Claude Code 2.1.326 runs Code Mode with agent teams on

Claude Code 2.1.326, released by Anthropic on November 16, lets the experimental Code Mode run while agent teams are enabled. If you had already turned Code Mode on, sessions with teams on now start in it. The same update changes hooks: a hook that always prints a “block” decision may start blocking calls after the update, because that decision now holds in more cases.
Most of the release notes on GitHub are fixes. The permission changes matter more, though: some commands that used to run without asking will now stop for approval or be refused, and setups that run with nobody watching are the ones that get refused.
What may stop working
Start with hooks. When more than one hook answers a tool call, the input a PreToolUse hook rewrites (updatedInput) is now kept in more cases. A PreToolUse hook that exits with code 127 now produces an error saying its check may not have run. HTTP hooks now ignore Content-Length and Transfer-Encoding in their headers setting and take the request’s length from its body.
Auto mode now stops some shell commands that ran before, and it ignores some wildcard shell allow rules, such as ones that name no program or start with $HOME, so its safety check decides those commands. Wildcard rules whose program is a Windows backslash path are now judged as well, so those calls can be blocked. Exact rules still allow what they name, and the changelog adds: “we plan to restore rules that start with .\”. Admins who set a managed autoMode.environment list without "$defaults" now replace every built-in entry in Claude Desktop and Agent SDK apps too. The fix the notes give is to add "$defaults" or copy over the entries that cover sensitive targets.
Some Bash commands on macOS now ask or get refused, some cp and mv commands ask first, and some tool calls in bypass permissions mode now ask for approval, or are denied when no one is there to answer. The PowerShell tool asks before more commands, and unattended runs refuse them unless an allow rule, auto mode or bypass permissions mode allows them. Claude’s file tools now refuse to write to Claude Code’s sign-in files, such as .credentials.json.
The Bash sandbox is tighter too. On macOS, Linux and WSL, sandboxed commands can’t write to some sensitive home folders by default. You can allow some of them again with sandbox.filesystem.allowWrite in user or managed settings. After each command, the sandbox also moves aside any unexpected entries it finds in a repository’s .git folder.
Some defaults changed as well. Terminal and VS Code sessions with no mode set now start in auto mode, instead of Manual mode, when the organization’s policy hasn’t loaded yet. Sessions that start in auto by default take a server-managed defaultMode once settings load, or Manual if settings fail with no stored copy and policy fails or requires it. Organizations with the Claude Code + Cowork HIPAA configuration now get auto mode by default in the terminal and VS Code, and disableAutoMode turns it off. In plan mode, automatic compaction now uses the model that is answering you, such as Opus with opusplan, so it can cost more.
Code Mode with agent teams
Code Mode arrived in version 2.1.323. In it, Claude writes JavaScript that makes its tool calls. In auto mode, the classifier reviews each piece of code instead of each tool call, and it doesn’t see values the code works out while it runs. This release also tightens Code Mode: more sandboxed commands ask or are refused, and the $ shell helper refuses some values with an error, in which case nothing runs. When Code Mode can’t be turned on, the message now says what you can do.
Fixes and new warnings
Among the fixes:
claude purgehid its prompts when stdout was redirected, so one Enter you never saw could delete a project’s data. The prompts now show in the terminal.- The spend limit menu removed the monthly limit without asking if Backspace and Enter were pressed as it opened.
- Resumed sessions lost earlier history when Claude Code ran low on memory, and lost recent turns after a crash if the previous turn ended with an empty reply.
- Parallel MCP tool calls could read another call’s saved result when two long results arrived in the same millisecond.
- grep in the Bash tool silently skipped some text files.
- Bash commands that delete or overwrite files, such as
find -deleteorcat a > b, were summed up in the transcript as a search or a read.
A few additions make problems easier to spot. A startup warning flags any --allowedTools, --disallowedTools or --tools entry that doesn’t look like a complete rule. A new row in /status says when your provider or a variable such as DISABLE_TELEMETRY turns features off, and Ctrl+O shows startup notices that were hidden. /update-config now uses about 13,000 tokens of context instead of 93,000.
Native installations update in the background. claude update updates by hand, and claude --version shows which version you’re running.
Up next
Microsoft shares slid as Claude got Sheets and Whiteboards
Microsoft stock fell 3% right after Anthropic put spreadsheets and whiteboards inside Claude, per CNBC, though Anthropic calls its Sheets “simple spreadsheet functionality.” When a quick budget or tracker is needed, which opens first, Claude or Excel?
Related stories
- Claude Code’s Code Mode runs tool calls as JavaScript
- Claude Code 2.1.323 lets Claude call tools via JavaScript
- Claude Code stops wiping .env files
- Claude Code Projects: proactive coordinator, /btw, /fork
- Cloud sessions in Claude Code now set up your repo
- Claude Code desktop gives each parallel task its own session
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

