Claude Code, Anthropic’s command-line coding agent, ships updates almost daily, and most of them are safe to install without reading the notes. Version 2.1.325, released on 13 November 2026, is not one of those. Besides a security fix in how zstd-compressed responses are decoded, it tightens several permission and sandbox defaults, which means scripts, scheduled tasks and CI jobs that ran unattended yesterday may now stop and wait for approval. This guide summarises what changed, what might break, and the exact settings to review.
Everything below comes from the official v2.1.325 release notes.
Update and check your version
claude update
claude --version
If you installed with npm:
npm install -g @anthropic-ai/claude-code@latest
Then run claude doctor. In this release it also lists an unrecognised teammateMode settings value as a skipped entry instead of silently ignoring it, which is a quick way to catch typos in your config.
The security fixes
The headline item is a fix for “a security issue in how zstd-compressed responses are decoded”. The release notes don’t give more detail, so the practical advice is simple: update. Many other fixes close permission gaps, including:
- A gap in Bash permission checks where some compound commands could run without asking. These now ask.
- A case where a Bash allow rule could run a different program than the one it names.
- Some deny and ask rules not applying to certain Bash commands.
- Relative deny and ask rules such as
Read(./.env)no longer covering the project after acd. They now also apply in added directories. - Cases where the Bash sandbox didn’t fully protect a worktree’s settings and hooks, or a linked worktree’s shared git folder.
- Sandboxed commands being able to write to the native updater’s download folder in some cases.
If you rely on deny rules to keep secrets like .env files away from the agent, this release makes them more reliable. It’s still worth re-testing them (see the checklist at the end).
Behaviour changes that can break automation
These are the items most likely to affect you:
1. Sandboxed Bash commands now ask by default
“Changed sandboxed Bash commands to follow the normal permission flow by default.” Previously, commands inside the sandbox could run without prompts. If you want the old behaviour, opt in explicitly in your settings:
{
"sandbox": {
"autoAllowBashIfSandboxed": true
}
}
For organisations that enable the sandbox through managed settings, the notes say to put this key in your highest-ranking managed source to keep commands running without prompts.
2. Scheduled tasks and /loop ask before running some skills
Scheduled tasks and /loop now ask before running a skill that declares allowed-tools or hooks. To let a specific skill run unattended, allow it by name:
{
"permissions": {
"allow": ["Skill(release-notes)"]
}
}
Replace release-notes with your skill’s name.
3. Read-only git commands in other repositories ask first
Read-only git commands now ask first in a repository that isn’t the project’s or one you’ve trusted. To run them without prompts, name the repository with the --add-dir flag (the notes say to use --add-dir, not the additionalDirectories setting):
claude --add-dir ../shared-library
4. “Always allow” works differently in Desktop, IDE and SDK hosts
A rule on its own is now offered only when saving it would actually allow that command again. A scheduled task that relied on such a rule may now wait for you.
5. Stricter sandbox and worktree rules
- The sandbox refuses writes to
.claudesettings, skills, agents, commands and hooks in added directories, as it already did in the launch folder. - Isolated worktree sessions are refused
git bisect run,git maintenance register|startand Scalar’s set-up commands. EnterWorktreerefuses aWorktreeCreatehook path that goes through a symlink inside the project.- Self-hosted runners refuse a session whose folder or repository path is a symbolic link (other than a
checkouthook’s checkout).
6. Secrets kept out of helper processes
The programs the sandbox starts for its own use (such as its network bridge) now run without Claude Code’s sign-in tokens and API keys in their environment. Ripgrep, local git commands and the sandbox runtime’s helpers also run without the mTLS client key’s passphrase. This is invisible unless a custom helper depended on those variables.
Fixes for everyday sessions
Plenty of quality-of-life bugs are fixed too:
- Subagents with their own model were moved to the plan-mode model when the model setting was
opusplanorhaiku. Fixed. - Subagents on a different model sized file reads and MCP results using the main session’s model. Fixed.
- Sessions on many third-party gateways and Foundry stopped with a context-length error every turn instead of compacting. Fixed.
- Recent turns sometimes went missing from a resumed session after an unclean exit. Fixed.
- Shell commands ran in the launch directory after the working directory was deleted; now the command fails with a message.
- Subagents cut off at the output token limit are now asked for the whole answer again, more briefly, instead of returning only the end.
If you use an LLM gateway with a custom model name, note one more change: CLAUDE_CODE_MAX_CONTEXT_TOKENS now also sets the context window for unrecognised model IDs that start with claude-, such as a gateway’s claude-sonnet-latest.
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=200000
claude
Pick a value that matches what your gateway actually supports.
For CI and headless users
If you run claude -p in pipelines, check these:
- With no cached account data,
claude -pnow waits for that data before its first request, so a first run in a clean environment can start a few seconds later. Adjust tight timeouts. claude -ptext output on a terminal no longer passes control sequences other than text styling. Pipes and files are unchanged.- In
-pand SDK runs inside a worktree or submodule mounted into a container,.claude/settings.local.jsonallow rules apply only once the folder is trusted. claude remote-controlnow exits with status 4 when it stops by itself and 69 when it can’t register, so a service manager can restart it.
A simple smoke test after upgrading:
claude -p "List the files in this repo's root and stop." --output-format text
echo "exit code: $?"
If it hangs waiting for approval, one of the new permission defaults is involved. Look at the prompt it’s waiting on and add a narrow allow rule rather than a broad one.
Plugin authors
Plugin MCP bundles (.mcpb, .dxt) now unpack under the plugins directory instead of .mcpb-cache inside the plugin folder, so read-only plugin folders load. Plugin loading skips some unsafe paths and links, and claude plugin validate reports them. Run validation before publishing your next version.
Upgrade checklist
- Run
claude update, thenclaude doctor. - Decide whether you want
sandbox.autoAllowBashIfSandboxed: true, and set it deliberately. - Add
Skill(name)allow rules for skills used by scheduled tasks or/loop. - Re-test deny rules such as
Read(./.env)after acdinto a subfolder. - Smoke-test every CI job that uses
claude -p. - If you use a gateway, set
CLAUDE_CODE_MAX_CONTEXT_TOKENSfor customclaude-*model IDs.
Takeaway
Claude Code 2.1.325 trades a little convenience for safer defaults: it fixes a decoding security issue, closes several permission gaps, and makes sandboxed and scheduled work ask before acting unless you opt in. Update now, then spend ten minutes reviewing your settings and CI jobs so nothing stalls. The full list of changes is in the release notes.
