How to handle 'Function is not found in the tools_dict' Error

[kxcode]

sometimes Gemini calls function with wrong tool name, will cause a ValueError, how can I handle this Error

line 447, in _run_one_step_async
    async for event in agen:
  File "/python3.12/site-packages/google/adk/flows/llm_flows/base_llm_flow.py", line 539, in _postprocess_async
    async for event in agen:
  File "/python3.12/site-packages/google/adk/flows/llm_flows/base_llm_flow.py", line 632, in _postprocess_handle_function_calls_async
    if function_response_event := await functions.handle_function_calls_async(
                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/python3.12/site-packages/google/adk/flows/llm_flows/functions.py", line 198, in handle_function_calls_async
    return await handle_function_call_list_async(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/python3.12/site-packages/google/adk/flows/llm_flows/functions.py", line 246, in handle_function_call_list_async
    function_response_events = await asyncio.gather(*tasks)
                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/python3.12/site-packages/google/adk/flows/llm_flows/functions.py", line 280, in _execute_single_function_call_async
    tool, tool_context = _get_tool_and_context(
                         ^^^^^^^^^^^^^^^^^^^^^^
  File "/python3.12/site-packages/google/adk/flows/llm_flows/functions.py", line 651, in _get_tool_and_context
    raise ValueError(
ValueError: Function add_-plan_step is not found in the tools_dict

[pinkponk]

+1

[Drinect]

I have the same issue.

I was thinking of using the 'after_model_callback' and creating a custom function that searches for the model's provided function definitions and checks the available functions (stored in the callback context or fetched from the adk context if that is possible. Don't know yet). If no match is found, you can do a closest string match (i.e., Levenshtein distance, or semantic similarity) or correct it with another model.

But to be honest, it would be nice if the adk handles this. At least with a simple retry mechanism, so the model is informed about its mistake.

[pinkponk]

yes it is crucial to be able to catch this value error. as the situation is now for me my entire agent flow just stops when this happens. Very annoying. I want to be able to catch this in a plugin.

[nicodelatorre7]

+1 Same error here using model="gemini-2.5-flash-lite" and model="gemini-2.5-flash".

idk how to catch this LLM errors

[NikhilKalloli]

Even I'm facing the same issue:

[NikhilKalloli]

Try adding this in the agent instructions. It has worked for my case

# Tool Selection Guide

**Tool Name Verification (Critical):** 
Before calling any tool, *always* meticulously verify its exact name against the `tool_code` definitions provided in the execution context. 
Use the precise, full name as it appears there, including any hyphens or prefixes (e.g., `firebase-firebase_list_projects`, not `firebase_list_projects`). 
Do not abbreviate or simplify tool names. 
Calling a tool with an incorrect name will result in an error."

[reyolivo]

This bug appears to have been introduced in version 1.9.0 of google-adk. I reverted to 1.8.0 and it now works fine again. This is not ideal though.

[EDIT] - we are actually seeing this error in older versions as well (1.7.0 and 1.8.0)

[NikhilKalloli]

This bug appears to have been introduced in version 1.9.0 of google-adk. I reverted to 1.8.0 and it now works fine again. This is not ideal though.

[EDIT] - we are actually seeing this error in older versions as well (1.7.0 and 1.8.0)

Yeah, it's a universal problem from the google adk. Hope they fix it soon

[iamsrinathks]

Same here. Would be great if adk can handle this

[reyolivo]

any update on this? We were able to reduce the frequency of the error by including tool specific instructions in the system prompt...but still not 100% resolved, and feels like a bad pattern to continue. Is there at least a better way to catch an incorrect tool from being called, and have the agent retry? An example would be very helpful

[NikhilKalloli]

Try switching to Vertex AI (using project ID and region rather than API key)
This helped in my case