Added#
The README catches up with eight releases. The profile section now
documents what the tool actually reports: the shape of the day and the Batch
API question it answers, what one conversation costs (median against p95),
conversations that never came back and the ceiling/fact split their figure
carries, and the coverage section. The CI-gate section gains the third gate
(--max-cache-loss-usd, reading the worst case), spend budgets in the
config, and the floor note every gate now prints; the output section names
--csv-out, directory mode and the documented --json contract.
The coverage section reaches the MCP and the browser. profile_usage
lists the missing fields with counts an agent can act on, and the web Bill
tab shows them in the gaps footer — same counts, same silence on a complete
log. An agent told "labelled" by a boolean would stop asking; told "12/40,000"
it does the right thing.
"What this log cannot answer yet." Every finding past the totals needs a
field the log format does not require, and a reader who never adds them sees a
report quietly missing half of itself — with no way to tell "nothing to
report" from "nothing recorded". The report now names each missing field with
what it would unlock, from exact counts: "label" on 0/40,000 records,
"session" on 12/40,000. Counts rather than booleans, because twelve
labelled records out of forty thousand is not a labelled log and a boolean
would call it one. Coverage is counted over records that parsed, priced or
not — whether a field is present is a property of the log, not of the price
catalogue — and the cache-TTL line is counted only over records that actually
wrote to the cache. A complete log gets no section at all: a paragraph of
things that are fine is the paragraph readers learn to skip.