Describe the bug
CodeAgent and planning with ToolCallingAgent are not passing add_generation_template=True when applying the chat template to the input messages of the model.

Note the <|endoftext|> token at the end of the prompt.

# End of CodeAgent prompt
<|system|> ...
Now Begin! If you solve the task correctly, you will receive a reward of $1,000,000.<|end|><|user|>New task:
Could you get me the title of the page at url "https://huggingface.co/blog"?<|end|><|endoftext|>

# End of planning step initial_facts prompt
<|user|> ...
Now Begin!<|end|><|endoftext|>

Not sure if it is a bug or this is how you intended for it to work, but from what I saw it confuses the model and just seems wrong.
The problem is that when calling the model, the condition to add the generation prompt is:

prompt_tensor = self.tokenizer.apply_chat_template(
    ...
    add_generation_prompt=True if tools_to_call_from else False,
)

And the tools_to_call_from are not passed to the call when using CodeAgent or when planning.

Code to reproduce the error

from smolagents import ToolCallingAgent, TransformersModel, CodeAgent

import torch


if __name__ == '__main__':
    model_id = 'microsoft/Phi-4-mini-instruct'
    model = TransformersModel(model_id=model_id, torch_dtype=torch.float16, device_map='cuda')
    agent = CodeAgent(tools=[], model=model, add_base_tools=True)
    agent.run('Could you get me the title of the page at url "https://huggingface.co/blog"?', reset=True)

Expected behavior
add_generation_prompt=True by default for all model calls?

Packages version:
smolagents==1.11.0

@albertvillanova saw you were involved in PRs around this issue, maybe you can shed some light on this matter. Thanks!