Should AI coding agents understand and maintain their own runtime?

[YizeSun]

After using Codex CLI heavily for several weeks, I noticed something unexpected.

My local environment had accumulated a large number of runtime artifacts, including:

• session history
• rollout logs
• SQLite databases
• caches
• temporary runtime files

At first, I assumed the storage growth was coming from Xcode or iOS Simulator.

After investigating, I realized that a significant portion was actually generated by Codex itself.

That led me to a broader question about AI coding agents.

Should AI coding agents maintain their own runtime?

Today, Codex can:

• write code
• edit repositories
• run tests
• execute shell commands
• operate fairly autonomously

But it doesn't help users understand the runtime artifacts it creates. To me, this isn't primarily a disk cleanup problem. It's a runtime understanding and management problem.

For example:

• What is this file?
• Why was it created?
• Is it still useful?
• Is it safe to remove?
• Which artifacts belong to Codex itself?

To explore this idea, I built a small prototype called Codex Doctor. The prototype isn’t really the point. It’s simply a way to explore the idea.

Maybe I’m overthinking this, but I couldn't help wondering whether AI coding agents should eventually understand and maintain the runtime they create.

I'd love to hear whether others think this is a real gap, or if I'm simply looking at it the wrong way.