[smakosh]

Closes #89

The problem

The Quick start names two ways to run the OpenDuet Bot: OpenAI and Anthropic. LLM Gateway serves models from many providers behind one OpenAI-compatible key, and OpenDuet already reaches it through OPENAI_BASE_URL and BOT_MODEL. Nothing says so, or which models will work.

The approach

One entry under To use another provider or model, beside OpenAI and Anthropic:

• Three variables: OPENAI_BASE_URL=https://api.llmgateway.io/v1, OPENAI_API_KEY, and BOT_MODEL with an ID from the gateway's model list. Unset, the Bot keeps gpt-5.6-terra from model.yaml.
• Which models: a model needs tools, for the Bot's computer, and JSON output, because skill choice, routing and action review call /chat/completions with response_format: json_object.
• Claude models: the gateway refuses OpenAI's JSON mode for them, so they go through the Anthropic settings instead: BOT_PROVIDER=anthropic, ANTHROPIC_BASE_URL=https://api.llmgateway.io. normalizeModelBaseUrls adds the /v1.
• Setup reads Not checked: checkModelAvailability asks only OpenAI and Anthropic about a key, and skips custom endpoints.

What is not covered

No code change and no new provider id. Everything goes through settings OpenDuet already reads. A first-class provider row would belong in OpenBot.

Verification

• bun run format:check and bun run lint in openbot/ pass.

• bun test tests/copilot.test.ts tests/routing-model-url.test.ts tests/model-availability-check.test.ts tests/openduet-setup.test.ts passes: 144 tests, 0 failures.

• A throwaway script, not committed, runs OpenDuet's own code with the README's settings: normalizeModelBaseUrls, runtimeModelForEnvironment, CopilotKit's resolveModel with a tool, and the router's createModelCompleter. Against a local capture server, the requests were:

  Settings	Bot request	Router / review request
  OpenAI, no BOT_MODEL	POST /v1/responses, gpt-5.6-terra, 1 tool, stream, Bearer	POST /v1/chat/completions, gpt-5.6-terra, json_object
  OpenAI, BOT_MODEL=gemini-3.6-flash	POST /v1/responses, gemini-3.6-flash, 1 tool, stream, Bearer	POST /v1/chat/completions, gemini-3.6-flash, json_object
  Anthropic, BOT_MODEL=claude-sonnet-5-5	POST /v1/messages, claude-sonnet-5-5, 1 tool, stream, x-api-key	POST /v1/messages, claude-sonnet-5-5

• The same script against https://api.llmgateway.io with a placeholder key got the gateway's own API key not found answer on all three endpoints, not a 404. So the URLs are the gateway's real routes.

• The public model list (GET /v1/models) lists gpt-5.6-terra, gemini-3.6-flash and claude-sonnet-5-5. The first two have tools and JSON output. claude-sonnet-5-5 has tools but json_output: false, which is why Claude goes through the Anthropic settings.

Not run: a full OpenDuet session with a live LLM Gateway key. This change is README only, so there is no screenshot or video.