Added#
Gemini's log shape, recognised. usageMetadata appears in no other
provider's response, which is what makes reading it honest — no guessing
which provider a field belongs to. promptTokenCount, candidatesTokenCount
and cachedContentTokenCount are read, with the cached half subtracted from
the prompt count through the same mechanism as OpenAI's cached_tokens,
because Gemini sets the same double-charge trap. finishReason: "MAX_TOKENS"
joins the truncation contract's three-way reading. The bundled catalogue has
priced gemini-2.5-pro and gemini-2.5-flash since the seven-provider work —
what was missing was reading the log their SDK actually writes, and a Gemini
model the catalogue does not know still parses and lands in unpricedModels
as everywhere.
--dry-run: what the log could answer, before you wire it in. The
question somebody has before putting Trazum in CI is not "what did we spend"
but "will this log support the gates I want" — and answering it with a full
report makes them read a bill to discover a field is missing. trazum profile <log> --dry-run states readiness per capability (totals, per-workload,
clock, sessions, stop reasons, the cache split), names unpriced models, and
produces no dollar figure at all, so nothing in it can be mistaken for spend.
It refuses to run beside a gate — a gate over a report that was never
produced would exit green having judged nothing — distinguishes "nothing
wrote to the cache" from a missing split, and exits non-zero when nothing
parses, because an unreadable log is not a ready one.