Skip to content
Trazum

1.24.0

How big, how uneven, and the day it spiked

Added#

The CI summary says what the terminal says. --markdown-out is what a pull request comment and a GitHub job summary show, which is where most people will ever read this report — and three findings the terminal makes were missing from it. A finding the summary omits is a finding nobody sees.

  • The doubled-bill warning, above the figures it would inflate rather than below them: a summary showing a total nobody can trust is worse than one showing no total.
  • How big the calls are, with the same four-times-the-median threshold and the same two sentences, because a summary that words a finding differently is a second opinion nobody asked for.
  • The --what-if repricing, with the assumption above the figure here too. A pull request comment is exactly where a dollar amount with the caveat underneath would be read as a recommendation and merged.

Fixed#

"1 lines are exact duplicates". The duplicate-line sentence took a pre-formatted count and could not agree with its own verb. It takes a number now and reads "1 line is an exact duplicate" — in both languages, where the Spanish had the same fault.

inputShapes: how big the calls actually are. outputShapes said where the output spend concentrates; input had a total and nothing else — and on a RAG or agent workload input is most of the invoice. "Input is 63% of this bill" is true and unactionable. Whether the p95 call carries twelve times the median call's input is the part somebody can act on, and the two shapes want opposite responses: a cap on whatever is growing, or a shorter prompt every call sends.

Everything the model read counts as input — fresh tokens, cache reads and cache writes — because that is the size of the request, which is what a context window and a retrieval cap are about. cachedShare then says how much of it was billed at the cache-read rate, so a large slice that is caching correctly does not read as an emergency.

Every figure is a bucket ceiling, never an interpolated percentile: "half the calls fit within 1,024 input tokens" is exact for the number named, where interpolating between buckets would invent a call nobody made. Slices with fewer than twenty calls are left out entirely rather than reported at a precision they do not have — a p95 over four calls is the largest of the four wearing a percentile's name.

In the terminal, in --json as inputShapes, and in the MCP profile_usage tool.

--max-day-usd: the gate a total cannot arm. A month at $3,000 against a $4,000 budget passes while one afternoon's runaway loop burned $900 of it. --max-usd gates the sum handed in; this gates the worst single UTC day inside it, which is the shape a loop, a bad deploy or a retry storm actually has — and it names that day's biggest label, so the spike arrives with a suspect attached.

Two refusals it inherits from the rest of the tool. A log with no clock at all cannot be judged by day and fails: "not measured" is not "under budget", and a gate that silently green-lights an unmeasurable log is worse than one that was never armed. And calls with no ts are in the bill and in none of the days, so a pass says how many were left out — the worst day is a floor by whatever they held, while a failure stands regardless.