fix(tray): show the tray before the speech model loads (#51)

The first model load runs on a background thread after the tray is up. While it loads the tray shows the processing icon and 'Loading model…', and a dictation request gets a 'still loading' notification. A failed load leaves the tray in the error state instead of exiting.

Fixes #50

src/vocalinux/main.py

@@ -621,6 +621,8 @@ def main():
             remote_api_key=saved_settings.get("remote_api_key", ""),
             remote_api_endpoint=saved_settings.get("remote_api_endpoint", "/inference"),
             remote_api_model=saved_settings.get("remote_api_model", "whisper-1"),
+            # The model loads in the background so the tray can appear first (#50).
+            defer_model_load=True,
         )
 
         # Initialize text injection system
@@ -1382,6 +1384,11 @@ def main():
             dictation_pad=dictation_pad,
         )
 
+        # Start the model load from the GTK main loop, after the tray's own idle
+        # callback has put the icon up. The load runs on its own thread, so the
+        # tray is usable while it loads (#50).
+        GLib.idle_add(speech_engine.load_model_in_background)
+
         # Start the GTK main loop
         indicator.run()
 

src/vocalinux/speech_recognition/recognition_manager.py

@@ -288,6 +288,9 @@ def resolve_language_preference(language: str) -> str:
 
 logger = logging.getLogger(__name__)
 
+# Engine names the constructor accepts; checked up front when the load is deferred (#50).
+_SUPPORTED_ENGINES = ("vosk", "whisper", "whisper_cpp", "parakeet", "faster_whisper", "remote_api")
+
 _WHISPERCPP_ABORT_CALLBACK_SUPPORT: dict[type, bool] = {}
 
 
@@ -570,6 +573,12 @@ class SpeechRecognitionManager:
         # cancel-and-retry cannot stack unaccounted request threads.
         self._download_openers: set[threading.Thread] = set()
         self._defer_download = defer_download
+        # Startup may defer the first model load to a background thread (#50):
+        # the engine reports model_loading until that load has finished.
+        self._model_loading = bool(kwargs.get("defer_model_load", False))
+        self._model_load_started = False
+        self._model_load_thread: Optional[threading.Thread] = None
+        self._model_load_callbacks: list[Callable[[], None]] = []
         self._model_initialized = False
         # True while a configured app/game blocks dictation, including pending unloads
         self._auto_paused = False
@@ -690,21 +699,14 @@ class SpeechRecognitionManager:
             f"Initializing speech recognition with {engine} engine, {self.language} language and {model_size} model"
         )
 
-        # Initialize the selected speech recognition engine
+        # Initialize the selected speech recognition engine. A deferred load is
-        if engine == "vosk":
+        # started later by load_model_in_background(); the engine name is still
-            self._init_vosk()
+        # checked now so a bad name fails at construction, as before.
-        elif engine == "whisper":
+        if self._model_loading:
-            self._init_whisper()
+            if engine not in _SUPPORTED_ENGINES:
-        elif engine == "whisper_cpp":
+                raise ValueError(f"Unsupported speech recognition engine: {engine}")
-            self._init_whispercpp()
-        elif engine == "parakeet":
-            self._init_parakeet()
-        elif engine == "faster_whisper":
-            self._init_faster_whisper()
-        elif engine == "remote_api":
-            self._init_remote_api()
         else:
-            raise ValueError(f"Unsupported speech recognition engine: {engine}")
+            self._init_selected_engine()
 
     @staticmethod
     def _normalize_language_candidates(value: str | list[str] | None) -> list[str]:
@@ -3300,6 +3302,83 @@ class SpeechRecognitionManager:
         for callback in self.state_callbacks:
             callback(new_state)
 
+    @property
+    def model_loading(self) -> bool:
+        """True from a deferred construction until its background load has finished (#50)."""
+        return bool(getattr(self, "_model_loading", False))
+
+    def register_model_load_callback(self, callback: Callable[[], None]) -> None:
+        """Call ``callback()`` on the loader thread once a deferred load has finished.
+
+        Fires whether the load succeeded, failed or was skipped. Consumers marshal
+        onto the GTK main loop themselves (#50).
+        """
+        self._model_load_callbacks.append(callback)
+
+    def load_model_in_background(self) -> None:
+        """Start the deferred first model load on a daemon thread (#50).
+
+        A no-op unless the manager was built with ``defer_model_load=True``, and
+        only the first call starts a load. The request token is taken here, on
+        the caller's thread, so an auto-pause that lands before the loader gets
+        the model lock supersedes the load instead of racing it.
+        """
+        if not self._model_loading or self._model_load_started:
+            return
+        self._model_load_started = True
+        request = self._new_model_request()
+        self._model_load_thread = threading.Thread(
+            target=self._run_background_model_load,
+            args=(request,),
+            name="voculinux-model-load",
+            daemon=True,
+        )
+        self._model_load_thread.start()
+
+    def _run_background_model_load(self, request: object) -> None:
+        """Initialize the configured engine on the loader thread (#50).
+
+        Shares the model lock and request ownership with resume, but skips the
+        resume-only bookkeeping: a startup load publishes the model and nothing
+        else. The engine's own loading ladder and fallbacks run unchanged.
+        """
+        started = time.monotonic()
+        try:
+            with self._model_lock:
+                # An auto-pause that is already active (a listed app is running)
+                # must keep the model out of memory; auto-resume loads it later.
+                if self._auto_paused or not self._is_current_model_request(request):
+                    logger.info("Speech model load skipped: auto-paused or superseded")
+                else:
+                    with self._model_request_context(request):
+                        self._init_selected_engine()
+                    logger.info("Speech model loaded in %.2fs", time.monotonic() - started)
+        except Exception as error:
+            logger.error("Speech model failed to load", exc_info=True)
+            if self._is_current_model_request(request):
+                self._update_state(RecognitionState.ERROR, model_request=request)
+                _show_notification(
+                    "Speech model failed to load",
+                    f"{error}. Restart Voculinux to try again.",
+                    "dialog-error",
+                )
+        finally:
+            self._model_loading = False
+            self._notify_model_load_callbacks()
+
+    def _is_current_model_request(self, request: object) -> bool:
+        """Whether no auto-pause or reconfigure has superseded ``request``."""
+        with getattr(self, "_model_request_lock", nullcontext()):
+            return self._model_request is request
+
+    def _notify_model_load_callbacks(self) -> None:
+        """Run each load-state callback; one failing consumer must not stop the rest."""
+        for callback in list(self._model_load_callbacks):
+            try:
+                callback()
+            except Exception:
+                logger.debug("Model load callback failed", exc_info=True)
+
     @property
     def model_ready(self) -> bool:
         """Check if the model is initialized and ready for recognition."""
@@ -3501,6 +3580,17 @@ class SpeechRecognitionManager:
             return False
 
         reload_in_background = False
+        # The first load (#50) is still running. Nothing can be dictated yet, and
+        # the missing-model offer below would be wrong, so say that instead.
+        if self.model_loading and not self.model_ready:
+            logger.info("Cannot start recognition: speech model is still loading")
+            play_error_sound()
+            _show_notification(
+                "Speech model loading",
+                "Dictation will be available once the speech model has finished loading.",
+                "dialog-information",
+            )
+            return False
         # Check if model is ready (lazy-reload after idle keep-alive unload)
         if not self.model_ready:
             if self._idle_unloaded:
@@ -3932,7 +4022,13 @@ class SpeechRecognitionManager:
             # select it.
             self._refuse_language_override(language)
             return False
-        if target != self.language and not self._serving_model_supports_language_switch():
+        # While the first load runs there is no serving model to judge yet; the
+        # start itself is refused with the loading notice.
+        if (
+            target != self.language
+            and not self.model_loading
+            and not self._serving_model_supports_language_switch()
+        ):
             self._refuse_language_override(language)
             return False
 

src/vocalinux/ui/settings_dialog.py

@@ -9346,6 +9346,10 @@ class SettingsDialog(AboutSectionMixin):
                     "Dictation is paused. Close the listed app or remove it from "
                     "Auto-Pause settings."
                 )
+            elif getattr(self.speech_engine, "model_loading", False) is True:
+                self.test_buffer.set_text(
+                    "The speech model is still loading. Try Test Dictation again in a moment."
+                )
             elif not getattr(self.speech_engine, "model_ready", True):
                 self.test_buffer.set_text(
                     "No speech model downloaded. Open the Speech Model page and "

src/vocalinux/ui/tray_indicator.py

@@ -161,6 +161,9 @@ _INPUT_SETTLE_SECONDS = 2
 _INPUT_MONITOR_CAP_SECONDS = 10
 _FALLBACK_KEYBOARD_RESTART_SECONDS = 6
 
+# Tray title while the speech model loads in the background (#50).
+LOADING_TITLE = "Loading model…"
+
 
 class TrayIndicator:
     """
@@ -282,6 +285,11 @@ class TrayIndicator:
 
         # Register for speech recognition state changes
         self.speech_engine.register_state_callback(self._on_recognition_state_changed)
+        # The model may still be loading in the background (#50); re-render when
+        # that load finishes.
+        register_load_callback = getattr(self.speech_engine, "register_model_load_callback", None)
+        if callable(register_load_callback):
+            register_load_callback(self._on_model_load_changed)
 
         # Offer the download instead of only telling the user a model is missing
         self._model_download_active = False
@@ -913,6 +921,25 @@ class TrayIndicator:
         if isinstance(indicator, GMenuIndicator):
             indicator.refresh_menu()
 
+    def _engine_model_loading(self) -> bool:
+        """True only while the engine reports its first model load in progress (#50)."""
+        return getattr(self.speech_engine, "model_loading", False) is True
+
+    def _on_model_load_changed(self) -> None:
+        """Model load finished. Runs on the loader thread, so marshal it (#50)."""
+        GLib.idle_add(self._apply_model_load_state)
+
+    def _apply_model_load_state(self) -> bool:
+        """Re-render after a model load changed state (GTK main loop only, #50)."""
+        if not self._engine_model_loading():
+            # Keep-alive refuses to unload while the load runs; restart its
+            # idle timer now that the model is in memory.
+            keepalive = getattr(self, "_model_keepalive", None)
+            if keepalive is not None:
+                keepalive.bump()
+        self._update_ui(getattr(self.speech_engine, "state", RecognitionState.IDLE))
+        return False  # One-shot idle callback
+
     def _on_recognition_state_changed(self, state: RecognitionState):
         """
         Handle changes in the speech recognition state.
@@ -1323,7 +1350,12 @@ class TrayIndicator:
         show_cancel = state == RecognitionState.PROCESSING and transcribing
         self._set_cancel_menu_item_visible(show_cancel)
 
-        if state == RecognitionState.IDLE:
+        if state == RecognitionState.IDLE and self._engine_model_loading():
+            self._set_indicator_title(LOADING_TITLE)
+            self._set_indicator_icon(self._icon_keys["processing"], "Loading speech model")
+            self._set_menu_item_enabled("Start Voice Typing", False)
+            self._set_menu_item_enabled("Stop Voice Typing", False)
+        elif state == RecognitionState.IDLE:
             self._set_indicator_title(TRAY_TITLE)
             self._set_indicator_icon(self._icon_keys["default"], "Microphone off")
             self._set_menu_item_enabled("Start Voice Typing", True)
@@ -1714,6 +1746,9 @@ class TrayIndicator:
 
     def _is_safe_for_keepalive_unload(self) -> bool:
         """Keep-alive may unload only when idle and not under auto-pause."""
+        if self._engine_model_loading():
+            # The first load (#50) is still putting the model in memory.
+            return False
         if getattr(self.speech_engine, "is_auto_paused", False) is True:
             return False
         if (

tests/test_main.py

@@ -257,6 +257,7 @@ class TestMainModule(unittest.TestCase):
                 remote_api_key="",
                 remote_api_endpoint="/inference",
                 remote_api_model="whisper-1",
+                defer_model_load=True,
             )
             mock_text.assert_called_once_with(wayland_mode=True)
             mock_action_handler.assert_called_once_with(mock_text_instance)
@@ -286,6 +287,69 @@ class TestMainModule(unittest.TestCase):
             # Verify the tray indicator was started
             mock_tray_instance.run.assert_called_once()
 
+    @patch("vocalinux.main.check_dependencies")
+    @patch("vocalinux.ui.dictation_pad.DictationPad")
+    @patch("vocalinux.ui.action_handler.ActionHandler")
+    @patch("vocalinux.speech_recognition.recognition_manager.SpeechRecognitionManager")
+    @patch("vocalinux.text_injection.text_injector.TextInjector")
+    @patch("vocalinux.ui.tray_indicator.TrayIndicator")
+    @patch("vocalinux.main.logging")
+    @patch("vocalinux.ui.config_manager.get_shared_config_manager")
+    @patch("vocalinux.ui.logging_manager.initialize_logging")
+    def test_tray_is_shown_before_the_model_load_is_scheduled(
+        self,
+        mock_init_logging,
+        mock_config_manager,
+        mock_logging,
+        mock_tray,
+        mock_text,
+        mock_speech,
+        mock_action_handler,
+        mock_dictation_pad,
+        mock_check_deps,
+    ):
+        """The tray is built first, then the model load is queued, then the loop runs (#50)."""
+        mock_check_deps.return_value = True
+        mock_config_instance = MagicMock()
+        mock_config_instance.get_bool.return_value = True
+        mock_config_instance.get_settings.return_value = {
+            "speech_recognition": {},
+            "general": {"first_run": False},
+        }
+        mock_config_instance.get_model_size_for_engine.return_value = "medium"
+        mock_config_instance.get_str.return_value = ""
+        mock_config_instance.is_dictate_to_pad_enabled.return_value = False
+        mock_config_manager.return_value = mock_config_instance
+
+        mock_speech_instance = MagicMock()
+        mock_speech.return_value = mock_speech_instance
+        mock_tray_instance = MagicMock()
+        mock_tray.return_value = mock_tray_instance
+
+        order = MagicMock()
+        order.attach_mock(mock_tray, "tray")
+        order.attach_mock(mock_tray_instance.run, "run")
+
+        with (
+            patch("vocalinux.main.parse_arguments") as mock_parse,
+            patch("gi.repository.GLib.idle_add") as mock_idle_add,
+        ):
+            order.attach_mock(mock_idle_add, "idle_add")
+            mock_args = MagicMock()
+            mock_args.debug = False
+            mock_args.model = "medium"
+            mock_args.engine = "whisper_cpp"
+            mock_args.language = "en-us"
+            mock_args.wayland = False
+            mock_parse.return_value = mock_args
+
+            main()
+
+        mock_speech.assert_called_once()
+        self.assertIs(mock_speech.call_args.kwargs["defer_model_load"], True)
+        mock_idle_add.assert_called_once_with(mock_speech_instance.load_model_in_background)
+        self.assertEqual([c[0] for c in order.mock_calls], ["tray", "idle_add", "run"])
+
     @patch("vocalinux.main.check_dependencies")
     @patch("vocalinux.speech_recognition.recognition_manager.SpeechRecognitionManager")
     @patch("vocalinux.text_injection.text_injector.TextInjector")

tests/test_model_load_in_background.py

@@ -0,0 +1,185 @@
+"""Startup model load runs on a background thread so the tray can show first (#50).
+
+Seams covered here, all at the SpeechRecognitionManager public surface:
+the deferred-load constructor flag, ``load_model_in_background`` (thread,
+loading flag, callbacks, failure reporting), the dictation guard while the
+model is loading, supersession by an auto-pause, and the language-shortcut
+path while loading.
+"""
+
+import threading
+from unittest.mock import patch
+
+import pytest
+
+MANAGER_MODULE = "vocalinux.speech_recognition.recognition_manager"
+
+
+@pytest.fixture
+def init_whispercpp():
+    """Stand in for the whisper.cpp load so no model file is read."""
+    from vocalinux.speech_recognition.recognition_manager import SpeechRecognitionManager
+
+    with patch.object(SpeechRecognitionManager, "_init_whispercpp") as init:
+        yield init
+
+
+def _deferred_manager():
+    from vocalinux.speech_recognition.recognition_manager import SpeechRecognitionManager
+
+    return SpeechRecognitionManager(
+        engine="whisper_cpp",
+        model_size="small",
+        language="en-us",
+        defer_model_load=True,
+    )
+
+
+class TestDeferredConstruction:
+    def test_construction_does_not_load_the_engine(self, init_whispercpp):
+        manager = _deferred_manager()
+
+        init_whispercpp.assert_not_called()
+        assert manager.model_loading is True
+        assert manager.model_ready is False
+
+    def test_default_construction_still_loads_synchronously(self, init_whispercpp):
+        from vocalinux.speech_recognition.recognition_manager import SpeechRecognitionManager
+
+        manager = SpeechRecognitionManager(
+            engine="whisper_cpp", model_size="small", language="en-us"
+        )
+
+        init_whispercpp.assert_called_once_with()
+        assert manager.model_loading is False
+
+    def test_unknown_engine_still_fails_at_construction(self, init_whispercpp):
+        from vocalinux.speech_recognition.recognition_manager import SpeechRecognitionManager
+
+        with pytest.raises(ValueError, match="Unsupported speech recognition engine"):
+            SpeechRecognitionManager(
+                engine="no-such-engine",
+                model_size="small",
+                language="en-us",
+                defer_model_load=True,
+            )
+
+
+class TestBackgroundLoad:
+    def test_load_runs_on_a_daemon_thread_and_returns_at_once(self, init_whispercpp):
+        entered = threading.Event()
+        release = threading.Event()
+        seen_threads = []
+
+        def slow_load(*_args, **_kwargs):
+            seen_threads.append(threading.current_thread())
+            entered.set()
+            release.wait(5)
+
+        init_whispercpp.side_effect = slow_load
+        manager = _deferred_manager()
+
+        manager.load_model_in_background()
+
+        assert entered.wait(5), "the load never started"
+        assert manager.model_loading is True
+        assert seen_threads[0] is not threading.main_thread()
+        assert seen_threads[0].daemon is True
+        release.set()
+        manager._model_load_thread.join(5)
+
+    def test_loading_flag_clears_and_callbacks_fire_after_success(self, init_whispercpp):
+        manager = _deferred_manager()
+        flags_seen_by_callback = []
+        manager.register_model_load_callback(
+            lambda: flags_seen_by_callback.append(manager.model_loading)
+        )
+
+        manager.load_model_in_background()
+        manager._model_load_thread.join(5)
+
+        init_whispercpp.assert_called_once_with()
+        assert manager.model_loading is False
+        assert flags_seen_by_callback, "the load-finished callback never ran"
+        assert flags_seen_by_callback[-1] is False
+
+    def test_failed_load_reports_error_instead_of_raising(self, init_whispercpp):
+        from vocalinux.common_types import RecognitionState
+
+        init_whispercpp.side_effect = RuntimeError("model file is corrupt")
+        manager = _deferred_manager()
+        states = []
+        manager.register_state_callback(states.append)
+
+        with patch(f"{MANAGER_MODULE}._show_notification") as notify:
+            manager.load_model_in_background()
+            manager._model_load_thread.join(5)
+
+        assert manager.model_loading is False
+        assert manager.state == RecognitionState.ERROR
+        assert states[-1] == RecognitionState.ERROR
+        notify.assert_called_once()
+        assert "failed to load" in notify.call_args.args[0].lower()
+
+    def test_auto_pause_already_active_at_startup_skips_the_load(self, init_whispercpp):
+        manager = _deferred_manager()
+        # A configured app is already running when the tray comes up. Its
+        # auto-pause runs before the load is scheduled, so the load must not
+        # keep the model resident; auto-resume loads it when the app exits.
+        manager._new_model_request(auto_pause=True)
+
+        manager.load_model_in_background()
+        manager._model_load_thread.join(5)
+
+        init_whispercpp.assert_not_called()
+        assert manager.is_auto_paused is True
+        assert manager.model_loading is False
+
+    def test_auto_pause_arriving_before_the_load_supersedes_it(self, init_whispercpp):
+        manager = _deferred_manager()
+
+        # The loader must wait for the model lock; an auto-pause takes a newer
+        # request while the lock is still held, so the load must not run.
+        with manager._model_lock:
+            manager.load_model_in_background()
+            manager._new_model_request(auto_pause=True)
+        manager._model_load_thread.join(5)
+
+        init_whispercpp.assert_not_called()
+        assert manager.is_auto_paused is True
+        assert manager.model_loading is False
+
+
+class TestDictationWhileLoading:
+    def test_start_is_refused_with_a_loading_notice(self, init_whispercpp):
+        from vocalinux.common_types import RecognitionState
+
+        manager = _deferred_manager()
+
+        with (
+            patch(f"{MANAGER_MODULE}._show_notification") as notify,
+            patch(f"{MANAGER_MODULE}.play_error_sound") as error_sound,
+            patch.object(manager, "_notify_model_missing") as model_missing_offer,
+        ):
+            started = manager.start_recognition()
+
+        assert started is False
+        assert manager.state == RecognitionState.IDLE
+        model_missing_offer.assert_not_called()
+        error_sound.assert_called_once()
+        assert "loading" in notify.call_args.args[0].lower()
+        init_whispercpp.assert_not_called()
+
+    def test_language_shortcut_during_load_reports_loading_not_unsupported(self, init_whispercpp):
+        manager = _deferred_manager()
+
+        with (
+            patch.object(manager, "_refuse_language_override") as refuse,
+            patch(f"{MANAGER_MODULE}._show_notification") as notify,
+            patch(f"{MANAGER_MODULE}.play_error_sound"),
+        ):
+            started = manager.start_recognition_with_language("de")
+
+        assert started is False
+        refuse.assert_not_called()
+        assert "loading" in notify.call_args.args[0].lower()

tests/test_tray_model_loading.py

@@ -0,0 +1,109 @@
+"""Tray rendering while the speech model loads in the background (#50).
+
+The tray is built before the model load starts, so it must show a loading
+state, re-render on the GTK main loop when the load finishes, and keep the
+idle keep-alive from unloading a model that is still being loaded.
+"""
+
+from unittest.mock import MagicMock, patch
+
+from vocalinux.common_types import RecognitionState
+from vocalinux.ui import tray_indicator
+
+LOADING_TITLE = "Loading model…"
+ICONS = {
+    "default": "default-icon",
+    "active": "active-icon",
+    "processing": "processing-icon",
+}
+
+
+def _engine(state=RecognitionState.IDLE, loading=False):
+    engine = MagicMock()
+    engine.state = state
+    engine.model_loading = loading
+    engine.is_auto_paused = False
+    return engine
+
+
+def _make_tray(engine):
+    """Build a TrayIndicator without running GTK setup; only rendering helpers are real."""
+    tray = tray_indicator.TrayIndicator.__new__(tray_indicator.TrayIndicator)
+    tray.speech_engine = engine
+    tray.indicator = MagicMock()
+    tray._icon_keys = dict(ICONS)
+    tray._indicator_title = tray_indicator.TRAY_TITLE
+    tray._language_shortcuts_refresh_pending = False
+    tray.overlay = None
+    tray.shortcut_manager = MagicMock()
+    tray._model_keepalive = MagicMock()
+    tray._auto_pause_monitor = MagicMock(paused=False)
+    tray._set_indicator_icon = MagicMock()
+    tray._set_indicator_title = MagicMock()
+    tray._set_menu_item_enabled = MagicMock()
+    tray._set_cancel_menu_item_visible = MagicMock()
+    return tray
+
+
+def test_loading_shows_processing_icon_and_loading_title_while_idle():
+    tray = _make_tray(_engine(loading=True))
+
+    tray._update_ui(RecognitionState.IDLE)
+
+    tray._set_indicator_icon.assert_called_with(ICONS["processing"], "Loading speech model")
+    tray._set_indicator_title.assert_called_with(LOADING_TITLE)
+    tray._set_menu_item_enabled.assert_any_call("Start Voice Typing", False)
+    tray._set_menu_item_enabled.assert_any_call("Stop Voice Typing", False)
+
+
+def test_finished_load_returns_to_idle_icon_and_enables_start():
+    tray = _make_tray(_engine(loading=False))
+
+    tray._update_ui(RecognitionState.IDLE)
+
+    tray._set_indicator_icon.assert_called_with(ICONS["default"], "Microphone off")
+    tray._set_menu_item_enabled.assert_any_call("Start Voice Typing", True)
+
+
+def test_loading_state_does_not_mask_an_active_session():
+    tray = _make_tray(_engine(state=RecognitionState.LISTENING, loading=True))
+
+    tray._update_ui(RecognitionState.LISTENING)
+
+    tray._set_indicator_icon.assert_called_with(ICONS["active"], "Microphone on")
+
+
+def test_load_change_is_marshalled_onto_the_gtk_main_loop():
+    tray = _make_tray(_engine(loading=False))
+
+    with patch.object(tray_indicator, "GLib") as glib:
+        tray._on_model_load_changed()
+
+    glib.idle_add.assert_called_once_with(tray._apply_model_load_state)
+    tray._set_indicator_icon.assert_not_called()
+
+
+def test_finishing_the_load_restarts_the_keep_alive_timer():
+    engine = _engine(loading=False)
+    tray = _make_tray(engine)
+
+    assert tray._apply_model_load_state() is False
+
+    tray._model_keepalive.bump.assert_called_once_with()
+    tray._set_indicator_icon.assert_called_with(ICONS["default"], "Microphone off")
+
+
+def test_loading_in_progress_does_not_restart_the_keep_alive_timer():
+    tray = _make_tray(_engine(loading=True))
+
+    tray._apply_model_load_state()
+
+    tray._model_keepalive.bump.assert_not_called()
+
+
+def test_keep_alive_never_unloads_while_the_model_is_loading():
+    loading = _make_tray(_engine(loading=True))
+    loaded = _make_tray(_engine(loading=False))
+
+    assert loading._is_safe_for_keepalive_unload() is False
+    assert loaded._is_safe_for_keepalive_unload() is True