Reliability & perf: atomic flags, WebSocket reconnect, zero-alloc eye curves, audio queue read-offset

- Make bAgentGenerating, bWaitingForAgentResponse, bWaitingForResponse,
  bFirstAudioResponseLogged, bAgentResponseStartedFired (std::atomic<bool>)
  and LastInterruptEventId (std::atomic<int32>) for thread-safety
- Add WebSocket auto-reconnection with exponential backoff (1s→30s cap,
  max 5 attempts), distinguishing intentional vs unexpected disconnects
- Add FillCurrentEyeCurves() zero-allocation method using FindOrAdd()
  to eliminate per-frame TMap heap allocation in anim thread
- Replace AudioQueue RemoveAt(0,N) with read-offset pattern — O(1) per
  underflow callback, periodic compaction when offset > half buffer

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
j.foucher 2026-03-03 12:19:45 +01:00
parent 8886e7a7a2
commit 35d217f6ec
7 changed files with 167 additions and 375 deletions

View File

@ -1,359 +0,0 @@
seem to less stuck but timeout often : LogDebuggerCommands: Repeating last play command: Selected Viewport
LogPlayLevel: PlayLevel: No blueprints needed recompiling
LogPlayLevel: Creating play world package: /Game/UEDPIE_0_TestMap
LogPlayLevel: PIE: StaticDuplicateObject took: (0.005478s)
LogPlayLevel: PIE: Created PIE world by copying editor world from /Game/TestMap.TestMap to /Game/UEDPIE_0_TestMap.TestMap (0.005520s)
LogUObjectHash: Compacting FUObjectHashTables data took 0.64ms
LogChaosDD: Creating Chaos Debug Draw Scene for world TestMap
LogPlayLevel: PIE: World Init took: (0.001821s)
LogAudio: Display: Creating Audio Device: Id: 4, Scope: Unique, Realtime: True
LogAudioMixer: Display: Audio Mixer Platform Settings:
LogAudioMixer: Display: Sample Rate: 48000
LogAudioMixer: Display: Callback Buffer Frame Size Requested: 1024
LogAudioMixer: Display: Callback Buffer Frame Size To Use: 1024
LogAudioMixer: Display: Number of buffers to queue: 1
LogAudioMixer: Display: Max Channels (voices): 32
LogAudioMixer: Display: Number of Async Source Workers: 4
LogAudio: Display: AudioDevice MaxSources: 32
LogAudio: Display: Audio Spatialization Plugin: None (built-in).
LogAudio: Display: Audio Reverb Plugin: None (built-in).
LogAudio: Display: Audio Occlusion Plugin: None (built-in).
LogAudioMixer: Display: Initializing audio mixer using platform API: 'XAudio2'
LogAudioMixer: Display: Using Audio Hardware Device Speakers (Realtek(R) Audio)
LogAudioMixer: Display: Initializing Sound Submixes...
LogAudioMixer: Display: Creating Master Submix 'MasterSubmixDefault'
LogAudioMixer: Display: Creating Master Submix 'MasterReverbSubmixDefault'
LogAudioMixer: FMixerPlatformXAudio2::StartAudioStream() called. InstanceID=4
LogAudioMixer: Display: Output buffers initialized: Frames=1024, Channels=2, Samples=2048, InstanceID=4
LogAudioMixer: Display: Starting AudioMixerPlatformInterface::RunInternal(), InstanceID=4
LogAudioMixer: Display: FMixerPlatformXAudio2::SubmitBuffer() called for the first time. InstanceID=4
LogInit: FAudioDevice initialized with ID 4.
LogAudio: Display: Audio Device (ID: 4) registered with world 'TestMap'.
LogAudioMixer: Initializing Audio Bus Subsystem for audio device with ID 4
LogLoad: Game class is 'GameModeBase'
LogWorld: Bringing World /Game/UEDPIE_0_TestMap.TestMap up for play (max tick rate 60) at 2026.02.20-17.38.08
LogWorld: Bringing up level for play took: 0.000919
LogOnline: OSS: Created online subsystem instance for: :Context_3
LogElevenLabsWS: Connecting to ElevenLabs: wss://api.elevenlabs.io/v1/convai/conversation?agent_id=agent_5301kc1qkq49fn2av43nrbsar65k
PIE: Server logged in
PIE: Play in editor total start time 0,075 seconds.
LogElevenLabsWS: WebSocket connected. Sending conversation_initiation_client_data...
LogElevenLabsWS: Sending initiation: {
"type": "conversation_initiation_client_data",
"conversation_config_override":
{
"agent":
{
"turn":
{
"turn_timeout": 1
}
},
"tts":
{
"optimize_streaming_latency": 3
}
},
"custom_llm_extra_body":
{
"enable_intermediate_response": true
}
}
LogElevenLabsWS: Received message type: conversation_initiation_metadata
LogElevenLabsWS: [T+0.00s] Conversation initiated. ID=conv_2901khxymx7xed88dq8tw35b2pfk
LogElevenLabsAgent: [T+0.00s] Agent connected. ConversationID=conv_2901khxymx7xed88dq8tw35b2pfk
LogBlueprintUserMessages: [test_AI_Actor_C_2] Connected
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+2.18s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+2.18s] [Turn 1] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+4.60s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+4.60s] [Turn 1] Mic closed — user spoke 2.42s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: [T+5.66s] Agent started generating (1061 ms after turn end — includes VAD silence timeout + LLM start).
LogElevenLabsAgent: [T+5.66s] [Turn 1] Agent generating. (1.06s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Agent Start Generating
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+7.95s] [LATENCY] First audio: 3346 ms after turn end (3346 ms after last chunk)
LogElevenLabsAgent: [T+7.95s] [Turn 1] Agent speaking — first audio chunk. (3.35s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+7.95s] [LATENCY] User transcript: 3346 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Hello, how are you?
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+9.60s] [LATENCY] Agent text response: 4994 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Hello! I'm doing well, thank you for asking. How can I assist you today?
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+18.84s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+18.84s] [Turn 2] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+21.67s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+21.67s] [Turn 2] Mic closed — user spoke 2.83s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: [T+22.45s] Agent started generating (777 ms after turn end — includes VAD silence timeout + LLM start).
LogElevenLabsAgent: [T+22.45s] [Turn 2] Agent generating. (0.78s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Agent Start Generating
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+23.85s] [LATENCY] First audio: 2180 ms after turn end (2180 ms after last chunk)
LogElevenLabsAgent: [T+23.85s] [Turn 2] Agent speaking — first audio chunk. (2.18s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+23.90s] [LATENCY] User transcript: 2227 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Nothing special, just talking.
LogElevenLabsAgent: Warning: [Turn 2] Agent silence hard-timeout (2s) without agent_response — declaring agent stopped.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogElevenLabsWS: Received message type: audio
LogElevenLabsAgent: [T+28.55s] [Turn 2] Agent speaking — first audio chunk. (6.88s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+28.58s] [LATENCY] Agent text response: 6910 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : That's perfectly fine! I'm here to chat if you have any questions or just want to talk. What's on your mind?
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogEOSSDK: LogEOS: Updating Product SDK Config, Time: 347.575653
LogEOSSDK: LogEOS: SDK Config Product Update Request Completed - No Change
LogEOSSDK: LogEOS: ScheduleNextSDKConfigDataUpdate - Time: 347.774902, Update Interval: 340.359497
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+41.83s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+41.83s] [Turn 3] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+44.86s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+44.86s] [Turn 3] Mic closed — user spoke 3.03s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: [T+45.20s] Agent started generating (343 ms after turn end — includes VAD silence timeout + LLM start).
LogElevenLabsAgent: [T+45.20s] [Turn 3] Agent generating. (0.34s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Agent Start Generating
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+48.22s] [LATENCY] First audio: 3361 ms after turn end (3361 ms after last chunk)
LogElevenLabsAgent: [T+48.22s] [Turn 3] Agent speaking — first audio chunk. (3.36s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+48.22s] [LATENCY] User transcript: 3361 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : I wanna... Hi agent!
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+48.25s] [LATENCY] Agent text response: 3393 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Hello! It's good to hear from you. How can I help you today?
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+57.55s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+57.55s] [Turn 4] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+60.33s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+60.33s] [Turn 4] Mic closed — user spoke 2.78s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: [T+60.39s] Agent started generating (61 ms after turn end — includes VAD silence timeout + LLM start).
LogElevenLabsAgent: [T+60.39s] [Turn 4] Agent generating. (0.06s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Agent Start Generating
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+62.44s] [LATENCY] First audio: 2111 ms after turn end (2111 ms after last chunk)
LogElevenLabsAgent: [T+62.44s] [Turn 4] Agent speaking — first audio chunk. (2.11s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+62.44s] [LATENCY] User transcript: 2112 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Do you speak French?
LogElevenLabsAgent: Warning: [Turn 4] Agent silence hard-timeout (2s) without agent_response — declaring agent stopped.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogElevenLabsWS: Received message type: audio
LogElevenLabsAgent: [T+66.24s] [Turn 4] Agent speaking — first audio chunk. (5.91s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+66.27s] [LATENCY] Agent text response: 5944 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Yes, I can communicate in French. Would you like to ask me something in French or perhaps practice a bit?
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+74.14s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+74.14s] [Turn 5] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+76.88s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+76.88s] [Turn 5] Mic closed — user spoke 2.75s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: [T+77.28s] Agent started generating (393 ms after turn end — includes VAD silence timeout + LLM start).
LogElevenLabsAgent: [T+77.28s] [Turn 5] Agent generating. (0.39s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Agent Start Generating
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+78.31s] [LATENCY] First audio: 1428 ms after turn end (1428 ms after last chunk)
LogElevenLabsAgent: [T+78.31s] [Turn 5] Agent speaking — first audio chunk. (1.43s after turn end)
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Speaking
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+78.34s] [LATENCY] User transcript: 1460 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Yes, tell me something in French.
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Received message type: audio
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsAgent: StartListening: interrupting agent (speaking) to allow user to speak.
LogElevenLabsWS: Sending interrupt — ignoring incoming content until server acks.
LogElevenLabsAgent: [T+84.24s] [Turn 5] Agent stopped speaking (spoke 5.93s, full turn round-trip 7.36s).
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Speaking
LogElevenLabsWS: [T+84.24s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+84.24s] [Turn 6] Mic opened — user speaking.
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+84.43s] [LATENCY] Agent text response: 7549 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Bien sûr! Je peux vous dire: "Bonjour, comment allez-vous aujourd'hui?" This means "Hello, how are you today?"
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+88.56s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+88.56s] [Turn 6] Mic closed — user spoke 4.32s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+92.37s] [LATENCY] First audio: 3811 ms after turn end (3811 ms after last chunk)
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+92.37s] [LATENCY] User transcript: 3811 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Mais c'est super! Tu parles très bien!
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+97.23s] [LATENCY] Agent text response: 8677 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Merci beaucoup! Je suis ravie que vous le pensiez. Y a-t-il autre chose que vous aimeriez savoir ou discuter en français?
LogElevenLabsAgent: Warning: [T+98.56s] [Turn 6] Response timeout — server did not start generating after 10.0s. Firing OnAgentResponseTimeout.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+110.60s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+110.60s] [Turn 7] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+113.82s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+113.82s] [Turn 7] Mic closed — user spoke 3.22s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+119.37s] [LATENCY] First audio: 5546 ms after turn end (5546 ms after last chunk)
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+119.37s] [LATENCY] User transcript: 5546 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : And do you speak German?
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+119.40s] [LATENCY] Agent text response: 5577 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Yes, I can also communicate in German. Would you like me to say something in German for you?
LogElevenLabsAgent: Warning: [T+123.83s] [Turn 7] Response timeout — server did not start generating after 10.0s. Firing OnAgentResponseTimeout.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+131.53s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+131.53s] [Turn 8] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+134.42s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+134.42s] [Turn 8] Mic closed — user spoke 2.88s. Waiting for server response (timeout 10s)...
LogElevenLabsAgent: Warning: [T+144.42s] [Turn 8] Response timeout — server did not start generating after 10.0s. Firing OnAgentResponseTimeout.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+148.56s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+148.56s] [Turn 9] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+150.35s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+150.35s] [Turn 9] Mic closed — user spoke 1.79s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+151.56s] [LATENCY] First audio: 1211 ms after turn end (1211 ms after last chunk)
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+151.59s] [LATENCY] User transcript: 1245 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Hello.
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+152.87s] [LATENCY] Agent text response: 2526 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : Hello! How can I help you today?
LogElevenLabsAgent: Warning: [T+160.36s] [Turn 9] Response timeout — server did not start generating after 10.0s. Firing OnAgentResponseTimeout.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Start Listening
LogElevenLabsWS: [T+162.59s] User turn started — mic open, audio chunks will follow.
LogAudioCaptureCore: Display: WasapiCapture AudioFormat SampeRate: 48000, BitDepth: 32-Bit Floating Point
LogElevenLabsMic: Capture device: Microphone Array (Intel® Smart Sound Technology for Digital Microphones) | Rate=48000 | Channels=2
LogElevenLabsMic: Audio capture started.
LogElevenLabsAgent: [T+162.59s] [Turn 10] Mic opened — user speaking.
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Stop Listening
LogElevenLabsMic: Audio capture stopped.
LogElevenLabsWS: [T+165.69s] User turn ended — server VAD silence detection started (turn_timeout=1s).
LogElevenLabsAgent: [T+165.69s] [Turn 10] Mic closed — user spoke 3.10s. Waiting for server response (timeout 10s)...
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: agent_chat_response_part
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Warning: [T+168.47s] [LATENCY] First audio: 2778 ms after turn end (2778 ms after last chunk)
LogElevenLabsWS: Received message type: user_transcript
LogElevenLabsWS: Warning: [T+168.50s] [LATENCY] User transcript: 2810 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Transcript : Why are you so slow?
LogElevenLabsAgent: Warning: [T+175.70s] [Turn 10] Response timeout — server did not start generating after 10.0s. Firing OnAgentResponseTimeout.
LogElevenLabsWS: Received message type: audio
LogElevenLabsWS: Received message type: agent_response
LogElevenLabsWS: Warning: [T+175.85s] [LATENCY] Agent text response: 10161 ms after turn end
LogBlueprintUserMessages: [test_AI_Actor_C_2] [AI Agent] Text Response : I apologize if my response seemed slow. I'm designed to be as quick as possible while still being helpful. Is there something specific I can do to improve your experience?
LogSlate: Updating window title bar state: overlay mode, drag disabled, window buttons hidden, title bar hidden
LogWorld: BeginTearingDown for /Game/UEDPIE_0_TestMap
Show less
Excellent logs — je vois deux bugs distincts :
Bug A — bIgnoreIncomingContent jamais effacé : après l'interruption au Turn 5, le serveur n'envoie pas d'interruption_event (ou on ne le reçoit pas). Le flag reste true → tout le contenu du Turn 6, 7, 8, 9, 10 est silencieusement ignoré → timeout systématique. Preuve : les [LATENCY] First audio du WebSocketProxy s'affichent (l'audio arrive) mais JAMAIS le Agent speaking — first audio chunk de l'AgentComponent (l'audio est jeté).
Bug B — HardSilenceTimeoutTicks = 120 (2s) trop court : l'agent répondait en deux batches audio avec 4.7s de gap (Turn 2 : T+23.85s puis T+28.55s). Le timeout de 2s s'est déclenché pendant le gap → faux "Stop Speaking" → Blueprint ouvre le mic → puis le 2e batch arrive → "Start Speaking" à nouveau.
Fixes :

View File

@ -323,7 +323,7 @@ void FAnimNode_PS_AI_ConvAgent_Posture::Update_AnyThread(const FAnimationUpdateC
// Instead, keep the last valid cached values as a hold-over.
if (PostureComponent.IsValid())
{
CachedEyeCurves = PostureComponent->GetCurrentEyeCurves();
PostureComponent->FillCurrentEyeCurves(CachedEyeCurves);
CachedHeadRotation = PostureComponent->GetCurrentHeadRotation();
CachedHeadCompensation = PostureComponent->GetHeadAnimationCompensation();
CachedEyeCompensation = PostureComponent->GetEyeAnimationCompensation();

View File

@ -151,6 +151,35 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::TickComponent(float DeltaTime, ELevel
}
}
// ── Reconnection ────────────────────────────────────────────────────────
if (bWantsReconnect && FPlatformTime::Seconds() >= NextReconnectTime)
{
++ReconnectAttemptCount;
if (ReconnectAttemptCount > MaxReconnectAttempts)
{
// Give up — full disconnect cleanup.
bWantsReconnect = false;
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Error,
TEXT("Reconnection failed after %d attempts — giving up."), MaxReconnectAttempts);
bNetIsConversing = false;
ApplyConversationPosture();
NetConversatingPlayer = nullptr;
NetConversatingPawn = nullptr;
OnAgentDisconnected.Broadcast(1006, TEXT("Reconnection failed"));
}
else
{
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Log,
TEXT("Reconnecting... attempt %d/%d"), ReconnectAttemptCount, MaxReconnectAttempts);
// Ensure proxy is in a clean state for reconnection.
if (WebSocketProxy)
{
WebSocketProxy->Disconnect();
}
StartConversation_Internal();
}
}
// Silence detection.
// ISSUE-8: broadcast OnAgentStoppedSpeaking OUTSIDE AudioQueueLock.
// OnProceduralUnderflow (audio thread) also acquires AudioQueueLock — if we broadcast
@ -162,7 +191,7 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::TickComponent(float DeltaTime, ELevel
if (bAgentSpeaking)
{
FScopeLock Lock(&AudioQueueLock);
if (AudioQueue.Num() == 0)
if (AudioQueue.Num() - AudioQueueReadOffset == 0)
{
SilentTickCount++;
@ -290,6 +319,10 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::EndConversation()
{
if (GetOwnerRole() == ROLE_Authority)
{
// Cancel any pending reconnection.
bWantsReconnect = false;
ReconnectAttemptCount = 0;
StopListening();
// ISSUE-4: StopListening() may set bWaitingForAgentResponse=true (normal turn end path).
// Cancel it immediately — there is no response coming because we are ending the session.
@ -299,6 +332,7 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::EndConversation()
if (WebSocketProxy)
{
bIntentionalDisconnect = true;
WebSocketProxy->Disconnect();
WebSocketProxy = nullptr;
}
@ -706,6 +740,15 @@ const FPS_AI_ConvAgent_ConversationInfo_ElevenLabs& UPS_AI_ConvAgent_ElevenLabsC
// ─────────────────────────────────────────────────────────────────────────────
void UPS_AI_ConvAgent_ElevenLabsComponent::HandleConnected(const FPS_AI_ConvAgent_ConversationInfo_ElevenLabs& Info)
{
// Reset reconnection state on successful connection.
if (bWantsReconnect)
{
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Log,
TEXT("Reconnection successful after %d attempt(s)."), ReconnectAttemptCount);
}
bWantsReconnect = false;
ReconnectAttemptCount = 0;
SessionStartTime = FPlatformTime::Seconds();
TurnIndex = 0;
LastClosedTurnIndex = 0;
@ -739,6 +782,17 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::HandleDisconnected(int32 StatusCode,
{
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Log, TEXT("Agent disconnected. Code=%d Reason=%s"), StatusCode, *Reason);
// During reconnection, a failed attempt also triggers HandleDisconnected.
// Just schedule the next attempt without full cleanup.
if (bWantsReconnect)
{
const double Delay = FMath::Min(FMath::Pow(2.0, static_cast<double>(ReconnectAttemptCount)), MaxReconnectDelay);
NextReconnectTime = FPlatformTime::Seconds() + Delay;
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Log,
TEXT("Reconnection attempt failed — retrying in %.0fs..."), Delay);
return;
}
// ISSUE-13: stop audio playback and clear the queue if the WebSocket drops while the
// agent is speaking. Without this the audio component kept playing buffered PCM after
// disconnect. StopAgentAudio also fires OnAgentStoppedSpeaking if bAgentSpeaking was true,
@ -757,6 +811,26 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::HandleDisconnected(int32 StatusCode,
FScopeLock Lock(&MicSendLock);
MicAccumulationBuffer.Reset();
}
// ── Auto-reconnect on unexpected disconnect ─────────────────────────────
// StatusCode 1000 = normal close (user called EndConversation).
// bIntentionalDisconnect covers EndConversation and ServerReleaseConversation.
if (!bIntentionalDisconnect && StatusCode != 1000
&& MaxReconnectAttempts > 0 && GetOwnerRole() == ROLE_Authority)
{
bWantsReconnect = true;
ReconnectAttemptCount = 0;
const double Delay = 1.0; // First attempt after 1 second.
NextReconnectTime = FPlatformTime::Seconds() + Delay;
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Warning,
TEXT("Unexpected disconnect — will attempt reconnection in %.0fs (max %d attempts)."),
Delay, MaxReconnectAttempts);
OnAgentError.Broadcast(TEXT("Connection lost — reconnecting..."));
// Keep bNetIsConversing / NetConversatingPawn so the NPC stays occupied.
return;
}
bIntentionalDisconnect = false;
// Reset replicated state on disconnect.
if (GetOwnerRole() == ROLE_Authority)
{
@ -772,6 +846,18 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::HandleDisconnected(int32 StatusCode,
void UPS_AI_ConvAgent_ElevenLabsComponent::HandleError(const FString& ErrorMessage)
{
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Error, TEXT("Agent error: %s"), *ErrorMessage);
// During reconnection, a connection error should schedule the next attempt
// (OnWsConnectionError fires but OnWsClosed may not follow).
if (bWantsReconnect)
{
const double Delay = FMath::Min(FMath::Pow(2.0, static_cast<double>(ReconnectAttemptCount)), MaxReconnectDelay);
NextReconnectTime = FPlatformTime::Seconds() + Delay;
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Log,
TEXT("Reconnection error — retrying in %.0fs..."), Delay);
return;
}
OnAgentError.Broadcast(ErrorMessage);
}
@ -783,7 +869,7 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::HandleAudioReceived(const TArray<uint
int32 QueueBefore;
{
FScopeLock Lock(&AudioQueueLock);
QueueBefore = AudioQueue.Num() / sizeof(int16);
QueueBefore = (AudioQueue.Num() - AudioQueueReadOffset) / sizeof(int16);
}
if (bDebug && DebugVerbosity >= 2)
{
@ -1107,13 +1193,28 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::OnProceduralUnderflow(
return;
}
if (AudioQueue.Num() > 0)
const int32 Available = AudioQueue.Num() - AudioQueueReadOffset;
if (Available > 0)
{
const int32 BytesRequired = SamplesRequired * sizeof(int16);
const int32 BytesToPush = FMath::Min(AudioQueue.Num(), BytesRequired);
const int32 BytesToPush = FMath::Min(Available, BytesRequired);
InProceduralWave->QueueAudio(AudioQueue.GetData(), BytesToPush);
AudioQueue.RemoveAt(0, BytesToPush, EAllowShrinking::No);
InProceduralWave->QueueAudio(AudioQueue.GetData() + AudioQueueReadOffset, BytesToPush);
AudioQueueReadOffset += BytesToPush;
// Compact when read offset exceeds half the buffer to prevent unbounded growth.
// This replaces the old RemoveAt(0,N) per-callback O(n) memmove with a periodic
// compaction that happens much less frequently (~every 0.5-1s instead of 60Hz).
if (AudioQueueReadOffset > AudioQueue.Num() / 2)
{
const int32 Remaining = AudioQueue.Num() - AudioQueueReadOffset;
if (Remaining > 0)
{
FMemory::Memmove(AudioQueue.GetData(), AudioQueue.GetData() + AudioQueueReadOffset, Remaining);
}
AudioQueue.SetNum(Remaining, EAllowShrinking::No);
AudioQueueReadOffset = 0;
}
// Log when queue recovers (new data arrived after being dry)
if (bQueueWasDry)
@ -1124,7 +1225,7 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::OnProceduralUnderflow(
const double T = FPlatformTime::Seconds() - SessionStartTime;
UE_LOG(LogPS_AI_ConvAgent_ElevenLabs, Log,
TEXT("[T+%.2fs] [Turn %d] AudioQueue recovered — feeding real data again (%d bytes remaining)."),
T, LastClosedTurnIndex, AudioQueue.Num());
T, LastClosedTurnIndex, AudioQueue.Num() - AudioQueueReadOffset);
}
}
}
@ -1269,6 +1370,7 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::StopAgentAudio()
{
FScopeLock Lock(&AudioQueueLock);
AudioQueue.Empty();
AudioQueueReadOffset = 0;
bAgentGenerating = false; // Always clear — covers interruptions during generation phase.
bAgentResponseReceived = false; // Reset — next response will re-confirm when done.
@ -1536,12 +1638,17 @@ void UPS_AI_ConvAgent_ElevenLabsComponent::ServerRequestConversation_Implementat
void UPS_AI_ConvAgent_ElevenLabsComponent::ServerReleaseConversation_Implementation()
{
// Cancel any pending reconnection.
bWantsReconnect = false;
ReconnectAttemptCount = 0;
StopListening();
bWaitingForAgentResponse = false;
StopAgentAudio();
if (WebSocketProxy)
{
bIntentionalDisconnect = true;
WebSocketProxy->Disconnect();
WebSocketProxy = nullptr;
}

View File

@ -472,7 +472,7 @@ void UPS_AI_ConvAgent_WebSocket_ElevenLabsProxy::HandleAudioResponse(const TShar
(*AudioEvent)->TryGetNumberField(TEXT("event_id"), EventId);
if (EventId > 0 && EventId <= LastInterruptEventId)
{
UE_LOG(LogPS_AI_ConvAgent_WS_ElevenLabs, Verbose, TEXT("Discarding audio event_id=%d (interrupted at %d)."), EventId, LastInterruptEventId);
UE_LOG(LogPS_AI_ConvAgent_WS_ElevenLabs, Verbose, TEXT("Discarding audio event_id=%d (interrupted at %d)."), EventId, LastInterruptEventId.load());
return;
}

View File

@ -175,6 +175,14 @@ public:
ToolTip = "Seconds to wait for a server response after the user stops speaking.\nFires OnAgentResponseTimeout if exceeded. Normal latency is 0.1-0.8s.\nSet to 0 to disable. Default: 10s."))
float ResponseTimeoutSeconds = 10.0f;
/** Maximum auto-reconnection attempts on unexpected WebSocket disconnect.
* Uses exponential backoff (1s, 2s, 4s, ... up to 30s).
* Set to 0 to disable auto-reconnection. */
UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "PS AI ConvAgent|ElevenLabs",
meta = (ClampMin = "0",
ToolTip = "Auto-reconnect attempts on unexpected disconnect.\n0 = disabled. Uses exponential backoff."))
int32 MaxReconnectAttempts = 5;
// ── Multi-agent / external mic ───────────────────────────────────────────
/** When true, StartListening/StopListening manage the turn state but do NOT
@ -537,10 +545,11 @@ private:
std::atomic<bool> bAgentSpeaking{false};
// True from the first agent_chat_response_part until the first audio chunk arrives.
// Used to block StartListening() while the server is processing the previous turn.
bool bAgentGenerating = false;
// Atomic: defensive — currently game-thread only, but documents thread-safety contract.
std::atomic<bool> bAgentGenerating{false};
// True between StopListening() and HandleAgentResponseStarted() (first chat response part).
// Used to fire OnAgentResponseTimeout if the server takes too long to start generating.
bool bWaitingForAgentResponse = false;
std::atomic<bool> bWaitingForAgentResponse{false};
// ── Turn tracking ─────────────────────────────────────────────────────────
// Monotonically increasing counter. Incremented each time StartListening() successfully
@ -561,7 +570,11 @@ private:
double AgentSpeakStart = 0.0; // Set in EnqueueAgentAudio (first chunk) — when audio begins.
// Accumulates incoming PCM bytes until the audio component needs data.
// Uses a read offset instead of RemoveAt(0,N) to avoid O(n) memmove every
// underflow callback (~60Hz). Compacted periodically when read offset
// exceeds half the buffer.
TArray<uint8> AudioQueue;
int32 AudioQueueReadOffset = 0;
FCriticalSection AudioQueueLock;
// Pre-buffer state: delay playback start to absorb TTS inter-chunk gaps.
@ -611,6 +624,15 @@ private:
* Formula: bytes = SampleRate * (ms / 1000) * BytesPerSample = 16000 * ms / 1000 * 2 = 32 * ms */
int32 GetMicChunkMinBytes() const { return MicChunkDurationMs * 32; }
// ── Reconnection ────────────────────────────────────────────────────────
// Exponential backoff reconnection when the WebSocket drops unexpectedly.
// Only active on the Authority (server/standalone) where the WebSocket lives.
bool bWantsReconnect = false;
bool bIntentionalDisconnect = false;
int32 ReconnectAttemptCount = 0;
double NextReconnectTime = 0.0;
static constexpr double MaxReconnectDelay = 30.0;
// ── Opus codec (network audio compression) ───────────────────────────────
TSharedPtr<IVoiceEncoder> OpusEncoder; // All: server encodes agent audio, clients encode mic audio
TSharedPtr<IVoiceDecoder> OpusDecoder; // All: clients decode agent audio, server decodes mic audio

View File

@ -232,7 +232,8 @@ public:
// ── Getters (read by AnimNode) ───────────────────────────────────────────
/** Get current eye gaze curves (8 ARKit eye look curves).
* Returns a COPY scaled by activation alpha safe to call from any thread. */
* Returns a COPY scaled by activation alpha safe to call from any thread.
* NOTE: Prefer FillCurrentEyeCurves() for hot paths to avoid per-frame allocation. */
UFUNCTION(BlueprintCallable, Category = "PS AI ConvAgent|Posture")
TMap<FName, float> GetCurrentEyeCurves() const
{
@ -244,6 +245,24 @@ public:
return Scaled;
}
/** Fill an existing TMap with current eye gaze curves (zero-alloc after first call).
* Reuses the caller's existing hash table no heap allocation when keys match.
* Thread-safe (takes PostureDataLock internally). */
void FillCurrentEyeCurves(TMap<FName, float>& OutCurves) const
{
FScopeLock Lock(&PostureDataLock);
if (CurrentActiveAlpha < 0.001f)
{
for (auto& Pair : OutCurves) { Pair.Value = 0.0f; }
return;
}
const float Alpha = (CurrentActiveAlpha >= 0.999f) ? 1.0f : CurrentActiveAlpha;
for (const auto& Pair : CurrentEyeCurves)
{
OutCurves.FindOrAdd(Pair.Key) = Pair.Value * Alpha;
}
}
/** Get current head rotation offset (applied by AnimNode as FQuat to avoid
* Euler round-trip that reintroduces parasitic tilt on diagonals).
* Thread-safe copy, blended by activation alpha. */

View File

@ -6,6 +6,7 @@
#include "UObject/NoExportTypes.h"
#include "PS_AI_ConvAgent_Definitions.h"
#include "IWebSocket.h"
#include <atomic>
#include "PS_AI_ConvAgent_WebSocket_ElevenLabsProxy.generated.h"
// ─────────────────────────────────────────────────────────────────────────────
@ -239,12 +240,13 @@ private:
// Timestamp when user turn ended (StopListening).
double UserTurnEndTime = 0.0;
// Whether we are waiting for the first response after user stopped speaking.
bool bWaitingForResponse = false;
// Atomic: defensive — documents thread-safety contract.
std::atomic<bool> bWaitingForResponse{false};
// Whether we already logged the first audio response latency for this turn.
bool bFirstAudioResponseLogged = false;
std::atomic<bool> bFirstAudioResponseLogged{false};
// Whether OnAgentResponseStarted has already been fired for the current turn.
// Reset at turn start so only the first agent_chat_response_part fires the event.
bool bAgentResponseStartedFired = false;
std::atomic<bool> bAgentResponseStartedFired{false};
// Timestamp when the conversation was initiated (conversation_initiation_metadata received).
// Used to compute [T+Xs] session-relative timestamps in all log messages.
@ -255,7 +257,8 @@ private:
// Audio events whose event_id <= LastInterruptEventId belong to the cancelled
// generation and must be discarded. Only AUDIO is filtered — transcripts,
// agent_response, agent_chat_response_part etc. are always processed.
int32 LastInterruptEventId = 0;
// Atomic: defensive — documents thread-safety contract.
std::atomic<int32> LastInterruptEventId{0};
public:
// Set by UPS_AI_ConvAgent_ElevenLabsComponent before calling Connect().