litellm - 💡(How to fix) Fix use_chat_completions_url_for_anthropic_messages config key not loaded from proxy YAML

Bug Description

The config key use_chat_completions_url_for_anthropic_messages is not respected when set in litellm_settings in the LiteLLM proxy configuration YAML.

Root Cause

In litellm/llms/anthropic/experimental_pass_through/messages/handler.py:

_RESPONSES_API_PROVIDERS = frozenset({"openai"})

def _should_route_to_responses_api(custom_llm_provider: Optional[str]) -> bool:
    if litellm.use_chat_completions_url_for_anthropic_messages:
        return False
    return custom_llm_provider in _RESPONSES_API_PROVIDERS

The module-level variable litellm.use_chat_completions_url_for_anthropic_messages is never populated from the proxy config YAML. Setting use_chat_completions_url_for_anthropic_messages: true under litellm_settings has no effect because the config parsing code does not map this key to the litellm module attribute.

Reproduction Steps

1. Configure LiteLLM proxy with an openai/-prefixed vLLM backend
2. Set use_chat_completions_url_for_anthropic_messages: true in litellm_settings in the proxy config YAML
3. Send an Anthropic Messages API request through the proxy (e.g., via Claude Code with ANTHROPIC_BASE_URL pointing to the proxy)
4. The request is incorrectly routed to the Responses API endpoint

Observed Behavior

vLLM receives a Responses API payload (input: [...]) instead of the expected chat completions format (messages: [...]), resulting in a 400 error.

Expected Behavior

When use_chat_completions_url_for_anthropic_messages: true is set in the proxy YAML, the litellm.use_chat_completions_url_for_anthropic_messages flag should be True, causing _should_route_to_responses_api() to return False and route through the Chat Completions API instead.

Environment

• LiteLLM version: 1.86.0
• Backend: vLLM (OpenAI-compatible)
• Client: Claude Code via ANTHROPIC_BASE_URL proxy

Current Workaround

Manually patch the source file:

_RESPONSES_API_PROVIDERS = frozenset()  # empty set bypasses the check

Suggested Fix

Wire the use_chat_completions_url_for_anthropic_messages config key from litellm_settings in the proxy YAML to the litellm.use_chat_completions_url_for_anthropic_messages module attribute during proxy startup.