Codex App: benchmark-driven context-drift warning and one-click HANDOFF.md migration to a new task

What variant of Codex are you using?

Codex App for Windows, version 26.727.6591.0.

What feature would you like to see?

Summary

Please add a benchmark-driven context-continuity warning and a one-click handoff to a fresh Codex task.

When OpenAI's long-session testing indicates that repeated compaction is approaching a meaningful risk of context drift, Codex should:

1. warn the user that continuing in the current task may reduce continuity;
2. let the user choose a handoff-summary depth (10%, 50%, or 100%);
3. generate a user-reviewable HANDOFF.md;
4. create a new Codex task;
5. make the new task automatically read and acknowledge the handoff; and
6. open the new task for the user so work can continue immediately.

Problem observed in real use

In a long Codex App task with hundreds of exchanges and many automatic compactions, I observed that:

• the task became slower and consumed more usage as the context remained full;
• the broad objective often survived, but detailed agreements, user corrections, assistant commitments, completed-versus-pending state, and the exact next action could gradually drift;
• setting a persistent goal helped, but did not fully solve the continuity problem;
• manually moving to a new task did not transfer the complete working context;
• reconstructing the old task required reading it by task ID, paging through history, and expanding long turns, and assistant messages could still be missed.

A single task can compact many times—even hundreds—so compaction count alone should not be treated as failure. Users also should not be expected to guess a universal safe number.

OpenAI should stress-test representative long-running, tool-heavy tasks across models and task types to measure when repeated compaction begins to cause meaningful continuity degradation. The product can then use those measured results to decide when to show an advisory handoff prompt. The trigger may be dynamic rather than one hard-coded number.

Proposed user flow

1. Detect a tested continuity-risk boundary

Use official long-session regression tests to evaluate whether Codex still preserves:

• the active objective and acceptance criteria;
• recent user corrections and constraints;
• decisions and rejected approaches;
• completed work and its evidence;
• assistant commitments that were stated but not executed;
• unresolved work and the exact next action.

When continuity risk reaches a tested threshold, show a non-blocking prompt such as:

This task has been compacted repeatedly and may begin losing working context. Create a handoff to a new task?

This should be a recommendation, not a forced cutoff after an arbitrary number of compactions.

2. Let the user choose handoff depth

Offer three clear choices:

• 10% — Compact handoff: current objective, current state, blockers, and next action.
• 50% — Standard handoff: current phase plus major decisions, corrections, evidence, completed work, and pending work.
• 100% — Full-history handoff summary: summarize the interaction from the beginning through the current turn, including important user requests, assistant commitments/actions, decisions, evidence, and unresolved items.

The percentages describe summary coverage/detail, not a promise to copy the raw transcript word-for-word.

3. Generate HANDOFF.md

The handoff should be user-reviewable and contain, at minimum:

• current goal and scope;
• durable constraints and permissions;
• major decisions and rejected approaches;
• completed work with evidence;
• assistant promises not yet executed;
• current files/artifacts and their status;
• unresolved questions or blockers;
• exact next safe action.

4. Create, initialize, and open the new task

After the user confirms:

• create a fresh Codex task;
• preserve a link back to the original task;
• automatically supply/read HANDOFF.md in the destination task;
• have the destination task confirm what it inherited before acting;
• open/jump to the destination task automatically;
• keep the original task intact and available for audit.

The user should not need to copy and paste, enter an old task ID, page through the old conversation, or manually expand every long turn.

Expected result

The user gets fresh context headroom without losing the working state of a long task. Handoff becomes a first-class product workflow instead of a manual recovery procedure.

Related issues

These cover important parts of the problem, but not the complete App workflow above:

• #29356 — context compaction can lose operational continuity;
• #34095 — repeated compaction can degrade the execution frontier;
• #22220 — compaction telemetry/context health;
• #33310 — atomic handoff during context rollover;
• #35032 — compaction may leave too little usable headroom.

This request specifically combines a tested drift-risk prompt, user-selectable summary depth, automatic HANDOFF.md generation, and one-click creation/opening of a new initialized Codex task.

Additional information

OpenAI recently described retained reasoning and compaction as settings used in ChatGPT and Codex, and showed that they improved continuity and efficiency in a long-running ARC-AGI-3 harness. This request does not argue against compaction; it asks for a product-level safety valve when very long, repeatedly compacted tasks begin to lose continuity:

https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/