Added#
profile reads a directory of rotated logs. Logs rotate — one file per
day for a month — and making somebody cat them together before a profile
will read them is a setup cost that gets a tool skipped. A directory is read
in name order (which for dated names is time order) as one bill, with the
number of files stated: a report over "the logs" that silently skipped one is
a total wrong by an unknown amount. A file with no trailing newline is joined
without gluing its last record to the next file's first. A directory with
nothing readable is an error naming the extensions it looked for, never an
empty report that would read as "you spent nothing".
profile --csv-out, the report for whoever signs off the bill. One row
per label and model — the grain a routing or budget decision is made at —
written from @trazum/core's profileToCsv. Three deliberate refusals:
no total row, because a total inside a data file gets summed with the data
and doubles every figure downstream; empty dollar cells for unpriced
models, never zeros, because 0 would claim those calls were free while
their tokens are real; and a label that starts with =, +, - or @ is
prefixed with an apostrophe, because a usage log is data and a spreadsheet
would otherwise run it.
Two bugs the feature's own tests caught, both now fixed: under --json the
side files (--csv-out, --markdown-out) were never written at all — a flag
that silently did nothing — and the "wrote to" notice went to stdout, turning
a parseable JSON document into a parse error. Both output paths now write the
files, and the notice goes to stderr under --json, the rule the money gates
already followed.
Money budgets that live in the repository: spend in trazum.config.json.
budgets gates the tokens a prompt file may hold; spend gates the dollars a
usage log records — maxUsd for the whole log, byLabel for each named
workload against its own limit in the same run. A per-workload budget is a
policy several people agree on, and a policy that lives in one CI invocation
is a policy nobody can read. Flags still win over the config, as everywhere.
Two refusals keep it honest. A budgeted label with no calls in the log is
reported as not measured, never as a pass: a workload that did not appear is
not a workload that came in under budget. And per-label budgets are not
applied under --since/--until, because a window makes "what this label
spent" mean a slice, and a budget written for the whole period would be gating
against something it does not describe.