Caveats

• POST /chat/completions with stream=true: streamed chunks are not in the standard OpenAI Chat Completions delta shape—do not point a strict OpenAI client at streaming here unless you accept that mismatch.
• Non-streaming POST /generate: the JSON body may not match what a minimal “single AG-UI blob” client expects when the workflow behaves like a chat-completions response—validate against your client assumptions.