Claude Code stops resetting your config file

Claude Code 2.1.329 is out, and like the last few releases it is mostly fixes. The one most people will care about: Claude Code could sometimes wipe its own settings file. Also on the list are a code review that could quietly check the wrong code and a change that means some scheduled tasks have to be set up again.
What was fixed
Your ~/.claude.json could sometimes be reset to defaults. If a network stall or a file lock made the file unreadable while Claude Code was saving it, the file was cleared. That no longer happens.
/code-review with a pull request number sometimes reviewed the branch you had checked out instead of the pull request, and didn’t tell you. It now reads the pull request itself with gh pr diff, or says when it can’t. With --fix, or when you give it a pull request URL, it now stops without reviewing if gh pr diff is refused or fails. Run with no target, on most models it now reviews the whole branch against the default branch, and it no longer loses the diff on branches with several commits.
On macOS, Claude Code kept signing people out when it could read the Keychain but not save to it. Now only that session stops when its login expires, and it says why. Other sessions stay signed in.
What you may need to redo
Scheduled tasks saved by earlier versions mostly won’t carry over: in new signed-in sessions, they need scheduling again. Resuming the session a task was created in with --resume still runs it. And “Exit and stop tasks” now ends the session’s scheduled tasks for good, so they don’t come back on resume, unless you exit while Claude is still responding. This comes right after version 2.1.328 fixed a bug where saving one scheduled task could wipe out a project’s other saved tasks.
For gateways, hooks and macOS
Version 2.1.326 had added an automatic retry when a request hit the limit named in a max_tokens error. That retry is gone. If you use Claude Code through a gateway whose model accepts fewer output tokens, set CLAUDE_CODE_MAX_OUTPUT_TOKENS to that limit yourself.
In Code Mode, the experimental setting where Claude writes JavaScript to call its tools, a PreToolUse hook that errors no longer refuses the tool call, which matches how hooks behave outside Code Mode. If a hook must block, set "onFailure": "block" on it.
Computer use on macOS now follows newer safety rules, so Claude will refuse some actions and some apps get less access than before.
Up next
OpenAI dots start reaching ChatGPT Plus users, Video
The always-on dots agents are leaving Pro: a small number of ChatGPT Plus users now test them, says OpenAI’s Thibault Sottiaux, and all Business Standard users get them. If dots reached your plan, what’s the first job you’d give one?
Related stories
- Claude Code 2.1.326 runs Code Mode with agent teams on
- Claude Code stopped erasing your scheduled tasks
- Claude Code’s Code Mode runs tool calls as JavaScript
- Claude Code Projects: proactive coordinator, /btw, /fork
- Claude Code 2.1.323 lets Claude call tools via JavaScript
- Cloud sessions in Claude Code now set up your repo
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

