GeezerVison

send_and_recv rebuilt data_recv as a new bytes object on every socket read, copying everything received so far. A reply spanning N reads cost O(N^2) copying, which dominates large GetImage requests.

Once a packet's length is known, hold incoming chunks in a list and join them only when the packet is complete. Connection setup and packets whose header hasn't arrived yet are handled as before.

GetImage of a full 3840x2160 ZPixmap on Xvfb: ~245 ms -> ~70 ms.
1920x1080: ~13.5 ms -> ~11.5 ms. Small replies are unchanged.

AI Disclaimer: Claude Opus 5.5 found this whilst testing performance improvements for https://github.com/Toufool/AutoSplit . I don't know what's these maintainers view on LLM-generated code. And the project doesn't seem actively maintained either. But I figured I may as well share the fix.