[BUG] [Malformed input request] Issue calling bedrock after loading tools from mcp server

[paulognr]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

0.1.8

Python Version

3.10.5

Operating System

Windows 11 Enterprise

Installation Method

pip

Steps to Reproduce

1. Create a MCPClient using awslabs.dynamodb-mcp-server
2. Create an Agent with tools provided by the mcp server.
3. Try to use it

Expected Behavior

Bedrock is able to parse the toolSpec and the API call is successful.

Actual Behavior

Bedrock API call crashes because of malformed input request.

Additional Context

ToolSpec example:

Error from Bedrock API:
.venv\lib\site-packages\botocore\client.py", line 1061, in _make_api_call raise error_class(parsed_response, operation_name) botocore.errorfactory.ValidationException: An error occurred (ValidationException) when calling the ConverseStream operation: The model returned the following errors: Malformed input request: #/toolConfig/tools/8/toolSpec/inputSchema/json: extraneous key [$defs] is not permitted, please reformat your input and try again.

Possible Solution

Sanitize all tools before calling bedrock.

Related Issues

No response

[dbschmigelski]

Hi, we are taking a look at this. It appears we have sanitization logic that has been applied on our decorator based tool loading which was not applied to MCP tool transformation. https://github.com/strands-agents/sdk-python/blob/main/src/strands/tools/decorator.py#L178

However, it appears that Bedrock should be able to handle the '$ref' as this is part of the JSON Schema spec https://json-schema.org/draft/2020-12 and no additional constraints are stated on Bedrock https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ToolInputSchema.html

Additionally, it seems our current sanitization limits tool definitions to simple types erroneously. For example

from pydantic import Field
from typing import TypedDict
from strands import Agent, tool

def test_complex_pydantic_case():
    class TimeToLiveSpecification(TypedDict):
        AttributeName: str  # The name of the TTL attribute used to store the expiration time for items
        Enabled: bool  # Indicates whether TTL is enabled (true) or disabled (false) on the table

    @tool
    def calculate_ttl_given_specification(
        time_to_live_specification: TimeToLiveSpecification = Field(
            description="The specification used to determine if time to live is enabled"
        )
    ) -> str:
        return "IT WORKED"
    
    agent = Agent(tools=[calculate_ttl_given_specification])
    result = agent("Calculate the ttl for an enabled attribute called 'Test'")

appears to fail as the schema drops the reference resulting in an invalid json schema.

---

We need to look into two things

1. Why is $ref not being accepted by Bedrock. Given that schema has been published and used in the MCP server, I would assume this is not a Bedrock issue, but instead on the Strands side.
2. Remove blind dropping of $ref, in favor of either keeping it or performing a dereference step using something like https://pypi.org/project/jsonref/

{'properties': {'time_to_live_specification': {'$ref': '#/$defs/TimeToLiveSpecification', 'description': 'Parameter time_to_live_specification'}}, 'type': 'object'}

[omeraha]

Plus 1!
Had the exact same issue - once I changed from the fancy schema (based on a designated request class) the agent invoked the tool correctly.
My original code relied heavily on annotations:

from typing import Annotated, List, Literal, Optional
from pydantic import BaseModel, Field
from fastmcp import FastMCP
from fastmcp.server.dependencies import get_http_headers
from fastmcp.exceptions import ToolError

class WeatherInfoReq(BaseModel):
  location: Annotated[str, Field(description="Weather location")] = "nyc"
  ...

class WeatherInfoRes(BaseModel):
  ....

@mcp.tool(name="fetch_weather_info",
          description=description)
async def fetch_weather_info(request: WeatherInfoReq) -> WeatherInfoRes:
  ...

And when I tried to use it with a Strands agent, it was calling the tool incorrectly (which resulted in input validation errors on the server). When I replaced this with description in the function declaration, it worked:

@mcp.tool(name="fetch_weather_info",
          description=description)
async def fetch_weather_info(location: str, ...) -> WeatherInfoRes:

[pgrayy]

I had success with the following script:

from strands.models.bedrock import BedrockModel


model = BedrockModel()

messages = [
    {
        "role": "user",
        "content": [
            {
                "text": "What is the current weather in New York."
            },
        ],
    },
]
tool_specs = [
    {
        "name": "current_weather",
        "description": "Get weather for a location",
        "inputSchema": {
            "json": {
                "$defs": {
                    "city": {"type": "string", "description": "City name"},
                },
                "properties": {
                    "city": {"$ref": "#/$defs/city"},
                },
                "required": ["city"],
                "type": "object",
            },
        },
    },
]

for chunk in model.converse(messages, tool_specs):
    print(chunk)

The following is the exact request sent to bedrock through the boto3 bedrock-runtime client:

{
  "modelId": "us.anthropic.claude-3-7-sonnet-20250219-v1:0",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "text": "What is the current weather in New York."
        }
      ]
    }
  ],
  "system": [],
  "toolConfig": {
    "tools": [
      {
        "toolSpec": {
          "name": "current_weather",
          "description": "Get weather for a location",
          "inputSchema": {
            "json": {
              "$defs": {
                "city": {
                  "type": "string",
                  "description": "City name"
                }
              },
              "properties": {
                "city": {
                  "$ref": "#/$defs/city"
                }
              },
              "required": [
                "city"
              ],
              "type": "object"
            }
          }
        }
      }
    ],
    "toolChoice": {
      "auto": {}
    }
  },
  "inferenceConfig": {}
}

Were you also testing with "us.anthropic.claude-3-7-sonnet-20250219-v1:0" (this is the default in Strands).

[dbschmigelski]

Hi, resolving as #294 addressed the issue. Thanks for reporting this.