Style disabled primary actions and native scrollbar tracks

Say what the local models are running on

.github/workflows/ui-screenshots.yml

@@ -0,0 +1,58 @@
+name: UI screenshots
+
+on:
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  capture:
+    timeout-minutes: 10
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            platform: xcb
+          - os: macos-latest
+            platform: cocoa
+          - os: windows-latest
+            platform: windows
+    runs-on: ${{ matrix.os }}
+    env:
+      QT_QPA_PLATFORM: ${{ matrix.platform }}
+      QT_SCALE_FACTOR: "2"
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.13"
+      - name: Install Linux display libraries
+        if: runner.os == 'Linux'
+        timeout-minutes: 5
+        run: |
+          apt="-o Acquire::Retries=3 -o Acquire::http::Timeout=20"
+          sudo apt-get $apt update
+          sudo apt-get $apt install --no-install-recommends -y \
+            xvfb xauth libegl1 libgl1 libxkbcommon0 libdbus-1-3 \
+            libglib2.0-0 libfontconfig1 libfreetype6 libgssapi-krb5-2 \
+            libxcb-cursor0 libxkbcommon-x11-0 libxcb-icccm4 \
+            libxcb-keysyms1 libxcb-shape0 libxcb-xinerama0 libxcb-render-util0
+      - name: Install PyQt6
+        run: python -m pip install --quiet PyQt6
+      - name: Capture with X11
+        if: runner.os == 'Linux'
+        run: xvfb-run -a -s "-screen 0 2400x1600x24" python -m tests.render_ui --output ui-screenshots --expect-platform xcb
+      - name: Capture with the native platform
+        if: runner.os != 'Linux'
+        run: python -m tests.render_ui --output ui-screenshots --expect-platform ${{ matrix.platform }}
+      - name: Upload screenshots and rendering details
+        if: ${{ !cancelled() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: ui-screenshots-${{ runner.os }}
+          path: ui-screenshots/
+          if-no-files-found: error
+          retention-days: 14

README.md

@@ -13,11 +13,17 @@ library, 3.11 or newer, and PyQt6.
 *[Türkçe README](README.tr.md)*
 
 <p align="center">
-  <img src="docs/settings-general.webp" width="820" alt="Dikte settings, General tab">
+  <img src="docs/home.webp" width="620" alt="Dikte, Nord theme">
+  <br><sub>Nord (default)</sub>
 </p>
 
+| Dracula | Classic dark | Classic light |
+|---|---|---|
+| <img src="docs/home-dracula.webp" width="270" alt="Dikte, Dracula"> | <img src="docs/home-dark.webp" width="270" alt="Dikte, Classic dark"> | <img src="docs/home-light.webp" width="270" alt="Dikte, Classic light"> |
+
 |  |  |
 |---|---|
+| <img src="docs/settings-general.webp" width="410" alt="General and themes"> | <img src="docs/settings-display.webp" width="410" alt="Nord, Dracula, dark, light"> |
 | <img src="docs/settings-api.webp" width="410" alt="API and models"> | <img src="docs/settings-cleanup.webp" width="410" alt="Cleanup rules"> |
 | <img src="docs/settings-agent.webp" width="410" alt="Agent"> | <img src="docs/settings-meeting.webp" width="410" alt="Meeting"> |
 | <img src="docs/settings-audio-file.webp" width="410" alt="Audio file"> | <img src="docs/settings-shortcuts.webp" width="410" alt="Shortcuts"> |
@@ -220,6 +226,11 @@ running.
   written for subtitles, so the lines keep their place and nothing is shortened.
 - **History** of every dictation under Settings → History, with a size limit and
   right-click to delete.
+- **The speech language is detected, not picked.** Auto is the default: whisper
+  on this machine says what it heard, the hosted providers transcribe in
+  whatever language comes in without being told, and the detected language
+  lands in the history and decides which cleanup prompt (Turkish or the
+  language-agnostic one) a run gets. A fixed language still overrides it.
 - **Turkish and English interface**, following the system locale by default.
 
 ## The global shortcuts, and the logout KDE needs

README.tr.md

@@ -12,11 +12,17 @@ Python standart kütüphanesi (3.11 veya üstü) ve PyQt6.
 *[English README](README.md)*
 
 <p align="center">
-  <img src="docs/settings-general.webp" width="820" alt="Dikte ayarları, Genel sekmesi">
+  <img src="docs/home.webp" width="620" alt="Dikte, Nord teması">
+  <br><sub>Nord (varsayılan)</sub>
 </p>
 
+| Dracula | Klasik karanlık | Klasik beyaz |
+|---|---|---|
+| <img src="docs/home-dracula.webp" width="270" alt="Dikte, Dracula"> | <img src="docs/home-dark.webp" width="270" alt="Dikte, Klasik karanlık"> | <img src="docs/home-light.webp" width="270" alt="Dikte, Klasik beyaz"> |
+
 |  |  |
 |---|---|
+| <img src="docs/settings-general.webp" width="410" alt="Genel ve temalar"> | <img src="docs/settings-display.webp" width="410" alt="Nord, Dracula, dark, light"> |
 | <img src="docs/settings-api.webp" width="410" alt="API ve modeller"> | <img src="docs/settings-cleanup.webp" width="410" alt="Temizleme kuralları"> |
 | <img src="docs/settings-agent.webp" width="410" alt="Ajan"> | <img src="docs/settings-meeting.webp" width="410" alt="Toplantı"> |
 | <img src="docs/settings-audio-file.webp" width="410" alt="Ses dosyası"> | <img src="docs/settings-shortcuts.webp" width="410" alt="Kısayollar"> |
@@ -214,6 +220,11 @@ olmasını ister.
   yerinde kalır, hiçbir şey kısaltılmaz.
 - **Geçmiş** Ayarlar → Geçmiş sekmesinde; boyut sınırı var, sağ tıklayıp
   silebilirsin.
+- **Konuşma dili seçilmez, algılanır.** Varsayılan otomatiktir: bu makinedeki
+  whisper ne duyduğunu söyler, bulut sağlayıcılar söylenmeden de hangi dilde
+  konuşuluyorsa o dilde yazar; algılanan dil geçmişe düşer ve bir kaydın hangi
+  temizleme promptunu alacağını belirler (Türkçe mi, dile duyarsız olanı mı).
+  Sabit bir dil yine de bunun önüne geçer.
 - **Türkçe ve İngilizce arayüz**, varsayılan olarak sistem dilini izler.
 
 ## Global kısayollar ve KDE'nin istediği oturum kapatma

dikte/__init__.py

@@ -10,4 +10,4 @@ business loading Qt to answer one question.
 # both the .dmg's Info.plist and the AppImage's file name are built from it. A
 # build off master rather than off a tag appends the commit to it, so that a
 # bug report from someone running "latest" names a commit.
-__version__ = "1.2.0"
+__version__ = "1.3.0"

dikte/api.py

@@ -367,7 +367,8 @@ def local_failure(service, server, exc):
 
 
 def _transcribe_request(target, audio_path, language, prompt, response_format,
-                        granularity=None, timeout=300, aborter=None):
+                        granularity=None, timeout=300, aborter=None,
+                        detect_language=False):
     if target.provider == "local":
         # The timeouts here are sized for a hosted API, where a slow answer is a
         # bill running. Locally the only thing being spent is time.
@@ -379,6 +380,12 @@ def _transcribe_request(target, audio_path, language, prompt, response_format,
     fields = [("model", target.model), ("response_format", response_format)]
     if language and language != "auto":
         fields.append(("language", language))
+    if detect_language:
+        # whisper.cpp was started with -nlp, which keeps the language
+        # probability sweep off every request. Detection is only worth that
+        # sweep for the run that asked for it, so it is switched back on here,
+        # per request, and reported in the verbose_json answer.
+        fields.append(("no_language_probabilities", "false"))
     # OpenRouter takes the hint field and throws it away, so spare it the bytes.
     # The same words still reach the cleanup model as a glossary. whisper.cpp
     # takes it as the initial prompt, the way OpenAI does.
@@ -546,6 +553,44 @@ def transcribe(target, audio_path, language="", prompt="", timeout=300, aborter=
     return text
 
 
+# whisper.cpp reports what it heard as a lowercase full name ("turkish",
+# "english", "german"…); the settings and the cleanup prompt speak in two-letter
+# codes. Only the handful Dikte offers as a fixed choice get a code; anything
+# else is left as the empty string, which the caller reads as "unknown" rather
+# than guessing at a language it has no label for.
+_DETECTED_TO_CODE = {
+    "english": "en", "turkish": "tr", "german": "de",
+    "french": "fr", "spanish": "es", "arabic": "ar",
+}
+
+
+def transcribe_detected(target, audio_path, language="", prompt="", timeout=300,
+                        aborter=None):
+    """(text, code) with the language the model heard.
+
+    The spoken language is only knowable when the transcription model reports
+    it, and only whisper.cpp does: the hosted endpoints accept "auto" but never
+    say what they heard. So detection is asked for exactly where it can be
+    answered, the local server in auto mode, and every other run transcribes
+    as before and hands back an empty code.
+    """
+    if target.provider == "local" and language == "auto":
+        data = _transcribe_request(
+            target, audio_path, language, prompt, "verbose_json",
+            detect_language=True, timeout=timeout, aborter=aborter,
+        )
+        text = _local_text(data.get("text") or "").strip()
+        if not text:
+            raise ApiError(t("Transcript came back empty."))
+        detected = data.get("detected_language")
+        code = _DETECTED_TO_CODE.get(
+            detected.strip().lower(), "") if isinstance(detected, str) else ""
+        return text, code
+    text = transcribe(target, audio_path, language=language, prompt=prompt,
+                      timeout=timeout, aborter=aborter)
+    return text, ""
+
+
 def transcribe_segments(target, audio_path, language="", prompt="", timeout=300,
                         aborter=None):
     """[(start_seconds, end_seconds, text)] using whisper-1's verbose response."""

dikte/app.py

@@ -58,6 +58,7 @@ from .i18n import t  # noqa: E402
 from .meeting import MeetingPipeline  # noqa: E402
 from .overlay import Overlay  # noqa: E402
 from .settings_ui import SettingsWindow  # noqa: E402
+from .home_ui import HomeWindow  # noqa: E402
 from .worker import Pipeline  # noqa: E402
 
 SERVER_NAME = ipc.SERVER_NAME
@@ -138,6 +139,8 @@ class Dikte:
         self.meeting_base = ""
         self.meeting_message = ""
         self.settings_window = None
+        self.home_window = None
+        self.home_messages = {}
         # The single-instance server, handed over once run_app has opened it, so
         # that a restart can stop answering before the replacement starts.
         self.server = None
@@ -165,13 +168,15 @@ class Dikte:
 
         self.overlay = Overlay(self.conf["overlay_corner"],
                                screen_name=self.conf["overlay_screen"],
-                               follow_pointer=self.conf["overlay_follows_pointer"])
+                               follow_pointer=self.conf["overlay_follows_pointer"],
+                               theme_name=self.conf["theme"])
         # The agent's indicator sits on top of the dictation one when both are
         # up, and drops into the corner when it is alone there.
         self.ask_overlay = Overlay(self.conf["overlay_corner"], below=self.overlay,
                                    dismissable=True,
                                    screen_name=self.conf["overlay_screen"],
-                                   follow_pointer=self.conf["overlay_follows_pointer"])
+                                   follow_pointer=self.conf["overlay_follows_pointer"],
+                               theme_name=self.conf["theme"])
         self.recorder = audio.Recorder()
         self.pipeline = Pipeline(self.conf)
         self.ask_pipeline = Pipeline(self.conf)
@@ -195,7 +200,7 @@ class Dikte:
         self.pipeline.stage.connect(self._on_stage)
         self.pipeline.finished.connect(self._on_finished)
         self.pipeline.failed.connect(self._on_pipeline_failed)
-        self.ask_pipeline.stage.connect(self.ask_overlay.show_busy)
+        self.ask_pipeline.stage.connect(self._on_ask_stage)
         self.ask_pipeline.finished.connect(self._on_ask_finished)
         self.ask_pipeline.failed.connect(self._on_ask_error)
         self.ask_pipeline.cancelled.connect(self._on_ask_cancelled)
@@ -294,6 +299,10 @@ class Dikte:
         self.unload_action.triggered.connect(self.unload_models)
         self.menu.addAction(self.unload_action)
 
+        self.home_action = QAction(t("Open Dikte"), self.menu)
+        self.home_action.triggered.connect(self.open_home)
+        self.menu.addAction(self.home_action)
+
         self.settings_action = QAction(t("Settings…"), self.menu)
         self.settings_action.triggered.connect(self.open_settings)
         self.menu.addAction(self.settings_action)
@@ -524,7 +533,7 @@ class Dikte:
     # than only able to press its buttons.
 
     def handle(self, request, reply):
-        cmd = str(request.get("cmd") or "settings").strip()
+        cmd = str(request.get("cmd") or "home").strip()
         if cmd in ("toggle", "start", "stop", "record"):
             self._dictation_request(cmd, request, reply)
         elif cmd == "ask":
@@ -541,6 +550,7 @@ class Dikte:
                 "ask-reset": self.reset_conversation,
                 "meeting-cancel": self.cancel_meeting,
                 "settings": self.open_settings,
+                "home": self.open_home,
                 "reload": self.reload_settings,
                 "restart": self.restart,
                 "quit": self.app.quit,
@@ -609,9 +619,28 @@ class Dikte:
 
     def _settle(self, kind, payload):
         """Tell whoever was waiting on this run how it ended."""
+        self._home_settled(kind, payload)
         for reply in self._waiters.pop(kind, []):
             reply(payload)
 
+    def _home_settled(self, kind, payload):
+        if not hasattr(self, "home_messages"):
+            self.home_messages = {}
+        self.home_messages.pop(kind + "_stage", None)
+        if payload.get("cancelled"):
+            message = t("Stopped.")
+        elif payload.get("error"):
+            message = t("Failed: {error}", error=payload["error"])
+        elif payload.get("warning"):
+            message = t("Completed with a warning: {error}", error=payload["warning"])
+        else:
+            message = t("Transcript ready") if kind == DICTATION else ""
+        self.home_messages[kind] = message
+        window = getattr(self, "home_window", None)
+        if window is not None:
+            window.refresh_results()
+            window.refresh()
+
     def _auto_stop(self, run):
         """The end of a `record --seconds`, if that recording is still the one."""
         if self._run_id == run and self.state == RECORDING:
@@ -630,6 +659,10 @@ class Dikte:
             "agent": assistant.display_name(self.conf),
             "provider": assistant.provider(self.conf),
             "listener": self.evdev.running,
+            # Whether each model on this machine is loaded, and what it ended up
+            # running on. Only this process knows: the servers are its children,
+            # and the command line has no way to ask them anything.
+            "local": self._local_state(),
             # Asked here rather than by the command line, because on macOS
             # there is no registry to read: a combination is held by this
             # process and by nothing else, so this is the only process that
@@ -638,9 +671,22 @@ class Dikte:
                           for name, spec in hotkey.SHORTCUTS.items()},
         }
 
+    def _local_state(self):
+        """ggml.state(), with a mark for the servers this setup actually uses.
+
+        A server that is neither wanted nor loaded is not worth a line anywhere;
+        one that is wanted and not loaded is exactly the line worth reading.
+        """
+        local = ggml.state()
+        local["whisper"]["used"] = self.conf["transcribe_provider"] == "local"
+        local["llama"]["used"] = self.conf.uses_local_llm()
+        return local
+
     def reload_settings(self):
         """Read the config file back after something outside changed it."""
         self.conf.load()
+        if self.settings_window is not None:
+            self.settings_window.refresh_configuration()
         self._apply_settings()
 
     def _toggle(self):
@@ -691,6 +737,9 @@ class Dikte:
         # recorder's.
         if self.state == RECORDING or self.recording:
             return
+        if isinstance(getattr(self, "home_messages", None), dict):
+            self.home_messages[DICTATION] = ""
+            self.home_messages.pop("dictation_stage", None)
         self.front_before = self._the_front()
         self.overlay.show_recording()
         self._begin_recording(DICTATION)
@@ -705,6 +754,9 @@ class Dikte:
     def start_ask(self):
         if self.ask_state != IDLE or self.recording:
             return
+        if isinstance(getattr(self, "home_messages", None), dict):
+            self.home_messages[ASK] = ""
+            self.home_messages.pop("ask_stage", None)
         self.front_before = self._the_front()
         self.ask_overlay.show_recording(asking=True)
         self._begin_recording(ASK)
@@ -1023,7 +1075,7 @@ class Dikte:
         self.overlay.show_error(t("Meeting failed: {error}", error=first_line))
         self.tray.showMessage(
             t("Dikte: the meeting could not be written up"),
-            t("{error}\n\nThe recording has been kept. Settings → Minutes can "
+            t("{error}\n\nThe recording has been kept. Meeting → Minutes can "
               "try again.", error=error),
             QSystemTrayIcon.MessageIcon.Warning, 12000,
         )
@@ -1065,7 +1117,14 @@ class Dikte:
             self.pipeline.run(wav_path, duration, rms_values,
                               paste=wants_paste, focus=focus)
 
+    def _on_ask_stage(self, message):
+        if isinstance(getattr(self, "home_messages", None), dict):
+            self.home_messages["ask_stage"] = message
+        self.ask_overlay.show_busy(message)
+
     def _on_stage(self, message):
+        if isinstance(getattr(self, "home_messages", None), dict):
+            self.home_messages["dictation_stage"] = message
         # The corner belongs to the recording when one is on: the previous
         # run's progress must not wipe the waveform mid-sentence.
         if self.state != RECORDING:
@@ -1084,30 +1143,30 @@ class Dikte:
         if not self._transcripts_pending:
             self._settle(DICTATION, payload)
 
-    def _on_finished(self, _raw, text, warning):
+    def _on_finished(self, _raw, text, warning, speech_language):
         if warning:
             # The text was still pasted, but cleanup did not run. Say so loudly:
             # a rejected key otherwise looks exactly like working dictation.
             if self.state != RECORDING:
                 self.overlay.show_warning(
-                    t("Pasted raw, cleanup failed: {error}",
+                    t("Completed with a warning: {error}",
                       error=warning.splitlines()[0])
                 )
             self.tray.showMessage(
-                t("Dikte: cleanup failed"), warning,
+                t("Dikte: completed with a warning"), warning,
                 QSystemTrayIcon.MessageIcon.Warning, 10000,
             )
         elif self.state != RECORDING:
             # While a new recording is on, the flash is skipped: the text
             # arriving where the cursor is says everything it would have.
-            action = t("Pasted") if self.conf["auto_paste"] else t("Copied")
             self.overlay.show_done(
-                t("{action}: {preview}", action=action, preview=_preview(text))
+                t("Transcript ready: {preview}", preview=_preview(text))
             )
         self._transcript_settled({"ok": True, "text": text, "raw": _raw,
-                                  "warning": warning})
+                                  "warning": warning,
+                                  "speech_language": speech_language})
 
-    def _on_ask_finished(self, _raw, text, warning):
+    def _on_ask_finished(self, _raw, text, warning, speech_language):
         agent = assistant.display_name(self.conf)
         if warning:
             # A tool the agent was not allowed to touch otherwise looks exactly
@@ -1128,7 +1187,8 @@ class Dikte:
             )
         self._set_ask_state(IDLE)
         self._settle(ASK, {"ok": True, "answer": text, "question": _raw,
-                           "warning": warning, "agent": agent})
+                           "warning": warning, "agent": agent,
+                           "speech_language": speech_language})
 
     def _on_ask_cancelled(self):
         self.ask_overlay.show_done(t("Stopped."), 2000)
@@ -1247,9 +1307,21 @@ class Dikte:
 
     # ---- settings ---------------------------------------------------------
 
+    def open_home(self):
+        if self.settings_window is None:
+            self._make_settings()
+        if getattr(self, "home_window", None) is None:
+            self.home_window = HomeWindow(self, self.settings_window)
+        self.home_window.show()
+        self.home_window.raise_()
+        self.home_window.activateWindow()
+
     def open_settings(self):
         if self.settings_window is None:
             self._make_settings()
+        else:
+            self.settings_window.refresh_configuration()
+            self.settings_window.refresh_sources()
         self.settings_window.show()
         self.settings_window.raise_()
         self.settings_window.activateWindow()
@@ -1264,8 +1336,9 @@ class Dikte:
         self.settings_window.finished.connect(self._settings_closed)
 
     def _settings_closed(self, *_):
-        # Don't drop the object while its own signal is still being delivered.
+        # Task pages share this controller and may still be processing a file.
-        QTimer.singleShot(0, lambda: setattr(self, "settings_window", None))
+        # Closing the configuration dialog keeps both its edits and jobs alive.
+        pass
 
     def _reopen_settings(self):
         """Replace the settings window, so a language change reaches it too.
@@ -1278,8 +1351,15 @@ class Dikte:
         old one stood, on the same tab.
         """
         old = self.settings_window
-        if old is None:
+        if old is None or old._work_in_flight():
             return
+        home = getattr(self, "home_window", None)
+        home_visible = home is not None and home.isVisible()
+        home_mode = home.mode if home is not None else "dictation"
+        home_geometry = home.geometry() if home is not None else None
+        if home is not None:
+            home.close()
+            self.home_window = None
         tab = old.tabs.currentIndex()
         geometry = old.geometry()
         # Replaced rather than merely closed: left connected, _settings_closed
@@ -1296,9 +1376,28 @@ class Dikte:
         # window does not come up at the default size and jump.
         self.settings_window.setGeometry(geometry)
         self.settings_window.tabs.setCurrentIndex(tab)
+        self.settings_window.file_path = old.file_path
+        self.settings_window.file_label.setText(old.file_label.text())
+        self.settings_window.file_output.setPlainText(old.file_output.toPlainText())
+        self.settings_window.file_segments = getattr(old, "file_segments", [])
+        self.settings_window.file_save_srt.setEnabled(bool(self.settings_window.file_segments))
+        self.settings_window.file_status.setText(old.file_status.text())
         self.settings_window.show()
         self.settings_window.raise_()
         self.settings_window.activateWindow()
+        for signal, callback in (
+            (self.meetings.progress, old._on_minutes_progress),
+            (self.meetings.finished, old._on_minutes_finished),
+            (self.meetings.failed, old._on_minutes_failed),
+        ):
+            signal.disconnect(callback)
+        if home is not None:
+            self.home_window = HomeWindow(self, self.settings_window)
+            self.home_window.setGeometry(home_geometry)
+            self.home_window.show_mode(home_mode)
+            if home_visible:
+                self.home_window.show()
+            home.deleteLater()
 
     def _apply_local(self):
         """Pass the local settings on, and hold the models ready if asked to.
@@ -1336,10 +1435,13 @@ class Dikte:
 
     def _apply_settings(self):
         for indicator in (self.overlay, self.ask_overlay):
+            indicator.set_theme(self.conf["theme"])
             indicator.corner = self.conf["overlay_corner"]
             indicator.screen_name = self.conf["overlay_screen"]
             indicator.follow_pointer = self.conf["overlay_follows_pointer"]
         self._apply_local()
+        if getattr(self, "home_window", None) is not None:
+            self.home_window.refresh()
         self._build_tray()
         self._refresh_tray()
         # Taken once here for _external: the answer cannot change under a
@@ -1515,11 +1617,11 @@ def _hand_over(command):
     """Give the running instance the attention this start was asking for.
 
     A start carrying a verb forwards only that verb; a bare double start asks
-    for the Settings window as the sign of life the click was looking for.
+    for the daily workspace as the sign of life the click was looking for.
     Retried for a moment, because the copy that won the lock may not be
     listening yet.
     """
-    verb = command or "settings"
+    verb = command or "home"
     deadline = time.monotonic() + 5
     while time.monotonic() < deadline:
         if ipc.send(verb) is not None:
@@ -1546,7 +1648,7 @@ def run_app(args):
         if command:
             ipc.send(command)
         else:
-            ipc.send("settings")
+            ipc.send("home")
         return 0
 
     app = QApplication(sys.argv)
@@ -1621,13 +1723,12 @@ def run_app(args):
     server.newConnection.connect(on_connection)
     app.aboutToQuit.connect(dikte.shutdown)
 
-    # No key for the chosen transcription provider means nothing can work yet,
+    # Explicit home requests and first setup open the daily workspace.
-    # so the settings window is the only useful thing to open.
+    # A configured --gui background start stays quiet for login and restart.
-    # A transcription provider that cannot run yet, whether that is a missing
+    if command == "settings":
-    # API key or a model nobody has downloaded, means nothing can work, so the
-    # settings window is the only useful thing to open.
-    if command == "settings" or not dikte.conf.transcribe_ready():
         dikte.open_settings()
+    elif command == "home" or not dikte.conf.transcribe_ready():
+        dikte.open_home()
     elif command == "toggle":
         QTimer.singleShot(0, dikte.toggle)
     elif command == "ask":

dikte/cli.py

@@ -30,6 +30,7 @@ from . import audio
 from . import cleanup
 from . import config as cfg
 from . import filetranscribe
+from . import ggml
 from . import hotkey
 from . import hub
 from . import ipc
@@ -44,7 +45,7 @@ NOT_RUNNING = 3
 # Verbs that start the application when none is running, which is what a
 # shortcut registered with the desktop has always relied on: press the key on a
 # fresh login and Dikte comes up recording.
-GUI_VERBS = {"", "settings", "toggle", "ask", "meeting"}
+GUI_VERBS = {"", "home", "settings", "toggle", "ask", "meeting"}
 
 # Asking a process that is not there to stop, cancel or quit is not a failure;
 # it is already in the state that was asked for.
@@ -836,6 +837,69 @@ def cmd_update(opts):
                f"{release.url}")
 
 
+# --- the models on this machine --------------------------------------------
+
+
+def _local_where(entry):
+    """Where a local model ran, in a phrase: the card, the processor, or neither.
+
+    The backend and the card keep the names the server printed for them. A
+    graphics card is a product somebody sells under that name, and translating
+    it would be inventing hardware.
+    """
+    kind = ggml.accel_kind({**entry, "running": True})
+    where = {"gpu": "the graphics card", "cpu": "the processor"}.get(
+        kind, "something it did not name")
+    detail = ggml.accel_detail(entry)
+    return where + (f" ({detail})" if detail else "")
+
+
+def _local_note(entry):
+    """What the log establishes when GPU use was requested but unavailable."""
+    if not entry.get("gpu_wanted"):
+        return ""
+    if ggml.accel_kind({**entry, "running": True}) != "cpu":
+        return ""
+    if not ggml.cpu_only_loaded(entry):
+        return "  - the graphics card is switched on but could not be used"
+    return ("  - only the CPU backend was loaded; check the server log for "
+            "graphics backend or driver errors")
+
+
+def _local_line(name, entry):
+    if not entry.get("running"):
+        return "not loaded"
+    model = entry.get("model") or ""
+    return (f"loaded on {_local_where(entry)}"
+            + (f", {model}" if model else "") + _local_note(entry))
+
+
+def _last_local(conf):
+    """What the local servers last ran on, read off the logs they left behind.
+
+    For a command line asking while nothing is running: there is no process to
+    put the question to, and the log outlives the process that wrote it. Every
+    entry says `running` is false, because this is an account of the last start
+    rather than a reading of a live one. The logs do not record the binary path
+    or requested GPU setting, so current settings cannot explain that run.
+    """
+    rows = {}
+    for program, used in (
+            (ggml.WHISPER, conf["transcribe_provider"] == "local"),
+            (ggml.LLAMA, conf.uses_local_llm())):
+        accel = ggml.last_accel(program)
+        rows[program.name] = {
+            # Whether one ever started here at all, which the backend cannot
+            # say on its own: a server that ran and named no backend and one
+            # that never ran both leave it empty.
+            "ran": ggml.server_log(program).exists(),
+            "running": False, "used": used,
+            "backend": accel.backend, "device": accel.device,
+            "layers": accel.layers, "available": list(accel.available),
+        }
+    return rows
+
+
 def cmd_status(opts):
     reply = ipc.send("status")
     if reply is None:
@@ -855,6 +919,11 @@ def cmd_status(opts):
         + (f"  {reply['meeting_message']}" if reply.get("meeting_message") else ""),
         f"listener:  {'on' if reply.get('listener') else 'off'}",
     ]
+    # Nothing for a setup that uses no model on this machine, and nothing at all
+    # from an instance too old to have been asked.
+    for name, entry in (reply.get("local") or {}).items():
+        if entry.get("used") or entry.get("running"):
+            lines.append(f"{name + ':':11}{_local_line(name, entry)}")
     return out(opts, reply, "\n".join(lines))
 
 
@@ -867,13 +936,16 @@ def cmd_doctor(opts):
     # Mac shells out for one half and Windows for neither. A row saying ydotool
     # is missing on a machine that would never have run it is not a diagnosis,
     # it is a red mark to explain away.
+    # Asked once and read twice: whether an instance is running, and what its
+    # local servers are doing, which is a question only that process can answer.
+    live = ipc.send("status") or {}
     here = paste.desktop()
     wanted = [here.clipboard, here.keyboard]
     if sys.platform.startswith("linux"):
         # Recording, the device list, and KDE's shortcut registry.
         wanted += ["pw-record", "pactl", "kwriteconfig6"]
     wanted += ["ffmpeg",
-               assistant.executable(assistant.provider(conf)) or "claude",
+               assistant.executable(assistant.provider(conf)),
                cleanup.executable(cleanup.provider(conf))]
     programs = {name: shutil.which(name) or "" for name in wanted if name}
     target = conf.transcribe_target()
@@ -908,8 +980,14 @@ def cmd_doctor(opts):
                     "ready": cleanup_ready},
         "agent": {"provider": assistant.provider(conf),
                   "directory": assistant.working_dir(conf)},
-        "running": ipc.send("status") is not None,
+        "running": bool(live),
+        # Live when there is an instance to ask, off the logs when there is not.
+        "local": live.get("local") or _last_local(conf),
     }
+    # An instance from before this field existed is not an instance saying
+    # nothing is loaded; it is one that cannot be asked, and the two must not
+    # print the same line.
+    stale = bool(live) and "local" not in live
     if target.provider == "local":
         transcribe_line = (f"{'✓' if transcribe_ready else '✗'} {target.service}, "
                            f"transcribing on {target.model or 'no model yet'}")
@@ -929,12 +1007,30 @@ def cmd_doctor(opts):
                         f"{cleanup.model(conf)}")
     lines = [f"{'✓' if path else '✗'} {name:14} {path or 'not on your PATH'}"
              for name, path in programs.items()]
-    lines += [
+    lines += [transcribe_line, cleanup_line]
-        transcribe_line,
+    # Only the models this setup actually uses: a machine transcribing in the
-        cleanup_line,
+    # cloud has nothing loaded here and no reason to read about it.
+    for name, entry in checks["local"].items():
+        if not entry.get("used"):
+            continue
+        if stale:
+            lines.append(f"· {name:14} the running instance is too old to say; "
+                         f"reload it with: dikte restart")
+        elif entry.get("running"):
+            lines.append(f"✓ {name:14} {_local_line(name, entry)}")
+        elif live:
+            lines.append(f"· {name:14} not loaded")
+        elif entry.get("backend"):
+            lines.append(f"· {name:14} last run on "
+                         f"{_local_where(entry)}")
+        elif entry.get("ran"):
+            lines.append(f"· {name:14} last run said nothing about what it "
+                         f"was running on")
+        else:
+            lines.append(f"· {name:14} never run here")
+    lines.append(
         f"{'✓' if checks['running'] else '·'} application "
-        + ("running" if checks["running"] else "not running"),
+        + ("running" if checks["running"] else "not running"))
-    ]
     return out(opts, {"ok": True, **checks}, "\n".join(lines))
 
 
@@ -1179,7 +1275,7 @@ def build_parser():
     updates.set_defaults(func=cmd_update)
 
     leaf(subs, "status", "what it is doing right now").set_defaults(func=cmd_status)
-    for name, help_text in (("settings", "open the settings window"),
+    for name, help_text in (("home", "open Dikte"), ("settings", "open the settings window"),
                             ("restart", "reload the running instance"),
                             ("quit", "shut it down")):
         leaf(subs, name, help_text).set_defaults(func=cmd_plain)
@@ -1210,8 +1306,8 @@ def run(argv):
                 pass
     parser = build_parser()
     opts = parser.parse_args(argv)
-    # No verb at all is the plain `dikte`, which means the settings window.
+    # No verb opens the daily workspace; settings remains an explicit verb.
-    opts.verb = opts.verb or ""
+    opts.verb = opts.verb or "home"
     # Every path here either talks over the socket or drives one of the workers,
     # and both want an event loop under them; a window is what none of them want.
     _app = QCoreApplication.instance() or QCoreApplication(sys.argv[:1])

dikte/config.py

@@ -418,6 +418,7 @@ da senin soracağın soruya verilecek bir yanıt yok.
   ve varsayımını bir yan cümlede söyle"""
 
 DEFAULTS = {
+    "theme": "nord",
     "ui_language": "auto",          # auto | tr | en
     "openai_api_key": "",
     "openai_base_url": "https://api.openai.com/v1",
@@ -438,7 +439,9 @@ DEFAULTS = {
     # What a timestamped run (subtitles) asks OpenRouter for: not every model
     # there returns segment times. Empty -> openai/whisper-1.
     "openrouter_file_model": "",
-    "language": "tr",
+    # A stored language overrides this default. Hosted providers receive no
+    # language hint in auto mode; local whisper also reports the detected code.
+    "language": "auto",
     "transcribe_prompt": "",
 
     # --- whisper.cpp, on this machine ---------------------------------------
@@ -773,13 +776,23 @@ class Config:
         return self["cleanup_provider"] == "local"
 
     def cleanup_prompt(self, with_timestamps=False, with_speakers=False,
-                       subtitles=False):
+                       subtitles=False, speech=""):
-        turkish = i18n.language() == "tr"
+        """`speech` is the two-letter code of the language that was heard, when
+        the transcription model reported one. The default prompts and the
+        glossary rule only exist in Turkish and English, so a detected Turkish
+        recording gets the Turkish prompt and any other detected language, or
+        none at all, the English one, which is written not to care what
+        language the transcript is in. Nothing else calls this with it, so the
+        interface language keeps deciding everywhere the speech was not asked
+        about."""
+        turkish = (speech == "tr") if speech else i18n.language() == "tr"
         if subtitles:
             prompt = (self["file_cleanup_prompt"].strip()
-                      or default_file_cleanup_prompt())
+                      or (FILE_CLEANUP_PROMPT_TR if turkish
+                          else FILE_CLEANUP_PROMPT_EN))
         else:
-            prompt = self["cleanup_prompt"].strip() or default_cleanup_prompt()
+            prompt = (self["cleanup_prompt"].strip()
+                      or (CLEANUP_PROMPT_TR if turkish else CLEANUP_PROMPT_EN))
         glossary = self["transcribe_prompt"].strip()
         if with_speakers:
             glossary = "\n".join(x for x in (glossary, self.participants()) if x)

dikte/ggml.py

@@ -35,6 +35,7 @@ import json
 import os
 import pathlib
 import platform
+import re
 import shutil
 import signal
 import socket
@@ -1065,6 +1066,162 @@ def _tail(path, lines=3):
     return " | ".join(found[-lines:])
 
 
+# --- what the server is running on ----------------------------------------
+
+# Both programs say where the model went, and neither is asked: it is printed
+# while they start and captured in the log Dikte already keeps. Reading it back
+# is the only way to tell a graphics card that was asked for from one that was
+# found, which is a difference the settings checkbox cannot make on its own.
+Accel = collections.namedtuple("Accel", "backend device layers available")
+
+NO_ACCEL = Accel("", "", "", ())
+
+# ggml loads each backend from a shared object and says which; whisper then says
+# whether it found a card, and llama says how many layers went onto it.
+_BACKEND_LOADED = re.compile(r"^load_backend: loaded (\w+) backend", re.M)
+_WHISPER_NO_GPU = "whisper_backend_init_gpu: no GPU found"
+# "using" precedes the attempt to initialise the backend. A later failure
+# invalidates it, even when model weights were already put on that device.
+_WHISPER_ATTEMPT = re.compile(
+    r"^whisper_backend_init_gpu: (using|failed to initialize) (\S+) backend",
+    re.M)
+_WHISPER_BUFFER = re.compile(r"^whisper_model_load:\s+(\S+) total size", re.M)
+# The device listing, read for the card's name rather than for the verdict.
+_WHISPER_DEVICE = re.compile(
+    r"^whisper_backend_init_gpu: device (\d+): (.+?) \(type: (\d+)\)", re.M)
+_LLAMA_OFFLOAD = re.compile(
+    r"^load_tensors: offloaded (\d+)/(\d+) layers to GPU", re.M)
+_LLAMA_BUFFER = re.compile(
+    r"^load_tensors:\s+(\S+) model buffer size\s*=\s*(\d+(?:\.\d+)?) MiB",
+    re.M)
+
+# whisper names the device by its ggml handle, "Vulkan0" or "CUDA0", which says
+# which slot rather than which card. Each backend prints the real name as it
+# enumerates, one line further up.
+_HANDLE = re.compile(r"^([A-Za-z]+?)(\d*)$")
+_BARE = re.compile(r"^(?:Vulkan|CUDA|ROCm|SYCL|Metal|GPU|CPU)\d*$", re.I)
+_METAL_DEVICE = re.compile(r"^ggml_metal.*picking default device: (.+)$", re.M)
+# The driver in brackets after the card's own name: "(radv)", "(nvidia)". The
+# name carries brackets of its own, but in capitals, so the case is what tells
+# a driver tag from part of the name.
+_DRIVER_TAG = re.compile(r"\s*\([a-z0-9_.\- ]+\)$")
+
+
+def _enumerated(text, backend, index):
+    """The name the backend printed for one of its own devices, by slot.
+
+    Asked by backend rather than by whichever listing came first: a machine
+    with both a CUDA build and a Vulkan loader prints two listings, and the
+    card named in the wrong one is somebody else's card.
+    """
+    listings = {
+        "vulkan": rf"^ggml_vulkan: {index} = (.+?) \| ",
+        "cuda": rf"^\s*Device {index}: (.+?), compute capability",
+        "rocm": rf"^\s*Device {index}: (.+?), compute capability",
+    }
+    pattern = listings.get(backend.lower())
+    found = re.search(pattern, text, re.M) if pattern else None
+    if found is None and backend.lower() == "metal":
+        found = _METAL_DEVICE.search(text)
+    if found is None:
+        return ""
+    return _DRIVER_TAG.sub("", found.group(1).strip())
+
+
+def _card_name(text, handle):
+    """The card behind a ggml handle like "Vulkan0", named the way it sells.
+
+    The backend's own enumeration is asked first because it is the only listing
+    indexed the way the handle is. whisper numbers every device it can see in
+    one sequence, so the Vulkan card can be its device 1 while being Vulkan0,
+    and reading that row by the handle's digit names whatever else was in slot
+    zero. The handle itself is never an answer: it says which slot, and a line
+    reading "Vulkan, Vulkan0" tells nobody which card is doing the work.
+    """
+    parts = _HANDLE.match(handle or "")
+    backend, index = (parts.group(1), parts.group(2) or "0") if parts else ("", "0")
+    found = _enumerated(text, backend, index)
+    if found:
+        return found
+    # Nothing enumerated: whisper's own listing is all there is, and a single
+    # named device in it can only be the one that ran.
+    named = [name.strip() for _slot, name, kind in _WHISPER_DEVICE.findall(text)
+             if kind != "0" and name.strip() and not _BARE.match(name.strip())]
+    return named[0] if len(named) == 1 else ""
+
+# The startup chatter is the first few hundred lines; the rest of the file is a
+# line per request and grows for as long as the server lives.
+_LOG_HEAD = 64 << 10
+
+
+def _read_accel(program, log_path):
+    """What the server that wrote `log_path` is running on.
+
+    An empty backend is a real answer rather than a failure: a whisper built by
+    hand on a Mac has Metal compiled in and prints no load_backend line at all,
+    and calling that "the processor" would be a confident lie about the one
+    thing this is here to be honest about.
+    """
+    try:
+        with open(log_path, encoding="utf-8", errors="replace") as fh:
+            text = fh.read(_LOG_HEAD)
+    except OSError:
+        return NO_ACCEL
+    # dict.fromkeys rather than a set: the order they were loaded in is the
+    # order they are worth showing in, and CPU is always one of them.
+    available = tuple(dict.fromkeys(_BACKEND_LOADED.findall(text)))
+    cards = [name for name in available if name.upper() != "CPU"]
+    if program is WHISPER:
+        handle, failed = "", False
+        for event, device in _WHISPER_ATTEMPT.findall(text):
+            if event == "using":
+                handle, failed = device, False
+            elif not handle or device == handle:
+                handle, failed = "", True
+        if failed:
+            return Accel("CPU", "", "", available)
+        buffered = _WHISPER_BUFFER.search(text)
+        if not handle and buffered:
+            handle = buffered.group(1)
+        if _WHISPER_NO_GPU in text or handle.upper().startswith("CPU"):
+            return Accel("CPU", handle or "", "", available)
+        if handle:
+            parts = _HANDLE.match(handle)
+            backend = parts.group(1) if parts else ""
+            # The backend as the build spells it, so "Vulkan" rather than the
+            # capitalisation the handle happened to use.
+            backend = next((name for name in cards
+                            if name.lower() == backend.lower()), backend)
+            return Accel(backend or "GPU", _card_name(text, handle), "",
+                         available)
+        if available and not cards:
+            # Nothing but a processor backend in this build: there was nowhere
+            # else the model could have gone.
+            return Accel("CPU", "", "", available)
+        return Accel("", "", "", available)
+    found = _LLAMA_OFFLOAD.search(text)
+    if found:
+        layers = f"{found.group(1)}/{found.group(2)}"
+        if int(found.group(1)) > 0:
+            # Loaded libraries do not identify the device holding the model.
+            # Host buffers such as CUDA_Host are not GPU allocations. When
+            # several devices hold weights, do not pretend only one ran.
+            handles = list(dict.fromkeys(
+                handle for handle, size in _LLAMA_BUFFER.findall(text)
+                if float(size) > 0 and _BARE.fullmatch(handle)
+                and not handle.upper().startswith("CPU")))
+            if len(handles) == 1:
+                handle = handles[0]
+                backend = _HANDLE.fullmatch(handle).group(1)
+                return Accel(backend, _card_name(text, handle), layers,
+                             available)
+            return Accel("GPU", "", layers, available)
+        return Accel("CPU", "", layers, available)
+    if available and not cards:
+        return Accel("CPU", "", "", available)
+    return Accel("", "", "", available)
+
+
 def _win_image_name(pid):
     """The full, lower-cased path of the process's executable, or ''.
 
@@ -1118,6 +1275,18 @@ class Server:
         self._port = 0
         self._log = ""
         self._key = None
+        # What the running child settled on, read out of its log once it was
+        # ready. Kept beside the process because it belongs to that process and
+        # to no other: a restart on new settings may land somewhere else.
+        self._accel = NO_ACCEL
+        # The settings the running child was started on, which is not what
+        # _settings holds: a change made while a start is in flight lands there
+        # first, and reporting the new model beside the old process would name
+        # a model this server is not running.
+        self._live = {}
+        # The copy that is running, resolved rather than configured: the
+        # setting is usually empty, meaning whichever one program_path finds.
+        self._binary = ""
         # The pid this instance last wrote to its pid file, so _forget never
         # removes a file some other Dikte wrote after us.
         self._pid = 0
@@ -1179,6 +1348,34 @@ class Server:
         with self._lock:
             return f"http://{HOST}:{self._port}/v1" if self._port else ""
 
+    def state(self):
+        """A snapshot of what this server is doing, for something to show.
+
+        Taken under the short lock rather than the start one, so the interface
+        is answered at once even while a model is being read in. Plain types
+        throughout, because this travels over the socket to the command line.
+        """
+        with self._lock:
+            up = self._proc is not None and self._proc.poll() is None
+            accel = self._accel if up else NO_ACCEL
+            # What it is running, when it is running; what it would run
+            # otherwise. The two differ for as long as a change waits for the
+            # restart that will pick it up.
+            settings = self._live if up else self._settings
+            return {
+                "running": up,
+                "pid": self._proc.pid if up else 0,
+                "port": self._port if up else 0,
+                "model": settings.get("model", ""),
+                "gpu_wanted": bool(settings.get("gpu")),
+                "backend": accel.backend,
+                "device": accel.device,
+                "layers": accel.layers,
+                "available": list(accel.available),
+                "binary": self._binary if up else "",
+                "downloaded": bool(up and is_downloaded(self._binary)),
+            }
+
     def error(self):
         """The last thing the server printed, for a failure after it started."""
         with self._lock:
@@ -1200,9 +1397,12 @@ class Server:
             self._stop_now()
             with self._lock:
                 settings, key = dict(self._settings), self._settings_key()
-            proc, port, log = self._launch(settings)
+            proc, port, log, accel = self._launch(settings)
             with self._lock:
                 self._proc, self._port, self._log, self._key = proc, port, log, key
+                self._accel, self._live = accel, settings
+                self._binary = program_path(self.program,
+                                            settings.get("binary", ""))
                 # Only the clock. The count is not this launch's to reset: a
                 # caller that took a hold and then asked for the address, which
                 # is what the local cleanup does, would have it wiped here and
@@ -1322,7 +1522,9 @@ class Server:
                 self._forget()
                 raise
             if reason == "ready":
-                return proc, port, str(log)
+                # Read now rather than on demand: the startup lines are at the
+                # head of a file a long-lived server keeps appending to.
+                return proc, port, str(log), _read_accel(self.program, log)
             last = _tail(log)
             self._forget()
             # Losing the port between the probe and the bind is the one
@@ -1423,6 +1625,8 @@ class Server:
         with self._lock:
             proc, self._proc = self._proc, None
             self._port, self._log, self._key = 0, "", None
+            self._accel, self._live = NO_ACCEL, {}
+            self._binary = ""
         self._kill(proc, gently=True)
         if proc is not None:
             self._forget()
@@ -1547,12 +1751,13 @@ def _whisper_args(settings):
         binary, "-m", str(model),
         "--inference-path", INFERENCE_PATH,
         # Whatever language the request does not name. api.py leaves the field
-        # out when the language is "auto", and the server's own default is
+        # out when the language is "auto", and the server's own language is
-        # English rather than detection.
+        # set here: "auto" makes whisper.cpp detect what it hears.
         "-l", "auto",
         # Stock phrases invented for near-silence come from non-speech tokens,
         # and verbose_json otherwise pays for a language probability sweep
-        # nothing here reads.
+        # nobody asked for. A request that wants the detected language switches
+        # that back on per request.
         "-sns", "-nlp",
     ]
     if int(settings["threads"]) > 0:
@@ -1606,6 +1811,80 @@ def sweep():
     return any([server.sweep() for server in SERVERS])
 
 
+def state():
+    """What each local server is doing, keyed by program name."""
+    return {server.program.name: server.state() for server in SERVERS}
+
+
+def is_downloaded(path):
+    """Whether `path` is a copy Dikte fetched rather than one the system has."""
+    return bool(path) and _under(path, BIN_DIR)
+
+
+def server_log(program):
+    """Where this program's server writes, which outlives the process."""
+    return DATA_DIR / f"{program.name}-server.log"
+
+
+def last_accel(program):
+    """What the last server for `program` ran on, from the log it left behind.
+
+    For a command line asking with nothing running: the log outlives the process
+    and is the only account of the last start there is.
+    """
+    return _read_accel(program, server_log(program))
+
+
+def accel_kind(state):
+    """"off" | "gpu" | "cpu" | "unknown", for a state() or an Accel.
+
+    A tag rather than a sentence, because the two places that show this write
+    their own: the command line answers in English and the settings window in
+    whatever language it was opened in.
+    """
+    if isinstance(state, Accel):
+        state = {"running": True, "backend": state.backend}
+    if not state.get("running"):
+        return "off"
+    backend = state.get("backend") or ""
+    if not backend:
+        return "unknown"
+    return "cpu" if backend.upper() == "CPU" else "gpu"
+
+
+def accel_detail(state):
+    """The backend, the card and the layers, joined, or "" when none were said.
+
+    Names as the server printed them: "CUDA", "Vulkan", the card's own model
+    name. Translating those would be inventing hardware nobody sells.
+    """
+    if isinstance(state, Accel):
+        state = state._asdict()
+    parts = [state.get("backend") or "", state.get("device") or ""]
+    if state.get("layers"):
+        parts.append(f"{state['layers']} layers")
+    # A whisper on the processor prints "CPU" as its device too, and saying it
+    # twice reads like two different things.
+    seen, out = set(), []
+    for part in parts:
+        if part and part.lower() not in seen:
+            seen.add(part.lower())
+            out.append(part)
+    return ", ".join(out)
+
+
+def cpu_only_loaded(state):
+    """Whether CPU is the only backend the log says was loaded.
+
+    A missing GPU backend and one that failed to load look the same here.
+    This cannot establish which backends the binary was built to support.
+    """
+    if isinstance(state, Accel):
+        state = state._asdict()
+    available = [name.upper() for name in (state.get("available") or [])]
+    return available == ["CPU"]
+
+
 def stop_all():
     for server in SERVERS:
         server.stop()

dikte/home_ui.py

@@ -0,0 +1,543 @@
+"""Task-first native desktop window backed by the application controllers."""
+
+import shutil
+
+from PyQt6.QtCore import Qt, QSize, QTimer
+from PyQt6.QtGui import QColor, QIcon, QPainter, QPen, QPixmap
+from PyQt6.QtWidgets import (
+    QApplication, QButtonGroup, QDialog, QFrame, QHBoxLayout, QLabel,
+    QPlainTextEdit, QPushButton, QScrollArea, QSizePolicy, QStackedWidget, QVBoxLayout, QWidget,
+)
+
+from . import api, assistant, audio, cleanup, config as cfg, ggml
+from .i18n import t
+from . import theme
+
+
+def _label(text="", name="", centered=False):
+    label = QLabel(text)
+    label.setTextFormat(Qt.TextFormat.PlainText)
+    label.setWordWrap(True)
+    label.setObjectName(name)
+    if centered:
+        label.setAlignment(Qt.AlignmentFlag.AlignCenter)
+    return label
+
+
+def _button(text, callback, name=""):
+    button = QPushButton(text)
+    button.setObjectName(name)
+    button.setAutoDefault(False)
+    button.setSizePolicy(QSizePolicy.Policy.Fixed, QSizePolicy.Policy.Fixed)
+    button.clicked.connect(callback)
+    return button
+
+
+def microphone_icon(recording=False, color="#172434"):
+    """A scalable microphone outline, independent of the desktop icon theme."""
+    pixmap = QPixmap(80, 80)
+    pixmap.fill(Qt.GlobalColor.transparent)
+    painter = QPainter(pixmap)
+    painter.setRenderHint(QPainter.RenderHint.Antialiasing)
+    painter.setPen(QPen(QColor(color), 4, Qt.PenStyle.SolidLine,
+                        Qt.PenCapStyle.RoundCap))
+    if recording:
+        painter.setBrush(QColor(color))
+        painter.drawRoundedRect(26, 26, 28, 28, 3, 3)
+    else:
+        painter.drawRoundedRect(32, 12, 16, 36, 8, 8)
+        painter.drawArc(22, 28, 36, 30, 180 * 16, 180 * 16)
+        painter.drawLine(40, 58, 40, 68)
+    painter.end()
+    return QIcon(pixmap)
+
+
+def settings_icon(color="#B2C1D1"):
+    pixmap = QPixmap(48, 48)
+    pixmap.fill(Qt.GlobalColor.transparent)
+    painter = QPainter(pixmap)
+    painter.setRenderHint(QPainter.RenderHint.Antialiasing)
+    painter.setPen(QPen(QColor(color), 3, Qt.PenStyle.SolidLine,
+                        Qt.PenCapStyle.RoundCap))
+    painter.translate(24, 24)
+    painter.drawEllipse(-12, -12, 24, 24)
+    painter.drawEllipse(-4, -4, 8, 8)
+    for _ in range(8):
+        painter.drawLine(0, -12, 0, -17)
+        painter.rotate(45)
+    painter.end()
+    return QIcon(pixmap)
+
+
+def _model_location(model, provider, local_state=None):
+    model = model or t("Model not selected")
+    if provider == "local":
+        kind = ggml.accel_kind(local_state or {})
+        location = t("Local GPU") if kind == "gpu" else t("Local CPU") if kind == "cpu" else t("Local")
+    else:
+        location = "CLI" if provider in ("claude", "codex", "agy") else "API"
+    return f"{model} ({location})"
+
+
+def processing_locations(conf, mode="dictation", file_cleanup=None, file_timestamps=None):
+    """Display configured model IDs and observed local acceleration, never keys."""
+    target = conf.transcribe_target()
+    local = ggml.state()
+    sound_model = target.model
+    timestamps = conf["file_timestamps"] if file_timestamps is None else file_timestamps
+    if mode == "meeting" or (mode == "file" and timestamps):
+        sound_model = api.timestamp_model(target.provider, target.model, target.file_model)
+    sound = _model_location(sound_model, target.provider, local.get("whisper"))
+    enabled = conf["cleanup_enabled"]
+    if mode == "file":
+        enabled = conf["file_cleanup"] if file_cleanup is None else file_cleanup
+    elif mode == "meeting":
+        enabled = conf["meeting_cleanup"]
+    elif mode == "ask":
+        enabled = conf["assistant_cleanup"]
+    provider = cleanup.provider(conf)
+    model = cleanup.model(conf)
+    if provider in ("codex", "agy") and not conf[f"cleanup_{provider}_model"].strip():
+        model = t("{name} default model", name="Codex" if provider == "codex" else "Antigravity")
+    text = _model_location(model, provider, local.get("llama")) if enabled else t("Editing off")
+    location = t("Dictation: {sound} / Cleanup: {text}", sound=sound, text=text)
+    if mode == "meeting":
+        location += " / " + t("Minutes: {model}", model=_model_location(conf["meeting_model"], "openrouter"))
+    elif mode == "ask":
+        provider = assistant.provider(conf)
+        model = assistant.model(conf)
+        if provider in ("codex", "agy") and not conf[f"assistant_{provider}_model"].strip():
+            model = t("{name} default model", name=assistant.display_name(conf))
+        location += " / " + t("Assistant: {model}", model=_model_location(model, provider))
+    return location
+
+
+class HomeWindow(QWidget):
+    """Own navigation and presentation; recording and processing stay in Dikte."""
+
+    def __init__(self, controller, settings):
+        super().__init__()
+        self.controller = controller
+        self.conf = controller.conf
+        self.settings = settings
+        self.mode = "dictation"
+        self._last_result = None
+        self._last_answer = None
+        font = self.font()
+        font.setPointSizeF(max(10.5, font.pointSizeF()))
+        self.setFont(font)
+        self.setObjectName("home")
+        self.setWindowTitle("Dikte")
+        theme.apply(self, self.conf["theme"])
+        self._theme_name = None
+        root = QVBoxLayout(self)
+        root.setContentsMargins(20, 16, 20, 12)
+        root.setSpacing(12)
+        navigation = QHBoxLayout()
+        navigation.setSpacing(6)
+        self.mode_group = QButtonGroup(self)
+        self.mode_buttons = {}
+        for name, title in (("dictation", "Dictation"), ("file", "File"),
+                            ("meeting", "Meeting"), ("ask", "Assistant")):
+            button = _button(t(title), lambda checked=False, name=name: self.show_mode(name), "mode")
+            button.setCheckable(True)
+            button.setSizePolicy(QSizePolicy.Policy.Expanding, QSizePolicy.Policy.Fixed)
+            self.mode_group.addButton(button)
+            navigation.addWidget(button, 1)
+            self.mode_buttons[name] = button
+        self.settings_button = _button("", controller.open_settings, "settings")
+        self.settings_button.setIcon(settings_icon())
+        self.settings_button.setIconSize(QSize(20, 20))
+        self.settings_button.setFixedSize(34, 34)
+        self.settings_button.setToolTip(t("Settings"))
+        self.settings_button.setAccessibleName(t("Settings"))
+        navigation.addWidget(self.settings_button)
+        root.addLayout(navigation)
+        self.pages = QStackedWidget()
+        root.addWidget(self.pages, 1)
+        self.mode_pages = {}
+        self.mode_pages["dictation"] = self._scrolled(self._capture_page())
+        self.mode_pages["file"] = settings.task_pages["file"]
+        self.mode_pages["meeting"] = self._scrolled(self._meeting_page())
+        self.mode_pages["ask"] = self._scrolled(self._assistant_page())
+        self.mode_pages["history"] = self._scrolled(self._history_page())
+        for page in self.mode_pages.values():
+            self.pages.addWidget(page)
+        self.footer = _label("", "footer", True)
+        root.addWidget(self.footer)
+        self._timer = QTimer(self)
+        self._timer.setInterval(500)
+        self._timer.timeout.connect(self.refresh)
+        self._timer.start()
+        settings.applied.connect(self.refresh)
+        settings.file_cleanup.toggled.connect(self.refresh)
+        settings.file_timestamps.toggled.connect(self.refresh)
+        settings.transcriber.finished.connect(self.refresh_results)
+        settings.history.model().rowsRemoved.connect(self.refresh_results)
+        settings.history.model().modelReset.connect(self.refresh_results)
+        self.resize(620, 560)
+        screen = self.screen()
+        if screen:
+            room = screen.availableGeometry()
+            self.resize(min(620, room.width() - 40), min(560, room.height() - 80))
+        self.setMinimumSize(420, 360)
+        self.show_mode("dictation")
+        self.refresh_results()
+
+    @staticmethod
+    def _scrolled(page):
+        page.setMaximumWidth(680)
+        area = QScrollArea()
+        area.setWidgetResizable(True)
+        area.setAlignment(Qt.AlignmentFlag.AlignHCenter | Qt.AlignmentFlag.AlignTop)
+        area.setFrameShape(QFrame.Shape.NoFrame)
+        area.setWidget(page)
+        return area
+
+    def _capture_page(self):
+        page = QWidget()
+        layout = QVBoxLayout(page)
+        layout.setContentsMargins(0, 10, 0, 0)
+        layout.setSpacing(8)
+        self.capture_status = _label("", "heading", True)
+        layout.addWidget(self.capture_status)
+        self.capture_button = _button("", self._capture, "capture")
+        self._mic_icon = microphone_icon()
+        self._stop_icon = microphone_icon(recording=True)
+        self.capture_button.setIcon(self._mic_icon)
+        self.capture_button.setIconSize(QSize(44, 44))
+        self.capture_button.setFixedSize(112, 112)
+        layout.addWidget(self.capture_button, 0, Qt.AlignmentFlag.AlignHCenter)
+        self.capture_shortcut = _label("", "muted", True)
+        layout.addWidget(self.capture_shortcut)
+        self.capture_models = _label("", "models", True)
+        layout.addWidget(self.capture_models)
+        controls = QHBoxLayout()
+        controls.addStretch()
+        self.pause_button = _button(t("Pause the recording"), self._pause)
+        self.cancel_button = _button(t("Discard the recording"), self._cancel_capture)
+        controls.addWidget(self.pause_button)
+        controls.addWidget(self.cancel_button)
+        controls.addStretch()
+        layout.addLayout(controls)
+        self.capture_error = _label()
+        layout.addWidget(self.capture_error)
+        card = QFrame()
+        card.setObjectName("result")
+        content = QVBoxLayout(card)
+        content.setContentsMargins(18, 12, 18, 12)
+        top = QHBoxLayout()
+        top.addWidget(_label(t("Latest text")))
+        top.addStretch(1)
+        top.addWidget(_button(t("History"), lambda: self.show_mode("history")))
+        content.addLayout(top)
+        self.latest_text = QPlainTextEdit()
+        self.latest_text.setReadOnly(True)
+        self.latest_text.setAccessibleName(t("Latest text"))
+        self.latest_text.setPlaceholderText(t("Your first transcript will appear here."))
+        self.latest_text.setMinimumHeight(84)
+        self.latest_text.setMaximumHeight(100)
+        content.addWidget(self.latest_text)
+        self.latest_warning = _label()
+        content.addWidget(self.latest_warning)
+        actions = QHBoxLayout()
+        self.latest_time = _label("", "muted")
+        actions.addWidget(self.latest_time, 1)
+        self.copy_button = _button(t("Copy"), lambda: QApplication.clipboard().setText(self.latest_text.toPlainText()))
+        self.open_button = _button(t("Open text"), lambda: self._open_text(self.latest_text.toPlainText()))
+        actions.addWidget(self.copy_button)
+        actions.addWidget(self.open_button)
+        content.addLayout(actions)
+        layout.addWidget(card)
+        layout.addStretch(1)
+        return page
+
+    def _meeting_page(self):
+        page = QWidget()
+        layout = QVBoxLayout(page)
+        layout.setContentsMargins(0, 0, 0, 0)
+        self.meeting_status = _label("", "heading")
+        layout.addWidget(self.meeting_status)
+        self.meeting_hint = _label("", "muted")
+        layout.addWidget(self.meeting_hint)
+        actions = QVBoxLayout()
+        self.meeting_button = _button(t("Record a meeting"), self._meeting, "primary")
+        self.meeting_cancel = _button(t("Discard the meeting"), self._cancel_meeting)
+        actions.addWidget(self.meeting_button)
+        actions.addWidget(self.meeting_cancel)
+        layout.addLayout(actions)
+        self.meeting_error = _label()
+        layout.addWidget(self.meeting_error)
+        layout.addWidget(_label(t("Minutes")))
+        minutes = self.settings.task_pages["minutes"]
+        minutes.setMinimumHeight(300)
+        layout.addWidget(minutes, 1)
+        minutes.show()
+        return page
+
+    def _assistant_page(self):
+        page = QWidget()
+        layout = QVBoxLayout(page)
+        layout.setContentsMargins(0, 0, 0, 0)
+        self.ask_status = _label("", "heading")
+        layout.addWidget(self.ask_status)
+        self.ask_scope = _label("", "muted")
+        layout.addWidget(self.ask_scope)
+        actions = QHBoxLayout()
+        self.ask_button = _button("", self._ask, "primary")
+        actions.addWidget(self.ask_button)
+        actions.addWidget(_button(t("Start a new conversation"), self.controller.reset_conversation))
+        actions.addStretch()
+        layout.addLayout(actions)
+        self.ask_pause = _button(t("Pause the recording"), self._pause)
+        layout.addWidget(self.ask_pause)
+        self.ask_cancel = _button(t("Stop"), self._cancel_ask)
+        layout.addWidget(self.ask_cancel)
+        self.ask_error = _label()
+        layout.addWidget(self.ask_error)
+        self.ask_output = QPlainTextEdit()
+        self.ask_output.setReadOnly(True)
+        self.ask_output.setAccessibleName(t("Assistant reply"))
+        self.ask_output.setPlaceholderText(t("The assistant's reply will appear here."))
+        self.ask_output.setMinimumHeight(160)
+        layout.addWidget(self.ask_output, 1)
+        layout.addWidget(_button(t("Copy"), lambda: QApplication.clipboard().setText(self.ask_output.toPlainText())))
+        return page
+
+    def _history_page(self):
+        page = QWidget()
+        layout = QVBoxLayout(page)
+        layout.setContentsMargins(0, 0, 0, 0)
+        layout.addWidget(_button(t("Back to dictation"), lambda: self.show_mode("dictation")))
+        layout.addWidget(self.settings.task_pages["history"], 1)
+        self.settings.task_pages["history"].show()
+        layout.addWidget(_button(t("Open selected text"), self._open_selected))
+        self.settings.history.itemDoubleClicked.connect(self._open_selected)
+        return page
+
+    def _open_selected(self, *_):
+        rows = self.settings._selected_rows()
+        if rows:
+            self._open_text("\n\n".join(row.get("text", "") for row in rows))
+
+    def _open_text(self, text):
+        if not text:
+            return
+        document = QDialog(self)
+        document.setWindowTitle(t("Transcript"))
+        document.resize(600, 500)
+        layout = QVBoxLayout(document)
+        editor = QPlainTextEdit()
+        editor.setReadOnly(True)
+        editor.setPlainText(text)
+        layout.addWidget(editor)
+        layout.addWidget(_button(t("Copy"), lambda: QApplication.clipboard().setText(text)))
+        document.setAttribute(Qt.WidgetAttribute.WA_DeleteOnClose)
+        document.show()
+
+    def show_mode(self, mode):
+        if mode not in self.mode_pages:
+            return
+        self.mode = mode
+        self.pages.setCurrentWidget(self.mode_pages[mode])
+        self.mode_buttons["dictation" if mode == "history" else mode].setChecked(True)
+        if mode == "history":
+            self.settings._load_history()
+        elif mode == "meeting":
+            self.settings._load_minutes()
+        self.refresh()
+
+    def showEvent(self, event):
+        super().showEvent(event)
+        self._timer.start()
+        self.refresh_results()
+        self.refresh()
+
+    def hideEvent(self, event):
+        self._timer.stop()
+        super().hideEvent(event)
+
+    def refresh_results(self, *_):
+        rows = cfg.read_history(self.conf["history_limit"])
+        result = next((row for row in reversed(rows) if row.get("mode") != "ask"), {})
+        answer = next((row for row in reversed(rows) if row.get("mode") == "ask"), {})
+        self._set_result(result, answer)
+
+    def _set_result(self, result, answer):
+        if result != self._last_result:
+            self._last_result = dict(result)
+            self.latest_text.setPlainText(result.get("text", ""))
+            self.latest_time.setText(result.get("ts", ""))
+            self.latest_warning.setText(result.get("cleanup_error", ""))
+            self.latest_warning.setVisible(bool(self.latest_warning.text()))
+            self.copy_button.setEnabled(bool(result.get("text")))
+            self.open_button.setEnabled(bool(result.get("text")))
+        if answer != self._last_answer:
+            self._last_answer = dict(answer)
+            self.ask_output.setPlainText(answer.get("text", ""))
+
+    def refresh(self, *_):
+        app, conf = self.controller, self.conf
+        if self._theme_name != conf["theme"]:
+            self._theme_name = conf["theme"]
+            theme.apply(self, self._theme_name)
+            colors = theme.palette(self._theme_name)
+            self._mic_icon = microphone_icon(color=colors["accent_text"])
+            self._stop_icon = microphone_icon(recording=True, color=colors["accent_text"])
+            self.settings_button.setIcon(settings_icon(colors["muted"]))
+        messages = getattr(app, "home_messages", {})
+        ready = conf.transcribe_ready()
+        recording = app.state == "recording"
+        busy = app.state == "busy"
+        title = t("Ready to speak") if ready else t("Set up transcription")
+        if recording:
+            seconds = int(app._recorded_seconds())
+            title = t("Paused") if app.paused else t("Recording")
+            title += f"  {seconds // 60:02d}:{seconds % 60:02d}"
+        elif busy:
+            title = messages.get("dictation_stage") or t("Transcribing…")
+        self.capture_status.setText(title)
+        action = t("Stop and transcribe") if recording else t("Start recording") if ready else t("Set up transcription")
+        self.capture_button.setAccessibleName(action)
+        self.capture_button.setIcon(self._stop_icon if recording else self._mic_icon)
+        self.capture_button.setToolTip(action)
+        self.capture_button.setEnabled(recording or not app.recording)
+        self.capture_shortcut.setText(" + ".join(part.strip() for part in conf["shortcut"].split("+")))
+        self.pause_button.setVisible(recording)
+        self.cancel_button.setVisible(recording)
+        self.pause_button.setText(t("Resume the recording") if app.paused else t("Pause the recording"))
+        warning = ""
+        if conf["cleanup_enabled"] and conf["cleanup_provider"] == "local" and not conf.local_llm_ready():
+            warning = t("The local editing model is missing. Set it up in Settings; the original transcript is kept if editing fails.")
+        self.capture_error.setText(messages.get("dictation", "") or warning)
+        self.capture_error.setVisible(bool(self.capture_error.text()))
+        details = processing_locations(conf, self.mode, self.settings.file_cleanup.isChecked(),
+                                       self.settings.file_timestamps.isChecked())
+        self.capture_models.setText(details if self.mode == "dictation" else "")
+        self.footer.setText(details)
+        self.footer.setVisible(self.mode != "dictation")
+        self._refresh_meeting(messages, ready)
+        self._refresh_ask(messages, ready)
+
+    def _refresh_meeting(self, messages, ready):
+        app = self.controller
+        state = app.meeting_state
+        supported = audio.sound().meetings
+        hint = t("Record your microphone and the other participants. Use headphones.")
+        if not supported:
+            hint = t("Meeting recording is not supported on this system. You can still transcribe a file.")
+        elif audio.sound() is audio.COREAUDIO:
+            hint = t("On macOS, set up BlackHole or Loopback and select the system audio source in Settings first.")
+        title = t("Meeting")
+        if state == "recording":
+            seconds = int(app.meeting_elapsed.elapsed() / 1000)
+            title = t("Recording") + f"  {seconds // 60:02d}:{seconds % 60:02d}"
+        elif state == "working":
+            title = app.meeting_message or t("Writing the meeting up…")
+        if supported and not self.conf.openrouter_key():
+            hint += "\n" + t("Connect OpenRouter in Settings to write minutes. The recording is kept if writing fails.")
+        self.meeting_status.setText(title)
+        self.meeting_hint.setText(hint)
+        self.meeting_button.setText(t("End the meeting and write it up") if state == "recording" else t("Record a meeting") if ready else t("Set up transcription"))
+        self.meeting_button.setEnabled(supported and state != "working")
+        self.meeting_cancel.setVisible(state == "recording")
+        self.meeting_error.setText(messages.get("meeting", ""))
+        self.meeting_error.setVisible(bool(self.meeting_error.text()))
+
+    def _refresh_ask(self, messages, ready):
+        app, conf = self.controller, self.conf
+        name = assistant.display_name(conf)
+        provider = conf["assistant_provider"]
+        directory = assistant.working_dir(conf)
+        if provider == "claude":
+            permission = {"auto": t("Automatic permission decisions"),
+                          "manual": t("Only actions that need no permission"),
+                          "bypassPermissions": t("All permissions allowed")}.get(conf["assistant_permission_mode"], conf["assistant_permission_mode"])
+        elif provider == "codex":
+            permission = {"workspace-write": t("Read files; write in the working directory"),
+                          "read-only": t("Read only"),
+                          "danger-full-access": t("No sandbox at all")}.get(conf["assistant_codex_sandbox"], conf["assistant_codex_sandbox"])
+        elif provider == "agy":
+            permission = t("Uses the CLI's configured permissions")
+        else:
+            permission = t("Chat provider; no local command execution")
+        self.ask_scope.setText(t("{name}\nPermissions: {permission}\nWorking directory: {directory}\nShortcut: {shortcut}\nThis button sends a spoken command to the assistant. Its reply is copied without automatic pasting.", name=name, permission=permission, directory=directory if provider in ("claude", "codex", "agy") else t("Not used"), shortcut=conf["assistant_shortcut"] or t("Not assigned")))
+        state = app.ask_state
+        self.ask_status.setText(t("Paused") if state == "recording" and app.paused else t("Recording") if state == "recording" else messages.get("ask_stage", t("Working…")) if state == "busy" else t("Assistant"))
+        available = self._assistant_available()
+        self.ask_button.setText(t("Stop and send command") if state == "recording" else t("Set up transcription") if not ready else t("Record a command") if available else t("Set up assistant"))
+        self.ask_button.setEnabled(state == "recording" or (state == "idle" and not app.recording))
+        self.ask_pause.setVisible(state == "recording")
+        self.ask_pause.setText(t("Resume the recording") if app.paused else t("Pause the recording"))
+        self.ask_cancel.setVisible(state != "idle")
+        self.ask_error.setText(messages.get("ask", "") or ("" if available else t("Install the selected assistant CLI or configure its connection in Settings.")))
+        self.ask_error.setVisible(bool(self.ask_error.text()))
+
+    def _assistant_available(self):
+        provider = self.conf["assistant_provider"]
+        binary = assistant.executable(provider)
+        if binary:
+            return bool(shutil.which(binary))
+        return bool(self.conf.opencode_key() if provider == "opencode" else self.conf.openrouter_key())
+
+    def _capture(self):
+        app = self.controller
+        if app.state == "recording":
+            # Clicking Stop puts this window in front of the original target.
+            app.paste_override["dictation"] = False
+            app.stop()
+        elif not app.recording:
+            if not self.conf.transcribe_ready():
+                self.settings.tabs.setCurrentIndex(self.settings.api_tab_index)
+                app.open_settings()
+                return
+            app.paste_override["dictation"] = False
+            app.start()
+            if app.state != "recording":
+                app.paste_override.pop("dictation", None)
+        self.refresh()
+
+    def _ask(self):
+        app = self.controller
+        if app.ask_state == "recording":
+            app.paste_override["ask"] = False
+            app.stop_ask()
+        elif app.ask_state == "idle" and not app.recording:
+            if not self.conf.transcribe_ready():
+                app.open_settings()
+                return
+            if not self._assistant_available():
+                self.settings.tabs.setCurrentIndex(4)
+                app.open_settings()
+                return
+            app.paste_override["ask"] = False
+            app.start_ask()
+            if app.ask_state != "recording":
+                app.paste_override.pop("ask", None)
+        self.refresh()
+
+    def _pause(self):
+        self.controller._toggle_pause()
+        self.refresh()
+
+    def _cancel_capture(self):
+        if self.controller.state == "recording":
+            self.controller._cancel()
+        self.refresh()
+
+    def _cancel_ask(self):
+        self.controller.cancel_ask()
+        self.refresh()
+
+    def _meeting(self):
+        if not audio.sound().meetings:
+            return
+        if not self.conf.transcribe_ready() and self.controller.meeting_state == "idle":
+            self.controller.open_settings()
+        else:
+            self.controller._toggle_meeting()
+        self.refresh()
+
+    def _cancel_meeting(self):
+        self.controller.cancel_meeting()
+        self.refresh()

dikte/hotkey.py

@@ -57,7 +57,7 @@ SHORTCUTS = {
                       "Dikte: pause/resume the recording", "pause_shortcut", ""),
     "cancel": Shortcut("cancel", CANCEL_DESKTOP_ID, "Dikte: discard the recording",
                        "cancel_shortcut", ""),
-    "ask": Shortcut("ask", ASK_DESKTOP_ID, "Dikte: ask Claude Code",
+    "ask": Shortcut("ask", ASK_DESKTOP_ID, "Dikte: ask the agent",
                     "assistant_shortcut", ""),
     "meeting": Shortcut("meeting", MEETING_DESKTOP_ID,
                         "Dikte: start/end a meeting recording",

dikte/i18n.py

@@ -60,6 +60,75 @@ def name(text, /, case=""):
 
 
 TR = {
+    "Theme": "Tema",
+    "Classic dark": "Klasik karanlık",
+    "Classic light": "Klasik beyaz",
+    "Local": "Yerel",
+    "Local GPU": "Yerel GPU",
+    "Local CPU": "Yerel CPU",
+    "Model not selected": "Model seçilmedi",
+    "{name} default model": "{name} varsayılan modeli",
+    "Dictation: {sound} / Cleanup: {text}": "Dikte: {sound} / Temizleme: {text}",
+    "Minutes: {model}": "Tutanak: {model}",
+    "Assistant: {model}": "Asistan: {model}",
+    "Completed with a warning: {error}": "Uyarıyla tamamlandı: {error}",
+    "Dikte: completed with a warning": "Dikte: uyarıyla tamamlandı",
+    "Connect OpenRouter in Settings to write minutes. The recording is kept if writing fails.": "Tutanak yazmak için ayarlardan OpenRouter bağlantısını kurun. Tutanak yazılamazsa kayıt korunur.",
+    "Set up assistant": "Asistanı ayarla",
+    "Install the selected assistant CLI or configure its connection in Settings.": "Seçili asistanın CLI uygulamasını kurun veya ayarlardan bağlantısını yapılandırın.",
+    "The local editing model is missing. Set it up in Settings; the original transcript is kept if editing fails.": "Yerel düzenleme modeli eksik. Ayarlardan yapılandırın; düzenleme başarısız olursa ham metin korunur.",
+    # Compact desktop workspace.
+    'Audio: {sound}\nText: {text}': 'Ses: {sound}\nMetin: {text}',
+    'This computer': 'Bu bilgisayar',
+    'Editing off': 'Düzenleme kapalı',
+    'Record your microphone and the other participants. Use headphones.': 'Mikrofonunuzu ve diğer katılımcıları kaydedin. Kulaklık kullanın.',
+    'Minutes: OpenRouter': 'Tutanak: OpenRouter',
+    'Use the shortcut in the app where you want to write.': 'Yazmak istediğiniz uygulamada kısayola basın.',
+    'Button recordings stay here and are copied to the clipboard.': 'Düğmeyle başlattığınız kayıtların metni burada kalır ve panoya kopyalanır.',
+    'Latest text': 'Son metin',
+    'Your first transcript will appear here.': 'İlk kaydınızın metni burada görünecek.',
+    'Open text': 'Metni aç',
+    'Assistant reply': 'Asistan yanıtı',
+    "The assistant's reply will appear here.": 'Asistanın yanıtı burada görünecek.',
+    'Ready to speak': 'Konuşmaya hazır',
+    'Set up transcription': 'Yazıya çevirmeyi ayarla',
+    'Automatic language': 'Dil otomatik',
+    '{mic} / {language}': '{mic} / {language}',
+    'Meeting recording is not supported on this system. You can still transcribe a file.': 'Bu sistemde toplantı kaydı desteklenmiyor. Ses dosyalarını yazıya çevirebilirsiniz.',
+    '{name}\nPermissions: {permission}\nWorking directory: {directory}\nShortcut: {shortcut}\nThis button sends a spoken command to the assistant. Its reply is copied without automatic pasting.': '{name}\nYetkiler: {permission}\nÇalışma dizini: {directory}\nKısayol: {shortcut}\nBu düğme asistana sesli komut gönderir. Yanıtı otomatik yapıştırılmadan kopyalanır.',
+    '{name} CLI (provider connection)': '{name} CLI (sağlayıcı bağlantısı)',
+    'Assistant: {name}': 'Asistan: {name}',
+    'Settings': 'Ayarlar',
+    'Back to dictation': 'Dikteye dön',
+    'Open selected text': 'Seçili metni aç',
+    'Paused': 'Duraklatıldı',
+    'Choose a connection or download a local model in Settings.': 'Ayarlardan bir bağlantı seçin veya yerel model indirin.',
+    'On macOS, set up BlackHole or Loopback and select the system audio source in Settings first.': 'macOS üzerinde önce BlackHole veya Loopback kurun ve ayarlardan sistem sesi kaynağını seçin.',
+    'Stop and send command': 'Bitir ve komutu gönder',
+    "Uses the CLI's configured permissions": 'CLI için yapılandırılmış yetkileri kullanır',
+    'Chat provider; no local command execution': 'Sohbet sağlayıcısı; yerel komut çalıştırmaz',
+    'Record a command': 'Sesli komut kaydet',
+    'Automatic permission decisions': 'İzin kararları otomatik',
+    'Only actions that need no permission': 'Yalnız izin gerektirmeyen işlemler',
+    'All permissions allowed': 'Tüm izinler açık',
+    'Not used': 'Kullanılmıyor',
+    'Not assigned': 'Atanmamış',
+    'Assistant': 'Asistan',
+    'Read files; write in the working directory': 'Dosyaları oku; çalışma dizinine yaz',
+    'Open Dikte': 'Dikteyi aç',
+    '{error}\n\nThe recording has been kept. Meeting → Minutes can try again.': '{error}\n\nKayıt korundu. Toplantı → Tutanaklar bölümünden yeniden deneyebilirsiniz.',
+    'Original text kept, cleanup failed: {error}': 'Ham metin korundu, düzenleme başarısız: {error}',
+    'Transcript ready: {preview}': 'Metin hazır: {preview}',
+    'Editing did not finish. The original text was kept. {error}': 'Düzenleme tamamlanamadı. Ham metin korundu. {error}',
+    'Transcript ready': 'Metin hazır',
+    'Text editing and dictionary': 'Metin düzenleme ve sözlük',
+    'Settings category': 'Ayar kategorisi',
+    'Apply changes': 'Değişiklikleri uygula',
+    'Discard changes': 'Değişikliklerden vazgeç',
+    'Unsaved changes': 'Kaydedilmemiş değişiklikler',
+    'Dictionary': 'Sözlük',
+    'File': 'Dosya',
+
     # --- tray ---------------------------------------------------------
     "Start recording": "Kaydı başlat",
     "Stop and transcribe": "Kaydı bitir ve yaz",
@@ -804,8 +873,22 @@ TR = {
         "deneyin.",
     "Local whisper": "Yerel whisper",
     "Local model": "Yerel model",
+    "Not loaded.": "Yüklü değil.",
+    "Loaded; it did not say what it is running on.":
+        "Yüklendi; neyin üzerinde çalıştığını söylemedi.",
+    "Loaded on the graphics card ({detail}).":
+        "Ekran kartına yüklendi ({detail}).",
+    "Loaded on the processor ({detail}).": "İşlemciye yüklendi ({detail}).",
+    "Loaded on the processor: only the CPU backend was loaded. Check the "
+    "server log for graphics backend or driver errors.":
+        "İşlemciye yüklendi: yalnızca CPU arka ucu yüklendi. Ekran kartı arka "
+        "ucu veya sürücü hataları için sunucu günlüğünü kontrol edin.",
+    "Loaded on the processor: the graphics card is switched on, but could not "
+    "be used.":
+        "İşlemciye yüklendi: ekran kartı açık, ama kullanılamadı.",
     "Not installed.": "Kurulu değil.",
     "Installed on the system: {path}": "Sistemde kurulu: {path}",
+    "Using custom build: {path}": "Özel derleme kullanılıyor: {path}",
     "Download again": "Yeniden indir",
     "Downloaded, version {version}.": "İndirildi, sürüm {version}.",
     "Downloaded, version {version}. There was no Vulkan build, "

dikte/icons/check-light.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="14" height="14" viewBox="0 0 14 14"><path d="m3.3 7 2.4 2.5 5-5" fill="none" stroke="#FFFFFF" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>

dikte/icons/check.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="14" height="14" viewBox="0 0 14 14"><path d="m3.3 7 2.4 2.5 5-5" fill="none" stroke="#172434" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>

dikte/icons/chevron-down-light.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="12" height="12" viewBox="0 0 12 12"><path d="m3 4.5 3 3 3-3" fill="none" stroke="#526174" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/></svg>

dikte/icons/chevron-down.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="12" height="12" viewBox="0 0 12 12"><path d="m3 4.5 3 3 3-3" fill="none" stroke="#B2C1D1" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/></svg>

dikte/icons/chevron-up-light.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="12" height="12" viewBox="0 0 12 12"><path d="m3 7.5 3-3 3 3" fill="none" stroke="#526174" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/></svg>

dikte/icons/chevron-up.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="12" height="12" viewBox="0 0 12 12"><path d="m3 7.5 3-3 3 3" fill="none" stroke="#B2C1D1" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"/></svg>

dikte/overlay.py

@@ -9,33 +9,18 @@ from PyQt6.QtGui import QColor, QCursor, QFont, QPainter, QPainterPath, QPen, QF
 from PyQt6.QtWidgets import QWidget, QApplication
 
 from . import mac_window
+from . import theme
 
 BARS = 22
-HEIGHT = 56
+HEIGHT = 48
 MIN_WIDTH = 210
 MAX_WIDTH = 460
 MARGIN = 28
 GAP = 10        # between two indicators sharing a corner
 FOLLOW_EVERY = 8   # ticks between two looks for the pointer: about four a second
 
-BG = QColor(22, 24, 29, 238)
+STATE_ROLES = {"recording": "rec", "asking": "ask", "meeting": "rec", "busy": "accent",
-BORDER = QColor(255, 255, 255, 28)
+               "done": "ok", "warning": "warn", "error": "err"}
-TEXT = QColor(235, 237, 242)
-MUTED = QColor(150, 156, 168)
-REC = QColor(240, 78, 82)
-BUSY = QColor(120, 170, 255)
-OK = QColor(80, 205, 140)
-ERR = QColor(240, 100, 90)
-WARN = QColor(240, 180, 80)
-THEM = QColor(110, 190, 255)   # the other side of a meeting
-
-ASK = QColor(150, 140, 255)    # recording a command rather than a dictation
-# Recording, but nothing is going in. The same amber a warning gets, and for
-# the same reason: it is the colour that stops you walking away from it.
-HELD = WARN
-
-STATE_COLORS = {"recording": REC, "asking": ASK, "meeting": REC, "busy": BUSY,
-                "done": OK, "warning": WARN, "error": ERR}
 LIVE = ("recording", "asking", "meeting")
 
 
@@ -94,8 +79,9 @@ class Overlay(QWidget):
     under way at the same time and still both be visible."""
 
     def __init__(self, corner="bottom-left", below=None, dismissable=False,
-                 screen_name="", follow_pointer=False):
+                 screen_name="", follow_pointer=False, theme_name=theme.DEFAULT):
         super().__init__(None)
+        self.set_theme(theme_name)
         self.corner = corner
         self.screen_name = screen_name
         # Whether it goes on following the pointer once it is up, rather than
@@ -388,6 +374,15 @@ class Overlay(QWidget):
 
     # ---- painting --------------------------------------------------
 
+    def set_theme(self, name):
+        self.colors = {key: QColor(value) for key, value in theme.palette(name).items()}
+        states = (dict(rec="#D52E3F", ok="#187B4B", err="#BD2735", warn="#946000",
+                       them="#2460A0", ask="#7048B4") if name == "light" else
+                  dict(rec="#F04E52", ok="#50CD8C", err="#F0645A", warn="#F0B450",
+                       them="#6EBEFF", ask="#968CFF"))
+        self.colors.update({key: QColor(value) for key, value in states.items()})
+        self.update()
+
     def paintEvent(self, _event):
         if self.state == "hidden":
             return  # translucent window, nothing drawn means nothing shown
@@ -397,14 +392,14 @@ class Overlay(QWidget):
         rect = QRectF(0.5, 0.5, self.width() - 1, self.height() - 1)
 
         path = QPainterPath()
-        path.addRoundedRect(rect, 15, 15)
+        path.addRoundedRect(rect, 12, 12)
-        painter.fillPath(path, BG)
+        painter.fillPath(path, self.colors["base"])
-        painter.setPen(QPen(BORDER, 1))
+        painter.setPen(QPen(self.colors["border"], 1))
         painter.drawPath(path)
 
-        accent = STATE_COLORS.get(self.state, MUTED)
+        accent = self.colors[STATE_ROLES.get(self.state, "muted")]
         if self._held:
-            accent = HELD
+            accent = self.colors["warn"]
         self._draw_indicator(painter, accent)
 
         if self.state in LIVE:
@@ -480,13 +475,13 @@ class Overlay(QWidget):
         gap = (right - left - BARS * bar_w) / max(1, BARS - 1)
         return left, bar_w, bar_w + gap
 
-    @staticmethod
+    def _bar_colour(self, shaped, accent):
-    def _bar_colour(shaped, accent):
+        color = QColor(accent if shaped > 0.04 else self.colors["muted"])
-        color = QColor(accent if shaped > 0.04 else MUTED)
         color.setAlphaF(0.35 + 0.65 * min(1.0, shaped * 2.2))
         return color
 
-    def _draw_waveform(self, painter, accent=REC):
+    def _draw_waveform(self, painter, accent=None):
+        accent = accent if accent is not None else self.colors["rec"]
         if self.state == "meeting":
             self._draw_dual_waveform(painter)
             return
@@ -512,7 +507,7 @@ class Overlay(QWidget):
         painter.setPen(Qt.PenStyle.NoPen)
         for i, (mine, theirs) in enumerate(zip(self.levels, self.levels2)):
             x = left + i * step
-            for level, accent, up in ((mine, REC, True), (theirs, THEM, False)):
+            for level, accent, up in ((mine, self.colors["rec"], True), (theirs, self.colors["them"], False)):
                 shaped = min(1.0, level ** 0.55)
                 h = 2.0 + shaped * 12.0
                 y = mid - 1.5 - h if up else mid + 1.5
@@ -524,7 +519,7 @@ class Overlay(QWidget):
         font.setPointSizeF(10.0)
         font.setFamilies(["monospace"])
         painter.setFont(font)
-        painter.setPen(MUTED)
+        painter.setPen(self.colors["muted"])
         mins, secs = divmod(int(self.seconds), 60)
         hours, mins = divmod(mins, 60)
         text = f"{hours}:{mins:02d}:{secs:02d}" if hours else f"{mins}:{secs:02d}"
@@ -543,7 +538,7 @@ class Overlay(QWidget):
         """A faint cross on the right: without it there is nothing to say the
         box can be clicked away, and a feature nobody can see is not one."""
         cx, cy = self.width() - 18.0, self.height() / 2
-        pen = QPen(QColor(MUTED), 1.6)
+        pen = QPen(QColor(self.colors["muted"]), 1.6)
         pen.setCapStyle(Qt.PenCapStyle.RoundCap)
         painter.setPen(pen)
         painter.setBrush(Qt.BrushStyle.NoBrush)
@@ -552,7 +547,7 @@ class Overlay(QWidget):
 
     def _draw_message(self, painter):
         painter.setFont(self._label_font())
-        painter.setPen({"error": ERR, "warning": WARN}.get(self.state, TEXT))
+        painter.setPen({"error": self.colors["err"], "warning": self.colors["warn"]}.get(self.state, self.colors["text"]))
         # Leave the cross its corner rather than running the text under it.
         box = QRectF(46, 0, self.width() - 60 - (18 if self._can_dismiss else 0),
                      self.height())

dikte/settings_ui.py

@@ -6,13 +6,13 @@ import shutil
 import sys
 import threading
 
-from PyQt6.QtCore import QEvent, QObject, QRect, Qt, QTimer, QUrl, pyqtSignal
+from PyQt6.QtCore import QEvent, QObject, QRect, Qt, QTimer, QUrl, QSignalBlocker, pyqtSignal
 from PyQt6.QtGui import QDesktopServices, QGuiApplication, QKeySequence, QShortcut
 from PyQt6.QtWidgets import (
     QAbstractItemView, QAbstractSpinBox, QCheckBox, QComboBox, QDialog,
     QDialogButtonBox, QFileDialog, QFormLayout, QGroupBox, QHBoxLayout, QLabel,
     QLineEdit, QListWidget, QListWidgetItem, QMenu, QMessageBox, QPlainTextEdit,
-    QPushButton, QScrollArea, QSpinBox, QTabWidget, QVBoxLayout, QWidget,
+    QPushButton, QScrollArea, QSizePolicy, QSpinBox, QTabWidget, QVBoxLayout, QWidget,
 )
 
 from . import __version__
@@ -32,6 +32,7 @@ from . import paste
 from . import update
 from .filetranscribe import FileTranscriber
 from .i18n import t
+from . import theme
 
 UI_LANGUAGES = [("Automatic (system)", "auto"), ("Turkish", "tr"), ("English", "en")]
 LANGUAGES = [
@@ -223,7 +224,9 @@ class WheelGuard(QObject):
     """
 
     def eventFilter(self, box, event):
-        if event.type() == QEvent.Type.Wheel and not box.hasFocus():
+        win = box.window()
+        focused = box.hasFocus() or (win is not None and win.focusWidget() is box)
+        if event.type() == QEvent.Type.Wheel and not focused:
             # Refused rather than swallowed. An unaccepted wheel event carries
             # on up the parents to the scroll area, so the page still moves.
             event.ignore()
@@ -253,9 +256,11 @@ class LocalModelBox(QGroupBox):
 
     changed = pyqtSignal()
 
-    def __init__(self, program, title, models, model_path, repos=None, parent=None):
+    def __init__(self, program, title, models, model_path, binary=None,
+                 repos=None, parent=None):
         super().__init__(title, parent)
         self.program = program
+        self._binary = binary          # () -> a path set by hand, or ""
         self._models = models          # () -> [hub.Item], or (repo) -> [hub.Item]
         self._model_path = model_path  # (name) -> Path
         self._repos = repos            # None, or () -> [repo id]
@@ -279,6 +284,8 @@ class LocalModelBox(QGroupBox):
         self._later.timeout.connect(self._later_fetch)
 
         form = QFormLayout(self)
+        form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
 
         self.program_label = WrappedLabel()
         self.install_button = QPushButton(t("Download"))
@@ -376,6 +383,8 @@ class LocalModelBox(QGroupBox):
         layout.setContentsMargins(0, 0, 0, 0)
         for index, widget in enumerate(widgets):
             layout.addWidget(widget, 1 if index == 0 else 0)
+        if widgets and all(isinstance(widget, QPushButton) for widget in widgets):
+            layout.addStretch()
         holder = QWidget()
         holder.setLayout(layout)
         return holder
@@ -391,13 +400,17 @@ class LocalModelBox(QGroupBox):
     def load(self, model, repo=""):
         """Show what is stored. What else is on offer is asked for on the way up.
 
-        Nothing is fetched here: building the settings window is not the same as
+        Hidden boxes defer fetching until shown. A catalog already fetched for
-        opening it, and a list nobody is looking at is not worth a request. What
+        this repository survives Apply; a visible box on a new repository
-        is already on this disk is shown straight away either way.
+        refreshes immediately. Installed files are available in either case.
         """
+        target_repo = repo or (ggml.suggested_llm()[0] if self._repos is not None else "")
+        reuse = self._answered and self._chosen_in == target_repo and self.repository() == target_repo
+        items = self._current_items() if reuse else []
+        self._later.stop()
         self._wanted = model
-        self._pending = True
+        self._pending = not reuse
-        self._answered = False
+        self._answered = reuse
         self._show_program()
         self._chosen_in = ""
         if self._repos is not None:
@@ -407,7 +420,10 @@ class LocalModelBox(QGroupBox):
             self.repo.setCurrentText(self._chosen_in)
             self.repo.blockSignals(False)
             self._fill_repos_box(suggested)
-        self._fill_models([])
+        self._fill_models(items)
+        if self._pending and self.isVisible():
+            self._pending = False
+            self._fetch_models(self.repository())
 
     def showEvent(self, event):
         super().showEvent(event)
@@ -417,13 +433,23 @@ class LocalModelBox(QGroupBox):
                 self._fill_repos(self.repository())
             self._fetch_models(self.repository())
 
+    def _program_path(self):
+        return ggml.program_path(self.program,
+                                 self._binary() if self._binary else "")
+
     def _show_program(self):
-        path = ggml.program_path(self.program)
+        path = self._program_path()
         if not path:
             self.program_label.setText(t("Not installed."))
             self.install_button.setText(t("Download"))
             self.install_button.setVisible(True)
             return
+        if self._binary and self._binary():
+            # Neither a system copy nor one Dikte fetched, and "Downloaded"
+            # over a build someone made themselves is not true.
+            self.program_label.setText(t("Using custom build: {path}", path=path))
+            self.install_button.setVisible(False)
+            return
         # A copy that is here is not a copy that is right. whisper.cpp releases
         # every few weeks, and a graphics card installed after Dikte was
         # changes which build this machine should be running; the button was
@@ -777,6 +803,10 @@ class LocalModelBox(QGroupBox):
 
     def _fill_models_from_current(self):
         """Redraw the rows without asking anybody anything again."""
+        self._wanted = self.selected()
+        self._fill_models(self._current_items())
+
+    def _current_items(self):
         # By name, because the recommended model has a row of its own at the
         # top as well as one in its group, and reading the rows back twice
         # would double it in the list every time a download finished.
@@ -786,8 +816,7 @@ class LocalModelBox(QGroupBox):
             if item is not None and item.name not in seen:
                 seen.add(item.name)
                 items.append(item)
-        self._wanted = self.selected()
+        return items
-        self._fill_models(items)
 
     def _delete(self):
         name = self.selected()
@@ -835,7 +864,7 @@ class LocalModelBox(QGroupBox):
                   repo=self.repository(), cap=ggml.human_size(ggml.GGUF_MAX_BYTES)))
         elif not name:
             self.status.setText(t("Nothing downloaded yet."))
-        elif here and not ggml.program_path(self.program):
+        elif here and not self._program_path():
             # The model alone runs nothing, and "Ready" over a missing program
             # reads as though it does.
             self.status.setText(t("{name} is here, but the program above is "
@@ -904,6 +933,7 @@ class SettingsWindow(QDialog):
         self._release_url = update.RELEASES_PAGE
         self.transcriber = FileTranscriber(conf, self)
         self.setWindowTitle(t("Dikte Settings"))
+        theme.apply(self, conf["theme"])
 
         # One for the whole window, parented to it so it outlives the boxes it
         # watches and goes when they do.
@@ -914,23 +944,46 @@ class SettingsWindow(QDialog):
         tabs.addTab(self._scrolled(self._display_tab()), t("Display"))
         self.api_tab_index = tabs.addTab(
             self._scrolled(self._api_tab()), t("API and models"))
-        tabs.addTab(self._scrolled(self._prompt_tab()), t("Cleanup rules"))
+        tabs.addTab(self._scrolled(self._prompt_tab()), t("Text editing and dictionary"))
         tabs.addTab(self._scrolled(self._assistant_tab()), t("Agent"))
         tabs.addTab(self._scrolled(self._meeting_tab()), t("Meeting"))
-        tabs.addTab(self._scrolled(self._minutes_tab()), t("Minutes"))
-        tabs.addTab(self._scrolled(self._file_tab()), t("Audio file"))
         tabs.addTab(self._scrolled(self._shortcut_tab()), t("Shortcuts"))
-        tabs.addTab(self._scrolled(self._history_tab()), t("History"))
+        self.task_pages = {
+            "file": self._scrolled(self._file_tab()),
+            "minutes": self._scrolled(self._minutes_tab()),
+            "history": self._scrolled(self._history_tab()),
+        }
+        for page in self.task_pages.values():
+            page.setParent(self)
+            page.hide()
+        retention = QGroupBox(t("History"))
+        retention_form = QFormLayout(retention)
+        retention_form.addRow(t("Keep at most"), self.history_limit)
+        general_layout = tabs.widget(0).widget().layout()
+        general_layout.addRow(retention)
+        tabs.tabBar().hide()
+        self.categories = QComboBox()
+        self.categories.setAccessibleName(t("Settings category"))
+        for index in range(tabs.count()):
+            self.categories.addItem(tabs.tabText(index))
+        self.categories.currentIndexChanged.connect(tabs.setCurrentIndex)
+        tabs.currentChanged.connect(self.categories.setCurrentIndex)
 
         # Save keeps the window open, so the window is closed with the titlebar
         # cross (or Escape) instead. A "Cancel" next to it would be a lie: the
         # settings are already on disk by then.
         buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Save)
-        buttons.button(QDialogButtonBox.StandardButton.Save).setText(t("Save"))
+        buttons.button(QDialogButtonBox.StandardButton.Save).setText(t("Apply changes"))
+        discard = buttons.addButton(t("Discard changes"), QDialogButtonBox.ButtonRole.ResetRole)
+        discard.clicked.connect(self._discard_changes)
+        self.dirty_label = QLabel("")
+        self.dirty_label.setObjectName("muted")
         buttons.accepted.connect(self._save)
 
         layout = QVBoxLayout(self)
+        layout.addWidget(self.categories)
         layout.addWidget(tabs)
+        layout.addWidget(self.dirty_label)
         layout.addWidget(buttons)
         self._size_to_screen(680, 640)
 
@@ -951,6 +1004,26 @@ class SettingsWindow(QDialog):
             self.meetings.finished.connect(self._on_minutes_finished)
             self.meetings.failed.connect(self._on_minutes_failed)
         self._load()
+        for button in self.findChildren(QPushButton):
+            button.setSizePolicy(QSizePolicy.Policy.Fixed, QSizePolicy.Policy.Fixed)
+        compact = self.tabs.findChildren(QSpinBox) + [
+            self.theme_choice,
+            self.ui_language, self.language, self.paste_shortcut, self.corner,
+            self.cleanup_reasoning, self.local_llm_reasoning,
+            self.assistant_reasoning, self.meeting_reasoning,
+        ]
+        for box in compact:
+            box.setSizePolicy(QSizePolicy.Policy.Fixed, QSizePolicy.Policy.Fixed)
+        self._saved_form = self._form_values()
+        for box in self.tabs.findChildren((QLineEdit, QComboBox, QCheckBox, QSpinBox, QPlainTextEdit)):
+            if isinstance(box, (QLineEdit, QPlainTextEdit)):
+                box.textChanged.connect(self._show_dirty)
+            elif isinstance(box, QComboBox):
+                box.currentTextChanged.connect(self._show_dirty)
+            elif isinstance(box, QCheckBox):
+                box.toggled.connect(self._show_dirty)
+            else:
+                box.valueChanged.connect(self._show_dirty)
         self._load_codex_models()
         self._load_agy_models()
         self._load_hosted_models()
@@ -962,6 +1035,115 @@ class SettingsWindow(QDialog):
         # because of that, so open it on the tab that fixes it.
         if not conf.transcribe_ready():
             self.tabs.setCurrentIndex(self.api_tab_index)
+        # A model takes up to ggml.STARTUP_TIMEOUT to load, so a line written
+        # once as the window opens would be wrong for most of the wait. Runs
+        # only while the window is on screen: there is nobody to read it
+        # otherwise, and it costs a lock and a poll() each time.
+        self._local_state_timer = QTimer(self)
+        self._local_state_timer.setInterval(2000)
+        self._local_state_timer.timeout.connect(self._show_local_state)
+        self._show_local_state()
+
+    def _form_values(self):
+        values = []
+        for box in self.tabs.findChildren((QLineEdit, QComboBox, QCheckBox, QSpinBox, QPlainTextEdit)):
+            if isinstance(box, QPlainTextEdit):
+                if not box.isReadOnly():
+                    values.append(box.toPlainText())
+            elif isinstance(box, QLineEdit):
+                values.append(box.text())
+            elif isinstance(box, QComboBox):
+                value = box.currentData()
+                values.append(box.currentText() if value is None else value)
+            elif isinstance(box, QCheckBox):
+                values.append(box.isChecked())
+            else:
+                values.append(box.value())
+        models = dict(self._models)
+        provider = self.transcribe_provider.currentData()
+        if provider in models:
+            models[provider] = self.transcribe_model.currentText().strip()
+        values.append(models)
+        return values
+
+    def refresh_configuration(self):
+        """Reload clean forms while keeping unsaved edits for a later merge."""
+        if self._form_values() == self._saved_form:
+            self._load()
+            self._saved_form = self._form_values()
+            self._show_dirty()
+
+    def refresh_sources(self):
+        """Discover new devices without replacing an in-progress selection."""
+        self._sources = audio.list_sources()
+        monitors = audio.list_monitors()
+        for box, title, sources in (
+            (self.mic, "Default microphone", self._sources),
+            (self.meeting_mic, "Same as dictation", self._sources),
+            (self.meeting_system, "Current output", monitors),
+        ):
+            selected = box.currentData() or ""
+            with QSignalBlocker(box):
+                box.clear()
+                box.addItem(t(title), "")
+                for name, description in sources:
+                    box.addItem(description, name)
+                if selected and box.findData(selected) < 0:
+                    box.addItem(t("{name} (not connected)", name=selected), selected)
+                self._select_data(box, selected)
+        self._show_dirty()
+
+    def _show_dirty(self, *_):
+        dirty = self._form_values() != getattr(self, "_saved_form", [])
+        self.dirty_label.setText(t("Unsaved changes") if dirty else "")
+
+    def _discard_changes(self):
+        self._load()
+        self._saved_form = self._form_values()
+        self._show_dirty()
+
+    def showEvent(self, event):
+        super().showEvent(event)
+        self._show_local_state()
+        self._local_state_timer.start()
+
+    def hideEvent(self, event):
+        self._local_state_timer.stop()
+        super().hideEvent(event)
+
+    def _show_local_state(self):
+        """What each model on this machine is loaded on, as it is now."""
+        local = ggml.state()
+        self.local_state.setText(
+            self._local_state_text(ggml.WHISPER, local.get("whisper", {})))
+        self.local_llm_state.setText(
+            self._local_state_text(ggml.LLAMA, local.get("llama", {})))
+
+    @staticmethod
+    def _local_state_text(program, entry):
+        """One line: whether the model is loaded, and what it ended up on.
+
+        Four answers rather than two, because "could not tell" is a real one: a
+        whisper built by hand on a Mac prints nothing about its backend, and
+        answering "the processor" there would be a confident lie about the one
+        thing this line exists to be honest about.
+        """
+        kind = ggml.accel_kind(entry)
+        if kind == "off":
+            return t("Not loaded.")
+        # The backend and the card keep the names the server printed for them.
+        detail = ggml.accel_detail(entry)
+        if kind == "unknown":
+            return t("Loaded; it did not say what it is running on.")
+        if kind == "gpu":
+            return t("Loaded on the graphics card ({detail}).", detail=detail)
+        if not entry.get("gpu_wanted"):
+            return t("Loaded on the processor ({detail}).", detail=detail)
+        if not ggml.cpu_only_loaded(entry):
+            return t("Loaded on the processor: the graphics card is switched "
+                     "on, but could not be used.")
+        return t("Loaded on the processor: only the CPU backend was loaded. "
+                 "Check the server log for graphics backend or driver errors.")
 
     def _scrolled(self, page):
         """A tab that scrolls instead of growing the window to fit."""
@@ -969,8 +1151,10 @@ class SettingsWindow(QDialog):
         # that height on as the window's minimum, and a tall one (the API tab
         # is the tallest, and taller still under a large interface font) then
         # pushes Save off the bottom of the screen with no way to shrink back.
+        page.setMaximumWidth(680)
         area = QScrollArea()
         area.setWidgetResizable(True)
+        area.setAlignment(Qt.AlignmentFlag.AlignHCenter | Qt.AlignmentFlag.AlignTop)
         area.setFrameShape(QScrollArea.Shape.NoFrame)
         area.setWidget(page)
         for box in page.findChildren((QComboBox, QAbstractSpinBox)):
@@ -1002,6 +1186,8 @@ class SettingsWindow(QDialog):
     def _general_tab(self):
         page = QWidget()
         form = QFormLayout(page)
+        form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
 
         self.ui_language = QComboBox()
         for label, code in UI_LANGUAGES:
@@ -1092,6 +1278,13 @@ class SettingsWindow(QDialog):
         page = QWidget()
         form = QFormLayout(page)
 
+        self.theme_choice = QComboBox()
+        for name, label in theme.NAMES.items():
+            self.theme_choice.addItem(t(label), name)
+        self.theme_choice.currentIndexChanged.connect(
+            lambda: theme.apply(self, self.theme_choice.currentData()))
+        form.addRow(t("Theme"), self.theme_choice)
+
         self.indicator_screen = QComboBox()
         # The active screen rather than the pointer, for the reason in
         # overlay._compositor_screen: it is what a compositor will answer for,
@@ -1158,6 +1351,8 @@ class SettingsWindow(QDialog):
 
         stt = QGroupBox(t("Speech to text"))
         stt_form = QFormLayout(stt)
+        stt_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.transcribe_provider = QComboBox()
         for label, value in TRANSCRIBE_PROVIDERS:
             self.transcribe_provider.addItem(t(label), value)
@@ -1191,7 +1386,8 @@ class SettingsWindow(QDialog):
 
         self.local_whisper = LocalModelBox(
             ggml.WHISPER, t("On this machine"),
-            ggml.whisper_models, ggml.whisper_model_path)
+            ggml.whisper_models, ggml.whisper_model_path,
+            binary=lambda: self.conf["local_binary"])
         stt_form.addRow(self.local_whisper)
 
         self.local_gpu = QCheckBox(t("Use the graphics card"))
@@ -1205,13 +1401,13 @@ class SettingsWindow(QDialog):
               "spends that once instead of on the first dictation, at the cost of "
               "the memory it sits in."))
         self.local_threads = QSpinBox()
-        self.local_threads.setRange(0, 64)
+        max_threads = max(1, os.cpu_count() or 1)
+        self.local_threads.setRange(0, max_threads)
         self.local_threads.setSpecialValueText(t("Automatic"))
-        # A spin box asks for room for its numbers, and 64 is two characters:
+        # A spin box asks for room for its numbers, and the word standing in for
-        # the word standing in for zero is what actually has to fit, and on
+        # zero is what actually has to fit, and on macOS, where the stepper sits
-        # macOS, where the stepper sits inside the frame, it does not. Widened
+        # inside the frame, it does not. Widened to the word rather than to a
-        # to the word rather than to a number picked by eye, so that it still
+        # number picked by eye, so that it still fits once the word is "Otomatik".
-        # fits once the word is "Otomatik".
         self.local_threads.setMinimumWidth(
             self.local_threads.fontMetrics()
             .horizontalAdvance(t("Automatic")) + 56)
@@ -1222,12 +1418,20 @@ class SettingsWindow(QDialog):
         options_form.addRow("", self.local_preload)
         options_form.addRow(t("Threads"), self.local_threads)
         stt_form.addRow(self.local_options)
+        # What the model is actually doing, as against what the boxes above
+        # ask for. The checkbox can only ask: whether a card was found is
+        # decided by the build and by the machine, and is read back off the
+        # server's own log once it has loaded.
+        self.local_state = WrappedLabel("")
+        stt_form.addRow(self.local_state)
 
         self.transcribe_provider.currentIndexChanged.connect(self._provider_changed)
         outer.addWidget(stt)
 
         orr = QGroupBox(t("Transcript cleanup"))
         orr_form = self.cleanup_form = QFormLayout(orr)
+        orr_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.cleanup_enabled = QCheckBox(t("Clean the transcript with a model"))
         orr_form.addRow("", self.cleanup_enabled)
 
@@ -1300,7 +1504,9 @@ class SettingsWindow(QDialog):
 
         self.local_llm = LocalModelBox(
             ggml.LLAMA, t("On this machine"),
-            ggml.llm_quants, ggml.llm_model_path, repos=ggml.llm_repos)
+            ggml.llm_quants, ggml.llm_model_path,
+            binary=lambda: self.conf["local_llm_binary"],
+            repos=ggml.llm_repos)
         orr_form.addRow(self.local_llm)
 
         self.local_llm_gpu = QCheckBox(t("Use the graphics card"))
@@ -1323,6 +1529,8 @@ class SettingsWindow(QDialog):
         llm_form.addRow("", self.local_llm_preload)
         llm_form.addRow(t("Thinking"), self.local_llm_reasoning)
         orr_form.addRow(self.local_llm_options)
+        self.local_llm_state = WrappedLabel("")
+        orr_form.addRow(self.local_llm_state)
 
         outer.addWidget(orr)
 
@@ -1393,6 +1601,7 @@ class SettingsWindow(QDialog):
                         "glossary, so it can repair the ones that still come out wrong."))
         hint.setWordWrap(True)
         layout.addWidget(hint)
+        layout.addWidget(QLabel(t("Dictionary")))
         self.transcribe_prompt = QPlainTextEdit()
         self.transcribe_prompt.setMaximumHeight(90)
         layout.addWidget(self.transcribe_prompt)
@@ -1461,6 +1670,8 @@ class SettingsWindow(QDialog):
         # be worse than none.
         self.claude_box = QGroupBox(t("Claude Code"))
         claude_form = QFormLayout(self.claude_box)
+        claude_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.assistant_model = QComboBox()
         self.assistant_model.setEditable(True)
         self.assistant_model.addItems(ASSISTANT_MODELS)
@@ -1478,6 +1689,8 @@ class SettingsWindow(QDialog):
 
         self.codex_box = QGroupBox(t("Codex"))
         codex_form = QFormLayout(self.codex_box)
+        codex_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.assistant_codex_model = QComboBox()
         self.assistant_codex_model.setEditable(True)
         self.assistant_codex_model.addItem(t("Codex's own default"), "")
@@ -1493,6 +1706,8 @@ class SettingsWindow(QDialog):
 
         self.openrouter_box = QGroupBox("OpenRouter")
         or_form = QFormLayout(self.openrouter_box)
+        or_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.assistant_openrouter_model = QComboBox()
         self.assistant_openrouter_model.setEditable(True)
         self.assistant_openrouter_model.addItems(ASSISTANT_OR_MODELS)
@@ -1510,6 +1725,8 @@ class SettingsWindow(QDialog):
 
         self.agy_box = QGroupBox(t("Antigravity"))
         agy_form = QFormLayout(self.agy_box)
+        agy_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.assistant_agy_model = QComboBox()
         self.assistant_agy_model.setEditable(True)
         self.assistant_agy_model.addItem(t("Antigravity's own default"), "")
@@ -1528,6 +1745,8 @@ class SettingsWindow(QDialog):
 
         self.opencode_box = QGroupBox("OpenCode Go")
         og_form = QFormLayout(self.opencode_box)
+        og_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.assistant_opencode_model = QComboBox()
         self.assistant_opencode_model.setEditable(True)
         self.assistant_opencode_model.addItems(OPENCODE_MODELS)
@@ -1667,6 +1886,8 @@ class SettingsWindow(QDialog):
 
         models = QGroupBox(t("Minutes"))
         models_form = QFormLayout(models)
+        models_form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self.meeting_model = QComboBox()
         self.meeting_model.setEditable(True)
         self.meeting_model.addItems(MEETING_MODELS)
@@ -1732,7 +1953,7 @@ class SettingsWindow(QDialog):
 
         self.minutes_list = QListWidget()
         self.minutes_list.setWordWrap(True)
-        self.minutes_list.setMaximumHeight(170)
+        self.minutes_list.setMaximumHeight(110)
         self.minutes_list.currentItemChanged.connect(self._show_minutes)
         layout.addWidget(self.minutes_list)
 
@@ -1743,6 +1964,7 @@ class SettingsWindow(QDialog):
         self.minutes_view = QPlainTextEdit()
         self.minutes_view.setReadOnly(True)
         self.minutes_view.setPlaceholderText(t("Pick a meeting to read it."))
+        self.minutes_view.setMinimumHeight(120)
         layout.addWidget(self.minutes_view, 1)
 
         copy = QPushButton(t("Copy"))
@@ -1765,10 +1987,10 @@ class SettingsWindow(QDialog):
         row = QHBoxLayout()
         row.addWidget(copy)
         row.addWidget(self.minutes_retry)
-        row.addStretch(1)
         row.addWidget(folder)
         row.addWidget(delete)
         row.addWidget(reload_)
+        row.addStretch()
         layout.addLayout(row)
         return page
 
@@ -1849,6 +2071,8 @@ class SettingsWindow(QDialog):
         # and two combination boxes starting at different places read as two
         # unrelated settings rather than the pair they are.
         form = QFormLayout()
+        form.setFieldGrowthPolicy(
+            QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow)
         self._shortcut_row(
             form, "toggle", t("Start and stop"),
             t("No global shortcut installed."), placeholder="Ctrl+Space",
@@ -1955,11 +2179,6 @@ class SettingsWindow(QDialog):
             "Once the history passes this many entries, the oldest one is dropped "
             "every time a new one arrives. Set it to 0 to keep everything."
         ))
-        limit_row = QHBoxLayout()
-        limit_row.addWidget(QLabel(t("Keep at most")))
-        limit_row.addWidget(self.history_limit)
-        limit_row.addStretch(1)
-        layout.addLayout(limit_row)
 
         copy = QPushButton(t("Copy selected to clipboard"))
         copy.clicked.connect(self._copy_history)
@@ -1972,10 +2191,13 @@ class SettingsWindow(QDialog):
         row = QHBoxLayout()
         row.addWidget(copy)
         row.addWidget(delete)
-        row.addStretch(1)
+        row.addStretch()
-        row.addWidget(clear)
-        row.addWidget(reload_)
         layout.addLayout(row)
+        more = QHBoxLayout()
+        more.addWidget(clear)
+        more.addWidget(reload_)
+        more.addStretch()
+        layout.addLayout(more)
         return page
 
     @staticmethod
@@ -2069,6 +2291,8 @@ class SettingsWindow(QDialog):
         layout.setContentsMargins(0, 0, 0, 0)
         for index, widget in enumerate(widgets):
             layout.addWidget(widget, 1 if index == 0 else 0)
+        if widgets and all(isinstance(widget, QPushButton) for widget in widgets):
+            layout.addStretch()
         holder = QWidget()
         holder.setLayout(layout)
         return holder
@@ -2078,7 +2302,9 @@ class SettingsWindow(QDialog):
     def _load(self):
         conf = self.conf
         self._select_data(self.ui_language, conf["ui_language"])
-        self._select_data(self.mic, conf["mic_target"])
+        self._select_data(self.theme_choice, conf["theme"])
+        theme.apply(self, self.theme_choice.currentData())
+        self._select_source(self.mic, conf["mic_target"])
         self._select_data(self.language, conf["language"])
         self.auto_paste.setChecked(conf["auto_paste"])
         self.paste_shortcut.setCurrentText(conf["paste_shortcut"])
@@ -2171,8 +2397,8 @@ class SettingsWindow(QDialog):
             conf["assistant_prompt"] or self._loaded_defaults["assistant"]
         )
 
-        self._select_data(self.meeting_mic, conf["meeting_mic_target"])
+        self._select_source(self.meeting_mic, conf["meeting_mic_target"])
-        self._select_data(self.meeting_system, conf["meeting_system_target"])
+        self._select_source(self.meeting_system, conf["meeting_system_target"])
         self.meeting_self_name.setText(conf["meeting_self_name"])
         self.meeting_other_name.setText(conf["meeting_other_name"])
         self.meeting_participants.setPlainText(conf["meeting_participants"])
@@ -2187,9 +2413,11 @@ class SettingsWindow(QDialog):
             conf["meeting_prompt"] or self._loaded_defaults["meeting"]
         )
 
-        self.file_timestamps.setChecked(conf["file_timestamps"])
+        with QSignalBlocker(self.file_timestamps), QSignalBlocker(self.file_cleanup):
-        self.file_cleanup.setChecked(conf["file_cleanup"])
+            self.file_timestamps.setChecked(conf["file_timestamps"])
-        self.file_path = ""
+            self.file_cleanup.setChecked(conf["file_cleanup"])
+        if not hasattr(self, "file_path"):
+            self.file_path = ""
 
         for which, (box, _status, _missing) in self._shortcut_rows.items():
             box.setCurrentText(conf[hotkey.SHORTCUTS[which].setting])
@@ -2202,10 +2430,13 @@ class SettingsWindow(QDialog):
         self._refresh_assistant_status()
         self._load_history()
         self._load_minutes()
+        self._loaded_config = dict(conf.data)
 
     def _save(self):
         conf = self.conf
+        before = dict(conf.data)
         conf["ui_language"] = self.ui_language.currentData() or "auto"
+        conf["theme"] = self.theme_choice.currentData() or theme.DEFAULT
         conf["mic_target"] = self.mic.currentData() or ""
         conf["language"] = self.language.currentData() or "auto"
         conf["auto_paste"] = self.auto_paste.isChecked()
@@ -2352,12 +2583,18 @@ class SettingsWindow(QDialog):
                                   or hotkey.default_combo(which))
         conf["evdev_hotkey"] = self.evdev_enabled.isChecked()
         conf["history_limit"] = self.history_limit.value()
+        # A retained form may predate a CLI reload. Only its edits take priority;
+        # unchanged fields keep the current runtime value.
+        for key, value in before.items():
+            if conf.data.get(key) == self._loaded_config.get(key):
+                conf.data[key] = value
         try:
             conf.save()
         except OSError as exc:
             # An antivirus or a sync tool holding the file for a beat is a
             # message, not an exit: an exception out of a Qt slot takes the
             # whole application down.
+            conf.data = before
             QMessageBox.warning(self, "Dikte",
                                 t("Could not save the settings: {error}",
                                   error=exc))
@@ -2367,7 +2604,9 @@ class SettingsWindow(QDialog):
             cfg.trim_history(conf["history_limit"])
         except OSError as exc:
             print(f"dikte: could not trim the history ({exc})")
-        self._load_history()  # the trim may just have dropped rows from the list
+        self._load()
+        self._saved_form = self._form_values()
+        self._show_dirty()
         self.applied.emit()
         # conf.save() has switched the language t() speaks, so the message box
         # already answers in the new one; the labels around it were translated
@@ -2391,6 +2630,12 @@ class SettingsWindow(QDialog):
                 or self.local_whisper._downloading
                 or self.local_llm._downloading)
 
+    @staticmethod
+    def _select_source(combo, value):
+        if value and combo.findData(value) < 0:
+            combo.addItem(t("{name} (not connected)", name=value), value)
+        SettingsWindow._select_data(combo, value)
+
     @staticmethod
     def _select_data(combo, value):
         index = combo.findData(value)
@@ -2410,6 +2655,7 @@ class SettingsWindow(QDialog):
         self.stt_form.setRowVisible(self.transcribe_status, not local)
         self.stt_form.setRowVisible(self.local_whisper, local)
         self.stt_form.setRowVisible(self.local_options, local)
+        self.stt_form.setRowVisible(self.local_state, local)
         self._refresh_local_box()
         if local:
             return
@@ -2919,6 +3165,7 @@ class SettingsWindow(QDialog):
                                         provider != "local")
         self.cleanup_form.setRowVisible(self.local_llm, provider == "local")
         self.cleanup_form.setRowVisible(self.local_llm_options, provider == "local")
+        self.cleanup_form.setRowVisible(self.local_llm_state, provider == "local")
         self._refresh_local_box()
         binary = cleanup.executable(provider)
         found = shutil.which(binary) if binary else ""

dikte/theme.py

@@ -0,0 +1,136 @@
+"""Built-in desktop palettes shared by windows and recording indicators."""
+
+from pathlib import Path
+
+from PyQt6.QtGui import QColor, QPalette
+
+DEFAULT = "nord"
+NAMES = {"nord": "Nord", "dark": "Classic dark", "light": "Classic light", "dracula": "Dracula"}
+
+# Nord keeps the original blue workspace colors for existing installations.
+PALETTES = {
+    "nord": dict(base="#172434", surface="#213247", border="#3C5168", text="#F1F5FA",
+                 muted="#B2C1D1", accent="#A5C7FA", accent_text="#172434", hover="#2B4059",
+                 disabled_text="#8998A9", disabled_bg="#1C2B3D", capture_disabled="#687B93"),
+    "dark": dict(base="#101010", surface="#202020", border="#474747", text="#F5F5F5",
+                 muted="#BBBBBB", accent="#DDDDDD", accent_text="#101010", hover="#303030",
+                 disabled_text="#888888", disabled_bg="#181818", capture_disabled="#606060"),
+    "light": dict(base="#FFFFFF", surface="#F2F4F7", border="#B6BEC9", text="#1D2633",
+                  muted="#526174", accent="#285DB5", accent_text="#FFFFFF", hover="#E2E7EF",
+                  disabled_text="#687385", disabled_bg="#E9EDF2", capture_disabled="#91A8CD"),
+    "dracula": dict(base="#282A36", surface="#303341", border="#626787", text="#F8F8F2",
+                    muted="#BBC0D9", accent="#BD93F9", accent_text="#282A36", hover="#44475A",
+                    disabled_text="#9096B0", disabled_bg="#282A36", capture_disabled="#706483"),
+}
+
+
+def palette(name=DEFAULT):
+    return PALETTES.get(name, PALETTES[DEFAULT])
+
+
+_STYLE = """
+QWidget { color: @text; }
+QDialog, QWidget#home, QScrollArea, QScrollArea > QWidget > QWidget {
+    background: @base;
+}
+QLabel { background: transparent; }
+QLabel#brand { font-size: 21px; font-weight: 700; }
+QLabel#heading { font-size: 24px; font-weight: 500; }
+QLabel#muted, QLabel#footer { color: @muted; }
+QLabel#footer { padding: 4px 0; }
+QLabel#models { color: @muted; font-size: 12px; }
+QFrame#result { background: @surface; border-radius: 12px; }
+QPushButton {
+    background: @surface; border: 1px solid @border; border-radius: 7px;
+    padding: 6px 10px; min-height: 18px;
+}
+QPushButton:hover { background: @hover; }
+QPushButton:pressed, QPushButton:checked { background: @border; }
+QPushButton:focus, QComboBox:focus, QLineEdit:focus, QPlainTextEdit:focus,
+QListWidget:focus { border: 2px solid @accent; }
+QPushButton:disabled { color: @disabled_text; background: @disabled_bg; }
+QPushButton#primary { background: @accent; color: @accent_text; font-weight: 600; }
+QPushButton#primary:disabled { background: @disabled_bg; color: @disabled_text; border-color: @border; }
+QPushButton#capture {
+    background: @accent; color: @accent_text; border: 6px solid @surface;
+    border-radius: 56px; padding: 0;
+    min-width: 100px; max-width: 100px; min-height: 100px; max-height: 100px;
+}
+QPushButton#capture:focus { border-color: @text; }
+QPushButton#capture:disabled { background: @capture_disabled; }
+QPushButton#mode { border: 1px solid @border; background: @surface; padding: 6px 4px; }
+QPushButton#mode:hover { background: @hover; }
+QPushButton#mode:checked { background: @accent; color: @accent_text; border-color: @accent; }
+QPushButton#mode:focus { border: 2px solid @accent; }
+QPushButton#settings { padding: 0; min-height: 30px; min-width: 32px; }
+QGroupBox { border: 1px solid @border; border-radius: 8px; margin-top: 14px; padding: 8px; }
+QGroupBox::title { subcontrol-origin: margin; left: 12px; padding: 0 5px; }
+QLineEdit, QPlainTextEdit, QListWidget, QComboBox, QSpinBox {
+    background: @surface; border: 1px solid @border; border-radius: 5px;
+    padding: 5px 7px; selection-background-color: @accent; selection-color: @accent_text;
+}
+QComboBox { padding-right: 28px; min-height: 18px; }
+QComboBox::drop-down {
+    subcontrol-origin: border; subcontrol-position: top right;
+    width: 26px; border: none; background: transparent;
+}
+QComboBox::down-arrow { image: url("__ICONS__/chevron-down@arrow_suffix.svg"); width: 12px; height: 12px; }
+QSpinBox { padding: 2px 20px 2px 6px; min-height: 18px; }
+QSpinBox > QLineEdit { border: none; background: transparent; padding: 0; }
+QSpinBox::up-button, QSpinBox::down-button {
+    subcontrol-origin: border; width: 20px; border: none; background: transparent;
+}
+QSpinBox::up-button { subcontrol-position: top right; border-top-right-radius: 5px; }
+QSpinBox::down-button { subcontrol-position: bottom right; border-bottom-right-radius: 5px; }
+QSpinBox::up-button:hover, QSpinBox::down-button:hover { background: @border; }
+QSpinBox::up-arrow { image: url("__ICONS__/chevron-up@arrow_suffix.svg"); width: 10px; height: 10px; }
+QSpinBox::down-arrow { image: url("__ICONS__/chevron-down@arrow_suffix.svg"); width: 10px; height: 10px; }
+QCheckBox { spacing: 7px; }
+QCheckBox::indicator {
+    width: 14px; height: 14px; border: 1px solid @border;
+    border-radius: 4px; background: @surface;
+}
+QCheckBox::indicator:hover { border-color: @accent; }
+QCheckBox::indicator:checked {
+    background: @accent; border-color: @accent;
+    image: url("__ICONS__/check@arrow_suffix.svg");
+}
+QCheckBox:focus::indicator { border-color: @text; }
+QCheckBox:disabled { color: @disabled_text; }
+QCheckBox::indicator:disabled { background: @disabled_bg; border-color: @disabled_text; }
+QCheckBox::indicator:checked:disabled { background: @disabled_text; }
+QComboBox QAbstractItemView { background: @surface; color: @text; selection-background-color: @border; }
+QTabWidget::pane { border: none; }
+QTabBar::tab { background: @surface; padding: 10px; }
+QTabBar::tab:selected { background: @border; }
+QMenu { background: @surface; color: @text; border: 1px solid @border; }
+QMenu::item:selected { background: @border; }
+QToolTip { background: @surface; color: @text; border: 1px solid @accent; }
+QScrollBar:vertical { background: @base; width: 12px; }
+QScrollBar::handle:vertical { background: @border; min-height: 24px; border-radius: 6px; }
+QScrollBar::add-line:vertical, QScrollBar::sub-line:vertical { height: 0; }
+QScrollBar::add-page:vertical, QScrollBar::sub-page:vertical { background: @base; }
+"""
+
+
+def stylesheet(name=DEFAULT):
+    colors = palette(name)
+    result = _STYLE.replace("@arrow_suffix", "-light" if name == "light" else "")
+    for role, color in sorted(colors.items(), key=lambda item: -len(item[0])):
+        result = result.replace("@" + role, color)
+    return result.replace("__ICONS__", (Path(__file__).parent / "icons").as_posix())
+
+
+def apply(widget, name=DEFAULT):
+    colors = palette(name)
+    native = QPalette(widget.palette())
+    for role, color in (
+        (QPalette.ColorRole.Window, "base"), (QPalette.ColorRole.Base, "surface"),
+        (QPalette.ColorRole.AlternateBase, "hover"), (QPalette.ColorRole.Button, "surface"),
+        (QPalette.ColorRole.WindowText, "text"), (QPalette.ColorRole.Text, "text"),
+        (QPalette.ColorRole.ButtonText, "text"), (QPalette.ColorRole.PlaceholderText, "muted"),
+        (QPalette.ColorRole.Highlight, "accent"), (QPalette.ColorRole.HighlightedText, "accent_text"),
+    ):
+        native.setColor(role, QColor(colors[color]))
+    widget.setPalette(native)
+    widget.setStyleSheet(stylesheet(name))

dikte/worker.py

@@ -37,7 +37,7 @@ _paste_lock = threading.Lock()
 
 class Pipeline(QObject):
     stage = pyqtSignal(str)          # human-readable progress line
-    finished = pyqtSignal(str, str, str)  # raw transcript, final text, warning
+    finished = pyqtSignal(str, str, str, str)  # raw, final text, warning, language
     failed = pyqtSignal(str)
     cancelled = pyqtSignal()
 
@@ -120,12 +120,24 @@ class Pipeline(QObject):
         try:
             self.stage.emit(t("Transcribing…"))
             target = conf.transcribe_target()
-            raw = api.transcribe(
+            # The spoken language is only knowable after the fact, and only the
-                target,
+            # local server says what it heard: auto mode asks it there, and
-                wav_path,
+            # every other run (a fixed language, or a hosted provider that
-                language=conf["language"],
+            # detects but stays silent) transcribes as before.
-                prompt=conf["transcribe_prompt"],
+            auto = conf["language"] == "auto"
-            )
+            if auto:
+                raw, detected = api.transcribe_detected(
+                    target, wav_path, language=conf["language"],
+                    prompt=conf["transcribe_prompt"],
+                )
+            else:
+                raw = api.transcribe(
+                    target,
+                    wav_path,
+                    language=conf["language"],
+                    prompt=conf["transcribe_prompt"],
+                )
+                detected = ""
 
             if conf["filter_hallucinations"] and vad.looks_like_hallucination(raw, duration):
                 self._discard(wav_path)
@@ -134,6 +146,10 @@ class Pipeline(QObject):
 
             text = raw
             warning = ""
+            # The language the run actually spoke, reported to the window, the
+            # clipboard path and the history alike: the detected code, or the
+            # configured one when nothing was detected to replace it.
+            speech_language = detected or conf["language"]
             # Remembered rather than re-derived at the history write below: the
             # ask path runs cleanup under a different setting, and the record
             # should say what happened, not what one of the two gates implies.
@@ -145,7 +161,7 @@ class Pipeline(QObject):
                 self.stage.emit(t("Cleaning up…"))
                 cleaned = True
                 try:
-                    text = cleanup.run(raw, conf, conf.cleanup_prompt())
+                    text = cleanup.run(raw, conf, conf.cleanup_prompt(speech=detected))
                 except api.ApiError as exc:
                     # Keep the transcript, but never let the failure pass unseen:
                     # a rejected key would otherwise look like working dictation.
@@ -183,6 +199,7 @@ class Pipeline(QObject):
                 "question": question,
                 "assistant": assistant.provider(conf) if ask else "",
                 "assistant_model": assistant.model(conf) if ask else "",
+                "speech_language": speech_language,
                 "raw": raw,
                 "text": text,
             }
@@ -221,7 +238,7 @@ class Pipeline(QObject):
                     time.sleep(0.35)
                     paste.copy_bytes(previous)
 
-            self.finished.emit(raw, text, warning)
+            self.finished.emit(raw, text, warning, speech_language)
 
         except assistant.Cancelled:
             self.cancelled.emit()

packaging/build-appimage.sh

@@ -55,6 +55,7 @@ cat > "$APPDIR/dikte.desktop" <<EOF
 [Desktop Entry]
 Type=Application
 Name=Dikte
+X-AppImage-Version=$VERSION
 Comment=Voice dictation: record, transcribe, clean up, paste
 Exec=dikte
 Icon=dikte

packaging/dikte.spec

@@ -49,6 +49,7 @@ UNUSED_QT = [
 analysis = Analysis(                            # noqa: F821
     [str(ROOT / "packaging" / "entry.py")],
     pathex=[str(ROOT)],
+    datas=[(str(ROOT / "dikte" / "icons" / "*.svg"), "dikte/icons")],
     hiddenimports=["PyQt6.QtNetwork"],
     # tkinter is the other GUI toolkit CPython ships and would be dead weight;
     # dikte's own tests have no business in a build at all.

tests/render_ui.py

@@ -0,0 +1,169 @@
+"""Capture native Qt client areas with isolated settings and sample data.
+
+Run with python -m tests.render_ui --output DIRECTORY. CI explicitly selects
+xcb, cocoa or windows before importing the test package; the default local
+test backend remains offscreen. No audio, API request or input is generated.
+Window decorations, native file dialogs and compositor effects are excluded.
+"""
+
+import argparse
+import html
+import json
+import platform
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+from unittest import mock
+
+from PyQt6.QtCore import QCoreApplication, QEvent, QT_VERSION_STR, Qt
+from PyQt6.QtGui import QFontInfo
+from PyQt6.QtWidgets import QApplication, QLineEdit, QScrollArea
+
+from tests.test_ui import Settings
+from dikte import config as cfg, i18n, overlay, settings_ui, theme
+from dikte.home_ui import HomeWindow
+
+SAMPLE = "Bir sonraki sürüm için kayıt kontrollerini tamamlayalım. Ayarları gözden geçirip uygulamayı üç platformda da deneyelim."
+CASES = (
+    ("dictation", 620, 560), ("dictation-wide", 1900, 1000),
+    ("file", 620, 640), ("meeting", 680, 760), ("ask", 680, 640),
+    ("settings-general", 720, 760), ("settings-display", 720, 640),
+    ("settings-api", 720, 760), ("settings-assistant", 720, 760),
+    ("settings-shortcuts", 720, 640), ("overlay", 220, 48),
+)
+SETTINGS_PAGES = {"general": 0, "display": 1, "api": 2, "assistant": 4, "shortcuts": 6}
+
+
+def settle():
+    for _ in range(5):
+        QApplication.processEvents()
+
+
+def capture(widget, path, width, height):
+    # Native styles and fonts still render; monitor size does not constrain
+    # wide-window cases, and the runner's other windows cannot cover them.
+    widget.setAttribute(Qt.WidgetAttribute.WA_DontShowOnScreen)
+    widget.resize(width, height)
+    widget.show()
+    settle()
+    pixmap = widget.grab()
+    if pixmap.isNull() or not pixmap.save(str(path), "PNG"):
+        raise RuntimeError(f"Could not capture {path.name}")
+    font = QFontInfo(widget.font())
+    result = {
+        "file": path.name, "requested_size": [width, height],
+        "logical_size": [widget.width(), widget.height()],
+        "pixel_size": [pixmap.width(), pixmap.height()],
+        "device_pixel_ratio": pixmap.devicePixelRatio(),
+        "font": font.family(), "font_points": font.pointSizeF(),
+        "horizontal_overflow": [],
+    }
+    for area in widget.findChildren(QScrollArea):
+        if area.isVisible() and area.horizontalScrollBar().maximum() > 0:
+            result["horizontal_overflow"].append(area.horizontalScrollBar().maximum())
+    widget.hide()
+    print(f"Captured {path.name}: {result['logical_size']}, {font.family()}", flush=True)
+    return result
+
+
+def capture_theme(name, output):
+    harness = Settings("runTest")
+    # Keep the actual platform's application branches as well as Qt's style.
+    harness.platform = sys.platform
+    harness.setUp()
+    try:
+        harness.enterContext(mock.patch.object(settings_ui.SettingsWindow, "_sources_once", return_value=[]))
+        conf = harness.config(
+            theme=name, ui_language="tr", transcribe_provider="openrouter",
+            openrouter_api_key="screenshot-only", openrouter_transcribe_model="openai/whisper-1",
+            cleanup_provider="openrouter", assistant_provider="openrouter",
+        )
+        i18n.set_language("tr")
+        cfg.append_history({"ts": "2026-09-09 14:32:00", "duration": 18,
+                            "elapsed": 2, "text": SAMPLE, "raw": SAMPLE})
+        settings = harness.window(conf)
+        settings.keep_audio.setText("Ses kayıtlarını sakla (örnek kayıt klasörü)")
+        settings.assistant_dir.setPlaceholderText("Proje klasörü")
+        for field in settings.findChildren(QLineEdit):
+            if field.isReadOnly() and "__main__.py toggle" in field.text():
+                field.setText("dikte toggle")
+        settings._saved_form = settings._form_values()
+        settings._show_dirty()
+        controller = SimpleNamespace(
+            conf=conf, state="idle", ask_state="idle", meeting_state="idle",
+            recording=False, paused=False, home_messages={}, meeting_message="",
+            paste_override={}, meeting_elapsed=SimpleNamespace(elapsed=lambda: 12000),
+            _recorded_seconds=lambda: 12, open_settings=settings.show,
+        )
+        for method in ("reset_conversation", "_toggle_pause", "_cancel", "cancel_ask",
+                       "cancel_meeting", "_toggle_meeting", "start", "stop", "start_ask", "stop_ask"):
+            setattr(controller, method, mock.Mock())
+        home = HomeWindow(controller, settings)
+        harness.addCleanup(home.deleteLater)
+        harness.addCleanup(home.close)
+        home._timer.stop()
+        settings.file_label.setText("örnek-kayıt.wav")
+        settings.file_output.setPlainText(SAMPLE)
+        home.ask_output.setPlainText("Örnek yanıt: Önce arayüzü doğrulayalım, ardından sürümü hazırlayalım.")
+        indicator = overlay.Overlay(theme_name=name)
+        harness.addCleanup(indicator.deleteLater)
+        harness.addCleanup(indicator.close)
+        images = []
+        for case, width, height in CASES:
+            if case.startswith("settings-"):
+                settings.tabs.setCurrentIndex(SETTINGS_PAGES[case.removeprefix("settings-")])
+                widget = settings
+            elif case == "overlay":
+                indicator.setAttribute(Qt.WidgetAttribute.WA_DontShowOnScreen)
+                indicator.show_recording()
+                indicator._anim.stop()
+                indicator.set_seconds(12)
+                indicator.levels = [0.15, 0.3, 0.7, 0.4] * (overlay.BARS // 4) + [0.2] * (overlay.BARS % 4)
+                widget = indicator
+                width, height = indicator.width(), indicator.height()
+            else:
+                home.show_mode("dictation" if case == "dictation-wide" else case)
+                widget = home
+            images.append(capture(widget, output / f"{name}-{case}.png", width, height))
+        return images
+    finally:
+        harness.doCleanups()
+        QCoreApplication.sendPostedEvents(None, QEvent.Type.DeferredDelete)
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--expect-platform")
+    args = parser.parse_args()
+    app = QApplication.instance()
+    backend = app.platformName()
+    if args.expect_platform and backend != args.expect_platform:
+        parser.error(f"Expected {args.expect_platform}, got {backend}")
+    output = args.output.resolve()
+    output.mkdir(parents=True, exist_ok=True)
+    manifest = {
+        "system": platform.system(), "python": platform.python_version(),
+        "qt": QT_VERSION_STR, "qpa_backend": backend,
+        "qt_style": app.style().objectName(), "language": "tr",
+        "scope": "Native Qt client-area renders with sample data and isolated settings. No real recording, API call, window decorations, native file dialogs or compositor validation.",
+        "images": [],
+    }
+    for name in theme.NAMES:
+        manifest["images"].extend(capture_theme(name, output))
+        (output / "manifest.json").write_text(json.dumps(manifest, ensure_ascii=False, indent=2), encoding="utf-8")
+    title = f"Dikte: {manifest['system']} / {backend} / Qt {QT_VERSION_STR}"
+    cards = "".join(
+        f'<figure><a href="{entry["file"]}"><img loading="lazy" src="{entry["file"]}"></a><figcaption>{html.escape(entry["file"])}</figcaption></figure>'
+        for entry in manifest["images"]
+    )
+    (output / "index.html").write_text(
+        '<!doctype html><meta charset="utf-8"><title>' + html.escape(title) + '</title>'
+        '<style>body{font:16px system-ui;background:#eee;color:#222;margin:24px}main{display:grid;grid-template-columns:repeat(auto-fit,minmax(300px,1fr));gap:20px}figure{margin:0}img{width:100%;height:340px;object-fit:contain;object-position:top}figcaption{padding:8px}</style>'
+        '<h1>' + html.escape(title) + '</h1><p>' + html.escape(manifest["scope"]) + '</p><main>' + cards + '</main>',
+        encoding="utf-8",
+    )
+
+
+if __name__ == "__main__":
+    main()

tests/support.py

@@ -92,6 +92,22 @@ class DikteTest(unittest.TestCase):
         # down when it last ran.
         self.patch_attr(assistant, "SESSION_FILE", data_dir / "assistant.json")
         self.patch_attr(update, "STATE_FILE", data_dir / "update.json")
+        # ggml resolves its own three from paths.DATA_DIR at import, the same
+        # way cfg does. Left alone, a test asking what is installed or what the
+        # last server ran on would be reading whatever this machine happens to
+        # have downloaded, and passing or failing on somebody's home directory.
+        # program_path prefers a whisper-server or llama-server on the PATH
+        # over the copy Dikte downloaded, so on a machine with whisper.cpp
+        # installed these tests would be answering from that copy instead of
+        # from the install they set up. Every other tool still resolves; the
+        # tests that are about the system build patch this again themselves.
+        _which = shutil.which
+        self.patch_attr(shutil, "which", lambda tool, *args, **rest: (
+            None if tool in ("whisper-server", "llama-server")
+            else _which(tool, *args, **rest)))
+        self.patch_attr(ggml, "DATA_DIR", data_dir)
+        self.patch_attr(ggml, "BIN_DIR", data_dir / "bin")
+        self.patch_attr(ggml, "MODELS_DIR", data_dir / "models")
 
         i18n.set_language("en")
         self.addCleanup(i18n.set_language, "en")

tests/test_api.py

@@ -827,6 +827,40 @@ class TranscribeHere(DikteTest):
             api.transcribe_segments(LOCAL, self.wav)
         self.assertEqual(multipart_fields(calls[0])["model"], "ggml-base.bin")
 
+    # ---- the detected language --------------------------------------------
+
+    def test_auto_mode_asks_whisper_for_the_detected_language(self):
+        # The -nlp the server was started with is switched back on for this one
+        # request, so whisper's verbose_json reports what it heard.
+        reply = {"text": " Merhaba dünya. ", "detected_language": "turkish"}
+        with fake_urlopen(reply) as calls:
+            text, code = api.transcribe_detected(LOCAL, self.wav, language="auto")
+        fields = multipart_fields(calls[0])
+        self.assertEqual(fields["response_format"], "verbose_json")
+        self.assertEqual(fields["no_language_probabilities"], "false")
+        self.assertNotIn("language", fields)
+        self.assertEqual(text, "Merhaba dünya.")
+        self.assertEqual(code, "tr")
+
+    def test_a_fixed_language_reports_no_detection(self):
+        with fake_urlopen({"text": "hello"}) as calls:
+            text, code = api.transcribe_detected(LOCAL, self.wav, language="tr")
+        self.assertNotIn("no_language_probabilities", multipart_fields(calls[0]))
+        self.assertEqual(text, "hello")
+        self.assertEqual(code, "")
+
+    def test_a_detected_language_without_a_code_stays_unknown(self):
+        with fake_urlopen({"text": "hello", "detected_language": "somali"}):
+            _text, code = api.transcribe_detected(LOCAL, self.wav, language="auto")
+        self.assertEqual(code, "")
+
+    def test_a_hosted_auto_run_transcribes_without_detection(self):
+        with fake_urlopen({"text": "hi"}) as calls:
+            text, code = api.transcribe_detected(OPENAI, self.wav, language="auto")
+        self.assertNotIn("no_language_probabilities", multipart_fields(calls[0]))
+        self.assertEqual(text, "hi")
+        self.assertEqual(code, "")
+
 
 class Stopping(unittest.TestCase):
     """The Stop button, from the far end: a request already blocked on a reply.

tests/test_cli.py

@@ -148,8 +148,8 @@ class Parser(unittest.TestCase):
     def parse(self, *argv):
         return cli.build_parser().parse_args(list(argv))
 
-    def test_no_verb_at_all_is_the_settings_window(self):
+    def test_no_verb_uses_the_plain_gui_command(self):
-        # argparse leaves the dest as None; run() is what turns it into "".
+        # argparse leaves the dest as None; run() selects home.
         opts = self.parse()
         self.assertIsNone(opts.verb)
         self.assertEqual(opts.func, cli.cmd_plain)
@@ -578,6 +578,21 @@ class Doctor(DikteTest):
                       self.run_doctor(as_json=False, cleanup_provider="codex",
                                       cleanup_codex_model="gpt-5.4"))
 
+    def test_agent_on_hosted_provider_does_not_ask_for_a_cli_program(self):
+        for provider in ("openrouter", "opencode"):
+            with self.subTest(provider=provider):
+                reply = self.run_doctor(assistant_provider=provider)
+                self.assertEqual(reply["agent"]["provider"], provider)
+                for cli_name in ("claude", "codex", "agy"):
+                    self.assertNotIn(cli_name, reply["programs"])
+
+    def test_agent_on_a_cli_asks_for_the_program(self):
+        for provider, binary in (("claude", "claude"), ("codex", "codex"), ("agy", "agy")):
+            with self.subTest(provider=provider):
+                reply = self.run_doctor(assistant_provider=provider)
+                self.assertEqual(reply["agent"]["provider"], provider)
+                self.assertIn(binary, reply["programs"])
+
 
 class Devices(DikteTest):
     def test_a_machine_with_nothing_names_its_own_missing_program(self):
@@ -685,10 +700,14 @@ class WithoutAnInstance(DikteTest):
         launch.assert_called_once_with("toggle")
 
     def test_every_verb_that_opens_a_window_can_start_it(self):
-        for verb in ("settings", "toggle", "ask", "meeting"):
+        for verb in ("home", "settings", "toggle", "ask", "meeting"):
             with self.subTest(verb=verb):
                 self.assertTrue(self.run_verb([verb])[3].called)
 
+    def test_bare_command_opens_the_daily_workspace(self):
+        _, _, _, launch = self.run_verb([])
+        launch.assert_called_once_with("home")
+
     def test_a_verb_asked_to_wait_starts_nothing(self):
         """There would be no run to wait for; the process would just be replaced."""
         _, _, _, launch = self.run_verb(["toggle", "--wait"])
@@ -746,6 +765,13 @@ class Replies(DikteTest):
         self.assertEqual(code, 0)
         self.assertEqual(out.strip(), "Book it for Thursday.")
 
+    def test_the_json_answer_carries_the_detected_language(self):
+        code, out, _ = self.run_verb(
+            ["--json", "record"],
+            {"ok": True, "text": "Selam", "speech_language": "tr"})
+        self.assertEqual(code, 0)
+        self.assertEqual(json.loads(out)["speech_language"], "tr")
+
     def test_a_dictation_that_failed(self):
         code, out, err = self.run_verb(["stop", "--wait"],
                                        {"ok": False, "error": "No speech detected"})
@@ -780,6 +806,145 @@ class Replies(DikteTest):
         self.assertFalse(launched.called)
 
 
+class LocalModels(DikteTest):
+    """Whether the model on this machine is loaded, and what it is loaded on."""
+
+    def status(self, local, **rest):
+        reply = {"ok": True, "running": True, "dictation": "idle", "ask": "idle",
+                 "meeting": "idle", "listener": True, "local": local, **rest}
+        with mock.patch.object(ipc, "send", return_value=reply), \
+                captured() as (out, _err):
+            cli.cmd_status(Options(json=False))
+        return out.getvalue()
+
+    def entry(self, **values):
+        base = {"running": True, "used": True, "pid": 7, "port": 4321,
+                "model": "ggml-small.bin", "gpu_wanted": True,
+                "backend": "CUDA", "device": "RTX 4070", "layers": "",
+                "available": ["CUDA", "CPU"]}
+        base.update(values)
+        return base
+
+    def test_a_loaded_model_says_what_it_is_loaded_on(self):
+        line = self.status({"whisper": self.entry()})
+        self.assertIn("whisper:", line)
+        self.assertIn("loaded on the graphics card (CUDA, RTX 4070)", line)
+        self.assertIn("ggml-small.bin", line)
+
+    def test_a_card_asked_for_and_not_found_is_said_out_loud(self):
+        line = self.status({"whisper": self.entry(
+            backend="CPU", device="CPU", available=["CPU"])})
+        self.assertIn("loaded on the processor", line)
+        self.assertIn("only the CPU backend was loaded", line)
+
+    def test_a_download_is_not_assumed_to_lack_gpu_support(self):
+        line = self.status({"whisper": self.entry(
+            backend="CPU", device="CPU", available=["CPU"], downloaded=True)})
+        self.assertIn("only the CPU backend was loaded", line)
+        self.assertIn("driver errors", line)
+        self.assertNotIn("has no GPU backend", line)
+
+    def test_a_card_the_build_could_have_used_says_something_else(self):
+        line = self.status({"whisper": self.entry(
+            backend="CPU", device="CPU", available=["CUDA", "CPU"])})
+        self.assertIn("could not be used", line)
+        self.assertNotIn("carries none", line)
+
+    def test_a_card_nobody_asked_for_is_not_a_complaint(self):
+        line = self.status({"whisper": self.entry(
+            backend="CPU", device="CPU", gpu_wanted=False, available=["CPU"])})
+        self.assertIn("loaded on the processor", line)
+        self.assertNotIn("switched on", line)
+
+    def test_a_model_that_is_wanted_and_not_loaded_says_so(self):
+        line = self.status({"whisper": self.entry(running=False)})
+        self.assertIn("whisper:", line)
+        self.assertIn("not loaded", line)
+
+    def test_a_model_neither_used_nor_loaded_is_not_worth_a_line(self):
+        line = self.status({"llama": self.entry(running=False, used=False)})
+        self.assertNotIn("llama", line)
+
+    def test_an_instance_too_old_to_have_been_asked_says_nothing(self):
+        reply = {"ok": True, "running": True, "dictation": "idle", "ask": "idle",
+                 "meeting": "idle", "listener": True}
+        with mock.patch.object(ipc, "send", return_value=reply), \
+                captured() as (out, _err):
+            cli.cmd_status(Options(json=False))
+        self.assertNotIn("whisper", out.getvalue())
+
+    # ---- doctor, which can be asked with nothing running -----------------
+
+    def doctor(self, as_json=False, **settings):
+        self.write_config(settings)
+        with mock.patch.object(ipc, "send", return_value=None), \
+                captured() as (out, _err):
+            cli.cmd_doctor(Options(json=as_json))
+        return json.loads(out.getvalue()) if as_json else out.getvalue()
+
+    def log(self, text):
+        path = ggml.DATA_DIR / "whisper-server.log"
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text(text)
+
+    def test_with_nothing_running_the_last_start_is_read_off_its_log(self):
+        self.log("load_backend: loaded CPU backend from /x.so\n"
+                 "whisper_backend_init_gpu: device 0: CPU (type: 0)\n"
+                 "whisper_backend_init_gpu: no GPU found\n")
+        line = self.doctor(transcribe_provider="local", local_gpu=True)
+        self.assertIn("last run on the processor", line)
+        self.assertNotIn("this build carries none", line)
+        self.assertNotIn("check the server log", line)
+
+    def test_old_cpu_log_does_not_diagnose_a_new_system_binary(self):
+        self.log("load_backend: loaded CPU backend from /old-download.so\n"
+                 "whisper_backend_init_gpu: no GPU found\n")
+        with mock.patch.object(ggml, "program_path",
+                               return_value="/usr/bin/whisper-server"):
+            line = self.doctor(transcribe_provider="local", local_gpu=True)
+            data = self.doctor(as_json=True, transcribe_provider="local",
+                               local_gpu=True)
+        self.assertIn("last run on the processor", line)
+        self.assertNotIn("carries none", line)
+        self.assertNotIn("gpu_wanted", data["local"]["whisper"])
+        self.assertNotIn("downloaded", data["local"]["whisper"])
+
+    def test_enabling_gpu_does_not_reinterpret_a_past_cpu_run(self):
+        self.log("load_backend: loaded Vulkan backend from /gpu.so\n"
+                 "load_backend: loaded CPU backend from /cpu.so\n"
+                 "whisper_init_with_params_no_state: use gpu = 0\n"
+                 "whisper_backend_init_gpu: no GPU found\n")
+        line = self.doctor(transcribe_provider="local", local_gpu=True)
+        self.assertIn("last run on the processor", line)
+        self.assertNotIn("none was found", line)
+        self.assertNotIn("could not be used", line)
+
+    def test_a_run_that_named_no_backend_is_not_read_as_no_run_at_all(self):
+        # A log with nothing recognisable in it still says a server started
+        # here once, which is a different thing from never having started.
+        self.log("whisper_model_load: model size    =  147.37 MB\n")
+        line = self.doctor(transcribe_provider="local")
+        self.assertIn("said nothing about what it was running on", line)
+        self.assertNotIn("never run here", line)
+
+    def test_a_machine_that_never_ran_one_is_not_made_up_a_history_for(self):
+        line = self.doctor(transcribe_provider="local")
+        self.assertIn("never run here", line)
+
+    def test_a_setup_that_transcribes_in_the_cloud_reads_about_none_of_it(self):
+        line = self.doctor(transcribe_provider="openai", cleanup_enabled=False)
+        self.assertNotIn("whisper ", line)
+        self.assertNotIn("never run here", line)
+
+    def test_an_instance_that_cannot_be_asked_is_not_read_as_a_no(self):
+        """It used to print "not loaded", which is a different claim."""
+        self.write_config({"transcribe_provider": "local"})
+        with mock.patch.object(ipc, "send", return_value={"ok": True}), \
+                captured() as (out, _err):
+            cli.cmd_doctor(Options(json=False))
+        self.assertIn("too old to say", out.getvalue())
+
+
 class TranscribeRunsHere(DikteTest):
     """`dikte transcribe` runs in this process, not in the instance."""
 

tests/test_config.py

@@ -274,6 +274,18 @@ class CleanupPrompt(DikteTest):
     def test_no_glossary_means_no_rule_about_one(self):
         self.assertEqual(cfg.Config().cleanup_prompt(), cfg.CLEANUP_PROMPT_EN)
 
+    def test_a_detected_turkish_recording_gets_the_turkish_prompt(self):
+        """Auto mode learns what was heard, and that decides the prompt rather
+        than the interface language."""
+        self.write_config({"ui_language": "en", "transcribe_prompt": "Paraşüt"})
+        conf = cfg.Config()
+        prompt = conf.cleanup_prompt(speech="tr")
+        self.assertEqual(prompt, cfg.CLEANUP_PROMPT_TR
+                         + cfg.GLOSSARY_RULE_TR.format(glossary="Paraşüt"))
+        self.assertIn("KONUŞMACININ KULLANDIĞI İSİM VE TERİMLER", prompt)
+        self.assertIn("NAMES AND TERMS THE SPEAKER USES",
+                      conf.cleanup_prompt(speech="de"))
+
     def test_subtitles_use_their_own_prompt(self):
         conf = cfg.Config()
         self.assertNotEqual(conf.cleanup_prompt(subtitles=True), conf.cleanup_prompt())

tests/test_ggml.py

@@ -735,6 +735,275 @@ class Catalogue(Local):
                          "model.gguf")
 
 
+# --- what it ended up running on ------------------------------------------
+
+
+# Trimmed from real logs. The first is this project's own bug report: the
+# graphics card is switched on, whisper asked for one, and the build had none
+# to give.
+WHISPER_CPU = """\
+load_backend: loaded CPU backend from /opt/whisper/libggml-cpu-haswell.so
+whisper_init_from_file_with_params_no_state: loading model from 'ggml-small.bin'
+whisper_init_with_params_no_state: use gpu    = 1
+whisper_model_load:          CPU total size =   189.49 MB
+whisper_backend_init_gpu: device 0: CPU (type: 0)
+whisper_backend_init_gpu: no GPU found
+"""
+
+WHISPER_CUDA = """\
+load_backend: loaded CUDA backend from /opt/whisper/libggml-cuda.so
+load_backend: loaded CPU backend from /opt/whisper/libggml-cpu-haswell.so
+whisper_init_with_params_no_state: use gpu    = 1
+whisper_model_load:        CUDA0 total size =   189.49 MB
+whisper_backend_init_gpu: device 0: NVIDIA GeForce RTX 4070 (type: 1)
+whisper_backend_init_gpu: using CUDA0 backend
+"""
+
+# A card listed, tried, and refused: whisper says so and carries on without it,
+# and the weights stay where they were put. Reading the listing alone would
+# report a graphics card that is doing nothing.
+WHISPER_GPU_FAILED = """\
+load_backend: loaded Vulkan backend from /usr/lib/ggml/libggml-vulkan.so
+load_backend: loaded CPU backend from /usr/lib/ggml/libggml-cpu-haswell.so
+whisper_model_load:          CPU total size =   189.49 MB
+whisper_backend_init_gpu: device 0: Vulkan0 (type: 1)
+whisper_backend_init_gpu: found GPU device 0: Vulkan0 (type: 1, cnt: 0)
+whisper_backend_init_gpu: using Vulkan0 backend
+whisper_backend_init_gpu: failed to initialize Vulkan0 backend
+"""
+
+# Both backends in one build. The Vulkan listing is there and is not the one
+# that ran, so naming the card out of it would name the wrong device.
+WHISPER_MIXED = """\
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = Intel UHD Graphics 770 (ANV TGL) (anv) | uma: 1
+load_backend: loaded CUDA backend from /opt/whisper/libggml-cuda.so
+load_backend: loaded Vulkan backend from /opt/whisper/libggml-vulkan.so
+load_backend: loaded CPU backend from /opt/whisper/libggml-cpu-haswell.so
+  Device 0: NVIDIA GeForce RTX 4070, compute capability 8.9, VMM: yes
+whisper_model_load:        CUDA0 total size =   189.49 MB
+whisper_backend_init_gpu: device 0: CUDA0 (type: 1)
+whisper_backend_init_gpu: using CUDA0 backend
+"""
+
+# The same start on a card whisper names only by its slot. The card's own name
+# is one line further up, printed by the backend as it enumerates.
+WHISPER_VULKAN = """\
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon RX 6600 (RADV NAVI23) (radv) | uma: 0 | fp16: dot2
+load_backend: loaded Vulkan backend from /usr/lib/ggml/libggml-vulkan.so
+load_backend: loaded CPU backend from /usr/lib/ggml/libggml-cpu-haswell.so
+whisper_model_load:      Vulkan0 total size =   189.49 MB
+whisper_backend_init_gpu: device 0: Vulkan0 (type: 1)
+whisper_backend_init_gpu: using Vulkan0 backend
+"""
+
+# A whisper built by hand on a Mac: Metal is compiled in rather than loaded, so
+# there is no line to read and no honest answer but "it did not say".
+WHISPER_QUIET = """\
+whisper_init_from_file_with_params_no_state: loading model from 'ggml-base.bin'
+whisper_model_load: model size    =  147.37 MB
+"""
+
+LLAMA_GPU = """\
+load_backend: loaded Vulkan backend from /opt/llama/libggml-vulkan.so
+load_backend: loaded CPU backend from /opt/llama/libggml-cpu.so
+load_tensors: offloading 28 repeating layers to GPU
+load_tensors: offloaded 29/29 layers to GPU
+load_tensors:      Vulkan0 model buffer size = 2048.00 MiB
+"""
+
+LLAMA_CPU = """\
+load_backend: loaded Vulkan backend from /opt/llama/libggml-vulkan.so
+load_backend: loaded CPU backend from /opt/llama/libggml-cpu.so
+load_tensors: offloaded 0/29 layers to GPU
+"""
+
+
+# A downloaded processor-only build pointed at the system's Vulkan backend
+# through GGML_BACKEND_PATH. whisper numbers every device it can see in one
+# sequence, so the card is its device 1 while still being Vulkan0.
+WHISPER_LENT_BACKEND = """\
+load_backend: loaded CPU backend from /data/bin/whisper/libggml-cpu-haswell.so
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon RX 6600 (RADV NAVI23) (radv) | uma: 0
+load_backend: loaded Vulkan backend from /usr/lib/ggml/libggml-vulkan.so
+whisper_model_load:      Vulkan0 total size =   189.49 MB
+whisper_backend_init_gpu: device 0: CPU (type: 0)
+whisper_backend_init_gpu: device 1: Vulkan0 (type: 1)
+whisper_backend_init_gpu: found GPU device 1: Vulkan0 (type: 1, cnt: 0)
+whisper_backend_init_gpu: using Vulkan0 backend
+"""
+
+# Two cards, and the one that ran is not the one in the slot the handle names.
+# Reading whisper's listing by the handle's digit would name the other card.
+WHISPER_TWO_CARDS = """\
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon RX 6600 (RADV NAVI23) (radv) | uma: 0
+load_backend: loaded CUDA backend from /opt/whisper/libggml-cuda.so
+load_backend: loaded Vulkan backend from /opt/whisper/libggml-vulkan.so
+load_backend: loaded CPU backend from /opt/whisper/libggml-cpu-haswell.so
+whisper_model_load:      Vulkan0 total size =   189.49 MB
+whisper_backend_init_gpu: device 0: NVIDIA GeForce RTX 4070 (type: 1)
+whisper_backend_init_gpu: device 1: Vulkan0 (type: 1)
+whisper_backend_init_gpu: using Vulkan0 backend
+"""
+
+
+class WhatItRunsOn(Local):
+    """Reading the backend back out of the log the server wrote."""
+
+    def log(self, text):
+        path = self.path("server.log")
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text(text)
+        return path
+
+    def read(self, program, text):
+        return ggml._read_accel(program, self.log(text))
+
+    def test_a_card_that_was_asked_for_and_not_found_is_the_processor(self):
+        accel = self.read(ggml.WHISPER, WHISPER_CPU)
+        self.assertEqual(accel.backend, "CPU")
+        self.assertEqual(ggml.accel_kind(accel), "cpu")
+
+    def test_only_loaded_backends_are_reported(self):
+        self.assertTrue(ggml.cpu_only_loaded(self.read(ggml.WHISPER, WHISPER_CPU)))
+        self.assertFalse(ggml.cpu_only_loaded(self.read(ggml.WHISPER, WHISPER_CUDA)))
+
+    def test_a_card_that_was_found_is_named(self):
+        accel = self.read(ggml.WHISPER, WHISPER_CUDA)
+        self.assertEqual(accel.backend, "CUDA")
+        self.assertEqual(accel.device, "NVIDIA GeForce RTX 4070")
+        self.assertEqual(ggml.accel_kind(accel), "gpu")
+        self.assertEqual(ggml.accel_detail(accel),
+                         "CUDA, NVIDIA GeForce RTX 4070")
+
+    def test_a_card_named_only_by_its_slot_is_looked_up(self):
+        accel = self.read(ggml.WHISPER, WHISPER_VULKAN)
+        self.assertEqual(accel.backend, "Vulkan")
+        # "Vulkan0" says which slot; the point of the line is which card.
+        self.assertEqual(accel.device, "AMD Radeon RX 6600 (RADV NAVI23)")
+
+    def test_the_driver_behind_the_card_is_not_part_of_its_name(self):
+        # "(radv)" is how it is reached; "(RADV NAVI23)" is what it is called.
+        self.assertNotIn("(radv)",
+                         self.read(ggml.WHISPER, WHISPER_VULKAN).device)
+
+    def test_a_card_numbered_one_way_and_handled_another_is_still_named(self):
+        accel = self.read(ggml.WHISPER, WHISPER_LENT_BACKEND)
+        self.assertEqual(accel.backend, "Vulkan")
+        self.assertEqual(accel.device, "AMD Radeon RX 6600 (RADV NAVI23)")
+
+    def test_the_card_named_is_the_one_the_handle_belongs_to(self):
+        # whisper's device 0 is the other card. The handle is Vulkan0, and
+        # Vulkan's own device 0 is the AMD one.
+        accel = self.read(ggml.WHISPER, WHISPER_TWO_CARDS)
+        self.assertEqual(accel.device, "AMD Radeon RX 6600 (RADV NAVI23)")
+        self.assertNotIn("NVIDIA", ggml.accel_detail(accel))
+
+    def test_a_card_that_failed_to_start_is_not_a_card_in_use(self):
+        # It was listed, it was tried, it did not work, and whisper went on
+        # without it. The listing alone would have called this a graphics card.
+        accel = self.read(ggml.WHISPER, WHISPER_GPU_FAILED)
+        self.assertEqual(accel.backend, "CPU")
+        self.assertEqual(ggml.accel_kind(accel), "cpu")
+
+    def test_failed_initialisation_overrides_weights_on_the_card(self):
+        log = WHISPER_GPU_FAILED.replace("CPU total size", "Vulkan0 total size")
+        self.assertEqual(self.read(ggml.WHISPER, log).backend, "CPU")
+
+    def test_a_later_successful_attempt_replaces_the_failed_one(self):
+        log = WHISPER_GPU_FAILED + (
+            "whisper_backend_init_gpu: using Vulkan0 backend\n")
+        self.assertEqual(self.read(ggml.WHISPER, log).backend, "Vulkan")
+
+    def test_the_card_named_is_the_one_that_ran(self):
+        accel = self.read(ggml.WHISPER, WHISPER_MIXED)
+        self.assertEqual(accel.backend, "CUDA")
+        self.assertEqual(accel.device, "NVIDIA GeForce RTX 4070")
+        self.assertNotIn("Intel", ggml.accel_detail(accel))
+
+    def test_a_slot_number_is_not_a_name(self):
+        # "Vulkan0" says which slot; with no listing to look it up in, saying
+        # nothing beats saying that.
+        self.assertEqual(self.read(ggml.LLAMA, LLAMA_GPU).device, "")
+
+    def test_a_log_that_says_nothing_is_not_guessed_at(self):
+        accel = self.read(ggml.WHISPER, WHISPER_QUIET)
+        self.assertEqual(accel.backend, "")
+        self.assertEqual(ggml.accel_kind(accel), "unknown")
+
+    def test_a_log_that_is_not_there_is_not_guessed_at_either(self):
+        self.assertEqual(ggml._read_accel(ggml.WHISPER, self.path("gone.log")),
+                         ggml.NO_ACCEL)
+
+    def test_the_layers_llama_offloaded_are_read_back(self):
+        accel = self.read(ggml.LLAMA, LLAMA_GPU)
+        self.assertEqual(accel.backend, "Vulkan")
+        self.assertEqual(accel.layers, "29/29")
+        self.assertEqual(ggml.accel_detail(accel), "Vulkan, 29/29 layers")
+
+    def test_llama_names_the_allocated_device_not_the_first_loaded_backend(self):
+        log = (
+            "load_backend: loaded CUDA backend from /x.so\n"
+            "  Device 0: NVIDIA RTX 4070, compute capability 8.9, VMM: yes\n"
+            "ggml_vulkan: 0 = Intel UHD Graphics | uma: 1\n"
+            "ggml_vulkan: 1 = AMD Radeon RX 6600 | uma: 0\n"
+            "load_tensors:     CUDA_Host model buffer size = 32.00 MiB\n"
+            "load_tensors:       Vulkan0 model buffer size = 0.00 MiB\n"
+            + LLAMA_GPU.replace("Vulkan0 model", "Vulkan1 model"))
+        accel = self.read(ggml.LLAMA, log)
+        self.assertEqual(accel.backend, "Vulkan")
+        self.assertEqual(accel.device, "AMD Radeon RX 6600")
+
+    def test_llama_without_buffer_evidence_does_not_guess_the_backend(self):
+        log = LLAMA_GPU.replace(
+            "load_tensors:      Vulkan0 model buffer size = 2048.00 MiB\n", "")
+        accel = self.read(ggml.LLAMA, log)
+        self.assertEqual((accel.backend, accel.device), ("GPU", ""))
+        self.assertEqual(accel.layers, "29/29")
+
+    def test_llama_split_across_cards_does_not_name_only_one(self):
+        log = LLAMA_GPU + (
+            "load_tensors:      Vulkan1 model buffer size = 1024.00 MiB\n")
+        accel = self.read(ggml.LLAMA, log)
+        self.assertEqual((accel.backend, accel.device), ("GPU", ""))
+
+    def test_llama_static_metal_build_can_be_identified_by_its_buffer(self):
+        log = (
+            "ggml_metal_init: picking default device: Apple M2\n"
+            "load_tensors: offloaded 29/29 layers to GPU\n"
+            "load_tensors:       Metal model buffer size = 2048.00 MiB\n")
+        accel = self.read(ggml.LLAMA, log)
+        self.assertEqual((accel.backend, accel.device), ("Metal", "Apple M2"))
+
+    def test_a_llama_that_offloaded_nothing_is_on_the_processor(self):
+        accel = self.read(ggml.LLAMA, LLAMA_CPU)
+        self.assertEqual(accel.backend, "CPU")
+        self.assertEqual(ggml.accel_kind(accel), "cpu")
+        # The build could have used the card; this run did not.
+        self.assertFalse(ggml.cpu_only_loaded(accel))
+
+    def test_the_processor_is_not_named_twice(self):
+        # whisper prints CPU as the backend and as the device, and saying it
+        # twice reads like two different things.
+        self.assertEqual(ggml.accel_detail(self.read(ggml.WHISPER, WHISPER_CPU)),
+                         "CPU")
+
+    def test_which_copy_is_running_decides_what_advice_is_worth_giving(self):
+        mine = ggml.BIN_DIR / "whisper" / "b1" / "whisper-server"
+        mine.parent.mkdir(parents=True, exist_ok=True)
+        mine.write_text("#!/bin/sh\n")
+        self.assertTrue(ggml.is_downloaded(str(mine)))
+        self.assertFalse(ggml.is_downloaded("/usr/bin/whisper-server"))
+        self.assertFalse(ggml.is_downloaded(""))
+
+    def test_nothing_is_running_is_not_a_backend(self):
+        self.assertEqual(ggml.accel_kind({"running": False, "backend": "CUDA"}),
+                         "off")
+
+
 # --- keeping a server alive -----------------------------------------------
 
 
@@ -754,6 +1023,18 @@ STAND_IN = textwrap.dedent("""
         print("could not load model: no such file")
         sys.exit(2)
 
+    # The startup chatter a real server prints before it binds, so that the
+    # log has something for _read_accel to find. Flushed, because stdout here
+    # is a file and nothing would reach it before the port opened.
+    if "--backend" in args:
+        print("load_backend: loaded " + opt("--backend") + " backend from /x.so",
+              flush=True)
+        print("whisper_backend_init_gpu: device 0: Test Card (type: 1)",
+              flush=True)
+        # The attempt is followed by no failure in this stand-in.
+        print("whisper_backend_init_gpu: using " + opt("--backend") + "0 backend",
+              flush=True)
+
     started = time.monotonic()
     healthy_after = float(opt("--healthy-after", "0"))
 
@@ -816,6 +1097,50 @@ class Servers(ServerCase):
         self.assertRegex(url, r"^http://127\.0\.0\.1:\d+/v1$")
         self.assertTrue(server.running)
 
+    def test_nothing_started_is_a_state_saying_so(self):
+        state = self.server().state()
+        self.assertFalse(state["running"])
+        self.assertEqual(ggml.accel_kind(state), "off")
+
+    def test_a_running_server_says_what_it_settled_on(self):
+        server = self.server(extra=["--backend", "CUDA"], gpu=True)
+        server.serve()
+        state = server.state()
+        self.assertTrue(state["running"])
+        self.assertIn(f":{state['port']}/v1", server.base_url())
+        self.assertEqual(state["backend"], "CUDA")
+        self.assertEqual(state["device"], "Test Card")
+        self.assertTrue(state["gpu_wanted"])
+        self.assertEqual(ggml.accel_kind(state), "gpu")
+
+    def test_a_setting_changed_mid_start_does_not_rename_what_is_running(self):
+        # A save that lands while the model is being read in finds no process
+        # to stop, so it changes the settings under a start already in flight.
+        # The line must name the model that is loaded, not the one that will be.
+        server = self.server(model="first")
+        launch = server._launch
+
+        def during(settings):
+            result = launch(settings)
+            server.configure(model="second")
+            return result
+
+        self.patch_attr(server, "_launch", during)
+        server.serve()
+        self.assertEqual(server.state()["model"], "first")
+        self.assertEqual(server.settings()["model"], "second")
+
+    def test_stopping_takes_the_backend_with_it(self):
+        server = self.server(extra=["--backend", "CUDA"])
+        server.serve()
+        server.stop()
+        self.assertEqual(server.state()["backend"], "")
+
+    def test_a_server_that_announced_nothing_is_not_guessed_at(self):
+        server = self.server()
+        server.serve()
+        self.assertEqual(ggml.accel_kind(server.state()), "unknown")
+
     def test_the_second_call_does_not_start_a_second_one(self):
         server = self.server()
         first = server.serve()

tests/test_home_ui.py

@@ -0,0 +1,404 @@
+"""Native workspace navigation, capture boundaries and persisted results."""
+
+from types import SimpleNamespace
+from unittest import mock
+
+from PyQt6.QtCore import Qt
+from PyQt6.QtWidgets import QApplication, QMessageBox, QPushButton
+
+from dikte import config as cfg, home_ui, i18n
+from dikte.app import Dikte
+from tests import test_ui
+from tests.support import DikteTest
+
+
+class Home(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.fixture = test_ui.Settings("runTest")
+        self.fixture.setUp()
+        self.addCleanup(self.fixture.doCleanups)
+        self.conf = self.fixture.config(transcribe_provider="openai", openai_api_key="test",
+                                        cleanup_enabled=False)
+        self.settings = self.fixture.window(self.conf)
+        self.controller = SimpleNamespace(
+            conf=self.conf, state="idle", ask_state="idle", meeting_state="idle",
+            meeting_message="", home_messages={}, recording=False, paused=False,
+            paste_override={}, open_settings=mock.Mock(), reset_conversation=mock.Mock(),
+            _recorded_seconds=lambda: 65, meeting_elapsed=SimpleNamespace(elapsed=lambda: 90000),
+        )
+        for name in ("start", "stop", "start_ask", "stop_ask", "_toggle_pause", "_cancel",
+                     "cancel_ask", "_toggle_meeting", "cancel_meeting"):
+            setattr(self.controller, name, mock.Mock())
+        self.window = home_ui.HomeWindow(self.controller, self.settings)
+        self.addCleanup(self.window.deleteLater)
+        self.addCleanup(self.window.close)
+        self.window.show()
+        QApplication.processEvents()
+
+    def test_daily_tasks_are_reachable_outside_configuration(self):
+        for mode in ("dictation", "file", "meeting", "ask", "history"):
+            self.window.show_mode(mode)
+            QApplication.processEvents()
+            self.assertEqual(self.window.pages.currentWidget(), self.window.mode_pages[mode])
+        self.window.show_mode("meeting")
+        self.assertTrue(self.settings.minutes_view.isVisible())
+        self.window.show_mode("history")
+        self.assertTrue(self.settings.history.isVisible())
+        self.assertEqual(self.settings.tabs.count(), 7)
+        self.assertFalse(self.settings.tabs.tabBar().isVisible())
+
+    def test_native_chrome_and_capture_geometry(self):
+        self.assertFalse(self.window.windowFlags() & Qt.WindowType.FramelessWindowHint)
+        self.assertEqual(self.window.capture_button.width(), self.window.capture_button.height())
+        self.assertGreaterEqual(self.window.capture_button.height(), 100)
+
+    def test_theme_preview_discard_and_save_keep_runtime_separate(self):
+        self.assertEqual(self.conf["theme"], "nord")
+        self.settings.theme_choice.setCurrentIndex(self.settings.theme_choice.findData("light"))
+        self.assertEqual(self.conf["theme"], "nord")
+        self.assertIn("#FFFFFF", self.settings.styleSheet())
+        self.assertNotEqual(self.window.styleSheet(), self.settings.styleSheet())
+        self.settings._discard_changes()
+        self.assertEqual(self.settings.theme_choice.currentData(), "nord")
+        self.settings.theme_choice.setCurrentIndex(self.settings.theme_choice.findData("dracula"))
+        self.settings._save()
+        self.assertEqual(cfg.Config()["theme"], "dracula")
+        self.assertEqual(self.window._theme_name, "dracula")
+        self.assertIn("#282A36", self.window.styleSheet())
+
+    def test_theme_colors_are_per_overlay_and_unknown_name_falls_back(self):
+        from dikte import overlay, theme
+        dark = overlay.Overlay(theme_name="dark")
+        light = overlay.Overlay(theme_name="light")
+        self.addCleanup(dark.deleteLater)
+        self.addCleanup(light.deleteLater)
+        self.assertEqual(light.colors["base"].name(), "#ffffff")
+        self.assertEqual(dark.colors["base"].name(), "#101010")
+        dark.set_theme("dracula")
+        self.assertEqual(light.colors["base"].name(), "#ffffff")
+        self.assertEqual(theme.stylesheet("unknown"), theme.stylesheet("nord"))
+
+    def test_wide_meeting_page_keeps_actions_compact(self):
+        self.window.resize(900, 700)
+        self.window.show_mode("meeting")
+        QApplication.processEvents()
+        self.assertLess(self.window.meeting_button.width(), 300)
+        self.assertLessEqual(self.settings.minutes_view.width(), 680)
+
+    def test_wide_windows_center_every_task_and_settings_page(self):
+        self.window.resize(1900, 1000)
+        for mode, area in self.window.mode_pages.items():
+            self.window.show_mode(mode)
+            QApplication.processEvents()
+            with self.subTest(mode=mode):
+                self.assertLessEqual(area.widget().width(), 680)
+                self.assertAlmostEqual(area.widget().geometry().center().x(),
+                                       area.viewport().rect().center().x(), delta=1)
+        self.settings.resize(1900, 1000)
+        self.settings.show()
+        for index in range(self.settings.tabs.count()):
+            self.settings.tabs.setCurrentIndex(index)
+            QApplication.processEvents()
+            area = self.settings.tabs.widget(index)
+            with self.subTest(settings=index):
+                self.assertAlmostEqual(area.widget().geometry().center().x(),
+                                       area.viewport().rect().center().x(), delta=1)
+
+    def test_meeting_and_assistant_actions_share_one_row(self):
+        self.window.resize(620, 760)
+        for mode, labels in (
+            ("meeting", ("Copy", "Write it up", "Open the folder", "Delete selected", "Reload")),
+            ("ask", (self.window.ask_button.text(), "Start a new conversation")),
+        ):
+            self.window.show_mode(mode)
+            QApplication.processEvents()
+            page = self.window.mode_pages[mode]
+            buttons = {b.text(): b for b in page.findChildren(QPushButton)}
+            positions = [buttons[label].mapTo(page, buttons[label].rect().center()).y()
+                         for label in labels]
+            with self.subTest(mode=mode):
+                self.assertLessEqual(max(positions) - min(positions), 1)
+
+    def test_empty_state_does_not_invent_a_transcript(self):
+        self.assertEqual(self.window.latest_text.toPlainText(), "")
+        self.assertFalse(self.window.copy_button.isEnabled())
+        self.assertFalse(self.window.open_button.isEnabled())
+
+    def test_clearing_history_removes_the_latest_preview(self):
+        cfg.append_history({"text": "Remove this preview"})
+        self.window.refresh_results()
+        self.window.show_mode("history")
+        with mock.patch.object(self.settings, "_confirm", return_value=True):
+            self.settings._clear_history()
+        self.window.show_mode("dictation")
+        self.assertEqual(self.window.latest_text.toPlainText(), "")
+        self.assertFalse(self.window.copy_button.isEnabled())
+
+    def test_missing_local_model_opens_setup_without_recording(self):
+        self.conf["transcribe_provider"] = "local"
+        self.window.refresh()
+        self.assertEqual(self.window.capture_status.text(), "Set up transcription")
+        self.window._capture()
+        self.controller.start.assert_not_called()
+        self.controller.open_settings.assert_called_once()
+        self.assertEqual(self.settings.tabs.currentIndex(), self.settings.api_tab_index)
+        self.assertEqual(self.controller.paste_override, {})
+
+    def test_button_capture_never_automatically_pastes(self):
+        def start():
+            self.assertIs(self.controller.paste_override["dictation"], False)
+            self.controller.state = "recording"
+            self.controller.recording = True
+        self.controller.start.side_effect = start
+        self.window._capture()
+        self.controller.start.assert_called_once()
+        self.assertFalse(self.controller.paste_override["dictation"])
+        self.assertIn("01:05", self.window.capture_status.text())
+        self.assertTrue(self.window.pause_button.isVisible())
+        self.window._capture()
+        self.controller.stop.assert_called_once()
+
+    def test_failed_capture_does_not_leak_a_paste_override(self):
+        self.window._capture()
+        self.assertEqual(self.controller.paste_override, {})
+
+    def test_busy_capture_can_queue_but_does_not_steal_assistant_microphone(self):
+        self.controller.state = "busy"
+        self.window.refresh()
+        self.assertTrue(self.window.capture_button.isEnabled())
+        self.controller.ask_state = "recording"
+        self.controller.recording = True
+        self.window.refresh()
+        self.assertFalse(self.window.capture_button.isEnabled())
+        self.window._capture()
+        self.controller.start.assert_not_called()
+
+    def test_pause_cancel_and_failures_are_visible(self):
+        self.controller.state = "recording"
+        self.controller.recording = True
+        self.controller.paused = True
+        self.window.refresh()
+        self.assertIn("Paused", self.window.capture_status.text())
+        self.window._pause()
+        self.controller._toggle_pause.assert_called_once()
+        self.window._cancel_capture()
+        self.controller._cancel.assert_called_once()
+        self.controller.home_messages["dictation"] = "Microphone permission denied"
+        self.window.refresh()
+        self.assertIn("permission denied", self.window.capture_error.text())
+
+    def test_real_latest_dictation_is_separate_from_assistant_answer(self):
+        cfg.append_history({"ts": "2026-09-09 12:00:00", "text": "Actual transcript"})
+        cfg.append_history({"mode": "ask", "text": "Actual answer"})
+        self.window.refresh_results()
+        self.assertEqual(self.window.latest_text.toPlainText(), "Actual transcript")
+        self.assertEqual(self.window.ask_output.toPlainText(), "Actual answer")
+        cursor = self.window.latest_text.textCursor()
+        cursor.setPosition(3)
+        self.window.latest_text.setTextCursor(cursor)
+        self.window.refresh()
+        self.window.refresh_results()
+        self.assertEqual(self.window.latest_text.textCursor().position(), 3)
+        cfg.clear_history()
+        self.window.refresh_results()
+        self.assertEqual(self.window.latest_text.toPlainText(), "")
+
+    def test_processing_summary_uses_full_models_and_actual_acceleration(self):
+        self.conf["transcribe_provider"] = "local"
+        self.conf["local_model"] = "ggml-large-v3-turbo-q5_0.bin"
+        self.conf["cleanup_enabled"] = True
+        self.conf["cleanup_provider"] = "local"
+        self.conf["local_llm_model"] = "gemma-3-4b-it-Q4_K_M.gguf"
+        self.conf["local_gpu"] = True
+        with mock.patch.object(home_ui.ggml, "state", return_value={
+            "whisper": {"running": True, "backend": "CPU"},
+            "llama": {"running": True, "backend": "Vulkan"},
+        }):
+            text = home_ui.processing_locations(self.conf)
+        self.assertIn("ggml-large-v3-turbo-q5_0.bin (Local CPU)", text)
+        self.assertIn("gemma-3-4b-it-Q4_K_M.gguf (Local GPU)", text)
+        self.assertNotIn("API", text)
+        with mock.patch.object(home_ui.ggml, "state", return_value={}):
+            text = home_ui.processing_locations(self.conf)
+        self.assertNotIn("GPU", text)
+        self.assertIn("(Local)", text)
+        self.assertIn(self.conf["meeting_model"], home_ui.processing_locations(self.conf, "meeting"))
+        self.conf["assistant_cleanup"] = True
+        self.conf["cleanup_provider"] = "gemini"
+        self.conf["assistant_provider"] = "codex"
+        text = home_ui.processing_locations(self.conf, "ask")
+        self.assertIn(self.conf["cleanup_gemini_model"] + " (API)", text)
+        self.assertIn("Codex default model (CLI)", text)
+
+    def test_timestamped_file_summary_uses_the_timestamp_model(self):
+        self.conf["transcribe_provider"] = "openrouter"
+        self.conf["openrouter_transcribe_model"] = "google/gemini-audio"
+        self.conf["openrouter_file_model"] = "openai/whisper-1"
+        text = home_ui.processing_locations(self.conf, "file", file_timestamps=True)
+        self.assertIn("openai/whisper-1 (API)", text)
+        self.assertNotIn("google/gemini-audio", text)
+
+    def test_meeting_summary_uses_segment_model_even_without_file_timestamps(self):
+        self.conf["transcribe_provider"] = "openai"
+        self.conf["transcribe_model"] = "gpt-4o-transcribe"
+        text = home_ui.processing_locations(self.conf, "meeting", file_timestamps=False)
+        self.assertIn("whisper-1 (API)", text)
+        self.assertNotIn("gpt-4o-transcribe", text)
+        self.conf["transcribe_provider"] = "openrouter"
+        self.conf["openrouter_file_model"] = "mistralai/voxtral-small-24b-2507"
+        text = home_ui.processing_locations(self.conf, "meeting", file_timestamps=False)
+        self.assertIn("mistralai/voxtral-small-24b-2507 (API)", text)
+
+    def test_unsupported_meeting_is_disabled(self):
+        self.enterContext(mock.patch.object(home_ui.audio, "sound", return_value=SimpleNamespace(meetings=False)))
+        self.window.show_mode("meeting")
+        self.assertFalse(self.window.meeting_button.isEnabled())
+        self.assertIn("not supported", self.window.meeting_hint.text())
+        self.window._meeting()
+        self.controller._toggle_meeting.assert_not_called()
+
+    def test_assistant_scope_uses_actual_shortcut_and_permissions(self):
+        self.conf["assistant_provider"] = "codex"
+        self.conf["assistant_shortcut"] = "Ctrl+Alt+A"
+        self.conf["assistant_dir"] = self.root
+        self.conf["assistant_codex_sandbox"] = "danger-full-access"
+        self.window.show_mode("ask")
+        self.assertIn("Ctrl+Alt+A", self.window.ask_scope.text())
+        self.assertIn(self.root, self.window.ask_scope.text())
+        self.assertIn("No sandbox at all", self.window.ask_scope.text())
+        self.assertEqual(self.window.ask_button.text(), "Set up assistant")
+        self.window._ask()
+        self.controller.start_ask.assert_not_called()
+        self.controller.open_settings.assert_called_once()
+
+    def test_failed_settings_save_keeps_runtime_config_and_form_edits(self):
+        before = dict(self.conf.data)
+        self.settings.auto_paste.setChecked(not self.conf["auto_paste"])
+        self.assertEqual(self.settings.dirty_label.text(), "Unsaved changes")
+        with mock.patch.object(self.conf, "save", side_effect=OSError("disk full")), mock.patch.object(QMessageBox, "warning"):
+            self.settings._save()
+        self.assertEqual(self.conf.data, before)
+        self.assertNotEqual(self.settings.auto_paste.isChecked(), self.conf["auto_paste"])
+        self.settings.file_path = "/tmp/chosen.wav"
+        self.settings._discard_changes()
+        self.assertEqual(self.settings.dirty_label.text(), "")
+        self.assertEqual(self.settings.file_path, "/tmp/chosen.wav")
+
+    def test_small_window_keeps_navigation_and_footer_accessible(self):
+        self.window.resize(460, 460)
+        QApplication.processEvents()
+        for mode in ("dictation", "file", "meeting", "ask", "history"):
+            self.window.show_mode(mode)
+            QApplication.processEvents()
+            if self.window.footer.isVisible():
+                self.assertTrue(self.window.rect().contains(self.window.footer.geometry()))
+            self.assertTrue(self.window.rect().contains(self.window.mode_buttons["dictation"].geometry().topLeft()))
+
+    def test_apply_merges_unrelated_cli_changes_and_keeps_user_edits(self):
+        self.settings.auto_paste.setChecked(False)
+        self.conf["shortcut"] = "Ctrl+Shift+F9"
+        self.conf["groq_transcribe_model"] = "external-model"
+        self.settings.refresh_configuration()
+        self.assertFalse(self.settings.auto_paste.isChecked())
+        self.settings._save()
+        self.assertFalse(self.conf["auto_paste"])
+        self.assertEqual(self.conf["shortcut"], "Ctrl+Shift+F9")
+        self.assertEqual(self.conf["groq_transcribe_model"], "external-model")
+        self.assertEqual(self.settings._shortcut_rows["toggle"][0].currentText(), "Ctrl+Shift+F9")
+        self.assertEqual(self.settings.dirty_label.text(), "")
+
+    def test_clean_form_refreshes_from_cli_without_changing_file_result(self):
+        self.settings.file_output.setPlainText("Existing file result")
+        self.conf["shortcut"] = "Ctrl+Alt+F9"
+        self.settings.refresh_configuration()
+        self.assertEqual(self.settings._shortcut_rows["toggle"][0].currentText(), "Ctrl+Alt+F9")
+        self.assertEqual(self.settings.file_output.toPlainText(), "Existing file result")
+        self.assertEqual(self.settings.dirty_label.text(), "")
+
+    def test_cached_provider_model_edits_remain_dirty_after_switching_back(self):
+        self.settings._select_data(self.settings.transcribe_provider, "groq")
+        self.settings.transcribe_model.setCurrentText("my-groq-model")
+        self.settings._select_data(self.settings.transcribe_provider, "openai")
+        self.assertEqual(self.settings.dirty_label.text(), "Unsaved changes")
+        self.settings._save()
+        self.assertEqual(self.conf["groq_transcribe_model"], "my-groq-model")
+
+    def test_cli_selected_new_source_is_resolved_when_settings_reopens(self):
+        self.conf["mic_target"] = "new-usb"
+        self.settings.refresh_configuration()
+        with mock.patch.object(home_ui.audio, "list_sources", return_value=[("new-usb", "New USB microphone")]), mock.patch.object(home_ui.audio, "list_monitors", return_value=[]):
+            self.settings.refresh_sources()
+        self.assertEqual(self.settings.mic.currentData(), "new-usb")
+        self.assertEqual(self.settings.mic.currentText(), "New USB microphone")
+        self.settings._save()
+        self.assertEqual(self.conf["mic_target"], "new-usb")
+
+    def test_language_rebuild_preserves_file_result_and_active_mode(self):
+        from dikte.meeting import MeetingPipeline
+        controller = Dikte.__new__(Dikte)
+        controller.__dict__.update(vars(self.controller))
+        controller.meetings = MeetingPipeline(self.conf)
+        controller._make_settings()
+        old_settings = controller.settings_window
+        old_home = home_ui.HomeWindow(controller, old_settings)
+        controller.home_window = old_home
+        old_home.show_mode("file")
+        old_home.show()
+        old_settings.file_path = "/tmp/chosen.wav"
+        old_settings.file_label.setText("chosen.wav")
+        old_settings.file_output.setPlainText("Retained transcript")
+        old_settings.file_segments = [{"text": "Retained transcript", "start": 0, "end": 2}]
+        i18n.set_language("tr")
+        controller._reopen_settings()
+        self.addCleanup(controller.settings_window.deleteLater)
+        self.addCleanup(controller.settings_window.close)
+        self.addCleanup(controller.home_window.deleteLater)
+        self.addCleanup(controller.home_window.close)
+        self.assertEqual(controller.home_window.mode, "file")
+        self.assertEqual(controller.settings_window.file_path, "/tmp/chosen.wav")
+        self.assertEqual(controller.settings_window.file_output.toPlainText(), "Retained transcript")
+        self.assertTrue(controller.settings_window.file_save_srt.isEnabled())
+        self.assertEqual(controller.home_window.mode_buttons["file"].text(), "Dosya")
+
+    def test_disconnected_configured_source_survives_an_unrelated_apply(self):
+        self.conf["mic_target"] = "disconnected-usb"
+        self.settings.refresh_configuration()
+        self.assertEqual(self.settings.mic.currentData(), "disconnected-usb")
+        self.settings.auto_paste.setChecked(False)
+        self.settings._save()
+        self.assertEqual(self.conf["mic_target"], "disconnected-usb")
+
+    def test_source_refresh_keeps_selection_and_discovers_hotplugged_devices(self):
+        self.settings.mic.addItem("Old microphone", "old")
+        self.settings.mic.setCurrentIndex(self.settings.mic.findData("old"))
+        with mock.patch.object(home_ui.audio, "list_sources", return_value=[("usb", "USB microphone")]), mock.patch.object(home_ui.audio, "list_monitors", return_value=[("loop", "Loopback")]):
+            self.settings.refresh_sources()
+        self.assertEqual(self.settings.mic.currentData(), "old")
+        self.assertGreaterEqual(self.settings.mic.findData("usb"), 0)
+        self.assertGreaterEqual(self.settings.meeting_mic.findData("usb"), 0)
+        self.assertGreaterEqual(self.settings.meeting_system.findData("loop"), 0)
+
+    def test_missing_assistant_directory_displays_the_actual_fallback(self):
+        self.conf["assistant_provider"] = "codex"
+        self.conf["assistant_dir"] = "/does/not/exist/dikte-test"
+        self.window.show_mode("ask")
+        self.assertNotIn(self.conf["assistant_dir"], self.window.ask_scope.text())
+        self.assertIn(home_ui.assistant.working_dir(self.conf), self.window.ask_scope.text())
+
+    def test_turkish_task_labels_and_runtime_status(self):
+        i18n.set_language("tr")
+        self.window.refresh()
+        self.assertEqual(self.window.capture_status.text(), "Konuşmaya hazır")
+        self.assertIn("Dikte:", self.window.capture_models.text())
+        self.assertIn("Temizleme:", self.window.capture_models.text())
+
+    def test_completed_run_refreshes_workspace_without_changing_controller_state(self):
+        controller = Dikte.__new__(Dikte)
+        controller.home_messages = {}
+        controller.home_window = self.window
+        controller._waiters = {}
+        cfg.append_history({"text": "Finished"})
+        controller._settle("dictation", {"ok": True, "text": "Finished"})
+        self.assertEqual(self.window.latest_text.toPlainText(), "Finished")
+        self.assertEqual(controller.home_messages["dictation"], "Transcript ready")

tests/test_ui.py

@@ -15,8 +15,8 @@ from typing import ClassVar
 from unittest import mock
 
 from PyQt6.QtCore import QPoint, QPointF, QRect, Qt
-from PyQt6.QtGui import QWheelEvent
+from PyQt6.QtGui import QHideEvent, QShowEvent, QWheelEvent
-from PyQt6.QtWidgets import QApplication, QMessageBox
+from PyQt6.QtWidgets import QApplication, QComboBox, QMessageBox, QSpinBox, QWidget
 
 from dikte import audio
 from dikte import cleanup
@@ -77,7 +77,7 @@ CHANGED = {
     "local_model": "ggml-small.bin",
     "local_gpu": False,
     "local_preload": False,
-    "local_threads": 6,
+    "local_threads": 1,
     "local_llm_model": "gemma-3-4b-it-Q4_K_M.gguf",
     "local_llm_repo": "ggml-org/gemma-4-E2B-it-GGUF",
     "local_llm_gpu": False,
@@ -179,10 +179,11 @@ class Settings(DikteTest):
                            Qt.KeyboardModifier.NoModifier,
                            Qt.ScrollPhase.NoScrollPhase, False)
 
-    def test_the_window_opens_with_every_tab_on_it(self):
+    def test_settings_keeps_configuration_and_exposes_separate_task_pages(self):
         window = self.window(cfg.Config())
         tabs = window.findChildren(settings_ui.QTabWidget)[0]
-        self.assertEqual(tabs.count(), 10)
+        self.assertEqual(tabs.count(), 7)
+        self.assertEqual(set(window.task_pages), {"file", "minutes", "history"})
         self.assertEqual(window.windowTitle(), "Dikte Settings")
 
     def test_no_tab_can_stretch_the_window_past_a_small_screen(self):
@@ -230,6 +231,51 @@ class Settings(DikteTest):
         QApplication.sendEvent(box, self.wheel())
         self.assertNotEqual(box.currentIndex(), before)
 
+    def test_the_wheel_uses_remembered_focus_in_an_inactive_window(self):
+        # Keep the window hidden so no desktop activation policy can give it
+        # keyboard focus. Its remembered focus still selects the wheel target.
+        for widget_type in (QComboBox, QSpinBox):
+            with self.subTest(widget=widget_type.__name__):
+                window = QWidget()
+                self.addCleanup(window.deleteLater)
+                box = widget_type(window)
+                other = QComboBox(window)
+                if isinstance(box, QComboBox):
+                    box.addItems(["first", "second", "third"])
+                    box.setCurrentIndex(1)
+                    value = box.currentIndex
+                else:
+                    box.setValue(5)
+                    value = box.value
+                box.setFocusPolicy(Qt.FocusPolicy.StrongFocus)
+                guard = settings_ui.WheelGuard(window)
+                box.installEventFilter(guard)
+                box.setFocus()
+                self.assertFalse(window.isActiveWindow())
+                self.assertFalse(box.hasFocus())
+                self.assertIs(window.focusWidget(), box)
+                before = value()
+                QApplication.sendEvent(box, self.wheel())
+                self.assertNotEqual(value(), before)
+                other.setFocus()
+                self.assertIs(window.focusWidget(), other)
+                before = value()
+                rolled = self.wheel()
+                QApplication.sendEvent(box, rolled)
+                self.assertEqual(value(), before)
+                self.assertFalse(rolled.isAccepted())
+
+    def test_the_wheel_is_refused_when_another_widget_has_focus(self):
+        window = self.window(cfg.Config())
+        box = window.ui_language
+        other = window.corner
+        other.setFocus()
+        before = box.currentIndex()
+        rolled = self.wheel()
+        QApplication.sendEvent(box, rolled)
+        self.assertEqual(box.currentIndex(), before)
+        self.assertFalse(rolled.isAccepted())
+
     def test_a_wrapped_label_keeps_the_room_its_lines_need(self):
         # The program path shares a row with a button, and a row is measured
         # before its width is known: the label has to claim the second line back
@@ -326,6 +372,56 @@ class Settings(DikteTest):
                 self.assertEqual(shown, [provider])
                 self.assertFalse(box.isHidden())
 
+    def test_editable_boxes_live_in_forms_that_grow_the_field_column(self):
+        window = self.window(cfg.Config())
+
+        def contains(layout, target):
+            for index in range(layout.count()):
+                item = layout.itemAt(index)
+                widget = item.widget()
+                if widget is target or (widget is not None and
+                                        widget.isAncestorOf(target)):
+                    return True
+                child = item.layout()
+                if child is not None and contains(child, target):
+                    return True
+            return False
+
+        forms = window.findChildren(settings_ui.QFormLayout)
+        boxes = [
+            window.paste_shortcut,
+            window.transcribe_model,
+            window.file_model,
+            window.cleanup_model,
+            window.cleanup_gemini_model,
+            window.cleanup_opencode_model,
+            window.cleanup_agy_model,
+            window.cleanup_claude_model,
+            window.cleanup_codex_model,
+            window.assistant_model,
+            window.assistant_agy_model,
+            window.assistant_opencode_model,
+            window.assistant_codex_model,
+            window.assistant_openrouter_model,
+            window.meeting_model,
+            *(box for box, _status, _missing in
+              window._shortcut_rows.values()),
+        ]
+        for box in boxes:
+            form = next((candidate for candidate in forms
+                         if contains(candidate, box)), None)
+            with self.subTest(box=box.objectName() or box.currentText()):
+                self.assertIsNotNone(form)
+                self.assertEqual(
+                    form.fieldGrowthPolicy(),
+                    settings_ui.QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow,
+                )
+
+        self.assertEqual(
+            window.local_llm.layout().fieldGrowthPolicy(),
+            settings_ui.QFormLayout.FieldGrowthPolicy.AllNonFixedFieldsGrow,
+        )
+
     def test_codex_answering_refills_both_of_its_boxes(self):
         """The list Codex gave replaces the built-in one, in both places, and
         neither loses what was already picked."""
@@ -1103,7 +1199,7 @@ class Overlay(DikteTest):
                 mock.patch.object(QApplication, "screenAt") as screen_at:
             widget._reposition()
         screen_at.assert_not_called()
-        self.assertEqual(widget.pos(), QPoint(1948, 995))
+        self.assertEqual(widget.pos(), QPoint(1948, 1003))
 
     def _screen(self, name, area):
         screen = mock.Mock()
@@ -1129,7 +1225,7 @@ class Overlay(DikteTest):
                 mock.patch.object(QApplication, "screenAt") as screen_at:
             widget._reposition()
         screen_at.assert_not_called()
-        self.assertEqual(widget.pos(), QPoint(1948, 995))
+        self.assertEqual(widget.pos(), QPoint(1948, 1003))
 
     def test_the_pointer_decides_when_the_compositor_will_not_say(self):
         """Every desktop but Plasma, and Plasma while KWin is being replaced."""
@@ -1141,7 +1237,7 @@ class Overlay(DikteTest):
                                   return_value=screens[0]) as screen_at:
             widget._reposition()
         screen_at.assert_called()
-        self.assertEqual(widget.pos(), QPoint(28, 995))
+        self.assertEqual(widget.pos(), QPoint(28, 1003))
 
     def _two_screens(self):
         return [self._screen("DP-1", settings_ui.QRect(0, 0, 1920, 1080)),
@@ -1163,10 +1259,10 @@ class Overlay(DikteTest):
         with mock.patch.object(overlay_module, "_kwin", kwin), \
                 mock.patch.object(QApplication, "screens", return_value=screens):
             widget.show_recording()
-        self.assertEqual(widget.pos(), QPoint(1948, 995))
+        self.assertEqual(widget.pos(), QPoint(1948, 1003))
         kwin.call.return_value.arguments.return_value = ["DP-1"]
         self._ticks_on(widget, screens, kwin)
-        self.assertEqual(widget.pos(), QPoint(28, 995))
+        self.assertEqual(widget.pos(), QPoint(28, 1003))
 
     def test_it_stays_where_it_appeared_unless_it_was_told_otherwise(self):
         """Left off, because an indicator that jumps desks mid-sentence is one
@@ -1179,7 +1275,7 @@ class Overlay(DikteTest):
             widget.show_recording()
         kwin.call.return_value.arguments.return_value = ["DP-1"]
         self._ticks_on(widget, screens, kwin)
-        self.assertEqual(widget.pos(), QPoint(1948, 995))
+        self.assertEqual(widget.pos(), QPoint(1948, 1003))
 
     def test_a_named_screen_is_never_left_for_the_pointer(self):
         """Naming one is the whole answer; following it would undo the naming."""
@@ -1190,7 +1286,7 @@ class Overlay(DikteTest):
             widget.show_recording()
         self._ticks_on(widget, screens, kwin)
         kwin.call.assert_not_called()
-        self.assertEqual(widget.pos(), QPoint(28, 995))
+        self.assertEqual(widget.pos(), QPoint(28, 1003))
 
     def test_the_one_on_top_goes_where_the_one_underneath_is(self):
         """Asking for itself would put the pair on two monitors, with this one
@@ -1204,8 +1300,8 @@ class Overlay(DikteTest):
             kwin.call.return_value.arguments.return_value = ["DP-1"]
             second = self.overlay(below=first)
             second.show_busy("Asking Claude…")
-        self.assertEqual(first.pos(), QPoint(1948, 995))
+        self.assertEqual(first.pos(), QPoint(1948, 1003))
-        self.assertEqual(second.pos(), QPoint(1948, 929))
+        self.assertEqual(second.pos(), QPoint(1948, 945))
 
     def test_the_compositor_is_asked_only_now_and_then(self):
         """Every tick would be thirty conversations a second about a hand
@@ -1309,6 +1405,7 @@ class LocalModels(DikteTest):
 
     def setUp(self):
         super().setUp()
+        self.enterContext(mock.patch.object(settings_ui.LocalModelBox, "_fetch_models"))
         # A machine Dikte is actually installed on would otherwise answer the
         # "nothing can transcribe" question from its real binary and model.
         self.patch_attr(ggml, "BIN_DIR", self.path("bin"))
@@ -1346,6 +1443,73 @@ class LocalModels(DikteTest):
             self.window(conf)._save()
         self.assertEqual(conf["local_model"], "ggml-large-v3-turbo-q5_0.bin")
 
+    def state(self, **values):
+        base = {"running": True, "pid": 3, "port": 4321, "model": "ggml-small.bin",
+                "gpu_wanted": True, "backend": "CUDA", "device": "RTX 4070",
+                "layers": "", "available": ["CUDA", "CPU"]}
+        base.update(values)
+        return base
+
+    def shown(self, **values):
+        """The line the window writes under the local model boxes."""
+        window = self.window(self.config(transcribe_provider="local"))
+        with mock.patch.object(ggml, "state",
+                               return_value={"whisper": self.state(**values),
+                                             "llama": self.state(running=False)}):
+            window._show_local_state()
+        return window.local_state.text(), window.local_llm_state.text()
+
+    def test_a_loaded_model_says_which_card_it_is_on(self):
+        whisper, llm = self.shown()
+        self.assertIn("graphics card", whisper)
+        self.assertIn("RTX 4070", whisper)
+        # The other box is about the other model, and that one is not loaded.
+        self.assertIn("Not loaded", llm)
+
+    def test_a_card_asked_for_and_missing_is_not_left_to_be_guessed_at(self):
+        whisper, _ = self.shown(backend="CPU", device="CPU", available=["CPU"])
+        self.assertIn("processor", whisper)
+        self.assertIn("only the CPU backend was loaded", whisper)
+
+    def test_a_download_is_not_assumed_to_lack_gpu_support(self):
+        whisper, _ = self.shown(backend="CPU", device="CPU", available=["CPU"],
+                                downloaded=True)
+        self.assertIn("only the CPU backend was loaded", whisper)
+        self.assertIn("driver errors", whisper)
+        self.assertNotIn("installing one", whisper)
+
+    def test_a_system_build_is_not_told_to_install_itself(self):
+        whisper, _ = self.shown(backend="CPU", device="CPU", available=["CPU"],
+                                downloaded=False)
+        self.assertIn("only the CPU backend was loaded", whisper)
+        self.assertNotIn("Dikte downloaded", whisper)
+
+    def test_a_build_that_could_have_used_one_says_the_other_thing(self):
+        whisper, _ = self.shown(backend="CPU", device="CPU",
+                                available=["CUDA", "CPU"])
+        self.assertIn("could not be used", whisper)
+        self.assertNotIn("no graphics backend", whisper)
+
+    def test_a_processor_nobody_argued_about_is_stated_plainly(self):
+        whisper, _ = self.shown(backend="CPU", device="CPU", gpu_wanted=False,
+                                available=["CPU"])
+        self.assertEqual(whisper, "Loaded on the processor (CPU).")
+
+    def test_a_server_that_said_nothing_is_not_answered_for(self):
+        """A whisper built by hand on a Mac prints no backend line at all."""
+        whisper, _ = self.shown(backend="", device="", available=[])
+        self.assertIn("did not say", whisper)
+
+    def test_the_line_stops_being_written_while_the_window_is_away(self):
+        # The events rather than show() and hide(): showing the window for real
+        # would send the same event down to the download boxes, which answer it
+        # by asking Hugging Face what models there are.
+        window = self.window(self.config(transcribe_provider="local"))
+        window.showEvent(QShowEvent())
+        self.assertTrue(window._local_state_timer.isActive())
+        window.hideEvent(QHideEvent())
+        self.assertFalse(window._local_state_timer.isActive())
+
     def test_nothing_is_fetched_for_a_window_nobody_opened(self):
         # DikteTest closes the network, so a request would fail the test. The
         # lists are asked for when the box is shown, not when it is built.
@@ -1461,6 +1625,29 @@ class LocalModels(DikteTest):
         self.assertNotIn("Ready", box.status.text())
         self.assertIn("program", box.status.text())
 
+    def test_a_program_set_in_the_settings_is_not_called_downloaded(self):
+        mine = self.path("my-whisper-server")
+        mine.write_text("#!/bin/sh\n")
+        mine.chmod(0o755)
+        self.patch_attr(ggml.shutil, "which", lambda name: None)
+        box = self.window(self.config(local_binary=str(mine))).local_whisper
+        self.assertIn(str(mine), box.program_label.text())
+        self.assertFalse(box.install_button.isVisibleTo(box))
+
+    def test_a_model_over_a_program_set_by_hand_is_ready(self):
+        # The program is there, it is just named by the settings rather than
+        # downloaded, and the status line looked past it.
+        mine = self.path("my-whisper-server")
+        mine.write_text("#!/bin/sh\n")
+        mine.chmod(0o755)
+        self.patch_attr(ggml.shutil, "which", lambda name: None)
+        path = ggml.whisper_model_path("ggml-small.bin")
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_bytes(b"not really a model")
+        box = self.window(self.config(local_binary=str(mine))).local_whisper
+        box.load("ggml-small.bin")
+        self.assertIn("Ready", box.status.text())
+
     def test_changing_the_publisher_changes_the_model(self):
         # The model chosen under the old publisher is not published by the new
         # one. Carried over, it was added back as "not downloaded" and selected
@@ -1502,6 +1689,25 @@ class LocalModels(DikteTest):
             time.sleep(0.05)
             _app.processEvents()
         self.assertEqual(fetch.call_count, 1)
+
+    def test_reloading_settings_keeps_the_fetched_model_choices(self):
+        repo = "ggml-org/SmolLM3-3B-GGUF"
+        box = self.window(self.config(local_llm_repo=repo)).local_llm
+        box._on_listed([("models", [self._item("first.gguf"), self._item("second.gguf")], repo)], "")
+        box.load("first.gguf", repo)
+        self.assertGreaterEqual(box.model.findData("second.gguf"), 0)
+        self.assertEqual(box.selected(), "first.gguf")
+
+    def test_reload_before_repository_debounce_does_not_reuse_previous_catalog(self):
+        first_repo = "ggml-org/SmolLM3-3B-GGUF"
+        second_repo = "ggml-org/gemma-3-4b-it-GGUF"
+        box = self.window(self.config(local_llm_repo=first_repo)).local_llm
+        box._on_listed([("models", [self._item("first.gguf"), self._item("second.gguf")], first_repo)], "")
+        box.repo.setCurrentText(second_repo)
+        box.load("first.gguf", second_repo)
+        self.assertLess(box.model.findData("second.gguf"), 0)
+        self.assertTrue(box._pending)
+        self.assertFalse(box._answered)
     def test_the_models_are_grouped_by_the_model_rather_than_by_size(self):
         # Sorted by size alone, the turbo files land between the two medium
         # ones, half a screen from the model they are a copy of.
@@ -1781,3 +1987,15 @@ class LocalModels(DikteTest):
                     # isHidden rather than isVisible: the window itself is never
                     # shown in a test, so nothing in it is ever visible.
                     self.assertEqual(other.isHidden(), name != chosen)
+
+    def test_local_threads_range_is_bounded_by_cpu_count(self):
+        with mock.patch("os.cpu_count", return_value=8):
+            window = self.window(cfg.Config())
+            self.assertEqual(window.local_threads.minimum(), 0)
+            self.assertEqual(window.local_threads.maximum(), 8)
+
+    def test_local_threads_range_has_safe_minimum_when_cpu_count_is_none(self):
+        with mock.patch("os.cpu_count", return_value=None):
+            window = self.window(cfg.Config())
+            self.assertEqual(window.local_threads.minimum(), 0)
+            self.assertEqual(window.local_threads.maximum(), 1)

tests/test_worker.py

@@ -35,7 +35,7 @@ class Chain(DikteTest):
                   cleaned="Book it for Thursday.",
                   cleanup_error=None, answer=("Booked.", ""), rms=None,
                   clipboard=b"what was there before", paste_error=None,
-                  focus=None):
+                  detected="en", focus=None):
         pipeline = worker.Pipeline(self.conf)
         done, failures, stages, cancels = [], [], [], []
         pipeline.finished.connect(lambda *args: done.append(args))
@@ -45,14 +45,19 @@ class Chain(DikteTest):
 
         cleanup = (mock.Mock(side_effect=cleanup_error) if cleanup_error
                    else mock.Mock(return_value=cleaned))
+        # Auto mode takes the detection path; a fixed language the plain one.
+        # Both are mocked so the chain runs either way without a server.
+        behavior = {"side_effect": transcribe_error} if transcribe_error \
+            else {"return_value": transcript}
+        detect_behavior = {"side_effect": transcribe_error} if transcribe_error \
+            else {"return_value": (transcript, detected)}
         calls = {}
         # The chain reports its own failures on stderr, which a test run has no
         # use for.
         with contextlib.redirect_stderr(io.StringIO()), \
-                mock.patch.object(
+                mock.patch.object(api, "transcribe", **behavior) as tr, \
-                    api, "transcribe",
+                mock.patch.object(api, "transcribe_detected",
-                    **({"side_effect": transcribe_error} if transcribe_error
+                                  **detect_behavior) as tdet, \
-                       else {"return_value": transcript})) as tr, \
                 mock.patch.object(api, "cleanup", cleanup), \
                 mock.patch.object(assistant, "ask", return_value=answer) as ask_call, \
                 mock.patch.object(paste, "copy") as copy, \
@@ -62,7 +67,8 @@ class Chain(DikteTest):
                                   return_value=clipboard) as read_clipboard, \
                 mock.patch.object(worker.time, "sleep", lambda seconds: None):
             press.side_effect = paste_error
-            calls = {"transcribe": tr, "cleanup": cleanup, "ask": ask_call,
+            calls = {"transcribe": tr, "transcribe_detected": tdet,
+                     "cleanup": cleanup, "ask": ask_call,
                      "copy": copy, "copy_bytes": copy_bytes, "press": press,
                      "read_clipboard": read_clipboard}
             pipeline._work(self.wav, duration,
@@ -77,7 +83,8 @@ class Chain(DikteTest):
         run = self.run_chain()
         self.assertEqual(run["failures"], [])
         self.assertEqual(run["done"][0],
-                         ("uh, book it for Thursday", "Book it for Thursday.", ""))
+                         ("uh, book it for Thursday", "Book it for Thursday.",
+                          "", "en"))
         run["copy"].assert_called_once_with("Book it for Thursday.")
         run["press"].assert_called_once_with(self.conf["paste_shortcut"],
                                              focus=None)
@@ -129,7 +136,7 @@ class Chain(DikteTest):
         self.conf["restore_clipboard"] = True
         run = self.run_chain(paste_error=paste.PasteError("not trusted"))
         self.assertEqual(run["failures"], [])
-        raw, text, warning = run["done"][0]
+        raw, text, warning, _lang = run["done"][0]
         self.assertIn("not trusted", warning)
         run["copy_bytes"].assert_not_called()
 
@@ -182,17 +189,41 @@ class Chain(DikteTest):
         self.assertEqual(run["transcribe"].call_args.kwargs["language"], "tr")
         self.assertEqual(run["transcribe"].call_args.kwargs["prompt"], "Paraşüt")
 
+    def test_auto_mode_asks_for_the_detected_language_and_records_it(self):
+        run = self.run_chain(detected="tr")
+        told = run["transcribe_detected"].call_args.kwargs
+        self.assertEqual(told["language"], "auto")
+        self.assertEqual(cfg.read_history()[0]["speech_language"], "tr")
+        self.assertEqual(run["done"][0][3], "tr")
+        run["transcribe"].assert_not_called()
+
+    def test_the_detected_language_is_told_to_the_cleanup_prompt(self):
+        # The mock stands in for api.cleanup, which the cleanup module calls
+        # with (text, key, model, system_prompt, …); the prompt is the fourth.
+        self.conf["transcribe_prompt"] = "Paraşüt"
+        run = self.run_chain(detected="tr")
+        prompt = run["cleanup"].call_args.args[3]
+        # Turkish was detected, so the Turkish glossary rule is appended.
+        self.assertIn("KONUŞMACININ KULLANDIĞI İSİM VE TERİMLER", prompt)
+
+    def test_a_fixed_language_needs_no_detection(self):
+        self.conf["language"] = "en"
+        run = self.run_chain()
+        run["transcribe"].assert_called_once()
+        run["transcribe_detected"].assert_not_called()
+        self.assertEqual(cfg.read_history()[0]["speech_language"], "en")
+
     # ---- silence and stock phrases ----------------------------------------
 
     def test_room_tone_costs_no_api_call(self):
         run = self.run_chain(rms=[0.00001] * 60)
-        run["transcribe"].assert_not_called()
+        run["transcribe_detected"].assert_not_called()
         self.assertIn("No speech", run["failures"][0])
 
     def test_the_silence_check_can_be_switched_off(self):
         self.conf["skip_silent"] = False
         run = self.run_chain(rms=[0.00001] * 60)
-        run["transcribe"].assert_called_once()
+        run["transcribe_detected"].assert_called_once()
 
     def test_a_stock_phrase_from_a_short_clip_is_thrown_away(self):
         run = self.run_chain(duration=2.0, transcript="Altyazı M.K.")
@@ -208,7 +239,7 @@ class Chain(DikteTest):
 
     def test_a_failed_cleanup_still_pastes_the_transcript(self):
         run = self.run_chain(cleanup_error=api.ApiError("rate limited"))
-        _raw, text, warning = run["done"][0]
+        _raw, text, warning, _lang = run["done"][0]
         self.assertEqual(text, "uh, book it for Thursday")
         self.assertIn("rate limited", warning)
         run["copy"].assert_called_once_with("uh, book it for Thursday")
@@ -220,6 +251,9 @@ class Chain(DikteTest):
         self.assertEqual(cfg.read_history()[0]["cleanup_error"], "bad key")
 
     def test_a_failed_transcription_ends_the_run(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)
@@ -231,6 +265,9 @@ class Chain(DikteTest):
         copy.assert_not_called()
 
     def test_a_clipboard_that_will_not_take_it(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)
@@ -243,6 +280,9 @@ class Chain(DikteTest):
         self.assertIn("wl-copy", failures[0])
 
     def test_an_unexpected_error_is_reported_rather_than_swallowed(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)
@@ -280,6 +320,9 @@ class Chain(DikteTest):
         run["press"].assert_not_called()
 
     def test_a_command_that_was_cancelled(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         cancels = []
         pipeline.cancelled.connect(lambda: cancels.append(True))
@@ -289,6 +332,9 @@ class Chain(DikteTest):
         self.assertEqual(cancels, [True])
 
     def test_an_agent_that_is_not_installed(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)