diff --git a/apps/api/app/services/orchestrator.py b/apps/api/app/services/orchestrator.py index 3741a48..436ef77 100644 --- a/apps/api/app/services/orchestrator.py +++ b/apps/api/app/services/orchestrator.py @@ -724,22 +724,10 @@ async def run_turn_stream( yield StreamEvent("token", {"text": guard.text}) latency_ms = int((time.perf_counter() - started) * 1000) - await record_llm_audit( - audit_hook, - session_id=ctx.session_id, - provider=str(stream_meta.get("provider") or engine.engine_mode), - model=str( - stream_meta.get("model") - or engine.default_model - or ENGINE_GATEWAY_DEFAULT_MODEL_SENTINEL - ), - tokens_in=_safe_int(stream_meta.get("tokens_in")), - tokens_out=_safe_int(stream_meta.get("tokens_out")), - cost_usd=_safe_float(stream_meta.get("cost_usd")), - inference_geo=_optional_str(stream_meta.get("inference_geo")), - latency_ms=latency_ms, - ) + # done을 먼저 내보내 사용자 체감 지연(글 표시·음성 준비)을 줄인다. + # 감사 기록(record_llm_audit)은 최선노력이며 예외를 삼키므로, done 뒤로 + # 미뤄도 사용자에게 보이는 결과는 바뀌지 않는다. yield StreamEvent( "done", { @@ -760,6 +748,22 @@ async def run_turn_stream( "output_error": output_error, }, ) + + await record_llm_audit( + audit_hook, + session_id=ctx.session_id, + provider=str(stream_meta.get("provider") or engine.engine_mode), + model=str( + stream_meta.get("model") + or engine.default_model + or ENGINE_GATEWAY_DEFAULT_MODEL_SENTINEL + ), + tokens_in=_safe_int(stream_meta.get("tokens_in")), + tokens_out=_safe_int(stream_meta.get("tokens_out")), + cost_usd=_safe_float(stream_meta.get("cost_usd")), + inference_geo=_optional_str(stream_meta.get("inference_geo")), + latency_ms=latency_ms, + ) except EngineGatewaySseDecodeError as e: yield StreamEvent( "error", diff --git a/apps/api/app/test_orchestrator_masking.py b/apps/api/app/test_orchestrator_masking.py index 1c044a4..b46a5db 100644 --- a/apps/api/app/test_orchestrator_masking.py +++ b/apps/api/app/test_orchestrator_masking.py @@ -609,5 +609,108 @@ class OrchestratorMaskingGateTest(unittest.IsolatedAsyncioTestCase): self.assertEqual(events[-1].data["output_error"], "client_reply_quality_retryable") +class StreamAuditAfterDoneTest(unittest.IsolatedAsyncioTestCase): + """USER-LATENCY: done 이벤트가 감사 저장보다 먼저 사용자에게 전달돼야 한다.""" + + def setUp(self) -> None: + self.client_affect_provider_patch = patch.object( + orchestrator.settings, + "client_affect_provider", + "legacy", + ) + self.client_affect_provider_patch.start() + self.addCleanup(self.client_affect_provider_patch.stop) + self.presidio_patch = patch.object( + guardrail, + "_try_load_presidio", + return_value=(None, None), + ) + self.presidio_patch.start() + guardrail.set_ko_pii_recognizer(None) + self.addCleanup(self.presidio_patch.stop) + self.addCleanup(guardrail.set_ko_pii_recognizer, None) + + async def test_done_event_is_observed_before_audit_hook_is_called(self) -> None: + ctx = _prepare_context() + engine = CaptureStreamEngine() + audit_payloads: list[dict[str, Any]] = [] + + async def audit_hook(payload: dict[str, Any]) -> None: + audit_payloads.append(payload) + + done_seen_hook_call_count: int | None = None + done_event: Any = None + async for event in orchestrator.run_turn_stream( + ctx, + engine, # type: ignore[arg-type] + audit_hook=audit_hook, + ): + if event.event == "done": + # 소비자가 done을 받는 바로 이 시점에 감사 hook은 아직 호출되지 + # 않아야 한다(지연의 핵심: done이 사용자에게 먼저 전달된다). + done_seen_hook_call_count = len(audit_payloads) + done_event = event + + self.assertIsNotNone(done_event) + self.assertEqual(done_seen_hook_call_count, 0) + self.assertEqual(len(audit_payloads), 1) + + async def test_audit_hook_called_once_with_same_payload_as_before(self) -> None: + ctx = _prepare_context() + engine = CaptureStreamEngine() + audit_payloads: list[dict[str, Any]] = [] + + async def audit_hook(payload: dict[str, Any]) -> None: + audit_payloads.append(payload) + + events = [ + event + async for event in orchestrator.run_turn_stream( + ctx, + engine, # type: ignore[arg-type] + audit_hook=audit_hook, + ) + ] + + done_event = events[-1] + self.assertEqual(done_event.event, "done") + self.assertEqual(len(audit_payloads), 1) + payload = audit_payloads[0] + self.assertEqual(payload["session_id"], ctx.session_id) + self.assertEqual(payload["provider"], "fake-provider") + self.assertEqual(payload["model"], "fake-model") + self.assertEqual(payload["tokens_in"], 5) + self.assertEqual(payload["tokens_out"], 6) + self.assertEqual(payload["cost_usd"], 0.0) + self.assertIsInstance(payload["latency_ms"], int) + self.assertGreaterEqual(payload["latency_ms"], 0) + self.assertEqual(payload["provider"], done_event.data["llm_provider"]) + self.assertEqual(payload["model"], done_event.data["model"]) + self.assertEqual(payload["tokens_in"], done_event.data["tokens_in"]) + self.assertEqual(payload["tokens_out"], done_event.data["tokens_out"]) + self.assertEqual(payload["cost_usd"], done_event.data["cost_usd"]) + + async def test_audit_hook_exception_does_not_break_the_stream(self) -> None: + ctx = _prepare_context() + engine = CaptureStreamEngine() + hook_calls: list[dict[str, Any]] = [] + + async def failing_audit_hook(payload: dict[str, Any]) -> None: + hook_calls.append(payload) + raise RuntimeError("audit sink unavailable") + + events = [ + event + async for event in orchestrator.run_turn_stream( + ctx, + engine, # type: ignore[arg-type] + audit_hook=failing_audit_hook, + ) + ] + + self.assertEqual([event.event for event in events], ["token", "done"]) + self.assertEqual(len(hook_calls), 1) + + if __name__ == "__main__": unittest.main()