Revert "feat(voice): start early replies before final asr"

This reverts commit a44a1b97a1.
2026-04-19 01:30:26 +05:00
parent 224973d840
commit aa64b6b94c
2 changed files with 100 additions and 469 deletions
@@ -1981,142 +1981,6 @@ def test_media_runtime_voice_v2_uses_early_plan_before_slow_final_decision():
    assert events.index(("main", "early reply")) < events.index("final_done")


-def test_media_runtime_voice_v2_starts_early_main_before_slow_final_asr():
-    events: list[str | tuple[str, str]] = []
-
-    class _SlowStreamingProvider(StreamingASRProvider):
-        name = "slow-streaming"
-        supports_streaming = True
-
-        def finalize(self, stream_id: str) -> ASRTranscription:
-            assert stream_id == "stream-1"
-            events.append("finalize_start")
-            time.sleep(0.35)
-            events.append("finalize_done")
-            return ASRTranscription(text="work schedule", language="ru", confidence=0.9)
-
-        def close_stream(self, stream_id: str) -> None:
-            assert stream_id == "stream-1"
-
-    def _process_turn(session_id, transcript_text, language, barge_in, metadata):
-        del session_id, transcript_text, barge_in
-        if metadata and metadata.get("reply_phase") == "early_plan":
-            events.append("early_plan")
-            return VoiceAITurnDecisionOut(
-                language=language or "ru",
-                intent="schedule",
-                reply_text="early reply",
-                confidence=0.8,
-                needs_handoff=False,
-                handoff_reason=None,
-                case_action="keep_open",
-                kb_refs=[],
-                summary_text="early ready",
-                model="early",
-                latency_ms=1,
-                status="active",
-            )
-        events.append("final_turn")
-        return VoiceAITurnDecisionOut(
-            language=language or "ru",
-            intent="schedule",
-            reply_text="final reply",
-            confidence=0.9,
-            needs_handoff=False,
-            handoff_reason=None,
-            case_action="keep_open",
-            kb_refs=[],
-            summary_text="final ready",
-            model="final",
-            latency_ms=1,
-            status="active",
-        )
-
-    runtime = AudioSocketMediaRuntime(
-        enabled=True,
-        host="127.0.0.1",
-        port=0,
-        frame_ms=20,
-        idle_timeout_seconds=2.0,
-        registration_wait_timeout_seconds=0.5,
-        min_speech_ms=40,
-        trailing_silence_ms=40,
-        max_turn_ms=2000,
-        asr_provider=_StubASRProvider(),
-        streaming_asr_provider=_SlowStreamingProvider(),
-        tts_provider=_StubTTSProvider(),
-        load_registration_by_media_uuid=lambda value: None,
-        mark_media_connected=lambda session_id, value: None,
-        mark_media_ended=lambda session_id, reason: None,
-        touch_media_frame=lambda session_id: None,
-        set_state=lambda session_id, state, handoff_reason, metadata: None,
-        get_pending_greeting=lambda session_id: None,
-        mark_reply_delivered=lambda session_id, text, is_greeting: None,
-        plan_reply=lambda session_id, text, metadata, kind: None,
-        process_turn=_process_turn,
-        request_handoff=lambda session_id, customer_request_text, decision: None,
-        handle_media_error=lambda session_id, message, metadata: None,
-    )
-
-    async def _fake_speak_reply(
-        actor: MediaActor,
-        text: str,
-        *,
-        is_greeting: bool,
-        style_hints: dict[str, object] | None = None,
-        reply_phase: str | None = "main",
-    ) -> None:
-        del actor, is_greeting, style_hints
-        events.append((str(reply_phase), text))
-
-    runtime._speak_reply = _fake_speak_reply  # type: ignore[method-assign]
-
-    async def _scenario() -> None:
-        actor = MediaActor(
-            registration=MediaRegistration(
-                voice_session_id="avs_media_runtime_early_before_final_asr",
-                call_id="call_media_runtime_early_before_final_asr",
-                interaction_id="int_media_runtime_early_before_final_asr",
-                ai_session_id="ais_media_runtime_early_before_final_asr",
-                language="ru",
-                media_uuid=str(uuid.uuid4()),
-                queue_code="voice_lab_ai",
-                queue_id="que_voice_lab_ai",
-                agent_profile="voice_support",
-                voice_v2_enabled=True,
-                voice_v2_ack_mode="immediate_short",
-                voice_v2_streaming_tts=True,
-                voice_v2_partial_asr=True,
-                voice_v2_duplex=True,
-                voice_v2_streaming_asr_backend="local_sidecar",
-            ),
-            reader=asyncio.StreamReader(),
-            writer=None,  # type: ignore[arg-type]
-            vad=EnergyVAD(frame_ms=20, min_speech_ms=40, trailing_silence_ms=40, max_turn_ms=2000),
-            frame_ms=20,
-            frame_bytes=320,
-        )
-        actor.asr_streaming_enabled = True
-        actor.asr_stream_id = "stream-1"
-        actor.partial_transcript = "work schedule"
-        actor.stable_partial_transcript = "work schedule"
-        actor.partial_intent = "schedule"
-        actor.stable_partial_intent = "schedule"
-        pcm_frame = (1000).to_bytes(2, "little", signed=True) * 160
-        await runtime._process_utterance(actor, pcm_frame * 40, False)
-        for _ in range(80):
-            if "final_turn" in events:
-                break
-            await asyncio.sleep(0.01)
-
-    asyncio.run(_scenario())
-
-    assert "finalize_start" in events
-    assert ("main", "early reply") in events
-    assert events.index(("main", "early reply")) < events.index("finalize_done")
-    assert ("main", "final reply") not in events
-
-
 def test_media_runtime_voice_v2_uses_partial_as_final_when_streaming_finalize_fails():
    captured: list[tuple[str, dict | None]] = []