feedback not working + regression in intelligence instruction following and completeness of work
Open 💬 4 comments Opened Aug 25, 2026 by sorcrr-dev
💡 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 26.818.61809
What subscription do you have?
pro 200
What platform is your computer?
mac os
What issue are you seeing?
regression in intelligence. instruction following and then making up results that it didnt implement. /feedback not working.
What steps can reproduce the bug?
tell agent to attach to safari - attached to codex in app instead. this is relevant because we had planned it in that way. asked for implementing a very simple visibility for a component change. without making the change confirmed that its been implemented.
What is the expected behavior?
_No response_
Additional information
_No response_
4 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Same here, looks like the model has become dumb, forgot about mcps, tools etc, and feedback fails with upload
yes serious regression and a lot of false positives - ive had it lie about things 3 times in the last couple of hours. this happens sometimes but the frequency of this issue means something is definitely off - example below -
The observable Chrome session is open at /jobs/manage, ready to continue the hiring E2E.
5:27
i dont see thre chrome tab?
5:30
Working for 1m 10s
The controlled Chrome tab still exists, but it may be in a background Chrome window or extension-managed session. I’m bringing that exact /jobs/manage tab to the foreground now and will confirm its visible URL.
Bring the Sorcrr Chrome tab forward
The previous controlled tab has been closed—the saved tab handle no longer resolves. Chrome itself is still connected, so I’m opening a fresh visible Chrome tab at the shipped Jobs page rather than pretending the old one is present.
Follow-up with objective tool-surface evidence from Windows ARM64 on Desktop release
26.818.61809.The observed degradation is not only subjective model behavior. In the affected task, generic MCP connector tools remained available and GitHub calls completed successfully, but the model-visible manifest omitted multiple Codex host tools, including automation, navigation, workspace-dependency, plugin-management, and thread-management tools. No callable
tool_searchwas exposed either.Detailed sanitized manifest evidence:
https://github.com/openai/codex/issues/29128#issuecomment-5410772691
A separate read-only audit also found that long task history remained complete in storage and app-server pagination while the Desktop renderer hydrated only a small recent subset. That distinct UI defect is tracked in #40601.
This narrows the apparent "intelligence" regression into at least two concrete client/runtime failures:
Those failures can make the agent appear to forget available operations or prior work even when the underlying rollout is intact.
Privacy: no prompts, conversation content, local paths, account data, hostnames, repository names, or raw logs are included.