BUG: CodeAgent outputs empty steps that ChatCompletionMessage(content=None,...)

[daviden1013]

Problem
agent.run often generates empty steps (steps with ChatCompletionMessage(content=None)). Results in repeating error message:

Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.

Steps to reproduce
Please provide a minimal, self-contained, and reproducible example of the bug.

1. Start vLLM OpenAI compatible server

vllm serve openai/gpt-oss-120b \
   --tensor-parallel-size 4 \
   --enable-prefix-caching \
   --gpu-memory-utilization 0.90 \
   --max-num-seqs 2 \
   --max-model-len 120000 \
   --reasoning-parser openai_gptoss \
   --tool-call-parser openai \
   --enable-auto-tool-choice

2. Define a simple CodeAgent and run it.

from smolagents import CodeAgent, OpenAIModel

model = OpenAIModel(model_id="openai/gpt-oss-120b",
                    api_key="", 
                    api_base="http://localhost:8000/v1",
                    reasoning_effort="low")

agent = CodeAgent(tools=[], 
                  model=model, executor_type='local', 
                  additional_authorized_imports = ["pandas", "numpy.*", "tableone", "scipy", "scipy.*", "sklearn", "sklearn.*", "statsmodels", "statsmodels.*", "matplotlib", "matplotlib.*", "PIL", "PIL.*"],
                  max_steps=10,
                  stream_outputs=False)

task = "hi"

res = agent.run(task, reset=False)

Actual behavior and error logs
A clear and concise description of what actually happened. Please include the full traceback if an exception was raised.

╭─────────────────────────────────────────────────────────────────────────────────── New run ────────────────────────────────────────────────────────────────────────────────────╮
│                                                                                                                                                                                │
│ hi                                                                                                                                                                             │
│                                                                                                                                                                                │
╰─ OpenAIModel - openai/gpt-oss-120b ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 1 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
Your code snippet is invalid, because the regex pattern <code>(.*?)</code> was not found in it.
Here is your code snippet:
Hello! How can I assist you today?</code>
Make sure to include code with the correct pattern, for instance:
Thoughts: Your thoughts
<code>
# Your python code here
</code>
Make sure to provide correct code blobs.
[Step 1: Duration 0.38 seconds| Input tokens: 2,089 | Output tokens: 26]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 2 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 2: Duration 0.51 seconds| Input tokens: 4,309 | Output tokens: 105]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 3 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 3: Duration 0.54 seconds| Input tokens: 6,582 | Output tokens: 189]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 4 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 4: Duration 0.48 seconds| Input tokens: 8,908 | Output tokens: 262]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 5 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 5: Duration 0.70 seconds| Input tokens: 11,287 | Output tokens: 369]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 6 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 6: Duration 0.73 seconds| Input tokens: 13,719 | Output tokens: 479]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 7 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 7: Duration 0.76 seconds| Input tokens: 16,204 | Output tokens: 594]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 8 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 8: Duration 0.81 seconds| Input tokens: 18,742 | Output tokens: 716]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 9 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 9: Duration 0.78 seconds| Input tokens: 21,333 | Output tokens: 833]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 10 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Error in code parsing:
expected string or bytes-like object, got 'NoneType'
Make sure to provide correct code blobs.
[Step 10: Duration 0.61 seconds| Input tokens: 23,977 | Output tokens: 922]
Reached max steps.
[Step 11: Duration 0.34 seconds| Input tokens: 24,712 | Output tokens: 950]

Expected behavior
The agent should generate string contents for Thought and Code.

Environment:
Please complete the following information:

• OS: Ubuntu 22.04.5
• Python version: 3.11.8
• Package version: (output of pip show smolagents)

Name: smolagents
Version: 1.23.0
Summary: 🤗 smolagents: a barebones library for agents. Agents write python code to call tools or orchestrate other agents.
Home-page:
Author:
Author-email: Aymeric Roucher <aymeric@hf.co>
License:
Location: /home/daviden1013/miniconda3/envs/NLP/lib/python3.11/site-packages
Requires: huggingface-hub, jinja2, pillow, python-dotenv, requests, rich

Additional context (optional)
I looked into the model.py OpenAIModel.generate method. I think the bug is related to stop_sequences sent to the self.client.chat.completions.create (line 1742)

completion_kwargs = self._prepare_completion_kwargs(
            messages=messages,
            stop_sequences=stop_sequences,
            response_format=response_format,
            tools_to_call_from=tools_to_call_from,
            model=self.model_id,
            custom_role_conversions=self.custom_role_conversions,
            convert_images_to_image_urls=True,
            **kwargs,
        )
        self._apply_rate_limit()
        response = self.retryer(self.client.chat.completions.create, **completion_kwargs)

I tried replacing the stop_sequences=stop_sequences with stop_sequences=[] and the issue is gone. I am not sure if this causes other issues.

I use this debugging class, and the issue is gone.

class DebuggingOpenAIModel(OpenAIModel):
    def generate(
    self,
    messages: list[ChatMessage | dict],
    stop_sequences: list[str] | None = None,
    response_format: dict[str, str] | None = None,
    tools_to_call_from: list[Tool] | None = None,
    **kwargs,
    ) -> ChatMessage:
        completion_kwargs = self._prepare_completion_kwargs(
            messages=messages,
            stop_sequences=[], # Remove stop sequences
            response_format=response_format,
            tools_to_call_from=tools_to_call_from,
            model=self.model_id,
            custom_role_conversions=self.custom_role_conversions,
            convert_images_to_image_urls=True,
            **kwargs,
        )
        self._apply_rate_limit()
        response = self.retryer(self.client.chat.completions.create, **completion_kwargs)
        print("Raw model response:", response)
        content = response.choices[0].message.content
        if stop_sequences is not None and not self.supports_stop_parameter:
            content = remove_content_after_stop_sequences(content, stop_sequences)
        return ChatMessage(
            role=response.choices[0].message.role,
            content=content,
            tool_calls=response.choices[0].message.tool_calls,
            raw=response,
            token_usage=TokenUsage(
                input_tokens=response.usage.prompt_tokens,
                output_tokens=response.usage.completion_tokens,
            ),
        )

model = DebuggingOpenAIModel(model_id="openai/gpt-oss-120b",
                    api_key="", 
                    api_base="http://localhost:8000/v1",
                    reasoning_effort="low")

agent = CodeAgent(tools=[], 
                  model=model, executor_type='local', 
                  additional_authorized_imports = ["pandas", "numpy.*", "tableone", "scipy", "scipy.*", "sklearn", "sklearn.*", "statsmodels", "statsmodels.*", "matplotlib", "matplotlib.*", "PIL", "PIL.*"],
                  max_steps=10,
                  stream_outputs=False)

task = "hi"

res = agent.run(task, reset=False)

---

Checklist

• I have searched the existing issues and have not found a similar bug report.
• I have provided a minimal, reproducible example.
• I have provided the full traceback of the error.
• I have provided my environment details.
• I am willing to work on this issue and submit a pull request. (optional)

[daviden1013]

After further exploration, I found it was the "</code>" in the stop_sequences causing the problem.

I manually set stop_sequences=["Observation:", "Calling tools:", "</code>"] and it triggers the bug.

class DebuggingOpenAIModel(OpenAIModel):
    def generate(
    self,
    messages: list[ChatMessage | dict],
    stop_sequences: list[str] | None = None,
    response_format: dict[str, str] | None = None,
    tools_to_call_from: list[Tool] | None = None,
    **kwargs,
    ) -> ChatMessage:
        completion_kwargs = self._prepare_completion_kwargs(
            messages=messages,
            stop_sequences=["Observation:", "Calling tools:", "</code>"],
            response_format=response_format,
            tools_to_call_from=tools_to_call_from,
            model=self.model_id,
            custom_role_conversions=self.custom_role_conversions,
            convert_images_to_image_urls=True,
            **kwargs,
        )
        self._apply_rate_limit()
        response = self.retryer(self.client.chat.completions.create, **completion_kwargs)
        print("stop_sequences:", stop_sequences)
        print("Raw model response:", response)
        content = response.choices[0].message.content
        if stop_sequences is not None and not self.supports_stop_parameter:
            content = remove_content_after_stop_sequences(content, stop_sequences)
        return ChatMessage(
            role=response.choices[0].message.role,
            content=content,
            tool_calls=response.choices[0].message.tool_calls,
            raw=response,
            token_usage=TokenUsage(
                input_tokens=response.usage.prompt_tokens,
                output_tokens=response.usage.completion_tokens,
            ),
        )

But if I set: stop_sequences=["Observation:", "Calling tools:"]. The bug is gone most of the time. If I further remove the remove_content_after_stop_sequences, the bug is gone.

class DebuggingOpenAIModel(OpenAIModel):
    def generate(
    self,
    messages: list[ChatMessage | dict],
    stop_sequences: list[str] | None = None,
    response_format: dict[str, str] | None = None,
    tools_to_call_from: list[Tool] | None = None,
    **kwargs,
    ) -> ChatMessage:
        completion_kwargs = self._prepare_completion_kwargs(
            messages=messages,
            stop_sequences=["Observation:", "Calling tools:"],
            response_format=response_format,
            tools_to_call_from=tools_to_call_from,
            model=self.model_id,
            custom_role_conversions=self.custom_role_conversions,
            convert_images_to_image_urls=True,
            **kwargs,
        )
        self._apply_rate_limit()
        response = self.retryer(self.client.chat.completions.create, **completion_kwargs)
        print("stop_sequences:", stop_sequences)
        print("Raw model response:", response)
        content = response.choices[0].message.content

        return ChatMessage(
            role=response.choices[0].message.role,
            content=content,
            tool_calls=response.choices[0].message.tool_calls,
            raw=response,
            token_usage=TokenUsage(
                input_tokens=response.usage.prompt_tokens,
                output_tokens=response.usage.completion_tokens,
            ),
        )

I found the </code> was added in agents.py line 1631

stop_sequences = ["Observation:", "Calling tools:"]
        if self.code_block_tags[1] not in self.code_block_tags[0]:
            # If the closing tag is contained in the opening tag, adding it as a stop sequence would cut short any code generation
            stop_sequences.append(self.code_block_tags[1])

I don't understand the reason for adding </code> to stop_sequences. Please help.