Read the explanation
The default five enabled skills consume four hundred fifty, six hundred, three hundred fifty, eight hundred, and three hundred tokens, totaling twenty five hundred. With context limit forty ninety six the ratio is about point six one zero three five; at sixteen thousand three hundred eighty four it is about point one five two five nine. At point zero one five pixels per context-limit token the bars span sixty one point four four and two hundred forty five point seven six. These are declared skill costs and limits, not actual tokenization or measured inference. The original D3 chart fails offline but arithmetic telemetry updates before that failure, and the source export remains locally usable. The same default skill set has a dilution penalty computed as point two five times its context ratio raised to one point three, capped at point three five. Expanding the context limit from forty ninety six to sixteen thousand three hundred eighty four reduces the displayed penalty from thirteen point two to two point two percent. The displayed net estimate rises from seventy seven point six to eighty eight point five percent. At two point five pixels per percentage point the bars span one hundred ninety four and two hundred twenty one point two five. These are heuristic formula outputs, not empirically validated success rates or a scientific conclusion that a larger context window improves every agent. With no enabled skills the source returns the base accuracy of seventy percent and zero skill penalties or gain. The default five skills produce seventy seven point six at the original context limit. At two point five pixels per percentage point the bars span one hundred ninety four and one hundred seventy five. For nonempty sets net success is base plus gross gain minus context dilution minus half the misfire heuristic, clamped between five and ninety nine percent. The optimal-count sweep checks first-k prefixes of the entire skill list rather than all enabled subsets, so its badge is not an exhaustive optimizer. These limitations remain in the local JSON; no live model evaluation is performed.