feat: add Responses API endpoint type

[tessaherself]

Summary

Adds support for the OpenAI Responses API (/v1/responses) as an opt-in endpoint type alongside the existing Chat Completions endpoint.

This aligns with HF's own Open Responses initiative and the broader ecosystem shift toward the Responses API standard:

• HF Inference Providers already serve /v1/responses via router.huggingface.co
• vLLM added Responses API support in Semantic Router v0.1
• llama.cpp has an active PR for OpenResponses compliance

What this enables

Feature	Chat Completions	Responses API
Tool calling	Function calls (bolted on)	Native tool orchestration
Conversation state	Manual (full history each turn)	previous_response_id chaining
File attachments	Base64 image_url only	input_file + input_image + file IDs
Reasoning	<think> tag hacks	Native reasoning output items
Streaming	Raw text deltas	Semantic events
Cache performance	Baseline	40-80% improvement (OpenAI benchmarks)

How it works

• New endpoint type "responses" — configure per-model:

  { "endpoints": [{ "type": "responses", "baseURL": "https://router.huggingface.co/v1" }] }

• Chat Completions (type: "openai") remains the default — zero breaking changes
• chatMessagesToResponsesInput() adapter converts existing message format → Responses API input
• Stream adapter maps Responses API events → existing TextGenerationStreamOutput type
• Uses the same OpenAI SDK v4 already in the project (openai@4.104.0)

Changes

• New: endpointResponses.ts — endpoint factory + message format adapter
• New: openAIResponsesToTextGenerationStream.ts — streaming + non-streaming response adapters
• New: endpointResponses.spec.ts — 6 unit tests for message conversion
• Modified: endpoints.ts — register responses type in endpoint registry + schema
• Modified: models.ts — dynamic endpoint type dispatch

Test plan

• npm run check — 0 type errors
• npm run lint — passes prettier + eslint
• npm run test — 161 tests passing (155 existing + 6 new)
• Manual: configure model with type: "responses" pointing at router.huggingface.co/v1
• Manual: verify streaming text generation
• Manual: verify multimodal (image) input
• Manual: verify existing type: "openai" endpoints unchanged

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a4a011542d

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Emit final text for incomplete Responses completions

When the Responses API finishes with response.completed status incomplete (for example after hitting max_output_tokens), this code sets generated_text to null because it only accepts status === "completed". The downstream generators only finalize an answer when generated_text is present, so truncated-but-successful runs can stream tokens to the client but never produce a final saved assistant message.

Useful? React with 👍 / 👎.