Gateway request log

One row per model call through /v1 (tool gateway calls are GET /client/usage/tools/calls): model, tokens, cost, latency, time to first token (ttftMs, streamed calls) and status. Narrow with from/to, keyId, model (one model, exactly), q (request id, or text in the model or endpoint) and band — errors, or slow: a streamed call whose first token took over 5s, or an unstreamed one over 30s. Keyset-paginated — pass the previous response's nextCursor as cursor. Requires usage:read.