feat(voice): add duplex streaming v2 pipeline

tests/test_ai_orchestrator_service.py

@@ -33,6 +33,7 @@ from services.shared.sql_models import (
     VoiceNameCollectionSettingsRow,
     VoiceTTSSettingsRow,
     VoiceAISessionRow,
+    VoiceTranscriptSegmentRow,
     WhatsAppThreadRow,
 )
 from services.telegram_adapter_service import app as telegram_module
@@ -1419,6 +1420,91 @@ def test_voice_v2_off_domain_request_returns_fast_operator_fallback_without_llm(
     assert decision["metadata"]["response_plan_id"] == "rsp_off_domain"
 
 
+def test_voice_v2_streaming_duplex_early_plan_returns_fast_safe_reply_without_llm(monkeypatch):
+    monkeypatch.setenv("AI_VOICE_POLICY_MODE", "v2_streaming_duplex")
+
+    def _unexpected_llm(messages):
+        raise AssertionError(f"LLM should not be called for early plan: {messages!r}")
+
+    monkeypatch.setattr(ai_module, "_request_structured_model_decision", _unexpected_llm)
+
+    decision = voice_module._voice_decision(
+        language="ru",
+        customer=None,
+        interaction=SimpleNamespace(interaction_id="int_voice_early_plan", status="new", queue_id="que_voice", subject="unknown"),
+        transcript_text="Расскажи, как устроен кондиционер",
+        transcript_window=[],
+        kb_results=[],
+        disclosure_required=False,
+        request_metadata={
+            "voice_v2_enabled": True,
+            "reply_phase": "early_plan",
+            "response_plan_id": "rsp_early",
+        },
+    )
+
+    assert decision["model"] == "voice_early_plan_off_domain"
+    assert decision["metadata"]["reply_phase"] == "early_plan"
+    assert decision["metadata"]["voice_v2_enabled"] is True
+    assert decision["metadata"]["response_plan_id"] == "rsp_early"
+    assert "кондиционер" not in decision["reply_text"].lower()
+    assert "оператор" in decision["reply_text"].lower()
+
+
+def test_turn_voice_session_early_plan_does_not_persist_partial_turns(monkeypatch):
+    monkeypatch.setenv("AI_VOICE_POLICY_MODE", "v2_streaming_duplex")
+
+    def _unexpected_llm(messages):
+        raise AssertionError(f"LLM should not be called for early plan: {messages!r}")
+
+    monkeypatch.setattr(ai_module, "_request_structured_model_decision", _unexpected_llm)
+
+    seeded = seed_voice_downstream_session(
+        marker=f"voice_early_plan_{new_id('seed')}",
+        name_status="name_not_obtained",
+        customer_display_name="+77010009999",
+    )
+
+    decision = voice_module.turn_voice_session(
+        seeded["session_id"],
+        VoiceAITurnIn(
+            voice_session_id=seeded["session_id"],
+            call_id=seeded["call_id"],
+            interaction_id=seeded["interaction_id"],
+            transcript_text="Расскажи, как устроен кондиционер",
+            language="ru",
+            sequence_no=1,
+            metadata={
+                "voice_v2_enabled": True,
+                "reply_phase": "early_plan",
+                "response_plan_id": "rsp_early_turn",
+            },
+        ),
+    )
+
+    assert decision.metadata["reply_phase"] == "early_plan"
+    assert decision.metadata["response_plan_id"] == "rsp_early_turn"
+    assert decision.status == "active"
+
+    session = get_session()
+    try:
+        voice_session = session.execute(
+            select(VoiceAISessionRow).where(VoiceAISessionRow.session_id == seeded["session_id"])
+        ).scalar_one()
+        ai_turns = session.execute(
+            select(AITurnRow).where(AITurnRow.interaction_id == seeded["interaction_id"])
+        ).scalars().all()
+        transcript_segments = session.execute(
+            select(VoiceTranscriptSegmentRow).where(VoiceTranscriptSegmentRow.session_id == seeded["session_id"])
+        ).scalars().all()
+        assert voice_session.ai_session_id is None
+        assert voice_session.status == "active"
+        assert ai_turns == []
+        assert transcript_segments == []
+    finally:
+        session.close()
+
+
 def test_ai_enqueue_creates_outbound_ai_reply_and_delivery_flow(monkeypatch):
     monkeypatch.setenv("AI_TELEGRAM_ENABLED", "1")
     monkeypatch.setenv("AI_PROVIDER", "stub")