Question: lightweight hosted OpenAI-compatible configuration across agent roles #19

I am evaluating a Tsubasa backend using https://api.tsubasa.sh/v1 and tsubasa-fast / tsubasa-pro through existing client configuration.

In the source below, OpenAILLM already accepts the configured model and OpenAI credentials, but importing its module also imports the local-serving runtime and Transformers unconditionally. pilot.py separately builds AutoGen role configurations. Testing a standalone OpenAI SDK client would therefore not establish that the complete role graph uses the intended endpoint, key and token limits.

Is a lightweight hosted-only installation and a shared custom-provider configuration still a maintained extension route? If so, would a small change that lazily imports only the local-serving dependencies, plus an example configuring every OpenAI/AutoGen role, be appropriate? I would retain the current local-serving behavior and add an import/configuration regression test before proposing code.

This is a source-based question, not a runtime or benchmark success claim. The intended profile would explicitly set the endpoint and credentials for every role and fit input plus output within Tsubasa's 32,768-token context. No hosted inference or full role graph has been qualified.

Source: src/hyperagent/agents/llms.py, src/hyperagent/pilot.py.

Tsubasa-affiliated, AI-assisted investigation.