feat: instructions options translated automatically by --instructions

1 files changed, 56 insertions, 0 deletions

diff --git a/app/tests/test_tts.py b/app/tests/test_tts.py
index 7170262..c0ba666 100644
--- a/app/tests/test_tts.py
+++ b/app/tests/test_tts.py
@@ -1042,6 +1042,14 @@ class AudioCppFamilyDetectionTests(unittest.TestCase):
         self.assertIs(client.profile, AUDIOCPP_DEFAULT_FAMILY_PROFILE)
         self.assertEqual(client.profile.language_style, AUDIOCPP_LANG_OMIT)
 
+    def test_breeze_tts_uses_the_option_instruction_channel(self):
+        # BreezeTTS 2 reads its instruction from the request options, not
+        # the OpenAI-style "instructions" field.
+        profile = AUDIOCPP_FAMILY_PROFILES["breeze_tts"]
+        self.assertEqual(profile.instruction_channel,
+                         audiocpp_client.AUDIOCPP_INSTRUCTION_OPTION)
+        self.assertEqual(profile.language_style, AUDIOCPP_LANG_OMIT)
+
     def test_speech_to_speech_only_family_rejected_at_connect(self):
         # A family whose model spec has no text-synthesis task
         # (PersonaPlex, s2s-only) cannot narrate regardless of its hosted
@@ -1612,6 +1620,54 @@ class AudioCppTTSClientRequestTests(unittest.TestCase):
         payload = json.loads(mock_urlopen.call_args[0][0].data.decode("utf-8"))
         self.assertEqual(payload["instructions"], "Read whisper quiet.")
 
+    def test_option_channel_family_sends_the_instruction_as_a_request_option(self):
+        # BreezeTTS 2 ignores the top-level OpenAI "instructions" field;
+        # the instruction goes inside "options" for those families.
+        client = self._make_client(
+            family="breeze_tts", preset_mode=True, voice="narrator",
+            instructions="Speak slowly with a restrained tone.")
+        with patch("converter.clients.faster.urllib.request.urlopen",
+                   return_value=self._post_response(self._wav_bytes())) as mock_urlopen:
+            client._request_wav("Hello.")
+        payload = json.loads(mock_urlopen.call_args[0][0].data.decode("utf-8"))
+        self.assertNotIn("instructions", payload)
+        self.assertEqual(payload["voice"], "narrator")
+        self.assertEqual(payload["options"],
+                         {"instruction": "Speak slowly with a restrained tone."})
+
+    def test_option_channel_family_keeps_other_request_options(self):
+        client = self._make_client(
+            family="breeze_tts", instructions="Calm and steady.",
+            request_options={"guidance_scale": "2.5"})
+        with patch("converter.clients.faster.urllib.request.urlopen",
+                   return_value=self._post_response(self._wav_bytes())) as mock_urlopen:
+            client._request_wav("Hello.")
+        payload = json.loads(mock_urlopen.call_args[0][0].data.decode("utf-8"))
+        self.assertEqual(payload["options"],
+                         {"guidance_scale": "2.5", "instruction": "Calm and steady."})
+
+    def test_option_channel_family_respects_the_callers_instruction_option(self):
+        # An explicit --option instruction=... is the caller's instruction
+        # and is not overwritten by the --instructions value.
+        client = self._make_client(
+            family="breeze_tts", instructions="from the flag",
+            request_options={"instruction": "from the option"})
+        with patch("converter.clients.faster.urllib.request.urlopen",
+                   return_value=self._post_response(self._wav_bytes())) as mock_urlopen:
+            client._request_wav("Hello.")
+        payload = json.loads(mock_urlopen.call_args[0][0].data.decode("utf-8"))
+        self.assertEqual(payload["options"]["instruction"], "from the option")
+
+    def test_option_channel_family_omits_options_when_no_instruction(self):
+        client = self._make_client(family="breeze_tts", preset_mode=True,
+                                   voice="narrator")
+        with patch("converter.clients.faster.urllib.request.urlopen",
+                   return_value=self._post_response(self._wav_bytes())) as mock_urlopen:
+            client._request_wav("Hello.")
+        payload = json.loads(mock_urlopen.call_args[0][0].data.decode("utf-8"))
+        self.assertNotIn("options", payload)
+        self.assertNotIn("instructions", payload)
+
     def test_preset_mode_sends_instructions_alongside_voice(self):
         # Clone + style control: both the server-side voice and the
         # instruction reach the model.