Merge pull request #272 from jamiepine/windows-support

Windows support: CUDA detection, cross-platform justfile, clean server shutdown

CONTRIBUTING.md

@@ -33,101 +33,41 @@ Thank you for your interest in contributing to Voicebox! This document provides
 
 ### Development Setup
 
-**Using `just` (recommended):**
+Install [just](https://github.com/casey/just) (`brew install just`, `cargo install just`, or `winget install Casey.Just`), then:
-
-Install [just](https://github.com/casey/just) (`brew install just` or `cargo install just`), then:
 
 ```bash
+git clone https://github.com/YOUR_USERNAME/voicebox.git
+cd voicebox
+
 just setup   # creates venv, installs Python + JS deps
-just dev     # starts backend + desktop app in one terminal
+just dev     # starts backend + desktop app
 ```
 
+`just setup` handles everything automatically, including:
+- Creating a Python virtual environment
+- Installing Python dependencies (with CUDA PyTorch on Windows if an NVIDIA GPU is detected)
+- Installing MLX dependencies on Apple Silicon
+- Installing JavaScript dependencies
+
+`just dev` starts the backend and desktop app together. If a backend is already running (e.g. from `just dev-backend` in another terminal), it detects it and only starts the frontend.
+
 Other useful commands:
 
 ```bash
 just dev-web       # backend + web app (no Tauri/Rust build)
 just dev-backend   # backend only
+just dev-frontend  # Tauri app only (backend must be running)
 just kill          # stop all dev processes
 just clean-all     # nuke everything and start fresh
 just --list        # see all available commands
 ```
 
-**Using the Makefile:** Run `make setup` then `make dev`. See `make help` for all commands.
+> **Note:** In dev mode, the app connects to a manually-started Python server.
+> The bundled server binary is only used in production builds.
 
-**Manual setup (required for Windows):**
+#### Windows Notes
 
-1. **Fork and clone the repository**
+The justfile works natively on Windows via PowerShell. No WSL or Git Bash required. On Windows with an NVIDIA GPU, `just setup` automatically installs CUDA-enabled PyTorch for GPU acceleration.
-   ```bash
-   git clone https://github.com/YOUR_USERNAME/voicebox.git
-   cd voicebox
-   ```
-
-2. **Install JavaScript dependencies**
-   ```bash
-   bun install
-   ```
-   This installs dependencies for:
-   - `app/` - Shared React frontend
-   - `tauri/` - Tauri desktop wrapper
-   - `web/` - Web deployment wrapper
-
-3. **Set up Python backend**
-   ```bash
-   cd backend
-   
-   # Create virtual environment
-   python -m venv venv
-   
-   # Activate virtual environment
-   source venv/bin/activate  # On macOS/Linux
-   # or
-   venv\Scripts\activate  # On Windows
-   
-   # Install Python dependencies
-   pip install -r requirements.txt
-   
-   # Install MLX dependencies (Apple Silicon only - for faster inference)
-   # On Apple Silicon, this enables native Metal acceleration
-   if [[ $(uname -m) == "arm64" ]]; then
-     pip install -r requirements-mlx.txt
-   fi
-   
-   # Install Qwen3-TTS (required for voice synthesis)
-   pip install git+https://github.com/QwenLM/Qwen3-TTS.git
-   ```
-
-4. **Start development servers**
-
-   Development requires two terminals: one for the Python backend, one for the Tauri app.
-
-   **Terminal 1: Backend server** (start this first)
-   ```bash
-   cd backend
-   source venv/bin/activate  # Activate venv if not already active
-   bun run dev:server
-   # Or manually: uvicorn main:app --reload --port 17493
-   ```
-   Backend will be available at `http://localhost:17493`
-
-   **Terminal 2: Desktop app**
-   ```bash
-   bun run dev
-   ```
-   This will:
-   - Create a placeholder sidecar binary (for Tauri compilation)
-   - Start Vite dev server on port 5173
-   - Launch Tauri window pointing to localhost:5173
-   - Connect to the Python server you started in Terminal 1
-   - Enable hot reload
-
-   > **Note:** In dev mode, the app connects to your manually-started Python server.
-   > The bundled server binary is only used in production builds.
-
-   **Optional: Web app**
-   ```bash
-   bun run dev:web
-   ```
-   Web app will be available at `http://localhost:5174`
 
 ### Model Downloads
 
@@ -139,25 +79,30 @@ First-time usage will be slower due to model downloads, but subsequent runs will
 
 ### Building
 
-**Build everything (recommended):**
+**Build production app:**
+
 ```bash
-bun run build
+just build        # Build CPU server binary + Tauri installer
 ```
-This automatically:
+
-1. Builds the Python server binary (`./scripts/build-server.sh`)
+On Windows, to build with CUDA support for local testing:
-2. Builds the Tauri desktop app (`cd tauri && bun run tauri build`)
+
+```bash
+just build-local  # Build CPU + CUDA server binaries + Tauri installer
+```
+
+This builds the CPU sidecar (bundled with the app), the CUDA binary (placed in `%APPDATA%/com.voicebox.app/backends/` for runtime GPU switching), and the installable Tauri app.
 
 Creates platform-specific installers (`.dmg`, `.msi`, `.AppImage`) in `tauri/src-tauri/target/release/bundle/`.
 
-**Note:** The build process detects your platform and includes the appropriate backend (MLX for Apple Silicon, PyTorch for others).
+**Individual build targets:**
 
-**Build server binary only:**
 ```bash
-bun run build:server
+just build-server       # CPU server binary only
-# or
+just build-server-cuda  # CUDA server binary only (Windows)
-./scripts/build-server.sh
+just build-tauri        # Tauri desktop app only
+just build-web          # Web app only
 ```
-Creates platform-specific binary in `tauri/src-tauri/binaries/`
 
 **Building with local Qwen3-TTS development version:**
 
@@ -165,17 +110,10 @@ If you're actively developing or modifying the Qwen3-TTS library, set the `QWEN_
 
 ```bash
 export QWEN_TTS_PATH=~/path/to/your/Qwen3-TTS
-bun run build:server
+just build-server
 ```
 
-This makes PyInstaller use your local qwen-tts version instead of the pip-installed package. Useful when testing changes to the TTS library before they're published to PyPI or when using an editable install (`pip install -e`).
+This makes PyInstaller use your local qwen-tts version instead of the pip-installed package.
-
-**Build web app:**
-```bash
-cd web
-bun run build
-```
-Output in `web/dist/`
 
 ### Generate OpenAPI Client
 

README.md

@@ -240,13 +240,24 @@ just dev     # starts backend + desktop app
 
 Install [just](https://github.com/casey/just): `brew install just` or `cargo install just`. Run `just --list` to see all commands.
 
-Also available via Makefile: `make setup && make dev` (run `make help` for all commands).
+**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org), [Tauri Prerequisites](https://v2.tauri.app/start/prerequisites/), and [Xcode](https://developer.apple.com/xcode/) on macOS.
 
-**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org), [XCode on macOS](https://developer.apple.com/xcode/), [Tauri Prerequisites](https://v2.tauri.app/start/prerequisites/).
+### Platform Notes
 
-**Performance:** 
+| Platform | GPU Backend | Notes |
-- **Apple Silicon (M1/M2/M3)**: Uses MLX backend with native Metal acceleration for 4-5x faster inference
+|----------|-------------|-------|
-- **Windows/Linux/Intel Mac**: Uses PyTorch backend (CUDA GPU recommended, CPU supported but slower)
+| macOS (Apple Silicon) | MLX (Metal) | 4-5x faster inference via Neural Engine |
+| Windows (NVIDIA) | PyTorch (CUDA) | `just setup` auto-installs CUDA PyTorch |
+| Windows/Linux (no NVIDIA) | PyTorch (CPU) | Works but slower |
+
+### Building Locally
+
+```bash
+just build          # Build CPU server binary + Tauri app
+just build-local    # (Windows) Build CPU + CUDA server binaries + Tauri app
+```
+
+`just build-local` produces a production-ready installer with the CUDA binary pre-placed for GPU switching.
 
 ### Project Structure
 

backend/build_binary.py

@@ -10,6 +10,7 @@ import PyInstaller.__main__
 import argparse
 import os
 import platform
+import sys
 from pathlib import Path
 
 
@@ -62,6 +63,10 @@ def build_server(cuda=False):
         '--hidden-import', 'backend.utils.hf_progress',
         '--hidden-import', 'backend.utils.validation',
         '--hidden-import', 'backend.cuda_download',
+        '--hidden-import', 'backend.effects',
+        '--hidden-import', 'backend.utils.effects',
+        '--hidden-import', 'backend.versions',
+        '--hidden-import', 'pedalboard',
         '--hidden-import', 'torch',
         '--hidden-import', 'transformers',
         '--hidden-import', 'fastapi',
@@ -91,9 +96,10 @@ def build_server(cuda=False):
             '--hidden-import', 'torch.backends.cudnn',
         ])
     else:
-        # Exclude NVIDIA CUDA packages from CPU-only builds to keep binary under 4GB.
+        # Exclude NVIDIA CUDA packages from CPU-only builds to keep binary small.
-        # On Linux, pip may pull CUDA-enabled PyTorch by default which includes ~3GB
+        # When building from a venv with CUDA torch installed, PyInstaller would
-        # of NVIDIA shared libraries that PyInstaller would bundle.
+        # bundle ~3GB of NVIDIA shared libraries. We exclude both the Python
+        # modules and the binary DLLs.
         nvidia_packages = [
             'nvidia', 'nvidia.cublas', 'nvidia.cuda_cupti', 'nvidia.cuda_nvrtc',
             'nvidia.cuda_runtime', 'nvidia.cudnn', 'nvidia.cufft', 'nvidia.curand',
@@ -127,7 +133,12 @@ def build_server(cuda=False):
     elif not cuda:
         print("Building for non-Apple Silicon platform - PyTorch only")
 
+    dist_dir = str(backend_dir / 'dist')
+    build_dir = str(backend_dir / 'build')
+
     args.extend([
+        '--distpath', dist_dir,
+        '--workpath', build_dir,
         '--noconfirm',
         '--clean',
     ])
@@ -135,12 +146,79 @@ def build_server(cuda=False):
     # Change to backend directory
     os.chdir(backend_dir)
     
+    # For CPU builds on Windows, ensure we're using CPU-only torch.
+    # If CUDA torch is installed (local dev), swap to CPU torch before building,
+    # then restore CUDA torch after. This prevents PyInstaller from bundling
+    # ~3GB of CUDA DLLs into the CPU binary.
+    restore_cuda = False
+    if not cuda and platform.system() == "Windows":
+        import subprocess
+        result = subprocess.run(
+            [sys.executable, "-c", "import torch; print(torch.version.cuda or '')"],
+            capture_output=True, text=True
+        )
+        has_cuda_torch = bool(result.stdout.strip())
+        if has_cuda_torch:
+            print("CUDA torch detected — installing CPU torch for CPU build...")
+            subprocess.run(
+                [sys.executable, "-m", "pip", "install", "torch", "torchvision", "torchaudio",
+                 "--index-url", "https://download.pytorch.org/whl/cpu", "--force-reinstall", "-q"],
+                check=True
+            )
+            restore_cuda = True
+
     # Run PyInstaller
-    PyInstaller.__main__.run(args)
+    try:
+        PyInstaller.__main__.run(args)
+    finally:
+        # Restore CUDA torch if we swapped it out (even on build failure)
+        if restore_cuda:
+            print("Restoring CUDA torch...")
+            import subprocess
+            subprocess.run(
+                [sys.executable, "-m", "pip", "install", "torch", "torchvision", "torchaudio",
+                 "--index-url", "https://download.pytorch.org/whl/cu126", "--force-reinstall", "-q"],
+                check=True
+            )
     
     print(f"Binary built in {backend_dir / 'dist' / binary_name}")
 
 
+def _get_cuda_dll_excludes():
+    """Get list of CUDA DLL filenames to exclude from CPU builds.
+    
+    When building locally with CUDA torch installed, PyInstaller bundles ~3GB of
+    CUDA DLLs from torch/lib/. Returns a list of DLL filenames to exclude.
+    """
+    try:
+        import torch
+        torch_lib = Path(torch.__file__).parent / 'lib'
+    except ImportError:
+        return []
+    
+    cuda_prefixes = (
+        'torch_cuda', 'cublas', 'cublasLt', 'cudnn', 'cusparse', 'cufft',
+        'cusolver', 'cusolverMg', 'curand', 'nvrtc', 'nvJitLink', 'nccl',
+        'nvperf', 'nvrtc-builtins',
+    )
+    
+    exclude_dlls = []
+    if torch_lib.exists():
+        for f in torch_lib.iterdir():
+            if f.suffix == '.dll' and any(f.name.startswith(p) for p in cuda_prefixes):
+                exclude_dlls.append(f.name)
+    
+    if exclude_dlls:
+        total_mb = sum(
+            (torch_lib / dll).stat().st_size 
+            for dll in exclude_dlls 
+            if (torch_lib / dll).exists()
+        ) / 1024 / 1024
+        print(f"CPU build: will exclude {len(exclude_dlls)} CUDA DLLs ({total_mb:.0f} MB)")
+    
+    return exclude_dlls
+
+
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(description="Build voicebox-server binary")
     parser.add_argument(

backend/main.py

@@ -108,6 +108,7 @@ _default_origins = [
     "http://127.0.0.1:17493",
     "tauri://localhost",         # Tauri webview (macOS)
     "https://tauri.localhost",   # Tauri webview (Windows/Linux)
+    "http://tauri.localhost",    # Tauri webview (Windows, some builds)
 ]
 _env_origins = os.environ.get("VOICEBOX_CORS_ORIGINS", "")
 _cors_origins = _default_origins + [o.strip() for o in _env_origins.split(",") if o.strip()]
@@ -142,6 +143,14 @@ async def shutdown():
     return {"message": "Shutting down..."}
 
 
+@app.post("/watchdog/disable")
+async def watchdog_disable():
+    """Disable the parent process watchdog so the server keeps running."""
+    from backend.server import disable_watchdog
+    disable_watchdog()
+    return {"message": "Watchdog disabled"}
+
+
 @app.get("/health", response_model=models.HealthResponse)
 async def health():
     """Health check endpoint."""

backend/server.py

@@ -6,6 +6,14 @@ absolute imports instead of relative imports.
 """
 
 import sys
+
+# Fast path: handle --version before any heavy imports so the Rust
+# version check doesn't block for 30+ seconds loading torch etc.
+if "--version" in sys.argv:
+    from backend import __version__
+    print(f"voicebox-server {__version__}")
+    sys.exit(0)
+
 import logging
 
 # Set up logging FIRST, before any imports that might fail
@@ -43,6 +51,100 @@ except Exception as e:
     logger.error(f"Failed to import required modules: {e}", exc_info=True)
     sys.exit(1)
 
+_watchdog_disabled = False
+
+
+def disable_watchdog():
+    """Disable the parent watchdog so the server keeps running after parent exits."""
+    global _watchdog_disabled
+    _watchdog_disabled = True
+
+
+def _start_parent_watchdog(parent_pid, data_dir=None):
+    """Monitor parent process and exit if it dies.
+
+    This is the clean shutdown mechanism: instead of the Tauri app trying to
+    forcefully kill the server (which spawns console windows on Windows),
+    the server monitors its parent and shuts itself down gracefully.
+    """
+    import os
+    import signal
+    import threading
+    import time
+
+    # Set up a file logger so we can debug in production
+    watchdog_logger = logging.getLogger("watchdog")
+    if data_dir:
+        try:
+            log_dir = os.path.join(data_dir, "logs")
+            os.makedirs(log_dir, exist_ok=True)
+            fh = logging.FileHandler(os.path.join(log_dir, "watchdog.log"))
+            fh.setFormatter(logging.Formatter('%(asctime)s - %(message)s'))
+            watchdog_logger.addHandler(fh)
+        except Exception:
+            pass
+    watchdog_logger.setLevel(logging.INFO)
+
+    def _is_pid_alive(pid):
+        """Check if a process with the given PID exists (cross-platform)."""
+        try:
+            if sys.platform == "win32":
+                import ctypes
+                kernel32 = ctypes.windll.kernel32
+                PROCESS_QUERY_LIMITED_INFORMATION = 0x1000
+                handle = kernel32.OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, False, pid)
+                if handle:
+                    # Check if process has actually exited
+                    STILL_ACTIVE = 259
+                    exit_code = ctypes.c_ulong()
+                    result = kernel32.GetExitCodeProcess(handle, ctypes.byref(exit_code))
+                    kernel32.CloseHandle(handle)
+                    if result and exit_code.value == STILL_ACTIVE:
+                        return True
+                    watchdog_logger.info(f"PID {pid}: exited with code {exit_code.value}")
+                    return False
+                # OpenProcess failed — check if it's an access error (process exists
+                # but we can't open it) vs process not found
+                error = ctypes.GetLastError()
+                ACCESS_DENIED = 5
+                if error == ACCESS_DENIED:
+                    return True  # process exists, we just can't open it
+                watchdog_logger.info(f"PID {pid}: OpenProcess failed, error={error}")
+                return False
+            else:
+                os.kill(pid, 0)
+                return True
+        except (OSError, PermissionError):
+            return False
+
+    def _watch():
+        watchdog_logger.info(f"Parent watchdog started, monitoring PID {parent_pid}, server PID {os.getpid()}")
+        # Verify parent is alive before starting the loop
+        alive = _is_pid_alive(parent_pid)
+        watchdog_logger.info(f"Parent PID {parent_pid} initial check: alive={alive}")
+        if not alive:
+            watchdog_logger.warning(f"Parent PID {parent_pid} not found on first check — disabling watchdog")
+            return
+        while True:
+            if _watchdog_disabled:
+                watchdog_logger.info("Watchdog disabled (keep server running), stopping monitor")
+                return
+            if not _is_pid_alive(parent_pid):
+                watchdog_logger.info(f"Parent process {parent_pid} gone, shutting down server...")
+                if sys.platform == "win32":
+                    # sys.exit triggers SystemExit, allowing uvicorn to run
+                    # shutdown handlers. os.kill(SIGTERM) on Windows calls
+                    # TerminateProcess which hard-kills without cleanup.
+                    os._exit(0)
+                else:
+                    os.kill(os.getpid(), signal.SIGTERM)
+                return
+            time.sleep(2)
+
+    t = threading.Thread(target=_watch, daemon=True)
+    t.start()
+
+
 if __name__ == "__main__":
     try:
         parser = argparse.ArgumentParser(description="voicebox backend server")
@@ -64,17 +166,21 @@ if __name__ == "__main__":
             default=None,
             help="Data directory for database, profiles, and generated audio",
         )
+        parser.add_argument(
+            "--parent-pid",
+            type=int,
+            default=None,
+            help="PID of parent process to monitor; server exits when parent dies",
+        )
         parser.add_argument(
             "--version",
             action="store_true",
-            help="Print version and exit",
+            help="Print version and exit (handled above, kept for argparse help)",
         )
         args = parser.parse_args()
 
-        if args.version:
+        if args.parent_pid is not None and args.parent_pid <= 0:
-            from backend import __version__
+            parser.error("--parent-pid must be a positive integer")
-            print(f"voicebox-server {__version__}")
-            sys.exit(0)
 
         # Detect backend variant from binary name
         # voicebox-server-cuda → sets VOICEBOX_BACKEND_VARIANT=cuda
@@ -87,6 +193,14 @@ if __name__ == "__main__":
             os.environ["VOICEBOX_BACKEND_VARIANT"] = "cpu"
             logger.info("Backend variant: CPU")
 
+        # Register parent watchdog to start after server is fully ready
+        if args.parent_pid is not None:
+            _parent_pid = args.parent_pid
+            _data_dir = args.data_dir
+            @app.on_event("startup")
+            async def _on_startup():
+                _start_parent_watchdog(_parent_pid, _data_dir)
+
         logger.info(f"Parsed arguments: host={args.host}, port={args.port}, data_dir={args.data_dir}")
 
         # Set data directory if provided

justfile

@@ -52,6 +52,7 @@ setup-python:
         {{ pip }} install -r {{ backend_dir }}/requirements-mlx.txt
     fi
     {{ pip }} install git+https://github.com/QwenLM/Qwen3-TTS.git
+    {{ pip }} install pyinstaller -q
     echo "Python environment ready."
 
 [windows]
@@ -74,6 +75,7 @@ setup-python:
     & "{{ pip }}" install -r {{ backend_dir }}/requirements.txt
     & "{{ pip }}" install --no-deps chatterbox-tts
     & "{{ pip }}" install git+https://github.com/QwenLM/Qwen3-TTS.git
+    & "{{ pip }}" install pyinstaller -q
     Write-Host "Python environment ready."
 
 # Install JavaScript dependencies
@@ -105,14 +107,14 @@ dev: _ensure-venv _ensure-sidecar
 
 [windows]
 dev: _ensure-venv _ensure-sidecar
-    $backendJob = $null
+    $backendJob = $null; \
     try { $null = Invoke-WebRequest -Uri "http://127.0.0.1:17493/health" -UseBasicParsing -TimeoutSec 2 -ErrorAction Stop; Write-Host "Backend already running on http://localhost:17493" } catch { \
         Write-Host "Starting backend on http://localhost:17493 ..."; \
-        $backendJob = Start-Job -ScriptBlock { & "{{ python }}" -m uvicorn backend.main:app --reload --port 17493 } -WorkingDirectory (Get-Location); \
+        $backendJob = Start-Process -PassThru -NoNewWindow -FilePath "{{ python }}" -ArgumentList "-m","uvicorn","backend.main:app","--reload","--port","17493"; \
         Start-Sleep -Seconds 2; \
-    }
+    }; \
-    Write-Host "Starting Tauri desktop app..."
+    Write-Host "Starting Tauri desktop app..."; \
-    try { Set-Location "{{ tauri_dir }}"; bun run tauri dev } finally { if ($backendJob) { Stop-Job $backendJob -ErrorAction SilentlyContinue; Remove-Job $backendJob -Force -ErrorAction SilentlyContinue } }
+    try { Set-Location "{{ tauri_dir }}"; bun run tauri dev } finally { if ($backendJob) { taskkill /PID $backendJob.Id /T /F 2>$null | Out-Null } }
 
 # Start backend only
 [unix]
@@ -124,9 +126,14 @@ dev-backend: _ensure-venv
     & "{{ python }}" -m uvicorn backend.main:app --reload --port 17493
 
 # Start Tauri desktop app only (backend must be running separately)
+[unix]
 dev-frontend: _ensure-sidecar
     cd {{ tauri_dir }} && bun run tauri dev
 
+[windows]
+dev-frontend: _ensure-sidecar
+    Set-Location "{{ tauri_dir }}"; bun run tauri dev
+
 # Start backend (if not already running) + web app (no Tauri)
 [unix]
 dev-web: _ensure-venv
@@ -149,14 +156,14 @@ dev-web: _ensure-venv
 
 [windows]
 dev-web: _ensure-venv
-    $backendJob = $null
+    $backendJob = $null; \
     try { $null = Invoke-WebRequest -Uri "http://127.0.0.1:17493/health" -UseBasicParsing -TimeoutSec 2 -ErrorAction Stop; Write-Host "Backend already running on http://localhost:17493" } catch { \
         Write-Host "Starting backend on http://localhost:17493 ..."; \
-        $backendJob = Start-Job -ScriptBlock { & "{{ python }}" -m uvicorn backend.main:app --reload --port 17493 } -WorkingDirectory (Get-Location); \
+        $backendJob = Start-Process -PassThru -NoNewWindow -FilePath "{{ python }}" -ArgumentList "-m","uvicorn","backend.main:app","--reload","--port","17493"; \
         Start-Sleep -Seconds 2; \
-    }
+    }; \
-    Write-Host "Starting web app..."
+    Write-Host "Starting web app..."; \
-    try { Set-Location "{{ web_dir }}"; bun run dev } finally { if ($backendJob) { Stop-Job $backendJob -ErrorAction SilentlyContinue; Remove-Job $backendJob -Force -ErrorAction SilentlyContinue } }
+    try { Set-Location "{{ web_dir }}"; bun run dev } finally { if ($backendJob) { taskkill /PID $backendJob.Id /T /F 2>$null | Out-Null } }
 
 # Kill all dev processes
 [unix]
@@ -175,23 +182,56 @@ kill:
 # Build everything (server binary + desktop app)
 build: build-server build-tauri
 
-# Build Python server binary
+# Build Python server binary (CPU)
 [unix]
 build-server: _ensure-venv
     PATH="{{ venv_bin }}:$PATH" ./scripts/build-server.sh
 
 [windows]
 build-server: _ensure-venv
-    $env:PATH = "{{ venv_bin }};$env:PATH"; & "{{ python }}" -m PyInstaller backend/voicebox-server.spec
+    $ErrorActionPreference = "Stop"; \
+    $env:PATH = "{{ venv_bin }};$env:PATH"; \
+    & "{{ python }}" backend/build_binary.py; \
+    if ($LASTEXITCODE -ne 0) { throw "build_binary.py failed with exit code $LASTEXITCODE" }; \
+    $triple = (rustc --print host-tuple); \
+    New-Item -ItemType Directory -Path "{{ tauri_dir }}/src-tauri/binaries" -Force | Out-Null; \
+    Copy-Item "backend/dist/voicebox-server.exe" "{{ tauri_dir }}/src-tauri/binaries/voicebox-server-$triple.exe" -Force; \
+    Write-Host "Copied sidecar: voicebox-server-$triple.exe"
+
+# Build CUDA server binary and place in app data dir for local testing
+[windows]
+build-server-cuda: _ensure-venv
+    $ErrorActionPreference = "Stop"; \
+    $env:PATH = "{{ venv_bin }};$env:PATH"; \
+    & "{{ python }}" backend/build_binary.py --cuda; \
+    if ($LASTEXITCODE -ne 0) { throw "build_binary.py --cuda failed with exit code $LASTEXITCODE" }; \
+    $dest = "$env:APPDATA/com.voicebox.app/backends"; \
+    New-Item -ItemType Directory -Path $dest -Force | Out-Null; \
+    Copy-Item "backend/dist/voicebox-server-cuda.exe" "$dest/voicebox-server-cuda.exe" -Force; \
+    Write-Host "Copied CUDA binary to $dest"
+
+# Build everything locally: CPU server + CUDA server + installable Tauri app
+[windows]
+build-local: build-server build-server-cuda build-tauri
 
 # Build Tauri desktop app
+[unix]
 build-tauri:
     cd {{ tauri_dir }} && bun run tauri build
 
+[windows]
+build-tauri:
+    Set-Location "{{ tauri_dir }}"; bun run tauri build
+
 # Build web app
+[unix]
 build-web:
     cd {{ web_dir }} && bun run build
 
+[windows]
+build-web:
+    Set-Location "{{ web_dir }}"; bun run build
+
 # ─── Code Quality ────────────────────────────────────────────────────
 
 # Run all checks (lint + format + typecheck)
@@ -213,8 +253,13 @@ fix:
 # ─── Database ─────────────────────────────────────────────────────────
 
 # Initialize SQLite database
+[unix]
 db-init: _ensure-venv
-    cd {{ backend_dir }} && {{ python }} -c "from database import init_db; init_db()"
+    {{ python }} -c "from backend.database import init_db; init_db()"
+
+[windows]
+db-init: _ensure-venv
+    & "{{ python }}" -c "from backend.database import init_db; init_db()"
 
 # Reset database (delete + reinit)
 [unix]

tauri/src-tauri/src/main.rs

@@ -282,6 +282,7 @@ async fn start_server(
         .ok_or_else(|| "Invalid data dir path".to_string())?
         .to_string();
     let port_str = SERVER_PORT.to_string();
+    let parent_pid_str = std::process::id().to_string();
     let is_remote = remote.unwrap_or(false);
 
     // Resolve the custom models directory from the parameter or stored state
@@ -294,7 +295,7 @@ async fn start_server(
     let spawn_result = if let Some(ref cuda_path) = cuda_binary {
         println!("Launching CUDA backend: {:?}", cuda_path);
         let mut cmd = app.shell().command(cuda_path.to_str().unwrap());
-        cmd = cmd.args(["--data-dir", &data_dir_str, "--port", &port_str]);
+        cmd = cmd.args(["--data-dir", &data_dir_str, "--port", &port_str, "--parent-pid", &parent_pid_str]);
         if is_remote {
             cmd = cmd.args(["--host", "0.0.0.0"]);
         }
@@ -304,7 +305,7 @@ async fn start_server(
         cmd.spawn()
     } else {
         // Use the bundled CPU sidecar
-        sidecar = sidecar.args(["--data-dir", &data_dir_str, "--port", &port_str]);
+        sidecar = sidecar.args(["--data-dir", &data_dir_str, "--port", &port_str, "--parent-pid", &parent_pid_str]);
         if is_remote {
             sidecar = sidecar.args(["--host", "0.0.0.0"]);
         }
@@ -490,67 +491,13 @@ async fn start_server(
     Ok(format!("http://127.0.0.1:{}", SERVER_PORT))
 }
 
-/// Check if a Windows process is still running
-#[cfg(windows)]
-fn is_process_running(pid: u32) -> bool {
-    use std::process::Command;
-    if let Ok(output) = Command::new("tasklist")
-        .args(["/FI", &format!("PID eq {}", pid), "/FO", "CSV", "/NH"])
-        .output()
-    {
-        // If process exists, tasklist returns it in output
-        let output_str = String::from_utf8_lossy(&output.stdout);
-        return !output_str.trim().is_empty() && output_str.contains(&pid.to_string());
-    }
-    false
-}
-
-/// Kill entire Windows process tree by enumerating children
-#[cfg(windows)]
-fn kill_windows_process_tree(parent_pid: u32) -> Result<(), String> {
-    use std::process::Command;
-
-    // Find all child processes using WMIC
-    let output = Command::new("wmic")
-        .args([
-            "process",
-            "where",
-            &format!("ParentProcessId={}", parent_pid),
-            "get",
-            "ProcessId"
-        ])
-        .output();
-
-    if let Ok(output) = output {
-        let output_str = String::from_utf8_lossy(&output.stdout);
-        for line in output_str.lines().skip(1) { // Skip header
-            if let Ok(child_pid) = line.trim().parse::<u32>() {
-                println!("Found child process: {}", child_pid);
-                // Recursively kill child's children
-                let _ = kill_windows_process_tree(child_pid);
-                // Kill the child
-                let _ = Command::new("taskkill")
-                    .args(["/PID", &child_pid.to_string(), "/F"])
-                    .output();
-            }
-        }
-    }
-
-    // Kill the parent process
-    let _ = Command::new("taskkill")
-        .args(["/PID", &parent_pid.to_string(), "/F"])
-        .output();
-
-    Ok(())
-}
-
 #[command]
 async fn stop_server(state: State<'_, ServerState>) -> Result<(), String> {
     let pid = state.server_pid.lock().unwrap().take();
     let _child = state.child.lock().unwrap().take();
     
     if let Some(pid) = pid {
-        println!("stop_server: Killing server process group with PID: {}", pid);
+        println!("stop_server: Stopping server with PID: {}", pid);
         
         #[cfg(unix)]
         {
@@ -569,62 +516,25 @@ async fn stop_server(state: State<'_, ServerState>) -> Result<(), String> {
             let _ = Command::new("kill")
                 .args(["-9", &pid.to_string()])
                 .output();
+            
+            println!("stop_server: Process group kill completed");
         }
         
         #[cfg(windows)]
         {
-            // Layer 1: Try graceful HTTP shutdown first
+            // Send graceful shutdown via HTTP — the server's parent-pid watchdog
-            println!("Attempting graceful shutdown via HTTP...");
+            // will also handle cleanup if this app process exits.
+            println!("Sending graceful shutdown via HTTP...");
             let client = reqwest::blocking::Client::builder()
                 .timeout(std::time::Duration::from_secs(2))
                 .build()
                 .unwrap();
 
-            let shutdown_result = client
+            let _ = client
                 .post(&format!("http://127.0.0.1:{}/shutdown", SERVER_PORT))
                 .send();
 
-            if shutdown_result.is_ok() {
+            println!("Shutdown request sent (server watchdog will handle cleanup)");
-                println!("HTTP shutdown sent, waiting for graceful exit...");
-                // Wait up to 3 seconds for graceful shutdown
-                for i in 0..30 {
-                    std::thread::sleep(std::time::Duration::from_millis(100));
-                    if !is_process_running(pid) {
-                        println!("Process exited gracefully after {}ms", i * 100);
-                        return Ok(());
-                    }
-                }
-                println!("Graceful shutdown timed out, forcing kill...");
-            } else {
-                println!("HTTP shutdown failed, forcing kill...");
-            }
-
-            // Layer 2: Kill process tree with enumeration
-            println!("Killing process tree for wrapper PID {}...", pid);
-            kill_windows_process_tree(pid)?;
-
-            // Layer 3: Verify and kill by name if still running
-            std::thread::sleep(std::time::Duration::from_millis(200));
-            if is_process_running(pid) {
-                println!("Process tree kill failed, killing by name...");
-                use std::process::Command;
-                let _ = Command::new("taskkill")
-                    .args(["/IM", "voicebox-server.exe", "/T", "/F"])
-                    .output();
-            }
-
-            // Layer 4: Final verification
-            std::thread::sleep(std::time::Duration::from_millis(200));
-            if is_process_running(pid) {
-                eprintln!("WARNING: Failed to kill server after all attempts");
-            } else {
-                println!("Server killed successfully");
-            }
-        }
-
-        #[cfg(unix)]
-        {
-            println!("stop_server: Process group kill completed");
         }
     }
     
@@ -798,9 +708,17 @@ pub fn run() {
             play_audio_to_devices,
             stop_audio_playback
         ])
-        .on_window_event(|window, event| {
+        .on_window_event({
+            let closing = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
+            move |window, event| {
             if let WindowEvent::CloseRequested { api, .. } = event {
-                // Prevent automatic close
+                // If we're already in the close flow, let it proceed
+                if closing.load(std::sync::atomic::Ordering::SeqCst) {
+                    return;
+                }
+                closing.store(true, std::sync::atomic::Ordering::SeqCst);
+
+                // Prevent automatic close so frontend can clean up
                 api.prevent_close();
 
                 // Emit event to frontend to check setting and stop server if needed
@@ -808,162 +726,77 @@ pub fn run() {
 
                 if let Err(e) = app_handle.emit("window-close-requested", ()) {
                     eprintln!("Failed to emit window-close-requested event: {}", e);
-                    // If event emission fails, allow close anyway
                     window.close().ok();
                     return;
                 }
 
                 // Set up listener for frontend response
                 let window_for_close = window.clone();
+                let closing_for_timeout = closing.clone();
                 let (tx, mut rx) = mpsc::unbounded_channel::<()>();
 
-                // Listen for response from frontend using window's listen method
                 let listener_id = window.listen("window-close-allowed", move |_| {
-                    // Frontend has checked setting and stopped server if needed
-                    // Signal that we can close
                     let _ = tx.send(());
                 });
 
-                // Wait for frontend response or timeout
-                // Use tauri::async_runtime::spawn instead of tokio::spawn to avoid
-                // panics when the Tokio runtime is being dropped during app shutdown
                 tauri::async_runtime::spawn(async move {
                     tokio::select! {
                         _ = rx.recv() => {
-                            // Frontend responded, close window
                             window_for_close.close().ok();
                         }
                         _ = tokio::time::sleep(tokio::time::Duration::from_secs(5)) => {
-                            // Timeout - close anyway
                             eprintln!("Window close timeout, closing anyway");
                             window_for_close.close().ok();
                         }
                     }
-                    // Clean up listener
                     window_for_close.unlisten(listener_id);
+                    closing_for_timeout.store(false, std::sync::atomic::Ordering::SeqCst);
                 });
             }
-        })
+        }})
         .build(tauri::generate_context!())
         .expect("error while building tauri application")
         .run(|app, event| {
+            let _ = &app; // used on unix
             match &event {
                 RunEvent::Exit => {
-                    println!("=================================================================");
-                    println!("RunEvent::Exit received - checking server cleanup");
                     let state = app.state::<ServerState>();
                     let keep_running = *state.keep_running_on_close.lock().unwrap();
-                    println!("keep_running_on_close = {}", keep_running);
+
-                    
+                    if keep_running {
-                    if !keep_running {
+                        // Tell the server to disable its watchdog so it survives
-                        // Get the stored PID for process group killing
+                        // after this process exits.
-                        let pid = state.server_pid.lock().unwrap().take();
+                        println!("Keep server running: disabling watchdog...");
-                        // Also take the child to clean up
+                        let client = reqwest::blocking::Client::builder()
-                        let _child = state.child.lock().unwrap().take();
+                            .timeout(std::time::Duration::from_secs(2))
-                        
+                            .build()
-                        if let Some(pid) = pid {
+                            .unwrap();
-                            println!("Killing server process group with PID: {}", pid);
+                        let _ = client
-                            
+                            .post(&format!("http://127.0.0.1:{}/watchdog/disable", SERVER_PORT))
-                            // Kill the entire process group on Unix systems
+                            .send();
-                            // Using negative PID sends signal to all processes in the group
+                    } else {
-                            #[cfg(unix)]
+                        // Server will self-terminate via parent-pid watchdog when
-                            {
+                        // this process exits. On Unix, also send SIGTERM for
+                        // immediate cleanup.
+                        println!("RunEvent::Exit - server will self-terminate via watchdog");
+
+                        #[cfg(unix)]
+                        {
+                            if let Some(pid) = state.server_pid.lock().unwrap().take() {
                                 use std::process::Command;
-                                // First try SIGTERM to the process group
+                                let _ = Command::new("kill")
-                                let pgid_kill = Command::new("kill")
                                     .args(["-TERM", "--", &format!("-{}", pid)])
                                     .output();
-                                
-                                match pgid_kill {
-                                    Ok(output) => {
-                                        if output.status.success() {
-                                            println!("SIGTERM sent to process group -{}", pid);
-                                        } else {
-                                            // Process group kill failed, try direct kill
-                                            println!("Process group kill failed, trying direct kill");
-                                            let _ = Command::new("kill")
-                                                .args(["-TERM", &pid.to_string()])
-                                                .output();
-                                        }
-                                    }
-                                    Err(e) => {
-                                        eprintln!("Failed to execute kill command: {}", e);
-                                    }
-                                }
-                                
-                                // Give it a moment, then force kill if needed
                                 std::thread::sleep(std::time::Duration::from_millis(100));
-                                
-                                // Force kill with SIGKILL
                                 let _ = Command::new("kill")
                                     .args(["-9", "--", &format!("-{}", pid)])
                                     .output();
                                 let _ = Command::new("kill")
                                     .args(["-9", &pid.to_string()])
                                     .output();
-                                
-                                println!("Server process group kill completed");
                             }
-                            
-                            #[cfg(windows)]
-                            {
-                                // Layer 1: Try graceful HTTP shutdown first
-                                println!("Attempting graceful shutdown via HTTP...");
-                                let client = reqwest::blocking::Client::builder()
-                                    .timeout(std::time::Duration::from_secs(2))
-                                    .build()
-                                    .unwrap();
-
-                                let shutdown_result = client
-                                    .post(&format!("http://127.0.0.1:{}/shutdown", SERVER_PORT))
-                                    .send();
-
-                                if shutdown_result.is_ok() {
-                                    println!("HTTP shutdown sent, waiting for graceful exit...");
-                                    // Wait up to 3 seconds for graceful shutdown
-                                    for i in 0..30 {
-                                        std::thread::sleep(std::time::Duration::from_millis(100));
-                                        if !is_process_running(pid) {
-                                            println!("Process exited gracefully after {}ms", i * 100);
-                                            println!("Server process tree kill completed");
-                                            return;
-                                        }
-                                    }
-                                    println!("Graceful shutdown timed out, forcing kill...");
-                                } else {
-                                    println!("HTTP shutdown failed, forcing kill...");
-                                }
-
-                                // Layer 2: Kill process tree with enumeration
-                                println!("Killing process tree for wrapper PID {}...", pid);
-                                let _ = kill_windows_process_tree(pid);
-
-                                // Layer 3: Verify and kill by name if still running
-                                std::thread::sleep(std::time::Duration::from_millis(200));
-                                if is_process_running(pid) {
-                                    println!("Process tree kill failed, killing by name...");
-                                    use std::process::Command;
-                                    let _ = Command::new("taskkill")
-                                        .args(["/IM", "voicebox-server.exe", "/T", "/F"])
-                                        .output();
-                                }
-
-                                // Layer 4: Final verification
-                                std::thread::sleep(std::time::Duration::from_millis(200));
-                                if is_process_running(pid) {
-                                    eprintln!("WARNING: Failed to kill server after all attempts");
-                                } else {
-                                    println!("Server killed successfully");
-                                }
-                                println!("Server process tree kill completed");
-                            }
-                        } else {
-                            println!("No server PID found (already stopped or never started)");
                         }
-                    } else {
-                        println!("Keeping server running per user setting");
                     }
-                    println!("=================================================================");
                 }
                 RunEvent::ExitRequested { api, .. } => {
                     println!("RunEvent::ExitRequested received");