429 Model Capacity Issues, No sufficient re-routing

[M-DEV-1]

What happened?

I am currently using Auto (Gemini 3) on a Pro account.

Sometimes, there is capacity issue on the connected server, because of which it might get stuck trying for a model (such as gemini-3-flash-preview) repeatedly even though there is a clear no capacity 429 error message.

What did you expect to happen?

While the existing backoff algorithm is fine, it should be exponential and should switch the other models available (such as gemini-3.1-pro or flash-lite) and retry as required to ensure that it doesn't silently spend ~10-20 minutes trying to access a model which is not currently available.

Screenshots (Dev Console)

Client information

Client Information

Run gemini to enter the interactive CLI, then run the /about command.

> /about
╭──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│                                                                                                                                                          │
│ About Gemini CLI                                                                                                                                         │
│                                                                                                                                                          │
│ CLI Version                                          0.41.2                                                                                              │
│ Model                                                Auto (Gemini 3)                                                                                     │
│ Sandbox                                              no sandbox                                                                                          │
│ OS                                                   win32                                                                                               │
│ Tier                                                 Gemini Code Assist in Google One AI Pro                                                             │
│                                                                                                                                                          │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

Login information

Google OAuth

Anything else we need to know?

I'm willing to open a PR with an improved backoff mechanism. Do let me know if the proposed solution works or if there are other suggestions that you may have to offer.

[M-DEV-1]

Clearly, max attempts are 10. But as per the timestamps, it simply doesn't stop trying. and all of this is on a single prompt.

Dev Console

My CLI

[gemini-cli]

This issue requests exponential backoff and better model routing when facing 429 capacity errors.

Effort Analysis:
The backoff mechanism is implemented in packages/core/src/utils/retry.ts within retryWithBackoff. Currently, it retries the same API call. To support routing to fallback models upon persistent 429 errors, the retry logic needs to be integrated with the model routing capabilities (e.g., packages/core/src/availability/modelPolicy.ts). This requires changes across the asynchronous flow of the API orchestration layer in packages/core, specifically passing routing state or throwing a specialized error to trigger the router higher up in the execution chain. This maps to the effort/medium criteria of 'Asynchronous Flow & Integration' and 'Cross-Component Refactors'.

[Marcus-Rise]

same in #24937