Repository navigation
[Bug]: Output speed in the TUI reads far too high #442
Description
The TUI prints an output speed on completed turns, like Completed in 4s with 6172 tok/s next to it. That turn produced 395 output tokens in a few seconds, so the real rate was in the low hundreds. Models do not decode at 6000 tokens per second.
Seen on Windows 11, MCode 3.0.60.108, model MiniMax-M3.1-Flash-Preview. Values run from a few thousand to tens of thousands, and they contradict the elapsed time printed beside them.
The rate is the provider's output token count divided by a measured decode window. The token count looks right, the window looks far too small. Worth a look.
Activity
[-]Output speed in the TUI reads far too high[/-][+][Bug]: Output speed in the TUI reads far too high[/+]hetaoBackend commentedon Oct 13, 2026
Thanks for catching this — you are right, and this is a real reporting bug worth fixing.
Triage action: labelled
bug+tui. Removed fromneeds-triage.What we need to reproduce it (please reply with these):
6172 tok/son a 395-token turn is far above any real decode rate, so the display is almost certainly counting wall-clock or a cumulative counter rather than this turn's tokens.Completed in Xsline with the displayedtok/s, plus the actual output token count if you can read it from/status.If you want to fix it yourself, a fork PR does not require collaborator access — please open an issue describing the scope first, and validation runs with
pnpm verify --profile platform.