AI service retry review Use approved redacted error metadata; omit prompts, credentials and sensitive payloads. Record provider/endpoint, SDK name and installed version, request ID, status/error type, Retry-After, timeout, retry configuration, total deadline, and observed attempt count. Check spend/quota/account errors separately from temporary rate/overload errors. Choose the single layer responsible for retries. Count SDK attempts and outer attempts. Confirm repeat safety for this operation; unknown outcome requires reconciliation. Define additional retry count, elapsed deadline, cancellation, concurrency and queue limits. Respect the provider wait hint or stop/defer; never shorten it just to fit a budget. For production, implement random jitter, monotonic timing and per-attempt timeouts. Handle stream interruptions and partial tool actions separately from HTTP status alone. Show a clear pending/failed state and request ID; keep manual recovery available. Test policy under controlled conditions before adopting it for real requests. Do not infer security, cost savings or recovery guarantees from the offline exercise.