feat: better tui menu option gating for models that support custom voices (qwen) and models that do not support instructions

1 files changed, 74 insertions, 21 deletions

diff --git a/app/tests/test_hub.py b/app/tests/test_hub.py
index 9842148..c5ca346 100644
--- a/app/tests/test_hub.py
+++ b/app/tests/test_hub.py
@@ -750,40 +750,75 @@ class ConvertFlowTests(unittest.TestCase):
                          [("audio.cpp [remote]", "audiocpp-remote")])
         # The model menu was fed from the live query (label, id).
         self.assertEqual(self._field("model_id")["choices"],
-                         [("higgs  (higgs_audio_tts, tts)", "higgs")])
+                         [("higgs  (higgs_audio_tts, clone)", "higgs")])
 
-    def test_audiocpp_qwen3_tts_voice_choices_lead_with_builtin_speaker(self):
+    def test_audiocpp_customvoice_entry_lists_builtin_speakers(self):
+        # A CustomVoice entry populates the Voice menu with the Qwen3-TTS
+        # built-in speakers and maps the pick to --speaker.
         self._patch_remote(
-            [{"id": "qwen", "family": "qwen3_tts", "task": "tts"}],
-            voices=["narrator"])
+            [{"id": "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF",
+              "family": "qwen3_tts", "task": "tts"}])
         with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
-            self._answer_form(backend="audiocpp-remote", model_id="qwen",
-                              audiocpp_voice="(built-in speaker)",
-                              instructions="")
+            self._answer_form(
+                backend="audiocpp-remote",
+                model_id="Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF",
+                audiocpp_voice="Ryan", instructions="")
             cmd = self._convert(
                 None, [self._remote("audiocpp", "audio.cpp")])
-        # The sentinel maps to "no voice" (built-in speaker).
+        # The speaker is passed as --speaker, not --voice.
         self.assertIsNone(cmd[2]["voice"])
+        self.assertEqual(cmd[2]["speaker"], "Ryan")
         fields = self.tui.forms_seen[0][1]
         voice_field = self._field("audiocpp_voice")
-        choices = voice_field["choices"](fields)
-        self.assertEqual(choices,
-                         [("(built-in speaker)", "(built-in speaker)"),
-                          ("narrator", "narrator")])
-
-    def test_audiocpp_remote_missing_family_treated_as_qwen3_tts(self):
-        # Legacy servers omit family/task; the converter defaults them to
-        # qwen3_tts/tts and so must the form (voice optional).
+        self.assertEqual(voice_field["choices"](fields),
+                         [(s, s) for s in hub.QWEN3_TTS_SPEAKERS])
+        # CustomVoice reads a style instruction, so the field stays visible.
+        instr = self._field("instructions")
+        self.assertTrue(instr["visible"](fields))
+        self.assertIsNone(instr["validate"](""))
+
+    def test_audiocpp_qwen3_tts_base_entry_lists_clone_voices(self):
+        # A Base entry populates the Voice menu with the server's clone
+        # voices only (no built-in speakers) and maps the pick to --voice.
+        self._patch_remote(
+            [{"id": "Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+              "family": "qwen3_tts", "task": "tts"}],
+            voices=["narrator"])
+        with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
+            self._answer_form(
+                backend="audiocpp-remote",
+                model_id="Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                audiocpp_voice="narrator", instructions="")
+            cmd = self._convert(
+                None, [self._remote("audiocpp", "audio.cpp")])
+        self.assertEqual(cmd[2]["voice"], "narrator")
+        self.assertIsNone(cmd[2]["speaker"])
+        self.assertIsNone(cmd[2]["instructions"])
+        fields = self.tui.forms_seen[0][1]
+        voice_field = self._field("audiocpp_voice")
+        self.assertEqual(voice_field["choices"](fields),
+                         [("narrator", "narrator")])
+        # Base (clone) ignores instructions, so the field is hidden.
+        instr = self._field("instructions")
+        self.assertFalse(instr["visible"](fields))
+
+    def test_audiocpp_remote_missing_family_is_clone_capable(self):
+        # A missing family is unknown — not guessed as qwen3_tts — so the
+        # entry is clone-only: it needs a --voice rather than offering a
+        # built-in speaker.
         self._patch_remote([{"id": "legacy", "family": "", "task": ""}],
                            voices=[])
         with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
             self._answer_form(backend="audiocpp-remote", model_id="legacy",
-                              audiocpp_voice="(built-in speaker)",
-                              instructions="")
+                              audiocpp_voice="", instructions="")
             cmd = self._convert(
                 None, [self._remote("audiocpp", "audio.cpp")])
         self.assertIsNotNone(cmd)
         self.assertIsNone(cmd[2]["voice"])
+        self.assertIsNone(cmd[2]["speaker"])
+        voice_field = self._field("audiocpp_voice")
+        # Clone-only: an empty voice is refused (no built-in speaker option).
+        self.assertIsNotNone(voice_field["validate"](""))
 
     def test_audiocpp_vdes_hides_voice_and_requires_instructions(self):
         self._patch_remote(
@@ -800,9 +835,24 @@ class ConvertFlowTests(unittest.TestCase):
         voice_field = self._field("audiocpp_voice")
         self.assertFalse(voice_field["visible"](fields))
         instr = self._field("instructions")
+        self.assertTrue(instr["visible"](fields))
         self.assertIsNotNone(instr["validate"](""))
         self.assertIsNone(instr["validate"]("describe me"))
 
+    def test_audiocpp_clone_drops_stale_instructions(self):
+        # A Base/clone entry ignores instructions: even if the form held a
+        # leftover value, the mapper must not send it to the model.
+        self._patch_remote(
+            [{"id": "higgs", "family": "higgs_audio_tts", "task": "tts"}],
+            voices=["narrator"])
+        with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
+            self._answer_form(backend="audiocpp-remote", model_id="higgs",
+                              audiocpp_voice="narrator",
+                              instructions="stale description")
+            cmd = self._convert(
+                None, [self._remote("audiocpp", "audio.cpp")])
+        self.assertIsNone(cmd[2]["instructions"])
+
     def test_audiocpp_required_voice_validates(self):
         # A non-qwen3_tts family needs a --voice; a blank value refuses.
         self._patch_remote(
@@ -1082,9 +1132,11 @@ class ConvertFlowTests(unittest.TestCase):
              "mode", "speaker", "clone", "output_format", "speed",
              "single_file", "debug"])
         # The form opens on the configured default (audio.cpp): its fields
-        # show, the other backend's hide.
-        for key in ("model_id", "audiocpp_voice", "instructions"):
+        # show, the other backend's hide. (Instructions is hidden too: the
+        # default higgs entry is clone-only, which ignores instructions.)
+        for key in ("model_id", "audiocpp_voice"):
             self.assertTrue(self._field(key)["visible"](fields))
+        self.assertFalse(self._field("instructions")["visible"](fields))
         for key in ("mode", "speaker", "clone"):
             self.assertFalse(self._field(key)["visible"](fields))
         # Picking qwen in the Backend field swaps which options show.
@@ -1100,8 +1152,9 @@ class ConvertFlowTests(unittest.TestCase):
             self.assertFalse(self._field(key)["visible"](fields))
         # And back to audio.cpp.
         fields[0]["value"] = "audiocpp"
-        for key in ("model_id", "audiocpp_voice", "instructions"):
+        for key in ("model_id", "audiocpp_voice"):
             self.assertTrue(self._field(key)["visible"](fields))
+        self.assertFalse(self._field("instructions")["visible"](fields))
         for key in ("mode", "speaker", "clone"):
             self.assertFalse(self._field(key)["visible"](fields))