## Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

## What happened?

I am using LiteLLM with a self-hosted vLLM backend that exposes the Anthropic-compatible Messages API.

The reason for using `custom_llm_provider: anthropic` is intentional: when using `hosted_vllm`, LiteLLM converts Anthropic `/v1/messages` requests into `/v1/chat/completions`, and that adapter path is not fully compatible with Claude Code / Anthropic Messages features such as content blocks, tool use, tool results, thinking blocks, and other Anthropic-native request shapes.

So the intended setup is:

```yaml model_list: - model_name: claude-opus-* litellm_params: custom_llm_provider: anthropic model: ultra api_base: http://<vllm-host>:<port>/v1 api_key: <vllm-api-key> ```

In this setup, `/v1/messages` should be forwarded to the configured Anthropic-compatible backend, not to Anthropic SaaS.

That works for normal message generation.

However, when Claude Code calls:

```text POST /v1/messages/count_tokens ```

LiteLLM does not forward the request to the configured `api_base`.

Instead, LiteLLM attempts to call:

…