- Providers get a default reasoning effort (low..max, empty = not sent);
chat can override it per message. OpenAI sends reasoning_effort,
Claude sends output_config.effort with adaptive thinking, and thinking
blocks are replayed during the tool loop.
- Show a thinking indicator while waiting for model output.
- Surface in-stream upstream errors (e.g. overloaded) instead of
breaking the SSE stream, and retry them up to 3 times when nothing
has been emitted yet.
