Truncated model response happened multiple times today

Open 💬 3 comments Opened Jul 8, 2026 by Meowzz95
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.623.141536 • Released Jul 7, 2026

What subscription do you have?

Pro $200

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

I have see multiple times (>3) that the model return truncated response, I have attached the thread ID

<img width="1518" height="572" alt="Image" src="https://github.com/user-attachments/assets/484be3c1-3573-47e8-9fd0-9a1e8a0fbf66" />

As a preview, something like the above screenshot

I have checked the session rollout file to confirm this is not a UI render issue, the agent_message is exactly like that

What steps can reproduce the bug?

Feedback ID: 019f40d9-0579-73b2-862b-418ed6e8fbd7

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 1 month ago

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

  • #31340

Powered by Codex Action

Meowzz95 · 1 month ago

I notice that it seems to always truncate at backtick '`'

<img width="1048" height="304" alt="Image" src="https://github.com/user-attachments/assets/91b93dcf-49d0-4167-b93c-1549b0fb9bcb" />

Meowzz95 · 1 month ago

Follow-up with the root cause I found locally: this does not look like model/API truncation. In the rollout, the completed response_item contains the full assistant answer, but the visible/event path (event_msg.agent_message) is shorter.

The cut happens when assistant text mentions the hidden memory-citation opener (<oai-mem-citation>) in normal inline prose/code. The assistant text stream parser treats that as real hidden citation markup; because there is no matching closing tag in the visible prose, it auto-closes at EOF and strips the rest of the message from the displayed/event text.

So the issue appears to be in the hidden-markup parsing/display path, not in response generation itself.