A tool call returns: {"is_error": true, "error_code": "RATE_LIMITED", "message": "Too many requests", "retry_after_seconds": 5, "retryable": true}. What is the correct model behavior?
Answer: Retry after waiting approximately 5 seconds, as indicated by retry_after_seconds
A transient error with retryable: true and retry_after_seconds: 5 signals that the tool will succeed if retried after the indicated delay. The model should communicate the delay to the user if appropriate and retry. Stopping the loop (A) wastes a valid recovery path. Retrying immediately (B) without the delay will likely hit the same rate limit, producing repeated failures. Switching tools (D) is not appropriate here — the error is not semantic (wrong tool); it's transient (rate limited). The retryable: true flag and retry_after_seconds are explicit signals for retry-with-backoff behavior.