ApigeeLlm Chat Completions drops output_schema and leaves json_schema unnamed #7864

Found while testing LLM provider integrations in agent frameworks: an automated tool flags candidates and each one is reproduced before it is reported. Reproduced on 58332a2.

Describe the Bug:

Structured output does not reach the request with ApigeeLlm on the Chat Completions path (apigee/openai/<model>). There are two problems in the same block:

1. LlmAgent(output_schema=Weather) sets config.response_schema (and response_mime_type="application/json"), but the Chat Completions path only reads response_json_schema. The schema is dropped and the request carries {"type": "json_object"}. OpenAI rejects JSON mode when no message contains the word "JSON", and an agent's instruction usually doesn't.
2. response_json_schema is placed directly under response_format.json_schema. Chat Completions expects {"name": ..., "schema": {...}} there, and name is required, so this request is rejected too.

Steps to Reproduce:

1. pip install google-adk (or a checkout at 58332a2)
2. Run the snippet below. It swaps the HTTP transport for an httpx.MockTransport and prints response_format for a request with output_schema and one with response_json_schema.

Expected Behavior:

Both requests carry {"type": "json_schema", "json_schema": {"name": "Weather", "schema": {...}}}.

Observed Behavior:

{"type": "json_object"}
{"type": "json_schema", "json_schema": {"properties": {"city": {"title": "City", "type": "string"}, "temperature_c": {"title": "Temperature C", "type": "integer"}}, "required": ["city", "temperature_c"], "title": "Weather", "type": "object"}}

I did not send these to a live endpoint. The JSON-mode rule is in the Structured Outputs guide ("the API will throw an error if the string "JSON" does not appear somewhere in the context"), and name is a required field of JSONSchema.

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (main at 58332a2)
• Desktop OS: Linux
• Python Version (python -V): 3.12.4

Model Information:

• Are you using LiteLLM: No
• Which model is being used: ApigeeLlm(model="apigee/openai/gpt-4o")

Additional Context:

CompletionsHTTPClient._map_config_parameters only looks at response_json_schema, passes it through unwrapped, and otherwise turns application/json into json_object:

A fix is to take the schema from response_json_schema or response_schema, convert it to a JSON Schema dict (lowercasing types.Schema types as elsewhere), and send {"type": "json_schema", "json_schema": {"name": <schema title or "response">, "schema": <dict>}}, passing through a value that is already in that wrapped form. The two existing test_construct_payload_response_format cases that assert the unwrapped shape would change with it.

Minimal Reproduction Code:

import asyncio, json, httpx
from pydantic import BaseModel
from google.adk.models import apigee_llm
from google.adk.models.llm_request import LlmRequest
from google.genai import types

def handler(request):
  print(json.dumps(json.loads(request.content).get('response_format')))
  return httpx.Response(200, json={'choices': [{'message': {'role': 'assistant', 'content': '{}'}}]})

real_client = httpx.AsyncClient
apigee_llm.httpx.AsyncClient = lambda *a, **kw: real_client(*a, **{**kw, 'transport': httpx.MockTransport(handler)})

class Weather(BaseModel):
  city: str
  temperature_c: int

llm = apigee_llm.ApigeeLlm(model='apigee/openai/gpt-4o', proxy_url='https://proxy.example.com')

def new_request(**config):
  return LlmRequest(model=llm.model, config=types.GenerateContentConfig(**config),
                    contents=[types.Content(role='user', parts=[types.Part(text='Weather in Paris?')])])

async def send(request):
  async for _ in llm.generate_content_async(request):
    pass

with_output_schema = new_request()
with_output_schema.set_output_schema(Weather)  # what LlmAgent(output_schema=Weather) does
asyncio.run(send(with_output_schema))
asyncio.run(send(new_request(response_json_schema=Weather.model_json_schema())))

How often has this issue occurred?:

• Always (100%)