Codex needs better long-session project continuity and context transition UX

Hi everyone,

after working with Codex on larger long-running projects, I noticed a major workflow issue that becomes very painful during real creative and development work.

The problem is not the model quality.

The problem is project continuity when the context window becomes full.

Right now, once the context is exhausted:

  • the conversation can suddenly stop
  • compacting may fail
  • forking may no longer work properly
  • continuing the same project becomes difficult
  • important thinking context can become fragmented

As a user, I expected the system to help guide this transition much earlier and more clearly.

What feels missing is a proper “project continuity” workflow.

For example:

  • automatic or assisted project snapshots
  • selective context cleanup
  • “continue project” transition flow
  • project-level persistence separate from raw chat context
  • graceful context compression before hard failure
  • lightweight project memory/state transfer between threads

The current system feels very conversation-based, while many real workflows are actually long-running thinking and development spaces.

This becomes especially important for:

  • creative work
  • architecture/design thinking
  • research workflows
  • iterative development
  • long-term collaboration with Codex

I think Codex could become dramatically more usable if project continuity and context transitions were treated as a first-class UX problem, not only as a token limit issue.

I’m sharing this because I believe many users working on larger projects will eventually hit the same wall.

Would love to hear how others currently manage long-running Codex projects and context transitions.

Welcome to the forum!

FYI

I have been using ChatGPT Pulse to reivew the commits for Codex and give me daily updates. The reasong being I have been watching the Codex commits manually for many months and there are too many to read, so I have ChatGPT analyze them and report back.

Todays report, similer to many for the last few weeks notes

The direction increasingly suggests:

  • isolated runtime identities,
  • dynamic workspace mounting,
  • multi-root execution,
  • safer agent orchestration,
  • and enterprise/team-oriented execution models.

and then notes many additial focus areas, one being

Shared execution/runtime infrastructure keeps consolidating

The ToolExecutor consolidation still looks strategically important.

Key commits:

  • feat: extract shared tool executor interface (#22359)
  • Refactor extension tools onto shared ToolExecutor (#22369)

This likely means:

  • standardized execution lifecycles,
  • centralized orchestration semantics,
  • and reduced duplication across:
    • app-server,
    • plugins,
    • MCP tooling,
    • desktop integrations,
    • and remote/mobile execution surfaces.

The repository increasingly resembles:

  • a generalized developer-agent runtime platform.

Notice: centralized orchestration semantics

HTH

Also see the GitHub OpenAI Codes Repository Discussions