Summary

• strip residual fenced blocks like image, audio, or ui before passing messages to marked
• add reusable markdown sanitizer and unit test to verify code blocks and <hr> still render
• wire sanitizer into chat message rendering and load via new module

Testing

• npm test (fails: textModels returns JSON — fetch failed, text(prompt) returns string — fetch failed, chat basic response — fetch failed, search convenience returns text — fetch failed, imageModels returns JSON — fetch failed, image fetch small blob — fetch failed, mcp generateImageBase64 returns base64 — fetch failed, vision with data URL — fetch failed, audio.tts returns audio blob — fetch failed, mcp list helpers return arrays/objects — fetch failed, pipeline end-to-end — fetch failed)
• node tests/markdown-sanitization.mjs

https://chatgpt.com/codex/tasks/task_b_68c5f4bcbf90832facf08d8e9f9a741d