fix(litellm): Pass response_format to LiteLLM completion for structured outputs

[dee-walia20]

Description

Summary of changes: Fix LiteLLM model to properly pass response_format to the underlying LiteLLM library when structured outputs are enabled.

Related issues: N/A (or link to any related issues)

Motivation and context:
When using LiteLLM model with output_schema, structured_outputs=True, and supports_native_structured_outputs=True, the agent was still falling back to JSON mode via prompt injection. This was because the response_format parameter was never passed to litellm.completion().

Environment or dependencies: No changes required

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Model update (Addition or modification of models)

Changes Made

1. Added explicit structured output flags with documentation:

   • Flags remain False by default (safe for providers like Anthropic, Ollama)
   • Added clear comments explaining when/how users should enable them
   • Users explicitly enable for models that support it (OpenAI, Gemini, etc.)

2. Added _get_response_format_param() helper method:

   • Converts Pydantic BaseModel classes to LiteLLM-compatible JSON schema format
   • Passes through dict formats unchanged
   • Returns None for None input

3. Added _add_additional_properties_false() helper method:

   • Recursively adds additionalProperties: false to all object schemas
   • Required by OpenAI's strict mode (without this, API returns 400 error)
   • Handles nested objects, arrays, $defs, definitions, anyOf/oneOf/allOf

4. Updated all invoke methods (invoke, invoke_stream, ainvoke, ainvoke_stream):

   • Added code to convert and add response_format to completion_kwargs
   • Only passes response_format when supports_native_structured_outputs=True
   • This prevents errors with providers like Anthropic that don't support it

5. Added tests for structured output functionality

Important Design Decision

The supports_native_structured_outputs flag is kept as False by default because:

• LiteLLM supports 100+ providers, not all support structured outputs
• Anthropic models don't support response_format parameter (causes 400 error)
• Ollama has limited/no structured output support
• Safe default prevents breaking existing usage

The response_format parameter is only passed to the API when supports_native_structured_outputs=True.
This ensures providers like Anthropic continue to work (falling back to JSON mode via prompt).

Users enable it explicitly when they know their model supports it:

# For OpenAI/Gemini models - enable structured outputs
model = LiteLLM(
    id='gpt-4o',
    supports_native_structured_outputs=True,
)

# For Anthropic models - leave as default (uses prompt-based JSON mode)
model = LiteLLM(
    id='claude-sonnet-4-20250514',
    # supports_native_structured_outputs=False  # Default
)

Checklist

• Adherence to standards: Code complies with Agno's style guidelines and best practices.
• Formatting and validation: You have run ./scripts/format.sh and ./scripts/validate.sh
• Self-review completed: A thorough review has been performed
• Documentation: Docstrings added for the new helper method
• Examples and guides: N/A (existing examples should now work correctly)
• Tested in a clean environment: Changes tested and verified
• Tests: Tests added for new functionality

Testing

Tested with:

• OpenAI models via LiteLLM (gpt-4o)
• Custom providers via SAP AI Core (sap/gpt-4o-mini)

Before fix:

LiteLLM: Params passed to completion() {..., 'response_format': None, ...}
WARNING  Failed to parse cleaned JSON: ...

After fix (with supports_native_structured_outputs=True):

LiteLLM: Params passed to completion() {..., 'response_format': {'type': 'json_schema', ...}, ...}