Released a context-window budgeter for long refactors

Close

Shows what is eating your context before the model forgets.

It reads a session transcript and breaks the token count down by file, by tool result, and by your own instructions. The usual finding is that one large generated file is taking a third of the window and contributing nothing.

Windsurf built the parser. What I learned building it: test output is almost always the biggest offender, and trimming it to the failing lines alone bought me roughly twice the working room on a long refactor.