Forward GPT-Live voice calls to the ChatGPT realtime/calls endpoint

[fancyspace78]

What changed

The router now forwards the ChatGPT desktop app's GPT-Live voice calls. POST /v1/live bodies (a multipart/form-data payload with an sdp part and a JSON session part) are parsed and re-encoded as the official JSON shape {"sdp": "...", "session": {...}}, then sent to chatgpt.com/backend-api/codex/realtime/calls.

To match the official client's request shape, the forwarded request includes:

• query params intent=quicksilver&architecture=avas
• header OpenAI-Alpha: quicksilver=v2 (the backend gate rejects anything else)
• the client's full original header set (cookies, integrity-state, DeviceCheck, …), because this endpoint sits behind stricter Cloudflare checks than /responses

Non-multipart POST /v1/live bodies are still passed through unchanged.

Why

Voice in the desktop app is driven by GPT-Live over WebRTC. Because the app routes all model traffic through the configured openai_base_url, those call-creation requests reach the router as POST /v1/live. Before this change the body was treated as JSON and forwarded to /backend-api/codex/v1/live, so voice failed with a series of upstream 400/403/404 responses. DeepSeek routing and the existing prefer_websockets = false upgrade handling (426) are unchanged; voice is never routed to DeepSeek.

Validation

• npm test: 88 pass, 2 skipped, 0 fail, including new tests for the multipart→JSON conversion, the single-path mapping, and the AVAS query/header injection
• Verified end-to-end on a real ChatGPT desktop session: a GPT-Live voice call completes through the router

Tested environment

• macOS (Apple Silicon) + ChatGPT desktop app
• Outbound traffic goes through a local HTTP proxy (VPN); the router reaches chatgpt.com through it, and voice completes end-to-end in that setup

Docs

Updated AGENTS.md and both READMEs so Voice is no longer described as an unrouted client-side feature.