Meanwhile GPT-5.3-Codex which just released recently is a huge change and much better. It now displays intermediate thinking summaries instead of being silent.
My experience using it from cursor has been fairly disappointing
In what way(s), if you can elaborate?
Claude 4.5 or 4.6 just one shots what I ask instead of getting stuck in random tangents.
Much better in the codex cli harness
Interesting, I can give that a try at some point.
There's one really confusing thing in Codex CLI from my perspective. How do I make it run unsandboxed but still ask me for approvals? I'm fine with it running bare on my machine but I like to approve first before it runs commands. But I only see how I can configure to have both or none. What am I missing?
--help explains this, -s for sandbox and -a for approvals
Thanks! That helped. Really strange though, that the slash commands inside the CLI do not allow for such a detailed configuration. It allows for the sandbox and approvals options but there's only "default" and "just allow everything" while the CLI flags or the config.toml allows for more nuanced options.