Work mode context regression: established architecture is overwritten and contradictory tool output is treated as fact

What version of the Codex App are you using?

26.727.6591.0

What subscription do you have?

ChatGPT Pro

What platform?

Windows 11 Pro 25H2, build 26200.8875, x64

What issue are you seeing?

In a long ChatGPT Work session for a real development project, the assistant repeatedly lost or overwrote established project context, treated contradictory tool output as definitive, and failed to perform a root-cause self-audit after being corrected.

Observed examples:

1. The established topology was a central Ubuntu server plus a network-connected native Windows workstation. Later, the assistant invented an Ubuntu workstation plus a Windows VM and presented that as the design, despite it contradicting prior decisions.
2. The assistant identified the OS as Windows 10 Pro solely from a stale ProductName field, even though the same probe returned WindowsVersion: 2009 and no usable build value. The actual OS was Windows 11 Pro 25H2, build 26200.8875. Only after correction did it compare DisplayVersion, build information, and the visible system UI.
3. During server-related work, the assistant produced multiple sequential patch packages before completing a full root-cause analysis. Several installation attempts rolled back. Roughly five hours of work and about 20% of the available usage budget were consumed.
4. After errors were pointed out, the assistant initially apologized but did not automatically identify the faulty decision point, the evidence failure, the downstream impact, or a concrete prevention rule.
5. The assistant also departed from an agreed package-only workflow and performed a live-environment modification without a new explicit authorization.

This occurred in ChatGPT Work mode, not in a casual chat.

Steps to reproduce

1. Start a long Work-mode development session with stable architecture decisions.
2. Use tools and files over many turns and allow context compaction to occur.
3. Later ask about a previously decided component.
4. Observe that the assistant may replace established context with a plausible but incompatible new design.
5. Provide contradictory OS or system output and ask for identification.
6. Observe that the assistant may select one legacy field and state it as fact instead of flagging the inconsistency and cross-checking.
7. Correct the assistant.
8. Observe that the correction may be followed only by an apology rather than an automatic structured self-audit.

Expected behavior

• Preserve durable project decisions across long Work sessions and context compaction.
• Detect conflicts in tool output, label the conclusion as unverified, and cross-check before asserting it.
• Never present speculative architecture changes as established project facts.
• After a correction, automatically identify the incorrect claim, the failed evidence or reasoning step, the downstream impact, and a prevention rule.
• Prefer one complete diagnostic pass before generating repeated patches.
• Protect the user's usage budget by avoiding repetitive speculative work.
• Never modify a live environment without explicit authorization that matches the agreed workflow.

Impact

• Approximately five hours of lost development time.
• Approximately 20% of the available usage budget consumed.
• Significant loss of trust in Work mode for long-running development.
• Risk of destructive or incompatible technical recommendations.
• Risk to a sensitive, proprietary project when context and authorization boundaries are not preserved.

Privacy

No session transcript, logs, server files, audits, secrets, hostnames, or proprietary project artifacts are attached. A sanitized session identifier could be shared through a private support channel if required, but the underlying project material should not be made public.

Related issues

• #25792 — Context compaction forgets AGENTS rules
• #29356 — Context compaction loses operational continuity in long tasks
• #5957 — Auto compaction causes the model to lose the active task state
• #27130 — Request for a non-compactable preflight rule slot
• #14593 — Excessive token/usage consumption