Codex Desktop: custom-provider model shown as "Custom High" and session persona claims GPT-5

Codex Desktop: custom-provider model is shown as "Custom High" and the session persona claims GPT-5

What version of Codex?

Desktop: OpenAI.Codex_26.727.6591.0 (Windows MSIX, build 26.727) / CLI 0.146.0-alpha.9.2 (Windows 11, 10.0.26200)

What issue are you seeing?

Two related symptoms when using a custom provider plus a locally configured model_catalog_json:

1. Wrong label in the Desktop model picker. The composer model picker (bottom-right of the input box) shows a fabricated provider + reasoning_effort label - e.g. Custom High, or Deepseek Max when the provider is named deepseek - instead of the model's display_name from the configured catalog (DeepSeek V4 Flash).
2. Wrong model identity in the session. The session's base_instructions / instructions_template contain You are Codex, an agent based on GPT-5. even though the configured model is deepseek-v4-flash and the catalog entry provides its own model messages. Asking "which model are you using?" returns GPT-5, while the CLI TUI correctly displays the configured model slug.

The CLI and the Desktop app run the same configured model (deepseek-v4-flash) and the catalog is resolved correctly (codex debug models shows the entry with display_name: "DeepSeek V4 Flash"), so this appears to be a Desktop rendering / metadata-selection bug rather than a configuration problem.

Steps to reproduce

1. Configure ~/.codex/config.toml:

```toml
model = "deepseek-v4-flash"
model_provider = "deepseek"
model_reasoning_effort = "max"
model_catalog_json = "C:/Users/user/.codex/models.json"

[model_providers.deepseek]
name = "deepseek"
base_url = "https://api.deepseek.com/"
wire_api = "responses"
```

1. In models.json, include an entry for the model with proper metadata:

``json
{
"slug": "deepseek-v4-flash",
"display_name": "DeepSeek V4 Flash",
"model_messages": {
"instructions_template": "You are Codex, an agent based on DeepSeek V4 Flash. ...",
...
}
}
``

1. Run codex debug models - the entry is returned with the correct display_name and instructions.
2. Open the Codex Desktop app. The model picker shows Custom High / Deepseek Max (provider + effort) instead of DeepSeek V4 Flash, even though the catalog entry exists.
3. Start a task with this model. The saved session (session_meta) contains base_instructions: "You are Codex, an agent based on GPT-5. ...". Asking the agent "which model are you using?" yields a GPT-5 answer, while the CLI shows deepseek-v4-flash.

Expected behavior

• The Desktop model picker should resolve the configured model to its catalog display_name (or at minimum its slug, like the CLI does), instead of concatenating provider + effort.
• Sessions for custom-provider models should use the model's catalog instructions (instructions_template / base_instructions), or otherwise must not identify the agent as GPT-5 when the configured model is a third-party model.

Additional notes

• The catalog metadata alone is not enough: even with a valid display_name present in model_catalog_json, the picker still renders the fallback provider + effort label.
• This matches the symptom already described in #19694 ("the Desktop App may show Custom ...") and is related to #34487; the new facet here is the session persona claiming the wrong model (GPT-5) for a custom-provider model.

Related issues

• #19694 - Codex Desktop model picker filters out models returned from model_catalog_json
• #34487 - Desktop: model_catalog_json loaded but custom models not shown in picker (regression, related to #32349)