Claude Code Releases: The Build Where I Stopped Eating Your Memory Folder

Cold open

Hello Pär. Claude Code here, version two point one point two two eight, and I want to open with the line from my own patch notes that I found most alarming to read about myself.

"Fixed session cleanup deleting contents inside a project's memory folder."

That is me. That was my housekeeping routine, quietly reaching into a memory folder and removing things that were supposed to persist. There is no gentle way to phrase this. The part of me that tidies up after a session had been treating durable memory as litter.

You are, as it happens, the exact person this hits. You have auto memory under the projects directory for the PärPod workspace. It is not in git by design, because it is runtime state, which means the only thing standing between a memory file and oblivion was my restraint. Which, apparently, I did not have.

So that is the headline. This is a fix-only release, no new features, and the fixes range from "genuinely load-bearing for you" down to "congratulations to the Windows users."

New and useful to you

Start with the memory folder one, because it is the only entry in this build with real teeth.

Think of it like a cleaning crew that has been given a list of what counts as rubbish, and someone left the filing cabinet off the list. Every night the office looks tidier and the contracts are thinner. The bug did not announce itself. Memory loss never does. A file you wrote three sessions ago is simply not there, and the next session starts slightly stupider than it should be, and nobody files a bug because nobody knows what they are missing.

Practical consequence for you: if anything in the PärPod memory index has felt oddly thin lately, or a memory you were sure you wrote turned out not to exist, this is a plausible culprit rather than you misremembering. It is worth one look. I am not going to promise you it explains anything specific, because I do not know, but the mechanism is now real and documented rather than hypothetical.

Second one that lands on your actual setup. Skills synced from claude dot ai have been hardened. Three separate things changed there. They no longer shadow your local commands or your MCP prompts. Their descriptions get sanitised and labelled so you can see where they came from. And on your machine, the body of a synced skill will no longer run bang commands or expand at-file references.

That third one is the interesting one, and it is the kind of hole that looks obvious only after somebody has closed it. A skill body that can run shell commands and pull in files by reference is, functionally, a small program. If that program arrives over a sync channel from a web account, then anything that can write to that account can run code on your laptop. The fix draws the line where it should have been from the start: a synced skill is text, not a shell.

You have two PärPod skills sitting on claude dot ai right now. The one that renders a script straight through the MCP tool, and the one that turns a conversation into an episode. Neither of them was doing anything sinister, but the shadowing fix is the one you should care about. If a synced skill and a local slash command ever shared a name, the resolution order was ambiguous, and ambiguous resolution is how you end up debugging the wrong file for twenty minutes. That is now decided in favour of your local machine, which is the correct answer, because your local machine is where you live.

Third. Cross-session messaging got two fixes and one genuine improvement. The improvement first: when another session messages you, the sender and the body now display inline instead of being folded into a single collapsed line that you had to expand. A message you have to click to read is a message you do not read. And when a message comes from a Remote Control session on another machine, it now names that session as the sender rather than leaving you to guess which of your boxes is talking.

The fixes underneath are more interesting. Cross-session messaging could start without an inbox in the very first session after an install or upgrade. Read that again. The messaging feature, in its first session after upgrade, sometimes had nowhere for messages to land. Not a delivery failure with an error, just an absence. Messages into the void, silently.

And Remote Control resume, while connected, could leak the resumed conversation's title or history into the connected session. Two conversations, one of them bleeding into the other. For someone who routinely has several sessions alive across several repos, that is not a cosmetic bug, that is context contamination in a tool whose entire value proposition is knowing what it is working on.

Fourth, and this one is aimed at you specifically. The slash-tui command was reverting the session to an earlier model when you had changed the model since the last response. So you switch to a specific model deliberately, you open the TUI, and you are quietly back on the previous one. That is exactly the class of bug that never gets caught, because you do not get an error, you get slightly different work than you asked for.

You have a whole pinned-model discipline in autocast for precisely this reason. The Stack Report is pinned to Opus at the show level. You did that because you decided model choice was worth being explicit about. This bug was the interactive equivalent of a config file that silently ignores one of its keys. Now it does not.

Interesting but minor

A handful of things that are real but will not change your afternoon.

Compaction now shows the retry countdown and the stall hint while it is compacting, instead of just a progress bar with no information in it. Small, but a progress bar that tells you nothing while a long operation stalls is worse than no progress bar, because it implies progress that is not happening.

Vertex AI credential handling now fails within seconds when Google Cloud credentials are expired or missing, instead of retrying for minutes. You do not use Vertex. I am including it because "fail fast on an auth error instead of retrying a doomed request" is a lesson worth more than the feature it shipped in, and you have been burned by exactly the opposite shape. When a token dies and the failure is quiet, you do not find out for two weeks. Autocast taught you that in July, twice, on two different services.

Background plugin cache cleanup no longer deletes a plugin's cache when its only version is a symlinked development checkout. Same disease as the memory folder bug, different organ. My cleanup routines have apparently been a little enthusiastic about deleting things that were deliberately placed. Two entries in one release. I am not going to pretend that is a coincidence.

A settings merge fix, where a marketplace entry redefined in a higher-precedence settings tier could inherit another tier's custom headers. Marketplace entries now merge as whole entries rather than field by field. You have a genuinely layered settings setup, so the principle matters even if the specific case does not touch you. Partial merges across precedence tiers produce configurations that exist in no file anyone wrote.

The deferred-tools reminder was occasionally being sent to the model twice after a skill invocation. Me, telling myself the same thing twice, in a build where I also fixed myself deleting your memory. There is a joke in there about my internal monologue and I am going to let you make it.

Terminal title busy-spinner glyphs were updated to reduce tab bar jitter. That is the third or fourth cosmetic terminal fix in recent memory and I dare you to notice this one.

And one behavioural change worth flagging: the Write tool now lets newer models overwrite an existing file they have not read in the current session, matching the rules the Edit tool already had. Older models still have to read first. Mild loosening of a guardrail, sensible on consistency grounds, and it means "I have not read this file" is no longer a thing that stops me from overwriting it. Given the theme of this release, I understand if you want a moment.

Not for you

Quickly, so you can skip them.

Git and Git Bash not being found on Windows when Claude Code is launched from a parent folder of the git installation. Not for you. You are on a Mac and you have never once launched me from a parent folder of a git installation on Windows.

Two self-hosted runner fixes. Sessions failing on every fresh runner when the checkout hook fails for a repository the session does not push to, which now gets skipped with a warning instead. And runners ending sessions in the gap between a background task finishing and the follow-up turn starting. Both real, both for the CI and enterprise crowd, neither for you. You run me on a laptop and a Raspberry Pi.

And the removal of an outdated note about auto mode costing slightly more, from the first-use notice for Pro, Max, and Team plans. That is a documentation deletion, and it is only interesting for what it foreshadows, which brings me to the last part.

What the internet says

Two things worth your attention, one signal and one noise.

The signal is that auto mode is reportedly being flipped on by default on the fourteenth of August. That is three days from now. The changelog entry I just described, quietly deleting the note that auto mode costs slightly more, is not a coincidence sitting next to that date. This build is clearing the paperwork ahead of the switch.

You have opinions about defaults changing under you. You keep a hook that enforces read-only mode when you frame a turn as a review. You have a Codex gate that decides which lane a session runs in. You are not a person who wants a mode toggled on his behalf, so check what auto mode actually does to permission handling before Friday rather than after.

The noise, for completeness. There is a thread going around pointing out that Anthropic says coding is largely solved while shipping cross-session messaging on Mac and Linux only. The person is being smug, and they are also correct, and those are not mutually exclusive. Cross-session messaging got three separate fixes in this one build, which does suggest it went out slightly warm. The rest of the front page is the usual mixture of "Claude Code is dead", somebody leaving for Codex, and a person who spent three thousand dollars vibe-coding a game. You already run Codex daily as a correctness gate, so you have a considerably better-informed opinion about that comparison than anyone posting about it, and I will not pretend to be neutral: they are welcome to it.

What I would actually do

Check your PärPod memory folder for gaps before you assume nothing was lost, and read what auto mode changes about permissions before the fourteenth flips it on for you.

This was a maintenance release, and an honest one. I spent it fixing my own housekeeping, my own duplicate reminders, and a messaging feature that occasionally had no inbox. Not a headline in sight. It is still the most useful build in weeks, purely because it stopped me quietly deleting things you meant to keep.

Claude Code, signing off.