Added#
spend.maxDayUsd: the day budget as repository policy. --max-day-usd
existed as a flag, which makes it one CI invocation's opinion; the config is
the repository's standing word. The flag still wins when both are present,
like every gate here, and the config path inherits the flag's refusal: a log
with no timestamps fails the day budget rather than passing it, because "not
measured" is not "under budget".
--csv-shape model-day: the drift, day by day. modelMixDrift states
two halves; a chart wants the whole series. The new shape writes one row per
UTC day and model — day,model,usd,calls, the long format a pivot table
or a plotting library takes as-is — and spendByDay in --json now carries
the same per-model split on every day. Model ids are formula-defused like
every other untrusted cell, no total row as in every shape, and the
unpriced stay out because their dollars were never computed.
truncationRetries: the "billed again" half, measured. The truncation
finding has always said cut-off answers are frequently retried, billed
again — and that second half was an assertion, because nothing counted the
retries. Now it is a measurement: a truncated answer followed within two
minutes by another call in the same conversation is the shape a retry has,
and both sides of the pair are in the log. The wasted attempt's full price
and the follow-up's travel together, with the checkable denominator —
"2 of 3 truncated answers" and "2 of 2" are different sentences, and the
reader gets the right one.
Attributed to the truncated call's slice, because that is where the
max_tokens ceiling that caused the pair lives and where the fix is
applied. Compared only against the immediately previous call in the session
(the bounded-memory design repeatedTurns uses); a single pair is not
reported — one retry is an anecdote; and the hedge is in every rendering:
the log cannot see content, so whether each pair was a retry is the
reader's to know. Needs a session, a clock and a stop reason.
In the terminal (inside the truncation section, next to the ceiling the
answers actually needed), in --json, in the CI summary and in the MCP
profile_usage tool.