429

One code covers three conditions, because the client action is identical: honour Retry-After. Which one you hit is in error.details. A spent rate window carries window_seconds — 60 for the burst floor, 18000 for the 5-hour budget — and retry_after_seconds, the same value the header carries; an exhausted 5-hour budget returns 18000, not 60, because retrying sooner just hits the same wall. The in-flight limit instead carries concurrency_limit (and no window_seconds), with a Retry-After of 1: your organization already has that many requests running on one API server, and no quota was spent.

Honour Retry-After exactly. Watch X-RateLimit-Remaining-5h as the primary signal and slow down before you get here. Prefer cursor pagination for large pulls: fewer, fuller pages cost fewer requests. If you parallelize, keep the pool to 10 or fewer concurrent requests.