Context compaction becomes much slower after pausing during automatic compaction and resuming #26826

Codex app version

Desktop Codex: 26.602.40724 — released Jun 5, 2026

Platform

• macOS Tahoe 26.5.1
• Apple Silicon MacBook Pro
• Pro plan

Model / mode

5.5 Fast

Related feedback ID

019e9e1e-35fb-7d10-a46b-9de92cd86904

Summary

I encountered a performance issue with automatic context compaction when a task is paused while compaction is happening, then later resumed.

Normally, when using 5.5 Fast, I would expect automatic context compaction to complete in roughly 30-40 seconds. However, in this case, I paused/stopped the task while automatic context compaction was in progress. After resuming the task, Codex automatically continued the compaction process, but it became much slower than expected.

I watched the timing: after resuming, the context compaction took at least about 2 minutes before the first Thinking state appeared. That means the compaction process itself lasted roughly two full minutes before the model began visibly continuing the task.

Observed behavior

1. A Codex task entered automatic context compaction.
2. I paused/stopped the task while the compaction process was happening.
3. Later, I resumed the task.
4. Codex continued/triggered automatic context compaction again.
5. The compaction process took at least about 2 minutes before the first visible Thinking state appeared.

Expected behavior

Context compaction should remain reasonably fast and predictable, even if the user pauses during compaction and resumes later.

For 5.5 Fast, a typical context compaction step should ideally complete in around 30-40 seconds, or at least the UI should clearly explain why it is taking much longer.

Actual behavior

After pausing during automatic compaction and then resuming, the automatic context compaction stage became much slower than expected. It took around 2 minutes before the task showed the first Thinking state.

Why this matters

Users may pause tasks for legitimate reasons, such as moving locations, temporarily closing the laptop, switching networks, or avoiding task execution during unstable conditions. If pausing during automatic context compaction causes the compaction stage to become much slower after resume, then task recovery becomes inefficient and confusing.

This also makes the resume experience feel unreliable, because the user cannot tell whether Codex is still compacting normally, stuck, or waiting on some internal state.

Request

Please investigate whether pausing/stopping a task during automatic context compaction can leave the compaction pipeline in a slower or inefficient state after resume.

It would also help if Codex displayed clearer progress during long context compaction, such as a status indicator, elapsed time, or an explanation when compaction is taking significantly longer than expected.