[playgrndlabs-bot]

Requested by PlaygrndLabs · project thread

Live test

A temporary Workers-only app, sub2api-test, ran with a real Claude login and a real ChatGPT login: 31 probes plus the Claude Code and Codex CLIs. The app is deleted now, along with its Durable Object.

Login	Works	Fails
Claude	Claude Code CLI (short and long streamed answers). Requests whose system prompt is Claude Code's: streaming, tools, extended thinking, images, count_tokens, on every model	Every other Anthropic request (plain SDK, own system prompt, Claude Code headers without its prompt): Anthropic answers 429 rate_limit_error. Chat clients: the gateway's 400
ChatGPT	Realtime WebSocket: session plus a full audio and transcript reply	Text on every route (Responses, Chat, Messages, Codex CLI, Claude Code with a GPT model) and WebRTC calls: chatgpt.com's 403 page

Changes

• README limits: ChatGPT text names the 403. Voice is now tested: WebSocket works, WebRTC gets the same 403. The Claude row says other apps get a 429 from Anthropic.
• worker/README: the same, in the feature list.
• Fix: a Responses client (such as the Codex CLI) asking for a claude-* model fell through to ChatGPT logins and got the 403. It now gets the "Claude logins serve Claude Code only" error, as Chat clients do. New test; it fails without the fix.

Security

• The routing change only narrows which credentials a request may use.
• The test app's keys were derived from the Cloudflare token, masked in logs, and died with the app. The temporary workflow is not in this branch.

Tests

npm run typecheck, npm run lint, npm test (254 passed).

---

Generated by Claude Code