error receiving from stream: unexpected end of JSON input when using MCP

[incoming-th]

Describe the bug

When running the docker cagent with the MCP DuckDuckGo toolset, the agent or MCP server fails to respond correctly or produces errors.

Version affected

version: "2"
agents:
  root:
    model: groq
    description: A helpful AI assistant
    instruction: |
      You are a knowledgeable assistant that helps users with various tasks.
      Be helpful, accurate, and concise in your responses. 
    
    toolsets:
      - type: mcp
        ref: docker:duckduckgo

models:
  groq:
    provider: openai
    model: llama-3.3-70b-versatile
    base_url: https://api.groq.com/openai/v1
    max_tokens: 8000
    temperature: 0.1

How To Reproduce

-Use the yaml above and run it with cagent run test.yaml --env-from-file .env

Expectation

The agent should return the output correctly.

Screenshots

OS and Terminal type

-Windows 11
-Windows PowerShell

Additional context

The issue occurs specifically with mcp type toolsets using docker:duckduckgo even if the MCP server is available in MCP Toolkit.

[dgageot]

Thanks, I was able to reproduce!

[dgageot]

I can only repro with this model so far. All the others I've tried are working ok

[dgageot]

It seems that this family of models has a pretty broken tool call experience on groq:

• [Bug] Tool calling unreliable/broken in Groq agno-agi/agno#4090
• [Bug]: Tool calling is not working with groq/llama-3.1-70b-versatile BerriAI/litellm#5195
• https://community.groq.com/t/tool-use-fail-rate-increases-recently/438

[dgageot]

I'm going to close this one because there's not much we can do.
As a side note, running tests with #744 merged gave me a much clearer error message.