Thanks for the fix! Before we merge, could you:

1. Share a reproduction scenario that triggers this in practice
2. Add a test that fails on main but passes with your fix

We want to make sure this is a real-world issue, not just a theoretical code path.

_{AI Disclaimer}

Rebased on upstream/main and added a regression test.

Reproduction: The race happens when the per-request response stream's send side is closed (e.g. receive loop's finally block calls aclose()) before send_request reads from it. response_stream_reader.receive() raises EndOfStream, which wasn't caught, so response_or_error is never assigned and the isinstance check on the next line hits UnboundLocalError.

In practice this shows up when the server disconnects abruptly and the receive loop's finally block and send_request race on the response stream — the loop closes the stream before or without sending the error object.

Test: test_response_stream_closed_raises_mcp_error reproduces this by closing the response stream's send side directly while the server connection stays open (so the receive loop never enters its finally block). Fails on main with unhandled EndOfStream, passes with the fix.