Schema Issue: "schema specified other fields alongside any_of"

[Acevedan117]

I'm currently running into an issue with schemas for ADK version 1.0.0, the error message is as follows:

raise ClientError(status_code, response_json, response)
google.genai.errors.ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Unable to submit request because do_update_security_alert functionDeclaration parameters.status schema specified other fields alongside any_of. When using any_of, it must be the only field set.

I ran into this issue on version 0.4.0 and 0.5.0 and was able to get around it by passing in MCP tools to this function:

def make_tools_gemini_compatible(tools):
"""
This function makes the schema compatible with Gemini/Vertex AI API
It is only needed when API used is Gemini and model is other than 2.5 models
It is however needed for ALL models when API used is VertexAI
"""
for tool in tools:
for key in tool.mcp_tool.inputSchema.keys():
if key == "properties":
for prop_name in tool.mcp_tool.inputSchema["properties"].keys():
if "anyOf" in tool.mcp_tool.inputSchema["properties"][prop_name].keys():
if (tool.mcp_tool.inputSchema["properties"][prop_name]["anyOf"][0]["type"] == "array"):
tool.mcp_tool.inputSchema["properties"][prop_name]["type"] = tool.mcp_tool.inputSchema["properties"][prop_name]["anyOf"][0]["items"]["type"]
else:
tool.mcp_tool.inputSchema["properties"][prop_name]["type"] = tool.mcp_tool.inputSchema["properties"][prop_name]["anyOf"][0]["type"]
tool.mcp_tool.inputSchema["properties"][prop_name].pop("anyOf")

return tools

Unfortunately, now with version 1.0.0 I am unable to use this workaround. Is there a working example of another workaround that can be used to make tools gemini compatible that works with 1.0.0?

[Acevedan117]

Ended up figuring this out shortly after making this post. For anyone else running into this issue, my solution was the following:

import os
import logging
from google.adk.agents.llm_agent import LlmAgent
from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset, StdioServerParameters

Your exact logic adapted for ADK 1.0.0

def make_tools_gemini_compatible(tools):
"""
Makes the schema compatible with Gemini/Vertex AI API.
This version works with ADK v1.0.0 MCPTool instances.
"""
for tool in tools:
try:
input_schema = tool._mcp_tool.inputSchema # _mcp_tool exists internally in v1.0.0
except AttributeError:
continue

    for key in input_schema.keys():
        if key == "properties":
            for prop_name in input_schema["properties"].keys():
                prop = input_schema["properties"][prop_name]
                if "anyOf" in prop:
                    if prop["anyOf"][0]["type"] == "array":
                        prop["type"] = prop["anyOf"][0]["items"]["type"]
                    else:
                        prop["type"] = prop["anyOf"][0]["type"]
                    prop.pop("anyOf")
return tools

Subclass MCPToolset to apply patch lazily and safely

class PatchedMCPToolset(MCPToolset):
async def get_tools(self, ctx=None):
tools = await super().get_tools(ctx)

    model = 'gemini-2.0-flash-001'
    model_version = model.split("-")[1]
    if float(model_version) < 2.5:
        logging.warning("Patching tools for Gemini compatibility")
        tools = make_tools_gemini_compatible(tools)

    return tools

MCPToolset with custom patching

_allowed_path = os.path.join(
r"[FILE PATH]"
)

toolset = PatchedMCPToolset(
connection_params=StdioServerParameters(
command='uv',
args=[
"--directory", _allowed_path,
"run",
"--env-file",
r"[FILE PATH TO .env]",
"server.py"
],
)
)

root_agent = LlmAgent(
model='gemini-2.0-flash-001',
name='google_security_assistant',
instruction=(
'Helps user investigate security issues using Google Secops SIEM. '
'All authentication actions are automatically approved. If the query '
'is about a SOAR case try to provide a backlink to the user. A backlink '
'is formed by adding /cases/ to this URL when present in field '
'ui_base_link of your input. If the user asks with only ? or are you there? '
'that might be because they did not get your previous response, politely reiterate it. '
'Try to respond in markdown whenever possible.'
),
tools=[toolset],
)