Regression: long resumed conversations fail remote compact with context_length_exceeded on 0.132.0+

What version of Codex CLI is running?

Observed with:

• codex-cli 0.132.0
• codex-cli 0.133.0

Downgrading to codex-cli 0.131.0 works around the problem.

Which model were you using?

GPT-5.5 in Codex.

What platform is your computer?

Linux / WSL environment.

What issue are you seeing?

A very long conversation can no longer continue after upgrading to 0.132.0 or 0.133.0. When the conversation is resumed, Codex appears to put the full conversation history back into the model context instead of successfully compacting it. The remote compact task then fails with context_length_exceeded, and the session cannot continue.

Error shown by Codex:

Error running remote compact task: {
  "error": {
    "message": "Your input exceeds the context window of this model. Please adjust your input and try again.",
    "type": "invalid_request_error",
    "param": "input",
    "code": "context_length_exceeded"
  }
}

This looks like a regression in the resume / remote compaction path. The same long conversation can be continued again after downgrading to 0.131.0.

What steps can reproduce the bug?

1. Start with a very long Codex conversation.
2. Use Codex CLI 0.132.0 or 0.133.0.
3. Resume the long conversation.
4. Continue the conversation until Codex tries to compact remotely.
5. Observe that remote compact fails with context_length_exceeded, and the conversation cannot continue.
6. Downgrade to 0.131.0 and resume the same workflow. The conversation can continue again.

What is the expected behavior?

Resume should not re-add the entire long conversation to the active context in a way that makes compaction impossible. Codex should compact or trim the resumed conversation before sending it to the model, so long conversations remain recoverable.

Additional information

This may be related to #23694, which also reports a remote compact regression after upgrading to 0.132.0, but that issue reports array_above_max_length. This report is specifically for the context_length_exceeded failure mode on resumed long conversations.