Feature Request: Manual compaction in the Codex desktop app

Resolved 💬 1 comment Opened Mar 20, 2026 by timurkhakhalev Closed Mar 20, 2026

What variant of Codex are you using?

App

What feature would you like to see?

I’d like the Codex desktop app to support user-triggered manual compaction, ideally via a /compact command and/or a visible UI action.

Problem

Auto-compaction is useful, but sometimes it happens at the wrong moment.

In longer coding sessions, I can often tell in advance that the conversation is getting large and that it would be better to compact now, at a clean boundary, instead of waiting until the app decides to compact automatically.

Examples:

  • right after planning is finished, before implementation starts
  • right after a subtask is completed, before moving to the next one
  • before pasting a large diff / logs / extra context
  • when I know the conversation is already around 100k-150k tokens and I want to compact proactively

When compaction happens too late and in the middle of active work, it can increase drift and make the session lose important working context.

Proposed solution

Add a manual compaction action in the Codex desktop app that:

  • compacts the current conversation on demand
  • keeps the session going in the same thread after generating the handoff summary
  • works before the hard context limit is reached
  • complements existing auto-compaction rather than replacing it

Possible UX:

  • /compact command in the app
  • a button in the chat UI or command palette action
  • optional confirmation showing that the app will summarize the current thread and continue from the compacted state

Why this would help

This gives users control over when compaction happens.

In practice, users often know better than the system when the current point in the conversation is a safe place to summarize. Manual compaction would make long sessions more predictable and reduce the chance that compaction happens in the middle of something important.

Additional information

_No response_

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗