Attribution tags

-H "x-vatan-prop-customer: acme" \
-H "x-vatan-prop-feature: summariser" \
-H "x-vatan-prop-env: production"

Every x-vatan-prop-<name> header is stored against the request, then filterable in the log and groupable on the usage page. This is what turns "you spent $40 on Sonnet" into "you spent $40 on Sonnet for this customer in this feature", which is the difference between usage data and something you can bill against.

Headers rather than a body field, deliberately: a body field would break the OpenAI SDKs' request types, and every SDK working unmodified is the entire pitch.

Up to 16 tags per request. Names are capped at 64 characters and values at 256, and a longer value is truncated rather than rejected: losing the tail of a label beats losing the request over one.