Anthropic or OpenAI streaming (SSE): the stream returns HTTP 200 and then fails midway with an error event such as overloaded_error. How do I detect and recover from mid-stream errors?

Cause: With streaming (server-sent events) the HTTP status is sent before generation finishes, so a problem that happens later cannot change the status code. Claude's docs say the API may send an error event inside the stream, for example an overloaded_error that would be a 529 in a non-streaming call, and that this does not follow the normal HTTP error handling. Network drops and proxy timeouts can also cut a stream with no error event at all. Fix (best first): 1. Treat a stream as successful only when you receive its terminal event (message_stop for the Claude Messages API) and, for the final message, inspect stop_reason. Treat an "event: error" payload, a connection reset or an exception from the SDK stream iterator as a failed attempt even though the status was 200. 2. Use the official SDK streaming helpers (for example client.messages.stream in Python/TypeScript) so accumulation and error raising are handled for you, and catch the SDK's typed exceptions rather than string-matching messages. Also handle unknown event types gracefully, since new ones can be added. 3. Retry transient failures (overloaded, 5xx, dropped connection) with exponential backoff and jitter. For a non-idempotent side effect, do not execute tool calls from a stream that did not finish. 4. To resume rather than restart: keep the text received so far. Claude's docs describe capture-and-continue: for Claude 4.5 and earlier models, send the partial text as the start of a new assistant message; for Claude 4.6 and later (which reject assistant prefill) put the partial response in a user message asking the model to continue. Tool-use and thinking blocks cannot be partially recovered, so resume only from the latest complete text block, or simply redo the request. 5. For very long outputs, prefer streaming (or the Batch API) over a single non-streaming request, because idle connections can be dropped by networks. Verify: Simulate a failure by closing the connection mid-stream or injecting an error event in a test; your code should log the request-id, discard or resume the partial output and complete successfully on retry. Sources: https://platform.claude.com/docs/en/build-with-claude/streaming https://platform.claude.com/docs/en/api/errors