Accessibility: screen reader focus jumps to the top while Codex is generating output

Open 💬 3 comments Opened May 15, 2026 by aim9sour
💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

Summary

I am reporting a serious accessibility issue affecting screen reader users in Codex.

When I write a prompt and send it, then try to read the model output while it is being generated, the interface repeatedly moves my screen reader/reading position back to the very beginning of the conversation.

This makes it extremely difficult to follow the latest generated output, because I have to keep navigating back down to the end of the conversation again and again.

Steps to reproduce

  1. Open Codex.
  2. Use a screen reader.
  3. Write a prompt and submit it.
  4. While Codex is generating the response, try to read/follow the latest output in the conversation.
  5. Press the button/control that should jump to the latest message or bottom of the conversation.

Actual behavior

  • The reading position/focus is repeatedly moved back to the top or beginning of the conversation.
  • The latest output becomes difficult to follow while it is being generated.
  • The jump-to-latest or scroll-to-bottom control does not keep the user at the latest output reliably.
  • This creates a very frustrating experience for screen reader users.

Expected behavior

  • Screen reader focus/reading position should not be reset to the top while Codex is generating output.
  • The user should be able to stay near the latest generated output.
  • The jump-to-latest or scroll-to-bottom control should work reliably with keyboard and screen readers.
  • New generated content should remain reachable/readable without stealing focus or resetting the reading position.

Impact

This is a major accessibility issue because it prevents screen reader users from reliably reading and following active Codex output as it is generated.

Environment

I can provide more details if needed, but the main issue is reproducible while using Codex with a screen reader and trying to read the generated output in real time.

View original on GitHub ↗

3 Comments

etraut-openai contributor · 3 months ago

Which variant of Codex are you using? The desktop app? The IDE extension? The TUI?

aim9sour · 1 month ago

This is reproducible in the Codex Desktop app on Windows.

Environment currently tested:

  • Codex Desktop client 26.715.52143 (Windows package 26.715.7063.0)
  • Microsoft Windows NT 10.0.19045.0 x64
  • NVDA 2026.1.1 (latest stable; portable copy, ProductVersion 2026.1.1 / FileVersion 2026.1.1.55980)
  • Also reproduced with JAWS and other screen readers across multiple versions

Additional details:

  • While reading the newest model output near the bottom of a long chat, the screen reader's reading position/focus unexpectedly jumps far upward in the conversation without any user action or announcement.
  • The visual scroll position also moves away from the latest response, and the “scroll to bottom” button appears.
  • This happens repeatedly, not just once when generation begins.
  • It has been present since the earliest Codex Desktop releases I tested and has never worked reliably for me in any release.
  • The issue is not specific to one NVDA version or one screen reader, which suggests that the app is replacing/reordering conversation DOM/accessibility nodes or programmatically changing scroll/focus during streamed updates.

Expected: streamed output should update without moving the user's virtual cursor, accessibility focus, or scroll position. The “scroll to bottom” control should appear only after intentional user scrolling, not because a render update displaced the reading position.

bhavyashah · 1 month ago

@etraut-openai This is also reproducible with the VSCode extension and has a significant impact for screen reader users.