22);
-  header.writeUInt32LE(sampleRate, 24);
-  header.writeUInt32LE(byteRate, 28);
-  header.writeUInt16LE(blockAlign, 32);
-  header.writeUInt16LE(bitsPerSample, 34);
-  header.write("data", 36, 4, "ascii");
-  header.writeUInt32LE(pcm.length, 40);
-
-  return Buffer.concat([header, pcm]);
-}

src/audio/pcm.ts

@@ -1,60 +0,0 @@
-export class Stereo48kToMono16kDownsampler {
-  private readonly pendingMono48k: number[] = [];
-
-  pushStereo48kChunk(chunk: Buffer): Int16Array {
-    if (chunk.length < 4) {
-      return new Int16Array();
-    }
-
-    for (let offset = 0; offset + 3 < chunk.length; offset += 4) {
-      const left = chunk.readInt16LE(offset);
-      const right = chunk.readInt16LE(offset + 2);
-      this.pendingMono48k.push(Math.round((left + right) / 2));
-    }
-
-    const outputLength = Math.floor(this.pendingMono48k.length / 3);
-    if (outputLength === 0) {
-      return new Int16Array();
-    }
-
-    const output = new Int16Array(outputLength);
-    let readIndex = 0;
-    for (let index = 0; index < outputLength; index += 1) {
-      const a = this.pendingMono48k[readIndex];
-      const b = this.pendingMono48k[readIndex + 1];
-      const c = this.pendingMono48k[readIndex + 2];
-      output[index] = Math.round((a + b + c) / 3);
-      readIndex += 3;
-    }
-
-    this.pendingMono48k.splice(0, readIndex);
-    return output;
-  }
-}
-
-export function int16ArrayToFloat32(input: Int16Array): Float32Array {
-  const output = new Float32Array(input.length);
-  for (let index = 0; index < input.length; index += 1) {
-    output[index] = input[index] / 32768;
-  }
-  return output;
-}
-
-export function float32ToPcm16Buffer(input: Float32Array): Buffer {
-  const buffer = Buffer.allocUnsafe(input.length * 2);
-  for (let index = 0; index < input.length; index += 1) {
-    const value = Math.max(-1, Math.min(1, input[index]));
-    const scaled = value < 0 ? value * 32768 : value * 32767;
-    buffer.writeInt16LE(Math.round(scaled), index * 2);
-  }
-  return buffer;
-}
-
-export function takeFrame(source: number[], frameSize: number): Int16Array | null {
-  if (source.length < frameSize) {
-    return null;
-  }
-
-  const values = source.splice(0, frameSize);
-  return Int16Array.from(values);
-}

src/audio/realtime-segmenter.ts

@@ -0,0 +1,148 @@
+interface RealtimeSegmenterOptions {
+  onSegment: (pcm16: Buffer) => void;
+  onLevel?: (peak: number) => void;
+  onSpeechStart?: (peak: number) => void;
+  onSpeechDiscarded?: (samples: number) => void;
+  onSpeechReady?: (samples: number) => void;
+  preRollSamples?: number;
+  speechStartThreshold?: number;
+  speechContinueThreshold?: number;
+  speechStartFrames?: number;
+  speechEndFrames?: number;
+  minSpeechSamples?: number;
+  maxSpeechSamples?: number;
+}
+
+export class RealtimeSegmenter {
+  private readonly pendingSamples: number[] = [];
+  private readonly preRoll: number[] = [];
+  private readonly speech: number[] = [];
+
+  private readonly frameSamples = 320;
+  private readonly preRollSamples: number;
+  private readonly speechStartThreshold: number;
+  private readonly speechContinueThreshold: number;
+  private readonly speechStartFrames: number;
+  private readonly speechEndFrames: number;
+  private readonly minSpeechSamples: number;
+  private readonly maxSpeechSamples: number;
+
+  private speechActive = false;
+  private speechCandidateFrames = 0;
+  private silenceFrames = 0;
+
+  constructor(private readonly options: RealtimeSegmenterOptions) {
+    this.preRollSamples = options.preRollSamples ?? 3200;
+    this.speechStartThreshold = options.speechStartThreshold ?? 900;
+    this.speechContinueThreshold = options.speechContinueThreshold ?? 450;
+    this.speechStartFrames = options.speechStartFrames ?? 2;
+    this.speechEndFrames = options.speechEndFrames ?? 24;
+    this.minSpeechSamples = options.minSpeechSamples ?? 7200;
+    this.maxSpeechSamples = options.maxSpeechSamples ?? 160000;
+  }
+
+  pushChunk(chunk: Buffer): void {
+    for (let offset = 0; offset + 1 < chunk.length; offset += 2) {
+      this.pendingSamples.push(chunk.readInt16LE(offset));
+    }
+
+    while (true) {
+      const frame = takeFrame(this.pendingSamples, this.frameSamples);
+      if (!frame) {
+        return;
+      }
+      this.processFrame(frame);
+    }
+  }
+
+  reset(): void {
+    this.pendingSamples.splice(0, this.pendingSamples.length);
+    this.preRoll.splice(0, this.preRoll.length);
+    this.speech.splice(0, this.speech.length);
+    this.speechActive = false;
+    this.speechCandidateFrames = 0;
+    this.silenceFrames = 0;
+  }
+
+  private processFrame(frame: Int16Array): void {
+    let peak = 0;
+    for (const sample of frame) {
+      const abs = Math.abs(sample);
+      if (abs > peak) {
+        peak = abs;
+      }
+    }
+
+    this.options.onLevel?.(peak);
+
+    if (!this.speechActive) {
+      appendWithCap(this.preRoll, frame, this.preRollSamples);
+      if (peak >= this.speechStartThreshold) {
+        this.speechCandidateFrames += 1;
+      } else {
+        this.speechCandidateFrames = 0;
+      }
+
+      if (this.speechCandidateFrames < this.speechStartFrames) {
+        return;
+      }
+
+      this.speechActive = true;
+      this.silenceFrames = 0;
+      this.speech.splice(0, this.speech.length, ...this.preRoll);
+      this.preRoll.splice(0, this.preRoll.length);
+      this.options.onSpeechStart?.(peak);
+    }
+
+    this.speech.push(...frame);
+
+    if (peak >= this.speechContinueThreshold) {
+      this.silenceFrames = 0;
+    } else {
+      this.silenceFrames += 1;
+    }
+
+    if (this.silenceFrames < this.speechEndFrames) {
+      if (this.speech.length < this.maxSpeechSamples) {
+        return;
+      }
+    }
+
+    const speechPcm = int16ArrayToBuffer(Int16Array.from(this.speech));
+    this.speechActive = false;
+    this.speech.splice(0, this.speech.length);
+    this.silenceFrames = 0;
+    this.speechCandidateFrames = 0;
+
+    if (speechPcm.length < this.minSpeechSamples * 2) {
+      this.options.onSpeechDiscarded?.(speechPcm.length / 2);
+      return;
+    }
+
+    this.options.onSpeechReady?.(speechPcm.length / 2);
+    this.options.onSegment(speechPcm);
+  }
+}
+
+function takeFrame(source: number[], size: number): Int16Array | null {
+  if (source.length < size) {
+    return null;
+  }
+  const samples = source.splice(0, size);
+  return Int16Array.from(samples);
+}
+
+function appendWithCap(target: number[], samples: Int16Array, cap: number): void {
+  target.push(...samples);
+  if (target.length > cap) {
+    target.splice(0, target.length - cap);
+  }
+}
+
+function int16ArrayToBuffer(input: Int16Array): Buffer {
+  const output = Buffer.allocUnsafe(input.length * 2);
+  for (let index = 0; index < input.length; index += 1) {
+    output.writeInt16LE(input[index]!, index * 2);
+  }
+  return output;
+}

src/config.ts

@@ -12,30 +12,46 @@ const emptyToUndefined = z.preprocess((value) => {
 }, z.string().min(1).optional());
 
 const envSchema = z.object({
-  DISCORD_BOT_TOKEN: emptyToUndefined,
-  DISCORD_APPLICATION_ID: emptyToUndefined,
-  DISCORD_COMMAND_GUILD_ID: emptyToUndefined,
-  OLLAMA_BASE_URL: z.string().min(1).default("http://localhost:11434"),
-  OLLAMA_MODEL: z.string().min(1).default("qwen3:0.6b"),
-  OLLAMA_KEEP_ALIVE: z.string().min(1).default("5m"),
-  OLLAMA_NUM_CTX: z.coerce.number().int().min(512).max(32768).default(4096),
   LOCAL_AI_VENV_PATH: z.string().min(1).default(".local-ai/.venv"),
-  LOCAL_AI_CACHE_DIR: z.string().min(1).default(".local-ai/cache"),
   LOCAL_AI_PYTHON: emptyToUndefined,
-  LOCAL_STT_MODEL: z.string().min(1).default("tiny"),
-  LOCAL_STT_DEVICE: z.string().min(1).default("auto"),
-  LOCAL_STT_COMPUTE_TYPE: z.string().min(1).default("auto"),
-  LOCAL_STT_BEAM_SIZE: z.coerce.number().int().min(1).max(8).default(1),
-  LOCAL_TTS_LANGUAGE: z.string().min(1).default("KR"),
-  LOCAL_TTS_SPEAKER: z.string().min(1).default("KR"),
-  LOCAL_TTS_DEVICE: z.string().min(1).default("auto"),
-  LOCAL_TTS_SPEED: z.coerce.number().min(0.8).max(1.6).default(1.12),
-  BOT_DEFAULT_LANGUAGE: z.string().min(2).default("ko"),
-  MAX_CONVERSATION_TURNS: z.coerce.number().int().min(4).max(30).default(12),
-  LOCAL_AUDIO_SOURCE: emptyToUndefined,
-  LOCAL_AUDIO_SINK: emptyToUndefined,
-  LOCAL_SPEAKER_NAME: z.string().min(1).default("local-user"),
-  DEBUG_TEXT_EVENTS: z
+  AUDIO_SOURCE: emptyToUndefined,
+  DOCKER_BIN: emptyToUndefined,
+  TTS_ENABLED: z
+    .string()
+    .optional()
+    .transform((value) => value?.trim().toLowerCase() !== "false"),
+  TTS_IMAGE: z.string().min(1).default("realtime-voice-bot-melotts:v0.1.2"),
+  TTS_LANGUAGE: z.string().min(1).default("KR"),
+  TTS_SPEAKER: z.string().min(1).default("KR"),
+  TTS_DEVICE: z.string().min(1).default("cpu"),
+  TTS_SPEED: z.coerce.number().min(0.5).max(2).default(1.18),
+  TTS_PLAYBACK_RATE: z.coerce.number().min(0.5).max(4).default(2.2),
+  TTS_SDP_RATIO: z.coerce.number().min(0).max(1).default(0.22),
+  TTS_NOISE_SCALE: z.coerce.number().min(0).max(2).default(0.55),
+  TTS_NOISE_SCALE_W: z.coerce.number().min(0).max(2).default(0.75),
+  TTS_CACHE_DIR: z.string().min(1).default(".local-ai/tts-cache"),
+  TTS_OUTPUT_DIR: z.string().min(1).default(".local-ai/tts-output"),
+  DEBUG: z
+    .string()
+    .optional()
+    .transform((value) => value?.trim().toLowerCase() === "true"),
+  OLLAMA_BASE_URL: z.string().min(1).default("http://127.0.0.1:11434"),
+  OLLAMA_MODEL: z.string().min(1).default("qwen3:8b"),
+  OLLAMA_KEEP_ALIVE: z.string().min(1).default("5m"),
+  MAX_CONVERSATION_TURNS: z.coerce.number().int().min(1).max(20).default(6),
+  WHISPER_MODEL: z.string().min(1).default("large-v3-turbo"),
+  WHISPER_LANGUAGE: z.string().min(1).default("ko"),
+  WHISPER_DEVICE: z.enum(["auto", "cuda", "cpu"]).default("auto"),
+  WHISPER_COMPUTE_TYPE: z.string().min(1).default("auto"),
+  WHISPER_BEAM_SIZE: z.coerce.number().int().min(1).max(8).default(2),
+  SEGMENT_START_THRESHOLD: z.coerce.number().int().min(100).max(10000).default(900),
+  SEGMENT_CONTINUE_THRESHOLD: z.coerce.number().int().min(50).max(10000).default(450),
+  SEGMENT_START_FRAMES: z.coerce.number().int().min(1).max(10).default(2),
+  SEGMENT_END_FRAMES: z.coerce.number().int().min(4).max(60).default(24),
+  SEGMENT_PREROLL_SAMPLES: z.coerce.number().int().min(320).max(16000).default(3200),
+  SEGMENT_MIN_SPEECH_SAMPLES: z.coerce.number().int().min(1600).max(64000).default(7200),
+  SEGMENT_MAX_SPEECH_SAMPLES: z.coerce.number().int().min(16000).max(320000).default(160000),
+  DEBUG_TRANSCRIPTS: z
     .string()
     .optional()
     .transform((value) => value === "true"),
@@ -43,32 +59,7 @@ const envSchema = z.object({
 });
 
 export type AppConfig = z.infer<typeof envSchema>;
-export type AssistantRuntimeConfig = AppConfig;
-export type DiscordRuntimeConfig = AssistantRuntimeConfig & {
-  DISCORD_BOT_TOKEN: string;
-  DISCORD_APPLICATION_ID: string;
-};
 
 export function loadConfig(): AppConfig {
   return envSchema.parse(process.env);
 }
-
-function requirePresent(value: string | undefined, name: string): string {
-  if (!value) {
-    throw new Error(`${name} 환경변수가 필요합니다.`);
-  }
-  return value;
-}
-
-export function requireAssistantRuntimeConfig(config: AppConfig): AssistantRuntimeConfig {
-  return config;
-}
-
-export function requireDiscordRuntimeConfig(config: AppConfig): DiscordRuntimeConfig {
-  const assistant = requireAssistantRuntimeConfig(config);
-  return {
-    ...assistant,
-    DISCORD_BOT_TOKEN: requirePresent(config.DISCORD_BOT_TOKEN, "DISCORD_BOT_TOKEN"),
-    DISCORD_APPLICATION_ID: requirePresent(config.DISCORD_APPLICATION_ID, "DISCORD_APPLICATION_ID"),
-  };
-}

src/discord-main.ts

@@ -1,238 +0,0 @@
-import process from "node:process";
-
-import {
-  GatewayIntentBits,
-  REST,
-  Routes,
-  SlashCommandBuilder,
-  type ChatInputCommandInteraction,
-  type Client,
-  type GuildMember,
-  type VoiceBasedChannel,
-} from "discord.js";
-import { Client as DiscordClient } from "discord.js";
-
-import { GuildVoiceSession } from "./audio/guild-voice-session.js";
-import { type DiscordRuntimeConfig } from "./config.js";
-import { Logger } from "./logger.js";
-import { LocalFasterWhisperSttService } from "./services/local-stt.js";
-import { LocalMeloTtsService } from "./services/local-tts.js";
-import { OllamaLlmService } from "./services/ollama-llm.js";
-
-export async function runDiscordBot(config: DiscordRuntimeConfig, logger: Logger): Promise<void> {
-  const commands = [
-    new SlashCommandBuilder().setName("join").setDescription("현재 들어가 있는 음성 채널에 봇을 입장시킵니다."),
-    new SlashCommandBuilder().setName("leave").setDescription("현재 음성 세션을 종료합니다."),
-    new SlashCommandBuilder().setName("status").setDescription("현재 음성 세션 상태를 확인합니다."),
-    new SlashCommandBuilder().setName("reset").setDescription("대화 문맥과 재생 큐를 초기화합니다."),
-    new SlashCommandBuilder()
-      .setName("say")
-      .setDescription("텍스트를 바로 음성으로 읽습니다.")
-      .addStringOption((option) =>
-        option.setName("text").setDescription("읽을 문장").setRequired(true).setMaxLength(400),
-      ),
-  ].map((command) => command.toJSON());
-
-  const client = new DiscordClient({
-    intents: [GatewayIntentBits.Guilds, GatewayIntentBits.GuildVoiceStates],
-  });
-
-  const stt = new LocalFasterWhisperSttService(config, logger);
-  const tts = new LocalMeloTtsService(config, logger);
-  const llm = new OllamaLlmService(config);
-  const sessions = new Map<string, GuildVoiceSession>();
-
-  await stt.warmup();
-  await tts.warmup();
-
-  function getVoiceChannel(interaction: ChatInputCommandInteraction): VoiceBasedChannel | null {
-    const member = interaction.member as GuildMember | null;
-    return member?.voice.channel ?? null;
-  }
-
-  async function registerCommands(_appClient: Client): Promise<void> {
-    const rest = new REST({ version: "10" }).setToken(config.DISCORD_BOT_TOKEN);
-    if (config.DISCORD_COMMAND_GUILD_ID) {
-      await rest.put(
-        Routes.applicationGuildCommands(config.DISCORD_APPLICATION_ID, config.DISCORD_COMMAND_GUILD_ID),
-        {
-          body: commands,
-        },
-      );
-      logger.info("Registered guild commands", config.DISCORD_COMMAND_GUILD_ID);
-      return;
-    }
-
-    await rest.put(Routes.applicationCommands(config.DISCORD_APPLICATION_ID), {
-      body: commands,
-    });
-    logger.info("Registered global commands");
-  }
-
-  async function createSession(interaction: ChatInputCommandInteraction): Promise<GuildVoiceSession> {
-    if (!interaction.guild) {
-      throw new Error("Guild interaction required");
-    }
-
-    const voiceChannel = getVoiceChannel(interaction);
-    if (!voiceChannel) {
-      throw new Error("먼저 음성 채널에 들어가 주세요.");
-    }
-
-    const existing = sessions.get(interaction.guild.id);
-    if (existing && existing.voiceChannelId === voiceChannel.id) {
-      existing.setTextChannel(interaction.channelId);
-      return existing;
-    }
-
-    if (existing) {
-      await existing.destroy();
-      sessions.delete(interaction.guild.id);
-    }
-
-    const session = await GuildVoiceSession.create({
-      client,
-      config,
-      logger,
-      guild: interaction.guild,
-      voiceChannel,
-      textChannelId: interaction.channelId,
-      stt,
-      tts,
-      llm,
-    });
-    sessions.set(interaction.guild.id, session);
-    return session;
-  }
-
-  async function handleJoin(interaction: ChatInputCommandInteraction): Promise<void> {
-    await interaction.deferReply({ ephemeral: true });
-
-    try {
-      const session = await createSession(interaction);
-      await interaction.editReply(
-        `음성 비서를 시작했습니다. 채널: ${session.statusSummary().split("\n")[1]?.replace("음성 채널: ", "") ?? "알 수 없음"}`,
-      );
-    } catch (error) {
-      const message = error instanceof Error ? error.message : "세션 생성에 실패했습니다.";
-      await interaction.editReply(message);
-    }
-  }
-
-  async function handleLeave(interaction: ChatInputCommandInteraction): Promise<void> {
-    const session = interaction.guild ? sessions.get(interaction.guild.id) : undefined;
-    if (!session) {
-      await interaction.reply({ content: "현재 활성화된 음성 세션이 없습니다.", ephemeral: true });
-      return;
-    }
-
-    await session.destroy();
-    sessions.delete(interaction.guildId!);
-    await interaction.reply({ content: "음성 세션을 종료했습니다.", ephemeral: true });
-  }
-
-  async function handleStatus(interaction: ChatInputCommandInteraction): Promise<void> {
-    const session = interaction.guild ? sessions.get(interaction.guild.id) : undefined;
-    if (!session) {
-      await interaction.reply({ content: "현재 활성화된 음성 세션이 없습니다.", ephemeral: true });
-      return;
-    }
-
-    await interaction.reply({
-      content: session.statusSummary(),
-      ephemeral: true,
-    });
-  }
-
-  async function handleReset(interaction: ChatInputCommandInteraction): Promise<void> {
-    const session = interaction.guild ? sessions.get(interaction.guild.id) : undefined;
-    if (!session) {
-      await interaction.reply({ content: "현재 활성화된 음성 세션이 없습니다.", ephemeral: true });
-      return;
-    }
-
-    session.clearConversation();
-    await interaction.reply({ content: "대화 문맥과 재생 큐를 초기화했습니다.", ephemeral: true });
-  }
-
-  async function handleSay(interaction: ChatInputCommandInteraction): Promise<void> {
-    await interaction.deferReply({ ephemeral: true });
-
-    const session = interaction.guild ? sessions.get(interaction.guild.id) : undefined;
-    if (!session) {
-      await interaction.editReply("먼저 `/join` 으로 음성 세션을 시작해 주세요.");
-      return;
-    }
-
-    const text = interaction.options.getString("text", true).trim();
-    await session.speakText(text);
-    await interaction.editReply("읽기 요청을 대기열에 추가했습니다.");
-  }
-
-  async function shutdown(exitCode = 0): Promise<void> {
-    logger.info("Shutting down");
-    for (const session of sessions.values()) {
-      await session.destroy().catch((error) => {
-        logger.warn("Session shutdown failed", error);
-      });
-    }
-    sessions.clear();
-    await Promise.allSettled([stt.destroy?.(), tts.destroy?.()]);
-    await client.destroy();
-    process.exit(exitCode);
-  }
-
-  client.once("ready", async () => {
-    logger.info("Discord client ready", client.user?.tag ?? "unknown");
-    try {
-      await registerCommands(client);
-    } catch (error) {
-      logger.error("Command registration failed", error);
-    }
-  });
-
-  client.on("interactionCreate", async (interaction) => {
-    if (!interaction.isChatInputCommand()) {
-      return;
-    }
-
-    try {
-      switch (interaction.commandName) {
-        case "join":
-          await handleJoin(interaction);
-          return;
-        case "leave":
-          await handleLeave(interaction);
-          return;
-        case "status":
-          await handleStatus(interaction);
-          return;
-        case "reset":
-          await handleReset(interaction);
-          return;
-        case "say":
-          await handleSay(interaction);
-          return;
-        default:
-          await interaction.reply({ content: "알 수 없는 명령입니다.", ephemeral: true });
-      }
-    } catch (error) {
-      logger.error("Interaction handler failed", error);
-      if (interaction.deferred || interaction.replied) {
-        await interaction.editReply("명령 처리 중 오류가 발생했습니다.").catch(() => null);
-        return;
-      }
-      await interaction.reply({ content: "명령 처리 중 오류가 발생했습니다.", ephemeral: true }).catch(() => null);
-    }
-  });
-
-  process.on("SIGINT", () => {
-    void shutdown(0);
-  });
-
-  process.on("SIGTERM", () => {
-    void shutdown(0);
-  });
-
-  await client.login(config.DISCORD_BOT_TOKEN);
-}

src/docker-runtime.ts

@@ -0,0 +1,93 @@
+import { spawn } from "node:child_process";
+import { constants as fsConstants } from "node:fs";
+import { access } from "node:fs/promises";
+import path from "node:path";
+import process from "node:process";
+
+import type { AppConfig } from "./config.js";
+
+async function fileExists(target: string): Promise<boolean> {
+  try {
+    await access(target, fsConstants.F_OK);
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+async function captureStdout(command: string, args: string[]): Promise<string | null> {
+  return await new Promise<string | null>((resolve) => {
+    const child = spawn(command, args, {
+      stdio: ["ignore", "pipe", "ignore"],
+      windowsHide: true,
+    });
+
+    let stdout = "";
+    child.stdout.on("data", (chunk: Buffer) => {
+      stdout += chunk.toString();
+    });
+
+    child.on("error", () => resolve(null));
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve(stdout);
+        return;
+      }
+      resolve(null);
+    });
+  });
+}
+
+async function resolveWithWhere(): Promise<string | null> {
+  const stdout = await captureStdout("cmd.exe", ["/d", "/s", "/c", "where docker"]);
+  if (!stdout) {
+    return null;
+  }
+
+  const candidates = stdout
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0);
+
+  for (const candidate of candidates) {
+    if (await fileExists(candidate)) {
+      return candidate;
+    }
+  }
+
+  return null;
+}
+
+export async function resolveDockerCommand(config: AppConfig): Promise<string> {
+  if (config.DOCKER_BIN && await fileExists(config.DOCKER_BIN)) {
+    return config.DOCKER_BIN;
+  }
+
+  if (process.platform !== "win32") {
+    return "docker";
+  }
+
+  const commonPaths = [
+    "C:\\Program Files\\Docker\\Docker\\resources\\bin\\docker.exe",
+    "C:\\Program Files\\Docker\\Docker\\resources\\bin\\docker-cli.exe",
+  ];
+
+  for (const candidate of commonPaths) {
+    if (await fileExists(candidate)) {
+      return candidate;
+    }
+  }
+
+  const found = await resolveWithWhere();
+  if (found) {
+    return found;
+  }
+
+  throw new Error(
+    [
+      "Docker 실행 파일을 찾지 못했습니다.",
+      "VSCode를 완전히 다시 열어 PATH를 새로 고치거나,",
+      ".env에 DOCKER_BIN=C:\\Program Files\\Docker\\Docker\\resources\\bin\\docker.exe 를 넣어주세요.",
+    ].join(" "),
+  );
+}

src/index.ts

@@ -1,31 +1,427 @@
 import process from "node:process";
+import { createInterface } from "node:readline";
 
-import { loadConfig, requireAssistantRuntimeConfig, requireDiscordRuntimeConfig } from "./config.js";
-import { runDiscordBot } from "./discord-main.js";
+import { loadConfig } from "./config.js";
 import { Logger } from "./logger.js";
-import { printLocalAudioDevices, runLocalAssistant } from "./local-main.js";
+import { printAudioDevices, spawnLoopbackCapture } from "./audio/capture.js";
+import { RealtimeSegmenter } from "./audio/realtime-segmenter.js";
+import { FasterWhisperSttService } from "./services/faster-whisper-stt.js";
+import { MeloTtsService } from "./services/melo-tts.js";
+import { OllamaLlmService } from "./services/ollama-llm.js";
 
-const mode = process.argv[2] ?? "discord";
-const config = loadConfig();
-const logger = new Logger(config.LOG_LEVEL);
+const mode = process.argv[2] ?? "test-stt";
+
+async function runSttTest(options: { enableLlm: boolean; enableTts: boolean }): Promise<void> {
+  const config = loadConfig();
+  const logger = new Logger(config.DEBUG ? config.LOG_LEVEL : "error");
+  const stt = new FasterWhisperSttService(config, logger);
+  const llm = options.enableLlm ? new OllamaLlmService(config, logger) : null;
+  let tts = options.enableTts && config.TTS_ENABLED ? new MeloTtsService(config, logger) : null;
+  let capture = null as ReturnType<typeof spawnLoopbackCapture> | null;
+  let shuttingDown: Promise<void> | null = null;
+  let suppressCapture = false;
+  let receivedChunks = 0;
+  let receivedBytes = 0;
+  let maxPeak = 0;
+  let lastChunkAt = 0;
+  let lastLevelLogAt = 0;
+  let sawSpeechStart = false;
+  let emittedSegmentCount = 0;
+
+  const shutdown = async (exitCode: number, reason: string, error?: unknown): Promise<void> => {
+    if (shuttingDown) {
+      return await shuttingDown;
+    }
+
+    shuttingDown = (async () => {
+      if (error) {
+        logger.error(`Shutting down: ${reason}`, error);
+      } else {
+        logger.info("Shutting down", reason);
+      }
+
+      if (capture && !capture.killed && capture.exitCode === null) {
+        capture.kill("SIGTERM");
+      }
+
+      await stt.destroy().catch((destroyError) => {
+        logger.warn("STT destroy failed", destroyError);
+      });
+      if (tts) {
+        await tts.destroy().catch((destroyError) => {
+          logger.warn("TTS destroy failed", destroyError);
+        });
+      }
+    })();
+
+    await shuttingDown;
+    process.exit(exitCode);
+  };
+
+  process.once("SIGINT", () => {
+    void shutdown(0, "SIGINT");
+  });
+  process.once("SIGTERM", () => {
+    void shutdown(0, "SIGTERM");
+  });
+  process.once("uncaughtException", (error) => {
+    void shutdown(1, "uncaughtException", error);
+  });
+  process.once("unhandledRejection", (reason) => {
+    void shutdown(1, "unhandledRejection", reason);
+  });
+  process.once("exit", () => {
+    if (capture && !capture.killed && capture.exitCode === null) {
+      capture.kill("SIGKILL");
+    }
+    void stt.destroy();
+    if (tts) {
+      void tts.destroy();
+    }
+  });
+
+  console.log("STT 준비중...");
+  await stt.warmup();
+  logger.info("STT warmup finished");
+  console.log("STT 준비 완료");
+  if (llm) {
+    console.log("LLM 준비중...");
+    await llm.warmup();
+    logger.info("LLM warmup finished");
+    console.log("LLM 준비 완료");
+  }
+  if (tts) {
+    console.log("TTS 준비중...");
+    try {
+      await tts.warmup();
+      logger.info("TTS warmup finished", {
+        image: config.TTS_IMAGE,
+        language: config.TTS_LANGUAGE,
+        speaker: config.TTS_SPEAKER,
+      });
+      console.log("TTS 준비 완료");
+    } catch (error) {
+      logger.warn("TTS warmup failed", error);
+      console.log("TTS 비활성화: bun run setup:tts 를 먼저 실행하세요.");
+      tts = null;
+    }
+  }
+
+  const transcriptionQueue: Array<{ pcm16: Buffer; queuedAt: number; index: number }> = [];
+  let transcribing = false;
+  let nextSegmentIndex = 1;
+
+  const runNext = async (): Promise<void> => {
+    if (transcribing) {
+      return;
+    }
+    const next = transcriptionQueue.shift();
+    if (!next) {
+      return;
+    }
+
+    transcribing = true;
+    try {
+      const startedAt = Date.now();
+      const text = await stt.transcribePcm16(next.pcm16);
+      logger.info("STT latency", {
+        index: next.index,
+        wait_ms: startedAt - next.queuedAt,
+        transcribe_ms: Date.now() - startedAt,
+      });
+      if (!text) {
+        logger.info("빈 전사 결과");
+      } else {
+        logger.info("Transcript", { index: next.index, text });
+        if (config.DEBUG) {
+          if (config.DEBUG_TRANSCRIPTS) {
+            console.log(`\n[text] ${text}\n`);
+          }
+        } else {
+          console.log(`사용자> ${text}`);
+        }
+
+        if (llm) {
+          const assessmentStartedAt = Date.now();
+          const assessment = await llm.assessReplyNeed(text);
+          logger.info("Reply assessment", {
+            index: next.index,
+            should_reply: assessment.shouldReply,
+            likely_needs_lookup: assessment.likelyNeedsLookup,
+            reason: assessment.reason,
+            assessment_ms: Date.now() - assessmentStartedAt,
+          });
+
+          if (!assessment.shouldReply) {
+            if (config.DEBUG) {
+              console.log(`[skip] ${assessment.reason}\n`);
+            }
+            return;
+          }
+
+          const llmStartedAt = Date.now();
+          const reply = await llm.generateReply(text, {
+            onProgress: (message) => {
+              if (config.DEBUG) {
+                console.log(`[assistant] ${message}`);
+                return;
+              }
+              console.log(`답변> ${message}`);
+            },
+          });
+          logger.info("LLM latency", {
+            index: next.index,
+            llm_ms: Date.now() - llmStartedAt,
+          });
+          logger.info("LLM reply", { index: next.index, text: reply });
+
+          if (config.DEBUG) {
+            if (config.DEBUG_TRANSCRIPTS) {
+              console.log(`[assistant] ${reply}\n`);
+            }
+          } else {
+            console.log(`답변> ${reply}`);
+          }
+
+          if (tts) {
+            suppressCapture = true;
+            segmenter.reset();
+            try {
+              await tts.speak(reply);
+            } catch (error) {
+              logger.warn("TTS playback failed", error);
+            } finally {
+              suppressCapture = false;
+              sawSpeechStart = false;
+              maxPeak = 0;
+            }
+          }
+        }
+      }
+    } catch (error) {
+      logger.error("STT/LLM failed", error);
+    } finally {
+      transcribing = false;
+      void runNext();
+    }
+  };
+
+  const segmenter = new RealtimeSegmenter({
+    preRollSamples: config.SEGMENT_PREROLL_SAMPLES,
+    speechStartThreshold: config.SEGMENT_START_THRESHOLD,
+    speechContinueThreshold: config.SEGMENT_CONTINUE_THRESHOLD,
+    speechStartFrames: config.SEGMENT_START_FRAMES,
+    speechEndFrames: config.SEGMENT_END_FRAMES,
+    minSpeechSamples: config.SEGMENT_MIN_SPEECH_SAMPLES,
+    maxSpeechSamples: config.SEGMENT_MAX_SPEECH_SAMPLES,
+    onLevel: (peak) => {
+      if (peak > maxPeak) {
+        maxPeak = peak;
+      }
+
+      const now = Date.now();
+      if (now - lastLevelLogAt >= 3000) {
+        lastLevelLogAt = now;
+        logger.info("Audio input heartbeat", {
+          chunks: receivedChunks,
+          bytes: receivedBytes,
+          peak: maxPeak,
+          speech_started: sawSpeechStart,
+          emitted_segments: emittedSegmentCount,
+        });
+        maxPeak = 0;
+      }
+    },
+    onSpeechStart: (peak) => {
+      sawSpeechStart = true;
+      logger.info("Speech start detected", { peak });
+    },
+    onSpeechDiscarded: (samples) => {
+      logger.info("Discarded short speech segment", { samples });
+    },
+    onSpeechReady: (samples) => {
+      emittedSegmentCount += 1;
+      logger.info("Speech segment ready", {
+        index: emittedSegmentCount,
+        samples,
+        ms: Math.round((samples / 16000) * 1000),
+      });
+    },
+    onSegment: (pcm16) => {
+      const index = nextSegmentIndex++;
+      transcriptionQueue.push({
+        pcm16,
+        queuedAt: Date.now(),
+        index,
+      });
+      logger.info("Queued segment for STT", {
+        index,
+        queue: transcriptionQueue.length,
+        bytes: pcm16.length,
+      });
+      void runNext();
+    },
+  });
+
+  capture = spawnLoopbackCapture(config, logger);
+  capture.stdout.on("data", (chunk: Buffer) => {
+    receivedChunks += 1;
+    receivedBytes += chunk.length;
+    lastChunkAt = Date.now();
+    if (suppressCapture) {
+      return;
+    }
+    segmenter.pushChunk(chunk);
+  });
+  capture.stderr.on("data", (chunk: Buffer) => {
+    const text = chunk.toString().trim();
+    if (text) {
+      logger.debug("[capture]", text);
+    }
+  });
+  capture.on("error", (error) => {
+    void shutdown(1, "capture-error", error);
+  });
+  capture.on("exit", (code, signal) => {
+    logger.warn("capture exited", { code, signal });
+    if (!shuttingDown) {
+      void shutdown(1, "capture-exit");
+    }
+  });
+
+  if (config.DEBUG) {
+    if (options.enableLlm && options.enableTts) {
+      console.log("실시간 출력장치 STT+LLM+TTS 테스트를 시작합니다. Ctrl+C 로 종료합니다.");
+    } else if (options.enableLlm) {
+      console.log("실시간 출력장치 STT+LLM 테스트를 시작합니다. Ctrl+C 로 종료합니다.");
+    } else {
+      console.log("실시간 출력장치 STT 테스트를 시작합니다. Ctrl+C 로 종료합니다.");
+    }
+    console.log(`source: ${config.AUDIO_SOURCE ?? "unset"}`);
+    console.log(`model: ${config.WHISPER_MODEL}`);
+    console.log(`language: ${config.WHISPER_LANGUAGE}`);
+    console.log(`beam: ${config.WHISPER_BEAM_SIZE}`);
+    if (options.enableLlm) {
+      console.log(`llm: ${config.OLLAMA_MODEL}`);
+    }
+    if (options.enableTts) {
+      console.log(`tts: ${config.TTS_IMAGE}`);
+    }
+  }
+
+  setInterval(() => {
+    const now = Date.now();
+    if (lastChunkAt === 0 && !shuttingDown) {
+      logger.warn("아직 캡처 PCM 데이터가 들어오지 않았습니다. AUDIO_SOURCE 가 잘못됐거나 loopback 입력이 아닌 장치일 수 있습니다.");
+      return;
+    }
+
+    if (lastChunkAt > 0 && now - lastChunkAt >= 5000 && !shuttingDown) {
+      logger.warn("최근 5초 동안 새 PCM chunk 가 들어오지 않았습니다.");
+    }
+  }, 5000).unref();
+}
+
+async function runLlmCli(): Promise<void> {
+  const config = loadConfig();
+  const logger = new Logger(config.DEBUG ? config.LOG_LEVEL : "error");
+  const llm = new OllamaLlmService(config, logger);
+
+  await llm.warmup();
+
+  console.log(`LLM CLI 테스트를 시작합니다. model=${config.OLLAMA_MODEL}`);
+  console.log("/exit 로 종료, /reset 으로 대화 초기화");
+
+  const rl = createInterface({
+    input: process.stdin,
+    output: process.stdout,
+    prompt: "you> ",
+  });
+
+  rl.prompt();
+
+  rl.on("line", async (line) => {
+    const text = line.trim();
+
+    if (!text) {
+      rl.prompt();
+      return;
+    }
+
+    if (text === "/exit") {
+      rl.close();
+      return;
+    }
+
+    if (text === "/reset") {
+      llm.resetConversation();
+      console.log("assistant> 대화 문맥을 초기화했습니다.");
+      rl.prompt();
+      return;
+    }
+
+    try {
+      const startedAt = Date.now();
+      const reply = await llm.generateReply(text, {
+        onProgress: (message) => {
+          console.log(`assistant> ${message}`);
+        },
+      });
+      logger.info("LLM latency", {
+        llm_ms: Date.now() - startedAt,
+      });
+      console.log(`assistant> ${reply}`);
+    } catch (error) {
+      console.error(error instanceof Error ? error.message : String(error));
+    }
+
+    rl.prompt();
+  });
+
+  rl.on("close", () => {
+    process.exit(0);
+  });
+}
+
+async function runTtsTest(): Promise<void> {
+  const text = process.argv.slice(3).join(" ").trim() || "안녕하세요. 로컬 티티에스 테스트입니다.";
+  const config = loadConfig();
+  const logger = new Logger(config.DEBUG ? config.LOG_LEVEL : "error");
+  const tts = new MeloTtsService(config, logger);
+
+  console.log("TTS 준비중...");
+  await tts.warmup();
+  console.log("TTS 준비 완료");
+  console.log(`재생 문장: ${text}`);
+  await tts.speak(text);
+}
 
 async function main(): Promise<void> {
   switch (mode) {
-    case "discord":
-      await runDiscordBot(requireDiscordRuntimeConfig(config), logger);
+    case "devices":
+      await printAudioDevices();
       return;
-    case "local":
-      await runLocalAssistant(requireAssistantRuntimeConfig(config), logger);
+    case "test-stt":
+      await runSttTest({ enableLlm: false, enableTts: false });
       return;
-    case "local-devices":
-      await printLocalAudioDevices();
+    case "test-sttllm":
+      await runSttTest({ enableLlm: true, enableTts: false });
+      return;
+    case "test-all":
+      await runSttTest({ enableLlm: true, enableTts: true });
+      return;
+    case "test-llm":
+      await runLlmCli();
+      return;
+    case "test-tts":
+      await runTtsTest();
       return;
     default:
-      throw new Error(`알 수 없는 실행 모드입니다: ${mode}. 사용 가능: discord, local, local-devices`);
+      throw new Error(`알 수 없는 실행 모드입니다: ${mode}. 사용 가능: test-stt, test-sttllm, test-all, test-llm, test-tts, devices`);
   }
 }
 
 void main().catch((error) => {
-  logger.error("Fatal startup error", error);
+  console.error(error instanceof Error ? error.message : String(error));
   process.exit(1);
 });

src/local-main.ts

@@ -1,111 +0,0 @@
-import { spawn } from "node:child_process";
-import process from "node:process";
-
-import type { AssistantRuntimeConfig } from "./config.js";
-import { Logger } from "./logger.js";
-import { LocalVoiceSession } from "./audio/local-voice-session.js";
-import { requireFfmpegPath } from "./audio/ffmpeg-path.js";
-import { LocalFasterWhisperSttService } from "./services/local-stt.js";
-import { LocalMeloTtsService } from "./services/local-tts.js";
-import { OllamaLlmService } from "./services/ollama-llm.js";
-
-export async function printLocalAudioDevices(): Promise<void> {
-  if (process.platform === "win32") {
-    const ffmpegPath = requireFfmpegPath();
-
-    console.log("\n=== ffmpeg dshow audio devices ===");
-    await new Promise<void>((resolve, reject) => {
-      const child = spawn(
-        ffmpegPath,
-        ["-hide_banner", "-list_devices", "true", "-f", "dshow", "-i", "dummy"],
-        {
-          stdio: ["ignore", "ignore", "inherit"],
-        },
-      );
-      child.on("exit", (code) => {
-        if (code === 0 || code === 1) {
-          resolve();
-          return;
-        }
-        reject(new Error(`ffmpeg exited with code ${code ?? "null"}`));
-      });
-      child.on("error", reject);
-    });
-
-    console.log("\n위 목록의 오디오 장치 이름을 `LOCAL_AUDIO_SOURCE` 에 그대로 넣으면 됩니다.");
-    console.log("Windows 로컬 모드는 현재 출력 장치 직접 선택 대신 시스템 기본 출력 장치를 사용합니다.");
-    return;
-  }
-
-  const runs = [
-    {
-      label: "wpctl status",
-      args: ["status"],
-    },
-    {
-      label: "wpctl status -n",
-      args: ["status", "-n"],
-    },
-  ] as const;
-
-  for (const run of runs) {
-    console.log(`\n=== ${run.label} ===`);
-    await new Promise<void>((resolve, reject) => {
-      const child = spawn("wpctl", run.args, {
-        stdio: ["ignore", "inherit", "inherit"],
-      });
-      child.on("exit", (code) => {
-        if (code === 0) {
-          resolve();
-          return;
-        }
-        reject(new Error(`wpctl exited with code ${code ?? "null"}`));
-      });
-      child.on("error", reject);
-    });
-  }
-}
-
-export async function runLocalAssistant(config: AssistantRuntimeConfig, logger: Logger): Promise<void> {
-  const stt = new LocalFasterWhisperSttService(config, logger);
-  const tts = new LocalMeloTtsService(config, logger);
-  const llm = new OllamaLlmService(config);
-
-  await stt.warmup();
-  await tts.warmup();
-
-  const session = new LocalVoiceSession({
-    config,
-    logger,
-    stt,
-    tts,
-    llm,
-  });
-
-  console.log(session.statusSummary());
-  console.log("로컬 음성 테스트를 시작합니다. Ctrl+C 로 종료합니다.");
-  if (process.platform === "win32") {
-    console.log("Windows 로컬 모드는 현재 시스템 기본 출력 장치로 재생됩니다.");
-  }
-  if (config.DEBUG_TEXT_EVENTS) {
-    console.log("텍스트 로그 출력이 켜져 있습니다.");
-  }
-
-  const shutdown = async (exitCode = 0) => {
-    await session.destroy().catch((error) => {
-      logger.warn("Local session shutdown failed", error);
-    });
-    await Promise.allSettled([stt.destroy?.(), tts.destroy?.()]);
-    process.exit(exitCode);
-  };
-
-  process.on("SIGINT", () => {
-    void shutdown(0);
-  });
-
-  process.on("SIGTERM", () => {
-    void shutdown(0);
-  });
-
-  await session.start();
-}

src/prompt-loader.ts

@@ -0,0 +1,20 @@
+import { readFileSync } from "node:fs";
+import path from "node:path";
+
+const cache = new Map<string, string>();
+
+export function loadPrompt(name: string): string {
+  const cached = cache.get(name);
+  if (cached) {
+    return cached;
+  }
+
+  const promptPath = path.resolve(process.cwd(), "prompts", name);
+  const content = readFileSync(promptPath, "utf8").trim();
+  if (!content) {
+    throw new Error(`프롬프트 파일이 비어 있습니다: ${promptPath}`);
+  }
+
+  cache.set(name, content);
+  return content;
+}

src/python-runtime.ts

@@ -1,90 +1,203 @@
-import { existsSync } from "node:fs";
-import { spawnSync } from "node:child_process";
+import { spawn } from "node:child_process";
+import { constants as fsConstants } from "node:fs";
+import { access } from "node:fs/promises";
 import path from "node:path";
+import process from "node:process";
 
 import type { AppConfig } from "./config.js";
 
-export interface PythonLaunch {
+export interface PythonCommandSpec {
   command: string;
   args: string[];
-  source: "venv" | "configured" | "system";
+  viaCmdShell?: boolean;
 }
 
-function splitCommandSpec(spec: string): string[] {
-  return spec.match(/(?:[^\s"]+|"[^"]*")+/g)?.map((part) => part.replace(/^"|"$/g, "")) ?? [];
+function shouldUseCmdShell(command: string): boolean {
+  if (process.platform !== "win32") {
+    return false;
+  }
+
+  const lower = command.toLowerCase();
+  return !path.isAbsolute(command) || lower.endsWith(".bat") || lower.endsWith(".cmd");
 }
 
-function canRun(command: string, args: string[]): boolean {
-  const result = spawnSync(command, [...args, "--version"], {
-    encoding: "utf8",
+function splitCommand(command: string): string[] {
+  const parts = command.match(/(?:[^\s"]+|"[^"]*")+/g) ?? [];
+  return parts.map((part) => part.replace(/^"(.*)"$/, "$1"));
+}
+
+function quoteWindowsCmdArg(value: string): string {
+  if (!/[ \t"&()<>^|]/.test(value)) {
+    return value;
+  }
+  return `"${value.replace(/"/g, '""')}"`;
+}
+
+function buildWindowsCommandLine(parts: string[]): string {
+  return parts.map((part) => quoteWindowsCmdArg(part)).join(" ");
+}
+
+export function buildPythonInvocation(spec: PythonCommandSpec, extraArgs: string[]): PythonCommandSpec {
+  if (process.platform === "win32" && spec.viaCmdShell) {
+    return {
+      command: "cmd.exe",
+      args: ["/d", "/s", "/c", buildWindowsCommandLine([spec.command, ...spec.args, ...extraArgs])],
+    };
+  }
+
+  return {
+    command: spec.command,
+    args: [...spec.args, ...extraArgs],
+  };
+}
+
+async function canRun(command: string, args: string[], viaCmdShell = false): Promise<boolean> {
+  const invocation = viaCmdShell
+    ? {
+        command: "cmd.exe",
+        args: ["/d", "/s", "/c", buildWindowsCommandLine([command, ...args, "--version"])],
+      }
+    : {
+        command,
+        args: [...args, "--version"],
+      };
+
+  return await new Promise<boolean>((resolve) => {
+    const child = spawn(invocation.command, invocation.args, {
+      stdio: ["ignore", "ignore", "ignore"],
+      windowsHide: true,
+    });
+
+    child.on("error", () => {
+      resolve(false);
+    });
+
+    child.on("exit", (code) => {
+      resolve(code === 0);
+    });
   });
-  return result.status === 0;
 }
 
-export function resolveLocalAiVenvPath(config: AppConfig): string {
-  return path.resolve(process.cwd(), config.LOCAL_AI_VENV_PATH);
+async function captureStdout(command: string, args: string[]): Promise<string | null> {
+  return await new Promise<string | null>((resolve) => {
+    const child = spawn(command, args, {
+      stdio: ["ignore", "pipe", "ignore"],
+      windowsHide: true,
+    });
+
+    let stdout = "";
+    child.stdout.on("data", (chunk: Buffer) => {
+      stdout += chunk.toString();
+    });
+
+    child.on("error", () => {
+      resolve(null);
+    });
+
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve(stdout);
+        return;
+      }
+      resolve(null);
+    });
+  });
 }
 
-export function resolveLocalAiCachePath(config: AppConfig): string {
-  return path.resolve(process.cwd(), config.LOCAL_AI_CACHE_DIR);
-}
-
-export function resolveVenvPythonPath(config: AppConfig): string {
-  const venvPath = resolveLocalAiVenvPath(config);
-  return process.platform === "win32"
-    ? path.join(venvPath, "Scripts", "python.exe")
-    : path.join(venvPath, "bin", "python");
-}
-
-export function resolvePythonLaunch(config: AppConfig, options?: { preferVenv?: boolean }): PythonLaunch {
-  const preferVenv = options?.preferVenv ?? true;
-  const venvPython = resolveVenvPythonPath(config);
-
-  if (preferVenv && existsSync(venvPython)) {
-    return {
-      command: venvPython,
-      args: [],
-      source: "venv",
-    };
+async function resolveWindowsExecutable(name: string): Promise<string | null> {
+  const stdout = await captureStdout("cmd.exe", ["/d", "/s", "/c", `where ${name}`]);
+  if (!stdout) {
+    return null;
   }
 
-  const configured = config.LOCAL_AI_PYTHON ? splitCommandSpec(config.LOCAL_AI_PYTHON) : [];
-  if (configured.length > 0 && canRun(configured[0]!, configured.slice(1))) {
-    return {
-      command: configured[0]!,
-      args: configured.slice(1),
-      source: "configured",
-    };
+  const candidates = stdout
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0);
+
+  for (const candidate of candidates) {
+    try {
+      await access(candidate, fsConstants.F_OK);
+      return candidate;
+    } catch {
+      // ignore
+    }
   }
 
-  const candidates =
-    process.platform === "win32"
-      ? [
-          ["py", "-3"],
-          ["python"],
-          ["python3"],
-        ]
-      : [
-          ["python3"],
-          ["python"],
-        ];
+  return null;
+}
 
-  for (const [command, ...args] of candidates) {
-    if (canRun(command, args)) {
+async function fileExists(target: string): Promise<boolean> {
+  try {
+    await access(target, fsConstants.X_OK);
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+export async function resolvePythonCommand(config: AppConfig): Promise<{ command: string; args: string[] }> {
+  return await resolveWorkerPythonCommand(config);
+}
+
+export async function resolveBasePythonCommand(config: AppConfig): Promise<PythonCommandSpec> {
+  const configured = config.LOCAL_AI_PYTHON?.trim();
+  if (configured) {
+    const [command, ...args] = splitCommand(configured);
+    if (!command) {
+      throw new Error("LOCAL_AI_PYTHON 값이 비어 있습니다.");
+    }
     return {
       command,
       args,
-        source: "system",
+      viaCmdShell: shouldUseCmdShell(command),
     };
   }
+
+  const venvPath = resolveVenvPythonPath(config);
+  if (await fileExists(venvPath)) {
+    return { command: venvPath, args: [] };
+  }
+
+  if (process.platform === "win32") {
+    return {
+      command: "python",
+      args: [],
+      viaCmdShell: true,
+    };
+  }
+
+  const unixCandidates = [
+    { command: "python3", args: [] as string[] },
+    { command: "python", args: [] as string[] },
+  ];
+
+  for (const candidate of unixCandidates) {
+    if (await canRun(candidate.command, candidate.args)) {
+      return candidate;
+    }
   }
 
-  throw new Error(
-    [
-      "Python 실행 파일을 찾지 못했습니다.",
-      "1. Python 3.11 이상을 설치",
-      "2. 필요하면 `.env` 에 `LOCAL_AI_PYTHON=python` 또는 `LOCAL_AI_PYTHON=py -3` 설정",
-      "3. 그 다음 `bun run setup:local-ai` 실행",
-    ].join("\n"),
-  );
+  throw new Error("사용 가능한 Python 실행기를 찾지 못했습니다. `python3 --version` 또는 `python --version` 이 먼저 동작해야 합니다.");
+}
+
+export async function resolveWorkerPythonCommand(config: AppConfig): Promise<PythonCommandSpec> {
+  const venvPath = resolveVenvPythonPath(config);
+  if (await fileExists(venvPath)) {
+    return { command: venvPath, args: [] };
+  }
+
+  return await resolveBasePythonCommand(config);
+}
+
+export function resolveVenvPythonPath(config: AppConfig): string {
+  const root = path.resolve(process.cwd(), config.LOCAL_AI_VENV_PATH);
+  if (process.platform === "win32") {
+    return path.join(root, "Scripts", "python.exe");
+  }
+  return path.join(root, "bin", "python");
+}
+
+export function resolveWorkerScript(name: string): string {
+  return path.resolve(process.cwd(), "python", name);
 }

src/services/audio-playback.ts

@@ -0,0 +1,104 @@
+import { spawn } from "node:child_process";
+import { rm } from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import process from "node:process";
+import { randomUUID } from "node:crypto";
+
+async function run(command: string, args: string[], env?: NodeJS.ProcessEnv): Promise<void> {
+  await new Promise<void>((resolve, reject) => {
+    const child = spawn(command, args, {
+      stdio: ["ignore", "inherit", "inherit"],
+      windowsHide: true,
+      env,
+    });
+
+    child.on("error", reject);
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve();
+        return;
+      }
+      reject(new Error(`${command} ${args.join(" ")} exited with code ${code ?? "null"}`));
+    });
+  });
+}
+
+function buildAtempoFilter(rate: number): string {
+  const filters: string[] = [];
+  let remaining = rate;
+
+  while (remaining > 2) {
+    filters.push("atempo=2.0");
+    remaining /= 2;
+  }
+
+  while (remaining < 0.5) {
+    filters.push("atempo=0.5");
+    remaining /= 0.5;
+  }
+
+  filters.push(`atempo=${remaining.toFixed(3)}`);
+  return filters.join(",");
+}
+
+async function applyPlaybackRate(filePath: string, playbackRate: number): Promise<string> {
+  if (Math.abs(playbackRate - 1) < 0.01) {
+    return filePath;
+  }
+
+  const targetPath = path.join(os.tmpdir(), `realtime-voice-bot-tts-${randomUUID()}.wav`);
+  const filter = buildAtempoFilter(playbackRate);
+
+  await run("ffmpeg", [
+    "-y",
+    "-hide_banner",
+    "-loglevel",
+    "error",
+    "-i",
+    filePath,
+    "-filter:a",
+    filter,
+    targetPath,
+  ]);
+
+  return targetPath;
+}
+
+async function playNativeWavFile(filePath: string): Promise<void> {
+  if (process.platform === "win32") {
+    const env = {
+      ...process.env,
+      TTS_WAV_PATH: filePath,
+    };
+
+    await run("powershell.exe", [
+      "-NoProfile",
+      "-NonInteractive",
+      "-ExecutionPolicy",
+      "Bypass",
+      "-Command",
+      [
+        "$path = $env:TTS_WAV_PATH",
+        "$player = New-Object System.Media.SoundPlayer $path",
+        "$player.Load()",
+        "$player.PlaySync()",
+      ].join("; "),
+    ], env);
+    return;
+  }
+
+  throw new Error(`지원하지 않는 플랫폼입니다: ${process.platform}`);
+}
+
+export async function playWavFile(filePath: string, playbackRate = 1): Promise<void> {
+  const playablePath = await applyPlaybackRate(filePath, playbackRate);
+
+  try {
+    await playNativeWavFile(playablePath);
+  } finally {
+    if (playablePath !== filePath) {
+      await rm(playablePath, { force: true }).catch(() => undefined);
+    }
+  }
+}

src/services/conversation.ts

@@ -1,77 +0,0 @@
-export interface ConversationTurn {
-  role: "user" | "assistant";
-  text: string;
-  speakerId?: string;
-  speakerName?: string;
-  createdAt: number;
-}
-
-export interface UserUtterance {
-  speakerId: string;
-  speakerName: string;
-  text: string;
-}
-
-export class ConversationMemory {
-  private readonly turns: ConversationTurn[] = [];
-
-  constructor(private readonly maxTurns: number) {}
-
-  addUserTurn(utterance: UserUtterance): void {
-    this.turns.push({
-      role: "user",
-      text: utterance.text,
-      speakerId: utterance.speakerId,
-      speakerName: utterance.speakerName,
-      createdAt: Date.now(),
-    });
-    this.trim();
-  }
-
-  addAssistantTurn(text: string): void {
-    this.turns.push({
-      role: "assistant",
-      text,
-      createdAt: Date.now(),
-    });
-    this.trim();
-  }
-
-  clear(): void {
-    this.turns.splice(0, this.turns.length);
-  }
-
-  recentTurns(): ConversationTurn[] {
-    return [...this.turns];
-  }
-
-  buildPrompt(currentUtterance: UserUtterance): string {
-    const recent = this.turns
-      .slice(-this.maxTurns)
-      .map((turn) => {
-        if (turn.role === "assistant") {
-          return `[assistant]\n${turn.text}`;
-        }
-        return `[user speaker_id=${turn.speakerId ?? "unknown"} speaker_name=${turn.speakerName ?? "unknown"}]\n${turn.text}`;
-      })
-      .join("\n\n");
-
-    const historyBlock = recent.length > 0 ? recent : "(이전 대화 없음)";
-
-    return [
-      "최근 대화:",
-      historyBlock,
-      "",
-      "이번 발화:",
-      `[user speaker_id=${currentUtterance.speakerId} speaker_name=${currentUtterance.speakerName}]`,
-      currentUtterance.text,
-    ].join("\n");
-  }
-
-  private trim(): void {
-    const overflow = this.turns.length - this.maxTurns;
-    if (overflow > 0) {
-      this.turns.splice(0, overflow);
-    }
-  }
-}

src/services/faster-whisper-stt.ts

@@ -0,0 +1,40 @@
+import type { AppConfig } from "../config.js";
+import type { Logger } from "../logger.js";
+import { PythonJsonWorker } from "./python-json-worker.js";
+
+interface PingResult {
+  model: string;
+  device: string;
+  compute_type: string;
+}
+
+interface TranscribeResult {
+  text: string;
+}
+
+export class FasterWhisperSttService {
+  private readonly worker: PythonJsonWorker;
+
+  constructor(
+    private readonly config: AppConfig,
+    private readonly logger: Logger,
+  ) {
+    this.worker = new PythonJsonWorker(config, logger, "loopback_stt_worker.py", "faster-whisper");
+  }
+
+  async warmup(): Promise<void> {
+    const result = await this.worker.request<PingResult>("ping", {});
+    this.logger.info("STT worker ready", result);
+  }
+
+  async transcribePcm16(pcm16: Buffer): Promise<string> {
+    const result = await this.worker.request<TranscribeResult>("transcribe", {
+      pcm16_base64: pcm16.toString("base64"),
+    });
+    return result.text.trim();
+  }
+
+  async destroy(): Promise<void> {
+    await this.worker.destroy();
+  }
+}

src/services/llm.ts

@@ -1,5 +0,0 @@
-import type { ConversationMemory, UserUtterance } from "./conversation.js";
-
-export interface LlmService {
-  generateReply(memory: ConversationMemory, utterance: UserUtterance): Promise<string>;
-}

src/services/local-stt.ts

@@ -1,43 +0,0 @@
-import type { AssistantRuntimeConfig } from "../config.js";
-import type { Logger } from "../logger.js";
-import { PythonJsonWorker } from "./python-json-worker.js";
-import type { SttService } from "./stt.js";
-
-interface TranscribeResult {
-  text?: string;
-}
-
-export class LocalFasterWhisperSttService implements SttService {
-  private readonly worker: PythonJsonWorker;
-
-  constructor(private readonly config: AssistantRuntimeConfig, logger: Logger) {
-    this.worker = new PythonJsonWorker(config, logger, "local_stt_worker.py", "local-stt", {
-      LOCAL_STT_MODEL: config.LOCAL_STT_MODEL,
-      LOCAL_STT_DEVICE: config.LOCAL_STT_DEVICE,
-      LOCAL_STT_COMPUTE_TYPE: config.LOCAL_STT_COMPUTE_TYPE,
-      LOCAL_STT_BEAM_SIZE: String(config.LOCAL_STT_BEAM_SIZE),
-    });
-  }
-
-  async warmup(): Promise<void> {
-    await this.worker.request("ping", {});
-  }
-
-  async transcribePcm16(pcm16MonoAudio: Buffer): Promise<string | null> {
-    if (pcm16MonoAudio.byteLength === 0) {
-      return null;
-    }
-
-    const result = await this.worker.request<TranscribeResult>("transcribe", {
-      audio_base64: pcm16MonoAudio.toString("base64"),
-      language: this.config.BOT_DEFAULT_LANGUAGE,
-    });
-
-    const transcript = result.text?.trim() ?? "";
-    return transcript.length > 0 ? transcript : null;
-  }
-
-  async destroy(): Promise<void> {
-    await this.worker.destroy();
-  }
-}

src/services/local-tts.ts

@@ -1,94 +0,0 @@
-import { Readable } from "node:stream";
-
-import prism from "prism-media";
-
-import type { AssistantRuntimeConfig } from "../config.js";
-import type { Logger } from "../logger.js";
-import { resolveFfmpegPath } from "../audio/ffmpeg-path.js";
-import { PythonJsonWorker } from "./python-json-worker.js";
-import type { PreparedSpeechAudio, TtsService } from "./tts.js";
-
-interface SynthesizeResult {
-  wav_base64?: string;
-}
-
-export class LocalMeloTtsService implements TtsService {
-  private readonly worker: PythonJsonWorker;
-
-  constructor(config: AssistantRuntimeConfig, logger: Logger) {
-    const resolvedFfmpegPath = resolveFfmpegPath();
-    if (resolvedFfmpegPath && !process.env.FFMPEG_PATH) {
-      process.env.FFMPEG_PATH = resolvedFfmpegPath;
-    }
-
-    this.worker = new PythonJsonWorker(config, logger, "local_tts_worker.py", "local-tts", {
-      LOCAL_TTS_LANGUAGE: config.LOCAL_TTS_LANGUAGE,
-      LOCAL_TTS_SPEAKER: config.LOCAL_TTS_SPEAKER,
-      LOCAL_TTS_DEVICE: config.LOCAL_TTS_DEVICE,
-      LOCAL_TTS_SPEED: String(config.LOCAL_TTS_SPEED),
-    });
-  }
-
-  async warmup(): Promise<void> {
-    await this.worker.request("ping", {});
-  }
-
-  async preparePlayback(text: string, signal?: AbortSignal): Promise<PreparedSpeechAudio> {
-    const result = await this.worker.request<SynthesizeResult>(
-      "synthesize",
-      {
-        text,
-      },
-      signal,
-    );
-
-    const wavBase64 = result.wav_base64;
-    if (!wavBase64) {
-      throw new Error("로컬 TTS가 빈 오디오를 반환했습니다.");
-    }
-
-    const input = Readable.from([Buffer.from(wavBase64, "base64")]);
-    const ffmpeg = new prism.FFmpeg({
-      args: [
-        "-analyzeduration",
-        "0",
-        "-loglevel",
-        "0",
-        "-i",
-        "pipe:0",
-        "-f",
-        "s16le",
-        "-ar",
-        "48000",
-        "-ac",
-        "2",
-        "pipe:1",
-      ],
-    });
-
-    if (signal) {
-      signal.addEventListener(
-        "abort",
-        () => {
-          input.destroy();
-          ffmpeg.destroy();
-        },
-        { once: true },
-      );
-    }
-
-    input.pipe(ffmpeg);
-
-    return {
-      stream: ffmpeg,
-      dispose: () => {
-        input.destroy();
-        ffmpeg.destroy();
-      },
-    };
-  }
-
-  async destroy(): Promise<void> {
-    await this.worker.destroy();
-  }
-}

src/services/melo-tts.ts

@@ -0,0 +1,364 @@
+import { spawn, type ChildProcessWithoutNullStreams } from "node:child_process";
+import { randomUUID } from "node:crypto";
+import { mkdir, rm } from "node:fs/promises";
+import { once } from "node:events";
+import path from "node:path";
+import { createInterface } from "node:readline";
+
+import type { AppConfig } from "../config.js";
+import { resolveDockerCommand } from "../docker-runtime.js";
+import type { Logger } from "../logger.js";
+import { playWavFile } from "./audio-playback.js";
+
+async function run(command: string, args: string[], stdio: "ignore" | "inherit" = "ignore"): Promise<void> {
+  const env = { ...process.env };
+  if (path.isAbsolute(command)) {
+    const dockerBinDir = path.dirname(command);
+    const currentPath = env.PATH ?? env.Path ?? "";
+    env.PATH = `${dockerBinDir}${path.delimiter}${currentPath}`;
+  }
+
+  await new Promise<void>((resolve, reject) => {
+    const child = spawn(command, args, {
+      stdio: ["ignore", stdio, "inherit"],
+      windowsHide: true,
+      shell: process.platform === "win32" && !path.isAbsolute(command),
+      env,
+    });
+
+    child.on("error", (error) => {
+      if ((error as NodeJS.ErrnoException).code === "ENOENT" && command === "docker") {
+        reject(new Error("Docker를 찾지 못했습니다. Docker Desktop을 설치하고 실행한 뒤 다시 시도하세요."));
+        return;
+      }
+      reject(error);
+    });
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve();
+        return;
+      }
+      reject(new Error(`${command} ${args.join(" ")} exited with code ${code ?? "null"}`));
+    });
+  });
+}
+
+interface RpcSuccess<T> {
+  id: string;
+  result: T;
+}
+
+interface RpcFailure {
+  id: string;
+  error: string;
+}
+
+type RpcResponse<T> = RpcSuccess<T> | RpcFailure;
+
+function isFailure<T>(value: RpcResponse<T>): value is RpcFailure {
+  return "error" in value;
+}
+
+interface TtsPingResult {
+  language: string;
+  speaker: string;
+  speaker_id: number;
+  device: string;
+  speed: number;
+  sdp_ratio: number;
+  noise_scale: number;
+  noise_scale_w: number;
+  speaker_count: number;
+}
+
+export class MeloTtsService {
+  private processRef: ChildProcessWithoutNullStreams | null = null;
+  private shuttingDown = false;
+  private warmedUp = false;
+  private readonly pending = new Map<
+    string,
+    {
+      resolve: (value: unknown) => void;
+      reject: (reason?: unknown) => void;
+    }
+  >();
+  private nextId = 1;
+
+  constructor(
+    private readonly config: AppConfig,
+    private readonly logger: Logger,
+  ) {}
+
+  async warmup(): Promise<void> {
+    if (this.warmedUp) {
+      return;
+    }
+
+    await mkdir(path.resolve(process.cwd(), this.config.TTS_CACHE_DIR), { recursive: true });
+    await mkdir(path.resolve(process.cwd(), this.config.TTS_OUTPUT_DIR), { recursive: true });
+    const docker = await resolveDockerCommand(this.config);
+
+    await run(docker, ["--version"]);
+    await run(docker, ["image", "inspect", this.config.TTS_IMAGE]);
+
+    await this.start();
+    const result = await this.request<TtsPingResult>("ping", {});
+    this.logger.info("TTS worker ready", result);
+
+    const warmupFileName = `warmup-${randomUUID()}.wav`;
+    const warmupHostPath = path.resolve(process.cwd(), this.config.TTS_OUTPUT_DIR, warmupFileName);
+    try {
+      await this.request("synthesize", {
+        text: "안녕하세요. 로컬 티티에스 준비 테스트입니다.",
+        output_path: `/work/output/${warmupFileName}`,
+      });
+    } finally {
+      await rm(warmupHostPath, { force: true }).catch(() => undefined);
+    }
+
+    this.warmedUp = true;
+  }
+
+  async speak(text: string): Promise<void> {
+    const trimmed = this.normalizeText(text);
+    if (!trimmed) {
+      return;
+    }
+
+    const fileName = `tts-${Date.now()}-${randomUUID()}.wav`;
+    const targetPath = path.resolve(process.cwd(), this.config.TTS_OUTPUT_DIR, fileName);
+
+    try {
+      await this.synthesizeToFile(trimmed, targetPath, fileName);
+      await playWavFile(targetPath, this.config.TTS_PLAYBACK_RATE);
+    } finally {
+      await rm(targetPath, { force: true }).catch(() => undefined);
+    }
+  }
+
+  async synthesizeToFile(text: string, targetPath: string, fileName?: string): Promise<void> {
+    await this.warmup();
+    const resolvedFileName = fileName ?? path.basename(targetPath);
+
+    this.logger.info("Starting MeloTTS synthesis", {
+      image: this.config.TTS_IMAGE,
+      language: this.config.TTS_LANGUAGE,
+      speaker: this.config.TTS_SPEAKER,
+      speed: this.config.TTS_SPEED,
+      playback_rate: this.config.TTS_PLAYBACK_RATE,
+      sdp_ratio: this.config.TTS_SDP_RATIO,
+      noise_scale: this.config.TTS_NOISE_SCALE,
+      noise_scale_w: this.config.TTS_NOISE_SCALE_W,
+      device: this.config.TTS_DEVICE,
+    });
+
+    await this.request("synthesize", {
+      text,
+      output_path: `/work/output/${resolvedFileName}`,
+    });
+  }
+
+  async destroy(): Promise<void> {
+    if (!this.processRef) {
+      return;
+    }
+
+    const child = this.processRef;
+    this.shuttingDown = true;
+
+    try {
+      child.stdin.end();
+    } catch {
+      // ignore
+    }
+
+    if (!child.killed && child.exitCode === null) {
+      child.kill("SIGTERM");
+    }
+
+    const timedWait = Promise.race([
+      once(child, "exit"),
+      new Promise<null>((resolve) => setTimeout(() => resolve(null), 1500)),
+    ]);
+
+    await timedWait;
+
+    if (child.exitCode === null && !child.killed) {
+      child.kill("SIGKILL");
+      await once(child, "exit").catch(() => null);
+    }
+
+    this.processRef = null;
+    this.shuttingDown = false;
+    this.warmedUp = false;
+  }
+
+  private normalizeText(input: string): string {
+    const collapsed = input
+      .replace(/[`*_#>\[\]\(\)]/g, " ")
+      .replace(/\s+/g, " ")
+      .trim();
+
+    if (!collapsed) {
+      return "";
+    }
+
+    if (/[.!?…]$/.test(collapsed)) {
+      return collapsed;
+    }
+
+    return `${collapsed}.`;
+  }
+
+  private async start(): Promise<void> {
+    if (this.processRef) {
+      return;
+    }
+    if (this.shuttingDown) {
+      throw new Error("tts worker is shutting down");
+    }
+
+    const docker = await resolveDockerCommand(this.config);
+    const outputDir = path.resolve(process.cwd(), this.config.TTS_OUTPUT_DIR);
+    const cacheDir = path.resolve(process.cwd(), this.config.TTS_CACHE_DIR);
+
+    await mkdir(outputDir, { recursive: true });
+    await mkdir(cacheDir, { recursive: true });
+
+    const args = [
+      "run",
+      "--rm",
+      "-i",
+      "-v",
+      `${outputDir}:/work/output`,
+      "-v",
+      `${cacheDir}:/cache`,
+      "-e",
+      "HF_HOME=/cache/huggingface",
+      "-e",
+      "HF_HUB_CACHE=/cache/huggingface/hub",
+      "-e",
+      "TRANSFORMERS_CACHE=/cache/transformers",
+      "-e",
+      `TTS_LANGUAGE=${this.config.TTS_LANGUAGE}`,
+      "-e",
+      `TTS_SPEAKER=${this.config.TTS_SPEAKER}`,
+      "-e",
+      `TTS_DEVICE=${this.config.TTS_DEVICE}`,
+      "-e",
+      `TTS_SPEED=${this.config.TTS_SPEED}`,
+      "-e",
+      `TTS_SDP_RATIO=${this.config.TTS_SDP_RATIO}`,
+      "-e",
+      `TTS_NOISE_SCALE=${this.config.TTS_NOISE_SCALE}`,
+      "-e",
+      `TTS_NOISE_SCALE_W=${this.config.TTS_NOISE_SCALE_W}`,
+      "--entrypoint",
+      "python",
+    ];
+
+    if (this.config.TTS_DEVICE !== "cpu") {
+      args.push("--gpus", "all");
+    }
+
+    args.push(
+      this.config.TTS_IMAGE,
+      "/opt/realtime-voice-bot/melo_tts_worker.py",
+    );
+
+    const env = { ...process.env };
+    if (path.isAbsolute(docker)) {
+      const dockerBinDir = path.dirname(docker);
+      const currentPath = env.PATH ?? env.Path ?? "";
+      env.PATH = `${dockerBinDir}${path.delimiter}${currentPath}`;
+    }
+
+    this.processRef = spawn(docker, args, {
+      stdio: ["pipe", "pipe", "pipe"],
+      windowsHide: true,
+      shell: process.platform === "win32" && !path.isAbsolute(docker),
+      env,
+    });
+
+    const rl = createInterface({
+      input: this.processRef.stdout,
+      crlfDelay: Infinity,
+    });
+
+    rl.on("line", (line) => {
+      this.handleStdoutLine(line);
+    });
+
+    this.processRef.stderr.on("data", (chunk: Buffer) => {
+      const text = chunk.toString().trim();
+      if (text.length > 0) {
+        this.logger.warn(`[melotts] ${text}`);
+      }
+    });
+
+    this.processRef.stdin.on("error", (error) => {
+      this.logger.debug("melotts stdin error", error);
+    });
+
+    this.processRef.on("exit", (code, signal) => {
+      const error = new Error(`melotts worker exited code=${code ?? "null"} signal=${signal ?? "null"}`);
+      for (const entry of this.pending.values()) {
+        entry.reject(error);
+      }
+      this.pending.clear();
+      this.processRef = null;
+    });
+  }
+
+  private async request<T>(method: string, params: Record<string, unknown>): Promise<T> {
+    await this.start();
+
+    if (!this.processRef) {
+      throw new Error("melotts worker is not running");
+    }
+
+    const id = String(this.nextId++);
+    const payload = JSON.stringify({
+      id,
+      method,
+      params,
+    });
+
+    const promise = new Promise<T>((resolve, reject) => {
+      this.pending.set(id, {
+        resolve: (value) => resolve(value as T),
+        reject,
+      });
+    });
+
+    this.processRef.stdin.write(`${payload}\n`);
+    return await promise;
+  }
+
+  private handleStdoutLine(line: string): void {
+    const trimmed = line.trim();
+    if (!trimmed) {
+      return;
+    }
+
+    let message: RpcResponse<unknown>;
+    try {
+      message = JSON.parse(trimmed) as RpcResponse<unknown>;
+    } catch (error) {
+      this.logger.warn("melotts stdout parse failed", error);
+      return;
+    }
+
+    const pending = this.pending.get(message.id);
+    if (!pending) {
+      return;
+    }
+
+    this.pending.delete(message.id);
+    if (isFailure(message)) {
+      pending.reject(new Error(message.error));
+      return;
+    }
+    pending.resolve(message.result);
+  }
+}

src/services/ollama-llm.ts

@@ -1,85 +1,550 @@
-import type { AssistantRuntimeConfig } from "../config.js";
-import type { ConversationMemory, UserUtterance } from "./conversation.js";
-import type { LlmService } from "./llm.js";
+import type { AppConfig } from "../config.js";
+import type { Logger } from "../logger.js";
+import { loadPrompt } from "../prompt-loader.js";
+import { webFetch, webSearch } from "./web-tools.js";
 
-const ASSISTANT_INSTRUCTIONS = [
-  "너는 디스코드 음성 채널 또는 로컬 마이크 테스트에서 동작하는 한국어 음성 비서다.",
-  "답변은 짧고 실용적으로 한다.",
-  "기본은 한 문장, 길어도 두 문장을 넘기지 않는다.",
-  "말투는 자연스러운 한국어로 유지한다.",
-  "speaker_id와 speaker_name은 화자 구분용이므로 필요할 때만 자연스럽게 반영한다.",
-  "잘 못 들었거나 의미가 불명확하면 짧게 다시 물어본다.",
-  "목록, 마크다운, 코드블록은 쓰지 않는다.",
-  "생각 과정을 드러내지 말고 최종 답변만 말한다.",
-].join(" ");
+interface OllamaChatMessage {
+  role: "system" | "user" | "assistant";
+  content: string;
+  tool_calls?: OllamaToolCall[];
+}
 
 interface OllamaChatResponse {
   message?: {
     content?: string;
-    thinking?: string;
+    tool_calls?: OllamaToolCall[];
   };
-  error?: string;
 }
 
-function normalizeReply(text: string): string {
-  const strippedThink = text.replace(/<think>[\s\S]*?<\/think>/gi, " ");
-  const compact = strippedThink.replace(/\s+/g, " ").trim();
-  if (compact.length <= 180) {
-    return compact;
-  }
-
-  const sentences = compact.match(/[^.!?]+[.!?]?/g);
-  if (!sentences || sentences.length === 0) {
-    return compact.slice(0, 180).trim();
-  }
-
-  return sentences.slice(0, 2).join(" ").trim().slice(0, 180).trim();
+interface OllamaToolCall {
+  type: "function";
+  function: {
+    name: string;
+    arguments: Record<string, unknown>;
+  };
 }
 
-export class OllamaLlmService implements LlmService {
-  constructor(private readonly config: AssistantRuntimeConfig) {}
+interface OllamaToolDefinition {
+  type: "function";
+  function: {
+    name: string;
+    description: string;
+    parameters: {
+      type: "object";
+      required?: string[];
+      properties: Record<string, unknown>;
+    };
+  };
+}
 
-  async generateReply(memory: ConversationMemory, utterance: UserUtterance): Promise<string> {
-    const url = new URL("/api/chat", this.config.OLLAMA_BASE_URL);
-    const response = await fetch(url, {
+interface OllamaToolResultMessage {
+  role: "tool";
+  tool_name: string;
+  content: string;
+}
+
+interface GenerateReplyOptions {
+  onProgress?: (message: string) => void;
+}
+
+export interface ReplyAssessment {
+  shouldReply: boolean;
+  likelyNeedsLookup: boolean;
+  reason: string;
+}
+
+const ASSISTANT_PROMPT = loadPrompt("assistant.md");
+const REPLY_GATE_PROMPT = loadPrompt("reply-gate.md");
+const REWRITE_KOREAN_PROMPT = loadPrompt("rewrite-korean.md");
+
+const TOOL_DEFINITIONS: OllamaToolDefinition[] = [
+  {
+    type: "function",
+    function: {
+      name: "get_current_time",
+      description: "현재 시스템 시간을 Asia/Seoul 기준 ISO 문자열과 사람이 읽기 쉬운 형식으로 반환한다.",
+      parameters: {
+        type: "object",
+        properties: {},
+      },
+    },
+  },
+  {
+    type: "function",
+    function: {
+      name: "get_runtime_settings",
+      description: "현재 로컬 LLM 및 STT 실행 설정의 핵심 값만 반환한다.",
+      parameters: {
+        type: "object",
+        properties: {},
+      },
+    },
+  },
+  {
+    type: "function",
+    function: {
+      name: "list_project_commands",
+      description: "현재 프로젝트에서 사용 가능한 주요 bun 스크립트 명령 목록을 반환한다.",
+      parameters: {
+        type: "object",
+        properties: {},
+      },
+    },
+  },
+  {
+    type: "function",
+    function: {
+      name: "evaluate_math",
+      description: "간단한 산술식을 정확히 계산한다. 숫자, 공백, 소수점, 괄호, + - * / % 만 허용한다.",
+      parameters: {
+        type: "object",
+        required: ["expression"],
+        properties: {
+          expression: {
+            type: "string",
+            description: "예: (11434+12341)*412",
+          },
+        },
+      },
+    },
+  },
+  {
+    type: "function",
+    function: {
+      name: "web_search",
+      description: "웹 검색 결과 제목, URL, 요약을 가져온다. 최신 정보, 뉴스, 사실 확인이 필요할 때만 사용한다.",
+      parameters: {
+        type: "object",
+        required: ["query"],
+        properties: {
+          query: {
+            type: "string",
+            description: "검색어",
+          },
+          max_results: {
+            type: "number",
+            description: "가져올 최대 결과 수. 보통 3~5",
+          },
+        },
+      },
+    },
+  },
+  {
+    type: "function",
+    function: {
+      name: "fetch_url",
+      description: "주어진 URL의 페이지 제목과 본문 텍스트를 읽어온다. 검색 결과 상세 확인에 사용한다.",
+      parameters: {
+        type: "object",
+        required: ["url"],
+        properties: {
+          url: {
+            type: "string",
+            description: "http 또는 https URL",
+          },
+          max_chars: {
+            type: "number",
+            description: "본문에서 가져올 최대 글자 수",
+          },
+        },
+      },
+    },
+  },
+];
+
+export class OllamaLlmService {
+  private history: OllamaChatMessage[] = [];
+
+  constructor(
+    private readonly config: AppConfig,
+    private readonly logger: Logger,
+  ) {}
+
+  async warmup(): Promise<void> {
+    const reply = await this.chat(
+      [
+        { role: "system", content: ASSISTANT_PROMPT },
+        { role: "user", content: "준비 상태 확인입니다. 한 단어로만 답하세요." },
+      ],
+    );
+    this.logger.info("LLM warmup finished", { model: this.config.OLLAMA_MODEL, reply: reply.content });
+  }
+
+  async assessReplyNeed(userText: string): Promise<ReplyAssessment> {
+    const heuristic = this.assessReplyNeedHeuristically(userText);
+    if (heuristic) {
+      return heuristic;
+    }
+
+    const reply = await this.chat([
+      { role: "system", content: REPLY_GATE_PROMPT },
+      { role: "user", content: userText },
+    ], { enableTools: false });
+
+    const parsed = this.parseAssessment(reply.content);
+    if (parsed) {
+      return parsed;
+    }
+
+    return {
+      shouldReply: true,
+      likelyNeedsLookup: this.mightNeedLookup(userText),
+      reason: "fallback",
+    };
+  }
+
+  async generateReply(userText: string, options?: GenerateReplyOptions): Promise<string> {
+    const messages: Array<OllamaChatMessage | OllamaToolResultMessage> = [
+      { role: "system", content: ASSISTANT_PROMPT },
+      ...this.history,
+      { role: "user", content: userText },
+    ];
+
+    const rawReply = await this.runAgentLoop(messages, options);
+    const reply = await this.repairReplyLanguageIfNeeded(rawReply, userText);
+
+    this.history.push({ role: "user", content: userText });
+    this.history.push({ role: "assistant", content: reply });
+    this.trimHistory();
+
+    return reply;
+  }
+
+  resetConversation(): void {
+    this.history = [];
+  }
+
+  private trimHistory(): void {
+    const maxMessages = this.config.MAX_CONVERSATION_TURNS * 2;
+    if (this.history.length <= maxMessages) {
+      return;
+    }
+    this.history = this.history.slice(-maxMessages);
+  }
+
+  private async runAgentLoop(
+    messages: Array<OllamaChatMessage | OllamaToolResultMessage>,
+    options?: GenerateReplyOptions,
+  ): Promise<string> {
+    let progressEmitted = false;
+
+    for (let step = 0; step < 6; step += 1) {
+      const response = await this.chat(messages, { enableTools: true });
+      const toolCalls = response.toolCalls ?? [];
+
+      messages.push({
+        role: "assistant",
+        content: response.content,
+        tool_calls: toolCalls.length > 0 ? toolCalls : undefined,
+      });
+
+      if (toolCalls.length === 0) {
+        return response.content;
+      }
+
+      for (const call of toolCalls) {
+        if (!progressEmitted) {
+          const progressMessage = this.getProgressMessage(call.function.name);
+          if (progressMessage) {
+            options?.onProgress?.(progressMessage);
+            progressEmitted = true;
+          }
+        }
+        const result = await this.executeTool(call);
+        this.logger.info("LLM tool call", {
+          name: call.function.name,
+          arguments: call.function.arguments,
+          result,
+        });
+        messages.push({
+          role: "tool",
+          tool_name: call.function.name,
+          content: result,
+        });
+      }
+    }
+
+    throw new Error("도구 호출 루프가 제한 횟수를 넘었습니다.");
+  }
+
+  private async chat(
+    messages: Array<OllamaChatMessage | OllamaToolResultMessage>,
+    options?: { enableTools: boolean },
+  ): Promise<{ content: string; toolCalls: OllamaToolCall[] }> {
+    const response = await fetch(`${this.config.OLLAMA_BASE_URL}/api/chat`, {
       method: "POST",
       headers: {
-        "Content-Type": "application/json",
+        "content-type": "application/json",
       },
       body: JSON.stringify({
         model: this.config.OLLAMA_MODEL,
-        messages: [
-          {
-            role: "system",
-            content: ASSISTANT_INSTRUCTIONS,
-          },
-          {
-            role: "user",
-            content: memory.buildPrompt(utterance),
-          },
-        ],
-        think: false,
+        messages,
+        tools: options?.enableTools ? TOOL_DEFINITIONS : undefined,
         stream: false,
+        think: false,
         keep_alive: this.config.OLLAMA_KEEP_ALIVE,
-        options: {
-          num_ctx: this.config.OLLAMA_NUM_CTX,
-          temperature: 0.4,
-          num_predict: 120,
-        },
       }),
     });
 
-    const body = (await response.json().catch(() => ({}))) as OllamaChatResponse;
-
     if (!response.ok) {
-      throw new Error(body.error ?? `Ollama request failed with status ${response.status}`);
+      const body = await response.text();
+      throw new Error(`Ollama API ${response.status}: ${body}`);
     }
 
-    const output = body.message?.content?.trim();
-    if (!output) {
-      return "잘 못 들었습니다. 한 번만 다시 말씀해 주세요.";
+    const payload = (await response.json()) as OllamaChatResponse;
+    const content = payload.message?.content?.trim() ?? "";
+    const toolCalls = payload.message?.tool_calls ?? [];
+
+    if (!content && toolCalls.length === 0) {
+      throw new Error("Ollama 응답에 message.content 와 tool_calls 가 모두 없습니다.");
     }
 
-    return normalizeReply(output);
+    return {
+      content,
+      toolCalls,
+    };
+  }
+
+  private async executeTool(call: OllamaToolCall): Promise<string> {
+    switch (call.function.name) {
+      case "get_current_time":
+        return JSON.stringify(this.getCurrentTime());
+      case "get_runtime_settings":
+        return JSON.stringify(this.getRuntimeSettings());
+      case "list_project_commands":
+        return JSON.stringify(this.listProjectCommands());
+      case "evaluate_math":
+        return JSON.stringify({
+          expression: this.getStringArg(call.function.arguments, "expression"),
+          result: this.evaluateMath(this.getStringArg(call.function.arguments, "expression")),
+        });
+      case "web_search":
+        return JSON.stringify(
+          await webSearch(
+            this.getStringArg(call.function.arguments, "query"),
+            Math.min(5, Math.max(1, Math.trunc(this.getNumberArg(call.function.arguments, "max_results", 4)))),
+          ),
+        );
+      case "fetch_url":
+        return JSON.stringify(
+          await webFetch(
+            this.getStringArg(call.function.arguments, "url"),
+            Math.min(10000, Math.max(1000, Math.trunc(this.getNumberArg(call.function.arguments, "max_chars", 6000)))),
+          ),
+        );
+      default:
+        return JSON.stringify({
+          error: `unknown tool: ${call.function.name}`,
+        });
+    }
+  }
+
+  private getCurrentTime(): { timezone: string; iso: string; local: string } {
+    const now = new Date();
+    return {
+      timezone: "Asia/Seoul",
+      iso: now.toISOString(),
+      local: new Intl.DateTimeFormat("ko-KR", {
+        timeZone: "Asia/Seoul",
+        dateStyle: "full",
+        timeStyle: "long",
+      }).format(now),
+    };
+  }
+
+  private getRuntimeSettings(): Record<string, unknown> {
+    return {
+      ollama_base_url: this.config.OLLAMA_BASE_URL,
+      ollama_model: this.config.OLLAMA_MODEL,
+      ollama_keep_alive: this.config.OLLAMA_KEEP_ALIVE,
+      max_conversation_turns: this.config.MAX_CONVERSATION_TURNS,
+      whisper_model: this.config.WHISPER_MODEL,
+      whisper_language: this.config.WHISPER_LANGUAGE,
+      whisper_device: this.config.WHISPER_DEVICE,
+      whisper_compute_type: this.config.WHISPER_COMPUTE_TYPE,
+      whisper_beam_size: this.config.WHISPER_BEAM_SIZE,
+      audio_source: this.config.AUDIO_SOURCE ?? null,
+      debug: this.config.DEBUG,
+    };
+  }
+
+  private listProjectCommands(): { commands: string[] } {
+    return {
+      commands: [
+        "bun run setup",
+        "bun run setup:stt",
+        "bun run setup:llm",
+        "bun run setup:tts",
+        "bun run setup:all",
+        "bun run devices",
+        "bun run test:stt",
+        "bun run test:sttllm",
+        "bun run test:all",
+        "bun run test:llm",
+        "bun run test:tts -- \"안녕하세요\"",
+      ],
+    };
+  }
+
+  private getStringArg(args: Record<string, unknown>, name: string): string {
+    const value = args[name];
+    if (typeof value !== "string" || value.trim().length === 0) {
+      throw new Error(`도구 인자 ${name} 가 비어 있습니다.`);
+    }
+    return value.trim();
+  }
+
+  private evaluateMath(expression: string): number {
+    if (!/^[0-9+\-*/%().\s]+$/.test(expression)) {
+      throw new Error("허용되지 않은 문자가 포함된 산술식입니다.");
+    }
+
+    const result = Function(`"use strict"; return (${expression});`)();
+    if (typeof result !== "number" || !Number.isFinite(result)) {
+      throw new Error("산술식 계산 결과가 유효하지 않습니다.");
+    }
+    return result;
+  }
+
+  private getNumberArg(args: Record<string, unknown>, name: string, fallback: number): number {
+    const value = args[name];
+    if (typeof value === "number" && Number.isFinite(value)) {
+      return value;
+    }
+    if (typeof value === "string") {
+      const parsed = Number(value);
+      if (Number.isFinite(parsed)) {
+        return parsed;
+      }
+    }
+    return fallback;
+  }
+
+  private async repairReplyLanguageIfNeeded(reply: string, userText: string): Promise<string> {
+    if (!this.needsLanguageRepair(reply)) {
+      return reply;
+    }
+
+    this.logger.warn("Reply language repair triggered", {
+      reply,
+      analysis: this.analyzeScriptUsage(reply),
+    });
+
+    const repaired = await this.chat(
+      [
+        {
+          role: "system",
+          content: REWRITE_KOREAN_PROMPT,
+        },
+        {
+          role: "user",
+          content: `원문 질문: ${userText}\n기존 답변: ${reply}`,
+        },
+      ],
+      { enableTools: false },
+    );
+
+    const normalized = repaired.content.trim();
+    if (!normalized) {
+      return reply;
+    }
+
+    return normalized;
+  }
+
+  private needsLanguageRepair(text: string): boolean {
+    const analysis = this.analyzeScriptUsage(text);
+    if (analysis.otherLetters > 0) {
+      return true;
+    }
+    if (analysis.hangul === 0 && analysis.latin > 0) {
+      return true;
+    }
+    return false;
+  }
+
+  private analyzeScriptUsage(text: string): { hangul: number; latin: number; otherLetters: number } {
+    let hangul = 0;
+    let latin = 0;
+    let otherLetters = 0;
+
+    for (const char of text) {
+      if (!/\p{Letter}/u.test(char)) {
+        continue;
+      }
+      if (/\p{Script=Hangul}/u.test(char)) {
+        hangul += 1;
+        continue;
+      }
+      if (/\p{Script=Latin}/u.test(char)) {
+        latin += 1;
+        continue;
+      }
+      otherLetters += 1;
+    }
+
+    return { hangul, latin, otherLetters };
+  }
+
+  private getProgressMessage(toolName: string): string | null {
+    switch (toolName) {
+      case "web_search":
+      case "fetch_url":
+        return "검색해볼게요.";
+      default:
+        return null;
+    }
+  }
+
+  private parseAssessment(content: string): ReplyAssessment | null {
+    const match = content.match(/\{[\s\S]*\}/);
+    if (!match) {
+      return null;
+    }
+
+    try {
+      const parsed = JSON.parse(match[0]) as Record<string, unknown>;
+      return {
+        shouldReply: parsed.should_reply === true || parsed.shouldReply === true,
+        likelyNeedsLookup: parsed.likely_needs_lookup === true || parsed.likelyNeedsLookup === true,
+        reason: typeof parsed.reason === "string" ? parsed.reason : "parsed",
+      };
+    } catch {
+      return null;
+    }
+  }
+
+  private assessReplyNeedHeuristically(userText: string): ReplyAssessment | null {
+    const normalized = userText.trim();
+
+    if (!normalized) {
+      return {
+        shouldReply: false,
+        likelyNeedsLookup: false,
+        reason: "empty",
+      };
+    }
+
+    if (/^(아+|어+|음+|으+|흠+|엉+|어어+|음음+|하+|호+|와+|오+|응+|네+|예+|끝\.?)$/u.test(normalized)) {
+      return {
+        shouldReply: false,
+        likelyNeedsLookup: false,
+        reason: "filler",
+      };
+    }
+
+    if (normalized.length <= 2 && !/[?？]/.test(normalized)) {
+      return {
+        shouldReply: false,
+        likelyNeedsLookup: false,
+        reason: "too_short",
+      };
+    }
+
+    return null;
+  }
+
+  private mightNeedLookup(text: string): boolean {
+    return /(최신|오늘|최근|뉴스|검색|찾아|알아봐|확인|업데이트|가격|날씨|현재|실시간)/u.test(text);
   }
 }

src/services/python-json-worker.ts

@@ -1,189 +1,184 @@
 import { spawn, type ChildProcessWithoutNullStreams } from "node:child_process";
+import { once } from "node:events";
 import { createInterface } from "node:readline";
-import path from "node:path";
 
-import type { AssistantRuntimeConfig } from "../config.js";
+import type { AppConfig } from "../config.js";
 import type { Logger } from "../logger.js";
-import { resolveLocalAiCachePath, resolvePythonLaunch } from "../python-runtime.js";
+import { buildPythonInvocation, resolveWorkerPythonCommand, resolveWorkerScript } from "../python-runtime.js";
 
-interface WorkerRequest {
-  id: number;
-  method: string;
-  params: Record<string, unknown>;
+interface RpcSuccess<T> {
+  id: string;
+  result: T;
 }
 
-interface WorkerResponse {
-  id: number;
-  ok: boolean;
-  result?: unknown;
-  error?: string;
+interface RpcFailure {
+  id: string;
+  error: string;
+}
+
+type RpcResponse<T> = RpcSuccess<T> | RpcFailure;
+
+function isFailure<T>(value: RpcResponse<T>): value is RpcFailure {
+  return "error" in value;
 }
 
 export class PythonJsonWorker {
-  private child: ChildProcessWithoutNullStreams | null = null;
-  private nextId = 1;
+  private processRef: ChildProcessWithoutNullStreams | null = null;
+  private shuttingDown = false;
   private readonly pending = new Map<
-    number,
+    string,
     {
       resolve: (value: unknown) => void;
-      reject: (error: Error) => void;
+      reject: (reason?: unknown) => void;
     }
   >();
+  private nextId = 1;
 
   constructor(
-    private readonly config: AssistantRuntimeConfig,
+    private readonly config: AppConfig,
     private readonly logger: Logger,
     private readonly scriptName: string,
-    private readonly label: string,
-    private readonly workerEnv: Record<string, string>,
+    private readonly logPrefix: string,
   ) {}
 
-  async request<T>(method: string, params: Record<string, unknown>, signal?: AbortSignal): Promise<T> {
-    const child = this.ensureStarted();
-    const id = this.nextId++;
-
-    return await new Promise<T>((resolve, reject) => {
-      if (signal?.aborted) {
-        reject(new Error(`${this.label} request aborted before start`));
+  async start(): Promise<void> {
+    if (this.processRef) {
       return;
     }
-
-      const abortHandler = () => {
-        this.pending.delete(id);
-        reject(new Error(`${this.label} request aborted`));
-      };
-
-      if (signal) {
-        signal.addEventListener("abort", abortHandler, { once: true });
+    if (this.shuttingDown) {
+      throw new Error(`${this.logPrefix} worker is shutting down`);
     }
 
-      this.pending.set(id, {
-        resolve: (value) => {
-          if (signal) {
-            signal.removeEventListener("abort", abortHandler);
-          }
-          resolve(value as T);
-        },
-        reject: (error) => {
-          if (signal) {
-            signal.removeEventListener("abort", abortHandler);
-          }
-          reject(error);
+    const python = await resolveWorkerPythonCommand(this.config);
+    const scriptPath = resolveWorkerScript(this.scriptName);
+    const invocation = buildPythonInvocation(python, [scriptPath]);
+
+    this.processRef = spawn(invocation.command, invocation.args, {
+      stdio: ["pipe", "pipe", "pipe"],
+      windowsHide: true,
+      env: {
+        ...process.env,
+        WHISPER_MODEL: this.config.WHISPER_MODEL,
+        WHISPER_LANGUAGE: this.config.WHISPER_LANGUAGE,
+        WHISPER_DEVICE: this.config.WHISPER_DEVICE,
+        WHISPER_COMPUTE_TYPE: this.config.WHISPER_COMPUTE_TYPE,
+        WHISPER_BEAM_SIZE: String(this.config.WHISPER_BEAM_SIZE),
       },
     });
 
-      const message: WorkerRequest = {
+    const rl = createInterface({
+      input: this.processRef.stdout,
+      crlfDelay: Infinity,
+    });
+
+    rl.on("line", (line) => {
+      this.handleStdoutLine(line);
+    });
+
+    this.processRef.stderr.on("data", (chunk: Buffer) => {
+      const text = chunk.toString().trim();
+      if (text.length > 0) {
+        this.logger.warn(`[${this.logPrefix}] ${text}`);
+      }
+    });
+
+    this.processRef.stdin.on("error", (error) => {
+      this.logger.debug(`${this.logPrefix} stdin error`, error);
+    });
+
+    this.processRef.on("exit", (code, signal) => {
+      const error = new Error(`${this.logPrefix} worker exited code=${code ?? "null"} signal=${signal ?? "null"}`);
+      for (const entry of this.pending.values()) {
+        entry.reject(error);
+      }
+      this.pending.clear();
+      this.processRef = null;
+    });
+  }
+
+  async request<T>(method: string, params: Record<string, unknown>): Promise<T> {
+    await this.start();
+
+    if (!this.processRef) {
+      throw new Error(`${this.logPrefix} worker is not running`);
+    }
+
+    const id = String(this.nextId++);
+    const payload = JSON.stringify({
       id,
       method,
       params,
-      };
-
-      child.stdin.write(`${JSON.stringify(message)}\n`);
     });
+
+    const promise = new Promise<T>((resolve, reject) => {
+      this.pending.set(id, {
+        resolve: (value) => resolve(value as T),
+        reject,
+      });
+    });
+
+    this.processRef.stdin.write(`${payload}\n`);
+    return await promise;
   }
 
   async destroy(): Promise<void> {
-    this.rejectAll(new Error(`${this.label} worker terminated`));
-
-    if (!this.child) {
+    if (!this.processRef) {
       return;
     }
 
-    const child = this.child;
-    this.child = null;
+    const child = this.processRef;
+    this.shuttingDown = true;
 
-    child.kill("SIGTERM");
-    await new Promise<void>((resolve) => {
-      child.once("exit", () => resolve());
-      setTimeout(resolve, 1_500);
-    });
-  }
-
-  private ensureStarted(): ChildProcessWithoutNullStreams {
-    if (this.child) {
-      return this.child;
-    }
-
-    const launch = resolvePythonLaunch(this.config);
-    const scriptPath = path.resolve(process.cwd(), "python", this.scriptName);
-    const cachePath = resolveLocalAiCachePath(this.config);
-    const recentStderr: string[] = [];
-
-    const child = spawn(launch.command, [...launch.args, scriptPath], {
-      stdio: ["pipe", "pipe", "pipe"],
-      env: {
-        ...process.env,
-        HF_HOME: cachePath,
-        TRANSFORMERS_CACHE: cachePath,
-        PYTHONIOENCODING: "utf-8",
-        BOT_DEFAULT_LANGUAGE: this.config.BOT_DEFAULT_LANGUAGE,
-        ...this.workerEnv,
-      },
-    });
-
-    createInterface({
-      input: child.stdout,
-      crlfDelay: Number.POSITIVE_INFINITY,
-    }).on("line", (line) => {
-      if (!line.trim()) {
-        return;
-      }
-
-      let payload: WorkerResponse;
     try {
-        payload = JSON.parse(line) as WorkerResponse;
-      } catch (error) {
-        this.logger.warn(`${this.label} stdout parse failed`, error);
+      child.stdin.end();
+    } catch {
+      // ignore
+    }
+
+    if (!child.killed && child.exitCode === null) {
+      child.kill("SIGTERM");
+    }
+
+    const timedWait = Promise.race([
+      once(child, "exit"),
+      new Promise<null>((resolve) => setTimeout(() => resolve(null), 1500)),
+    ]);
+
+    await timedWait;
+
+    if (child.exitCode === null && !child.killed) {
+      child.kill("SIGKILL");
+      await once(child, "exit").catch(() => null);
+    }
+
+    this.processRef = null;
+    this.shuttingDown = false;
+  }
+
+  private handleStdoutLine(line: string): void {
+    const trimmed = line.trim();
+    if (!trimmed) {
       return;
     }
 
-      const pending = this.pending.get(payload.id);
+    let message: RpcResponse<unknown>;
+    try {
+      message = JSON.parse(trimmed) as RpcResponse<unknown>;
+    } catch (error) {
+      this.logger.warn(`${this.logPrefix} stdout parse failed`, error);
+      return;
+    }
+
+    const pending = this.pending.get(message.id);
     if (!pending) {
       return;
     }
 
-      this.pending.delete(payload.id);
-      if (payload.ok) {
-        pending.resolve(payload.result);
+    this.pending.delete(message.id);
+    if (isFailure(message)) {
+      pending.reject(new Error(message.error));
       return;
     }
-
-      pending.reject(new Error(payload.error ?? `${this.label} worker error`));
-    });
-
-    child.stderr.on("data", (chunk: Buffer) => {
-      const text = chunk.toString().trim();
-      if (text.length > 0) {
-        recentStderr.push(text);
-        if (recentStderr.length > 20) {
-          recentStderr.shift();
-        }
-        this.logger.warn(`[${this.label}]`, text);
-      }
-    });
-
-    child.on("exit", (code, signal) => {
-      if (this.child === child) {
-        this.child = null;
-      }
-
-      const detail = recentStderr.length > 0 ? `\n${recentStderr.join("\n")}` : "";
-      this.rejectAll(new Error(`${this.label} worker exited code=${code ?? "null"} signal=${signal ?? "null"}${detail}`));
-    });
-
-    child.on("error", (error) => {
-      this.rejectAll(error as Error);
-    });
-
-    this.child = child;
-    return child;
-  }
-
-  private rejectAll(error: Error): void {
-    const pending = [...this.pending.values()];
-    this.pending.clear();
-    for (const item of pending) {
-      item.reject(error);
-    }
+    pending.resolve(message.result);
   }
 }

src/services/stt.ts

@@ -1,4 +0,0 @@
-export interface SttService {
-  transcribePcm16(pcm16MonoAudio: Buffer): Promise<string | null>;
-  destroy?(): Promise<void>;
-}

src/services/tts.ts

@@ -1,11 +0,0 @@
-import type { Readable } from "node:stream";
-
-export interface PreparedSpeechAudio {
-  stream: Readable;
-  dispose: () => void;
-}
-
-export interface TtsService {
-  preparePlayback(text: string, signal?: AbortSignal): Promise<PreparedSpeechAudio>;
-  destroy?(): Promise<void>;
-}

src/services/web-tools.ts

@@ -0,0 +1,105 @@
+interface WebSearchResult {
+  title: string;
+  url: string;
+  snippet: string;
+}
+
+interface WebFetchResult {
+  url: string;
+  title: string;
+  content: string;
+}
+
+function stripTags(html: string): string {
+  return html
+    .replace(/<script[\s\S]*?<\/script>/gi, " ")
+    .replace(/<style[\s\S]*?<\/style>/gi, " ")
+    .replace(/<noscript[\s\S]*?<\/noscript>/gi, " ")
+    .replace(/<[^>]+>/g, " ");
+}
+
+function decodeEntities(text: string): string {
+  return text
+    .replace(/&amp;/g, "&")
+    .replace(/&lt;/g, "<")
+    .replace(/&gt;/g, ">")
+    .replace(/&quot;/g, '"')
+    .replace(/&#39;/g, "'")
+    .replace(/&nbsp;/g, " ");
+}
+
+function normalizeWhitespace(text: string): string {
+  return decodeEntities(text).replace(/\s+/g, " ").trim();
+}
+
+function extractTitle(html: string): string {
+  const match = html.match(/<title[^>]*>([\s\S]*?)<\/title>/i);
+  return normalizeWhitespace(match?.[1] ?? "");
+}
+
+function extractSearchResults(html: string, maxResults: number): WebSearchResult[] {
+  const results: WebSearchResult[] = [];
+  const pattern =
+    /<a[^>]*class="result__a"[^>]*href="([^"]+)"[^>]*>([\s\S]*?)<\/a>[\s\S]*?(?:<a[^>]*class="result__snippet"[^>]*>|<div[^>]*class="result__snippet"[^>]*>)([\s\S]*?)(?:<\/a>|<\/div>)/gi;
+
+  for (const match of html.matchAll(pattern)) {
+    const url = match[1]?.trim();
+    const title = normalizeWhitespace(stripTags(match[2] ?? ""));
+    const snippet = normalizeWhitespace(stripTags(match[3] ?? ""));
+    if (!url || !title) {
+      continue;
+    }
+    results.push({ title, url, snippet });
+    if (results.length >= maxResults) {
+      break;
+    }
+  }
+
+  return results;
+}
+
+export async function webSearch(query: string, maxResults = 5): Promise<{ query: string; results: WebSearchResult[] }> {
+  const url = `https://duckduckgo.com/html/?q=${encodeURIComponent(query)}`;
+  const response = await fetch(url, {
+    headers: {
+      "user-agent":
+        "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/136.0.0.0 Safari/537.36",
+    },
+  });
+
+  if (!response.ok) {
+    throw new Error(`web search failed: ${response.status}`);
+  }
+
+  const html = await response.text();
+  const results = extractSearchResults(html, Math.min(Math.max(maxResults, 1), 8));
+  return { query, results };
+}
+
+export async function webFetch(url: string, maxChars = 6000): Promise<WebFetchResult> {
+  if (!/^https?:\/\//i.test(url)) {
+    throw new Error("http 또는 https URL만 허용됩니다.");
+  }
+
+  const response = await fetch(url, {
+    headers: {
+      "user-agent":
+        "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/136.0.0.0 Safari/537.36",
+    },
+    redirect: "follow",
+  });
+
+  if (!response.ok) {
+    throw new Error(`web fetch failed: ${response.status}`);
+  }
+
+  const html = await response.text();
+  const title = extractTitle(html);
+  const content = normalizeWhitespace(stripTags(html)).slice(0, Math.max(500, maxChars));
+
+  return {
+    url,
+    title,
+    content,
+  };
+}

src/setup-llm.ts

@@ -0,0 +1,38 @@
+import process from "node:process";
+import { spawn } from "node:child_process";
+
+import { loadConfig } from "./config.js";
+
+async function run(command: string, args: string[]): Promise<void> {
+  await new Promise<void>((resolve, reject) => {
+    const child = spawn(command, args, {
+      stdio: "inherit",
+      windowsHide: true,
+      shell: process.platform === "win32",
+    });
+
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve();
+        return;
+      }
+      reject(new Error(`${command} ${args.join(" ")} exited with code ${code ?? "null"}`));
+    });
+
+    child.on("error", reject);
+  });
+}
+
+export async function setupLlm(): Promise<void> {
+  const config = loadConfig();
+  console.log(`Ollama 모델 준비: ${config.OLLAMA_MODEL}`);
+  await run("ollama", ["pull", config.OLLAMA_MODEL]);
+  console.log("Ollama LLM 환경 준비 완료");
+}
+
+if (import.meta.main) {
+  void setupLlm().catch((error) => {
+    console.error(error instanceof Error ? error.message : String(error));
+    process.exit(1);
+  });
+}

src/setup-local-ai.ts

@@ -1,88 +0,0 @@
-import { existsSync } from "node:fs";
-import { mkdir } from "node:fs/promises";
-import { spawn } from "node:child_process";
-import path from "node:path";
-
-import { loadConfig } from "./config.js";
-import { resolveLocalAiCachePath, resolveLocalAiVenvPath, resolvePythonLaunch, resolveVenvPythonPath } from "./python-runtime.js";
-
-async function run(command: string, args: string[], extraEnv?: NodeJS.ProcessEnv): Promise<void> {
-  await new Promise<void>((resolve, reject) => {
-    const child = spawn(command, args, {
-      stdio: "inherit",
-      env: {
-        ...process.env,
-        ...extraEnv,
-      },
-    });
-
-    child.on("exit", (code) => {
-      if (code === 0) {
-        resolve();
-        return;
-      }
-      reject(new Error(`${command} ${args.join(" ")} exited with code ${code ?? "null"}`));
-    });
-    child.on("error", reject);
-  });
-}
-
-async function ensurePip(pythonBin: string, env: NodeJS.ProcessEnv): Promise<void> {
-  await new Promise<void>((resolve, reject) => {
-    const child = spawn(pythonBin, ["-m", "pip", "--version"], {
-      stdio: "ignore",
-      env,
-    });
-    child.on("exit", (code) => {
-      if (code === 0) {
-        resolve();
-        return;
-      }
-      reject(new Error("pip missing"));
-    });
-    child.on("error", reject);
-  }).catch(async () => {
-    await run(pythonBin, ["-m", "ensurepip", "--upgrade"], env);
-  });
-}
-
-async function main(): Promise<void> {
-  const config = loadConfig();
-  const venvPath = resolveLocalAiVenvPath(config);
-  const venvPython = resolveVenvPythonPath(config);
-  const cachePath = resolveLocalAiCachePath(config);
-  const requirementsPath = path.resolve(process.cwd(), "python", "requirements.txt");
-  const baseEnv = {
-    HF_HOME: cachePath,
-    TRANSFORMERS_CACHE: cachePath,
-    PYTHONIOENCODING: "utf-8",
-  };
-
-  await mkdir(cachePath, { recursive: true });
-
-  if (!existsSync(venvPython)) {
-    const launch = resolvePythonLaunch(config, { preferVenv: false });
-    console.log(`기본 Python 확인: ${launch.command} ${launch.args.join(" ")}`.trim());
-    console.log(`가상환경 생성: ${venvPath}`);
-    await run(launch.command, [...launch.args, "-m", "venv", venvPath], baseEnv);
-  }
-
-  await ensurePip(venvPython, {
-    ...process.env,
-    ...baseEnv,
-  });
-
-  console.log("로컬 AI 의존성 설치를 시작합니다.");
-  await run(venvPython, ["-m", "pip", "install", "--upgrade", "pip", "setuptools", "wheel"], baseEnv);
-  await run(venvPython, ["-m", "pip", "install", "-r", requirementsPath], baseEnv);
-
-  console.log("설치가 끝났습니다.");
-  console.log("다음 순서:");
-  console.log("1. bun run devices");
-  console.log("2. bun run start:local");
-}
-
-void main().catch((error) => {
-  console.error(error instanceof Error ? error.message : String(error));
-  process.exit(1);
-});

src/setup-python.ts

@@ -0,0 +1,55 @@
+import process from "node:process";
+import { mkdir } from "node:fs/promises";
+import path from "node:path";
+import { spawn } from "node:child_process";
+
+import { loadConfig } from "./config.js";
+import { buildPythonInvocation, resolveBasePythonCommand, resolveVenvPythonPath } from "./python-runtime.js";
+
+async function run(command: string, args: string[], cwd: string): Promise<void> {
+  await new Promise<void>((resolve, reject) => {
+    const child = spawn(command, args, {
+      cwd,
+      stdio: "inherit",
+      windowsHide: true,
+    });
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve();
+        return;
+      }
+      reject(new Error(`${command} ${args.join(" ")} exited with code ${code ?? "null"}`));
+    });
+    child.on("error", reject);
+  });
+}
+
+export async function setupSttPython(): Promise<void> {
+  const config = loadConfig();
+  const python = await resolveBasePythonCommand(config);
+  const venvRoot = path.resolve(process.cwd(), config.LOCAL_AI_VENV_PATH);
+  const requirementsPath = path.resolve(
+    process.cwd(),
+    "python",
+    process.platform === "win32" ? "requirements-windows.txt" : "requirements.txt",
+  );
+
+  await mkdir(path.dirname(venvRoot), { recursive: true });
+
+  console.log(`가상환경 생성: ${venvRoot}`);
+  const createVenv = buildPythonInvocation(python, ["-m", "venv", venvRoot]);
+  await run(createVenv.command, createVenv.args, process.cwd());
+
+  const venvPython = resolveVenvPythonPath(config);
+  await run(venvPython, ["-m", "pip", "install", "--upgrade", "pip", "setuptools", "wheel"], process.cwd());
+  await run(venvPython, ["-m", "pip", "install", "-r", requirementsPath], process.cwd());
+
+  console.log("Python STT 환경 준비 완료");
+}
+
+if (import.meta.main) {
+  void setupSttPython().catch((error) => {
+    console.error(error instanceof Error ? error.message : String(error));
+    process.exit(1);
+  });
+}

src/setup-tts.ts

@@ -0,0 +1,76 @@
+import process from "node:process";
+import { mkdir } from "node:fs/promises";
+import path from "node:path";
+import { spawn } from "node:child_process";
+
+import { loadConfig } from "./config.js";
+import { resolveDockerCommand } from "./docker-runtime.js";
+import { Logger } from "./logger.js";
+import { MeloTtsService } from "./services/melo-tts.js";
+
+async function run(command: string, args: string[], cwd = process.cwd()): Promise<void> {
+  const env = { ...process.env };
+  if (path.isAbsolute(command)) {
+    const dockerBinDir = path.dirname(command);
+    const currentPath = env.PATH ?? env.Path ?? "";
+    env.PATH = `${dockerBinDir}${path.delimiter}${currentPath}`;
+  }
+
+  await new Promise<void>((resolve, reject) => {
+    const child = spawn(command, args, {
+      cwd,
+      stdio: "inherit",
+      windowsHide: true,
+      shell: process.platform === "win32" && !path.isAbsolute(command),
+      env,
+    });
+
+    child.on("error", (error) => {
+      if ((error as NodeJS.ErrnoException).code === "ENOENT" && command === "docker") {
+        reject(new Error("Docker를 찾지 못했습니다. Docker Desktop을 설치하고 실행한 뒤 다시 시도하세요."));
+        return;
+      }
+      reject(error);
+    });
+    child.on("exit", (code) => {
+      if (code === 0) {
+        resolve();
+        return;
+      }
+      reject(new Error(`${command} ${args.join(" ")} exited with code ${code ?? "null"}`));
+    });
+  });
+}
+
+export async function setupTts(): Promise<void> {
+  const config = loadConfig();
+  const logger = new Logger(config.DEBUG ? config.LOG_LEVEL : "error");
+  const docker = await resolveDockerCommand(config);
+  const dockerContext = path.resolve(process.cwd(), "docker", "melotts");
+  const cacheDir = path.resolve(process.cwd(), config.TTS_CACHE_DIR);
+  const outputDir = path.resolve(process.cwd(), config.TTS_OUTPUT_DIR);
+
+  await mkdir(cacheDir, { recursive: true });
+  await mkdir(outputDir, { recursive: true });
+
+  console.log(`MeloTTS Docker 이미지 빌드: ${config.TTS_IMAGE}`);
+  await run(docker, ["build", "-t", config.TTS_IMAGE, dockerContext]);
+
+  const tts = new MeloTtsService(config, logger);
+
+  console.log("MeloTTS 모델 워밍업...");
+  try {
+    await tts.warmup();
+  } finally {
+    await tts.destroy().catch(() => undefined);
+  }
+
+  console.log("로컬 TTS 환경 준비 완료");
+}
+
+if (import.meta.main) {
+  void setupTts().catch((error) => {
+    console.error(error instanceof Error ? error.message : String(error));
+    process.exit(1);
+  });
+}

src/setup.ts

@@ -0,0 +1,18 @@
+import process from "node:process";
+
+import { setupLlm } from "./setup-llm.js";
+import { setupSttPython } from "./setup-python.js";
+import { setupTts } from "./setup-tts.js";
+
+async function main(): Promise<void> {
+  await setupSttPython();
+  await setupLlm();
+  await setupTts();
+}
+
+if (import.meta.main) {
+  void main().catch((error) => {
+    console.error(error instanceof Error ? error.message : String(error));
+    process.exit(1);
+  });
+}