1.18.0 - the setting judged against the work

1.18.0 - the setting judged against the work, not against the window

Everything in this plugin has been budget-triggered: it speaks when a window is
filling. That is the wrong trigger for the question Gev actually asked - notice
the model and effort are bigger than the job needs, and act - because a
mechanical hour at the top setting is waste at 10 per cent used exactly as much
as at 80. The only difference is that at 80 somebody notices.

usage.settingFit() measures what the current effort costs against the cheapest
level with real evidence behind it, on this account, and the line reports it
from the first prompt at any percentage: "ultra, measured at 6 times the cost of
medium a turn (8 turns against 6)".

The plugin cannot judge how hard the work is; only the agent reading it can. So
it supplies the measured half and asks for the judgement, naming all three
levers: drop the effort, hand the stretch to a cheaper model, or do less of it
at this setting - a fan-out multiplies the setting across every agent, so six
agents at ultra is six ultra turns rather than one. Put it back when the work
gets hard again.

Two things it will not do. It never invents a ratio: with one effort level
measured, or fewer than three turns either side, it says nothing, because a
number off a price list would have an agent drop effort on a hunch and call it
evidence. And it asks once per setting per session rather than every prompt -
the question only changes when the setting does.

Found while wiring it: settingFit read the previously-asked setting after
overwriting it, so it compared the new answer against itself and could never
have fired; and it iterated an event list that a cache hit does not have.

633 tests.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com