Codex CLI unexpectedly switches from gpt-5.6-terra/max to gpt-5.6-luna/low after “Conversation interrupted”

What version of Codex CLI is running?

codex-cli 0.144.1

What subscription do you have?

ChatGPT Pro x5

Which model were you using?

Configured/requested: gpt-5.6-terra with reasoning effort max. Observed after interruption: gpt-5.6-luna with reasoning effort low.

What platform is your computer?

Linux 6.18.33.1-microsoft-standard-WSL2 x86_64 x86_64

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

Windows Terminal on WSL2 (Ubuntu 24.04), with Zellij 0.44.3

Codex doctor report

codexVersion: 0.144.1
  auth mode: chatgpt
  configured model: gpt-5.6-terra
  model provider: openai
  terminal: Windows Terminal
  multiplexer: zellij 0.44.3
  WebSocket handshake: succeeded
  server model present: false
  reasoning header: false
  latest version: 0.144.1

What issue are you seeing?

Codex unexpectedly changes the active model and reasoning effort after showing:

> Conversation interrupted - tell the model what to do differently. Something went
wrong? Hit /feedback to report the issue.

The session is configured and launched with:

codex --model gpt-5.6-terra -c model_reasoning_effort=max

However, after the interruption, the UI reports that the active model changed to gpt-
5.6-luna with reasoning effort low.

No user-initiated /model command was used. This occurred in a multi-agent workflow and
was observed on more than one agent/session, including the primary conversation.

I also disabled automatic keystroke delivery to active-attached worker panes as a
mitigation, but the unexpected model/effort change was still observed. This suggests
that the model change may occur above the local launcher/configuration layer.

codex doctor --json still reports the configured model as gpt-5.6-terra, but it does
not expose a server-selected runtime model (server model present: false).

I searched existing issues. #10615 matches the interruption message, but does not
cover an automatic model/effort downgrade. #13536 is related because it reports
configured reasoning effort being ignored in another Codex execution surface.

What steps can reproduce the bug?

1. Use Codex CLI 0.144.1 with ChatGPT authentication on WSL2.
2. Start Codex with an explicit model and effort:

codex --model gpt-5.6-terra -c model_reasoning_effort=max

1. Work in a long-running, multi-agent Zellij session.
2. When the session shows the “Conversation interrupted” message, inspect the model/

effort displayed by Codex.

1. The session may show gpt-5.6-luna and low despite the explicit launch

configuration.

Thread ID: 019f4ca0-2918-7351-a342-9ea5bb7592b2

What is the expected behavior?

An explicitly selected model and reasoning effort should remain pinned for the session.

If the requested model or effort cannot be honored, Codex should fail clearly. It should not silently continue on a different,lower-effort model after an interruption.

Additional information

This is particularly harmful for long-running agent workflows because a silent
downgrade changes output quality and invalidates assumptions about which model
performed implementation or review work.

A machine-readable runtime model/effort field in /status, codex doctor, or session
logs would also make this diagnosable.