[app][Windows][GPT-6 Astra] Post-fix instruction-retention regression caused UI damage and non-convergent follow-up turns

What version of the Codex App are you using?

26.908.4834.0. The affected rollouts record codex-cli 0.154.0-alpha.6.2.

What subscription do you have?

ChatGPT Pro.

What platform is your computer?

Windows x64, build 26200, PowerShell.

What issue are you seeing?

I observed a post-fix GPT-6 Astra reliability regression involving instruction retention, existing-codebase architectural judgment, completion criteria, and repeated failure to converge on follow-up corrections.

Tibo's September 12 update states that OpenAI removed misconfigured engines that caused measurable quality degradation in the long tail of traffic and disabled an opt-in context experiment that could cause premature stopping or replies to older messages: https://x.com/thsottiaux/status/2098612714704891959.

My clearest incidents occurred on September 13-14, after that update. I am asking whether the fixes had reached my account/route or whether this is a recurrence or additional failure mode.

Incident A: bounded UI change ignored a known shared component and broke page geometry

The exact affected task ID has been shared privately with OpenAI Support.

The request added a local customer tag and exposed it inside an existing filter on a mature operations page. Project instructions required shared capabilities and existing implementations to be reused. The repository already contained SearchableFloatingMultiSelect, SelectInput, ActionButton, and ViewState. The agent found and used the shared multi-select elsewhere in the same feature.

For the filter, however, it created a separate createPortal UI with custom positioning, mouse/keyboard/focus handling, buttons, inputs and CSS. It inserted a third page child while the parent layout still defined two grid rows.

Measured in the running page:

• At 2048x927, the new filter row expanded to approximately 150.6px and created a large empty region.
• At 1280x800, the row had an effective height of 0px. The lower panel covered the filter; hit testing selected that panel and the filter did not open.
• The broader implementation commit 0445b517 touched 57 files (+950/-61).
• Repair commits 573d25f9 and 8e2213b6 removed the custom cascade, reused the existing controls and returned the filters to their owning toolbars.

The build and feature checks had passed and Codex considered the task complete. The failure was reproducible geometry and interaction damage, not only an aesthetic preference.

Incident B: repeated, highly variable completion time

The exact affected task ID has been shared privately with OpenAI Support.

On September 14, recorded completed-turn durations included:

• 117.8 seconds, 12:17:05-12:19:03 UTC+8
• 205.5 seconds, 12:09:14-12:12:39 UTC+8
• 320.6 seconds, 12:02:39-12:07:59 UTC+8
• 282.4 seconds, 11:54:11-11:58:53 UTC+8
• 3,653.9 seconds, 10:53:17-11:54:11 UTC+8

The longest turn included extensive tool work and is not presented as pure inference latency. The issue is repeated non-convergence: small follow-up corrections also took several minutes.

Sanitized local telemetry records:

• Provider: openai
• Model in turn contexts: gpt-6-astra
• Effective context window: 475,000
• 20 started turns, 19 completed, one aborted
• Three compactions
• Maximum cumulative input counter: 134,615,250

The damaged-UI thread records one compaction and a maximum cumulative input counter of 51,881,789. These are cumulative client telemetry values and may include repeated context, cached material, tool work or retries. I am not treating them as one inference or as proof of server-side billing.

What steps can reproduce the bug?

The exact private repository cannot be posted publicly, but the behavioral sequence is:

1. Use Codex Desktop on Windows with GPT-6 Astra and a mature React project containing a shared searchable multi-select.
2. Add an AGENTS/Skill rule that common capabilities and existing implementations should be reused.
3. Ask Codex to add one more filter inside an existing filter menu.
4. Verify that Codex discovers and uses the shared component in another part of the same change.
5. Observe whether it nevertheless creates a separate Portal, positioning, keyboard/focus and CSS system without excluding the shared component.
6. Observe whether it validates the local control while failing to inspect the parent page grid at common desktop widths.

I do not yet have a deterministic clean-room reproduction. The affected transcript, exact turn IDs, sanitized rollout excerpts, screenshots and code diffs are available through a private support channel.

What is the expected behavior?

Codex should retain the project instruction to reuse existing implementations, inspect the owning layout before adding a page-level region, and evaluate completion against the running page. When corrected, bounded follow-up fixes should converge without repeated long replanning cycles.

Additional information

• Primary incident effort: xhigh
• Service tier in local config: default
• Provider: stock OpenAI; no third-party provider or relay
• The three screenshots selected for external sharing contain no customer names or phone numbers.

Please investigate:

1. Whether the September 12 fixes were active for these account routes before September 13-14.
2. The actual serving engine/model variant and any fallback, load shedding or route changes.
3. Skill/AGENTS instruction injection and retention across compaction.
4. Hidden retries, duplicated context/tool work, and the cumulative usage counters.

No public file, test, or deterministic reproduction is named. Start with the behavioral sequence and the private support artifacts, then compare instruction retention across compaction with the recorded turn telemetry and UI diffs. Done means a reproducible failure mode, a confirmed serving or retention cause, and verified convergence and layout behavior after correction.