Fix OpenClaw realtime transcript item identity

services/athena-realtime-voice/test_smoke.py

@@ -18,7 +18,7 @@ from av import AudioFrame
 from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PrivateKey
 from cryptography.hazmat.primitives.serialization import Encoding, PublicFormat
 
-from server import FRAME_SAMPLES, SAMPLE_RATE, create_app, decode_public_key_token, make_wav
+from server import FRAME_SAMPLES, SAMPLE_RATE, RealtimeSession, create_app, decode_public_key_token, make_wav
 
 
 class Microphone(MediaStreamTrack):
@@ -56,6 +56,36 @@ async def serve(app):
 
 
 class VoiceSmokeTest(unittest.IsolatedAsyncioTestCase):
+    async def test_transcript_item_announced_before_completion(self):
+        async def stt(_):
+            return web.json_response({"text": "Hallo Athena"})
+
+        fake = web.Application()
+        fake.router.add_post("/audio/transcriptions", stt)
+        runner, base = await serve(fake)
+        os.environ["ATHENA_API_BASE_URL"] = base
+        events = []
+
+        class Channel:
+            readyState = "open"
+
+            def send(self, message):
+                events.append(json.loads(message))
+
+        try:
+            async with ClientSession() as http:
+                session = RealtimeSession(None, http, None)
+                session.channel = Channel()
+                await session.transcribe_and_consult(b"\0\0" * SAMPLE_RATE)
+            committed = next((index, event) for index, event in enumerate(events)
+                             if event.get("type") == "input_audio_buffer.committed")
+            completed = next((index, event) for index, event in enumerate(events)
+                             if event.get("type") == "conversation.item.input_audio_transcription.completed")
+            self.assertLess(committed[0], completed[0])
+            self.assertEqual(committed[1]["item_id"], completed[1]["item_id"])
+        finally:
+            await runner.cleanup()
+
     async def test_public_key_session_token(self):
         private = Ed25519PrivateKey.generate()
         pem = private.public_key().public_bytes(Encoding.PEM, PublicFormat.SubjectPublicKeyInfo)
@@ -163,6 +193,12 @@ class VoiceSmokeTest(unittest.IsolatedAsyncioTestCase):
             self.assertTrue(any(event.get("type") ==
                                 "conversation.item.input_audio_transcription.completed"
                                 for event in events))
+            committed = next((index, event) for index, event in enumerate(events)
+                             if event.get("type") == "input_audio_buffer.committed")
+            completed = next((index, event) for index, event in enumerate(events)
+                             if event.get("type") == "conversation.item.input_audio_transcription.completed")
+            self.assertLess(committed[0], completed[0])
+            self.assertEqual(committed[1]["item_id"], completed[1]["item_id"])
             self.assertTrue(any(event.get("type") ==
                                 "response.output_audio_transcript.done" for event in events))
         finally: