When output_pydantic / response_model is set on a task, the response_model
was being forwarded to the LLM on every iteration of the tool-calling loop.
This causes many non-OpenAI LLMs to skip tool calls entirely because tools
and response_format parameters conflict.
Fix: Don't pass response_model to the LLM during tool-calling loop
iterations when the agent has tools. Structured output conversion is
already handled as post-processing via Task._export_output().
Changes:
- crew_agent_executor.py: _invoke_loop_react, _invoke_loop_native_tools,
_ainvoke_loop_react, _ainvoke_loop_native_tools now pass
response_model=None when tools are present
- experimental/agent_executor.py: call_llm_and_parse, call_llm_native_tools
similarly updated
- Added regression tests in TestResponseModelNotLeakedDuringToolLoop
Co-Authored-By: João <joao@crewai.com>
892fc4ad3b