Codex app duplicates words and phrases throughout assistant responses

Open 💬 2 comments Opened Jul 18, 2026 by joe-cole1

Summary

The Codex app duplicated words, phrases, punctuation, and larger spans throughout assistant responses. The corruption affected both normal prose and a fenced text/code block, and persisted into the next response in the same thread.

It is unclear whether the duplication occurs in generated output, streaming assembly, or client rendering.

Actual behavior

Examples visible in the captured screenshots:

You’reYou’re right right.. I I turned turned “detailed” into “detailed” into a giant context dump giant context dump,, and the and the formatting formatting became became mangled. mangled.

And inside a fenced block:

ContinueContinue my my Home Assistant Home Assistant automation automation audit audit using using the the live live server server as as the the only only source source of of truth truth.

The duplication varies between individual words and multi-word spans. It is not limited to Markdown formatting.

Expected behavior

Each word and phrase should appear once, with normal punctuation and Markdown rendering.

Reproduction

  1. Open a long-running Codex app thread.
  2. Ask Codex to generate a detailed handoff prompt for a follow-up chat.
  3. Observe the response while/after it streams.
  4. In this case, extensive duplication appeared throughout the response.
  5. Send a follow-up asking for a corrected shorter response.
  6. The next response also exhibited the same duplication.

Environment

  • Codex app connected to a Windows desktop workspace
  • Model label shown in the UI: GPT-5.6-Sol High
  • Review mode shown: Auto-review
  • Occurred on 2026-07-17
  • Affected an established, long-context thread
  • Two screenshots were captured showing the corruption

Impact

The responses became effectively unreadable and unusable. The duplication also made an already long response substantially larger, which may unnecessarily consume visible/context capacity if the duplicated text is present server-side rather than being only a rendering defect.

Additional notes

  • The issue reproduced across two consecutive assistant messages.
  • It affected both prose and monospaced fenced content.
  • The screenshots were provided in the originating Codex thread, but the available GitHub issue API did not support uploading the local image files. They can be supplied if maintainers provide an upload path.

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗