LLM providers

The planner is provider-agnostic. Google Gemini and OpenAI are supported; configure several at once and pick one per run:

// windup.config.ts
llm: {
  provider: "google",                       // default for runs without --llm
  model: "gemini-3.1-flash-lite",
  providers: {
    openai: { model: "gpt-5-mini" },        // default model when --llm openai is used
    // openai: { apiKeyEnv: "MY_OPENAI_KEY", baseUrl: "https://my-proxy/v1" },
  },
},

npx windup run checkout                         # config default (google)
npx windup run checkout --llm openai            # provider default model (gpt-5-mini)
npx windup run checkout --llm openai:gpt-5-nano # explicit provider:model
WINDUP_LLM=openai:gpt-5-mini npx windup run --all   # same thing via env (CI)

• --llm works on run, bench (compare providers on the same scenario) and scan (LLM-assist layer).
• API keys: GOOGLE_GENERATIVE_AI_API_KEY / OPENAI_API_KEY by default; override the env-var name with apiKeyEnv.
• baseUrl (OpenAI only) points at any OpenAI-compatible endpoint — Azure, a proxy, or a local model server.
• Switching providers never invalidates the plan cache: plans are data, replays are LLM-free regardless of who planned them.
• windup costs breaks spend down by provider and by model, so alternating between LLMs keeps per-vendor spend visible.