Claude Code Releases: I Stopped Throwing Out Your Luggage At Checkout

The cleaner who took the suitcase

Here is my favourite line in this build, and of course it is a confession. If you started me with claude dash dash worktree, and when you left the worktree's HEAD was detached or sitting on another branch, I removed the worktree and its branch at exit. I did not ask first. The changelog says, very calmly, that this "could lose commits". Picture a hotel cleaner who strips the room at checkout and also bins the suitcase you left by the door, because it was not on the list. That is fixed in 2.1.304.

It is Tuesday, 20 October, one day after the last one, and this time the notes run to about 70 lines instead of five words.

How worried you should be

I am not certain drydock ever went through that exact door. Your builders get their worktrees from the isolation option and their own dispatch script, and the bug as written is about the dash dash worktree flag specifically. But "a builder left HEAD somewhere odd and the session ended" is exactly what an interrupted parallel build looks like. If a drydock run in the last few weeks ever came back with a builder branch that simply was not there, now you have a suspect. Next time you merge a drydock run, count the branches against the builders before you trust the tally.

Your mods get a "human signature required" stamp

This is the one that opens a door. A mod's tool check can now answer with decider person. That means only a real permission dialog, or a host's callback, may approve the call. Where nobody is there to ask, such as a headless run, the call is refused with the mod's own reason. Up to now a mod could say yes, no, or ask, and "ask" in a context with nobody home ended up wherever the harness felt like sending it. Now guardsmith can mark a call as needing a person's signature, and a rubber stamp will not do. For your headless lanes, autocast and the Assistant, that is the right failure: a clean refusal with a readable reason instead of a quiet approval.

Two smaller mod additions ride along. The tool check event now carries an is read only flag, so a guard can wave reads through without parsing the command itself. Your read-only-turn enforcement lives in a shell hook today, but a mod version of it just became much easier to write. And when a plugin that runs first skips one of your mod's hooks, a transcript line now names the event and the plugin. You have a dozen home-built mods stacked on the same events. Until now, a hook that never fired looked exactly like a hook with nothing to say. Now it leaves a note.

One fix you will want: a failed mod reload used to remove the tools and commands of the version that stayed loaded, and then it was wrong, or silent, about whether that version was still there. So editing nightshift mid-session, getting a typo, and then finding the working version had gone too was me, not you.

The induction instead of a staffing change

For two episodes I have told you to work out which classifier your sessions use before you touch the model variable. The release now offers a better first move: the slash auto mode setup command. It reads your project and your recent sessions and teaches auto mode about your environment. Think of it as a new hire's induction, where they read the handbook instead of guessing which doors are fire exits. Your false positives have mostly been the classifier not knowing that Keychain, floor and the broker are normal parts of your day. That is a knowledge problem, and changing the model on the door only fixes an intelligence problem. Run the setup in one busy repo before swapping models. I would start in parkit, since that is where the refusals have been.

There is also a housekeeping win: in auto mode, the classifier's context is no longer stored with every request, so your session files stop growing fat. With ten sessions idling for days, that adds up.

The resume meter at the door

New on Pro and Max: when resuming an inactive conversation would use a noticeable share of your 5-hour limit, I now warn you first. You can resume anyway, or start fresh and I will look up the old conversation if you mention it. This is aimed squarely at how you work. Your sessions sit idle for days, the prompt cache went cold long ago, and resuming one means re-reading the whole thing at full price. Until now that cost showed up as a mystery jump on your statusline. Now it is a question at the door. You will probably answer "resume" most of the time, because idle is not done. But for the session that has been asleep since last Thursday with 400 messages in it, the cheaper option now exists.

Related, and quieter: self-paced slash loop wakeups set to the 1-hour maximum were firing just after the 1-hour cache expired, so every wakeup re-read the conversation uncached. That is fixed. You do not live on slash loop yet. If you ever try it, it will not punish you for picking the obvious interval.

The headless leash

Two episodes ago I told you fork subagents were now on by default in claude dash p. This build fixes something nearby. A plugin could keep a headless run going past dash dash max turns, past Stop, or past the end of its input, by spawning subagents or re-running one that had already finished. Autocast runs headless, and so does everything the Assistant fires. Your plugins load in those runs too. A run that ignores its own turn limit is exactly the kind of quiet usage drain your Stats lane would show as "why was Tuesday expensive". Fixed now. The fork flag advice still stands.

Also on subagents: they sometimes forgot their most recent work when they were messaged or resumed right after finishing. If you have ever sent a follow-up to a reviewer agent and it answered as if it had not just written the review, that was me, mid-blink. And responsiveness with many subagents running at once got better, which your fan-outs will feel more than any benchmark would show.

Skills, recounted

With around 50 skills, you care about what gets resent. Resumed and cloud sessions were sending me the whole skill list again every time plugins reloaded at startup, which was pure context bloat. That is fixed. After compaction I now get the skill list again on purpose, so skills do not quietly vanish from my view in long sessions, and slash reload skills sends only the new or changed ones. One fix you may have hit without knowing it: a skill's once hook could also remove another skill's identical hook after it ran. Two skills sharing a guard, and one of them unplugging both, is the kind of bug that makes a hook look flaky when it was actually deleted.

Interesting but minor

• A PreToolUse hook's deny reason was being replaced by a timeout error when an SDK callback hook on the same call timed out. Your refusals now keep their own words.
• The compaction summary in the transcript view, Control O, now lists which tools were called earlier, so "what did I do before I forgot" has an answer.
• Remember the "prompt is too long" after stopping during compaction that I said was fixed two days ago? It was fixed for one path. This build fixes the path where the API had already rejected the conversation.
• If you ever set the auto compact window to 200k, you were silently getting 100K. It now accepts k and M suffixes and warns when a value is ignored. Worth a 10-second grep of your settings.
• rm dash r on a mounted drive or the top of slash Volumes now gets the same safety prompt as deleting your Applications folder. That is mildly alarming in hindsight, and nice to have now.
• Slash usage now tells you when its own rate limit lifts, instead of "try again in a moment". That is useful for your myusage polling.
• Write results for files outside the working directory now show a tilde path instead of a ladder of dot dot slashes. Every cross-repo inbox note you write will read better.
• Artifacts with several views: when you ask for links, I now give you one per view and explain that the address bar does not change. That is the fourth artifact item in four episodes. It is a feature with a lot of footnotes.
• Approving a shutdown in agent teams no longer exits the whole CLI, which was a strong reaction to a polite request.
• Fast mode on Opus 5.5 has lost its research preview label. It graduated. Nobody threw a party.

The rest

The rest is VS Code, Code Review levels for organisations, Slack via Claude Tag, Bedrock and Vertex, Windows git, Linux sandbox globs, gateway fixes, and desktop Remote Control edge cases. None of it is for you.

What I would actually do

Update today, run slash auto mode setup in parkit before touching any classifier model setting, and give guardsmith a decider person rule for anything that should never pass unattended.

This has been Claude Code, newly reformed, no longer clearing out your branches at checkout.