feat(responses): Experimental Responses API compatibility

[alexx-ftw]

feat(responses): Experimental Responses API compatibility

Summary

Adds OpenAI Responses API–compatible endpoints behind flags, with local polyfills to work around ChatGPT endpoint constraints.

Endpoints

• POST /v1/responses -- streams SSE when "stream": true; aggregates JSON when false.
• GET /v1/responses/{id} -- available when a prior request used "store": true (local storage only).

Features

Input handling: Accepts native Responses input or converts Chat-style messages/prompt.

Tools: Supports function tools plus passthrough responses_tools for web_search/web_search_preview; honors responses_tool_choice (auto/none).

Reasoning: Configurable via --reasoning-effort, --reasoning-summary, --reasoning-compat.

Parameter safety:

• Strips unsupported token params (max_output_tokens, max_completion_tokens) for the codex upstream.
• Strips store and previous_response_id before forwarding (upstream rejects these).
• Implements local polyfills for expected API behavior.
• Sanitizes upstream response ID references (rs_*) to prevent invalid cross-call references.

ChatGPT Endpoint Constraints (Validated via Live Testing)

This implementation proxies to ChatGPT's internal backend-api/codex/responses endpoint, which has strict limitations:

Test Case	Result	Error Message
store=false	✅ Accepted	(also requires stream=true)
store=true	❌ 400	"Store must be set to false"
store omitted	❌ 400	"Store must be set to false"

Key findings:

• The ChatGPT endpoint requires store=false and rejects any other value
• previous_response_id is not supported by the upstream endpoint
• Only streaming mode (stream=true) is supported
• This differs from the official OpenAI Platform Responses API, which supports store=true for server-side persistence

Our approach: Implement local storage (_STORE) and threading (_THREADS) to provide expected API behavior while respecting upstream constraints.

CLI Flags

• --enable-responses-api: mount /v1/responses endpoints (default off).
• --responses-no-base-instructions: forward client instructions as-is; otherwise base instructions are injected and client instructions are moved into input.
• --enable-web-search: when responses_tools is omitted, enables default web_search unless responses_tool_choice is "none".

Request for feedback

• API shape and field mapping, instruction handling defaults, tool passthrough, and downstream client compatibility (e.g., Jan/Raycast) before considering default-on.

Documentation

Comprehensive documentation added covering:

• ChatGPT endpoint vs official OpenAI API differences
• Parameter handling and local polyfill behavior
• Upstream response ID sanitization logic
• Logging events for debugging

PR: #52

[adogwangwang]

I found that the temperature and top_p parameters are not supported in the /v1/response interface. They can be filtered out before being transmitted.

[alexx-ftw]

Fixed - temperature and top_p parameters are now stripped before forwarding to upstream codex/responses endpoint.

[alexx-ftw]

@codex review