有办法临时绕过 normalization 逻辑吗？我设置使用 deepseek-v4-pro 结果给我改成 chat 了

Yes — I verified this against the current main-branch normalization behavior.

On main, using the native DeepSeek provider currently normalizes:

deepseek-v4-pro + provider=deepseek -> deepseek-chat

But using a custom OpenAI-compatible endpoint bypasses the DeepSeek-specific normalization and preserves the model id:

deepseek-v4-pro + provider=custom -> deepseek-v4-pro

So a temporary workaround before this PR lands is to configure DeepSeek as a custom OpenAI-compatible provider, for example:

model:
  provider: custom
  default: deepseek-v4-pro
  base_url: https://api.deepseek.com/v1

Or, if using a named custom provider / custom_providers entry:

custom_providers:
  - name: deepseek-custom
    base_url: https://api.deepseek.com/v1
    key_env: DEEPSEEK_API_KEY
    api_mode: chat_completions
    model: deepseek-v4-pro

The important bit is that the provider should be custom, not deepseek, so the model name is passed through unchanged. This PR makes the native deepseek provider also pass deepseek-v4-flash and deepseek-v4-pro through unchanged.

Superseded by #14946. After #14934 landed the v4 model ids on the DeepSeek native provider, the only remaining gap was the normalization path — I've extracted that into a smaller, fix-scoped PR based on latest main. Closing this one.