Model picker exposes hidden codex-auto-review with GPT-5.6 Terra description

Open 💬 2 comments Opened Jul 27, 2026 by theogottwald

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

Description

The Codex CLI model picker exposes codex-auto-review as the quick model and displays the description:

Balanced agentic coding model for everyday work.

That description belongs to gpt-5.6-terra in the official model catalog. The catalog describes codex-auto-review as "Automatic approval review model for Codex." and marks it with visibility: "hide".

The same picker simultaneously shows:

All models (current) ... (current: gpt-5.6-terra)

This makes it appear that codex-auto-review is a normal, selectable everyday coding model and creates uncertainty about which model and pricing apply.

Screenshot evidence

A user-provided screenshot shows this exact model picker state:

Select Model
Pick a quick auto mode or browse all models.

1. codex-auto-review    Balanced agentic coding model for everyday work.
2. All models (current) Choose a specific model and reasoning level (current: gpt-5.6-terra)

Screenshot dimensions: 945 × 186 pixels. No private information is visible.

Steps to reproduce

Start an interactive Codex CLI session.
Open the model picker with /model.
Observe the quick model entry.

Actual behavior

codex-auto-review is visible as a user-selectable quick model and is paired with the GPT-5.6 Terra description.

Expected behavior

The quick model entry should display gpt-5.6-terra if Terra is intended.

Alternatively, if codex-auto-review is shown intentionally, it should use its catalog description and clearly indicate that it is the internal automatic approval reviewer. A model whose catalog visibility is hide should not appear as a normal user-facing coding-model option.

Environment

Surface: Codex CLI
Active model shown by the picker: gpt-5.6-terra
CLI version: not captured in the screenshot
Date observed: 2026-07-27

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 1 month ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#35493
#35305

Powered by Codex Action

ItsukiYoshida · 24 days ago

I can reproduce the same picker state on Codex CLI 0.146.0 while the current session model is gpt-5.6-sol xhigh, so the symptom is not specific to Terra being the active model:

Select Model
Pick a quick auto mode or browse all models.

1. codex-auto-review     Balanced agentic coding model for everyday work.
2. All models (current)  Choose a specific model and reasoning level (current: gpt-5.6-sol)

There also appears to be a client-side classification issue in the TUI, separate from the remote catalog metadata that made this model visible.

The picker first filters on show_in_picker, then classifies every visible slug beginning with codex-auto- as a quick auto mode:

https://github.com/openai/codex/blob/7750465934d97dd3cbcb3b1655d2f622744010d3/codex-rs/tui/src/chatwidget/model_popups.rs#L74-L89

fn is_auto_model(model: &str) -> bool {
    model.starts_with("codex-auto-")
}

However, the immediately following ordering function only names the three user-facing quick modes: codex-auto-fast, codex-auto-balanced, and codex-auto-thorough:

https://github.com/openai/codex/blob/7750465934d97dd3cbcb3b1655d2f622744010d3/codex-rs/tui/src/chatwidget/model_popups.rs#L176-L185

As a result, any internal or future model whose slug happens to match codex-auto-* is promoted into the quick-mode picker whenever a catalog/cohort regression sets show_in_picker = true. That includes codex-auto-review, which is the automatic approval reviewer rather than a normal coding mode.

A client-side fix could use explicit model metadata for quick modes, or at minimum allowlist the three supported quick-mode slugs instead of relying on the prefix. A regression test could expose codex-auto-review in a model fixture and verify that it is not classified as a quick auto mode, while fast/balanced/thorough still are.

Related catalog/visibility reports: #35493 and this issue. This comment adds the TUI-side mechanism and a defense-in-depth fix, rather than attributing the behavior only to the Terra description.