fix: stop output_pydantic from leaking into tool-calling loop (#5472)

[mahasarabesh]

Fixes #5472

What's wrong

Since v1.9.0, Task(output_pydantic=...) and Task(output_json=...) get mapped onto response_model in the agent executor. This means every LLM call in the tool loop sends both tools and response_format at the same time.

OpenAI kinda handles this, but vLLM, Gemini, Anthropic, and anything going through LiteLLM's InternalInstructor -- they all treat response_format as higher priority. The LLM returns structured JSON immediately and never calls any tools.

The post-processing path (Task._export_output() -> convert_to_model()) already handles converting raw output to Pydantic models after the loop. That's how it worked before v1.9.0 and it works fine.

What this PR does

1. agent/core.py -- Only task.response_model (explicit opt-in) gets passed as response_model. output_pydantic and output_json stay in post-processing where they belong.

2. crew_agent_executor.py -- Pass response_model=None in _invoke_loop_native_tools (sync + async). When tools are present, structured output shouldn't compete with them.

3. experimental/agent_executor.py -- Same fix for call_llm_native_tools().

4. Tests -- 4 regression tests covering output_pydantic, output_json, explicit response_model, and _update_executor_parameters.

How I tested

• All 4 new tests pass
• Full test_agent.py suite passes (84 passed, pre-existing failures from missing optional deps unchanged)

Note

There are two existing PRs for this -- #5680 (stops the mapping at source) and #5767 (suppresses during loop iterations). This PR combines both approaches: fix the mapping AND defensively suppress during tool loops as a safety net.

Summary by CodeRabbit

• Bug Fixes

  • Executor response-model is now derived only from an explicit task response model and no longer falls back to legacy output formats.
  • Disabled response-model parsing during native tool invocation to prevent unnecessary validation of tool-call responses.

• Tests

  • Added regression tests to ensure correct response-model behavior across task variants and executor updates.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 0e9636f4-06e4-4b04-9c24-49a6f3ac3adf

📥 Commits

Reviewing files that changed from the base of the PR and between fc9f4c2 and 3ecf73d.

📒 Files selected for processing (4)

• lib/crewai/src/crewai/agent/core.py
• lib/crewai/src/crewai/agents/crew_agent_executor.py
• lib/crewai/src/crewai/experimental/agent_executor.py
• lib/crewai/tests/agents/test_agent.py

🚧 Files skipped from review as they are similar to previous changes (4)

• lib/crewai/src/crewai/experimental/agent_executor.py
• lib/crewai/src/crewai/agents/crew_agent_executor.py
• lib/crewai/tests/agents/test_agent.py
• lib/crewai/src/crewai/agent/core.py

---

📝 Walkthrough

Walkthrough

This PR prevents Task.output_pydantic/output_json from setting the agent executor's response_model and ensures native tool-calling paths call get_llm_response with response_model=None so structured-output validation does not interfere with tool invocation.

Changes

Response Model Separation from Tool-Calling Loop

Layer / File(s)	Summary
Agent executor response_model initialization and updates lib/crewai/src/crewai/agent/core.py	create_agent_executor and _update_executor_parameters now derive executor response_model only from explicit task.response_model (or None), removing the fallback to task.output_pydantic and task.output_json.
Native tool execution response_model fixes lib/crewai/src/crewai/agents/crew_agent_executor.py, lib/crewai/src/crewai/experimental/agent_executor.py	Native tool-calling paths in _invoke_loop_native_tools, _ainvoke_loop_native_tools, and call_llm_native_tools now pass response_model=None to get_llm_response to prevent structured output validation from interfering with tool invocation.
Regression test coverage lib/crewai/tests/agents/test_agent.py	New TestOutputPydanticDoesNotLeakIntoResponseModel class with four test methods verifying that output_pydantic/output_json do not set executor response_model, explicit response_model is still respected, and executor parameter updates with different task types produce correct final values.

Sequence Diagram(s)

sequenceDiagram
  participant AgentExecutor
  participant LLM
  participant ToolRunner
  AgentExecutor->>LLM: get_llm_response(..., response_model=None)
  LLM->>AgentExecutor: return tool-call list OR final-answer
  AgentExecutor->>ToolRunner: run indicated tools (if any)
  ToolRunner->>AgentExecutor: tool outputs
  AgentExecutor->>LLM: (final formatting or post-process) optional

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Suggested labels

size/M

Suggested reviewers

• greysonlalonde

Poem

🐰 I nibbled code beneath the moon,
Where JSON schemas hummed a tune.
I freed the tools to run and play,
Then shaped the answer on its way.
Hooray — the crew can work today!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main fix: preventing output_pydantic from leaking into the tool-calling loop, which is the primary objective of this PR.
Linked Issues check	✅ Passed	The PR fully addresses the objectives from issue #5472: removes output_pydantic/output_json mapping to response_model in agent/core.py, passes response_model=None during tool loops in crew_agent_executor.py and experimental/agent_executor.py, and adds comprehensive regression tests covering all scenarios.
Out of Scope Changes check	✅ Passed	All changes are directly related to fixing the output_pydantic leakage issue: agent/core.py fixes response_model assignment, executor files fix tool-loop response_model passing, and tests verify the fix. No unrelated changes detected.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}