Rate Limit Guard#

Before Expensive Calls#

Check: provider/model, recent failures, concurrency, context size.

On 429#

1. Set concurrency to 1.
2. Stop parallel retries.
3. Shrink context.
4. Send one minimal probe.
5. Back off (exponential).

Rules#

• Never retry a giant request unchanged.
• Never fallback-spam premium providers.
• Batch corpus work in 10–20 item chunks.

Output#

Next action: proceed / recovery mode / pause.