ApigeeLlm (Chat Completions) ignores FunctionCallingConfig.allowed_function_names and the VALIDATED mode #8100

Describe the Bug:

ApigeeLlm (Chat Completions) ignores FunctionCallingConfig.allowed_function_names and the VALIDATED mode. _map_tools in src/google/adk/models/apigee_llm.py (lines 765-789 at b8836bdb, 815-822 on current main) maps only mode: ANY -> "required", NONE -> "none", AUTO -> "auto". So "call lookup_order" becomes "call any tool", and "only these tools" becomes no restriction. No warning is logged. We found this while testing how SDKs carry tool-call constraints to the provider.

At the same commit, AnthropicLlm ignored tool_config entirely and always sent tool_choice: {"type": "auto"} (anthropic_llm.py:1068-1072). This looks fixed on main by _build_tool_choice. One case may remain: when llm_request.tools_dict is empty (a direct generate_content_async call with tools only in config.tools), tool_choice is still NOT_GIVEN even if tool_config is set. We have not rerun that on main.

Steps to Reproduce:

1. Start a local HTTP server that logs request bodies.
2. Create an ApigeeLlm with proxy_url pointing at it and api_type="chat_completions".
3. Run an LlmAgent with two tools and the config below.
4. Look at tool_choice in the captured request.

Expected Behavior:

mode=ANY, allowed_function_names=["lookup_order"] -> "tool_choice": {"type": "function", "function": {"name": "lookup_order"}}. mode=VALIDATED (or AUTO) with allowed_function_names -> an allowed_tools choice, or the other tools removed from the request, or an error saying the restriction cannot be sent.

Observed Behavior:

• ANY + ["lookup_order"]: "tool_choice": "required", both tools declared. In our test the server answered with issue_refund and the agent ran it.
• VALIDATED + ["lookup_order"]: no tool_choice, both tools declared. The agent ran issue_refund.

Environment Details:

• ADK Library Version: google-adk 2.9.2 (commit b8836bdb). Apigee code unchanged on main at 9687bae1.
• Desktop OS: Linux (Ubuntu 24.04)
• Python Version: 3.12

Model Information:

• Are you using LiteLLM: No
• Which model is being used: apigee/openai/gpt-4o (Chat Completions) and claude-sonnet-4-5 through AnthropicLlm

Minimal Reproduction Code:

from google.adk.agents import LlmAgent
from google.adk.models.apigee_llm import ApigeeLlm
from google.adk.runners import InMemoryRunner
from google.genai import types

def lookup_order(order_id: str) -> str:
    return "delivered"

def issue_refund(order_id: str, amount: int) -> str:
    return "refunded"

cfg = types.GenerateContentConfig(tool_config=types.ToolConfig(
    function_calling_config=types.FunctionCallingConfig(
        mode=types.FunctionCallingConfigMode.ANY, allowed_function_names=["lookup_order"])))

agent = LlmAgent(name="app", instruction="Use the tools.",
                 model=ApigeeLlm(model="apigee/openai/gpt-4o", proxy_url="http://127.0.0.1:8080/v1",
                                 api_type="chat_completions"),
                 tools=[lookup_order, issue_refund], generate_content_config=cfg)
runner = InMemoryRunner(agent=agent)
# run one turn with runner.run_async(...) and inspect the captured request

How often has this issue occurred?: Always (deterministic, request construction)

Additional Context:

Impact: an application that forces or limits tool calls through FunctionCallingConfig expects the model to be held to that. On these backends the restriction is not sent, so any declared tool can be called.

Suggested fix: map allowed_function_names with one name to a named function choice, map a subset to allowed_tools (or filter tools), and raise or warn for combinations that cannot be expressed.