Claude Code stops wiping .env files

Claude Code 2.1.327, released by Anthropic on November 17, fixes four bugs that could delete or overwrite your files. In one of them, stopping Remote Control in worktree mode wiped a running session’s git-ignored files, such as .env. If you use Remote Control worktrees, self-hosted runners or project-scoped MCP servers, this update closes those gaps. Auto mode users should also read the new notice: an upcoming release removes the fallback to billed classifier requests.
The release notes on GitHub list well over a hundred changes. Most are small fixes to resumed sessions, subagents and the terminal display. The ones below change what you do, or what you can trust Claude Code to leave alone.
Files that could disappear
The Remote Control bug hit a running session’s git-ignored files when the server stopped in worktree mode. Git keeps no copy of a file it ignores, so a .env lost that way had no undo.
Self-hosted runners started with --use-anthropic-git-proxy had a similar problem. At startup they deleted a non-empty directory named by GIT_CONFIG_GLOBAL. The runner now keeps that directory, ignores the variable and shows a warning.
With project scope, claude mcp add and claude mcp remove deleted other top-level keys from .mcp.json, such as $schema or another tool’s servers. A new guard covers a related case. If .mcp.json lists its servers at the top level with no mcpServers key, claude mcp add --scope project now stops instead of overwriting the file.
/export overwrote an existing file without asking. Now the dialog asks first, and /export <name> leaves the existing file alone and tells you to run /export instead. Several silent failures now report themselves too. /rewind says so when some files could not be restored, and a settings change from /config that fails to save gets a line under the prompt. ExitWorktree asks before removing a worktree that may hold unsaved work. In don’t-ask mode, or with nobody to answer, it refuses unless an allow rule, auto mode or bypass permissions mode allows it.
Secrets and the first Bash command
The new claude mcp add-json <name> - reads a server’s JSON from standard input, so tokens no longer have to appear on the command line. It doesn’t work inside a Claude Code session, so run it from your own shell. On newer models and in PowerShell, Claude is now told not to commit files that likely contain secrets. Claude in Chrome now refuses to upload some credential files; you attach those to the page yourself.
One speed change works in the other direction. To get a session’s first Bash command running sooner, Claude Code may now run your shell’s start-up files before you approve that command, even if you then refuse it. To turn this off, set CLAUDE_CODE_DISABLE_SHELL_HEAD_START=1.
Fixes you’ll notice day to day
Resumed sessions get the biggest share of the fixes. Several cases are closed where messages, an answer or tasks were lost after a crash, a killed process or /clear. A transcript read that ends early is now retried once, so messages are lost less often. Other changes:
- After a longer outage such as a VPN drop, most remote MCP servers on other machines reached by hostname over https are now retried in the background for about a day.
claude -pno longer hangs when fetching a model’s settings stalls, for example at a proxy. After 10 seconds it carries on without them.- Claude Code starts again when
https_proxyorHTTPS_PROXYis a bare host:port and neither its input nor its output is a terminal (the desktop app, an SDK, cron). - In-process agent teammates on a different model from the lead now compact at their own context limit and run on their own model, billed at that model’s rate.
- On unrecognized models, a reply cut at 32,000 output tokens now continues with a limit of 128,000.
- When a mod’s
$.model.completeor$.model.classifycall hits a usage limit, it now gets the API’s error at once instead of hanging the turn. Mods should handle that error. - In VS Code, “Claude Code: Add to Chat” in the Explorer’s right-click menu @-mentions the selected files and folders in the chat.
Where auto mode is heading
Auto mode’s notice about falling back to billed classifier requests now says an upcoming release removes that fallback. After that, each action the fallback would have checked will be asked about or denied. Some setups already work this way. Since version 2.1.319, sessions with telemetry or feature flags off, gateways included, have had no built-in classifier fallback, and calls the server couldn’t check are asked about or denied.
This release tightens auto mode in other ways too. It ignores more allow rules that run any package, such as Bash(pnpm dlx *), plus path or version forms of rules it already ignored (/usr/bin/npm run:*). The safety check now decides those calls. Some symbolic links are now refused in path checks, and some paths no longer count as inside a working directory, so they ask for permission.
With permissions.blockReadsOutsideWorkingDirectories on, more Bash commands now get a prompt in bypass and auto modes, and some are denied in auto mode. A Bash command that names a path under a Read or Edit deny rule is now refused more often, where it used to get a prompt.
Up next
Perplexity puts Claude, GPT and Gemini on one team, Video
Claude, GPT and Gemini now work as one team in Perplexity: a lead agent splits a task among up to 16 agents that check each other’s work. It’s for Max subscribers on the web. Would you trust one model on hard work, or a team that checks itself?
Related stories
- Claude Code 2.1.326 runs Code Mode with agent teams on
- Claude Code 2.1.319 drops auto mode fallback behind gateways
- Long-running jobs and mid-task questions reach Claude Code
- OAuth login to MCP servers is fixed
- Microsoft rolls out Autopilot agent in Frontier preview
- Replit says its agent swarm built Radar in 2 days for $3,000
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

