Greptile Summary

This PR fixes five regressions introduced by #1784's OpenAI Audio API integration: broken SSE Accept-header negotiation, default-streaming bypassing external audio storage, empty binary TTS streams not triggering retries, bare speech.audio.done events counted as content, and full audio payloads accumulating in persistence buffers.

• Accept header fix (DoStream): forces text/event-stream when the outbound Accept is empty or application/json, preserving explicit */* for binary TTS.
• Default streaming removed (audio_inbound): omitting stream_format now stays non-streaming so UpdateRequestCompletedWithAudio can persist the full audio file; only explicit \"sse\" or \"audio\" engages streaming.
• Memory fix (SummarizeBinaryChunk): persistence/live-preview buffers store a size-only summary instead of the raw audio bytes, with chunk.Size used as fallback in aggregators when Data has been elided.
• Empty-response detection (checkEmptyResponse + hasResponseContent): recognises freshly-constructed {Object:\"[DONE]\"} sentinels and restricts SpeechStreamEvent content check to AudioBase64 != \"\".

Confidence Score: 5/5

Safe to merge. All five bug fixes have targeted, well-isolated changes and the new binary streaming path is covered end-to-end by integration tests.

Each regression is addressed with the minimal necessary change: the Accept-header fix, the default-streaming removal, the binary decoder registration, the memory summarisation at every persistence boundary, and the empty-response sentinel recognition. The test suite covers all four empty-response scenarios and the full binary streaming round-trip. No correctness issues were found during review.

No files require special attention. The most complex new code is binaryChunkDecoder and SummarizeBinaryChunk, both of which are fully unit-tested.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    Client -->|POST /audio/speech| CreateSpeech

    CreateSpeech -->|stream_format absent| NonStream[Non-streaming path\nChatCompletionWithRequest]
    CreateSpeech -->|stream_format=sse| SSEPath[SSE path\nWriteSSEStream]
    CreateSpeech -->|stream_format=audio| BinaryPath[Binary path\nWriteBinaryStream]

    NonStream -->|Do HTTP| ProviderHTTP[Provider HTTP response\nbinary audio body]
    ProviderHTTP -->|UpdateRequestCompletedWithAudio| ExternalStorage[(External Storage)]

    SSEPath -->|DoStream Accept: text/event-stream| ProviderSSE[Provider SSE stream]
    ProviderSSE -->|SSE decoder| SSEDecoder[speech.audio.delta / done events]
    SSEDecoder -->|transformSpeechStreamChunk| LLMResponseSSE[llm.Response SpeechStreamEvent]

    BinaryPath -->|DoStream Accept: */*| ProviderBinary[Provider chunked audio]
    ProviderBinary -->|binaryChunkDecoder| BinaryDecoder[audio/mpeg chunks + binary.done]
    BinaryDecoder -->|transformSpeechBinaryChunk| LLMResponseBin[llm.Response SpeechAudioChunk / DONE]

    LLMResponseSSE --> CheckEmpty{checkEmptyResponse}
    LLMResponseBin --> CheckEmpty

    CheckEmpty -->|has content| InboundTransform[inbound.TransformStream]
    CheckEmpty -->|DONE with no content| Retry[ErrEmptyResponse - retry]

    InboundTransform -->|SpeechAudioChunk| StreamEventBin[StreamEvent Type=audio/mpeg Data=bytes]
    InboundTransform -->|SpeechStreamEvent| StreamEventSSE[StreamEvent Type=speech.audio.delta]
    InboundTransform -->|binary DONE| StreamEventDone[StreamEvent Type=binary.done]

    StreamEventBin -->|InboundPersistentStream| SummarizePersist[SummarizeBinaryChunk\nSize only stored]
    SummarizePersist --> DB[(DB chunks\naudio_bytes count)]
    StreamEventBin -->|WriteBinaryStream| ClientResponse[Binary audio response]

    StreamEventDone -->|WriteBinaryStream skip| ClientResponse

_{Reviews (2): Last reviewed commit: "fix: 处理 TTS 流式评审问题" | Re-trigger Greptile}