[BUG] Using DeepSeek R1 reasoning model from Bedrock and FileSessionManager in Strands agent causes 'ValidationException error that says user messages cannot contain reasoning content'

[aryan835-datainflexion]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.4.0

Python Version

3.13.3

Operating System

macOS 15.5

Installation Method

pip

Steps to Reproduce

1. Create agent using Bedrock DeepSeek R1 model and FileSessionManager or S3SessionManager

With FileSessionManager:

from strands import Agent
from strands.models import BedrockModel
from strands.session.file_session_manager import FileSessionManager

bedrock_model = BedrockModel(model_id="us.deepseek.r1-v1:0")
session_manager = FileSessionManager(session_id="test-session")

agent = Agent(model=bedrock_model, session_manager=session_manager)

With S3SessionManager:

from strands import Agent
from strands.models import BedrockModel
from strands.session.s3_session_manager import S3SessionManager

bedrock_model = BedrockModel(model_id="us.deepseek.r1-v1:0")
session_manager = S3SessionManager(session_id="test-session", bucket="test-bucket")

agent = Agent(model=bedrock_model, session_manager=session_manager)

2. Send multiple messages and get ValidationException error

Expected Behavior

Should not give error and remember previous messages and continue conversation.

Actual Behavior

Gives error "An error occurred (ValidationException) when calling the ConverseStream operation: User messages cannot contain reasoning content. Please remove the reasoning content and try again."

Additional Context

No response

Possible Solution

I think the FileSessionManager is appending the reasoning generated content as well to the user messages which is leading to this error.

Related Issues

No response

[aryan835-datainflexion]

I've raised PR #652 to solve this bug, please self-assign this issue to me, thanks!

[lekhnath]

Until the PR gets merged and released, I was able to get things working with the following customizations.
This basically intercepts and strips any reasoningContent from assistant messages at three points:

• When messages are intercepted via a hook
• When messages are appended to the session (Optional, if session storage is configured for agent)
• When sessions are initialized from storage (Optional, if session storage is configured for agent)

That way, the rest of the pipeline never sees reasoning traces.

from strands import Agent
from strands.types.content import Message
from strands.hooks import HookProvider, HookRegistry, MessageAddedEvent
from strands.session.repository_session_manager import RepositorySessionManager
from strands.session.file_session_manager import FileSessionManager

def reasoning_free_message(message: Message) -> Message:
    if message["role"] == "assistant":
        message["content"] = [
            content
            for content in message["content"]
            if "reasoningContent" not in content
        ]
    return message


class MessageInterceptor(HookProvider):
    def register_hooks(self, registry: HookRegistry) -> None:
        registry.add_callback(MessageAddedEvent, self.intercept_tool)

    def intercept_tool(self, event: MessageAddedEvent) -> None:
        reasoning_free_message(event.message)

# Optional 
class ReasoningFreeSessionManagerMixin(RepositorySessionManager):
    def append_message(self, message: Message, agent: Agent, **kwargs: Any) -> None:
        super().append_message(reasoning_free_message(message), agent, **kwargs)

    def initialize(self, agent: Agent, **kwargs: Any) -> None:
        super().initialize(agent, **kwargs)
        agent.messages = [reasoning_free_message(msg) for msg in agent.messages]


class CustomFileSessionManager(ReasoningFreeSessionManagerMixin, FileSessionManager):
    pass


session_manager = CustomFileSessionManager(
    session_id="<uuid>",
    storage_dir=path.join("<project-root-dir>", "<chat-sessions-dir>/"),
)

# End Optional

agent = Agent(
    session_manager=session_manager, # Optional
    hooks=[MessageInterceptor()],
    ...
)

[aryan835-datainflexion]

@zastrowm @dbschmigelski I've already raised PR #652 as was sent in the message above which fixes this, please can you check and review it. Thanks!

[dbschmigelski]

@zastrowm @dbschmigelski I've already raised PR #652 as was sent in the message above which fixes this, please can you check and review it. Thanks!

Yep absolutely will review! Thanks for raising this. Assigning here is just for us to track internally who is responsible for making sure the issue gets resolved, not necessarily who will raise the PR

[aryan835-datainflexion]

@dbschmigelski Alright got it thank you!

[dbschmigelski]

Hi just confirming that this is NOT SessionManager related as it is reproducible minimally even with

from strands import Agent
from strands.models import BedrockModel

bedrock_model = BedrockModel(model_id="us.deepseek.r1-v1:0")

agent = Agent(model=bedrock_model)
agent("tell me a story")
agent("tell me another")

[dbschmigelski]

Update, Bedrock is not going to change the behavior on their end so we will be proceeding with this fix.

[lekhnath]

@dbschmigelski

Yes, this is not related to SessionManager. If SessionManager is used, the effect reaches saving and loading of messages, so we should handle that condition optionally. If SessionManager is not used, intercepting the messages alone is enough.

[zastrowm]

This should be fixed by #652 in the next release; the next planned release is for Wed, tomorrow