Limits, scheduling, and retries

Distinguish different constraints

Requests per minute, tokens per minute, context size, and concurrency describe different resource constraints. Ask support for actual values, scope, and reset behavior; this site does not supply speculative quotas. Plan for bursts and long tasks rather than estimating demand only from average call counts across a quiet period.

Bound retries by cause and deadline

Use application queues to bound active work and choose randomized backoff according to actual response information. Correct authentication and parameter errors instead of endlessly resubmitting them. Give each task a deadline and cancellation path so expired work cannot keep consuming resources after the user has abandoned the result.

Keep useful operational evidence

Observe queue time, timeouts, retries, and completion states while retaining redacted request identifiers. If a stream interrupts or completion is uncertain, assess duplicate-execution risks before retrying. These are application-side recommendations, not claims that a console quota manager, automatic traffic controls, or reserved capacity is already available from UBTL.