A streaming document analysis application uses the Messages API with stream=True. Users report that the UI occasionally displays partial JSON mid-stream (e.g., '{"analysis": "The contract is') before the stream abruptly stops, leaving the UI in a broken state. The engineering team's server logs show these events correlate with HTTP 529 (overloaded) errors arriving mid-stream after approximately 2,000-4,000 tokens. What is the correct architectural fix?

Mid-stream 529 errors require a retry architecture, not a buffering workaround. The correct architectural fix is the exponential backoff retry handler that, on detecting a mid-stream 529 error, cancels the current stream, waits with jitter, and retries the full request while maintaining client-side state to resume rendering from token 0 on retry. In practice, you detect the error event in the stream, gracefully close the connection, apply exponential backoff with jitter, and retry the full request. The client must restart rendering from the beginning of the retry response. This is the standard pattern for handling transient overload errors in streaming systems. The client-side stream buffer that accumulates all chunks before parsing defeats the purpose of streaming entirely — if users must wait for the full response before anything renders, the UX is equivalent to non-streaming but with additional complexity. Crucially, 529s can happen after thousands of tokens, meaning users wait for nothing. Switching from streaming to non-streaming requests is factually incorrect — 529 overloaded errors are not streaming-specific; they can occur on any request type. The structured output mode with a defined JSON schema does not change how mid-stream errors are handled at the transport layer — the API can still return a 529 mid-stream regardless of output structure, and structured outputs do not imply buffering behavior.