Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Claude Code 2.1.326 tightens the sandbox

Written November 17, 2026

Claude Code 2.1.326 stops sandboxed commands writing to sensitive home folders, refuses writes to sign-in files, and runs Code Mode with teams.

Claude Code 2.1.326, released on November 16, 2026, is a long fix list with a clear centre of gravity: the sandbox and the permission checks around it get stricter. The v2.1.326 release notes describe sandboxed commands that write to fewer places than before, file tools that will not touch Claude Code's own sign-in files, and permission checks in Code Mode that stop more commands for a decision. If you run Claude Code with the sandbox on, some writes that landed last week will now stop or ask. This page covers the changed defaults, the one setting that re-allows a folder, and the Code Mode change that removes an old conflict with agent teams.

Sandboxed commands write to fewer home folders

Sandboxed commands on macOS, Linux and WSL no longer write to some sensitive home folders by default. The notes put it plainly, and they name the way back for folders you do want writable: some of them can be allowed again with sandbox.filesystem.allowWrite, set in user or managed settings. That split is the practical shape of the change. The default tightened for everyone, and the exception is a deliberate settings entry, not a prompt you can wave through in the moment. If a sandboxed build or script of yours writes into a home folder as part of its normal work, expect it to fail until you add that folder through the setting. The release also improves Bash sandbox protections on the same three platforms more generally, and changes edits in a session's mods folder so they ask more often. On the mods folder, the notes add a specific hint: if every edit asks, remove any link there that points to an unreadable or network path.

Sign-in files are off limits

Claude's file tools now refuse writes to Claude Code's sign-in files, such as .credentials.json. This is a refusal, not a prompt. The tools will not write there even if the edit looks routine, which closes a route by which a session could alter the credentials it itself runs on. The Bash side moves the same way: the Bash permission check now asks before more commands that reach Claude Code's sign-in files. Between the two changes, the credential files sit behind both a hard stop for the file tools and a wider ask-first net for shell commands. For most users nothing about daily work changes here, because few legitimate tasks write to those files. For anyone who scripted around them, that scripting stops working in this release.

Code Mode asks more, and runs with teams

Permission checks in Code Mode are improved, in the notes' wording, so more sandboxed commands now ask or are refused. Combined with the tighter Bash checks, where some commands now ask first or are refused where no one can answer, the direction is consistent: unattended and sandboxed work gets less silent latitude in 2.1.326. The other Code Mode change removes a restriction instead of adding one. You can now turn on Code Mode while agent teams are enabled, and if Code Mode was already on, sessions with teams enabled now start in Code Mode. The two features no longer exclude each other, which matters if you held off on one of them to keep the other. How Code Mode itself works is covered in the page linked below.

Three smaller changes worth knowing

A startup warning now flags entries in --allowedTools, --disallowedTools and --tools that do not look like complete rules, so a malformed rule is named at launch instead of silently doing less than you intended. Background sessions that lose the LLM gateway your shell exported, after a crash, a reopen from another folder, or a background service that lacks it, now stop before any model request instead of going ahead without it. And Remote Control replies now stream to your phone or browser as they are written, for sessions from claude remote-control, Claude Desktop and VS Code.

What to do

The rest of the release is a very long list of fixes across the terminal, fullscreen rendering, Remote Control and sign-in screens. Sample it if one of your own annoyances sounds familiar. The changes that alter what Claude Code is allowed to do, though, are the ones above, and they all point the same way: narrower writes, earlier questions, and credentials kept out of reach.

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Claude Code 2.1.323: Code Mode writes the tool calls