High API error rate during remote compaction; now ALL Codex CLI API calls fail with “high demand” error last ~2hrs

What version of Codex CLI is running?

v0.115.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.4, gpt-5.3 codex

What platform is your computer?

WSL and Linux vm both give the same error

What terminal emulator and version are you using (if applicable)?

WSL and Xfce

What issue are you seeing?

For about the last five days, I have been experiencing a significantly elevated rate of API failures in Codex CLI, especially when compaction is involved.

The problem first showed up around Friday or Saturday and has progressively worsened. By Monday, my rough estimate was that only about 50% of remote auto-compactions were succeeding. Manual compaction via /compact also frequently failed.

The two most common errors I have seen are:

1. "We're currently experiencing high demand, which may cause temporary errors" 2. "stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses)"

An annoying temporary workaround was to switch to a cheaper, lower-reasoning model, ask for a status report (which appears to trigger a compaction), and then switch back to the original model. That sometimes helped earlier, but it was unreliable.

For roughly the last two hours and counting, the issue became much worse: none of my API calls were succeeding at all, not just compaction-related ones. At that point, every request was failing with the “high demand” error above. I switched to my linux VM and got the same issue.

This appears to be a persistent backend or compaction-path reliability issue affecting both automatic compaction and manual /compact, and in the worst state it seems to affect all API calls in the session.

I already submitted some feedback through Codex CLI while this was happening, but I did not keep the feedback IDs. The version and affected models were supplied separately in the issue form fields.

What steps can reproduce the bug?

Steps to reproduce

1. Start a normal Codex CLI session and work in it long enough for auto-compaction to be triggered.
2. Observe whether remote auto-compaction succeeds or fails.
3. Also try manual compaction using /compact.
4. Repeat this over multiple attempts in the same general period of elevated failures.
5. Observe that compaction often fails instead of succeeding reliably.

Common failure messages observed:

• "We're currently experiencing high demand, which may cause temporary errors"
• "stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses)"

Optional workaround that sometimes helped earlier:

1. Switch to a cheaper, lower-reasoning model.
2. Ask for a status report, which in my workflow appears to trigger compaction.
3. Switch back to the original model.
4. Observe that this workaround was only intermittently successful and eventually stopped helping.

Escalated failure mode:

1. Continue trying normal API calls after compaction failures begin occurring.
2. Observe that, in the worst state, even ordinary API calls fail and all requests return the “high demand” error.

What is the expected behavior?

Auto-compaction and manual /compact should complete reliably, and normal API calls should continue working. Temporary backend load should not cause repeated compaction failures over multiple days or escalate into a state where all requests in the session fail over a period of hours.

Additional information

_No response_