User-configurable rolling context limit

What variant of Codex are you using?

iOS app, macOS app, linux cli

What feature would you like to see?

Please add a user-configurable rolling context limit to ChatGPT Work tasks.

Long-running Work tasks can accumulate ever-increasing conversation/tool context, which substantially increases usage and cost even when most of the older context is no longer useful.

Ideally, Work should let users configure:

• a maximum active context size, e.g. 64k/128k tokens;
• an automatic compaction threshold;
• retention of fixed task instructions, compacted task state, and only the most recent turns/tool results.

Codex already supports automatic context compaction, so exposing equivalent controls in ChatGPT Work would make persistent and recurring tasks much more economical and predictable to run.

Simply instructing a task to ignore old context does not solve the problem if that context is still sent to the model and counted towards usage.

Additional information

No response

The issue does not name implementation files, tests, or entry points. Start by locating Codex's automatic context compaction and the handling of ChatGPT Work task context; done should include configurable maximum context size, compaction threshold, and retention of task instructions, compacted state, and recent turns or tool results.