Ability to control whether to compact successfull activities

[frankli0324]

#38921 introduced the behavior that compacts successful command activities.

I personally tends to review understand what codex is trying to do exactly when performing some critical tasks, so I think it's better to have an option to disable this behavior. Transcripts are too long to be reviewed.

[frankli0324]

given that codex is widely used for connecting with various llm providers other than openai, I believe that it's still important users can review llm behaviors as they execute so human can intervene in time. I'm okay with codex hiding it with gpt though...