Added#
--csv-shape day|hour, the series a spreadsheet gets asked to chart.
--csv-out wrote one table: label and model. The time series is what somebody
actually pastes into a chart, so it is now a choice rather than extra columns
— one row shape per file, because a spreadsheet that has to filter before
it can sum is a spreadsheet somebody sums wrong. The day table carries each
day's biggest label; a day whose calls carried no label at all leaves those
cells empty rather than inventing a name. No total row in any shape, still,
and a shape it does not know is refused by name.
Relative windows: --since 7d, --until now. "The last week" is what a
nightly job asks for, and computing a date in a shell to say it is the step
that gets skipped. Days and hours (7d, 24h), plus now for the other
bound. Measured against this machine's clock, not the log's — a real
difference on an exported log — and said out loud beside the window line; when
a relative window finds nothing, the refusal names the clock difference as the
likely reason instead of leaving the reader to guess.
The token budget, against what actually goes up the wire. budgets gates
a prompt file; the log records what the call carried — system prompt,
retrieved context, conversation history, tool results. When labels maps a
workload to a file and a budget covers it, the report now states the gap: a
2,000-token budget on calls carrying 50,000 input tokens governs roughly 4% of
what is sent, and nobody looking at a green build would know it. Only said
when the budget covers less than half the call, because a budget doing its job
is not news; the share is named as approximate, since the budget counts a
file with the estimator while the log counts what the provider billed; and it
never says the budget is wrong — only that it is smaller than the bill.
Cached tokens count towards the call, because a cached token was still sent.