[BUG] structured_output fails with Nova / OpenAI models when using file_read tool

[praven80]

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

1.9.0

Tools Package Version

0.2.8

Tools used

1. file_read

Python Version

3.12.9

Operating System

SageMaker JupyterLab Notebook

Installation Method

pip

Steps to Reproduce

When using structured_output with the file_read tool, the behavior is inconsistent across models:

• With Claude models, the call executes (though results hallucinate extra questions not present in the file).
• With Nova or OpenAI models, the execution fails with:

ValueError: Model returned stop_reason: end_turn instead of "tool_use".

Code to Reproduce

from strands import Agent
from strands_tools import file_read
from strands.models import BedrockModel
from typing import List
from pydantic import BaseModel, Field

class Question(BaseModel):
    question_number: int = Field(..., description="The number of the question.")
    question_text: str = Field(..., description="The text of the question.")

class QuestionList(BaseModel):
    questions: List[Question] = Field(..., description="A list of questions from the file.")

# Using OpenAI model (also reproducible with Nova)
model = BedrockModel(model_id='openai.gpt-oss-20b-1:0')
json_agent = Agent(model=model, tools=[file_read])

def read_questions_from_file(file_path_in: str):
    """
    Reads questions from a file and returns them in a structured JSON format.
    """
    prompt = f"""
    The file {file_path_in} has a numbered list of questions.
    Read the file and return a list of questions in the specified JSON structure.
    """
    
    response = json_agent.structured_output(QuestionList, prompt)
    return response

file_path_in = "questions.txt"
response = read_questions_from_file(file_path_in)
print(response.model_dump_json(indent=2))

Create a simple questions.txt file with:

1. What is the capital of France?
2. What is 2+2?
3. Who wrote Hamlet?

Run the above script using openai.gpt-oss-20b-1:0 or nova.* models.

Observe error.

Expected Behavior

The tool should execute consistently across all Bedrock models.

Either structured output should work the same way as with Claude, or if unsupported, the error should be caught and surfaced clearly.

Actual Behavior

With Claude models → executes, but hallucinates.

With Nova/OpenAI models → raises error:

ValueError: Model returned stop_reason: end_turn instead of "tool_use".

Additional Context

This seems related to how different model providers handle tool invocation and structured_output.

Request: Please clarify whether structured_output + tools is expected to work across all models (Claude, Nova, OpenAI), or only a subset.

Possible Solution

No response

Related Issues

No response

[afarntrog]

Thanks for reporting this issue, @praven80 ! So this is related to the current approach Strands takes towards Structured Output. The structured output method, as of now, should be viewed as a helper function rather than invoking a normal agent. Since it's just a helper function - we don't get the features that come as part of a regular agent invocation and that's why the tools aren't working.

To accomplish both, we can update your referenced code to the following:

from strands import Agent
from strands_tools import file_read
from strands.models import BedrockModel
from typing import List
from pydantic import BaseModel, Field

class Question(BaseModel):
    question_number: int = Field(..., description="The number of the question.")
    question_text: str = Field(..., description="The text of the question.")

class QuestionList(BaseModel):
    questions: List[Question] = Field(..., description="A list of questions from the file.")

# Using OpenAI model (also reproducible with Nova)
model = BedrockModel(model_id='openai.gpt-oss-20b-1:0')
json_agent = Agent(model=model, tools=[file_read])

def read_questions_from_file(file_path_in: str):
    """
    Reads questions from a file and returns them in a structured JSON format.
    """
    prompt = f"""
    The file {file_path_in} has a numbered list of questions.
    Read the file and return a list of questions in the specified JSON structure.
    """
    json_agent(prompt) # <--- this is the regular agent call that will work with tools etc
    response = json_agent.structured_output(QuestionList, "parse the the messages to the QuestionList")
    return response

file_path_in = "questions.txt"
response = read_questions_from_file(file_path_in)
print(response.model_dump_json(indent=2))

Note. We are working on revamping structured output so that it is more native to a regular agent invocation. Tools and everything should work as regular. See the following issue: strands-agents/sdk-python#891