Error with AWS Bedrock Claude 3.7 with thinking enabled

[dvejsada]

What happened?

When using Claude 3.7 from AWS Bedrock with thinking enabled, some conversation will throw error, usually after tool call. error:

[api/server/controllers/agents/client.js #sendCompletion] Operation aborted The model returned the following errors: messages.1.content.0: When providing `thinking` or `redacted_thinking` blocks, all preceding prompt sections (including system, tools, and thinking configuration) must match the parameters during the original request.

Error occurs on latest build, previously (last update approx. 2-3 weeks ago) it worked fine.

Version Information

latest:dev

Steps to Reproduce

1. Have agent with Claude 3.7 from AWS Bedrock
2. Have thinking enabled and tools connected to agent
3. Ask question requiring use of tools

What browsers are you seeing the problem on?

No response

Relevant log output

2025-04-22 07:08:32 error: [api/server/controllers/agents/client.js #sendCompletion] Unhandled error type The model returned the following errors: messages.1.content.0: When providing `thinking` or `redacted_thinking` blocks, all preceding prompt sections (including system, tools, and thinking configuration) must match the parameters during the original request.

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Not having the issue here

[dvejsada]

Thanks @danny-avila, unfortunately the issue keeps happening. Not at every message but every third/fourth message to the point that I had to disable thinking until the solution is found. Usually it happens after (successful) tool call. Do you have any idea where the issue might be?

[step-bauer]

Hello,
did you find any solution?

I face the same issue using langgraphs create_react_agent and providing a ChatBedrockConverse LLM with thinking enabled. It fails most times but some times it runs through calling up to 15 tools. Next time it fails after just some tool calls or even the first.
Once I remove the thinking it works fine.

I'm using
langchain 0.3.25
langchain-aws 0.2.23
langchain-community 0.3.23
langchain-core 0.3.59
langgraph 0.4.3
langgraph-checkpoint 2.0.25
langgraph-prebuilt 0.1.8

regards, Stephan

[dvejsada]

Unfortunately, no. I have tried again today after updating to the latest-dev container. The issue still persist, but does not occur every time. Below you can see similar question, similar tools use (both successful), but one results with error and the other will not.

@danny-avila Would you consider having another look at this please? Maybe it can be connected to handling of redacted_thinking blocks, if they are present?