Codex Desktop loses cumulative operator constraints during long repository sessions

Codex App version

26.707.9981.0 (Windows x64)

Surface and session

Codex Desktop on Windows, long-running local repository task.
Session ID: 019f638a-d8b8-7092-a6b5-0660f3aa9982

What issue are you seeing?

During a long task, Codex preserved only part of a cumulative operator contract. After context compression/continuation, it could recite parts of the contract but acted from an incomplete summary: it omitted mandatory documentation-before-code rules, reintroduced explicitly excluded work, and reduced the active delivery task to a few generic steps.

This was not a single missed detail. The user repeatedly restated the full cumulative contract, but later turns still behaved as if only the most recent message or an incomplete summary mattered.

Steps to reproduce

1. Start a long Codex Desktop repository task with many explicit cumulative constraints.
2. Add corrections over multiple turns. State that later messages modify only named parts and all other rules remain active.
3. Continue through tool calls, subagents, and context compression.
4. Ask Codex to resume the task without restating every constraint.
5. Observe that Codex can describe the task approximately while omitting or contradicting earlier mandatory constraints during execution.

Expected behavior

Codex should preserve a structured cumulative task contract across long sessions and compaction. Later instructions should modify only the addressed portions unless the user explicitly replaces the whole contract. Before acting after compaction, Codex should re-anchor to the complete active contract, not an incomplete prose summary.

Actual behavior

The agent drifted into repeated planning and disallowed work categories, omitted required documentation steps, and had to be corrected repeatedly. This caused substantial duplicated work and token consumption.

Related reports

Related to #8310 and #21468, but this reproduction is Codex Desktop during one continuously active long repository session, not only rate-limit resume or manual /compact visibility.

Privacy

The public report intentionally excludes repository secrets and private transcript contents. OpenAI staff can use the session ID for the full reproduction.