Added#
The most expensive day, and the bill as a CI gate. spendByDay buckets
exact per-record dollars per UTC day — the delta each record adds to the total,
so the day arithmetic can never drift from the bill's — with the top label
attached. The report names the peak day against the median day (a mean
would let the spike inflate its own yardstick), loud only past twice it. And
--max-usd exits 1 when the log spent more than its budget, --max-growth-usd
(with --against) when the bill grew past the limit — no period assumed, both
firing under --json because CI reads the exit code there. Alone,
--max-growth-usd is an error rather than a flag that silently gates nothing.
The clock reaches every rendering. --markdown-out gains the span line,
the peak day and the TTL verdicts with the failing ones loud; the gap, day and
median helpers live once, shared by both renderings, so they cannot drift. The
web Bill tab draws spend per day — a bar per UTC day, divs rather than a chart
library, the peak bar in the warning colour — verified in a real browser.
The ceilings a max_tokens cap actually wants. OutputShape gains
medianWithinTokens and p95WithinTokens: the bucket ceiling at least half
and 95% of the measured answers fit within. Exact over the histogram, null
for the open-ended last bucket rather than an invented ceiling, and said in one
line per slice: measured on these calls, promised for nothing.
profile --label, the drill-down. Once the full report has named a
suspect, the same command profiles that workload alone — every section, the
gates included, over one label's calls. A label that matches nothing is an
error naming the labels that exist, never a silent report over zero calls that
would read as "this workload is free". With --against, both logs are
filtered, so the comparison stays one workload.
The spend gate, packaged into the GitHub Action. Hand the action a
usage-log instead of a target and it runs profile with max-usd,
against and max-growth-usd — the money gates in a workflow, with the report
in the run summary and a failing gate still writing it. The two modes are
mutually exclusive: one run gates tokens before the money is spent or the
spend itself, and folding both verdicts into one exit code would leave nobody
knowing which gate fired. Self-tested in CI with hand-checkable arithmetic: a
$5.00 log passes a $9 budget, a $15.00 log fails it, +$10.00 growth fails a $5
limit, and naming both modes at once is refused.
The comparison reaches the browser, and the drill-down reaches the MCP.
The web Bill tab takes a second usage log and renders "against the previous
log" — the sign convention stated before the first figure, drivers derived over
the union of labels so appeared and vanished workloads are named, and a
previous log with nothing priced reported as its own answer rather than zero
growth. Verified in a real browser with exact arithmetic and zero network
requests. profile_usage on the MCP gains the optional label, with the
drill-down's rule riding along; the web output shapes gain the max_tokens
ceilings.