Codex Desktop model picker shows two indistinguishable GPT-5.6 Luna entries

What version of the Codex App are you using?

Codex Desktop with embedded CLI 0.146.0-alpha.3.1 on Windows. The local model catalog was refreshed immediately before capture.

What platform are you using?

Windows 11

What issue are you seeing?

The model selector lists two indistinguishable “5.6 Luna” options. Both are selectable, have the same visible label, and no description or badge distinguishes them.

The local ~/.codex/models_cache.json catalog contains two different entries with visibility: "list":

{ "slug": "gpt-5.6-luna",    "display_name": "GPT-5.6-Luna" }
{ "slug": "codex-auto-review", "display_name": "GPT-5.6-Luna" }

They also share the same public description and supported reasoning levels.

These are not simply duplicate catalog rows: comparing their metadata shows that the codex-auto-review entry carries an additional automatic-review / Guardian policy configuration, while the normal gpt-5.6-luna entry does not. This suggests codex-auto-review is an internal reviewer lane that has been exposed as a normal user-selectable model with a misleading label.

Steps to reproduce

1. Open Codex Desktop's model selector.
2. Observe two entries named “5.6 Luna”.
3. Select either entry; the UI provides no way to know which backend model slug or policy profile was selected.
4. Inspect ~/.codex/models_cache.json and compare the entries above.

Expected behavior

The model picker should either:

1. hide codex-auto-review from normal user model selection, or
2. render it as a clearly distinct option such as “5.6 Luna (Auto Review)” with an explanation of its approval-review role.

Actual behavior

Two visually identical Luna choices are presented as ordinary models, despite different backend slugs and policy behavior. This makes model selection ambiguous and can accidentally place a task on the automatic-review profile.

Additional information

Related reports establish that codex-auto-review is used as a dedicated automatic-approval reviewer model, but I could not find an existing public report specifically for the duplicate, indistinguishable entries:

• #34976
• #24879
• #33146