feat: production-ready OpenAI Responses API support

[tasoo-oos]

PR Checklist

• Required Checks
  • Have you added type definitions?
  • Have you tested your changes?
  • Have you checked that it won't break any existing features?
• If your PR uses models¹, check the following:
  • Have you checked if it works normally in all models?
  • Have you checked if it works normally in all web, local, and node-hosted versions? If it doesn't, have you blocked it in those versions?
• If your PR is highly AI generated², check the following:
  • Have you understood what the code does?
  • Have you cleaned up any unnecessary or redundant code?
  • Is it not a huge change?
    • We currently do not accept highly AI generated PRs that are large changes.

Summary

Rewrites and significantly expands the OpenAI Responses API support. The previous implementation was a minimal stub: no streaming, no tool calls, no structured output, no reasoning, no NanoGPT wiring, and several correctness issues.

Related Issues

None.

Changes

Refactor

• requestOpenAIResponseAPI was extracted from requests.ts into a new responses.ts module and re-exported.
• LocalNetworkRequestOptions / getLocalNetworkRequestOptions extracted into a new shared.ts used by both modules.
• ResponseOutputItem.status type corrected from 'complete' to 'completed'.

Behavior added/changed

• Streaming support for Responses API (SSE TransformStream parser).
• Function/tool call execution with multi-turn continuation.
• Sanitized continuation payloads for store: false: server IDs and reasoning items are stripped before re-submission.
• Reasoning summaries and reasoning_text content extracted and wrapped in <Thoughts>...</Thoughts>.
• Refusal-only responses surfaced as text.
• JSON schema structured output via text.format.
• reasoning_effort and verbosity model parameters.
• Web search tool via web_search_preview.
• Multimodal inputs (input_image, input_file).
• developer role conversion for models with the DeveloperRole flag.
• NanoGPT Responses endpoint, model selection, auth, and provider header.
  • It is not exposed to the UI, so end user cannot select it.
• Reverse proxy and custom provider endpoint autofill, additional parameters, and headers.
• Local-network routing and streaming timeout support.
• Null-safe aiModel?.startsWith('xcustom:::') in the Chat Completions path.

Tests added

• requests.responses.test.ts: covers request body construction, reasoning summaries, custom/reverse proxy params, NanoGPT, text extraction, incomplete/failed responses, tool continuation sanitization (non-streaming and streaming), SSE chunk parsing, streamed reasoning, and streaming error events.

Impact

• The Chat Completions path (requestOpenAI) is unchanged.
• All LLMFormat.OpenAIResponseAPI models now correctly support streaming, tool calls, structured output, and reasoning.
• NanoGPT Responses format is properly wired (was dead code before).
  • It is not exposed to the UI yet.
• Reverse proxy and custom providers support the Responses format with the same additional-params/headers behavior as Chat Completions.
• Streaming for Responses API is entirely new.

Additional Notes

Manual E2E tests completed:

• With OpenAI's official Responses API
  • Plain Streaming / Non-streaming request
  • Multimodal Input (inlay image file)
  • Integrated tool use (web_search_preview via Chat bot -> Others -> Tools -> Search)
  • External tool use (Risuai Access MCP / RisuAI Dice MCP) with Streaming / Non-streaming request
  • Summarized Reasoning Inclusion with tool use
• With OpenRouter Responses API (by Custom API + https://openrouter.ai/api/v1/responses endpoint)
  • Plain Streaming / Non-streaming request
  • Multimodal Input (inlay image file with hasImageInput custom flags)
  • Tool use (RisuAI Dice MCP) with Streaming / Non-streaming request
  • Reasoning Inclusion with tool use
• With NanoGPT Responses API (by Custom API + https://nano-gpt.com/api/v1/responses endpoint)
  • Plain Streaming / Non-streaming request
  • Multimodal Input (inlay image file with hasImageInput custom flags)
  • Tool use (RisuAI Dice MCP) with Streaming / Non-streaming request

Footnotes

1. Modifies the behavior of prompting, requesting, or handling responses from AI models. ↩

2. Over 80% of the code is AI generated. ↩