Reproducible quality regression in Codex since Aug 28 (Pro, sol / high)

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

0.151.0-alpha.7.2

What subscription do you have?

Pro

What platform is your computer?

Windows 11

What issue are you seeing?

Since Aug 28 there's been a consistent, ongoing drop in output quality that hasn't improved since. Symptoms: instructions frequently ignored, generated code often doesn't run, context lost mid-task, tasks reported as "complete" when they clearly aren't. Same machine/setup, happening across multiple unrelated projects/sessions — not a one-off.

Already submitted diagnostics/logs via /feedback (ID above). Could you confirm whether anything changed with model routing, quantization, or capacity allocation for this model/tier around Aug 28? Happy to provide a concrete repro example if that would help.

No single reproducible code snippet captured yet — this shows up as a general behavioral regression across unrelated day-to-day tasks/prompts rather than one specific script.

Feedback ID: 01a056e8-a7c8-78a0-b2cc-94a66be12daf (diagnostics already submitted via /feedback)

Pattern observed, consistently since Aug 28: the model ignores instructions in the prompt/AGENTS.md, generated code frequently doesn't run, context is lost partway through a task, and the task is reported as "complete" when it clearly isn't. Happens across multiple sessions/projects on the same machine, not a one-off.

Session/Thread: 01a056e8-a7c8-78a0-b2cc-94a66be12daf
Context: 83% left (43,677 used / 258K)
7d limit: 93% left (resets Sep 6)

Happy to capture another session's /status output if a fresh example would help narrow it down.

What steps can reproduce the bug?

Feedback ID: 01a056e8-a7c8-78a0-b2cc-94a66be12daf

What is the expected behavior?

No response

Additional information

No response

No source file or test is identified. Start by reviewing feedback/session ID 01a056e8-a7c8-78a0-b2cc-94a66be12daf and the submitted diagnostics, then capture the suggested /status output and a concrete reproducible example. Done would require isolating and correcting the reported quality regression, but the issue does not define a specific expected behavior.