The AI coding signal I watch before restarting a session

One thing I keep noticing with AI coding: the session usually gets expensive before it feels broken.

It starts with a few pasted logs, a retry loop, an old file that keeps riding along, or staying on a bigger model after the hard part is already solved. The output still looks fine, but the context is getting heavier every prompt.

That is the workflow reason behind TokenBar. It puts live token and spend visibility in the macOS menu bar so you can catch that bloat while you are still working, not after the bill or usage limit shows up.

The signal I care about most is simple: tokens per useful step. If that number starts climbing but the task is not getting clearer, it is usually time to restart the session, trim context, or switch models.

Curious how other AI builders handle this. Do you restart when the chat feels stale, or do you wait until the model starts failing?