Automatic function calling corrupts MCP func names with session-specific UUIDs

[maxkydanchuk]

Environment details
Programming language: Python
OS: macOS Darwin 24.6.0 / Linux
Language runtime version: Python 3.11 / 3.13
Package version: google-genai==1.53.0
MCP packages: mcp==1.9.4, databricks-mcp==0.3.0
Model: gemini-2.5-pro

Issue details
When using automatic function calling with MCP tools that have session-specific UUID identifiers, the Google GenAI SDK randomly corrupts func names during the function calling process. This causes KeyError exceptions when the SDK attempts to look up the corrupted name in the function_map.
The SDK corrupts func names in unpredictable ways during automatic function calling:

• Leading/trailing whitespace: Adds spaces before func names.
• UUID corruption: Inserts underscores, hyphens, or modifies characters within the UUID portion
• Character replacement: Changes valid hex digits to invalid characters

Examples
MCP session creates the name poll_response_01f0aa74324312bbaef790169f0ef410
Error from SDK

KeyError: 'poll_response_01f0aa74324312bbaef790169f0ef4-10'

Expected name: poll_response_01f0aa74324312bbaef790169f0ef410
Actual lookup name: poll_response_01f0aa74324312bbaef790169f0ef4-10

More detailed error:

Traceback (most recent call last):
  File "/usr/local/lib/python3.11/site-packages/google/genai/models.py", line 7054, in generate_content
    await _extra_utils.get_function_response_parts_async(
  File "/usr/local/lib/python3.11/site-packages/google/genai/_extra_utils.py", line 384, in get_function_response_parts_async
    func = function_map[func_name]
           │            └ 'poll_response_01f0aa74324312bbaef790169f0ef4-10'  # , <- CORRUPTED
           └ {
                'query_space_01f0aa74324312bbaef790169f0ef410': <McpToGenAiToolAdapter>,
                'poll_response_01f0aa74324312bbaef790169f0ef410': <McpToGenAiToolAdapter>  # <- CORRECT
              }

KeyError: 'poll_response_01f0aa74324312bbaef790169f0ef4-10'

Expected Behavior

• When using automatic function calling, func names from MCP's list_tools() should be used exactly as provided when:
• The LLM generates a function call in its response
• The SDK looks up the tool in function_map.
• The SDK calls session.call_tool(name, arguments)
• The SDK should perform exact string matching without any transformations.

Actual Behavior

• The SDK corrupts names between the LLM response and the function_map lookup, causing KeyError

Corruption patterns observed:

• Character replacement: poll_response_...ef410 -> poll_response_...ef4-10 (last 3 chars become -10)
• Leading whitespace: poll_response_... -> poll_response_... (space added at start)
• UUID mangling: poll_response_01f083de8626... -> poll_response_01f_083_de_8626... (underscores inserted randomly)
  Key finding: The corruption is non-deterministic - the same func name may work correctly on one call but get corrupted on the next.

Steps to Reproduce

1. Create an MCP session with a server that generates UUID-based tool names (e.g., databricks genie MCP)
2. Create a Gemini client with automatic function calling enabled
3. Make the first request passing the MCP session as a tool
4. Automatic function calling completes (may succeed or fail randomly)
5. Close the MCP session
6. Open a NEW MCP session (with new UUID-based tool names)
7. Build conversation history from the first request
8. Make the second request with the new MCP session and conversation history
9. Bug may occur - SDK corrupts the UUID portion of the func name
10. KeyError is raised when SDK tries to look up the corrupted func name

Expected Behavior

The SDK should preserve the exact func name returned by the current MCP session.

Actual Behavior

The SDK non-deterministically corrupts UUID portions of func names. The bug can occur on any request (1st, 2nd, 3rd, 4th) randomly:

• Sometimes fails on 1st request
• Sometimes succeeds on 1st but fails on 2nd
• Sometimes succeeds on 1st and 2nd but fails on 3rd
• Sometimes fails on 2nd and 3rd but succeeds on 4th
• Failure rate: approximately 30% of requests

Common corruption patterns include character splits with hyphens, leading spaces, or random underscores inserted into UUIDs.

[janasangeetha]

Hey @maxkydanchuk
Thanks for reaching out!
To assist us in effectively investigating and reproducing the issue, please provide a minimal, runnable code example.

[maxkydanchuk]

hey @just, here is the code

# You need to provide your own MCP session here

import asyncio
import os
import traceback

from databricks.sdk import WorkspaceClient
from databricks_mcp import DatabricksOAuthClientProvider
from dotenv import load_dotenv
from google import genai
from mcp import ClientSession
from mcp.client.streamable_http import streamablehttp_client

load_dotenv()


async def main():
    databricks_host = os.getenv("DATABRICKS_HOST")
    databricks_token = os.getenv("DATABRICKS_TOKEN")
    genie_space_id = os.getenv("GENIE_SPACE_ID")

    mcp_server_url = f"{databricks_host}/api/2.0/mcp/genie/{genie_space_id}"

    workspace_client = WorkspaceClient(
        host=databricks_host,
        token=databricks_token,
    )

    async with streamablehttp_client(
        mcp_server_url,
        auth=DatabricksOAuthClientProvider(workspace_client),
        timeout=300,
        sse_read_timeout=300,
    ) as (read, write, _):
        async with ClientSession(read, write) as session:
            await session.initialize()

            tools = await session.list_tools()
            print("MCP Tools:")
            for tool in tools.tools[:3]:
                print(f"  {tool.name}")

            client = genai.Client(api_key=os.getenv("GOOGLE_API_KEY"))

            config = genai.types.AutomaticFunctionCallingConfig(
                disable=False,
                maximum_remote_calls=60,
            )

            conversation_history = []

            query_1 = "your query 1"

            print("=" * 60)
            print(f"Query 1: {query_1}")
            print("=" * 60)

            try:
                response1 = await client.aio.models.generate_content(
                    model="gemini-2.5-pro",
                    contents=query_1,
                    config=genai.types.GenerateContentConfig(
                        automatic_function_calling=config,
                        tools=[session],
                    ),
                )

                print("Query 1 SUCCESS")
                print(f"Response: {response1.text[:200]}\n")

                conversation_history.append(
                    {
                        "role": "user",
                        "parts": [{"text": f"{query_1}"}],
                    }
                )

                conversation_history.append(
                    {"role": "model", "parts": [{"text": response1.text}]}
                )

            except KeyError as e:
                print("ERROR in Query 1!")
                print(f"Error: {e}")
                return

            query_2 = "your query 2"
            print("=" * 60)
            print(f"Query 2: {query_2}?")
            print("=" * 60)

            try:
                conversation_history.append(
                    {
                        "role": "user",
                        "parts": [{"text": f"{query_2}"}],
                    }
                )

                response2 = await client.aio.models.generate_content(
                    model="gemini-2.5-pro",
                    contents=conversation_history,
                    config=genai.types.GenerateContentConfig(
                        automatic_function_calling=config,
                        tools=[session],
                    ),
                )

                print("Query 2 SUCCESS")
                print(f"Response: {response2.text[:200]}\n")

                conversation_history.append(
                    {"role": "model", "parts": [{"text": response2.text}]}
                )

            except KeyError as e:
                print("Error in Query 2!")
                print(f"Error: {e}")
                return

            query_3 = "your query 3"
            print("=" * 60)
            print(f"Query 3: {query_3}?")
            print("=" * 60)

            try:
                conversation_history.append(
                    {
                        "role": "user",
                        "parts": [{"text": f"{query_3}"}],
                    }
                )

                response3 = await client.aio.models.generate_content(
                    model="gemini-2.5-pro",
                    contents=conversation_history,
                    config=genai.types.GenerateContentConfig(
                        automatic_function_calling=config,
                        tools=[session],
                    ),
                )

                print("Query 3 SUCCESS")
                print(f"Response: {response3.text[:200]}.\n")

            except KeyError as e:
                print("Error in Query 3!")
                print(f"Error: {e}")
                return


if __name__ == "__main__":
    # run multiple times to see the non-deterministic failure
    for i in range(5):
        print(f"\n{'=' * 60}")
        print(f"Attempt {i + 1}/5")
        print("=" * 60)
        try:
            asyncio.run(main())
        except Exception as e:
            print(f"Error: {type(e).__name__}: {str(e)}")

            traceback.print_exc()

Just ran this script and got an error from the 1st iteration

============================================================
Attempt 1/5
============================================================
MCP Tools:
  query_space_01f083de86261bde8e79144948aaa4ca
  poll_response_01f083de86261bde8e79144948aaa4ca

============================================================
Query 1: your query 1 (placeholder for my query)
============================================================
ERROR in Query 1!
Error: 'poll_response_01f083de86261bde8e7-9144948aaa4ca'

The UUID in poll response was updated, and - was added to the UUID: 7-9.

The 2nd test run was successful

3rd test run failed

MCP Tools:
  query_space_01f083de86261bde8e79144948aaa4ca
  poll_response_01f083de86261bde8e79144948aaa4ca
============================================================
Query 1: your query 1
============================================================
Query 1 SUCCESS
Response: response
============================================================
Query 2: your query 2
============================================================
Query 2 SUCCESS
Response: response

============================================================
Query 3: your query 3
============================================================
Error in Query 3!
Error: 'poll_response_01f_083_de86261bde8e79144948aaa_4ca'

poll response UUID is corrupted - poll_response_01f_083_de86261bde8e79144948aaa_4ca