Lifestyle

Claude Code 2.1.323 Update: Experimental Code Mode, Plugin Version Limits and Stricter Sandbox Permissions

Per Anthropic's changelog, Claude Code 2.1.323 adds an experimental Code Mode and plugin version-range checks, and tightens sandbox, MCP and permission behavior. Developers, plugin authors and team admins should know which settings and workflows may be affected before upgrading.

News date: About 6 min read

Share on LINE
Contents
  1. 1. Key points of this update
  2. 2. Changes to security and permission behavior
  3. 3. Settings that may be affected before you upgrade
  4. 4. Minor enterprise and admin additions

Key points of this update

Anthropic says in its changelog that Claude Code 2.1.323 adds Code Mode, an experimental, opt-in feature in which Claude writes JavaScript to call its tools. The changelog points to a documentation page for turning the feature on, or for turning it off across an entire organization.

For plugins (extensions that add features to Claude Code), Anthropic says the plugin.json manifest now supports engines["claude-code"]: a plugin outside its declared Claude Code version range will not be installed, updated or loaded, and installs and updates pick the latest version that fits the range. The claude plugin validate command, used to check plugins, also now type-checks a plugin's TypeScript hooks module (hooks are custom programs that run automatically at specific moments). Each finding appears as a warning, and adding --strict makes it fail.

Other additions include: .docx support for Artifact tool asset uploads; an options menu (stop, switch to usage credits, upgrade) when a model needs usage credits that are turned off or used up; and, in claude -p and SDK sessions, a progress heartbeat every 30 seconds while hooks run for a long time. Anthropic says this keeps hosts such as Claude Desktop from treating the turn as stuck (once they pick up this version).

Changes to security and permission behavior

Anthropic's changelog lists several tightening measures. A sandbox is an isolated environment that limits which files and resources a command can access. In the macOS and Linux sandbox, git worktree add, move and remove now fail as read-only (a worktree is an additional working folder for the same git repository), and so do commits in a worktree the session wasn't started in or given, even one inside the project. Allow rules created by "don't ask again" that appear to contain secrets, or that would be saved to a git-tracked settings.local.json, now apply only to that session and are not saved.

An MCP server is an interface that lets Claude connect to external tools and services. The tool Claude uses to finish an MCP server sign-in now asks the user to confirm, even in bypass and auto modes; when no one is available to ask (for example in dontAsk mode), /mcp must be used. Approval of MCP servers from a project's .mcp.json now matches the exact name, so approvals written under a different spelling will prompt again. In addition, worktree creation now refuses checkouts whose .git folder is a link or contains hard links, and Anthropic recommends using a plain clone instead.

Anthropic also says that Claude Code's own ripgrep and most internal git commands now run without most sign-in tokens and API keys, so credential helpers or hooks under those git commands can no longer find them. One more change: running /logout in a background session previously had no effect; it now signs out every session on the computer, deletes stored MCP logins and plugin secrets, and ends that session.

Settings that may be affected before you upgrade

The table below lists some changes in the changelog that may require adjusting settings. The "context window" is the amount of content a model can process at one time.

Selected changes compiled from Anthropic's changelog
ItemChange per the changelogWho may be affected
Context window for unknown modelsAssumed size changes from 200K to 1M; set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep 200KAnyone using a model ID Claude Code does not recognize
Smaller-window models from other providersOnce the window fills, every turn may fail until /clearSessions using other providers' models
Proxy serversOrganization settings and policy requests no longer use a proxy configured only via ALL_PROXY; HTTPS_PROXY must be setTeams on proxied networks
Self-hosted runnersOn-demand runner exit code for a superseded work order reverts to 1 (2.1.321 and 2.1.322 had changed it to 0)Automation that judges results by exit code
Subagent status lineThe subagentStatusLine setting is removed; settings files containing the key still load as usualUsers with a custom status line
Plugin agent settingDeprecated; still works but warns at startup and elsewherePlugin authors

Minor enterprise and admin additions

Anthropic says sessions that Claude Tag starts in Slack are now covered by the Enterprise plan's Compliance API (beta), letting admins list sessions and read transcripts. Code Review admin settings add a Download CSV button that exports review costs by month, week or day for a single repository, a single GitHub organization or all of them. The VS Code extension adds a trust prompt: it asks before the first session in a folder you haven't trusted, and if you decline, that folder's settings and MCP servers are not used.

Frequently asked questions

Is Code Mode on by default?

No. Per Anthropic's changelog, Code Mode is an experimental, opt-in feature; organizations can also turn it off as described on its documentation page.

Will my plugin stop working because of this update?

Per Anthropic, if a plugin declares a version range with engines["claude-code"] in plugin.json, Claude Code versions outside that range will not install, update or load it. The changelog does not mention an effect on plugins that declare no range.

Why might some git-related hooks fail to find credentials?

Anthropic says Claude Code's own ripgrep and most internal git commands now run without most sign-in tokens and API keys, so credential helpers or hooks under them can no longer find them.

What should I watch for when using other providers' models?

Anthropic says the assumed window for unrecognized model IDs is now 1M; if the model's window is smaller, every turn may fail once it fills, and you need /clear. The changelog suggests setting CLAUDE_CODE_DISABLE_1M_CONTEXT=1 beforehand.

Where can I see the full list of changes?

This article covers only selected highlights from Anthropic's official Claude Code changelog; rely on the official changelog and documentation for the complete list and details.

Browse the latest news in this topic

Found this guide helpful? Share it with someone who needs it

Share on LINE

Sources

Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.

Lifestyle