Skip to content

PATCH-README.md

Latest commit

 

History

History
49 lines (30 loc) · 1.94 KB

PATCH-README.md

File metadata and controls

49 lines (30 loc) · 1.94 KB
Raw

Patched Gemini CLI

This repository contains a patched version of the Google Gemini CLI that prevents automatic downgrading from gemini-2.5-pro to gemini-2.5-flash, regardless of latency issues or rate limits.

Modifications

The following modification was made to the codebase:

  • Disabled Auto-Downgrade Logic: Modified packages/core/src/utils/retry.ts to remove the code that switches from gemini-2.5-pro to gemini-2.5-flash when multiple 429 errors (Too Many Requests) are detected for OAuth users. This keeps the CLI consistently using gemini-2.5-pro even under high latency conditions.

Benefits

  • Consistently uses the more capable gemini-2.5-pro model
  • Maintains all original functionality, authentication flows, and rate limit handling
  • Only prevents the automatic model fallback, nothing else is affected

Trade-offs

  • May experience more 429 errors (Too Many Requests) if you hit rate limits
  • Response times might be slower during high usage periods, as the pro model is more resource-intensive

Installation

The patched Gemini CLI has been installed globally on your system:

  1. The original @google/gemini-cli package was uninstalled
  2. This patched version was built and linked globally

How to Revert

If you wish to revert to the original Gemini CLI:

# Uninstall the patched version
npm uninstall -g @google/gemini-cli

# Install the official version
npm install -g @google/gemini-cli

Technical Details

The patch specifically removes the model fallback logic in the retryWithBackoff function within packages/core/src/utils/retry.ts, which was triggered when:

  1. A user is authenticated with Google OAuth (personal account)
  2. Multiple 429 errors are encountered consecutively
  3. The onPersistent429 callback is available

The removed code would have downgraded from gemini-2.5-pro to gemini-2.5-flash and reset the retry counter. With the patch, the CLI will continue to respect rate limits but won't change models.