fix(cli): surface local STT model preparation

cli.py

@@ -11826,14 +11826,37 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
         threading.Thread(target=_refresh_level, daemon=True).start()
 
     def _voice_stt_model(self) -> Optional[str]:
-        """STT model override from config, or None for the provider default."""
+        """STT model override from config, or None for the provider default.
+
+        For the local provider, prefer stt.local.model (default ``base``) so the
+        CLI passes a real model name into the local STT backend.
+        """
         try:
             from hermes_cli.config import load_config
             stt_config = load_config().get("stt", {})
-            return stt_config.get("model") if isinstance(stt_config, dict) else None
+            if not isinstance(stt_config, dict):
+                return None
+            provider = str(stt_config.get("provider") or "").strip().lower()
+            if provider == "local":
+                local_config = stt_config.get("local") or {}
+                if not isinstance(local_config, dict):
+                    local_config = {}
+                return local_config.get("model") or "base"
+            return stt_config.get("model")
         except Exception:
             return None
 
+    def _voice_stt_provider(self) -> str:
+        """Configured STT provider name (lowercased), or empty string."""
+        try:
+            from hermes_cli.config import load_config
+            stt_config = load_config().get("stt", {})
+            if not isinstance(stt_config, dict):
+                return ""
+            return str(stt_config.get("provider") or "").strip().lower()
+        except Exception:
+            return ""
+
     def _voice_restart_recording_async(self) -> None:
         """Restart continuous-mode recording off-thread (start() can block)."""
         def _restart_recording():
@@ -11880,10 +11903,18 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
             # _voice_processing is already True (set atomically above)
             if hasattr(self, '_app') and self._app:
                 self._app.invalidate()
-            _cprint(f"{_DIM}Transcribing...{_RST}")
+
+            stt_model = self._voice_stt_model()
+            if self._voice_stt_provider() == "local":
+                _cprint(
+                    f"{_DIM}Preparing local STT model '{stt_model}' "
+                    f"(first use may download it from Hugging Face)...{_RST}"
+                )
+            else:
+                _cprint(f"{_DIM}Transcribing...{_RST}")
 
             from tools.voice_mode import transcribe_recording
-            result = transcribe_recording(wav_path, model=self._voice_stt_model())
+            result = transcribe_recording(wav_path, model=stt_model)
 
             if result.get("success") and result.get("transcript", "").strip():
                 transcript = result["transcript"].strip()

tests/tools/test_voice_cli_integration.py

@@ -1299,6 +1299,56 @@ class TestVoiceStopAndTranscribeReal:
         cli._voice_stop_and_transcribe()
         cli._voice_start_recording.assert_not_called()
 
+    @pytest.mark.parametrize(
+        ("stt_config", "expected_model"),
+        [
+            ({"provider": "local", "model": "whisper-1", "local": {"model": "small"}}, "small"),
+            ({"provider": "local", "local": {"model": "tiny"}}, "tiny"),
+            ({"provider": "local", "model": "whisper-1"}, "base"),
+        ],
+    )
+    def test_local_stt_shows_model_download_status(self, stt_config, expected_model):
+        recorder = MagicMock()
+        recorder.stop.return_value = "/tmp/test.wav"
+        cli = _make_voice_cli(_voice_recording=True, _voice_recorder=recorder)
+
+        with patch("cli._cprint") as mock_print, \
+             patch("cli.os.path.isfile", return_value=False), \
+             patch("hermes_cli.config.load_config", return_value={"stt": stt_config}), \
+             patch("tools.voice_mode.transcribe_recording",
+                   return_value={"success": True, "transcript": "hello"}) as mock_transcribe, \
+             patch("tools.voice_mode.play_beep"):
+            cli._voice_stop_and_transcribe()
+
+        messages = [call.args[0] for call in mock_print.call_args_list]
+        assert any(
+            f"local STT model '{expected_model}'" in message
+            and "first use may download it from Hugging Face" in message
+            for message in messages
+        )
+        mock_transcribe.assert_called_once_with("/tmp/test.wav", model=expected_model)
+
+    def test_non_local_stt_keeps_generic_transcribing_status(self):
+        recorder = MagicMock()
+        recorder.stop.return_value = "/tmp/test.wav"
+        cli = _make_voice_cli(_voice_recording=True, _voice_recorder=recorder)
+
+        with patch("cli._cprint") as mock_print, \
+             patch("cli.os.path.isfile", return_value=False), \
+             patch(
+                 "hermes_cli.config.load_config",
+                 return_value={"stt": {"provider": "openai", "model": "whisper-1"}},
+             ), \
+             patch("tools.voice_mode.transcribe_recording",
+                   return_value={"success": True, "transcript": "hello"}) as mock_transcribe, \
+             patch("tools.voice_mode.play_beep"):
+            cli._voice_stop_and_transcribe()
+
+        messages = [call.args[0] for call in mock_print.call_args_list]
+        assert any("Transcribing..." in message for message in messages)
+        assert all("Hugging Face" not in message for message in messages)
+        mock_transcribe.assert_called_once_with("/tmp/test.wav", model="whisper-1")
+
     @patch("cli._cprint")
     @patch("cli.os.unlink")
     @patch("cli.os.path.isfile", return_value=True)