LiteLLM wrapper incompatible with LiteLLM v1.81.3+ for Gemini 2.0+ models (responseJsonSchema change)

[oi-manabu]

🔴 Required Information

Describe the Bug:

ADK's LiteLlm wrapper is incompatible with LiteLLM v1.81.3+ when using Gemini 2.0+ models. The _to_litellm_response_format function always sends response_schema (OpenAPI format), but LiteLLM v1.81.3+ automatically converts this to responseJsonSchema (JSON Schema format) for Gemini 2.0+ models. This format mismatch causes structured output inconsistencies, preventing agent exit conditions from being satisfied.

Steps to Reproduce:

1. Install google-adk==1.23.0 and litellm>=1.81.3
2. Configure LiteLLM proxy with Gemini 2.0+ model (e.g., gemini-3.0-flash-preview)
3. Create an agent that uses structured output with response_schema
4. Run the agent - exit conditions are never satisfied due to structured output inconsistencies

Expected Behavior:

ADK should send the appropriate response format based on the model version:

• Gemini 1.5: response_schema (OpenAPI format)
• Gemini 2.0+: json_schema type with JSON Schema format

Observed Behavior:

Structured output fields are not properly populated, causing agent exit conditions to never be satisfied.

Environment Details:

• ADK Library Version: 1.23.0
• Desktop OS: macOS
• Python Version: 3.11

Model Information:

• Are you using LiteLLM: Yes (via proxy)
• Which model is being used: gemini-3.0-flash-preview

---

🟡 Optional Information

Regression:

This likely worked with LiteLLM < v1.81.3 (before PR #19314 was merged).

Logs:

N/A

Additional Context:

LiteLLM PR #19314 introduced automatic responseJsonSchema usage for Gemini 2.0+ models:

Model	Parameter	Format
Gemini 1.5	response_schema	OpenAPI (uppercase types)
Gemini 2.0+	response_json_schema	JSON Schema (lowercase types)

ADK's current implementation in lite_llm.py (lines 1473-1477, function _to_litellm_response_format):

if _is_litellm_gemini_model(model):
    return {
        "type": "json_object",
        "response_schema": schema_dict,  # Always uses old format
    }

Minimal Reproduction Code:

from google.adk.models.lite_llm import LiteLlm
from google.adk.agents import LlmAgent
from pydantic import BaseModel

class OutputSchema(BaseModel):
    result: str
    is_complete: bool

model = LiteLlm(
    model="openai/gemini-3.0-flash-preview",
    api_base="https://your-litellm-proxy.com",
)

agent = LlmAgent(
    name="TestAgent",
    model=model,
    output_schema=OutputSchema,
    instructions="Process the input and set is_complete=True when done.",
)

# Exit conditions are never satisfied due to structured output inconsistencies

How often has this issue occurred?:

• Always (100%) - when using LiteLLM v1.81.3+ with Gemini 2.0+ models