Anthropic published three Claude Code releases on GitHub between November 12 and 16: v2.1.323, v2.1.325 and v2.1.326. Two changes stand out for teams that run Claude Code unattended, in CI or under managed settings. Starting with the builds shipped on November 13, sandboxed Bash commands “follow the normal permission flow by default”, where the old default approved them without a prompt. And v2.1.323 added Code Mode, “an experimental opt-in in which Claude writes JavaScript to call its tools”. v2.1.323 notes · v2.1.325 notes · v2.1.326 notes
Everything below comes from the GitHub release notes, the repository changelog, the npm registry, Anthropic’s Claude Code documentation and two user-filed GitHub issues, read on November 17. S5 Labs has not installed or run these versions. The notes are long, and this article covers mainly entries that change a default or a permission decision, plus the security fixes. These are client changes, and the API side is in our pieces on the 2026-10-21 API version and on linked billing and self-hosted Managed Agents.
Which release carries what
| Version | GitHub release | What it carries |
|---|---|---|
| 2.1.323 | November 12 | Code Mode; unrecognized model IDs assumed to have a 1M window; v1 MCP client removed |
| 2.1.324 | None (npm only, November 13) | Per the settings reference, the first version with the new sandbox default |
| 2.1.325 | November 13 | The sandbox change as the notes list it; new prompts for scheduled tasks, git and “Always allow” |
| 2.1.326 | November 16 | Code Mode with agent teams; hook "decision": "block" change; more commands ask or are refused |
The sandbox change is easy to misdate. The note sits under 2.1.325, but the settings reference dates the change to the build before it: “Default: false. Before v2.1.324, the default was true”. Version 2.1.324 was published to npm on November 13 at 18:42 UTC, about an hour and a half before 2.1.325, and has no GitHub release, no git tag and no changelog entry. So a machine on 2.1.323 has Code Mode available and still has the old sandbox default. Settings reference · npm versions · Changelog
On npm, the latest tag points to 2.1.326 and stable to 2.1.311 as of November 17, so stable-channel npm installs have none of this yet. We did not confirm the native installer’s stable version against Anthropic’s setup docs. npm versions
Sandboxed Bash now goes through the permission flow
The release note reads: “Changed sandboxed Bash commands to follow the normal permission flow by default; set sandbox.autoAllowBashIfSandboxed: true to keep running them without prompts”. v2.1.325 notes
This affects only setups that use the sandbox. The sandboxing doc says “The sandbox is off by default”, and it is turned on with /sandbox or sandbox.enabled: true. For those setups the old default, which the docs call auto-allow mode, approved a command with no prompt when it ran inside the sandbox. Under the new default, “sandboxed commands go through the regular permission flow, so your allow rules and permission mode decide”. A pipeline that relied on the sandbox as its approval now needs allow rules, a permission mode that covers its commands, or the setting put back. Sandboxing · Settings reference
The notes give managed deployments their own line: “to keep commands running without prompts, set sandbox.autoAllowBashIfSandboxed: true in your highest-ranking managed source”. The sandboxing doc adds that even when the sandbox is admin-required, a repository’s files can still supply the autoAllowBashIfSandboxed value, and says to “Set the key in managed settings to keep a repository from changing it”. Admins who want the prompts should therefore set false explicitly and not rely on the default. v2.1.325 notes · Sandboxing
Putting the setting back to true does not switch off permission rules. Per the settings reference, “deny rules and content-scoped ask rules such as Bash(git push *) still apply, but a bare Bash ask rule is skipped for sandboxed commands”, and CLAUDE_CODE_SUBPROCESS_ENV_SCRUB turns auto-allow off. Auto-allow is also a different feature from auto mode, the classifier-based permission mode. The docs say the two “work independently and can be combined”, with listed exceptions. Settings reference · Sandboxing
Other changed defaults that can stall an unattended run
| Release | What changed | Remedy in the notes |
|---|---|---|
| 2.1.325 | Scheduled tasks and /loop ask before running a skill that declares allowed-tools or hooks | Add Skill(name) to permissions.allow |
| 2.1.325 | Read-only git commands ask first in a repository that isn’t the project’s or one you trusted | Name it with --add-dir, not the additionalDirectories setting |
| 2.1.325 | “Always allow” in Desktop, IDE and SDK hosts offers a rule alone only when saving it would allow that command again | Nothing given; “a scheduled task that relied on one can now wait for you” |
| 2.1.323 | Context window assumed for an unrecognized model ID goes from 200K to 1M | /autocompact to compact earlier, or CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep 200K |
| 2.1.323 | The v1 MCP client is removed | MCP_SDK_GENERATION=v1 now runs the v2 client in a compatibility mode |
| 2.1.323 | Organization settings and policy requests no longer use a proxy set only by ALL_PROXY | Set HTTPS_PROXY |
| 2.1.326 | Some tool calls in bypass permissions mode now ask for approval, or are denied “where no one can be asked” | Nothing given |
| 2.1.326 | With auto-allow off, the Bash sandbox asks before some lines that set environment variables | Nothing given; “unattended runs can refuse them” |
| 2.1.326 | Auto mode ignores some wildcard shell allow rules, such as ones that name no program or that start with $HOME | Nothing given; “the safety check decides those commands” |
v2.1.323 notes · v2.1.325 notes · v2.1.326 notes
The model-ID change matters behind gateways. The 2.1.323 notes warn that sessions on other vendors’ models with a smaller window, once full, “can fail every turn until /clear”. Version 2.1.325 then made CLAUDE_CODE_MAX_CONTEXT_TOKENS set the window for unrecognized IDs that start with claude-, “such as a gateway’s claude-sonnet-latest”. v2.1.323 notes · v2.1.325 notes
Two 2.1.326 entries change which mode terminal and VS Code sessions start in. Sessions with no mode set now start in auto mode, not Manual, “when your organization’s policy hasn’t loaded”. Sessions in organizations with the Claude Code + Cowork HIPAA configuration start in auto mode by default, and disableAutoMode turns that off. v2.1.326 notes
The hook entry is the vaguest: “Changed a hook’s "decision": "block" to hold in more cases, so a hook that always prints one may start blocking calls”. The note does not say which events or cases changed, and the hooks reference carries no version note for it. v2.1.326 notes · Hooks reference
Security fixes in the same releases
The first line of the 2.1.325 notes is “Fixed a security issue in how zstd-compressed responses are decoded”. The notes give no detail and name no CVE or advisory. We did not check GitHub’s security advisories. v2.1.325 notes
Most of the other security-relevant fixes close gaps in permission checks. In 2.1.325: “some deny and ask rules not applying to certain Bash commands”, “some compound commands could run without asking”, “a Bash allow rule could run a different program than the one it names”, and a committed .claude/settings.local.json that could be treated as the user’s own when Claude Code was started in certain subfolders of a repository. In 2.1.323: claude -p and the Agent SDK “running Bash unsandboxed” when settings that turn the sandbox on fail validation (Claude Code now stops at start-up), “a case where the Bash sandbox could switch off during a session”, and --permission-mode auto starting in a mode that could allow more when auto mode was unavailable. That last one now starts in Manual mode. v2.1.323 notes · v2.1.325 notes
Pinning 2.1.323 to avoid the new sandbox default keeps the 2.1.325 gaps, and pinning anything older keeps the 2.1.323 ones too. For a pipeline that depends on deny rules or the sandbox, the fixes are a stronger reason to upgrade than the prompts are to stay put.
What Code Mode does to permission review
In Code Mode, Claude writes a piece of JavaScript that makes several tool calls, and Claude Code runs it in the session. The Code Mode doc calls it “experimental and off by default” and says it “may change or be removed in any release”. Code Mode doc
Most unattended setups cannot use it. It works only in an interactive terminal session, and the doc excludes -p, the Agent SDK, the VS Code panel, the Desktop app and cloud sessions. It is unavailable through a gateway, Amazon Bedrock, Claude Platform on AWS, Google Cloud or Microsoft Foundry, and in organizations on the HIPAA configuration or zero data retention. It needs Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2, and it will not start with an API key helper, OpenTelemetry export or DISABLE_TELEMETRY set.
A user turns it on with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 in the shell or in the env block of ~/.claude/settings.json. Claude Code ignores the variable in a project’s settings files, so a cloned repository cannot enable it. Admins can block it with the same variable set to 0 in managed settings, but “You can’t turn on Code Mode for your users from managed settings.”
Rules and hooks still run per call. Each tool call in the code goes through deny and ask rules and PreToolUse hooks under the tool’s usual name, such as Bash.
In auto mode, the classifier “reviews each piece of code before it runs, not each tool call. It doesn’t see values computed at run time, such as a path taken from search output, so keep rules or hooks for calls you must control”. In the other permission modes, the code starts without a review and you are prompted for each call that would prompt in that mode.
So in an auto-mode session, a call whose arguments are computed inside the script gets less review than the same call made outside Code Mode, because the classifier reads the script before those arguments exist. The doc also warns that Bash rules match command text, so a denied command can run “in a form the rule doesn’t match”, with or without Code Mode, and it points to the sandbox for limits that don’t depend on command text.
A hook can also receive a command in a different form. A git push in a folder named app can reach a hook as cd 'app' && git push, so a hook that allows or blocks shell commands has to check each part of a compound command. To inspect the code itself, match REPL in a PreToolUse hook and read tool_input.code. A PreToolUse hook that fails to run or times out refuses the call even when onFailure is unset. Code Mode doc
Version 2.1.326 lets Code Mode run with agent teams and adds this line: “Improved permission checks in Code Mode; more sandboxed commands now ask or are refused”. v2.1.326 notes
What the sources leave open
The release notes and the docs we read do not say when the stable channel will move. The Code Mode doc gives no figures for time or tokens saved, and we did not review any third-party testing.
Early GitHub issues are user reports, not vendor statements. #110568, opened November 17 by a user on 2.1.325, says auto mode blocks Bash commands rewritten by a PreToolUse hook, and its author had not verified a cause. A 2.1.326 entry changes how a PreToolUse hook’s rewritten input is kept when more than one hook answers a call, but nothing we read ties it to the report. #110714 reports Code Mode tool calls hanging. We found no report of CI breakage from the sandbox change, though our search was not exhaustive.
What to check after upgrading
Version. Run claude --version on every runner image. Version 2.1.323 has the old sandbox default, and 2.1.324 and later have the new one.
The sandbox setting. If sandbox.enabled is true anywhere, set sandbox.autoAllowBashIfSandboxed explicitly to the value you want. In managed settings, put it in the highest-ranking source.
Scheduled tasks and /loop. List the skills they call. Any that declare allowed-tools or hooks need a Skill(name) entry in permissions.allow, and tasks that leaned on an “Always allow” answer may now wait for a person.
Hooks. Find hooks that always print "decision": "block" and test them on 2.1.326.
Gateways. For other vendors’ models with smaller windows, the 2.1.323 notes say to set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before sessions fill. For gateway aliases that start with claude-, set CLAUDE_CODE_MAX_CONTEXT_TOKENS.
Code Mode policy. Decide before a developer enables it. To block it, set CLAUDE_CODE_EXPERIMENTAL_CODE_MODE to 0 in managed env. To allow it, write deny rules for what must never run, add a PreToolUse hook on REPL if you need to inspect the code, and keep the sandbox on.
Then run one unattended job end to end on 2.1.326 and read its log for refused calls, because several of these changes turn a silent approval into a prompt, or into a refusal when nobody is there to answer.
Sources
Release notes and registry
- v2.1.323, v2.1.325, v2.1.326, CHANGELOG.md — Anthropic on GitHub
- @anthropic-ai/claude-code versions — npm
Claude Code documentation
- Code Mode, Sandboxing, Settings reference, Hooks reference — Claude Code docs
User reports
- Issue #110568, Issue #110714 — anthropics/claude-code on GitHub
