[BUG] Problems with OpenRouter thinking models

[JCaratt]

Description

Hello to everyone.
I have several orchestrated agents for the generation of academic proposals. I have tested many models (GLM, Qwen 3.6 Plus, GPT 5.4) and everything works fine, but when I use Sonnet 4.5, Gemini 3.1 Pro Preview, or Opus 4.5, these errors usually occur.
It is worth noting that I am running this in Google Colab, using OpenRouter APIs.
I would greatly appreciate your help, since Anthropic and Google are very good at academic writing.

Steps to Reproduce

1. Set up orchestrated agents for academic proposal generation in Google Colab.
2. Configure OpenRouter API with model set to anthropic/claude-sonnet-4-5 or anthropic/claude-opus-4-5.
3. Send a multi-agent chained request for academic writing generation.
4. Observe the error in the API response.

Expected behavior

The model should respond successfully to the academic proposal generation
prompts, as other models (GLM, Qwen 3.5 Plus, GPT 4.5) do under the
same setup and configuration.

Screenshots/Code snippets

Operating System

Ubuntu 20.04

Python Version

3.10

crewAI Version

1.14.2

crewAI Tools Version

1.14.2

Virtual Environment

Conda

Evidence

---

ValueError Traceback (most recent call last)
/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in execute_task(self, task, context, tools)
760 else:
--> 761 result = self._execute_without_timeout(task_prompt, task)
762

27 frames
/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _execute_without_timeout(self, task_prompt, task)
827 dict[str, Any],
--> 828 self.agent_executor.invoke(
829 {

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in invoke(self, inputs)
224 if self.ask_for_human_input:
--> 225 formatted_answer = self._handle_human_feedback(formatted_answer)
226

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _handle_human_feedback(self, formatted_answer)
1581 provider = get_provider()
-> 1582 return provider.handle_feedback(formatted_answer, self)
1583

/usr/local/lib/python3.12/dist-packages/crewai/core/providers/human_input.py in handle_feedback(self, formatted_answer, context)
186
--> 187 return self._handle_regular_feedback(formatted_answer, feedback, context)
188

/usr/local/lib/python3.12/dist-packages/crewai/core/providers/human_input.py in _handle_regular_feedback(self, current_answer, initial_feedback, context)
262 context.messages.append(context._format_feedback_message(feedback))
--> 263 answer = context._invoke_loop()
264 feedback = self._prompt_input(context.crew)

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _invoke_loop(self)
306 if use_native_tools:
--> 307 return self._invoke_loop_native_tools()
308

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _invoke_loop_native_tools(self)
552 handle_unknown_error(PRINTER, e, verbose=self.agent.verbose)
--> 553 raise e
554 finally:

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _invoke_loop_native_tools(self)
466 if has_reached_max_iterations(self.iterations, self.max_iter):
--> 467 formatted_answer = handle_max_iterations_exceeded(
468 None,

/usr/local/lib/python3.12/dist-packages/crewai/utilities/agent_utils.py in handle_max_iterations_exceeded(formatted_answer, printer, messages, llm, callbacks, verbose)
301 )
--> 302 raise ValueError("Invalid response from LLM call - None or empty.")
303

ValueError: Invalid response from LLM call - None or empty.

During handling of the above exception, another exception occurred:

ValueError Traceback (most recent call last)
/tmp/ipykernel_6575/2898347091.py in <cell line: 0>()
62 print()
63
---> 64 resultado = propuesta_crew.kickoff()
65
66 print("\n✅ Generación completada. Procediendo a crear el .docx...")

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in kickoff(self, inputs, input_files, from_checkpoint)
932
933 if self.process == Process.sequential:
--> 934 result = self._run_sequential_process()
935 elif self.process == Process.hierarchical:
936 result = self._run_hierarchical_process()

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _run_sequential_process(self)
1375 def _run_sequential_process(self) -> CrewOutput:
1376 """Executes tasks sequentially and returns the final output."""
-> 1377 return self._execute_tasks(self.tasks)
1378
1379 def _run_hierarchical_process(self) -> CrewOutput:

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _execute_tasks(self, tasks, start_index, was_replayed)
1472
1473 context = self._get_context(task, task_outputs)
-> 1474 task_output = task.execute_sync(
1475 agent=exec_data.agent,
1476 context=context,

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in execute_sync(self, agent, context, tools)
522 """Execute the task synchronously."""
523 self.start_time = datetime.datetime.now()
--> 524 return self._execute_core(agent, context, tools)
525
526 @Property

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _execute_core(self, agent, context, tools)
819 self.end_time = datetime.datetime.now()
820 crewai_event_bus.emit(self, TaskFailedEvent(error=str(e), task=self))
--> 821 raise e # Re-raise the exception after emitting the event
822 finally:
823 clear_task_files(self.id)

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _execute_core(self, agent, context, tools)
726 self, TaskStartedEvent(context=context, task=self)
727 )
--> 728 result = agent.execute_task(
729 task=self,
730 context=context,

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in execute_task(self, task, context, tools)
772 raise e
773 except Exception as e:
--> 774 result = self._handle_execution_error(e, task, context, tools)
775
776 return self._finalize_task_execution(task, result)

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _handle_execution_error(self, e, task, context, tools)
682 """
683 self._check_execution_error(e, task)
--> 684 return self.execute_task(task, context, tools)
685
686 async def _handle_execution_error_async(

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in execute_task(self, task, context, tools)
772 raise e
773 except Exception as e:
--> 774 result = self._handle_execution_error(e, task, context, tools)
775
776 return self._finalize_task_execution(task, result)

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _handle_execution_error(self, e, task, context, tools)
681 Result from retried execution.
682 """
--> 683 self._check_execution_error(e, task)
684 return self.execute_task(task, context, tools)
685

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _check_execution_error(self, e, task)
661 ),
662 )
--> 663 raise e
664
665 def _handle_execution_error(

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in execute_task(self, task, context, tools)
759 )
760 else:
--> 761 result = self._execute_without_timeout(task_prompt, task)
762
763 except TimeoutError as e:

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _execute_without_timeout(self, task_prompt, task)
826 result = cast(
827 dict[str, Any],
--> 828 self.agent_executor.invoke(
829 {
830 "input": task_prompt,

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in invoke(self, inputs)
210
211 try:
--> 212 formatted_answer = self._invoke_loop()
213 except AssertionError:
214 if self.agent.verbose:

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _invoke_loop(self)
305
306 if use_native_tools:
--> 307 return self._invoke_loop_native_tools()
308
309 return self._invoke_loop_react()

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _invoke_loop_native_tools(self)
551 continue
552 handle_unknown_error(PRINTER, e, verbose=self.agent.verbose)
--> 553 raise e
554 finally:
555 self.iterations += 1

/usr/local/lib/python3.12/dist-packages/crewai/agents/agent_builder/base_agent.py in _invoke_loop_native_tools(self)
465 try:
466 if has_reached_max_iterations(self.iterations, self.max_iter):
--> 467 formatted_answer = handle_max_iterations_exceeded(
468 None,
469 printer=PRINTER,

/usr/local/lib/python3.12/dist-packages/crewai/utilities/agent_utils.py in handle_max_iterations_exceeded(formatted_answer, printer, messages, llm, callbacks, verbose)
300 color="red",
301 )
--> 302 raise ValueError("Invalid response from LLM call - None or empty.")
303
304 formatted = format_answer(answer=answer)

ValueError: Invalid response from LLM call - None or empty.

Possible Solution

No idea

Additional context

NA

[JCaratt]

I'm using Google Colab and litellm 1.83.10

[PHclaw]

For the OpenRouter thinking model issue:

OpenRouter's reasoning models return a different response format with a reasoning_content field. CrewAI likely doesn't handle this extra field.

Fix: Add handling for the reasoning response:

def parse_response(self, response):
    content = response.choices[0].message.content
    reasoning = getattr(response.choices[0].message, 'reasoning_content', None)
    
    if reasoning:
        # Include reasoning in metadata but not in output
        self._last_reasoning = reasoning
    
    return content

Also check that the model name is passed correctly: openrouter/meta-llama/llama-3.1-8b-instruct:free