gemini-cli - 💡(How to fix) Fix Default fallback policy chain missing gemini-2.5-flash-lite — users lose 1,000 RPD of available capacity [1 comments, 2 participants]

Summary

The default fallback policy chain includes only gemini-2.5-pro → gemini-2.5-flash. The gemini-2.5-flash-lite model is missing entirely. When both Pro and Flash quotas are exhausted, the CLI fails hard—even though Flash-Lite has 10x Pro's daily quota and is the intended last-resort model for free-tier users.

Environment

• OS: Windows 11 Pro (10.0.26200)
• Node.js: v22.19.0
• gemini-cli: 0.41.2
• Auth: Google OAuth (gemini login) — AuthType.USE_GEMINI (free tier)

Root Cause Analysis

1. Incomplete Policy Chain

The default fallback chain is defined in chunk-6DSAZLFF.js (~line 270025) as:

definePolicy({ model: DEFAULT_GEMINI_MODEL }),
definePolicy({ model: DEFAULT_GEMINI_FLASH_MODEL, isLastResort: true })

The DEFAULT_GEMINI_FLASH_LITE_MODEL (gemini-2.5-flash-lite) is not present in the chain.

2. Quota Impact

When Pro (100) + Flash (250) are exhausted, the CLI fails despite 1,000 Flash-Lite requests remaining. The tool appears "broken" to the user when ~74% of their daily capacity is actually untouched.

Steps to Reproduce

1. Auth: Log in via gemini login.
2. Exhaust Quota: Exhaust RPD limits on both gemini-2.5-pro and gemini-2.5-flash.
3. Execute: Run gemini -y -p "hello"
4. Observe: QUOTA_EXHAUSTED error, exit code 1. No fallback to Flash-Lite.
5. Verify: Run gemini -y --model gemini-2.5-flash-lite -p "hello". This succeeds, proving Flash-Lite quota was available.

Expected Behaviour

The fallback chain should be: gemini-2.5-pro → gemini-2.5-flash → gemini-2.5-flash-lite.

Actual Behaviour

The chain stops at Flash. The CLI exits with QUOTA_EXHAUSTED despite 1,000 Flash-Lite RPD remaining.

Suggested Fix

Update the policy chain to include the Flash-Lite model as the final resort:

definePolicy({ model: DEFAULT_GEMINI_MODEL }),
definePolicy({ model: DEFAULT_GEMINI_FLASH_MODEL }),
definePolicy({ model: DEFAULT_GEMINI_FLASH_LITE_MODEL, isLastResort: true })

Model IDs confirmed working via --model flag:

• gemini-2.5-pro
• gemini-2.5-flash
• gemini-2.5-flash-lite

Note: This issue compounds with a related bug (reported separately) where QUOTA_EXHAUSTED does not trigger fallback in headless -p mode at all. Together, these bugs effectively block free-tier users from CLI access prematurely.