feat: Local | With AI switch, small settings gear, quick provider/model selection

.claude/skills/gui-usability-review/SKILL.md

@@ -0,0 +1,51 @@
+---
+name: gui-usability-review
+description: Reproducible usability review of the Blitztext main window - offscreen clicks, screenshots, hard accessibility checks and a blind LLM rating. Use after changing app/main_window.py, theme.py or popup/toolbar layout.
+---
+
+# GUI usability review
+
+## Purpose
+Measure how quickly a user reaches common goals in the Qt main window and catch
+accessibility defects, without a desktop session.
+
+## Trigger
+Any change to `app/main_window.py`, `app/theme.py`, toolbar/popup controls or the
+controller methods the window calls.
+
+## Inputs
+- `scripts/gui_usability_check.py` (task list `TASKS`, click budgets)
+- `tests/test_gui_usability.py` (hard gates)
+
+## Procedure
+1. Run the driver: `QT_QPA_PLATFORM=offscreen .venv/bin/python scripts/gui_usability_check.py`.
+   It prints a fresh temp directory containing `report.json` and one PNG per click step; copy what you want to keep.
+2. Hard gates (no LLM): `WHISPER_GUI_TESTS=1 QT_QPA_PLATFORM=offscreen .venv/bin/python -m pytest -q tests/test_gui_usability.py`.
+   Click budgets, goal reached, accessible names, keyboard focus, minimum target 22 px.
+3. Blind LLM rating, at least 3 independent runs per task. The reviewer gets ONLY
+   the task text and the PNG sequence of that task - never code, never this file's
+   results. Use a fresh subagent per run.
+4. Reviewer output, JSON only:
+   `{"task": id, "score": 1-5, "findings": [{"widget": "...", "severity": "low|medium|high", "issue": "..."}], "suggestion": "..."}`
+5. Score = mean over runs. Store under `.omc/uebergabe/gui-review/<date>/`, compare with the frozen baseline.
+
+## Rubric (score 1-5 per task)
+Findability of the control, number of steps, label clarity, feedback after the
+action (is the new state visible?), error/edge handling, target size and contrast.
+5 = obvious and immediate, 3 = works but needs hunting, 1 = most users fail.
+
+## Verification
+Hard gates green; mean score not below the baseline; no high-severity finding unaddressed.
+
+## Rollback / abort
+The driver only reads the UI with a fake controller; nothing to roll back. Abort the
+review if the screenshots are blank (check `QT_QPA_PLATFORM=offscreen`).
+
+## Known pitfalls
+- A drop-down counts as two clicks (open, choose); the driver sets the combo index directly, so budgets are assumed clicks, not measured ones.
+- Offscreen renders no native Wayland decorations or real fonts; judge layout, not font metrics.
+- Setting a combo to its current value emits no signal; set a different state first.
+- Reviewer prompts must never include the source, or the rating is no longer blind.
+
+## References
+`app/main_window.py`, `scripts/gui_usability_check.py`, Nielsen's 10 usability heuristics.

README.de.md

@@ -209,7 +209,7 @@ Das Hauptfenster ist ein kompaktes, rahmenloses Desktop-Widget mit einheitlicher
       <td align="center" width="25%">
         <img src="docs/screenshots/linux/main-menu-de.png" width="170"><br><br>
         <b>Aktions-Menü</b><br>
-        <i>Workflows und Schnellzugriff direkt am Widget.</i>
+        <i>Details zur gewählten Seite: Aktion, Stil, KI-Anbieter, Whisper-Modell.</i>
       </td>
     </tr>
   </table>
@@ -220,12 +220,13 @@ Das Hauptfenster ist ein kompaktes, rahmenloses Desktop-Widget mit einheitlicher
   - **Bereit (IDLE):** Grüner Statuspunkt und Zeitanzeige (`Bereit: 00:00`).
   - **Aufnahme (RECORDING):** Gelber Statuspunkt, sekundengenaue Zeitanzeige und animierte Wellenbalken links und rechts des Mikrofons.
   - **Verarbeitung (TRANSCRIBING):** Rotierender goldener Kreisbogen um das Mikrofon (`Transkribiere…`).
+- **Umschalter Lokal | Mit KI:** Zwei Knöpfe direkt unter dem Mikrofon. **Lokal** = nur lokale Whisper-Transkription, nichts verlässt den Rechner. **Mit KI** = lokal transkribieren, danach bearbeitet dein Cloud-KI-Anbieter den Text. Die gewählte Seite ist gold; die Zustandszeile darunter (zum Beispiel `Lokal · Whisper base` oder `Mit KI · OpenAI · Verbessern`) zeigt, was die nächste Aufnahme tut, und öffnet per Klick das Detail-Popup. Hotkey-Workflows stellen den Umschalter auf die passende Seite, während der Aufnahme ist er gesperrt.
 - **Kompakte Werkzeugleiste:**
   - ✍ **Compose-Button** (hervorgehobenes Stift-Symbol): Öffnet das Entwurfsfenster für Texteingabe und KI-Verfeinerung.
   - 🕒 **Verlauf-Button** (mit Zähler-Badge): Öffnet den Transkript- und Diktat-Verlauf.
   - 🔊 **Vorlesen-Button** (Lautsprecher-Symbol): Öffnet das TTS-Fenster zur Sprachausgabe.
-  - ⚙ **Einstellungen-Button** (Zahnrad): Öffnet den Konfigurations-Dialog.
-  - ⋯ **Mehr-Button**: Öffnet das Popup-Menü für *Diktieren & verbessern*, Text-Aktionen (*Kürzen*, *Ausformulieren*, *Tonfall ändern*), *Eigene Anweisung*, *Sammel-Diktat* und *Verwerfen*.
+  - ⋯ **Mehr-Button** (gleiches Popup wie die Zustandszeile): Zeigt nur die Felder, die zur gewählten Seite passen — *Lokal*: Whisper-Modell; *Mit KI*: Aktion (*Diktieren & verbessern*, *Sachlich formulieren*, *Emojis ergänzen*), Schreibstil für *Diktieren & verbessern* (*Kürzen*, *Ausformulieren*, *Tonfall ändern*, *Eigene Anweisung*), KI-Anbieter (OpenAI oder OpenRouter; einen eigenen Endpunkt änderst du nur in den Einstellungen) und Whisper-Modell. Ein **Fertig**-Knopf schließt das Popup; *Sammel-Diktat* und *Verwerfen* stehen darunter. Der Anbieterwechsel braucht die passende API-Key-Variable, größere Whisper-Modelle fragen vorher, weil sie heruntergeladen werden können.
+- **Einstellungs-Zahnrad:** Kleiner ⚙-Knopf in der Titelleiste neben Minimieren und Schließen; öffnet den Konfigurations-Dialog.
 
 *Das Widget öffnet sich beim Start sowie über den Tray-Eintrag **Fenster anzeigen** oder einen Klick auf das Tray-Icon. Schließen blendet das Widget aus — die App läuft im Hintergrund im Tray weiter.*
 

README.md

@@ -208,7 +208,7 @@ The main window is a compact, frameless desktop widget with a consistent dark lo
       <td align="center" width="25%">
         <img src="docs/screenshots/linux/main-menu-en.png" width="170"><br><br>
         <b>Options menu</b><br>
-        <i>Workflows and quick actions.</i>
+        <i>Details for the chosen side: action, style, AI provider, Whisper model.</i>
       </td>
     </tr>
   </table>
@@ -219,12 +219,13 @@ The main window is a compact, frameless desktop widget with a consistent dark lo
   - **Idle:** Green status dot with duration counter (`Ready: 00:00`).
   - **Recording:** Amber indicator, live recording timer, and animated stereo wave bars flanking the mic.
   - **Processing:** Rotating gold progress ring around the microphone (`Transcribing…`).
+- **Local | With AI switch:** Two buttons right under the microphone. **Local** = local Whisper transcription only, nothing leaves your machine. **With AI** = local transcription, then the text is edited by your cloud AI provider. The chosen side is gold; the state line below it (for example `Local · Whisper base` or `With AI · OpenAI · Improve`) shows what the next recording does and opens the details popup when clicked. Hotkey workflows move the switch to the matching side, and it is locked while recording.
 - **Tool row:**
   - ✍ **Compose button** (highlighted pen icon): Opens the Compose window for drafting and rewriting.
   - 🕒 **History button** (with counter badge): Opens the transcript history with recent entries.
   - 🔊 **Read aloud button** (speaker icon): Opens the TTS window to synthesize speech.
-  - ⚙ **Settings button** (gear icon): Opens the settings dialog.
-  - ⋯ **More options button**: Pops up quick access to *Dictate & improve*, text improvement presets (*Shorten*, *Expand*, *Change tone*), *Custom instruction*, *Batch dictation*, and *Discard*.
+  - ⋯ **More options button** (same popup as the state line): Shows only the fields that apply to the chosen side — *Local*: Whisper model; *With AI*: action (*Dictate & improve*, *Make it factual*, *Add emojis*), the writing style for *Dictate & improve* (*Shorten*, *Expand*, *Change tone*, *Custom instruction*), the AI provider (OpenAI or OpenRouter; a custom endpoint is changed in the settings only), and the Whisper model. A **Done** button closes the popup; *Batch dictation* and *Discard* sit below it. Switching the AI provider needs the matching API key variable; larger Whisper models ask before they may be downloaded.
+- **Settings gear:** Small ⚙ button in the title bar next to minimize and close; opens the settings dialog.
 
 *The widget opens at startup, via the tray entry **Show window**, or by clicking the tray icon. Closing only hides the widget — the app keeps running in the background tray.*
 

app/blitztext_linux.py

@@ -1067,6 +1067,95 @@ def main_window_preset_changed(self, key: str) -> None:
         self._refresh_preset_menu()
         logger.info("Writing preset changed via main window: %s", key)
 
+    def main_window_provider_changed(self, provider: str) -> None:
+        """Switch between OpenAI and OpenRouter from the main window.
+
+        A custom endpoint is only changed in the settings dialog: leaving it here would
+        overwrite the user's URL and key variable, and entering it here could pair a stale
+        URL with an OpenAI key. Without a usable key the switch is refused.
+        """
+        cfg = self.config
+        if provider == cfg.llm_provider:
+            return
+        previous = (cfg.llm_provider, cfg.llm_base_url, cfg.openai_api_key_env)
+        problem = None
+        if provider == "custom" or cfg.llm_provider == "custom":
+            problem = (t("settings.save_error.title"), t("mainwindow.provider.custom_in_settings.message"))
+        else:
+            if cfg.llm_provider == "openai":
+                # Remember a personal OpenAI key variable so the way back restores it.
+                self._openai_key_env_memo = cfg.openai_api_key_env
+            cfg.llm_provider = provider
+            if provider == "openrouter":
+                cfg.llm_base_url = "https://openrouter.ai/api/v1"
+                cfg.openai_api_key_env = "OPENROUTER_API_KEY"
+            else:
+                cfg.llm_base_url = ""
+                cfg.openai_api_key_env = getattr(
+                    self, "_openai_key_env_memo", DEFAULTS["openai_api_key_env"]
+                )
+            if not cfg.resolve_llm_api_key():
+                problem = (
+                    t("mainwindow.provider.no_key.title"),
+                    t("mainwindow.provider.no_key.message").format(env=cfg.openai_api_key_env),
+                )
+        if problem is None:
+            try:
+                cfg.save()
+            except Exception as exc:
+                logger.error("Failed to save LLM provider: %s", exc)
+                problem = (
+                    t("settings.save_error.title"),
+                    t("settings.save_error.message").format(error=exc),
+                )
+        if problem is not None:
+            cfg.llm_provider, cfg.llm_base_url, cfg.openai_api_key_env = previous
+            if self._main_window is not None:
+                self._main_window.set_llm_provider(cfg.llm_provider)
+            self.show_tray_warning(*problem)
+            return
+        self._rebuild_llm_service()
+        self.update_menu_availability()
+        if self._main_window is not None:
+            self._main_window.set_llm_provider(provider)
+        logger.info("LLM provider changed via main window: %s", provider)
+
+    def _confirm_model_download(self, model: str) -> bool:
+        if self._main_window is not None:
+            # The selector popup is a Qt.Popup; a modal box over it would be unreliable.
+            self._main_window._options_popup.hide()
+        answer = QMessageBox.question(
+            self._main_window,
+            t("mainwindow.model.download.title"),
+            t("mainwindow.model.download.message").format(model=model),
+        )
+        return answer == QMessageBox.StandardButton.Yes
+
+    def main_window_model_changed(self, model: str) -> None:
+        """Switch the Whisper model; larger models ask first because they may download."""
+        cfg = self.config
+        if model == cfg.model:
+            return
+        previous = cfg.model
+        if model not in ("tiny", "base") and not self._confirm_model_download(model):
+            if self._main_window is not None:
+                self._main_window.set_model(previous)
+            return
+        cfg.model = model
+        try:
+            cfg.save()
+        except Exception as exc:
+            cfg.model = previous
+            if self._main_window is not None:
+                self._main_window.set_model(previous)
+            logger.error("Failed to save model: %s", exc)
+            self.show_tray_warning(
+                t("settings.save_error.title"),
+                t("settings.save_error.message").format(error=exc),
+            )
+            return
+        logger.info("Whisper model changed via main window: %s", model)
+
     def start_hotkey_worker(self) -> None:
         self.stop_hotkey_worker()
 
@@ -1120,6 +1209,8 @@ def show_settings_dialog(self) -> None:
             self._refresh_preset_menu()
             if self._main_window is not None:
                 self._main_window.set_preset(self.config.writing_preset)
+                self._main_window.set_llm_provider(self.config.llm_provider)
+                self._main_window.set_model(self.config.model)
 
             # Restart hotkey listener if mode or key changed
             if self.hotkey_worker and (
@@ -1194,6 +1285,13 @@ def gui_toggle_recording(self, workflow: WorkflowType) -> None:
         Fallback wenn der Hotkey nicht greift oder keine Tastatur zur Hand ist.
         """
         if self.state == "IDLE":
+            if workflow in LLM_WORKFLOWS and not self.llm_service.is_available():
+                # Fail before speaking: otherwise the transcript is lost at the rewrite step.
+                self.show_tray_warning(
+                    t("mainwindow.provider.no_key.title"),
+                    t("mainwindow.provider.no_key.message").format(env=self.config.openai_api_key_env),
+                )
+                return
             self._start_recording(workflow)
         elif self.state == "RECORDING":
             self._stop_recording_and_process()
@@ -1456,6 +1554,8 @@ def _ensure_main_window(self) -> MainWindow:
                 window.set_history_count(self._history_panel.entry_count)
             window.update_state(self.state, self.current_workflow, self._tray_error_message)
             window.set_preset(self.config.writing_preset)
+            window.set_llm_provider(self.config.llm_provider)
+            window.set_model(self.config.model)
             self._main_window = window
         return self._main_window
 

app/config.py

@@ -159,14 +159,21 @@ def save(self) -> None:
             tmp = self.config_file.with_suffix(".json.tmp")
             flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC
             fd = os.open(str(tmp), flags, 0o600)
+            # Fix the mode before the file can replace the config: a failure after the
+            # replace would report an error although the new content is already live.
+            try:
+                os.fchmod(fd, 0o600)
+            except OSError:
+                os.close(fd)
+                tmp.unlink(missing_ok=True)
+                raise
             payload = _deep_copy_without_legacy_key(self._data)
             if not self._tts_env_explicit:
                 payload.pop("tts_openai_api_key_env", None)
             with os.fdopen(fd, "w", encoding="utf-8") as f:
                 json.dump(payload, f, indent=2, ensure_ascii=False)
                 f.write("\n")
             tmp.replace(self.config_file)
-            self.config_file.chmod(0o600)
             self._data = payload
             self._legacy_openai_api_key_present = False
             self._legacy_openai_api_key_value = ""