fix: prevent output_pydantic from injecting tool schema when LLM lacks function calling

[giulio-leone]

Summary

Fixes #4695

When a Task has output_pydantic set and the LLM does not support native function calling (e.g., Ollama models), the executor correctly falls back to the ReAct text-based loop. However, response_model was still forwarded to the LLM call, causing InternalInstructor to inject the Pydantic schema as a tool via instructor.from_litellm(completion) in Mode.TOOLS mode — which models without function-calling support cannot handle.

Root Cause

_invoke_loop_react() and _ainvoke_loop_react() both passed self.response_model to get_llm_response() / aget_llm_response(). This triggered the instructor path in LLM._process_call_params() (line 1153), which wraps the LLM with instructor.from_litellm(completion) using the default Mode.TOOLS — injecting the Pydantic model as a native tool regardless of whether the model actually supports function calling.

Fix

Pass response_model=None in both _invoke_loop_react() and _ainvoke_loop_react(). The task-level convert_to_model() in task.py (line 1021-1040) already handles text → Pydantic conversion after the ReAct loop finishes, so instructor-based extraction is unnecessary and harmful in this code path.

Changes

• lib/crewai/src/crewai/agents/crew_agent_executor.py: Set response_model=None in both sync and async ReAct loops
• lib/crewai/tests/agents/test_react_response_model.py: Added 2 regression tests:
  • Verifies response_model=None is passed to get_llm_response in ReAct mode
  • Verifies _invoke_loop_react is selected (not native tools) when function calling is unsupported

Test Results

• 2 new tests pass ✅
• 29 existing sync executor tests pass ✅
• No regressions

---

Note

Medium Risk
Changes how output_pydantic is handled in the ReAct execution path, which can affect structured output behavior for models without native tool/function calling. Risk is mitigated by added regression tests covering the selection of ReAct mode and ensuring response_model is not forwarded.

Overview
Prevents output_pydantic from being forwarded to the LLM call when the executor falls back to the ReAct text loop, by forcing response_model=None in both sync and async ReAct invocations so non-function-calling models don’t receive an injected tool schema.

Adds regression tests ensuring ReAct mode is chosen when supports_function_calling() is false and that get_llm_response is called with response_model=None in this path.

^{Written by Cursor Bugbot for commit 16be9ac. This will update automatically on new commits. Configure here.}

⚠️ This reopens #4828 which was accidentally closed due to fork deletion.