Prompt parameters added by defaultParamsEndpoint: 'anthropic' do not work

[jaceli-stripe]

What happened?

Hi, we are using LiteLLM as a provider for LibreChat. Currently, we have anthropic/claude models on their own custom endpoint:

custom:
    - name: "litellm-bedrock-claude"
      apiKey: "<our-litellm-api-key>"
      baseURL: "<internal-litellm-url>/v1"
      models:
        default: ["claude-opus-4.1", "claude-sonnet-4", "claude-3.5-sonnet-v2", "claude-3.5-haiku"]
      titleConvo: true
      titleModel: "claude-3.5-haiku"
      summarize: false
      summaryModel: "claude-3.5-haiku"
      forcePrompt: false
      modelDisplayLabel: "litellm-bedrock-claude"
      customParams:
        defaultParamsEndpoint: 'anthropic'

Due to defaultParamsEndpoint: 'anthropic', we are able to see prompt parameters such as Thinking and Thinking Budget using the models. However, they do not actually modify the request sent and therefore do not work. In particular, toggling Thinking does not meaningfully change the request being sent out and doesn't show the reasoning/chain of thought of the model as expected.

Version Information

v0.7.9

Steps to Reproduce

1. Apply librechat.yaml config
2. Use model with Thinking and Thinking Budget parameters
3. Send prompt and inspect request on browser
4. Observe the lack of "Thought" section of chat log

What browsers are you seeing the problem on?

Chrome

Relevant log output

Request payload with `Thinking` enabled and `Thinking Budget` of `2000`:
{"text":"hi","sender":"User","clientTimestamp":"2025-08-07T16:55:48","isCreatedByUser":true,"parentMessageId":"00000000-0000-0000-0000-000000000000","conversationId":"935947a8-91f9-4437-b649-3b75675619eb","messageId":"4d80f6a1-6866-4781-9ea5-4626b3fa69aa","error":false,"responseMessageId":"ebc6e4c9-ed97-4a28-a007-558f81ce8e1e_","overrideParentMessageId":"8f24c973-bb9b-4ab2-b85c-bb1f6eb4ba88","endpoint":"litellm-bedrock-claude","endpointType":"custom","model":"claude-opus-4.1","resendFiles":true,"key":"never","modelDisplayLabel":"litellm-bedrock-claude","isTemporary":false,"isRegenerate":true,"isContinued":false,"ephemeralAgent":{"execute_code":false,"web_search":false,"file_search":false,"artifacts":false,"mcp":[]}}

Screenshots

No response

Code of Conduct

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

[danny-avila]

Thanks for providing a config to test with. I think I see the issue here and will resolve it soon!

[jaceli-stripe]

hi danny, any updates on this? 🙏

[danny-avila]

Have a couple higher priority items, I will update on this thread when I've addressed it.

[danny-avila]

@jaceli-stripe just to confirm, when you set these parameters, do you expect the exact payload as it would be for anthropic for thinking/thinking-budget? i.e. https://docs.anthropic.com/en/api/messages#body-thinking

    "thinking": {
        "type": "enabled",
        "budget_tokens": 10000
    },

[jaceli-stripe]

the request sent by the browser to the librechat backend does not need to be in the anthropic format. the actual model invocation request that librechat sends to our llm gateway (LiteLLM) would have to be in the anthropic format, yes, in order to use the extended-thinking mode.

[danny-avila]

Got it, thanks this helps me plan a solution

[jaceli-stripe]

could modelspecs https://www.librechat.ai/docs/configuration/librechat_yaml/object_structure/model_specs be used as a workaround for now? is there a way to expose different parameters on the UI (and send them in provider-specific formats to our endpoint gateway) on a per-model basis?

[jaceli-stripe]

one of the reasons we are testing out having the claude models on their own custom endpoint in LibreChat (while pointing to the same LLM gateway as our OpenAI models) is that all the custom endpoint models are treated as OpenAI models and lack the thinking or budget_tokens params in LibreChat by default unless we put them in a different endpoint with defaultParamsEndpoint: 'anthropic'. an even better user experience would probably be to have all the models on the same custom endpoint, but be able to specify the provider format and params for each model. eg:

librechat.yaml

custom:
    - name: "litellm-unified"
      apiKey: "<our-litellm-api-key>"
      baseURL: "<internal-litellm-url>/v1"
      models:
        default: 
           - "gpt-5"
               - provider_format: "openai"
           - "claude-opus-4.1"
               - provider_format: "anthropic"

[danny-avila]

Just wanted to give you a heads up that we started on this here and looking to implement this weekend:
#9415

[danny-avila]

Ran out of time today but got to this point:

Still need to test the branch more but getting there

[danny-avila]

Closed by #9415