Litellm provider using openai model name defaults

[myaple]

Describe the bug

The litellm provider uses crates/goose/src/providers/formats/openai.rs as it's create_request (https://github.com/block/goose/blob/main/crates/goose/src/providers/litellm.rs#L170), but this create_request function does very specific openai processing, like here: https://github.com/block/goose/blob/main/crates/goose/src/providers/formats/openai.rs#L580

This means that if my litellm proxy has a model name that starts with "o", like "open-mistral-small-3.1", then it will send a reasoning_effort parameter regardless of whether or not it's a reasoning model.

To Reproduce
Steps to reproduce the behavior:

1. Stand up litellm with a model that starts with "o"
2. Send a request through goose
3. See that it sends a reasoning_effort parameter regardless of model

I've written a quick test for this behavior that fails on the current main branch: https://github.com/myaple/goose/blob/myaple/litellm-reasoning-bug/crates/goose/src/providers/litellm.rs#L339

Expected behavior
Reasoning parameters should only be sent to be models that support reasoning. Additionally, there should be an environment variable to override the reasoning parameter for the openai models or disable it completely.

Please provide following information:

• OS & Arch: ubuntu 24
• Interface: cli
• Version: 1.5.0
• Extensions enabled: n/a
• Provider & Model: litellm, open-mistral-small-3.1

[myaple]

I'm more than happy to give this issue a shot if you guys would like. My initial idea would be to add a parameter to the create_request function that takes in a reasoning effort override of high, medium, low, or disabled, and sets the reasoning effort appropriately without doing the model based parsing. This could then be passed in as a configuration parameter like GOOSE_REASONING_EFFORT if you'd like it done at a global level, or LITELLM_REASONING_EFFORT for provider level.

We could also create a new create_request specifically for litellm, but I think it'd be easier in the long run to leave the openai format alone until otherwise necessary rather than duplicate code.

[michaelneale]

thanks @myaple - added a comment on the PR. Agree ideally could just be minor tweaks to do reasoning or not, even a list of big models. I think general idea is not pass in reasoning_effort other than exceptional cases you would say? Longer term I think reasoning like this goes away, as the models work it out themselves with gating, routing and mixtures of experts (like gpt-5 basically does now, but I guess there is some space in api to be specific, but I think that may be short lived)

[DOsinga]

should be fixed after #5084

[myaple]

That is a dramatic improvement on how providers are handled now, thanks for letting me know. With it, is there a plan to add a the reasoning effort param to the model info? Just looking at the second half of this where I'd like an idea on how to provide a fixed reasoning param to arbitrary models without having to modify the model name on the API I'm calling.

[DOsinga]

I wouldn't say there's a plan for that, but once we have this in, we can ad all kinds of flags and features to the custom providers yeah