payload:
    override:
      - models:
          - name: deepseek-v4-flash-0731
          - name: deepseek-v4-flash
        params:
          thinking.type: adaptive
          output_config.effort: max

    filter:
      - models:
          - name: deepseek-v4-flash-0731
          - name: deepseek-v4-flash
        params:
          - thinking.budget_tokens

然后走 Claude 的 
不走 v1/response 
这样指定就行了