Add tests/fix bug for single planning step

[g-eoj]

Fixes issue hit while testing #300, which has flatten_messages_as_text set to True. Setting a planning_interval would fail.

Adds test coverage for all issues found.

agent = CodeAgent(model=mlx_model, tools=[UserInputTool()], add_base_tools=True, max_steps=3, planning_interval=1)
answer = agent.run("Hi!")

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[3], line 1
----> 1 answer = agent.run("Hi!")

File ~/.cache/uv/archive-v0/3MaM4p_ZnzhOGFENeDNZr/lib/python3.12/site-packages/smolagents/agents.py:565, in MultiStepAgent.run(self, task, stream, reset, single_step, images, additional_args)
    563     return self._run(task=self.task, images=images)
    564 # Outputs are returned only at the end as a string. We only look at the last step
--> 565 return deque(self._run(task=self.task, images=images), maxlen=1)[0]

File ~/.cache/uv/archive-v0/3MaM4p_ZnzhOGFENeDNZr/lib/python3.12/site-packages/smolagents/agents.py:586, in MultiStepAgent._run(self, task, images)
    584 try:
    585     if self.planning_interval is not None and self.step_number % self.planning_interval == 0:
--> 586         self.planning_step(
    587             task,
    588             is_first_step=(self.step_number == 0),
    589             step=self.step_number,
    590         )
    591     self.logger.log(
    592         Rule(
    593             f"[bold]Step {self.step_number}",
   (...)
    597         level=LogLevel.INFO,
    598     )
    600     # Run one step!

File ~/.cache/uv/archive-v0/3MaM4p_ZnzhOGFENeDNZr/lib/python3.12/site-packages/smolagents/agents.py:661, in MultiStepAgent.planning_step(self, task, is_first_step, step)
    648             message_prompt_facts = {
    649                 "role": MessageRole.SYSTEM,
    650                 "content": SYSTEM_PROMPT_FACTS,
    651             }
    652             message_prompt_task = {
    653                 "role": MessageRole.USER,
    654                 "content": f"""Here is the task:
   (...)
    658 Now begin!""",
    659             }
--> 661             answer_facts = self.model([message_prompt_facts, message_prompt_task]).content
    663             message_system_prompt_plan = {
    664                 "role": MessageRole.SYSTEM,
    665                 "content": SYSTEM_PROMPT_PLAN,
    666             }
    667             message_user_prompt_plan = {
    668                 "role": MessageRole.USER,
    669                 "content": USER_PROMPT_PLAN.format(
   (...)
    674                 ),
    675             }

File ~/.cache/uv/archive-v0/3MaM4p_ZnzhOGFENeDNZr/lib/python3.12/site-packages/smolagents/models.py:485, in MLXModel.__call__(self, messages, stop_sequences, tools_to_call_from, **kwargs)
    478 def __call__(
    479     self,
    480     messages: List[Dict[str, str]],
   (...)
    483     **kwargs,
    484 ) -> ChatMessage:
--> 485     completion_kwargs = self._prepare_completion_kwargs(
    486         flatten_messages_as_text=True,  # mlx-lm doesn't support vision models
    487         messages=messages,
    488         stop_sequences=stop_sequences,
    489         tools_to_call_from=tools_to_call_from,
    490         **kwargs,
    491     )
    492     messages = completion_kwargs.pop("messages")
    493     stop_sequences = completion_kwargs.pop("stop", [])

File ~/.cache/uv/archive-v0/3MaM4p_ZnzhOGFENeDNZr/lib/python3.12/site-packages/smolagents/models.py:263, in Model._prepare_completion_kwargs(self, messages, stop_sequences, grammar, tools_to_call_from, custom_role_conversions, convert_images_to_image_urls, flatten_messages_as_text, **kwargs)
    254 """
    255 Prepare parameters required for model invocation, handling parameter priorities.
    256 
   (...)
    260 3. Default values in self.kwargs
    261 """
    262 # Clean and standardize the message list
--> 263 messages = get_clean_message_list(
    264     messages,
    265     role_conversions=custom_role_conversions or tool_role_conversions,
    266     convert_images_to_image_urls=convert_images_to_image_urls,
    267     flatten_messages_as_text=flatten_messages_as_text,
    268 )
    270 # Use self.kwargs as the base configuration
    271 completion_kwargs = {
    272     **self.kwargs,
    273     "messages": messages,
    274 }

File ~/.cache/uv/archive-v0/3MaM4p_ZnzhOGFENeDNZr/lib/python3.12/site-packages/smolagents/models.py:228, in get_clean_message_list(message_list, role_conversions, convert_images_to_image_urls, flatten_messages_as_text)
    226 else:
    227     if flatten_messages_as_text:
--> 228         content = message["content"][0]["text"]
    229     else:
    230         content = message["content"]

TypeError: string indices must be integers, not 'str'

[albertvillanova]

Thank you for the fix, @g-eoj.

I am sorry, I missed your PR and fixed the issue with:

• Fix MultiStepAgent.planning_step message content #437

[g-eoj]

No worries.