[BUG] Expected `thinking` or `redacted_thinking`, but found `text` when using Claude Sonnet 3.7 Thinking

[noahzuiuc]

Description

litellm.exceptions.BadRequestError: litellm.BadRequestError: AnthropicException - {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.type: Expected thinkingorredacted_thinking, but found text. When thinkingis enabled, a finalassistantmessage must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disablethinking. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking"}}

This error is thrown during crew kickoff. I tested litellm on its own and everything works as expected. I have defined the llm backend for my agent as
LLM(model=f"{<provider>}/{<model_name>}", base_url=<base_url>, api_key=<api_key>, temperature=1, max_tokens=24000, thinking={ "type": "enabled", "budget_tokens": 16000 })

Steps to Reproduce

Run CrewAI with Claude Sonnet 3.7 Thinking as the backend

Expected behavior

The crew should run without errors

Screenshots/Code snippets

litellm.exceptions.BadRequestError: litellm.BadRequestError: AnthropicException - {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.type: Expected thinkingorredacted_thinking, but found text. When thinkingis enabled, a finalassistantmessage must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disablethinking. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking"}}

Operating System

Other (specify in additional context)

Python Version

3.12

crewAI Version

0.105.0

crewAI Tools Version

0.37.0

Virtual Environment

Venv

Evidence

ERROR:root:LiteLLM call failed: litellm.BadRequestError: AnthropicException - {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking"}}
Error during LLM call: litellm.BadRequestError: AnthropicException - {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking"}}
Traceback (most recent call last):
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/llms/anthropic/chat/handler.py", line 412, in completion
response = client.post(
^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/llms/custom_httpx/http_handler.py", line 553, in post
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/llms/custom_httpx/http_handler.py", line 534, in post
response.raise_for_status()
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/httpx/_models.py", line 763, in raise_for_status
raise HTTPStatusError(message, request=request, response=self)
httpx.HTTPStatusError: Client error '400 Bad Request' for url 'https://api.anthropic.com/v1/messages'
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/400

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/main.py", line 1846, in completion
response = anthropic_chat_completions.completion(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/llms/anthropic/chat/handler.py", line 427, in completion
raise AnthropicError(
litellm.llms.anthropic.common_utils.AnthropicError: {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking"}}

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/Users/noahz/Documents/lumyn/.venv/bin/run_crew", line 10, in
sys.exit(run())
^^^^^
File "/Users/noahz/Documents/lumyn/src/lumyn/main.py", line 140, in run
LumynCrew().crew().kickoff(inputs=inputs)
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/agent_analytics/instrumentation/traceloop/sdk/tracing/opentelemetry_instrumentation_crewai/patch.py", line 91, in traced_method
result = wrapped(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/crew.py", line 576, in kickoff
result = self._run_sequential_process()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/crew.py", line 683, in _run_sequential_process
return self._execute_tasks(self.tasks)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/crew.py", line 781, in _execute_tasks
task_output = task.execute_sync(
^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/agent_analytics/instrumentation/traceloop/sdk/tracing/opentelemetry_instrumentation_crewai/patch.py", line 91, in traced_method
result = wrapped(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/task.py", line 302, in execute_sync
return self._execute_core(agent, context, tools)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/task.py", line 366, in _execute_core
result = agent.execute_task(
^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/agent_analytics/instrumentation/traceloop/sdk/tracing/opentelemetry_instrumentation_crewai/patch.py", line 91, in traced_method
result = wrapped(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agent.py", line 254, in execute_task
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agent.py", line 243, in execute_task
result = self.agent_executor.invoke(
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agents/crew_agent_executor.py", line 112, in invoke
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agents/crew_agent_executor.py", line 102, in invoke
formatted_answer = self._invoke_loop()
^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agents/crew_agent_executor.py", line 160, in _invoke_loop
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agents/crew_agent_executor.py", line 140, in _invoke_loop
answer = self._get_llm_response()
^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agents/crew_agent_executor.py", line 210, in _get_llm_response
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/agents/crew_agent_executor.py", line 201, in _get_llm_response
answer = self.llm.call(
^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/llm.py", line 291, in call
response = litellm.completion(**params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/agent_analytics/instrumentation/traceloop/sdk/tracing/opentelemetry_instrumentation_litellm/patch.py", line 291, in traced_method
result = wrapped(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/utils.py", line 1154, in wrapper
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/utils.py", line 1032, in wrapper
result = original_function(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/main.py", line 3068, in completion
raise exception_type(
^^^^^^^^^^^^^^^
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/litellm_core_utils/exception_mapping_utils.py", line 2201, in exception_type
raise e
File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/litellm/litellm_core_utils/exception_mapping_utils.py", line 526, in exception_type
raise BadRequestError(
litellm.exceptions.BadRequestError: litellm.BadRequestError: AnthropicException - {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking"}}
An error occurred while running the crew: Command '['uv', 'run', 'run_crew']' returned non-zero exit status 1.

Possible Solution

None

Additional context

MacOS Sequoia

[Vidit-Ostwal]

Can you share how you are calling the litellm with Claude Sonnet 3.7 Thinking.

[noahzuiuc]

The call happens inside CrewAI, File "/Users/noahz/Documents/lumyn/.venv/lib/python3.12/site-packages/crewai/llm.py", line 291, in call
response = litellm.completion(**params)

[Vidit-Ostwal]

Yes, I want to understand, the normal litellm call which has worked.

I tested litellm on its own and everything works as expected. I have defined the llm backend for my agent as

[noahzuiuc]

import litellm

kwargs = {
    "model": f"{provider}/{model_name}",
    "api_key": api_key,
    "api_base": base_url,
    "temperature": 1,
    "max_tokens": 20000,
    "messages": [{ "role": "user", "content": "what is the fastest way to travel to every city in the US?" }],
    "thinking": { "type": "enabled", "budget_tokens": 16000 }
}

print(litellm.completion(**kwargs))

[berkybear]

FYI, this same error is happening in pure langgraph when a thinking node is invoked/handed off by another node. Basically whenever there is a list of messages present in the input that don't have the thinking block before the tool blocks seems to trigger this.

[github-actions]

This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[Vidit-Ostwal]

Commenting to keep this alive.