Codex is deprecating plan mode news update

OpenAI closed developer request issue 18490 on its Codex command line interface repository without planning an implementation, rejecting user proposals to add a manual context compaction choice to plan mode menus. The discussion, documented in github.com's report, centered on whether engineers should receive a dedicated terminal action to compress session history rather than wiping previous context when executing multi-step architecture plans.

This report examines the friction between automated context handling and developer control across large software repositories containing thousands of files. Engineering teams using Codex terminal interfaces and reasoning models can evaluate how OpenAI manages context window exhaustion during complex refactoring workflows.

01Context management options in plan mode

OpenAI recently introduced a terminal prompt offering engineers the choice to clear context and implement plans during interactive sessions. Contributor guidedways opened issue 18490 on April 18, 2026, requesting an alternative option to compact context instead of erasing it. The developer noted that wiping previous turns forces the coding assistant to re-examine project directories and re-learn architectural details across repositories.

The issue tracker tagged the submission under terminal user interface, context compaction, and plan mode categories while flagging issue 18489 as a potential duplicate. Five community participants quickly reacted with approval, including Alexander Riccio, Nicholas Blauch, enniofilicicchia, Manu, and John-Henry Lim. Contributors argued that a compacting trigger preserves the agent's working memory while reducing token volume before code generation starts.

The original poster explained that utilizing plan mode with high reasoning effort produces detailed implementation steps that guide multi-file refactoring. Without a compacted history option, developers must choose between runaway token consumption or complete amnesia regarding earlier architectural discoveries. The proposed enhancement sought to bridge that gap directly inside the terminal interface menus.

02OpenAI transparent compaction roadmap philosophy

OpenAI maintainer etraut-openai responded to the request on April 18, 2026, clarifying the team's architectural direction for context management in Codex. The maintainer stated that automated compaction must remain fully transparent so that developers no longer monitor token counters or execute manual cleaning commands. Adding extra toggles to the terminal interface contradicts the goal of removing operational overhead from developers.

The maintainer acknowledged that earlier model releases and harness implementations forced developers to actively manage memory constraints during extended sessions. Recent advances in automated compaction routines now handle context decay internally without requiring interactive user steering. OpenAI indicated that manual compaction buttons would only receive consideration if an overwhelming volume of community votes demanded the configuration.

Although community interest expanded, the maintainers formally marked the enhancement closed as not planned on the repository. OpenAI leadership emphasized that investing development resources into automated context engines delivers superior long-term reliability over exposing manual levers. The stance reinforces the company's commitment to eliminating prompt engineering and memory management burdens from daily developer workflows.

03Developer friction with clear versus compact

Community members highlighted contradictions between OpenAI's automated vision and the recently shipped prompt to clear context before implementation. Developer ariccio commented on May 5, 2026, noting that clearing context discards valuable project discoveries that never made it into the written plan document. The contributor reported avoiding the clear prompt in both Codex and Claude Code to prevent catastrophic context loss.

On May 10, 2026, contributor yeliex asked for clarification on recommended best practices when plan mode consumes over half of the context window. Maintainer etraut-openai responded that automated routines handle large sessions reliably, noting personal use of plan mode without manual intervention. Community members countered that agents running in cleared environments waste unnecessary tokens re-reading local directory trees.

Contributor enniofilicicchia reiterated on June 11, 2026, that clearing active sessions causes models to inspect unchanged folders repeatedly. Retaining compacted summaries preserves execution speed while preventing the assistant from stalling on large multi-directory projects. Developers argued that offering compaction alongside clearing would solve immediate workflow degradation without adding interface complexity.

04External memory tools and plan persistence

Workflow limitations prompted independent developers to share complementary tooling and persistence scripts within the discussion thread. Participant mahdikayvan introduced AgaveCore on May 10, 2026, offering early access to mitigate agent drift through structured progress logging. The project relies on explicit commands named remember-progress, remember-this, and consult-brain to ground coding agents as session history grows.

Contributor pattty847 noted on May 20, 2026, that plan documents generated near context limits risk permanent deletion during aggressive compaction. Preserving the verbatim text of planning documents alongside automated summaries prevents the loss of implementation specifications. On August 2, 2026, developer jask-aran confirmed that server-side compaction preserves plans accurately and shared custom scripts to automate backup routines.

05Multi-file workflows and context retention limits

Engineering teams working on enterprise codebases face unique challenges when sessions span millions of lines across thousands of files. Contributor guidedways described using plan mode with high reasoning effort to decompose large architectural challenges into manageable execution stages. When context resets, the assistant loses indexed file relationships and historical debug attempts, forcing repetitive file reads that slow progress.

While OpenAI closed issue 18490 as not planned, the debate reveals persistent tension between invisible platform automation and explicit developer controls. Engineers managing extensive repositories continue to request fine-grained levers to balance token economy against historical recall. As command line coding harnesses evolve, balancing seamless automated compaction with predictable context preservation remains an active operational challenge.