@nevescloud/chat-compat

One source of truth for the OpenAI-Codex (chatgpt.com Responses API) and Gemini-Code-Assist (cloudcode-pa.googleapis.com) ↔ Chat Completions translation. Pure transforms, no I/O — runtime-agnostic (CJS), so the same logic runs unchanged in a Node http server and a Cloudflare Worker (fetch). Each runtime keeps its own streaming I/O and wraps these.

npm i @nevescloud/chat-compat

const { openai, gemini, drainSSE } = require('@nevescloud/chat-compat');         // Node
// import cc from '@nevescloud/chat-compat'; const { openai, gemini, drainSSE } = cc; // ESM/Worker (default import — CJS)

• openai: normalizeModel · resolveCodexModel · chatToResponsesPayload · responsesToChatCompletion · newCodexStream + transformCodexEvent (per-SSE-event → Chat-Completion chunks).
• gemini: resolveGeminiModel · toCodeAssistRequest · isRetryableGeminiError · extractGeminiText · geminiTextChunk.
• drainSSE(buffer) → { events, rest } — drains \n\n-delimited SSE events, keeps the remainder.

Streaming pattern (per runtime): accumulate bytes → drainSSE → per event call the transform → write returned chunks → on done/stream-end emit data: [DONE]\n\n.

npm test — golden tests lock the exact output.