[BUG] ADK LiteLlm tool calling fails with Groq — model outputs XML tags instead of function_call JSON #1532
Description
Environment
- kagent-adk v0.8.0-beta9
- Google ADK with LiteLlm model adapter
- Groq API (llama-3.3-70b-versatile)
- MCP tools via StreamableHTTPConnectionParams
Description
When using kagent-adk with Groq's llama-3.3-70b-versatile via LiteLlm, MCP tool calling fails. The model outputs XML-style tool call tags instead of proper OpenAI function_call JSON.
Groq rejects with:
```
{"error":{"message":"Failed to call a function. Please adjust your prompt. See 'failed_generation'",
"type":"invalid_request_error","code":"tool_use_failed",
"failed_generation":"<function=run_command {"command": "kubectl get deployments -n demo"} "}}
```
Root Cause
LiteLLM itself works correctly with Groq. Direct LiteLLM call inside the kagent container with the same model and same tool schema returns proper `tool_calls` JSON with `finish_reason: "tool_calls"`.
The issue is in how ADK converts MCP tool schemas through `FunctionDeclaration` and structures the messages/system prompt before passing to LiteLLM. Something in that conversion makes the model fall back to XML-style tool calling.
Reproduction
```python
This works (direct LiteLLM):
import litellm
result = litellm.completion(
model="groq/llama-3.3-70b-versatile",
messages=[{"role": "user", "content": "Get pods in namespace demo"}],
tools=[{"type": "function", "function": {"name": "run_command", "description": "Execute command", "parameters": {"type": "object", "required": ["command"], "properties": {"command": {"type": "string"}}}}}]
)
Returns proper tool_calls JSON ✓
This fails (via ADK):
from google.adk import Agent
from google.adk.models.lite_llm import LiteLlm
from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset, StreamableHTTPConnectionParams
agent = Agent(
model=LiteLlm(model="groq/llama-3.3-70b-versatile"),
tools=[MCPToolset(connection_params=StreamableHTTPConnectionParams(url="http://mcp-server:3001/mcp"))],
instruction="Use run_command to execute kubectl commands"
)
Model outputs: <function=run_command {...}
Groq rejects: "Failed to call a function" ✗
```
Full reproduction
https://github.com/vitas/evidra-kagent-bench — Docker Compose stack with kagent + Evidra MCP server.
Related
- [BUG] Groq API fails to call tools for ToolCallingAgents huggingface/smolagents#1119 — same Groq XML tag issue in different framework
- Groq not working well with the Litellm definition of tools. BerriAI/litellm#11001 — Groq tool calling issues (but LiteLLM itself works fine here)
Expected
Agent executes tool calls via proper function calling protocol.
Actual
Model falls back to XML-style tool call text, Groq rejects it with 400.