Case:
requests #187 (gemini-4-argon), #302 (gemini-4-carbon), and #303 (gemini-4-carbon) on chat_completions:https://api.githubcopilot.com each hung silently for ~27s / 52s / 55s while Gemini was thinking, emitted a single opening chunk with response_id (Started(id) → streamed_events = 1, thinking_bytes = 0, output_bytes = 0), and then Copilot closed the SSE stream with data: [DONE] without ever emitting a finish_reason.
So add a test case for this & make it retry instead.
Other folks meet this as well https://github.com/agentgateway/agentgateway/issues/4220 https://github.com/earendil-works/pi/issues/4345