Why You Keep Seeing 'Rate Limited — Switching to Fallback' and How to Fix It

In some cases, those 429 errors are specifically tied to long‑context requests. If you're deep into a long conversation and your prompt context has grown very large, the upstream provider may impose tighter quotas or outright refuse the request.

The Most Suspicious Part of This Setup

The fallback label

sg-claude-opus-4.7 via custom

This matters because if your primary and fallback are both pointing to the same upstream gateway, API pool, or backend, you'll see the provider switch in the UI but experience no real relief — both paths hit the same exhausted quota. While we don't have enough evidence to confirm they share a pool in your specific case, it's a very high probability worth checking directly in your configuration.

Hermes stores custom endpoints in config.yaml, and the fallback chain lives there too under the fallback_providers key.

Where to Look First

Before you start pulling your hair out, check these items in order:

1. Identify exactly what your primary model is and where the fallback chain points. Open your ~/.hermes/config.yaml and look at the fallback_providers list.
2. Check if the primary and

   sg-claude-opus-4.7 via custom

   share the same backend gateway, pool, or API key. If they do, they'll both hit the same rate limit wall.
3. If you're running through an OpenClaw gateway or custom gateway, run

   openclaw gateway probe

   to see whether the real error is a 429, an authentication failure, or a stale client‑gateway mismatch.
4. If your gateway or service reads keys from environment variables, confirm the key is correctly placed on the machine running the gateway, and restart the gateway after any key or config changes.
5. If the error mostly appears during long conversations, check your long‑context limits. OpenClaw docs explicitly note that 429s can arise because "extra usage is required for long context requests."

The Bottom Line

This message loops because Hermes is faithfully executing its per‑turn fallback design while your primary model remains rate‑limited. To stop seeing it, you need to fix the underlying cause — whether that's adjusting your provider quotas, reconfiguring your fallback to point to a truly independent provider, or reducing your prompt context size.

If you want a precise diagnosis, you can read your active config.yaml file and trace exactly which primary is failing, where the sg-* fallback actually routes to, and why it insists on retrying the primary on every turn.