Claude Code Gets Safer Defaults and Better Cache Shields

Lachlan Reed

Alright, if you've been running Claude Code in tmux lately and- and felt like your terminal screen was throwing a absolute rave, you're not alone. This is the Claude Code Changelog, brought to you by Jellypod AI. I'm Lachlan.

James Turner

And I'm James. And- and seriously, that- that tmux flickering was driving me up the wall, Lachlan. But let's- let's actually dive into the big safety shift in v2.1.200 first. They've flipped the default permission mode to 'Manual'.

Lachlan Reed

Right, right. So no more Claude just... wandering off and running commands while you're, I don't know, grabbing a cuppa. It's- it's like putting a lock on the back gate, you know? If it hits one of those AskUserQuestion dialogs, it just- it sits there. It pauses indefinitely.

James Turner

Yeah, it- it won't auto-continue at all now unless you go out of your way to configure an explicit idle timeout via /config. It's a massive deal because if you leave your terminal open, you don't want an autonomous agent just deciding, 'Hey, let's push this' without a human eye on it.

Lachlan Reed

Oh, absolutely. I mean, back in my early days, I nearly tanked a- a client's site with a dodgy midnight update. Having an AI doing that on autopilot? No thanks, that gives me the absolute collywobbles.

James Turner

Exactly. But then, right on its heels, we got v2.1.201, and this is the- the real under-the-hood performance beauty. It's all about protecting prompt caching for Sonnet 5.

Lachlan Reed

Right, because if you're working in a massive codebase, prompt caching is what keeps your API bill from- from absolutely skyrocketing, yeah? If you bust that cache, you're paying for thousands of tokens over and over.

James Turner

Right, and what was happening under the hood was that Claude Code was injecting these mid-conversation system role messages—basically 'harness reminders' to keep the agent aligned. But because those system messages were being popped in dynamically mid-chat, they completely invalidated the prompt cache prefix!

Lachlan Reed

Oh, mate, that's- that's like... imagine you're reading a massive manual, and every five minutes, someone taps your shoulder and rewrites the- the preface of the book. You'd have to start reading the whole bloody thing from page one again!

James Turner

Exactly! Sonnet has to re-evaluate the entire context from scratch. By removing those mid-conversation reminders in 2.1.201, the cache prefix stays completely untouched. Your latency drops, and your API usage stays incredibly efficient.

Lachlan Reed

That is brilliant. Truly. And- and what about those- those quality of life tweaks? The tmux fix?

James Turner

Right, so they added synchronized terminal output. Instead of rendering the screen in messy, staggered chunks—which is what caused that rapid flickering in tmux—it now coordinates the render so it draws the terminal in one single, clean sweep.

Lachlan Reed

Ah, nice. No more rave in the terminal. And- and there's a new installer check too, right? For- for memory?

James Turner

Yeah, it's a smart diagnostic warning for out-of-memory, or OOM, conditions. If you're trying to install this on a resource-constrained environment—like a tiny VPS or a docker container—the installer will flag it upfront instead of letting Node crash silently with a cryptic stack trace.

Lachlan Reed

Incredible. Safer defaults, cheaper runs, and a screen that doesn't melt your eyes. Good stuff. Talk soon, mate.

James Turner

Yep, see ya.