Added#
A ts field on the usage record, and the two findings only a clock can make.
ISO 8601, epoch seconds or milliseconds, or OpenAI's created; parsed under the
same three-state rule as the counts — absent is null, present-and-unreadable
rejects the line into skippedLines, because a silently dropped timestamp would
mis-measure every gap it touches.
The span. The report states what period the log covers — This log covers 2026-08-01 → 2026-08-14 (13.0 days) — and deliberately stops there. Stated,
never extrapolated: the span makes the reader's own monthly arithmetic valid,
while a per-month figure from a partial month would be Trazum doing the guessing
it exists to end. Partial coverage is said in the same breath, and a test asserts
the span alone conjures no monthly figure anywhere.
Does the cache TTL fit how fast the turns come? A cache entry lives 5
minutes, or an hour at 2x the write price, and whether either fits depends on a
number the bill never shows: how long the workload waits between turns. Measured
as the median gap between consecutive turns of the same conversation, sorted by
the recorded clock so the answer is independent of the order of the log. Five
states: expires before reuse — the mechanism behind a losing cache, with
both honest ways out named; overlong TTL — turns seconds apart paying 2x for
an hour of endurance they never use, priced exactly as the same tokens at the
other published rate; unsettled when the unrecorded TTL decides it; fits
said out loud; and could not be measured over writes with no clock, rather
than silence. Rendered in the CLI (both locales), the MCP profile_usage tool
and the web Bill tab; --json carries span and cacheTtlFit.
The recording recipe gains ts in the README, the onboarding message and the
docs-pinned fixture, so following the documented recipe still produces a report
that asks for nothing more.