Claude Code ships constantly, and the official release notes are written for people who already know what every line refers to. This page covers the changes worth knowing about and, for each one, what it changes about the way you work.
Covering releases through 2.1.278 · last reviewed · official changelog
Projects that use the cross-tool AGENTS.md convention are read directly, rather than needing a CLAUDE.md to exist first.
What it means: If you work in a repo shared with people on other AI coding tools, you can keep one instructions file instead of maintaining two that drift apart. If your project already has a CLAUDE.md, nothing changes for you.
How project instructions are read →For Claude API and Enterprise accounts, the decision about which actions need approval is now made server-side, and /status shows where that classifier is running.
What it means: Permission prompts get more consistent between machines, because the decision no longer depends on your local build. If you have been tuning permission rules by hand, re-check them — the behaviour you were working around may be gone.
Configuring permissions →A sandboxed command can be granted its own list of reachable domains instead of inheriting one blanket network policy.
What it means: The practical win is letting a build step reach your package registry without opening the whole network for everything else. This is the setting to reach for instead of disabling the sandbox.
When a command changes files, the result shows what changed rather than only the command output.
What it means: You can see the effect of a script without running git diff yourself, which matters most when a command does more than you expected and you want to catch it immediately.
A dedicated send key interrupts and submits, separate from Enter.
What it means: Useful once you have run /terminal-setup and rebound Shift+Enter to newline — you get an explicit send that does not depend on where the cursor sits.
Run /terminal-setup first →A command for changing output style without editing configuration files.
What it means: Worth knowing if you switch between exploratory work and writing that goes into a document, where you want different verbosity from the same session.
Sessions now warn before memory exhaustion rather than failing without explanation.
What it means: Long agentic runs on large repositories were the usual way to hit this. The warning gives you a chance to clear context before losing the session.
A setting that bounds how much reasoning effort any request may use, for every provider.
What it means: This is a cost and latency control. If you have people on a shared plan running long agentic sessions, this is the lever that stops a single task spending far more than intended.
Effort control and thinking modes →Fast mode, which speeds up output without dropping to a smaller model, now works in Remote Control.
What it means: Relevant if you drive Claude Code from your phone or another machine — the remote session no longer feels slower than the one on your desk.
Voice and remote control →A command that runs an evaluation suite against a plugin and reports how it performed.
What it means: Plugins have been testable only by using them. If you maintain one, this is the difference between "it seemed fine" and a result you can check before publishing.
Plugins and hooks →Skills and plugins associated with your account are pulled down rather than installed separately on each machine.
What it means: Setting up a second machine, or a remote box, stops being a manual re-install of everything you built. Worth checking what syncs before you assume a machine is configured.
Working with skills →We read the release notes so you don't have to, and send a short digest when something actually changes how you work. No release-note dumps.
New tutorials & tips weekly
No spam. Unsubscribe anytime.