Add a dedicated model configuration for conversation compaction

What version of Codex CLI is running?

codex-cli 0.135.0

What subscription do you have?

pro 20x

Which model were you using?

gpt-5.5-xhigh

What platform is your computer?

Microsoft Windows NT 10.0.18363.0 x64

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

powershell

Codex doctor report

What issue are you seeing?

Conversation compaction currently fails when using gpt5.5, but the same compaction flow succeeds when using gpt5.4 / gpt5.4-xhigh.

From what I can tell, compaction currently reuses the active conversation model. That means if the main session is using gpt5.5, the compaction request also uses gpt5.5, with no way to override it independently.

What steps can reproduce the bug?

auto compact \ manual compact

What is the expected behavior?

## Expected behavior

It should be possible to keep the main conversation model as gpt5.5, while configuring compaction to use a different model, for example:

```toml
model = "gpt5.5"
model_auto_compact_model = "gpt5.4-xhigh"

## Actual behavior

The compaction request uses the same model as the current conversation. If that model cannot successfully compact the conversation, automatic compaction fails.

## Proposed solution

Add a dedicated configuration option for compaction, such as:

model_auto_compact_model = "gpt5.4-xhigh"

When this option is set, both remote and inline compaction paths should use this configured model instead of the active conversation model.

If the option is not set, the current behavior should remain unchanged.

## Rationale

This would allow users to:

• Use the latest/default model for normal turns.
• Use a known working model for compaction.
• Avoid failed auto-compaction when a newer model has issues with summary/compaction behavior.
• Work around model-specific regressions without changing the entire session model.

Additional information

I am used three-part api，all commands is ok but /compact