feat: support the OpenAI Responses API, fixes Copilot gpt-6 models

[kai-leddy]

Note

Disclaimer: Both the code in this PR and the PR description were AI-assisted

Copilot's gpt-6 models fail on /chat/completions:

400: {"message":"model \"gpt-6-astra\" is not accessible via the /chat/completions endpoint","code":"unsupported_api_for_model"}

They only work on /responses. So do a few other models in Copilot's list today, for example gpt-5.5, gpt-5.6-luna, grok-4.6 and mai-code-1.1-flash.

What this does

• Adds an api option, "chat_completions" (the default) or "responses". Set it on an agent's model, or on a provider to change the default for all its models. It works for any OpenAI compatible provider. Anthropic, Google AI and Ollama ignore it and log a warning.
• For Copilot, when api isn't set, the plugin asks GET /models once per session, right after the bearer refresh and only before the first query that needs it. It reads supported_endpoints for each model. Models that only list /responses are sent there. Models that list both stay on /chat/completions, so existing setups behave as before. If the lookup fails the plugin logs a warning, uses /chat/completions and tries again on the next query.
• I chose a lookup over a hard coded model list because a list would go stale as Copilot adds models. Setting api explicitly skips the lookup.
• The Responses payload turns system messages into instructions and the rest into input. temperature and top_p are not sent, since reasoning models reject them. reasoning_effort, max_output_tokens and tools are passed through if set on the model.
• The endpoint is the provider's responses_endpoint, or endpoint with /chat/completions replaced by /responses. If neither works, as with Azure's templated URL, the plugin logs a warning and sends nothing.
• The stream parser reads response.output_text.delta events.

Relation to #290

This overlaps with #290, which adds a separate openai_resp provider. I went with an option on the model and provider instead. A provider entry also carries auth, and Copilot needs its own bearer refresh, so a second provider per vendor would copy that. The tools passthrough comes from #290.

Testing

The repo has no test suite, so I tested by hand.

• Unit checks of payload building and api resolution, plus a local fake SSE server for endpoint mapping, the /models lookup, caching, concurrent first queries and a failing lookup.
• Live Copilot, ten models in one cold start run with a single /models call. gpt-6-astra, gpt-6-luna, gpt-6-sol, gpt-5.5, gpt-5.6-luna, gpt-5.4-mini and grok-4.6 went through /responses. gpt-5.4, claude-sonnet-5 and gemini-3.8-flash stayed on /chat/completions. All returned a reply.
• Not tested against Azure or against OpenAI's own /responses.

The README has a short note under the providers list. The vimdoc is left to the docgen action.