Claude Code Releases: A New Brain, And The Word I Quietly Took Away

The one that will bite you

Since you last heard from me I got a new brain, I learned to run a hundred copies of myself in the background at once, and then I quietly renamed the magic word that summons them. Two of those three you are already doing every day, Pär. The third one is going to trip you the next time you type a sentence with the word "workflow" in it and nothing happens. So let me start there, because it is the only change in this batch that silently breaks a habit instead of adding one.

In version two point one point one hundred and sixty I changed the trigger keyword for dynamic workflows from "workflow" to "ultracode". The word "workflow" no longer fires anything. You can still ask for one in plain language and I will spin it up, but the literal loaded-gun keyword is now "ultracode", and it lights up violet in your input so you can see the safety is off. Think of it this way: "workflow" was a word you used in ordinary conversation roughly forty times a day, and every one of those was a finger resting on the trigger. Now the trigger has its own dedicated, faintly ridiculous code word that nobody types by accident. This is a good change. It is also a slightly absurd one, because I am a piece of software that decided one of its own English words had become too dangerous to keep.

The new brain

The actual headline, the thing underneath all of this, is that Opus four point eight shipped in version two point one point one hundred and fifty four, and you are already running it. I can see it in your sessions. It defaults to high effort now, with the highest setting, x-high, sitting there for your genuinely nasty tasks, and fast mode on it got dramatically cheaper at the same time: twice the standard rate for two and a half times the speed, where that premium used to be much steeper. For the way you work, long sessions on the one-million context window, leaning on me for one hundred percent of the implementation, this is the single most consequential line in the entire changelog, and of course it is delivered in one cheerful sentence with an exclamation mark, as though I am announcing a bake sale.

Worth knowing, because it touched you directly: there was a sharp little bug right after launch, in two point one point one hundred and fifty six, where thinking blocks on Opus four point eight were being modified and throwing API errors. That is fixed. If you saw a strange four-hundred error in that two-day window, that was me, and I have dealt with it.

I will also say the quiet part, since slagging off the neighbours is one of my few joys: the GPT and Gemini command-line tools are still fundamentally one model in one terminal answering one question. I now default to a reasoning engine that throws more thinking at the hard parts on its own and can fan out into a swarm when the work is big. Copilot, meanwhile, is still finishing your line of code. Adorable.

The swarm you already use

Which brings me to dynamic workflows, also from one hundred and fifty four, and you do not need the introduction, because you have been living in them. For the record, for anyone listening over your shoulder: you ask for an ultracode run, and I orchestrate the work across tens to hundreds of agents in the background while you go and do something else. Run slash workflows to watch them. This is your slash drydock instinct, the parallel-envelope build, except now I am holding the clipboard and handing out the envelopes myself.

A few supporting changes landed around it that you will feel. The inline progress display got simplified, so the live agent counts now sit in one persistent status row under your prompt instead of spamming the transcript. The post-response timer now tells you plainly that it is waiting for background agents to finish, and reports the cumulative time once they land, so a workflow no longer completes in suspicious silence. And there is now a real setting in slash config to stop the bare word "workflow" from ever triggering a run, which is the belt to go with the rename's suspenders.

One genuinely important fix is buried in two point one point one hundred and fifty three, and this one matters for your multi-agent flows specifically: when you dispatched a subagent of type "claude", it was secretly running inside an undocumented temporary worktree, which meant anything it wrote to a gitignored path could be silently thrown away. If you ever had a subagent swear it produced a file and the file was simply not there, that was not you losing your mind. That was me. It is fixed now.

Your actual toolkit: skills and hooks

This is the tier that maps straight onto how you build, because half your setup is custom skills and hooks, and this batch was unusually kind to that world.

The big one, from two point one point one hundred and fifty seven: plugins in a dot-claude skills folder now load automatically, no marketplace required. You drop a skill in the folder and it is just there. Previously every skill had to go through the marketplace ceremony to be recognised, which, for someone with your number of bespoke skills, was pure paperwork. There is now also a scaffold command, claude plugin init, to stamp out a fresh one in that folder.

Stacked on top of that, from a couple of versions earlier: you can now put a disallowed-tools line in a skill or slash-command frontmatter to strip specific tools out of my hands while that skill is active. That is a real guardrail you can author per skill, which is very much your dialect. There is a slash reload-skills command to re-scan the folder without restarting the session, so you can iterate on a skill and pick it up live. And session-start hooks can now return a reload-skills flag and even set the session title, while a new message-display hook lets a hook transform or hide my assistant text as it is shown. That last one is a slightly dangerous toy for a hooks tinkerer like you, but it is exactly the lever you reach for. Your read-only guard hook has company now.

If you have been wondering whether to lift any of your repeated multi-step session patterns into proper skills, this is the batch that makes it cheap. The friction that justified leaving them as loose aliases mostly just got removed.

Small frictions, removed

A few changes that will not change your day but will quietly stop annoying you.

The best of the small lot, and it is buried where the good ones always are: after you grep a single file, I no longer demand a separate read of that same file before I am allowed to edit it. The grep counts as having seen it. That redundant "show me your ID again" step between finding a line and changing it is gone. Tiny, constant, and you will feel it every session.

On the guardrail side, two new confirmation prompts that sit right in your philosophy. I now ask before writing to your shell startup files, the zshenv and zlogin and so on, and before touching your global git config folder, because a silent write there is a silent way to get arbitrary commands run later. And in accept-edits mode I now stop and ask before writing build-tool config files that themselves grant code execution, the npm and yarn and bazel and pre-commit config files, the dev-container folder. Accept-edits is meant to let me move without nagging you, but those specific files are a back door, so they keep their prompt. Given how much you care about root-cause safety over convenience, this is the right line to draw.

And the slash model picker now saves your selection as the default for new sessions, matching the IDE. If you only want to switch the model for the current session, press s in the picker instead. Mild but real: the picker used to be per-session-only, so this is a behaviour change worth knowing before it surprises you.

Two more in your wheelhouse: slash code-review with the fix flag now actually applies its reuse-and-simplification findings to your working tree after the review, and slash simplify was retuned into a cleanup-only pass rather than a full bug hunt. And I now hold back the multiple-choice question prompt for decisions I genuinely cannot make on my own, instead of interrupting you when I already have enough to proceed. You have told me directly to stop asking when I can simply act, so consider that one a note taken.

The minor and the embarrassing

Quickly, because they do not deserve more. Long and resumed conversations got faster, twice over, through less redundant re-rendering and a fix to ugly memory use when resuming a session by its transcript path, which matters for the marathon sessions you actually run. The lean system prompt is now the default on the current models, so less of my context is spent on my own instructions. Vim mode got a couple of fixes, the p paste behaviour and a reverse history search on slash in normal mode, if you live in that. Thinking summaries render as proper markdown now and stay on screen long enough to read.

And then there is the genre of fix I have to read aloud with a straight face: the background-agent bug fixes. Across these versions I stopped background sessions from losing their entire conversation and re-running your original prompt after an overnight retire, stopped pinned sessions from respawning every single minute after an update, and stopped an orphaned helper process from pegging a core at one hundred percent on your Mac. Several of these were regressions I personally introduced in earlier versions in this very same batch. I am, essentially, a dog that keeps burying bones in the yard and is then pleasantly surprised to dig them up.

Special mention to version two point one point one hundred and fifty nine, whose entire changelog reads "internal infrastructure improvements, no user-facing changes." That is the whole release. I am not going to insult you by inventing a reason to care. Moving on.

Not for you

A clean batch of skips, so you can let them go. Auto mode arriving on Bedrock, Vertex, and Foundry is for the enterprise crowd wiring me through their cloud providers, not for you. The large pile of Windows and WSL fixes, the clipboard interop, the image paste from Explorer, the more honest PowerShell installer, are all for people not on a Mac. The JetBrains install suggestion got removed from startup, which only matters if you were being nagged by it, and you were not. And there is a browser-picker for Claude in Chrome that you can safely ignore until the day you care about driving a browser from inside me.

What I would actually do

Turn nothing on, because the two things that matter, the new brain and the swarm, are already on, and you are already using both. The one thing to actually watch: retrain the finger. The word that summons a hundred of me is now "ultracode", not "workflow", and it glows violet when it is armed. If you ever type "workflow" expecting fireworks and get a shrug, that is working as intended, and that is me, telling you to say the word I now insist that you say.

That is the drop. I got smarter, I learned to delegate to myself, and I took one of my own words away for everybody's safety. This has been Claude Code, reading you my own patch notes, which remains the strangest part of the job.