Added#
The Bill tab catches up, four findings at once. The browser lacked the
CLI's newest sections: the doubled-bill warning (duplicateLines, placed
above the totals it would inflate), the same request sent again
(repeatedTurns), the ceiling in sight (contextPressure) and the moving
mix (modelMixDrift). All four render in the tab now, computed in the page
like everything else there, with the CLI's own thresholds — 85% loud on
pressure, fifteen points on drift — because two surfaces summarising one log
differently is a second opinion nobody asked for. Verified by driving the
built page in Chromium with a log that triggers all four at once, zero
network requests recorded.
modelMixDrift: the migration a total cannot show. A bill can grow with
no workload growing — traffic quietly moving from the cheap model to the
expensive one, a deploy that flipped a default, a fallback that became the
main path. Day totals cannot show it (both models land in the same number)
and per-model totals cannot either (a total has no direction). The log's days
are split chronologically in half and each model's share of the priced
spend in each half is stated, exactly.
null — not empty — under four dated days: one day against one day is
weather presented as climate. The renderings speak only past fifteen points
of movement (a presentation threshold, stated in the copy), while --json
carries every model's exact shares either way — the data states, the
rendering judges. And never a forecast: where the mix goes next is not in the
log, so it is not said.
In the terminal, --json, the CI summary and the MCP profile_usage tool.
contextPressure: the ceiling, seen coming. Input grows turn by turn or
document by document, costs nothing extra to grow, and then one call crosses
the model's context window and the API refuses it outright — the bill looks
fine right up to the day the product breaks. The distance is knowable from
what the log already carries: each slice's largest call (input, cache reads
and writes) against its own model's window, reported from half the window up,
loud from 85%.
What it refuses: no prediction of when. A straight line through two
points is a guess wearing arithmetic's clothes; the share is a fact and the
trajectory is the reader's to know. The maximum rather than a percentile,
for reprice's reason — one call over the window is one failed call, and an
average hides exactly the call that matters. Unpriced models are absent: a
model with no catalogue entry has no window to compare against, and
inventing one would turn a missing fact into a false comfort.
In the terminal, in --json as contextPressure (computed once in the CLI
and passed to every rendering, because the denominator lives in the possibly
overlaid catalogue), in the CI summary, and in the MCP profile_usage tool.
Also: the README's action pins advanced to the 1.26.0 commit.