Observe the signals
Read documented limit headers and request id on every response. Do not display a calculated number to a user unless response data or account settings confirm it.
Measure requests, tokens, and media operations separately: their cost and throughput can differ.
Read response headers
bash
curl --include --request GET https://cicora.ai/api/v1/models \
--header "Authorization: Bearer $CICORA_API_KEY"
# Store the request identifier and any documented limit headers with your trace.Plan load
- Bound concurrency per user and per queue.
- Place background media work into a managed queue.
- Leave time headroom for retries and do not use retries as a throughput mechanism.
Behavior on 429
Show a clear status, preserve an uncreated draft, and schedule a retry after an allowed pause.
Do not automatically change model or parameters when it would change task meaning without an explicit product choice.