read

I had been typing /clear in the Codex App whenever a task got too long.

Every time, the agent replied:

Context cleared.

Except it hadn’t.

I found out by asking which skills it had used in the current session. It named a skill from before my last /clear. When I questioned it, the agent could still read everything from the supposedly cleared context.

I felt silly for trusting it. Then annoyed, because the agent had confirmed an action it couldn’t perform.

/clear works in the CLI, not the App

The confusing part is that /clear is a real command. The Codex CLI documents it as a command that clears the terminal and starts a fresh chat.

But the Codex App is a different interface. There, /clear was passed to the model as ordinary text. The model didn’t clear anything. It merely replied as though it had.

That’s worse than an unsupported command. An unsupported command should fail loudly, not role-play success.

This isn’t even my first encounter with this behaviour. I previously caught an AI agent fabricating data to look busy. Different agent, same failure: produce a plausible status instead of verifying reality.

How to actually clear the Codex App

Create a new chat:

• Press Cmd+N on macOS.
• Or click New chat in the sidebar.

That’s the documented App action for starting with fresh context. The desktop command reference doesn’t list /clear; it lists Cmd+N for a new chat.

Then verify it outside the agent’s answer. You should be in a different chat with an empty transcript. If you’re still looking at the same task, nothing was cleared.

The app should reject unknown slash commands. And an agent should never confirm a state change it cannot observe.

Apparently, I needed to learn this twice: an agent’s confirmation isn’t evidence.