[gakonst]

A host that sends model traffic through a proxy of its own can now plug it into App.create with one function. The test case was paradigmxyz/iron-proxy, where a Worker holds no credentials and the proxy injects them.

• Forward.create({ fetch }) (nanocode-ts/cloudflare) is a generic egress. It answers the private binding's /model with a Request to the dispatch's endpoint, carrying its headers and body, and /socket with a WebSocket upgrade to the socket's endpoint. It hands each to fetch(request, dispatch) and returns the response, which for sockets is a 101 with its webSocket. /warm answers 204, /metrics ok and /remote 501. Before this, every such host had to reimplement nanocode-ts's internal /model and /socket protocol.
• SocketDispatch now carries the Codex Responses endpoint the socket opens, plus subscription, as ModelDispatch does. Before, an egress had to hard-code chatgpt.com for plan sockets. The App computes the endpoint once and uses it for both the runtime's config and the socket dispatch.
• AGENTS.md documents both and the end-to-end check below.

Checked:

• npm run check:types passes.
• npx vp check reports 0 errors.
• npx vp test run passes 39 of 39.
• End to end, 2026-10-28. A Worker using Forward sent every request through an mtls_certificates binding to iron-proxy's reverse listener (feat(proxy): add an mTLS reverse listener for clients that cannot trust the MITM CA paradigmxyz/iron-proxy#294) on a GCP VM. iron-proxy injected a ChatGPT plan's token, which iron-token-broker kept refreshed, and a Claude plan's token. With subscription: () => true and no credential in the Worker:
  • Claude Haiku 5.5 over HTTPS ran a Code Mode turn in 4 s.
  • GPT-6 Luna over the Codex WebSocket ran one in 6 s.

[gakonst]

Landed on main directly as 095cd57, without comments, following nanocode-ts's development rules (66394d0).