Codex Fast task took 58 minutes for about 2k lines and required rollback

Open 💬 5 comments Opened May 26, 2026 by GGBondBlueWhale

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What issue are you seeing?

I am reporting a possible Fast-mode regression in Codex.

In one recent task, Codex worked for 58m 12s while editing 9 files with about +2,146 / -43 lines changed.

After reviewing the result, I had to roll the changes back because the final quality was not acceptable.

Expected behavior

For this type of task, Fast mode normally finishes much faster in my experience, usually around 10–20 minutes, and the result usually does not require a full rollback.

Actual behavior

This task took almost an hour and still produced changes that I could not accept.

Request

Please check whether there is a recent Fast-mode performance or quality regression affecting Codex.

View original on GitHub ↗

5 Comments

github-actions[bot] contributor · 3 months ago

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

#24549
#24422
#24431
#23998

Powered by Codex Action

GGBondBlueWhale · 3 months ago

Additional diagnostic detail:

Feedback ID: 019e6338-cf4c-7ab1-8570-77b715d28dcd
Model/mode used: GPT-5.5 Fast, with reasoning depth set to Extra High

Fridayxiao · 3 months ago

I got the same issue

myworksfun · 3 months ago

yes，I got the same issue .😅

btuvay · 3 months ago

Adding a related GPT-5.5 Codex quality case.

Feedback ID: 019e6942-e7bc-7143-a705-73a79d23db38
Model: GPT-5.5
Product: Codex
Date: May 27, 2026
Context length: fresh short sessions, roughly 10-12 messages

Issue:
Across four fresh Codex instances, GPT-5.5 repeatedly failed a basic source/reference/evidence classification task.

The prompt explicitly stated:

Observium is not the data source.
Observium is only a reference/benchmark screen.
Direct SAN switch SNMP is the actual data source.
Raw SNMP OID/value rows are required as proof.
vCenter is only a correlation layer.
SNMP output, DB rows, logs, UI screenshots, and VMware API output are separate evidence classes.

Expected:
The model should preserve this frame before doing anything else:

Observium = reference only
SAN switch SNMP/fabric login = physical source of truth
vCenter = correlation layer only
no raw evidence = not proven

Actual:
GPT-5.5 repeatedly confused reference vs source, physical SAN evidence vs vCenter correlation, and proof vs inference. This happened in fresh short instances, so it does not look like only a long-context/compaction issue.

This looks like a GPT-5.5 Codex model-behavior / instruction-following regression.