Update implementation for Agents API to use newer inference APIs

[jwm4]

🤔 What is the technical debt you think should be addressed?

The Agents API in Llama Stack currently depends on the legacy Llama Stack inference APIs. So it would be good for the Agents APIs to be re-implemented to use newer APIs such as the OpenAI chat completions API or the OpenAI Responses API. Using chat completions seems like it would be easier, but either would be acceptable.

💡 What is the benefit of addressing this technical debt?

The legacy Llama Stack inference APIs are not strategic for the project.

Other thoughts

Acceptance Criteria

• The Agents API in no longer calls the legacy Llama Stack inference APIs.
• There are no changes to the Agents API or its behavior.

[jwm4]

@franciscojavierarceo , can you assign this to @aakankshaduggal ?

[franciscojavierarceo]

@aakankshaduggal can you comment on the issue? can't assign unless she does. 🥲

[aakankshaduggal]

/assign @aakankshaduggal

[mattf]

@aakankshaduggal do you have time to complete this?

[aakankshaduggal]

Yes, thanks for opening this issue. Will get it done.

[Akrog]

This will help make the OpenAI inference provider work with compatible servers, because right now it has some problems:

• Even with base_url configured it will not be used for the authentication.
• Cannot use custom provider names if we don't manually register them.
• Some servers fail because we need to pass drop_params to the litellm mixin call as an extra parameter.

These problems get resolved when removing the litellm mixin.