## What happened

When passing a provider-specific parameter (e.g. `chat_template_kwargs` for Scaleway/Qwen thinking mode) via `extra_body` or as a top-level kwarg, the request fails with:

``` litellm.APIConnectionError: ScalewayException - AsyncCompletions.create() got an unexpected keyword argument 'chat_template_kwargs' ```

The param arrives as a bare top-level kwarg on the OpenAI SDK call instead of being nested inside `extra_body`.

## Root Cause

There is a mismatch between **routing** and **parameter-formatting** for JSON-configured providers.

### Routing (main.py ~2580) — Scaleway IS treated as OpenAI-compatible

```python or custom_llm_provider in litellm.openai_compatible_providers or JSONProviderRegistry.exists(custom_llm_provider) # ← catches Scaleway ```

Scaleway is a JSON-configured provider (`providers.json`) so `JSONProviderRegistry.exists("scaleway")` is `True`, and it is correctly routed to the OpenAI client path.

### Parameter formatting (utils.py `add_provider_specific_params_to_optional_params` ~4762) — Scaleway is NOT treated as OpenAI-compatible

```python if custom_llm_provider in ["openai", "azure", "text-completion-openai"] + litellm.openai_compatib…