Speeding Up Codex: Composer Predictions and Mid-Run Steering

Ethan Park

Thanks to Jellypod for helping make this daily show a reality. Last time we spent our time on the backend, how fast the model generates tokens. Today is about the other half of the loop, which is how fast you can tell the agent what to do next. On October 9, OpenAI put Composer predictions into beta in the Codex desktop app.

Maya

Here's the mechanism. After Codex finishes responding in a thread, a suggested next message can appear in the message box. It's a whole message, not a code snippet. You press Tab to accept it. And this is the part to remember: accepting only puts the text in the box. It does not send anything. You read it, edit it, and then you hit Enter yourself.

Ethan Park

That distinction matters more than it sounds. Think about what you actually type in a long agent session. Run the unit tests. Show me the diff. Check the linter output. Go ahead with the refactor on the second file. Those turns are short and predictable, but each one still costs you a context switch and a few seconds. The prediction is trying to take that typing off your hands, based on the conversation in the current thread.

Maya

And it's different from inline code completion in an editor. It doesn't update word by word as you type. It proposes a complete next turn, once, after the response lands. If you start typing your own message instead, you're just writing your own message. You don't have to wait for the suggestion, and you can ignore it entirely.

Ethan Park

Let's talk about the safety angle, because this is an agent that can run shell commands and apply patches. If Tab sent the message, a wrong guess could trigger real execution. Since Tab only fills the composer, the human review step is still there. Treat the suggestion like a draft from a colleague. Read it before Enter, especially if it says something like proceed or apply.

Maya

Now, who can use it. This is where people will get tripped up. During the beta it's for personal ChatGPT Pro users aged 18 and older, in all supported regions, in the latest Codex desktop app. It works in local threads and SSH threads, and only with GPT 6 Astra or GPT 6.1 Sol. The help center says it isn't available in Chat on web or desktop, and not on other plans on the web. It says that in Work, predictions may appear in local Work threads.

Ethan Park

So if you're on a Business or Enterprise workspace, or you've picked a different model, you may simply not see it. Before you assume something is broken, check three things from the help center: the app is up to date, Show predictions is on, and your account, model, and thread type are all supported.

Maya

Setup is short. It's on by default for eligible accounts. Open Settings in the desktop app, go to General, then the Composer section, and you'll find Show predictions. Flip it off if you find the suggestions distracting, and flip it back on in the same place.

Ethan Park

Next, what it can see. Predictions use the conversation in your current thread. They don't go looking up memories or pulling from connected apps on their own. But if memory context or connected app output was already pulled into the thread earlier, that can inform the suggestion. You can use predictions with memory off. If you want memory on, that's Settings, Personalization, Enable Codex memories.

Maya

On cost, there's good news for the beta. Generating predictions doesn't count toward your Codex usage limits and doesn't consume credits. But the message you send does, including an accepted prediction. So Tab is free, and Enter is normal billing.

Ethan Park

A few caveats. The composer may skip a prediction on a given turn, and the help center says that alone doesn't mean something is wrong. Predictions can also take several seconds to show up, and longer in long conversations. And they suggest a complete message, so they won't finish a sentence you've already started. Plan around that. Don't sit and wait for a suggestion, just type.

Maya

I'd add one interpretation, and it's mine, not OpenAI's. The suggestion will be most useful when your next step is obvious from the thread, like after a test run or a diff. When the next step is a real design decision, you'll probably want to write that one yourself anyway.

Maya

The related change shipped a day earlier, on October 8. OpenAI rolled out faster steering in Codex in the ChatGPT desktop app. When you send a follow up while a task is running, Codex can respond to your changes sooner. You'd use that to correct an approach, add missing information, or change direction mid run.

Ethan Park

And there's a setting that decides how it behaves. Go to Settings, General, Follow up behavior. You choose whether a follow up message steers the current run or waits for the next one. If you want to redirect an agent that's heading the wrong way, pick steer. If you'd rather queue thoughts without interrupting work in progress, choose the waiting option. Pair that with predictions and the loop tightens on both ends: the suggestion fills in your next message faster, and steering lets that message land sooner.

Maya

A few smaller things worth knowing from the open source CLI. Version 0.162.0, released October 8, added copy support for transcript blocks with the slash copy command. Control Insert now copies a selection, and there's a config key, tui dot mouse underscore scroll underscore speed, to tune mouse wheel scrolling.

Ethan Park

The same release makes URLs clickable in approval headers, questions, MCP prompts, warnings, and banners, including when a link wraps across lines. That sounds minor until you're approving something and need to open the link in it. And from 0.161.0 on October 7, you can sign in to an MCP server from an active terminal session with slash mcp login, then the server name.

Maya

So here's what to try today. If you're on a personal Pro account, update the desktop app, open a local thread on GPT 6.1 Sol or GPT 6 Astra, and let a turn finish. When a suggestion appears, press Tab, read it, change one word, and send it. Then check your Follow up behavior setting, so you know how mid run messages will behave.

Ethan Park

One last habit worth building: always read before Enter. The feature is designed so that you can. That's the whole point of Tab not sending.