Skip to content
Trazum

1.17.0

What the report cannot see

Added#

The conversation cost reaches the MCP and the browser. profile_usage speaks the median, the p95 and the tail sentence; the web Bill tab renders them in the conversation card, the tail in the warning colour. Same thresholds, same refusal of a mean, from the same core measurement.

What one conversation costs. The question a total cannot answer: whether $4,000 is forty thousand cheap conversations or four hundred expensive ones — the figure a per-seat price or a quota is set from. sessionCosts reports the median conversation, its p95 and its maximum per label and model, exact billed counts summed per conversation, with a mean deliberately refused (one runaway agent loop drags a mean up and hides the ordinary case). A p95 past ten times the median is called out as a tail a quota can catch; a p95 beside the median is a workload that is simply expensive, and the report says that instead. Slices with fewer than five conversations are dropped — a median over four is one of the four. Session keys group turns and never appear.

A gate says when its figure is a floor. A gate can only judge the money it can see, and three things hide spend from it: unreadable lines, models the price table does not know, and clockless calls dropped by a time window. When any of them is present the gates now say so beside the verdict — a pass means "the part I could read fits", never "the bill fits". Silent passes on partial data were the flattering omission still left in the money gates.

--against warns when the two logs cover the same days. Comparing overlapping periods puts the same calls on both sides of the subtraction, so part of the reported growth is the same money counted twice. Warned between the totals line and the drivers built from it, and only when both logs carry a clock — unknown stays silent rather than reassuring.

The comparison reaches --markdown-out. The section the terminal has had since 1.11: totals with the sign convention, the overlap warning, drivers per label and per model, and a previous log with nothing priced reported as its own answer. A CI summary that compared two logs was showing only one of them.