Builtin tools that have tool_context as a parameter don't get a description from the doc-string

[efunneko]

Describe the bug
If you add a built-in tool (python function) to your agent and that tool contains one of the 'ignored parameters', such as tool_context, in its parameters, the ADK will rebuild that function without that parameter. When it does this, it neglects to copy the doc attribute to the new function. Then when the tool declaration is built for the LLM API call, its description is empty. This is particularly bad if you are calling an Anthropic model on Bedrock because it demands a description.

To Reproduce
Steps to reproduce the behavior:

1. Create a very basic agent with a single 'python' tool type
2. Create a python function that has a 'tool_context' parameter
3. Run an inference and look at the tool list that is sent to the LLM (your tool will have 'description: ""'

Expected behavior
The tool list will contain a description for each tool

Desktop (please complete the following information):

• OS: [e.g. iOS]
• Python version(python -V): 3.11.10
• ADK version(pip show google-adk): 0.4.0

Possible Fix

def build_function_declaration(
    func: Union[Callable, BaseModel],
    ignore_params: Optional[list[str]] = None,
    variant: Literal['GOOGLE_AI', 'VERTEX_AI', 'DEFAULT'] = 'GOOGLE_AI',
) -> types.FunctionDeclaration:
  signature = inspect.signature(func)
  new_func = None
  if not ignore_params:
    ignore_params = []
  for name, _ in signature.parameters.items():
    if name in ignore_params:
      should_update_signature = True
      break
  if should_update_signature:
    new_params = [
        param
        for name, param in signature.parameters.items()
        if name not in ignore_params
    ]
    if isinstance(func, type):
      fields = {
          name: (param.annotation, param.default)
          for name, param in signature.parameters.items()
          if name not in ignore_params
      }
      new_func = create_model(func.__name__, **fields)
    else:
      new_sig = signature.replace(parameters=new_params)
      new_func = FunctionType(
          func.__code__,
          func.__globals__,
          func.__name__,
          func.__defaults__,
          func.__closure__,
      )
      new_func.__signature__ = new_sig

      # FIX
      new_func.__doc__ = func.__doc__

  return (
      from_function_with_options(func, variant)
      if not should_update_signature
      else from_function_with_options(new_func, variant)
  )

[efunneko]

PR submitted: #552

[efunneko]

@hangfei - just curious about an update for this and whether you need more from me. I keep having to repatch the ADK library to allow Sonnet models to work. Also, even Gemini models suffer in that the description of the tool is not included. The model does an okay job of guessing at the purpose based on the function and parameter names, but it still is a large limitation to lose the description of the tool

[efunneko]

I just want to raise the priority of this issue. This prevents the use of functions requiring tool_context to work with Anthropic models and silently makes the performance worse for Gemini ones. The system performs much better with this very small fix in place.

[seanzhou1023]

fixed in 7fdc6b4