Interactive Ultra selection produces inconsistent Default/Plan reasoning state

[htregidgo]

What version of Codex CLI is running?

0.146.0

What subscription do you have?

ChatGPT Pro

Which model were you using?

gpt-5.6-sol

What platform is your computer?

Darwin 25.5.0 arm64 arm

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

PyCharm integrated terminal on macOS; no terminal multiplexer

Codex doctor report

What issue are you seeing?

Selecting the ultra reasoning level interactively through /model behaves differently from selecting other reasoning levels and from configuring ultra through CLI overrides.

A separate Default/Plan configuration can be created successfully with:

codex
-c model_reasoning_effort='"medium"'
-c plan_mode_reasoning_effort='"ultra"'

This produces the expected state:

• Default: Medium
• Plan: Ultra

However, selecting Ultra interactively through /model, from either Default or Plan mode, does not show the normal scope-selection dialogue and changes both Default and Plan to Ultra.

More importantly, this resulting Ultra/Ultra state behaves differently from an Ultra/Ultra state created with separate CLI configuration. For example:

• With Ultra configured explicitly as the Plan override, selecting XHigh from Default produces Default=XHigh, Plan=Ultra.
• After selecting Ultra interactively and thereby reaching Ultra/Ultra, selecting XHigh from Default produces Default=XHigh, Plan=XHigh.

The same distinction is visible from Plan mode: choosing “Plan only” changes only Plan when the Ultra override came from CLI configuration, but changes both levels when Ultra/Ultra was reached through interactive Ultra selection.

Max and XHigh behave consistently with the CLI-configured Default/Plan split when tested from fresh sessions. Ultra behaves differently, producing a state that subsequently behaves differently despite displaying the same reasoning levels.

What steps can reproduce the bug?

Uploaded thread: 019fad6d-ff20-72f2-a3fb-1c66652f5fc4

What is the expected behavior?

Reasoning-level selection should be internally consistent.

Two sessions displaying the same Default and Plan reasoning levels should exhibit the same subsequent behaviour regardless of how those levels were reached (CLI configuration versus interactive /model selection).

If different semantics are intended for interactive Ultra selection, they should be reflected consistently in the UI and documented. Otherwise, interactive Ultra selection should participate in the same Default/Plan state model as the other reasoning levels.

Additional information

The CLI configuration itself works correctly; the issue concerns the interactive /model transition to Ultra.

The behaviour was tested across fresh Codex invocations. Fresh-state tests showed that Max and XHigh preserve explicit Plan overrides and expose the scope dialogue from Plan mode.

The automatic feedback thread was uploaded and is referenced in the reproduction field.

[htregidgo]

During investigation I collected a behavioural matrix to better understand the issue.

The important observation is not any individual transition, but that two sessions displaying identical reasoning levels (Default=Ultra, Plan=Ultra) can subsequently behave differently depending on how that state was reached.

Behaviour summary

Initial state	Action	Result
CLI: Default=Medium, Plan=Ultra	Select Ultra interactively	Both become Ultra (no scope dialogue shown)
CLI: Default=Ultra, Plan=Ultra	From Default, select XHigh	Default=XHigh, Plan=Ultra
Interactive: Default=Ultra, Plan=Ultra	From Default, select XHigh	Default=XHigh, Plan=XHigh
CLI: Default=Ultra, Plan=Ultra	From Plan, select XHigh, choose Plan only	Default=Ultra, Plan=XHigh
Interactive: Default=Ultra, Plan=Ultra	From Plan, select XHigh	Both become XHigh (no separate Plan-only behaviour observed)

The first and third rows are the key observations.

Although both sessions display the same reasoning levels before selecting XHigh (Default=Ultra, Plan=Ultra), subsequent /model changes behave differently depending on whether that state was created:

• by CLI configuration, or
• by selecting Ultra interactively.

This suggests that the two visible states are not internally equivalent, or that interactive Ultra selection follows different state-transition logic from the CLI configuration.

I intentionally avoided drawing conclusions about the intended behaviour, since I could not establish what the expected semantics should be. The core observation is simply that identical displayed reasoning states exhibit different subsequent behaviour depending on how they were reached.