Support non-Gemini LLM providers (OpenAI-compatible, Anthropic, etc.)

[cosgroveb]

Problem

AsyncReview is hardcoded to Gemini — virtual_runner.py force-prepends gemini/ to model names, GEMINI_API_KEY is the only auth path, and there's no way to point at a different API endpoint. This is unnecessary since DSPy/LiteLLM already handle multi-provider routing via model prefixes.

Proposed solution

Trust the MAIN_MODEL/SUB_MODEL strings as-is and let LiteLLM do the routing. Add optional LLM_API_BASE/LLM_API_KEY env vars for proxy endpoints.

Usage

# Gemini (unchanged)
MAIN_MODEL=gemini/gemini-2.5-pro
GEMINI_API_KEY=...

# OpenAI-compatible proxy
MAIN_MODEL=openai/claude-sonnet-4-5
SUB_MODEL=openai/gemini-2.5-flash

# Anthropic direct
MAIN_MODEL=anthropic/claude-sonnet-4-5
ANTHROPIC_API_KEY=...

Alternatives considered

N/A