What
The "Context low" warning and the "% until auto-compact" footer now run the current model through a helper first. From the result they work out three things: the warning level, the usable context size and the point where auto-compact starts. Context is how much of the conversation the model can hold at once. Auto-compact is when Claude Code shortens the conversation to make room. The wording of the warning is unchanged.
Why
The warning's numbers now come from the model after this extra step, so the percentage shown can differ for some models even though the text looks the same.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
The finding does not say what the helper does to the model, so it is unclear for which models the numbers change.