[playgrndlabs-bot]

What

OpenAI Responses clients (/v1/responses) could only reach Codex keys and ChatGPT logins. Now they reach every provider, and so do OpenAI Chat clients, by pivoting through Anthropic Messages:

• free/src/responses-claude.ts (new), ported from CPA's translator/claude/openai/responses:
  • Request: instructions and system/developer items become the system prompt; input_text/output_text/refusal/input_image/input_file parts; reasoning items replay as thinking only when they carry a Claude signature (OpenAI's gAAAA… tokens are dropped); function_call and custom_tool_call become tool_use (custom tools take a one-string {input} schema); tool outputs keep the last result per call id; missing results are filled in as an "interrupted" error and orphaned ones become text, so Claude always gets valid pairing; a trailing thought is stripped; reasoning.effort, max_output_tokens, text.format, tool_choice (incl. allowed_tools, custom), parallel_tool_calls. Built-in tools (web search etc.) are dropped.
  • Response: ResponsesStream turns Anthropic events into Responses SSE (response.created … output_item.added/done, text and reasoning-summary deltas, function-call argument deltas, custom tool input, response.completed/incomplete/failed, sequence numbers, usage with cached tokens). responsesResponse does the same for non-streaming clients.
• free/src/app.ts: viaClaudeKey becomes viaClaudeMessages, used for Chat and Responses clients. It translates to Claude, then sends to a Claude key directly or through the existing Claude→Chat/Gemini/Responses translators (viaTranslation split so its core takes a parsed body), and converts back. Upstream errors come back in the client's error shape. servable() now lets Chat and Responses clients use every provider; Chat clients asking for a ChatGPT model also get the ChatGPT login. Gemini clients are unchanged (next PR).
• free/src/openai-claude.ts: helpers exported for reuse, no behaviour change.

Security and safety

• No new upstream hosts: requests still go only to the configured credentials' base URLs or the ChatGPT endpoint. The client's own headers are never forwarded on translated requests (a bare request is used, as before).
• Image and file URLs in the input are passed to Anthropic as URL sources; the Worker never fetches them, so there's no SSRF.
• Claude logins are still not offered to non-Claude-Code clients (no cloaking), with a clear error.
• Body size cap (32 MB) and the one-parse rule are unchanged. A 600 KB Responses conversation translates in about 2–3 ms warm, in line with the Chat translator. Chained routes (for example Responses → Claude → Chat) re-parse the stream's events once more, which costs a little extra CPU per event.
• Upstream error messages are passed through as before; nothing new is logged or stored.

Tests

npm run verify: typecheck, lint and 170 tests pass. New free/test/responses-claude.test.ts covers request translation, streaming order and sequence numbers, custom tools, incomplete/failed streams and whole responses. app.test.ts adds Responses → Claude key (streamed and whole), Responses → OpenAI-compatible key, errors in the Responses shape, and Chat → ChatGPT login.

---

Generated by Claude Code