Feature request: Allow routing to local OpenAI-compatible backends (Ollama / vLLM / LM Studio)

[Ban921]

Summary

Today agent/core/llm_params.py::_resolve_llm_params hard-routes every
model id into one of three branches:

• anthropic/... → Anthropic API
• openai/... → OpenAI API (no api_base override)
• everything else → forced onto https://router.huggingface.co/v1
  with an HF token

There's no escape hatch for a local OpenAI-compatible server (Ollama,
vLLM, LM Studio, llama.cpp's server, etc.), even though litellm
itself supports all of them out of the box.

Motivation

Smaller local models almost certainly won't match frontier-model
quality on this agent loop, and that's fine — the value is in opening
up more room for experimentation:

• Cheap / unlimited iteration without per-token costs
• Offline use and stronger privacy for sensitive datasets / papers
• A testbed for the community to benchmark open-weights models on the
  same agent harness
• Lower barrier to entry for students and hobbyists who want to study
  the agent end-to-end

Suggested shape

Something minimally invasive, e.g.:

• recognize an ollama/... prefix and pass it through to litellm, or
• honor an optional api_base field in main_agent_config.json /
  LITELLM_API_BASE env var on the "fallback" branch instead of
  hard-coding the HF router URL.

Related to #56, but specifically about local execution rather than
hosted open-source model APIs.

Happy to test or send a PR if this is on the roadmap.

[Pro777]

@Ban921 You can build the PR or I can. :)

Dealer's choice.

[Pro777]

Investigating a downstream implementation in the Spitfire-Cowboy fork that adds generic OpenAI-compatible provider routing (including local backends). Will follow up with a cleaner upstream slice if it makes sense relative to #55 / #60.