v0.4.1

Fixed

• GPT-5.6 Luna (OpenAI OAuth) now works — gpt-5.6-luna was selectable but failed at inference (404 Model not found / "No output generated") because it requires OpenAI's Codex Responses-Lite over WebSocket transport (wss://chatgpt.com/backend-api/codex/responses), while relay-ai only spoke the standard HTTP Responses path. relay-ai now opens an outbound WebSocket per request for models the backend flags this way, forwarding the ChatGPT subscription headers (ChatGPT-Account-Id, originator, version, x-openai-internal-codex-responses-lite) and the OpenAI-Beta: responses_websockets opt-in, and streams the event frames back as SSE. One socket per request, so concurrent Claude Code requests (e.g. background title generation) can't be crossed. Thanks to @tonyb760 for the detailed report (#18). (store: false and the ChatGPT-Account-Id/originator headers were already shipped in 0.4.0.)

Added

• Backend-driven transport selection for ChatGPT Codex models — relay-ai reads the use_responses_lite / prefer_websockets capability flags the ChatGPT Codex model endpoint reports and routes each model accordingly, instead of hardcoding model names. Future Responses-Lite models are picked up automatically with no code change. The static seed carries the same flags as a fallback for a discovery-endpoint outage.