Some messages output gibberish

Resolved 💬 3 comments Opened Apr 4, 2026 by LachlanStuart Closed Apr 4, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

codex-cli 0.118.0

What subscription do you have?

Plus

Which model were you using?

5.4

What platform is your computer?

Darwin 25.4.0 arm64 arm

What terminal emulator and version are you using (if applicable)?

iTerm2

What issue are you seeing?

The agent will be fine and then suddenly will switch to gibberish for a few messages. It seems to recover after I send another message. It has happened 3-4 times today. Sometimes includes random characters from other scripts (Chinese, Arabic, etc.) interspersed through the output.

What steps can reproduce the bug?

Uploaded thread: 019d5529-41aa-7142-b080-dfd7fdd91ec2

What is the expected behavior?

_No response_

Additional information

Examples:
1.

• Searching the web • Searched • Called codex_apps.github_add_reaction_to_issue_comment({}) └ Error: user cancelled MCP tool call ──────────────────── • The wrapper staytempsno trunc on—with_ni VEG identity whatsoever. I chased pre55 now: the RAP marker decoded internally, not directional. Forecast “on_step that 90 more characters” appears upstream-ish /://. So it “platform style gesture” rather than us subtracting—our new Botanical `_proxy notС text. So: no formatting beyond split-lines/strip in this trims- _emitcu. The exact q-phrase * pre/post longer: would move insofar. It's likely once browser rather.
  1. When asked about which GitHub issue it tried to react to:
• Ias, no infoman neither content report interaction orgdin in this session. The computer based read simply shoes twogi.

(Later probing showed it had made the tool call with no arguments)

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 4 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #16526
  • #15622
  • #15152

Powered by Codex Action

etraut-openai contributor · 4 months ago

Thanks for the bug report. We'll share the feedback with the team responsible for training our codex models. If you see other model behavior that you'd like to report, please use the /feedback command. That will provide us with additional details including logs and conversation text.

LachlanStuart · 4 months ago

The duplicates were correct. Pretty sure this is an infra issue, not a model training issue though.

The model was clearly trying to get its train of thought back on track but couldn't. Each burst of corruption only lasted 1 Turn, and then the model suddenly started working again after the next message from me. Sounds like RAM corruption somewhere, and the message from me was enough delay to get my context moved to another node.