Streaming and connection checks
SSE, proxy buffering and timeout diagnosis.
Read response chunks
Set stream: true for Chat Completions, read SSE data lines and choices[].delta, then finish on [DONE]. Use curl -N to disable buffering. Allow long connections in clients and proxies; first-token latency includes upstream reasoning.
Disconnects and empty replies
Start with a minimal non-streaming request to isolate authentication and model errors, then check proxy buffering, timeouts and finish_reason. A small token limit can be consumed by reasoning; parameters depend on the model.