When I use **oneapi** to relay OpenRouter, I can't show the thinking model's thinking label in custom endpoint

[harrisonhxy]

What happened?

OpenRouter provides a lot of models, I mainly use one of the thinking models, such as deepseek, grok-3-mini-beta, etc. When I plug OpenRouter into one-api, and use custom endpoint in librechat.yml to customize the endpoints, the thinking model loses the thinking tag, and doesn't show the contents of the thinking, only the result of the answer.

Version Information

ghcr.io/danny-avila//librechat latest 9b0706aeaa4a 20 hours ago 958MB
ghcr.io/danny-avila/librechat-rag-api-dev-lite latest 817faebddae9 2 weeks ago 1.31GB
ghcr.io/rubentalstra/librechat-prom-exporter latest 05eee494c72a 5 weeks ago 245MB

Steps to Reproduce

For example, when I forward DeepSeek from OpenRouter through oneapi, my custom endpoint is set up as follows:

    - name: "deepseek"
      apiKey: "${DEEPSEEK_API_KEY}"
      baseURL: "[https://My oneapi proxy address.com/v1]()"
      iconURL: "https://icon.com/file/1736692987753_deepseek.svg"
      models:
        default: [
          "deepseek-v3",
          "deepseek-r1"
          ]
        fetch: false
      userIdQuery: false
      titleConvo: true
      titleMethod: "completion"
      titleModel: "deepseek-v3"
      summarize: false
      summaryModel: "deepseek-v3"
      forcePrompt: false
      modelDisplayLabel: "deepseek"

DeepSeek at this point would not have a thinking process

DeepSeek got the thinking process going when I set the name field in the custom endpoint, with openrouter (case-insensitive). the custom endpoint settings are shown below:

    - name: "openrouterdeepseek"
      apiKey: "${DEEPSEEK_API_KEY}"
      baseURL: "[https://My oneapi proxy address.com/v1]()"
      iconURL: "https://icon.com/file/1736692987753_deepseek.svg"
      models:
        default: [
          "deepseek-v3",
          "deepseek-r1"
          ]
        fetch: false
      userIdQuery: false
      titleConvo: true
      titleMethod: "completion"
      titleModel: "deepseek-v3"
      summarize: false
      summaryModel: "deepseek-v3"
      forcePrompt: false
      modelDisplayLabel: "deepseek"

I know that the name field of the custom endpoint is case-insensitive, so I'm guessing that when I set name: "deepseek", it will automatically go DeepSeek's route. And when I set it to name: "openrouterdeepseek", it may use OpenRouter's routes. This in undoubtedly limits the advantages of custom endpoint, because there are many vendors that provide DeepSeek models, but LibreChat separates them by different vendor channels. even if I use oneapi aggregation on the backend, I can't achieve polling for multiple vendor channels under the same model name

What browsers are you seeing the problem on?

Chrome

Relevant log output

Librechat's container logs don't show any error messages

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Seems like a oneapi issue

[harrisonhxy]

Indeed, I tested this with a simple conversion on the backend using a cloudflare worker, and the root cause of this is still because OpenRouter's reasoning parameter is unique to it and doesn't conform to the OpenAI parameter specification.

But I wonder if LibreChat could consider a judgment for OpenRouter's routing, i.e. at the time of the custom endpoint request, rather than based on the name field defined by the custom endpoint. such a dynamic request would not only enable multi-channel polling of the same model, but also no longer be constrained by the limitations of the back-end API aggregation platform 😊.