chore(release): v0.3.4 (#269)

Patch release. Version bumped across all sources + lock files; [0.3.4] CHANGELOG.

Ships:
- #255 — PyTorch-Whisper backend works as a standalone fallback (no cuDNN 8,
  no OMNIVOICE_PRELOAD_TTS_ASR=1), unblocking Windows+NVIDIA users hitting the
  cudnn_ops_infer64_8.dll error.

Tagging v0.3.4 triggers release.yml (desktop) + docker.yml.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.github/workflows/release.yml

@@ -540,7 +540,13 @@ jobs:
 
           # Gather artifact paths per matrix leg's `bundles` (msi/app/dmg/deb/appimage/updater).
           # `find` is portable across all three runners (Git Bash on Windows).
-          mapfile -t ARTIFACTS < <(find "$BUNDLE_DIR" -type f \
+          # NB: macOS runners use /bin/bash 3.2, which has no `mapfile` (a bash 4+
+          # builtin) — using it 127'd this step and dropped the macOS SHA256SUMS
+          # for v0.3.1 and v0.3.2. A `while read` loop is portable to bash 3.2.
+          ARTIFACTS=()
+          while IFS= read -r artifact; do
+            ARTIFACTS+=("$artifact")
+          done < <(find "$BUNDLE_DIR" -type f \
             \( -name "*.dmg" -o -name "*.app.tar.gz" -o -name "*.app.tar.gz.sig" \
                -o -name "*.msi" -o -name "*.msi.sig" \
                -o -name "*.AppImage" -o -name "*.AppImage.sig" \

CHANGELOG.md

@@ -6,6 +6,68 @@ The format is loosely based on [Keep a Changelog](https://keepachangelog.com/).
 Versions track the desktop app (`tauri.conf.json` + `frontend/src-tauri/Cargo.toml`).
 The bundled TTS model package (`pyproject.toml`) is versioned independently.
 
+## [0.3.4] — 2026-06-03
+
+### Fixed
+- **Transcription on Windows + NVIDIA failed with `Could not locate
+  cudnn_ops_infer64_8.dll`.** WhisperX/faster-whisper need cuDNN 8 (via
+  CTranslate2); when the side-loaded `cudnn8_compat` libs are missing, the
+  **PyTorch Whisper** backend (Settings → Models) now works as a drop-in
+  fallback — it builds its own transformers pipeline on PyTorch's cuDNN-9
+  stack, with no CTranslate2/cuDNN-8 dependency and no
+  `OMNIVOICE_PRELOAD_TTS_ASR=1` required. (#255)
+
+## [0.3.3] — 2026-06-03
+
+### Fixed
+- **Settings → About showed the wrong architecture in the Docker/web build.**
+  The "Architecture" row rendered the *client browser's* platform
+  (`navigator.platform` → e.g. "Win32"); it now reports the **server's** CPU
+  architecture from the backend (`platform.machine()`), correct for both the
+  desktop app and Docker. The blank version/GPU/RAM/VRAM in the same report
+  were the loopback-gate 403s already fixed in v0.3.2. (#262)
+
+### CI
+- The release SHA-256 checksum step no longer uses `mapfile` (a bash 4+
+  builtin) — it broke on the macOS runner's bash 3.2 and dropped the macOS
+  `SHA256SUMS` for v0.3.1/v0.3.2. Now portable to bash 3.2.
+
+## [0.3.2] — 2026-06-03
+
+### Fixed
+- **"Loopback origin required" all over the Docker UI** (and a blank version).
+  The `/system/*` and `/api/settings/*` routes are restricted to a loopback
+  origin, but Docker's NAT makes every request look non-loopback, so the gate
+  403'd the operator out of the admin UI — including `/system/info` (blanking
+  the version) and HF-token entry. The Docker image now runs with
+  `OMNIVOICE_SERVER_MODE=1`, which relaxes the gate for the headless
+  deployment; exposure is governed by the `-p` port mapping plus the optional
+  share PIN. Desktop builds are unaffected — their loopback boundary (and the
+  denial of admin routes to LAN share guests) is unchanged. (#261)
+
+## [0.3.1] — 2026-06-03
+
+First tagged build of the 0.3 line off `main` — it ships the accumulated
+`[0.3.0]` work below plus the fixes here. (The `[0.3.0]` milestone heading is
+kept for the qualitative "actually useful" release.)
+
+### Fixed
+- **Voice-clone / export download crashed in the Docker & browser build** with
+  `TypeError: Cannot read properties of undefined (reading 'invoke')`. The
+  export button called the Tauri save dialog unconditionally; outside the
+  desktop shell it now falls back to a standard browser download of the file
+  served at `/audio/<path>`. (#256)
+- **Docker container showed no version** (a dash) in Settings → About, and the
+  desktop-only update-channel toggle appeared in the web build. The running
+  version is now read from the backend (`/system/info` `app_version`, `/health`
+  `version`); the updater UI is hidden outside Tauri. Also corrected the
+  version-check command in the Docker docs (`omnivoice`, not
+  `omnivoice-studio`). (#249)
+- **Transcription failures were masked** by a generic "Transcribe stream
+  dropped" message. The transcribe SSE stream now surfaces the real, sanitized
+  cause (with an actionable hint) instead of silently dropping when model load
+  or VRAM offload fails. (#255)
+
 ## [0.3.0] — Unreleased
 
 ### Added

backend/api/dependencies.py

@@ -5,9 +5,12 @@ These are intentionally tiny — one concern per dependency — so they can be
 composed at the route or router level without surprises.
 
 Currently exposed:
-- `require_loopback`: 403 unless the request came from a loopback origin.
+- `require_loopback`: 403 unless the request came from a loopback origin
+  (bypassed in explicit server mode — see `_server_mode`).
 """
 
+import os
+
 from fastapi import HTTPException, Request
 
 
@@ -19,6 +22,28 @@ from fastapi import HTTPException, Request
 # the guard: nothing here matches a non-loopback origin.
 _LOOPBACK_HOSTS = frozenset({"127.0.0.1", "::1", "localhost"})
 
+_TRUTHY = frozenset({"1", "true", "yes", "on"})
+
+
+def _server_mode() -> bool:
+    """Whether this process is a headless server deployment (Docker image).
+
+    In Docker the loopback gate is *unenforceable*: Docker's network NAT
+    rewrites ``request.client.host`` to the bridge gateway (e.g. 172.17.0.1)
+    even for a localhost-only ``-p 127.0.0.1:3900:3900`` mapping, so every
+    request looks non-loopback and the gate 403s the operator out of the
+    system/settings routes they need (issue #261 — incl. ``/system/info``,
+    which blanks the version display).
+
+    The Docker image sets ``OMNIVOICE_SERVER_MODE=1`` to opt out of the gate.
+    Network exposure then rests on the operator's port mapping plus the
+    optional share PIN (``NetworkAccessMiddleware`` still 401s unauthenticated
+    non-loopback clients whenever a PIN is set). The desktop build never sets
+    this, so its loopback boundary — including denying LAN share guests access
+    to admin routes — is unchanged. Read at call time so it stays testable.
+    """
+    return os.environ.get("OMNIVOICE_SERVER_MODE", "").strip().lower() in _TRUTHY
+
 
 def require_loopback(request: Request) -> None:
     """Reject any request whose `client.host` is not a loopback address.
@@ -35,7 +60,14 @@ def require_loopback(request: Request) -> None:
     Returns None on success (FastAPI dependency convention). Raises 403
     on rejection — the response body is `{"detail": "loopback origin required"}`
     so existing tests for `/system/set-env` keep passing without modification.
+
+    In server mode (Docker, see `_server_mode`) the gate is a no-op: the
+    loopback origin is unenforceable there and exposure is governed by the
+    deployment's port mapping + the optional share PIN instead.
     """
     host = request.client.host if request.client else None
-    if host not in _LOOPBACK_HOSTS:
-        raise HTTPException(status_code=403, detail="loopback origin required")
+    if host in _LOOPBACK_HOSTS:
+        return
+    if _server_mode():
+        return
+    raise HTTPException(status_code=403, detail="loopback origin required")

backend/api/routers/dub_core.py

@@ -384,24 +384,38 @@ async def dub_transcribe_stream(job_id: str):
     if not job:
         preflight_error = "Job not found. It may have been cleaned up or was never created."
     else:
-        _model = await get_model()
-        asr_audio_target = job.get("vocals_path")
-        if not asr_audio_target or not os.path.exists(asr_audio_target):
-            asr_audio_target = job.get("audio_path")
-        if not asr_audio_target or not os.path.exists(asr_audio_target):
-            preflight_error = "No audio available for transcription."
-        else:
-            from services.asr_backend import get_active_asr_backend
-            try:
-                _asr_backend = get_active_asr_backend(asr_pipe=getattr(_model, "_asr_pipe", None))
-                if _asr_backend.id == "pytorch-whisper" and getattr(_model, "_asr_pipe", None) is None:
-                    preflight_error = (
-                        "No ASR backend is ready. Install WhisperX/faster-whisper/MLX Whisper "
-                        "or set OMNIVOICE_PRELOAD_TTS_ASR=1 before launch to use the PyTorch fallback."
+        # Guard the model load: if it raises, the SSE stream would otherwise die
+        # before emitting any event, and the UI shows a misleading generic
+        # "stream dropped" message instead of the real cause (issue #255).
+        try:
+            _model = await get_model()
+        except Exception as e:
+            logger.exception("transcribe preflight: model load failed (job=%s)", job_id)
+            from core.failure import build_failure
+            f = build_failure(e, stage="transcribe-preflight", include_diagnostic=False)
+            preflight_error = f["reason"] + (f" — {f['hint']}" if f.get("hint") else "")
+            _model = None
+        if _model is not None:
+            asr_audio_target = job.get("vocals_path")
+            if not asr_audio_target or not os.path.exists(asr_audio_target):
+                asr_audio_target = job.get("audio_path")
+            if not asr_audio_target or not os.path.exists(asr_audio_target):
+                preflight_error = "No audio available for transcription."
+            else:
+                from services.asr_backend import get_active_asr_backend
+                try:
+                    # The PyTorch-Whisper backend lazily builds its own pipeline
+                    # when no preloaded `_asr_pipe` is present (issue #255), so it
+                    # no longer needs OMNIVOICE_PRELOAD_TTS_ASR=1 — don't reject it
+                    # here; any load failure surfaces per-chunk with a real cause.
+                    _asr_backend = get_active_asr_backend(asr_pipe=getattr(_model, "_asr_pipe", None))
+                except Exception as e:
+                    from core.failure import build_failure
+                    f = build_failure(e, stage="transcribe-preflight", include_diagnostic=False)
+                    preflight_error = "ASR backend initialization failed: " + f["reason"] + (
+                        f" — {f['hint']}" if f.get("hint") else ""
                     )
-            except Exception as e:
-                preflight_error = f"ASR backend initialization failed: {e}"
-            scene_cuts = job.get("scene_cuts") or []
+                scene_cuts = job.get("scene_cuts") or []
 
     async def gen():
         if preflight_error:
@@ -429,7 +443,12 @@ async def dub_transcribe_stream(job_id: str):
 
         # Free VRAM: move TTS model to CPU so WhisperX + VAD can fit.
         # Only offloads when free GPU memory is < 4 GB (e.g. laptop GPUs).
-        await loop.run_in_executor(_cpu_pool, offload_tts_for_asr)
+        # Non-fatal: an offload failure must not drop the stream (#255) —
+        # transcription can still proceed (it just has less headroom).
+        try:
+            await loop.run_in_executor(_cpu_pool, offload_tts_for_asr)
+        except Exception as e:
+            logger.warning("offload_tts_for_asr failed (continuing): %s", e)
 
         all_segments: list[dict] = []
         detected_lang = None
@@ -540,15 +559,23 @@ async def dub_transcribe_stream(job_id: str):
         # whisperx's VAD load, or an unsupported audio format.
         if not all_segments:
             # Deduplicate while preserving order so one root cause doesn't
-            # repeat N times in the UI toast.
+            # repeat N times in the UI toast. Sanitize each message so home
+            # paths / tokens from a backend traceback never leak (#255).
+            from core.failure import sanitize, build_failure
             seen = set()
             uniq: list[str] = []
             for msg in chunk_errors:
-                if msg and msg not in seen:
-                    seen.add(msg)
-                    uniq.append(msg)
+                s = sanitize(msg)
+                if s and s not in seen:
+                    seen.add(s)
+                    uniq.append(s)
             if uniq:
                 detail = "Transcription produced no segments. " + " | ".join(uniq[:3])
+                # Add the actionable hint for a recognized failure class
+                # (e.g. pkg_resources missing → install setuptools).
+                hint = build_failure(" ".join(uniq), stage="transcribe", include_diagnostic=False).get("hint")
+                if hint:
+                    detail += f" — {hint}"
             else:
                 detail = (
                     "Transcription produced no segments. The audio may be silent, "

backend/api/routers/system.py

@@ -1,5 +1,6 @@
 import os
 import sys
+import platform
 import uuid
 import psutil
 import asyncio
@@ -15,6 +16,7 @@ import torch
 import shutil
 
 from core.config import OUTPUTS_DIR, DATA_DIR, CRASH_LOG_PATH, LOG_PATH, IDLE_TIMEOUT_SECONDS
+from core.version import APP_VERSION
 from services.model_manager import get_model_status, get_best_device
 from services.ffmpeg_utils import find_ffmpeg, run_ffmpeg
 
@@ -168,6 +170,7 @@ def system_info():
     try:
         _ffmpeg = find_ffmpeg()
         return {
+            "app_version": APP_VERSION,
             "data_dir": DATA_DIR,
             "outputs_dir": OUTPUTS_DIR,
             "crash_log_path": CRASH_LOG_PATH,
@@ -179,6 +182,7 @@ def system_info():
             "device": get_best_device(),
             "python": sys.version.split()[0],
             "platform": sys.platform,
+            "arch": platform.machine(),
             "ffmpeg_ok": bool(_ffmpeg),
             "ffmpeg_path": _ffmpeg or "",
             "proxy_url": os.environ.get("HTTP_PROXY") or os.environ.get("http_proxy") or "",
@@ -193,6 +197,7 @@ def system_info():
     except Exception as e:
         logger.exception("system_info failed — returning safe defaults")
         return {
+            "app_version": APP_VERSION,
             "data_dir": DATA_DIR,
             "outputs_dir": OUTPUTS_DIR,
             "crash_log_path": str(CRASH_LOG_PATH),
@@ -204,6 +209,7 @@ def system_info():
             "device": "cpu",
             "python": sys.version.split()[0],
             "platform": sys.platform,
+            "arch": platform.machine(),
             "proxy_url": "",
             "share_enabled": network_share.get_state().enabled,
             "share_port": network_share.get_state().share_port,

backend/api/schemas.py

@@ -25,6 +25,7 @@ class SystemInfoResponse(BaseModel):
     """GET /system/info"""
     model_config = ConfigDict(extra="allow")
 
+    app_version: str = ""
     data_dir: str
     outputs_dir: str
     crash_log_path: str
@@ -36,6 +37,7 @@ class SystemInfoResponse(BaseModel):
     device: str = "cpu"
     python: str = ""
     platform: str = ""
+    arch: str = ""
     error: str | None = None
     ffmpeg_ok: bool = False
     ffmpeg_path: str = ""

backend/core/version.py

@@ -12,4 +12,4 @@ from importlib.metadata import PackageNotFoundError, version
 try:
     APP_VERSION = version("omnivoice")
 except PackageNotFoundError:  # non-installed source checkout
-    APP_VERSION = "0.3.0"
+    APP_VERSION = "0.3.4"

backend/main.py

@@ -606,7 +606,7 @@ def health():
     elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
         device = "mps"
 
-    return {"status": "ok", "device": device}
+    return {"status": "ok", "device": device, "version": APP_VERSION}
 
 
 app.include_router(system.router)

backend/services/asr_backend.py

@@ -577,22 +577,32 @@ class PyTorchWhisperBackend(ASRBackend):
     def _ensure_pipe(self):
         if self._pipe is not None:
             return
-        # Fall back to grabbing the TTS model's ASR head.
-        import asyncio
-        from services.model_manager import get_model
-        try:
-            loop = asyncio.get_running_loop()
-            if loop.is_running():
-                raise RuntimeError(
-                    "PyTorchWhisperBackend needs the ASR pipe — pass it via constructor "
-                    "when calling from an async context."
-                )
-            model = loop.run_until_complete(get_model())
-        except RuntimeError:
-            model = asyncio.run(get_model())
-        self._pipe = getattr(model, "_asr_pipe", None)
-        if self._pipe is None:
-            raise RuntimeError("Loaded TTS model has no `_asr_pipe` attribute.")
+        # Build a standalone transformers Whisper pipeline on demand. This runs
+        # on PyTorch's own stack (cuDNN 9 ships with torch), so it works as a
+        # fallback on machines where WhisperX / faster-whisper can't load
+        # cuDNN 8 (the `cudnn_ops_infer64_8.dll` failure, issue #255) — and it
+        # needs neither OMNIVOICE_PRELOAD_TTS_ASR=1 nor a loaded TTS model.
+        # When the TTS model already has an ASR head, dub_core passes it via the
+        # constructor and this path is skipped.
+        import torch
+        from transformers import pipeline as hf_pipeline
+        from services.model_manager import get_best_device
+
+        model_name = os.environ.get(
+            "OMNIVOICE_PYTORCH_ASR_MODEL", "openai/whisper-large-v3-turbo"
+        )
+        device = get_best_device()
+        asr_dtype = torch.float16 if str(device).startswith("cuda") else torch.float32
+        logger.info(
+            "PyTorchWhisperBackend: loading standalone ASR pipeline %s on %s",
+            model_name, device,
+        )
+        self._pipe = hf_pipeline(
+            "automatic-speech-recognition",
+            model=model_name,
+            dtype=asr_dtype,
+            device_map=device,
+        )
 
     def transcribe(self, audio_path: str, *, word_timestamps: bool = True) -> dict:
         import soundfile as sf

deploy/Dockerfile

@@ -29,6 +29,13 @@ ENV HF_HOME=/app/omnivoice_data/huggingface
 # Allow bare imports (from core.config, from services.*, etc.) when
 # uvicorn is started as `backend.main:app` from WORKDIR /app.
 ENV PYTHONPATH=/app/backend
+# Headless server deployment: relax the desktop-only loopback origin gate.
+# Docker's network NAT rewrites the client host to the bridge gateway, so the
+# gate would otherwise 403 the operator out of /system/* and /api/settings/*
+# ("Loopback origin required", issue #261). Exposure is governed by the
+# operator's `-p` port mapping plus the optional share PIN. Desktop builds
+# never set this, so their loopback boundary is unchanged.
+ENV OMNIVOICE_SERVER_MODE=1
 
 # Install system dependencies (FFmpeg is critical for torchaudio/scene splitting)
 RUN apt-get update && apt-get install -y --no-install-recommends \

deploy/docker-compose.yml

@@ -46,6 +46,12 @@ services:
       # OMNIVOICE_BIND_HOST=0.0.0.0 here only opens the container's own
       # interface. The backend default is 127.0.0.1 (see backend/main.py).
       - OMNIVOICE_BIND_HOST=0.0.0.0
+      # Headless server: relax the desktop-only loopback origin gate so the
+      # web UI's /system/* and /api/settings/* routes work through Docker's
+      # NAT (issue #261). Already baked into the image; shown here so it's
+      # discoverable. If you front the container with your own auth proxy on
+      # loopback, set this to 0 to re-enable the strict gate.
+      - OMNIVOICE_SERVER_MODE=1
     healthcheck:
       test: ["CMD", "curl", "-sf", "http://localhost:3900/health"]
       interval: 30s
@@ -76,6 +82,9 @@ services:
       # service above. The host-side `127.0.0.1:3900:3900` mapping keeps
       # LAN reachability off by default.
       - OMNIVOICE_BIND_HOST=0.0.0.0
+      # See the CPU service above — relaxes the loopback origin gate for the
+      # headless Docker deployment (issue #261). Set to 0 to re-enable it.
+      - OMNIVOICE_SERVER_MODE=1
     healthcheck:
       test: ["CMD", "curl", "-sf", "http://localhost:3900/health"]
       interval: 30s

docs/install/docker.md

@@ -115,7 +115,18 @@ Two paths are worth persisting across container restarts:
 - **Container reports 0.2.7 but image is tagged 0.3.x:** This was a workflow bug
   (fixes #249, #251) — the `:latest` tag was not being updated on release tag
   pushes. Pull the image again after the fix is merged: `docker pull ghcr.io/debpalash/omnivoice-studio:latest`.
-- **Checking which version is running:** `docker exec omnivoice python -c "import importlib.metadata; print(importlib.metadata.version('omnivoice-studio'))"` or visit the `/health` endpoint which includes the version field.
+  The running version is now shown in **Settings → About → Version** (read live
+  from the backend), so the web UI no longer displays a dash in Docker.
+- **Checking which version is running:** `docker exec omnivoice python -c "import importlib.metadata; print(importlib.metadata.version('omnivoice'))"`, or hit the `/health` endpoint — it returns `{"status": "ok", "device": ..., "version": "0.3.x"}`.
+- **"Loopback origin required" errors (and a blank version):** the desktop
+  build restricts the `/system/*` and `/api/settings/*` routes to a loopback
+  origin, but Docker's NAT makes every request look non-loopback, so the gate
+  used to 403 the whole admin UI (issue #261). The image now ships with
+  `OMNIVOICE_SERVER_MODE=1`, which relaxes that gate for the headless
+  deployment — exposure is instead governed by your `-p` port mapping (keep the
+  `127.0.0.1:` prefix to stay local) plus the optional share PIN. If you front
+  the container with your own auth proxy on loopback, set `OMNIVOICE_SERVER_MODE=0`
+  to re-enable the strict gate.
 - **Media-preview 404 in LAN mode:** see the [LAN access](#lan-access) section
   above — the `window.location.host` fix shipped in v0.3.
 - **GPU not detected:** verify `docker run --rm --gpus all nvidia/cuda:12.8.0-base-ubuntu22.04 nvidia-smi` succeeds first.

docs/install/troubleshooting.md

@@ -121,7 +121,24 @@ falling back to faster-whisper`.
 path and is still fast. If you want the latest CT2 wheels, run `uv sync`
 from a fresh source checkout.
 
-## 10. IndexTTS / CosyVoice / ChatterboxTTS clash
+## 10. Windows: `Could not locate cudnn_ops_infer64_8.dll` during transcription
+
+**Symptom:** on Windows + NVIDIA, transcription/dubbing fails and the backend
+log shows `Could not locate cudnn_ops_infer64_8.dll`. Settings → Models shows
+WhisperX or faster-whisper selected.
+
+**Cause:** WhisperX and faster-whisper run on **CTranslate2**, which needs
+**cuDNN 8**, but PyTorch 2.8 ships cuDNN 9. OmniVoice side-loads a cuDNN-8 copy
+from `.venv\Lib\site-packages\cudnn8_compat\`; if that folder is missing
+(some upgrade paths don't install it), CTranslate2 can't find the DLL.
+
+**Fix:** switch the ASR backend to **PyTorch Whisper** in **Settings → Models**.
+It runs on PyTorch's own stack (cuDNN 9, bundled with torch) and needs no
+cuDNN-8 DLL — it loads its Whisper pipeline on demand (no extra env var). To
+keep using faster-whisper/WhisperX instead, reinstall to restore the bundled
+`cudnn8_compat` libraries.
+
+## 11. IndexTTS / CosyVoice / ChatterboxTTS clash
 
 **Symptom:** installing one of these engines breaks the others — e.g. after
 installing CosyVoice, IndexTTS errors out with import conflicts.

frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "omnivoice-studio",
   "private": true,
-  "version": "0.3.0",
+  "version": "0.3.4",
   "type": "module",
   "scripts": {
     "dev": "vite",

frontend/src-tauri/Cargo.lock

@@ -2878,7 +2878,7 @@ dependencies = [
 
 [[package]]
 name = "omnivoice-studio"
-version = "0.3.0"
+version = "0.3.4"
 dependencies = [
  "dirs-next",
  "enigo",

frontend/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "omnivoice-studio"
-version = "0.3.0"
+version = "0.3.4"
 description = "OmniVoice Studio – AI voice cloning & dubbing desktop app"
 authors = ["Debpalash"]
 license = "AGPL-3.0"

frontend/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "../node_modules/@tauri-apps/cli/config.schema.json",
   "productName": "OmniVoice Studio",
-  "version": "0.3.0",
+  "version": "0.3.4",
   "identifier": "com.debpalash.omnivoice-studio",
   "build": {
     "frontendDist": "../dist",

frontend/src/App.jsx

@@ -60,6 +60,7 @@ import { saveProject as apiSaveProject, loadProject as apiLoadProject, deletePro
 import { exportAction, exportReveal, exportRecord } from './api/exports';
 
 import { isTauri, doubleClickMaximize, fileToMediaUrl, playBlobAudio, playPing } from './utils/media';
+import { browserDownload } from './utils/download';
 import { checkForUpdate, fetchAppVersion } from './utils/updater';
 import { syncChannel } from './utils/channelControl';
 import i18n from './i18n';
@@ -506,6 +507,25 @@ function App() {
 
   const handleNativeExport = async (e, sourceIdentifier, fallbackName, mode) => {
     if (e) { e.preventDefault(); e.stopPropagation(); }
+    // Browser / Docker web build: there is no Tauri shell, so the native save
+    // dialog is unavailable — invoking it throws "Cannot read properties of
+    // undefined (reading 'invoke')" (issue #256). Fall back to a plain HTTP
+    // blob download of the file already served at /audio/<path>.
+    if (!isTauri) {
+      const niceName = (fallbackName || sourceIdentifier || 'audio').split('/').pop();
+      try {
+        const finalName = await browserDownload(`${API}/audio/${sourceIdentifier}`, niceName);
+        toast.success(i18n.t('app.toast_downloaded', { name: finalName }));
+        try {
+          await exportRecord({ filename: finalName, destination_path: `~/Downloads/${finalName}`, mode });
+          loadExportHistory();
+        } catch (err) { console.warn('exportRecord (browser export path) failed:', err); }
+      } catch (err) {
+        console.error(err);
+        toast.error(i18n.t('app.toast_export_failed', { message: err?.message || err }));
+      }
+      return;
+    }
     try {
       const { save } = await import('@tauri-apps/plugin-dialog');
       const ext = fallbackName.includes('.') ? fallbackName.split('.').pop() : 'wav';
@@ -527,14 +547,6 @@ function App() {
       toast.error(i18n.t('app.toast_open_folder_failed', { message: err.message }));
     }
   };
-  const parseFilenameFromContentDisposition = (header) => {
-    if (!header) return null;
-    const utf8 = header.match(/filename\*=(?:UTF-8|utf-8)''([^;]+)/i);
-    if (utf8) { try { return decodeURIComponent(utf8[1].trim().replace(/^"|"$/g, '')); } catch { /* ignore */ } }
-    const plain = header.match(/filename="?([^";]+)"?/i);
-    return plain ? plain[1].trim() : null;
-  };
-
   const triggerDownload = async (url, fallbackName) => {
     const extGuess = (fallbackName.includes('.') ? fallbackName.split('.').pop() : 'bin').toLowerCase();
     const modeGuess = ['mp4','mov','mkv','webm'].includes(extGuess)
@@ -573,19 +585,7 @@ function App() {
     // Browser path: standard blob download.
     try {
       toast.loading(i18n.t('app.toast_processing', { name: fallbackName }), { id: fallbackName });
-      const response = await fetch(url);
-      if (!response.ok) throw new Error("Download failed");
-      const serverName = parseFilenameFromContentDisposition(response.headers.get('content-disposition'));
-      const finalName = serverName || fallbackName || 'download';
-      const blob = await response.blob();
-      const localUrl = URL.createObjectURL(blob);
-      const a = document.createElement('a');
-      a.href = localUrl;
-      a.download = finalName;
-      document.body.appendChild(a);
-      a.click();
-      document.body.removeChild(a);
-      URL.revokeObjectURL(localUrl);
+      const finalName = await browserDownload(url, fallbackName);
       toast.success(i18n.t('app.toast_downloaded', { name: finalName }), { id: fallbackName });
       try {
         await exportRecord({ filename: finalName, destination_path: `~/Downloads/${finalName}`, mode: modeGuess });

frontend/src/api/types.ts

@@ -60,6 +60,7 @@ export interface SystemInfo {
   app_version?: string;
   python?: string;
   platform?: string;
+  arch?: string;
   device?: string;
   data_dir?: string;
   outputs_dir?: string;

frontend/src/pages/Settings.jsx

@@ -1352,10 +1352,10 @@ export default function Settings() {
         <section className="settings-section">
           <h2><Info size={16} color="#8ec07c" /> {t('settings.about')}</h2>
           <Row label={t('about.app')}             value="OmniVoice Studio" />
-          <Row label={t('about.version')}         value={appVersion || '—'} mono />
+          <Row label={t('about.version')}         value={appVersion || info?.app_version || '—'} mono />
           <Row label={t('about.tauri_runtime')}   value={tauriVersion || (isTauri() ? '—' : t('about.web_preview'))} mono />
           <Row label={t('about.platform')}        value={info?.platform || '—'} />
-          <Row label={t('about.architecture')}    value={typeof navigator !== 'undefined' ? (navigator.userAgentData?.platform || navigator.platform || '—') : '—'} mono />
+          <Row label={t('about.architecture')}    value={info?.arch || '—'} mono />
           <Row label={t('about.python')}          value={info?.python || '—'} mono />
           <Row label={t('about.compute_device')}  value={info?.device || '—'} mono />
           <Row label={t('about.gpu_active')}      value={hw?.gpu_active
@@ -1371,41 +1371,49 @@ export default function Settings() {
           <Row label={t('about.data_dir')}        value={info?.data_dir || '—'} mono />
           <Row label={t('about.outputs')}         value={info?.outputs_dir || '—'} mono />
           <Row label={t('about.crash_log')}       value={info?.crash_log_path || '—'} mono />
-          <Row
-            label={t('about.update_channel')}
-            value={
-              <Segmented
-                size="xs"
-                value={updateChannel}
-                onChange={changeChannel}
-                items={[
-                  { value: 'stable',  label: t('about.channel_stable') },
-                  { value: 'preview', label: t('about.channel_preview') },
-                ]}
+          {/* Auto-updater + channel toggle are desktop-only (Tauri). The Docker
+              web build updates by pulling a new image tag, so hide these rows
+              there to avoid a non-functional control (issue #249). */}
+          {isTauri() && (
+            <>
+              <Row
+                label={t('about.update_channel')}
+                value={
+                  <Segmented
+                    size="xs"
+                    value={updateChannel}
+                    onChange={changeChannel}
+                    items={[
+                      { value: 'stable',  label: t('about.channel_stable') },
+                      { value: 'preview', label: t('about.channel_preview') },
+                    ]}
+                  />
+                }
               />
-            }
-          />
-          <Row
-            label={t('about.update_endpoint')}
-            value={updateChannel === 'preview'
-              ? 'releases/download/preview/latest.json'
-              : 'releases/latest/download/latest.json'}
-            mono
-          />
-          {updateChannel === 'preview' && (
-            <p className="settings-muted">{t('about.channel_preview_hint')}</p>
+              <Row
+                label={t('about.update_endpoint')}
+                value={updateChannel === 'preview'
+                  ? 'releases/download/preview/latest.json'
+                  : 'releases/latest/download/latest.json'}
+                mono
+              />
+              {updateChannel === 'preview' && (
+                <p className="settings-muted">{t('about.channel_preview_hint')}</p>
+              )}
+            </>
           )}
           <div className="settings-link-row">
-            <Button
-              variant="primary"
-              size="md"
-              leading={<Download size={12} />}
-              onClick={checkForUpdates}
-              loading={updateState === 'checking' || updateState === 'downloading'}
-              disabled={!isTauri()}
-            >
-              {updateState === 'downloading' ? t('about.downloading') : t('about.check_updates')}
-            </Button>
+            {isTauri() && (
+              <Button
+                variant="primary"
+                size="md"
+                leading={<Download size={12} />}
+                onClick={checkForUpdates}
+                loading={updateState === 'checking' || updateState === 'downloading'}
+              >
+                {updateState === 'downloading' ? t('about.downloading') : t('about.check_updates')}
+              </Button>
+            )}
             <Button
               variant="subtle"
               size="md"

frontend/src/utils/download.js

@@ -0,0 +1,57 @@
+// Shared browser-download helpers.
+//
+// Used by every "save this file" path that runs outside the Tauri desktop
+// shell (browser dev mode + the Docker web-server build). In Tauri we use the
+// native save dialog instead; calling that dialog when no Tauri runtime is
+// present throws "Cannot read properties of undefined (reading 'invoke')"
+// (issue #256), so callers must guard on isTauri and route here otherwise.
+
+/**
+ * Parse a download filename out of a Content-Disposition header.
+ * Returns null when the header is absent or unparseable.
+ */
+export function parseFilenameFromContentDisposition(header) {
+  if (!header) return null;
+  const utf8 = header.match(/filename\*=(?:UTF-8|utf-8)''([^;]+)/i);
+  if (utf8) {
+    try {
+      return decodeURIComponent(utf8[1].trim().replace(/^"|"$/g, ''));
+    } catch {
+      /* fall through to the plain match */
+    }
+  }
+  const plain = header.match(/filename="?([^";]+)"?/i);
+  return plain ? plain[1].trim() : null;
+}
+
+/**
+ * Fetch `url` and trigger a standard browser blob download via a temporary
+ * <a download> element. Prefers the server-provided Content-Disposition
+ * filename, falling back to `fallbackName`. Returns the filename used.
+ *
+ * `deps` lets tests inject fetch/document/url without a real DOM + network.
+ */
+export async function browserDownload(url, fallbackName, deps = {}) {
+  const _fetch = deps.fetch ?? globalThis.fetch;
+  const doc = deps.document ?? globalThis.document;
+  const urlApi = deps.url ?? globalThis.URL;
+
+  const response = await _fetch(url);
+  if (!response.ok) throw new Error('Download failed');
+
+  const serverName = parseFilenameFromContentDisposition(
+    response.headers?.get?.('content-disposition'),
+  );
+  const finalName = serverName || fallbackName || 'download';
+
+  const blob = await response.blob();
+  const localUrl = urlApi.createObjectURL(blob);
+  const a = doc.createElement('a');
+  a.href = localUrl;
+  a.download = finalName;
+  doc.body.appendChild(a);
+  a.click();
+  doc.body.removeChild(a);
+  urlApi.revokeObjectURL(localUrl);
+  return finalName;
+}

frontend/src/utils/download.test.js

@@ -0,0 +1,66 @@
+import { describe, it, expect, vi } from 'vitest';
+import { parseFilenameFromContentDisposition, browserDownload } from './download';
+
+describe('parseFilenameFromContentDisposition', () => {
+  it('returns null for missing/empty headers', () => {
+    expect(parseFilenameFromContentDisposition(null)).toBe(null);
+    expect(parseFilenameFromContentDisposition('')).toBe(null);
+    expect(parseFilenameFromContentDisposition('inline')).toBe(null);
+  });
+
+  it('parses a plain filename', () => {
+    expect(parseFilenameFromContentDisposition('attachment; filename="clip.wav"')).toBe('clip.wav');
+    expect(parseFilenameFromContentDisposition('attachment; filename=clip.wav')).toBe('clip.wav');
+  });
+
+  it('prefers and decodes the RFC 5987 UTF-8 form', () => {
+    expect(parseFilenameFromContentDisposition("attachment; filename*=UTF-8''my%20clip.wav")).toBe('my clip.wav');
+  });
+});
+
+describe('browserDownload', () => {
+  function makeDeps({ ok = true, disposition = null } = {}) {
+    const anchor = { href: '', download: '', click: vi.fn() };
+    const body = { appendChild: vi.fn(), removeChild: vi.fn() };
+    const fetch = vi.fn(async () => ({
+      ok,
+      headers: { get: () => disposition },
+      blob: async () => new Blob(['data']),
+    }));
+    const document = { createElement: vi.fn(() => anchor), body };
+    const url = { createObjectURL: vi.fn(() => 'blob:local'), revokeObjectURL: vi.fn() };
+    return { deps: { fetch, document, url }, anchor, body, fetch, url };
+  }
+
+  it('downloads via a temporary <a> using the fallback name', async () => {
+    const { deps, anchor, url } = makeDeps();
+    const name = await browserDownload('http://x/audio/foo.wav', 'foo.wav', deps);
+    expect(name).toBe('foo.wav');
+    expect(anchor.download).toBe('foo.wav');
+    expect(anchor.href).toBe('blob:local');
+    expect(anchor.click).toHaveBeenCalledOnce();
+    expect(url.revokeObjectURL).toHaveBeenCalledWith('blob:local');
+  });
+
+  it('prefers the server-provided Content-Disposition filename', async () => {
+    const { deps, anchor } = makeDeps({ disposition: 'attachment; filename="server.mp3"' });
+    const name = await browserDownload('http://x/audio/foo.wav', 'foo.wav', deps);
+    expect(name).toBe('server.mp3');
+    expect(anchor.download).toBe('server.mp3');
+  });
+
+  it('throws when the response is not ok (so callers can surface an error toast)', async () => {
+    const { deps } = makeDeps({ ok: false });
+    await expect(browserDownload('http://x/missing', 'foo.wav', deps)).rejects.toThrow('Download failed');
+  });
+
+  // Regression for #256: in the Docker/browser build there is no Tauri shell,
+  // so the download path must never call the native save dialog (which throws
+  // "Cannot read properties of undefined (reading 'invoke')"). This helper is
+  // pure HTTP + DOM and works without any Tauri runtime present.
+  it('works with no Tauri globals defined', async () => {
+    expect(typeof window === 'undefined' || window.__TAURI_INTERNALS__).toBeFalsy();
+    const { deps } = makeDeps();
+    await expect(browserDownload('http://x/audio/foo.wav', 'foo.wav', deps)).resolves.toBe('foo.wav');
+  });
+});

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "omnivoice"
-version = "0.3.0"
+version = "0.3.4"
 description = "OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models"
 readme = "README.md"
 # Source-available under FSL-1.1-ALv2 (see LICENSE); each release converts to

tests/test_dub_transcribe.py

@@ -109,6 +109,43 @@ def _seed_job(dc_module, tmp_path: Path, duration: float, scene_cuts=None) -> st
 # Tests
 # ---------------------------------------------------------------------------
 
+def test_transcribe_stream_surfaces_model_load_failure(tmp_path, monkeypatch):
+    """Regression #255: when the model fails to load, the SSE transcribe stream
+    must emit a structured `error` event carrying the real cause — not silently
+    drop the connection (the UI renders a dropped stream as a misleading generic
+    "Transcribe stream dropped … Likely ASR backend failed to load").
+
+    Drives the route's async generator directly (no TestClient/lifespan) — the
+    preflight-error path yields a single event with no executor/Queue, so it
+    stays isolated from the app event loop.
+    """
+    import asyncio
+    from api.routers import dub_core as dc
+
+    job_id = "t_modelfail"
+    dc._dub_jobs[job_id] = {"audio_path": str(tmp_path / "a.wav"), "vocals_path": None}
+
+    async def _boom():
+        raise RuntimeError("CUDA driver init failed: simulated")
+
+    monkeypatch.setattr(dc, "get_model", _boom)
+
+    async def _collect():
+        resp = await dc.dub_transcribe_stream(job_id)
+        parts = []
+        async for chunk in resp.body_iterator:
+            parts.append(chunk.decode() if isinstance(chunk, (bytes, bytearray)) else str(chunk))
+        return "".join(parts)
+
+    try:
+        body = asyncio.run(_collect())
+    finally:
+        dc._dub_jobs.pop(job_id, None)
+
+    assert "event: error" in body, body
+    assert "CUDA driver init failed: simulated" in body, body
+
+
 @pytest.mark.xfail(
     reason="dub_core._transcribe was refactored to route through "
            "services.asr_backend.get_active_asr_backend; the MagicMock fixture "

tests/test_loopback_server_mode.py

@@ -0,0 +1,55 @@
+"""`require_loopback` gate contract (issue #261).
+
+The gate must stay strict on the desktop build (non-loopback → 403, which is the
+PR #81 trust boundary), but become a no-op in the headless Docker server mode,
+where Docker's NAT makes the loopback origin unenforceable and exposure is
+governed by the port mapping + the share PIN instead.
+"""
+from types import SimpleNamespace
+
+import pytest
+from fastapi import HTTPException
+
+from api.dependencies import require_loopback
+
+
+def _req(host):
+    """Minimal stand-in for a Starlette Request — the gate only reads client.host."""
+    return SimpleNamespace(client=SimpleNamespace(host=host) if host else None)
+
+
+@pytest.fixture(autouse=True)
+def _clear_server_mode(monkeypatch):
+    # Start each test from the desktop default regardless of the ambient env.
+    monkeypatch.delenv("OMNIVOICE_SERVER_MODE", raising=False)
+
+
+@pytest.mark.parametrize("host", ["127.0.0.1", "::1", "localhost"])
+def test_loopback_always_allowed(host):
+    require_loopback(_req(host))  # must not raise
+
+
+def test_non_loopback_rejected_by_default():
+    with pytest.raises(HTTPException) as exc:
+        require_loopback(_req("172.17.0.1"))  # Docker bridge gateway
+    assert exc.value.status_code == 403
+    assert "loopback" in str(exc.value.detail).lower()
+
+
+def test_missing_client_rejected_by_default():
+    with pytest.raises(HTTPException):
+        require_loopback(_req(None))
+
+
+@pytest.mark.parametrize("val", ["1", "true", "TRUE", "yes", "on"])
+def test_server_mode_allows_non_loopback(monkeypatch, val):
+    monkeypatch.setenv("OMNIVOICE_SERVER_MODE", val)
+    require_loopback(_req("172.17.0.1"))  # must not raise
+    require_loopback(_req("127.0.0.1"))   # loopback still fine
+
+
+@pytest.mark.parametrize("val", ["0", "false", "no", "", "off"])
+def test_falsey_server_mode_keeps_gate_strict(monkeypatch, val):
+    monkeypatch.setenv("OMNIVOICE_SERVER_MODE", val)
+    with pytest.raises(HTTPException):
+        require_loopback(_req("10.0.0.5"))

tests/test_pytorch_whisper_fallback.py

@@ -0,0 +1,83 @@
+"""PyTorch-Whisper backend must work as a standalone fallback (issue #255).
+
+On machines where WhisperX / faster-whisper can't load cuDNN 8
+(`cudnn_ops_infer64_8.dll` missing), the PyTorch-Whisper backend should build
+its own transformers pipeline on demand — without OMNIVOICE_PRELOAD_TTS_ASR=1
+and without loading the full TTS model.
+"""
+import sys
+import types
+
+import pytest
+
+from services import asr_backend as ab
+
+
+def test_is_available_when_transformers_present():
+    ok, msg = ab.PyTorchWhisperBackend.is_available()
+    assert ok is True
+    assert msg == "ready"
+
+
+def test_reuses_constructor_pipe_without_building(monkeypatch):
+    sentinel = object()
+    be = ab.PyTorchWhisperBackend(asr_pipe=sentinel)
+
+    def _boom(*a, **k):
+        raise AssertionError("must not build a pipeline when one was passed in")
+
+    # transformers.pipeline is imported lazily inside _ensure_pipe.
+    fake_tf = types.ModuleType("transformers")
+    fake_tf.pipeline = _boom
+    monkeypatch.setitem(sys.modules, "transformers", fake_tf)
+
+    be._ensure_pipe()
+    assert be._pipe is sentinel
+
+
+def test_lazy_builds_standalone_pipeline(monkeypatch):
+    """No preloaded pipe → build a standalone transformers ASR pipeline, with no
+    call into the TTS model loader (get_model)."""
+    captured = {}
+
+    def fake_pipeline(task, **kw):
+        captured["task"] = task
+        captured["kw"] = kw
+        return lambda *a, **k: {"chunks": []}
+
+    fake_tf = types.ModuleType("transformers")
+    fake_tf.pipeline = fake_pipeline
+    monkeypatch.setitem(sys.modules, "transformers", fake_tf)
+    monkeypatch.setattr("services.model_manager.get_best_device", lambda: "cpu")
+
+    # Guard: building the standalone pipe must NOT pull in the full TTS model.
+    import services.model_manager as mm
+
+    def _no_get_model(*a, **k):
+        raise AssertionError("standalone ASR build must not call get_model()")
+
+    monkeypatch.setattr(mm, "get_model", _no_get_model, raising=False)
+
+    be = ab.PyTorchWhisperBackend(asr_pipe=None)
+    be._ensure_pipe()
+
+    assert be._pipe is not None
+    assert captured["task"] == "automatic-speech-recognition"
+    assert captured["kw"]["model"]  # a concrete model name was chosen
+
+
+def test_pytorch_asr_model_overridable_via_env(monkeypatch):
+    captured = {}
+
+    def fake_pipeline(task, **kw):
+        captured["kw"] = kw
+        return object()
+
+    fake_tf = types.ModuleType("transformers")
+    fake_tf.pipeline = fake_pipeline
+    monkeypatch.setitem(sys.modules, "transformers", fake_tf)
+    monkeypatch.setattr("services.model_manager.get_best_device", lambda: "cpu")
+    monkeypatch.setenv("OMNIVOICE_PYTORCH_ASR_MODEL", "openai/whisper-small")
+
+    ab.PyTorchWhisperBackend(asr_pipe=None)._ensure_pipe()
+    assert captured["kw"]["model"] == "openai/whisper-small"

tests/test_router_smoke.py

@@ -33,6 +33,22 @@ def test_system_info_smoke(client):
     body = r.json()
     assert "data_dir" in body
     assert "device" in body
+    # The Docker/web build has no Tauri getVersion(); Settings → About reads the
+    # running version from here so it shows the real version, not a dash (#249).
+    from core.version import APP_VERSION
+    assert body["app_version"] == APP_VERSION
+    # Settings → About → Architecture must reflect the SERVER's machine, not the
+    # client browser's navigator.platform (which showed "Win32" in Docker, #262).
+    import platform as _pf
+    assert body["arch"] == _pf.machine()
+
+
+def test_health_exposes_version(client):
+    """`/health` is the zero-auth way to confirm the running version (#249)."""
+    from core.version import APP_VERSION
+    body = client.get("/health").json()
+    assert body["status"] == "ok"
+    assert body["version"] == APP_VERSION
 
 
 def test_system_logs_smoke(client):

uv.lock

@@ -3058,7 +3058,7 @@ wheels = [
 
 [[package]]
 name = "omnivoice"
-version = "0.3.0"
+version = "0.3.4"
 source = { editable = "." }
 dependencies = [
     { name = "accelerate" },