Issue Summary

Tool responses from MCP (Model Context Protocol) servers are double-serialized when converted to LiteLLM message format, creating triple-nested JSON that prevents Claude, GPT, and other non-Gemini models from parsing and presenting tool results to users.

Environment

• ADK Version: 1.19.0
• Affected Models: All models via LiteLLM (Claude via Vertex AI, Azure OpenAI GPT-5, etc.)
• Tool Integration: MCP servers (e.g., Google Workspace via Gluon Link)
• Working Models: Native Gemini models (not affected - use different conversion path)

Problem Description

When MCP tools return responses, the data comes as already-serialized JSON strings. However, the _content_to_message_param() function in lite_llm.py (line 313) unconditionally calls _safe_json_serialize() on these responses, which performs json.dumps() on already-JSON strings. This creates escaped, triple-nested JSON that models cannot parse.

Root Cause

File: lite_llm.py
Line: 313

tool_messages.append(
    ChatCompletionToolMessage(
        role="tool",
        tool_call_id=part.function_response.id,
        content=_safe_json_serialize(part.function_response.response),  # Double serialization
    )
)

The _safe_json_serialize() helper (lines 180-189) calls json.dumps() without checking if the input is already a JSON string, causing:

Original MCP response: {"type": "files", "count": 2, ...}
After first serialization (MCP): '{"type": "files", "count": 2, ...}'
After second serialization (line 313): '"{\\"type\\": \\"files\\", \\"count\\": 2, ...}"'
Final in conversation: '{"content": [{"type": "text", "text": "{\\\"type\\\": ..."}]}'  # Triple-nested!

Impact

Models cannot parse tool results:

• Claude receives "{\\\"type\\\": \\\"files\\\"..." instead of clean JSON
• GPT receives the same malformed triple-escaped structure
• Models fail to extract actual tool data (file listings, search results, etc.)
• Users see incomplete responses: "I'll list files," but no actual file data is displayed
• Models can call tools successfully, but cannot present results to users

This breaks the entire tool usage workflow for non-Gemini models.

Steps to Reproduce

1. Configure an ADK agent with LiteLLM model and MCP tools:

   model = LiteLlm(model="vertex_ai/claude-sonnet-4-5@20250929", stream=True)
   toolset = McpToolset(mcp_servers={"google": mcp_config})
   agent = Agent(model=model, toolset=toolset)

2. Send a request that triggers tool execution:

   "List my recent files from Google Drive"
   

3. Observe in logs:

   • Tool executes successfully
   • MCP returns valid JSON: {"type": "files", "items": [...]}
   • Conversion to LiteLLM format double-serializes
   • Model receives: '{"content": [{"type": "text", "text": "{\\\"type\\\"..."}]}'

4. Model response:

   • Says: "I'll list the files from your Drive root directory."
   • Shows: No actual file data (cannot parse the triple-nested JSON)

Example Log Evidence

From uvicorn_debug.log (line 3355):

{
  "content": [
    {
      "type": "text",
      "text": "{\n  \"type\": \"files\",\n  \"path\": \"/\",\n  \"count\": 2,\n  \"items\": [\n    {\n      \"name\": \"mydoc_v2.pdf\",\n      \"owner\": \"my_user\",\n      \"sharing\": \"Private\",\n      \"modified\": \"2025-06-13\",\n      \"size\": \"1.6 MB\",\n      \"link\": \"https://drive.google.com/file/d/...\",\n      \"id\": \"1F4y...\"\n    }\n  ]\n}"
    }
  ]
}

Notice: The entire JSON payload is inside a quoted string within the "text" field, with escaped quotes \". This is the result of double serialization.

What Claude/GPT receives:

'{"content": [{"type": "text", "text": "{\\n  \\"type\\": \\"files\\",\\n ..."}]}'

They cannot parse this as structured data.

Expected Behavior

Tool responses should be passed through as-is when they're already JSON strings, or serialized only once if they're Python objects:

# Correct handling:
if isinstance(part.function_response.response, str):
    content = part.function_response.response  # Already serialized by MCP
else:
    content = _safe_json_serialize(part.function_response.response)  # Serialize Python objects

Proposed Fix

File: lite_llm.py
Lines: 306-320

# BUGFIX: Check if response is already a string before serializing
tool_messages = []
for part in content.parts:
    if part.function_response:
        # If response is already a string (from MCP), don't serialize again
        response_content = (
            part.function_response.response
            if isinstance(part.function_response.response, str)
            else _safe_json_serialize(part.function_response.response)
        )
        tool_messages.append(
            ChatCompletionToolMessage(
                role="tool",
                tool_call_id=part.function_response.id,
                content=response_content,
            )
        )
if tool_messages:
    return tool_messages if len(tool_messages) > 1 else tool_messages[0]