Fix Union type handling in Pydantic output generation

[devin-ai-integration]

Fix Union type handling in Pydantic output generation

Summary

Fixed a bug in generate_model_description where all Union types were incorrectly wrapped in Optional[], even when None was not part of the Union. This caused imprecise schema descriptions to be sent to LLMs when generating structured outputs.

Before:

• Union[str, int] → Optional[Union[str, int]] ❌
• Union[str, int, None] → Optional[Union[str, int]] ✓

After:

• Union[str, int] → Union[str, int] ✓
• Union[str, int, None] → Optional[Union[str, int]] ✓

Key Changes:

• Added logic to check if None is actually present in the Union before wrapping with Optional
• Added support for Python 3.10+ pipe syntax (int | str | None)
• Added comprehensive test coverage for Union type scenarios
• Updated existing test expectation to match corrected behavior

Fixes #3735

Review & Testing Checklist for Human

• Verify schema generation: Run pytest tests/test_union_types.py -vv to confirm all Union type tests pass (13 tests should pass)
• Check format consistency: Review the changes in converter.py to ensure the Union/Optional logic is correct and handles edge cases properly
• VCR cassette updates: Two tests (test_converter_with_llama3_1_model and test_converter_with_llama3_2_model) will fail with cassette mismatches because the schema format changed. These cassettes should be regenerated or the tests marked as expected to change
• End-to-end validation: Test with a real crew/task using output_pydantic with Union types to ensure LLMs correctly understand the new schema format

Recommended test plan:

from typing import Union
from pydantic import BaseModel
from crewai import Agent, Task, Crew

class SuccessResponse(BaseModel):
    status: str = "success"
    data: str

class ErrorResponse(BaseModel):
    status: str = "error"
    message: str

class Response(BaseModel):
    result: Union[SuccessResponse, ErrorResponse]

# Test that the agent can correctly return either variant
task = Task(
    description="Return a success response with data='test'",
    expected_output="A structured response",
    output_pydantic=Response,
    agent=your_agent
)

Notes

• The change improves precision in schema descriptions, which should help LLMs generate more accurate structured outputs
• All existing converter tests pass except for 2 VCR tests that need cassette regeneration
• Session: https://app.devin.ai/sessions/fe3b5f22bf65457abebb13cdc9d73362
• 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