Added#
profile --by-source: the fleet. One service's logs merge into one
honest bill; twelve services' logs merge into a bill that hides which one is
bleeding. The config's new sources block names services as glob patterns
over log paths; --by-source walks the directory recursively, assigns each
file to the most specific matching pattern (the budget patterns' own
precedence rule), and renders one summary per service plus a rollup that
names the source where the money is.
The findings are the ones a merged bill cannot make: the same workload running on different models in different sources (judged on each source's dearest model for the label, so one stray experiment call is not reported as a migration); caching that pays for the fleet in aggregate while losing money in a named source (reported only when the aggregate pays — when the aggregate itself loses, the whole-fleet report already shouts, and repeating it per source would be the same alarm in pieces); and sources whose logs cover different periods flagged in the copy, because the shares compare totals and a 3-day log looking cheap beside a 30-day one has nothing to do with cost.
spend.bySource gates each service against its own budget in the same run,
failing with the service named — the fleet total can be fine while one
service bleeds. A budgeted source with no logs is named, not passed: a
service that did not appear is not a service under budget. Files matching no
source pattern are named loudly — a log in no report is spend missing from
every bill. --json emits the fleet document: full per-source reports plus
the rollup, documented in docs/json-output.md.
New core module fleet.ts (assignSources, fleetRollup), browser-safe —
it reads no files; the CLI keeps its monopoly on I/O.