[BUG] crewai 0.120.0 doesn’t think

[MoShiha]

Description

Litellm expects “thinking” but got text instead error is returned when using crewai 0.120.0 with Claude AI sonet 3.7 and reasoning is enabled.

Steps to Reproduce

!pip install “crewai==0.120.0”

Instantiate LLM with Claude Sonnet 3.7 as your model and set reasoning to any value (e.g., “high”)

Once the crew is kicked off it fails with the above mentioned error.

Expected behavior

Run normally as previous stable version. I was able to get the code up and running again by reverting to a specific version of crewai (0.114.0) from GitHub repo.

Screenshots/Code snippets

Cannot share (internal to my company)

Operating System

Other (specify in additional context)

Python Version

3.10

crewAI Version

0.120.0

crewAI Tools Version

NA

Virtual Environment

Conda

Evidence

You can try it yourself. It’s few lines of code.

Possible Solution

Investigate the difference between the 0.120.0 and 0.114.0

Additional context

Also tried on AWS SageMaker and the same problem occurred

[MoShiha]

The following crewai version is working without any problem with Claude AI sonet 3.7
{"url": "https://github.com/crewAIInc/crewAI.git", "vcs_info": {"commit_id": "16eb4df5568b685be92ae4e305523046ce686425", "vcs"

[Vidit-Ostwal]

Can you try setting verbose=True at crew and agent level and share the error trace once.

[MoShiha]

[Vidit-Ostwal]

Can you share the litellm version you are using, a while back this issue was raised, and this was resolved by updating the litellm verison?

#2323

Although this shouldn't be an issue, but still double-checking it to be sure.

[MoShiha]

True.. it was an issue then got resolved and it is back again .. I didn’t have the time to investigate the changes that have been committed since the one I shared above.

Anyway here is a brief overview of the packages on my environment:

Name: litellm
Version: 1.60.2

Name: pydantic
Version: 2.11.4

Name: langchain
Version: 0.3.21

[Vidit-Ostwal]

True.. it was an issue then got resolved and it is back again .. I didn’t have the time to investigate the changes that have been committed since the one I shared above.

Anyway here is a brief overview of the packages on my environment:

Name: litellm Version: 1.60.2

Name: pydantic Version: 2.11.4

Name: langchain Version: 0.3.21

ahh, I see the issue here, can you bump up the litellm version to 1.68.0
pip install --upgrade litellm==1.68.0
That's the version pinned in the latest uv.lock file.

[MoShiha]

strange! This same setup is working with the crewai version 0.114.0 I install from GitHub as mentioned above. Let me give this a try and update you.

[MoShiha]

When upgraded to crewai 0.121 and litellm 1.68.0, running the same code is throwing Error: maximum recursion depth exceeded described in another issue.

Reverted back to crewai 0.120.0 and litllm 1.68.0 but this didn't resolve the issue and gave the error I raised this issue about

[MoShiha]

───────────────────────────────────────────────── Crew Failure ──────────────────────────────────────────────────╮
│ │
│ Crew Execution Failed │
│ Name: crew │
│ ID: f91b3095-68c3-4ef9-80f7-dfc6326cb9d1 │
│ │
│ │
╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
Failed to complete the workflow!
Error: litellm.BadRequestError: BedrockException - {"message":"The model returned the following errors:
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"}

[Vidit-Ostwal]

Duplicated: #2323

Bumping up the version should have reslved it.