| kind | added | |
|---|---|---|
| title | codeaf app reports spending by local day, model and subject | |
| pr | 0 | |
| surface |
|
Was true: the only ways to see the usage ledger were spend.export (raw
calls) and spend.ledger (rows by run); a calendar or model breakdown had to
page every row and add them up itself.
Is true now: spend.report takes spend.ledger's filter (both days required,
at most 366 apart) and answers cells — one per stored local day, model and
subject, with a task's conversation — plus exact totals, the distinct models,
and gaps (records this process dropped, lines skipped, days derived). The
engine folds nothing. The ledger records no billing provider, task outcome or
crew, so no cell carries one and not_recorded says so; an unnamed model has
no model, an unpriced cell no cost. More than 8 192 cells, an estimated 4 MiB,
or a 4 MiB reply is refused as bad_params, never cut. It is one linear pass
over the ledger's length at the start, with no cache; the byte figure is an
estimate, not a memory promise. spend.ledger is unchanged. Engines without
the capability answer unknown_method.