Fix A2A LLM parameter forwarding in InternalInstructor (Issue #3927)

[devin-ai-integration]

Summary

Fixes issue #3927 where A2A (Agent-to-Agent) delegation was losing LLM configuration parameters when checking if remote agents are relevant.

The bug was in InternalInstructor.to_pydantic() which only passed the model name to the instructor client, dropping critical parameters like api_key, api_base, and temperature. This caused A2A to fail when using LiteLLM proxy configurations.

Changes:

• Modified to_pydantic() to build a params dict and forward LLM configuration for LiteLLM instances
• Only forwards parameters when is_litellm=True to avoid breaking non-LiteLLM paths
• Filters out None values to prevent errors
• Prefers max_tokens over max_completion_tokens when both are present
• Removed unused type: ignore[import-untyped] comment

Review & Testing Checklist for Human

⚠️ IMPORTANT: No automated tests were added for this fix due to mocking complexity. Manual testing is critical.

• Test A2A delegation with custom LLM parameters - Verify that A2A works correctly when using LiteLLM with custom api_key, api_base, and temperature settings (this is the core issue from [BUG] A2A looses information about LLM provided when checking if A2A remote agent is relevant #3927)
• Test non-LiteLLM paths - Verify that regular OpenAI/Anthropic/etc. LLM usage still works correctly (no regression)
• Verify parameter completeness - Check that the forwarded parameter list matches what LiteLLM expects and that no critical parameters are missing
• Test edge cases - Verify behavior with None values, both max_tokens and max_completion_tokens set, and string LLM values
• Review CI test results - Ensure existing A2A tests pass and no regressions were introduced

Recommended Test Plan

1. Set up a LiteLLM proxy with custom configuration
2. Create a crew with A2A enabled using the proxy (custom api_base, api_key)
3. Trigger A2A delegation and verify it works correctly
4. Test without custom parameters to ensure backward compatibility
5. Test with non-LiteLLM providers (OpenAI, Anthropic) to ensure no regression

Notes

• The fix is designed to be backward compatible - it only adds parameters for LiteLLM instances
• The parameter list was derived from LLM._prepare_completion_params() (lines 588-635 in llm.py)
• Lint (ruff) and type checks (mypy) pass on the modified file
• Session: https://app.devin.ai/sessions/5d5b415b1eb2487480c58f8d48f97c9b
• 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