Fix: Allow new Gemini models (gemini-3-pro-preview) to route to native provider

[devin-ai-integration]

Fix: Allow new Gemini models to route to native provider without constants

Summary

Fixes #3949 where LLM(model='google/gemini-3-pro-preview') was failing to route to the native Gemini provider because the model wasn't in the GEMINI_MODELS constants list.

Changes:

• Modified LLM._validate_model_in_constants() to allow any model starting with gemini-, gemma-, or learnlm- prefixes to route to the native Gemini provider, even if not explicitly listed in constants
• This prevents breakage when Google releases new Gemini models before the constants list is updated
• The Google SDK will surface a clear error if a model truly doesn't exist
• Added 4 comprehensive tests covering:
  • New preview models routing correctly (including gemini-3-pro-preview)
  • Case-insensitive prefix validation
  • Fallback to LiteLLM for non-matching models
  • Existing models still working correctly

Technical approach:
Instead of strictly requiring models to be in GEMINI_MODELS, the validation now:

1. First checks if the model is in the constants list (fast path for known models)
2. If not found, checks if the model name starts with known Gemini prefixes (case-insensitive)
3. Falls back to LiteLLM for models that don't match any prefix

This approach is more future-proof and aligns with how Azure models are handled (which also don't have strict validation).

Review & Testing Checklist for Human

⚠️ Risk Level: YELLOW - Unable to test locally due to environment issues, but changes follow existing patterns

• Verify the fix resolves issue [BUG] Gemini 3.0 Pro Preview doesn't work #3949: Test that LLM(model='google/gemini-3-pro-preview') now routes to GeminiCompletion instead of falling back to LiteLLM
• Test existing Gemini models: Verify that known models like google/gemini-2.0-flash-001 and google/gemini-1.5-pro still work correctly
• Test invalid model handling: Verify that LLM(model='google/unknown-model-xyz') properly falls back to LiteLLM or surfaces an appropriate error
• Review CI test results: Ensure all new tests pass, especially test_gemini_allows_new_preview_models_without_constants
• Consider long-term approach: Evaluate whether prefix-based validation is the right strategy vs. maintaining an exhaustive constants list

Test Plan

# Test 1: Verify the reported issue is fixed
llm = LLM(model='google/gemini-3-pro-preview')
assert llm.__class__.__name__ == "GeminiCompletion"
assert llm.provider == "gemini"

# Test 2: Verify existing models still work
llm = LLM(model='google/gemini-2.0-flash-001')
assert isinstance(llm, GeminiCompletion)

# Test 3: Verify invalid models fall back appropriately
llm = LLM(model='google/unknown-model-xyz')
assert llm.is_litellm == True

Notes

• Unable to run tests locally: The uv.lock file had parsing errors, preventing local test execution. Changes were verified through code analysis and pattern matching with existing code.
• Scope limited to Gemini: This change only affects Gemini model routing. OpenAI, Anthropic, and Bedrock providers maintain strict constant validation.
• Case-insensitive matching: The implementation uses .lower() for prefix matching, which should handle various capitalizations (e.g., "Gemini-3-Pro", "GEMINI-3-FLASH").

---

Link to Devin run: https://app.devin.ai/sessions/eb9b5faf184046608c23eb6de4e7a067
Requested by: João (joao@crewai.com)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring