Codex Desktop overrides `model_reasoning_summary = "concise"` with `detailed`

What version of the Codex App are you using (From “About Codex” dialog)?

Desktop build identifier observed in the injected app context: 26.721.41059

Bundled core recorded by the rollout:

codex-cli 0.146.0-alpha.3.1

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Windows 11 x64

Which model were you using?

gpt-5.6-terra, medium reasoning effort

What issue are you seeing?

Codex Desktop appears to ignore the explicit user-level configuration:

model_reasoning_summary = "concise"

A fresh Desktop thread records the effective thread setting as:

"reasoning_summary": "detailed"

The rollout contains this thread_settings_applied event twice, including after a rollback/retry:

{
  "type": "thread_settings_applied",
  "thread_settings": {
    "model": "gpt-5.6-terra",
    "reasoning_effort": "medium",
    "reasoning_summary": "detailed"
  }
}

The same rollout identifies its origin as:

{
  "originator": "Codex Desktop",
  "source": "vscode",
  "cli_version": "0.146.0-alpha.3.1"
}

Reasoning response items were subsequently persisted with generated summary sections, consistent with the effective detailed setting.

This is not the older behavior from #2376 / #2625 where a model rejected concise with HTTP 400. Here the request succeeds, but the effective Desktop thread configuration becomes detailed instead of the explicit configured value.

What steps can reproduce the bug?

1. Close Codex Desktop.
2. Set the following as a top-level value in the user config at ~/.codex/config.toml:

``toml
model_reasoning_summary = "concise"
``

1. Reopen Codex Desktop.
2. Start a new thread using gpt-5.6-terra with medium reasoning effort.
3. Run a multi-step coding task.
4. Inspect the generated rollout JSONL under ~/.codex/sessions/....
5. Find the thread_settings_applied event.
6. Observe "reasoning_summary":"detailed" rather than "concise".

Observed thread/session ID:

019f9da5-7cac-7590-822e-ddaebe723274

What is the expected behavior?

An explicit user-level model_reasoning_summary = "concise" should remain the effective thread setting unless the user explicitly selects a different thread-level override in the Desktop UI.

The rollout should record:

"reasoning_summary": "concise"

Alternatively, if the Desktop client has no explicit user-selected override, it should omit the turn/start.summary or thread-level summary override and allow normal config resolution to use model_reasoning_summary.

Additional information

The official config reference currently documents auto, concise, detailed, and none as valid values for model_reasoning_summary.

This may be related to the config-precedence observation in #30793. That issue notes that turn/start.summary is an override and that a frontend-provided fallback can override ~/.codex/config.toml. The value in this report is detailed rather than none, but the likely failure mode may be similar: the Desktop frontend or model preset supplies an explicit summary override even though the user did not choose one.

Two possible causes worth checking:

1. Codex Desktop always sends a model/preset-derived summary: detailed override for new turns or threads.
2. The model catalog's default_reasoning_summary incorrectly takes precedence over an explicit Config.model_reasoning_summary value in the Desktop/app-server path.

A sanitized rollout can be provided if needed.