DeVora Voice Agent UI (Web)

DeVora = UI suara untuk coding agent. OpenCode hanya core agent di belakang; UI, mic, TTS, dan pipeline voice semuanya milik DeVora.

⚠️ Keamanan: jangan pernah commit opencode.json / opencode.server.json (berisi API key). Sudah di .gitignore. Contoh config ada di opencode.example.json.

Arsitektur

Browser (React web app: mic, playback, voice screen)
   │  WebSocket /ws (audio PCM16 16k + events)
   ▼
DeVora Server (packages/server Bun)
   │  VoiceController: VAD → STT → narration → piper TTS
   │  @opencode-ai/sdk
   ▼
opencode serve (core agent)  ── LLM provider (9router / openrouter / dll)

Requirements

• Linux + systemd (server), sudo untuk hotspot setup
• bun ≥ 1.x
• opencode CLI
• piper + voice id_ID-news_tts-medium
• Python 3 + faster-whisper (STT streaming) pip install faster-whisper
• LLM gateway (opsional, contoh 9router di 127.0.0.1:20128)

Download Model

Model	Untuk	Ukuran	Download
piper id_ID-news_tts-medium	TTS suara Indonesia	~60MB	https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/id/id_ID/news_tts-medium/id_ID-news_tts-medium.onnx (+ .onnx.json)
faster-whisper-base	STT streaming (id, real-time)	~145MB	https://huggingface.co/Systran/faster-whisper-base (HuggingFace Hub)
whisper base.pt (opsional)	STT batch fallback	~140MB	https://openaipublic.azureedge.net/main/whisper/models/ed3a0b6b1c0edf879ad9b11b1af5a0e6ab5db9205f891f668f8b0e6c6326e34e/base.pt

Lokasi default (models/ di repo):

mkdir -p models/id models/faster-whisper-base

# piper voice id
curl -L "https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/id/id_ID/news_tts-medium/id_ID-news_tts-medium.onnx" -o ~/.local/share/piper/models/id/id_ID-news_tts-medium.onnx
curl -L "https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/id/id_ID/news_tts-medium/id_ID-news_tts-medium.onnx.json" -o ~/.local/share/piper/models/id/id_ID-news_tts-medium.onnx.json

# faster-whisper (streaming STT)
python -m pip install faster-whisper huggingface_hub
python -c "from huggingface_hub import snapshot_download; snapshot_download('Systran/faster-whisper-base', local_dir='models/faster-whisper-base')"

Menjalankan

Server

cd packages/server
PROJECT=$HOME/DeVora bun run src/index.ts

• HTTP + WS: port 3000 (http://127.0.0.1:3000, /ws)
• Otomatis spawn opencode serve (port acak internal)
• Butuh: piper + voice id (PATH), model STT di models/, .venv (python), opencode di PATH, LLM gateway (9router 127.0.0.1:20128)

Env opsional:

Var	Default	Fungsi
PROJECT	cwd	project directory opencode
PORT	3000	port server
DEVORA_PIPER_MODEL/CONFIG	~/.local/share/piper/models/id/...	model piper
DEVORA_STT_MODEL	models/faster-whisper-base	model STT
WHISPER_LANG	id	bahasa STT
STT_TRAILING_MS	2000	window partial STT
DEVORA_TLS_CERT/KEY		aktifkan HTTPS langsung

Web frontend (dev)

cd apps/web
bun install
bunx --bun vite --host 0.0.0.0
# buka http://<host>:5173 proxy /ws ke 127.0.0.1:3000

Build produksi

cd apps/web && bunx --bun vite build   # → dist/ (diserve server)

Akses dari luar (VPS firewall)

# SSH tunnel (cepat, sudah teruji):
ssh -f -N -L 3000:127.0.0.1:3000 anlap05
# buka http://127.0.0.1:3000

Protocol WS (client → server)

message	field	fungsi
start		mulai listening (VAD aktif)
stop		berhenti listening
audio	data (base64 PCM16 16k mono)	kirim chunk mic
interrupt		barge-in / stop agent
submit	text	keyboard fallback
create_session	directory, title?, model? {providerID, modelID}	buat sesi baru di project directory
select_session	id, model?	pindah ke sesi yang sudah ada
set_model	providerID, modelID	ganti model di tengah percakapan
list_dir	path?	browse folder (project picker)

Server → client: {type:"state", snapshot}, {type:"audio", wav}, {type:"error", message}, {type:"context", context} (Project/MCP/Skills/Models), {type:"sessions", sessions}, {type:"session-selected", sessionId}, {type:"dir_list", path, parent, dirs}, {type:"model", model}.

Test

cd packages/devora && bun test   # 48 test (unit + E2E mock)

Instalasi otomatis

# via package manager (setelah publish)
npm install -g devora
pnpm add -g devora
bun add -g devora
npx devora install

# atau script shell langsung
./install.sh                    # install dasar
sudo ./install.sh --hotspot     # + bind9 DNS + nginx TLS + CA (gateway hotspot)

# hotspot mode menghasilkan:
#   DNS: devora.local → 10.42.0.1 (bind9 listen gateway)
#   TLS: https://devora.local  (CA: https://devora.local/ca.crt)

Systemd service

Template: packages/installer/scripts/devora.service (ganti USER dan HOME_DIR).

# install unit
sudo cp packages/installer/scripts/devora.service /etc/systemd/system/devora.service
sudo sed -i "s/USER/$(whoami)/g; s|HOME_DIR|$HOME|g" /etc/systemd/system/devora.service
sudo systemctl daemon-reload

# manage
sudo systemctl enable devora    # auto-start saat boot
sudo systemctl start devora
sudo systemctl stop devora
sudo systemctl restart devora
sudo systemctl disable devora
systemctl status devora         # cek status + log

Log server: journalctl -u devora -f

Config LLM buat opencode.json sendiri (jangan di-commit):

{
  "model": "9router/bai/deepseek-v4-flash",
  "provider": {
    "9router": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "9Router",
      "options": {
        "baseURL": "http://127.0.0.1:20128/v1",
        "apiKey": "YOUR_KEY_HERE"
      },
      "models": {
        "bai/deepseek-v4-flash": { "name": "DeepSeek V4 Flash" }
      }
    }
  }
}

DNS + SSL (hotspot)

• DNS: bind9 listen 10.42.0.1:53 devora.local → 10.42.0.1 (client hotspot set DNS 10.42.0.1)
• TLS: CA lokal Devora Local CA + fullchain di /etc/ssl/devora/; nginx https://devora.local → 127.0.0.1:3000 (WebSocket upgrade)
• Trust CA: unduh https://devora.local/ca.crt, install di perangkat → https://devora.local tanpa warning + Web Speech API aktif (secure context)