opus-ratelimit-fanout-retry-on-sonnet-throttled-waves

What problem does it solve? When a large parallel fan-out of Opus subagents (Workflow parallel/pipeline or many concurrent Agent calls) mass-fails with HTTP 429 "Server is temporarily limiting requests (not your usage limit)", only a handful of agents survive and the run returns far fewer results than dispatched. Waiting and retrying on Opus often re-hits the same capacity throttle. ## Core Features & Use Cases - Model-swap recovery: Re-runs the same fan-out with model: 'sonnet' on the worker agents, since Sonnet draws from a separate capacity pool and sidesteps the Opus-specific server throttle. - Wave throttling: Replaces one wide N-wide burst with sequential waves of ~8 parallel agents, spacing request bursts so 50 simultaneous requests never hit at once. - Scope guidance: Keeps the orchestrator and final synthesis/judge on Opus while moving only triage, review, code-tracing, or classification fan-out workers to Sonnet. - Use Case: A ~50-issue triage fan-out on Opus collapses with nearly all agents returning "Rate limited · 429"; re-running the identical fan-out on Sonnet in waves of 8 completes all ~50 with zero missing. ## Quick Start Ask the AI to re-run the failed Opus fan-out with the worker agents set to Sonnet and chunked into sequential waves of about eight parallel agents each.