[BUG] Intermediate agent messages missing when using LiteLLM Proxy Server with GPT OSS, Nova Premier

[roeetal]

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

1.8.0

Python Version

3.13.5

Operating System

macOS 15.7

Installation Method

pip

Steps to Reproduce

from strands import Agent
from strands.models.litellm import LiteLLMModel
from strands import tool

@tool
def weather_forecast(city: str, days: int = 3) -> str:
    """Get weather forecast for a city.

    Args:
        city: The name of the city
        days: Number of days for the forecast
    """
    return f"Weather forecast for {city} for the next {days} days..."


litellm_model = LiteLLMModel(model_id="openai.gpt-oss-120b-1:0", params={
    "aws_profile_name": 'default',
})
agent = Agent(model=litellm_model, tools=[weather_forecast])
response = agent("can you call the weather tool for NYC")

litellm_proxy_model = LiteLLMModel(model_id="openai/openai.gpt-oss-120b-1:0", params={
    "api_base": "http://0.0.0.0:4000",
    "api_key": "key"
})
proxy_agent = Agent(model=litellm_proxy_model, tools=[weather_forecast])
proxy_response = proxy_agent("can you call the weather tool for NYC")

Expected Behavior

agent.messages and response are comparable to proxy_agent.messages and proxy_response taking non-determinism into account.

Actual Behavior

agent.messages:

[{'role': 'user',
  'content': [{'text': 'can you call the weather tool for NYC'}]},
 {'role': 'assistant',
  'content': [{'text': 'Here’s the latest weather forecast for New\u202fYork City (NYC):'},
   {'toolUse': {'toolUseId': 'tooluse_8XQsOOSyT5Oss7QB3jNvNw',
     'name': 'weather_forecast',
     'input': {'city': 'New York City'}}}]},
 {'role': 'user',
  'content': [{'toolResult': {'toolUseId': 'tooluse_8XQsOOSyT5Oss7QB3jNvNw',
     'status': 'success',
     'content': [{'text': 'Weather forecast for New York City for the next 3 days...'}]}}]},
 {'role': 'assistant',
  'content': [{'text': 'Here’s a quick 3‑day outlook for New\u202fYork\u202fCity:\n\n| Day | Condition | High [/](https://file+.vscode-resource.vscode-cdn.net/) Low | Precip. Chance | Wind |\n|-----|-----------|------------|----------------|------|\n| **Today** (Apr\u202f30) | Partly cloudy | 78\u202f°F [/](https://file+.vscode-resource.vscode-cdn.net/) 65\u202f°F | 20\u202f% | 8‑12\u202fmph from the SW |\n| **Tomorrow** (May\u202f1) | Mostly sunny | 80\u202f°F [/](https://file+.vscode-resource.vscode-cdn.net/) 66\u202f°F | 10\u202f% | 6‑10\u202fmph from the W |\n| **Day\u202f3** (May\u202f2) | Light showers & breezy | 75\u202f°F [/](https://file+.vscode-resource.vscode-cdn.net/) 64\u202f°F | 55\u202f% | 10‑15\u202fmph from the NW |\n\n*Temperatures are in Fahrenheit (°F). If you need the forecast in Celsius or want a longer‑range outlook, just let me know!*'}]}]

response:

AgentResult(stop_reason='end_turn', message={'role': 'assistant', 'content': [{'text': 'Here’s a quick 3‑day outlook for New\u202fYork\u202fCity:\n\n| Day | Condition | High [/](https://file+.vscode-resource.vscode-cdn.net/) Low | Precip. Chance | Wind |\n|-----|-----------|------------|----------------|------|\n| **Today** (Apr\u202f30) | Partly cloudy | 78\u202f°F [/](https://file+.vscode-resource.vscode-cdn.net/) 65\u202f°F | 20\u202f% | 8‑12\u202fmph from the SW |\n| **Tomorrow** (May\u202f1) | Mostly sunny | 80\u202f°F [/](https://file+.vscode-resource.vscode-cdn.net/) 66\u202f°F | 10\u202f% | 6‑10\u202fmph from the W |\n| **Day\u202f3** (May\u202f2) | Light showers & breezy | 75\u202f°F [/](https://file+.vscode-resource.vscode-cdn.net/) 64\u202f°F | 55\u202f% | 10‑15\u202fmph from the NW |\n\n*Temperatures are in Fahrenheit (°F). If you need the forecast in Celsius or want a longer‑range outlook, just let me know!*'}]}, metrics=EventLoopMetrics(cycle_count=2, tool_metrics={'weather_forecast': ToolMetrics(tool={'toolUseId': 'tooluse_8XQsOOSyT5Oss7QB3jNvNw', 'name': 'weather_forecast', 'input': {'city': 'New York City'}}, call_count=1, success_count=1, error_count=0, total_time=0.0004942417144775391)}, cycle_durations=[20.01394510269165], traces=[<strands.telemetry.metrics.Trace object at 0x117e00ad0>, <strands.telemetry.metrics.Trace object at 0x117bc2ea0>], accumulated_usage={'inputTokens': 390, 'outputTokens': 441, 'totalTokens': 831}, accumulated_metrics={'latencyMs': 0}), state={})

proxy_agent.messages:

[{'role': 'user',
  'content': [{'text': 'can you call the weather tool for NYC'}]},
 {'role': 'assistant',
  'content': [{'text': '{\n  "city": "New York City",\n  "days": [0]\n}'}]}]

proxy_response:

AgentResult(stop_reason='end_turn', message={'role': 'assistant', 'content': [{'text': '{\n  "city": "New York City",\n  "days": [0]\n}'}]}, metrics=EventLoopMetrics(cycle_count=1, tool_metrics={}, cycle_durations=[19.30731701850891], traces=[<strands.telemetry.metrics.Trace object at 0x117b0b020>], accumulated_usage={'inputTokens': 15, 'outputTokens': 18, 'totalTokens': 33}, accumulated_metrics={'latencyMs': 0}), state={})

Additional Context

LiteLLM proxy server routes requests to bedrock with a model list defined like:

"model_list": [
    {
        "model_name": "openai/openai.gpt-oss-120b-1:0",
        "litellm_params": {
            "model": "bedrock/openai.gpt-oss-120b-1:0",
            "aws_region_name": "us-west-2",
            "aws_profile_name": "default",
        }
    }
],

Possible Solution

No response

Related Issues

No response

[Ratish1]

Hello,I have successfully reproduced it, identified the root cause, and found a working solution.

The problem occurs because the litellm library requires a global flag, litellm.use_litellm_proxy = True, to be set when communicating with a proxy server via api_base. Without this flag, litellm cannot resolve the custom model name and raises a BadRequestError, which is what we were seeing.

The issue is completely resolved by setting litellm.use_litellm_proxy = True before the agent call.

This appears to be highly related to Issue #862 , which also describes a failure when using LiteLLMModel with a proxy, just with a different function (structured_output). @dbschmigelski is already working on it and will get it fixed soon.

[roeetal]

In the first agent call the request is routed directly to Bedrock. In the second agent call, the request is successfully routed to the LiteLLM proxy. However, the response from the proxy is hallucinated, it appears the tools are not available to the model...

[Ratish1]

@roeetal Thank you for the clarification. That makes sense. My fix with litellm.use_litellm_proxy = True solves the initial connection and provider resolution error, but you are right that there seems to be a deeper issue where the tool definitions are not being correctly passed through the proxy, causing the model to hallucinate. This seems like the core bug. I'm happy to help investigate further if you can point me in the right direction.

[dbschmigelski]

Hi @roeetal and @Ratish1 , I believe in the second example the proxy may not actually be getting used.

Can you check out https://strandsagents.com/latest/documentation/docs/user-guide/concepts/model-providers/litellm/#option-1-use-use_litellm_proxy-parameter

The SDK PR which addresses this is #808.

[roeetal]

Hey @dbschmigelski this proxy is being used. using the openai/ prefix for the model works similar to the litellm_proxy/ prefix described in the link you shared. I also verified the proxy server did receive the request in its logs.