Codex fails to reconstruct the global plan after repeated corrections across dependent deliverables

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

Version 26.818.61809

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Summary

During a long task involving several dependent deliverables, Codex applied each human correction locally but failed to reconstruct the global document architecture.

The task involved an email, a PDF, and several dependent examples whose medium, role, order, and content were defined by a common plan. After one element was corrected, Codex produced a plausible local repair but introduced a contradiction elsewhere. This happened repeatedly.

Four human corrections were required before Codex returned to the original plan and restored the intended roles and ordering of the deliverables.

This report is not about hidden reasoning content failing to display. It concerns failure to detect and recover from global architectural drift during a multi-document task.

Actual behavior

Codex treated each correction as a local editing request. Each repair was individually plausible, but it could invalidate another part of the deliverable set. Codex did not detect the accumulating contradictions, reconstruct the canonical plan, or perform a global consistency check.

Impact

The output can remain locally convincing while becoming globally inconsistent. This creates repeated correction cycles, wasted time, and reduced confidence in complex multi-document work.

What steps can reproduce the bug?

1. Start a task containing several dependent deliverables, such as an email, a PDF, and examples derived from the same canonical plan.
2. Define distinct roles, media, ordering, and dependencies for those deliverables.
3. Ask Codex to produce or revise them over several exchanges.
4. Correct an error affecting the role, order, medium, or function of one deliverable.
5. Continue making corrections when the local repair introduces contradictions in other deliverables.
6. Observe that Codex handles the corrections separately without reconstructing and validating the complete architecture.

No session ID or private task content is included in this public report.

What is the expected behavior?

When several corrections affect the order, medium, role, or function of dependent deliverables, Codex should detect a risk of global architectural drift.

It should then:

1. pause local execution;
2. reconstruct the latest canonical plan;
3. map every deliverable to its intended role, medium, dependencies, and order;
4. identify contradictions created by previous local repairs;
5. ask for clarification if the canonical intent is uncertain;
6. perform a global consistency check before continuing;
7. consider a deeper reasoning pass when the dependency structure has materially changed.

Additional information

Related issues:

• #40584 concerns the feedback-submission failure encountered while trying to report this behavior.
• #39892 covers the feedback-upload and delivery-state problem.
• #40553 concerns reasoning content display and is not a duplicate of this issue.
• #40591 is adjacent instruction-following behavior, but it does not describe failure to reconstruct a shared architecture across dependent deliverables.

This issue intentionally excludes personal information, private task content, local paths, account data, logs, and private files.

No source files, tests, or entry points are named. Start by reproducing the multi-deliverable correction sequence described in the issue and inspect the handling of dependent plans and revisions. Done means the behavior detects architectural drift, reconstructs the canonical plan, identifies contradictions, and validates consistency before continuing.