Add backward-compatible OpenAI Responses API support with URL auto-detection

[Copilot]

PR #277 proposed switching the default OpenAI endpoint from /chat/completions to /responses, but the change was incomplete: the Responses API has a different request format (input vs messages) and response shape (output[0].content[0].text vs choices[0].message.content), making the PR as-is a silent breaking change.

Changes

• createOpenAILanguageModel extended — accepts an optional fifth parameter useResponsesApi?: boolean and auto-detects the API variant from the endpoint URL (path ending with /responses):

  • useResponsesApi: true — forces Responses API regardless of URL
  • useResponsesApi: false — forces Chat Completions regardless of URL
  • useResponsesApi omitted (default) — inferred from the endpoint URL

• URL auto-detection — any endpoint whose path ends with /responses (e.g. https://api.openai.com/v1/responses or an Azure deployment URL like https://{resource}.openai.azure.com/.../responses?api-version=...) automatically uses the Responses API request/response format; all other URLs use Chat Completions

• createOpenAIResponsesLanguageModel removed — the separate function was never published and has been deleted; all Responses API usage goes through createOpenAILanguageModel

• createLanguageModel simplified — the OPENAI_USE_RESPONSES_API env var is removed; to opt into the Responses API set OPENAI_ENDPOINT to a /responses URL and URL auto-detection handles the rest

• retry-after header support — on transient error responses (429, 503, etc.), the Retry-After header value (seconds) is used as the back-off delay, capped at retryPauseMs × retryMaxAttempts to stay within the configured total retry budget

• Improved documentation — createResponsesFetchLanguageModel now includes a sample response JSON, links to OpenAI and Azure OpenAI Responses API docs, and full @param documentation

// Existing code — unchanged, backward compatible
const model = createOpenAILanguageModel(apiKey, "gpt-4");

// New: explicit Responses API via URL (auto-detected)
const model = createOpenAILanguageModel(apiKey, "gpt-4", "https://api.openai.com/v1/responses");

// New: explicit Responses API via flag
const model = createOpenAILanguageModel(apiKey, "gpt-4", undefined, undefined, true);

// Via createLanguageModel — set OPENAI_ENDPOINT to a /responses URL
// OPENAI_ENDPOINT=https://api.openai.com/v1/responses  →  uses /v1/responses
// (default)                                             →  uses /v1/chat/completions
const model = createLanguageModel(process.env);

• Tests — 19 unit tests (typescript/tests/model.test.mjs) using Node's built-in node:test with mocked fetch, covering both API paths, URL auto-detection, the useResponsesApi flag, retry-after handling (429 and 503) for both paths, and createLanguageModel routing; "test" script added to package.json

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.