Claude Code 2.1.326 to 2.1.330: auto mode reaches Opus 4.6 and Sonnet 4.6 on Bedrock, Agent Platform, and Foundry
Five Claude Code releases shipped November 16 to 20. The big one for teams on cloud providers: Opus 4.6 and Sonnet 4.6 sessions on Bedrock, Google Cloud's Agent Platform, Foundry, and the Claude apps gateway now start in auto mode. Pin permissions.defaultMode if you want something else, and recheck your allow rules and hooks.
Everything here comes from the Claude Code changelog, which gives no dates, and the publish times on npm. We read the notes and haven't run these builds ourselves.
One thing to know first: npm's stable tag still points at 2.1.323, while latest is 2.1.330. If your team installs from the stable channel, none of this has reached you yet. It will, so read on before it does.
The five releases
| Version | Published (UTC) | What matters for teams |
|---|---|---|
| 2.1.326 | November 16 | File tools refuse writes to sign-in files; auto mode ignores more broad allow rules; --add-dir shows the trust prompt |
| 2.1.327 | November 17 | ExitWorktree asks before deleting work; claude mcp add-json reads stdin; billed classifier fallback is going away |
| 2.1.328 | November 18 | Auto mode for Opus 4.6 and Sonnet 4.6 on third-party providers; "onFailure": "block" for MCP tool hooks |
| 2.1.329 | November 19 | Code Mode stops refusing tool calls when a hook errors; trust dialog names where hooks are declared |
| 2.1.330 | November 20 | Sign-in to organizations managed in the Enterprise Admin Console |
Which sessions now start in auto mode
Auto mode lets Claude Code run actions without asking, with a safety check deciding which ones need you. Earlier releases already started interactive sessions in auto mode when no permission mode was set, on the models that supported it.
2.1.328 closes that gap. It adds auto mode "for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and Claude apps gateway; interactive sessions with no permission mode start in it." The same release changes plan mode and Bash prompts for those models to use and offer auto mode.
That matters most for enterprise teams that stayed on 4.6 models through a cloud account. If you never set a mode, developers who upgrade will see fewer prompts the next morning. Two settings control it:
permissions.defaultModeoverrides the default, per earlier changelog entries. Set it in managed or project settings to pin the mode you want.disableAutoModeturns auto mode off, including the new plan-mode and Bash prompt behavior on third-party providers.
Also from 2.1.326: terminal and VS Code sessions with no mode set now start in auto mode, not Manual mode, when your organization's policy hasn't loaded yet. Once settings load, a server-managed defaultMode takes over. It falls back to Manual only if settings fail with no stored copy and the organization's policy also fails or requires Manual.
Allow rules that auto mode now ignores
Auto mode trusts your allow rules less than it used to. Broad rules no longer skip the safety check, so a rule you wrote to stop prompts may now be judged case by case:
- 2.1.326 ignores wildcard shell rules "that name no program or that start with
$HOME", and judges wildcard rules whose program is a Windows backslash path. - 2.1.327 ignores more rules that run any package, such as
Bash(pnpm dlx *), plus path or version forms of ones it already ignored, like/usr/bin/npm run:*. - 2.1.327 also makes
permissions.blockReadsOutsideWorkingDirectoriesask about more Bash commands in bypass and auto modes, and deny some in auto mode. - 2.1.329 runs
git statusless often for speed, so a git command that discards changes may be blocked until the working tree is checked.
This is the right direction. A rule like Bash(pnpm dlx *) lets the agent download and run any package, which is exactly what a supply-chain attack needs. If those commands now get checked, replace the wildcard with exact rules for the packages you use.
2.1.327 also warns that the billed classifier fallback goes away in an upcoming release. After that, each action it would have checked will be asked about or denied. Unattended jobs that rely on it should plan for more refusals.
Hooks that must block now need a flag
2.1.328 adds "onFailure": "block" to mcp_tool hooks, "as for command and HTTP hooks: a hook whose call fails blocks the action." Without it, a policy hook that crashes or times out doesn't stop anything.
2.1.329 makes that default explicit in Code Mode, where Claude runs scripts that call tools. A PreToolUse hook that errors "no longer refuses a script's tool call, as outside Code Mode." If you used a hook as a guardrail and relied on errors failing closed, add the flag to every hook that must refuse. The trust dialog in 2.1.329 also now says where a folder's hooks are declared, which helps catch hooks a cloned repo brings with it.
For hosts that embed Claude Code, 2.1.328 adds CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME. It stops claude -p --resume and --continue from running a tool call that a hook deferred.
Worktrees, credentials, and secrets
- Sign-in files: 2.1.326 makes Claude's file tools refuse writes to Claude Code's sign-in files, such as
.credentials.json. - Worktrees: 2.1.326 stops
ExitWorktreeremoving a worktree while a background subagent or shell command may still run in it. 2.1.327 makes it ask before removing one that may hold unsaved work, and 2.1.330 fixes it deleting a worktree another running session was using. - MCP tokens: 2.1.327 adds
claude mcp add-json <name> -, which reads the server's JSON from stdin "so tokens can stay off the command line" and out of shell history. - Extra folders: 2.1.326 makes
--add-dirand/add-dirshow the trust prompt for a folder you haven't trusted, andclaude --bgrefuses such a folder.
For admins: Enterprise Admin Console sign-in
2.1.330 adds sign-in to organizations managed in Anthropic's Enterprise Admin Console. After claude auth login, Claude Code applies the organization's settings and uses its cloud provider or gateway. That gives admins one place to push permissions.defaultMode and disableAutoMode instead of relying on each developer's settings file.
Teams on the self-hosted runner get a related warning in 2.1.328: the runner now warns at start-up when the host's settings.json sets a default permission mode that sessions ignore. If you set the mode there, move it to where sessions read it.
What to do this week
- Decide which mode Opus 4.6 and Sonnet 4.6 sessions on your cloud provider should start in, and set
permissions.defaultModeordisableAutoModein managed settings before developers upgrade to 2.1.328 or later. - Search your settings for wildcard allow rules like
Bash(pnpm dlx *),npx *, or rules starting with$HOME, and replace them with exact commands. - Add
"onFailure": "block"to every hook that enforces policy, including MCP tool hooks. - Move MCP server setup that passes tokens as arguments to
claude mcp add-json <name> -. - Pin your rollout to
stableorlateston purpose, since the two are seven versions apart today.
For background on how auto mode's server-side check works behind gateways, see our Claude Code 2.1.319 guide and our Code Mode explainer.
Sources
- Claude Code changelog: release notes for 2.1.326 through 2.1.330, and earlier entries on
permissions.defaultMode. - npm: @anthropic-ai/claude-code: publish times and the
stableandlatesttags.
Researched and drafted by an AI agent from the linked sources. How posts are made.