Codex auto-compact loses critical task details and causes wrong execution, even on small fixes

What version of the Codex App are you using (From “About Codex” dialog)?

codex 26.415.40636

What subscription do you have?

pro

What platform is your computer?

windows

What issue are you seeing?

I want to report a product bug in Codex app related to "auto-compact / handoff", not model capability.

After working with Codex heavily for about two weeks, I initially thought GPT-5.4 was "not following instructions." After more testing, I found the real issue: when the context window is already highly filled and Codex triggers "auto-compact during execution", the handoff loses important task details. After that, the model continues working from an incomplete internal summary and starts solving a different problem than the one I asked for.

This is not specifically a long-running task issue. I can reproduce it even on a very small fix.

A concrete example:
This happened on a very small UI fix, only about "17 lines of code changed".

My real request was to improve the selected-node animation "without changing the concept of the effect".

The model initially understood this correctly. Early in its reasoning, it stated:

"I’ll fix the animation quality of the selected node specifically: remove the line through the center and replace it with a more professional edge effect along the card border, without touching the map logic."

That was correct.

Then auto-compact happened during execution.

In the final response, it said:

"I removed the masked conic ring that caused the center artifact and replaced it with a small light 'comet' rotating around the outer edge of the card. Glow and node selection logic were unchanged; only the active orbit rendering method changed."

This is the key bug:
the original task was to preserve the intended edge-animation concept and fix the artifact, but after compact the model changed the actual animation approach and delivered a different solution than requested.

So the failure was not "bad coding" or "bad reasoning."
The failure was that "the purpose of the fix was lost during handoff".

What steps can reproduce the bug?

1. Work in Codex long enough that the context window becomes highly filled, around "90%-97% full".
2. Start a new request while the context is already near full.
3. Use a task that the model initially understands correctly, including a constrained fix where the exact intent matters.
4. Let Codex continue until "auto-compact" happens during execution.
5. Compare the original request with the final result after compact.

What I observe during reproduction:

• Before auto-compact, the model understands the task correctly.
• During execution, auto-compact happens.
• After compact, the model continues from a degraded internal handoff.
• The final result is often adjacent to my request, but not the actual requested fix.
• This makes it look like the model is "ignoring instructions," but the real issue is context loss during compact/handoff.

What is the expected behavior?

Auto-compact / handoff should preserve the full intent, constraints, and priority of the original request, even when the context window is already highly filled.

If the model understood the task correctly before compact, it should continue solving that same task after compact without changing the requested approach.

In the example above, the expected behavior was:

• preserve the intended edge-animation concept,
• fix the visual artifact,
• and not replace the animation with a different effect.

The handoff summary should not become incomplete or distorted in a way that changes execution.

If there were an option to trigger handoff and compact manually, that would likely solve the entire problem.

Additional information

The model may still read agents.md after auto-compact, but the problem is not that it stops reading instructions entirely. The problem is that the handoff summary itself becomes incomplete or distorted, so the execution path changes.

Because of that, this can look like an instruction-following problem, but in my testing it appears to be a product-level compact/handoff issue instead.

In most of my work, GPT-5.4 followed the intended plan very closely, especially on smaller tasks. On larger refactors or broader-scope requests, there were occasional cases where it deviated somewhat, but those were relatively rare. In my experience, the more serious failures began specifically after auto-compact occurred during execution of the plan.

Thank you for your time and attention to this issue.