Your company runs Claude Code through Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry, and the model is pinned to Claude Opus 4.6 or Sonnet 4.6. Until this week, every new terminal session asked before it edited a file or ran a command. After an update on Thursday, it stopped asking. Nobody changed a setting.
That is one line in the Claude Code changelog. Version 2.1.328, published on November 19, turned on auto mode for those two models on those providers, and made it the mode new sessions start in. This post explains which sessions changed, what auto mode does instead of asking you, what it adds to your cloud bill, and the settings that pick the mode on purpose.
What 2.1.328 says
The Claude Code changelog has two entries about it under 2.1.328:
Added 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
Changed plan mode and Bash prompts for Opus 4.6 and Sonnet 4.6 on third-party providers to use and offer auto mode, as on newer models;
disableAutoModeturns both off
The GitHub release for v2.1.328 is dated November 19, 2026. Versions 2.1.329 and 2.1.330 followed and do not undo it.
Why your sessions used to ask
Since version 2.1.283, auto mode has been the built-in starting mode for interactive terminal and VS Code sessions on every plan and provider. The permission modes page says so, and it also says what happens when the built-in default picks auto but auto mode is not available to the session: Claude Code starts in Manual mode instead.
A model that auto mode does not support is one of the reasons it can be unavailable. When we read the page on November 22, its requirements list still said that on Bedrock, Agent Platform, Foundry and the Claude apps gateway, auto mode works only with Sonnet 5 or later, Opus 4.7 or later, Haiku 5.5 and the Fable models. So a session on Opus 4.6 or Sonnet 4.6 on those providers fell back to Manual every time, and Manual asks before most edits, shell commands and network calls. The changelog is newer than that list. Go by the changelog until the page catches up.
From 2.1.328 on, the same session has auto mode available, so the built-in default applies and it starts in auto.
Who is affected
All of these have to be true for a session to change:
- It is an interactive session in a terminal or through the VS Code extension.
- It runs on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or a signed-in Claude apps gateway session.
- The model is Claude Opus 4.6 or Sonnet 4.6.
- No
--permission-modeflag and nopermissions.defaultModein a settings file picks a mode. - No settings file sets
disableAutoModeto"disable".
Scripts are not affected. For claude -p and the Agent SDK, the built-in default is default, which is Manual, according to the same page. That was the change in 2.1.302, which we covered in what 2.1.302 changed for claude -p.
The docs say Claude Code shows a notice the first time the built-in default starts one of your sessions in auto mode: once, at the top of the session, in a terminal, and as a card on the new-conversation screen in VS Code. If you dismissed it without reading, that was the change.
What auto mode does in place of a prompt
In auto mode a second model, the classifier, reviews each action before it runs. The docs say it blocks anything that goes beyond your request, targets infrastructure it does not recognize, or looks driven by hostile content Claude read. Deny rules block in every mode. Ask rules are checked before the classifier, so they still prompt you.
The auto mode configuration page is specific about the trust boundary. By default the classifier trusts only the working directory and the current repository's configured remotes. Claude can push to that repository, including its default branch, without a prompt, and open pull requests there when the task asks for one. A push to your company's other repositories or a write to a team bucket is blocked until you list it under autoMode.environment.
The same docs carry a warning worth repeating: auto mode reduces prompts but does not guarantee safety. A boundary you state in chat ("don't push until I review") is re-read from the transcript on each check and can be lost when the conversation is compacted. A rule in a settings file is not lost that way. We wrote about one way the classifier was steered by a repository file in the AGENTS.md auto mode bypass.
What it costs on a cloud provider
Each review is a model call. The permission modes page says classifier calls count toward your token usage on Bedrock, Agent Platform and Foundry, and each check sends a portion of the transcript plus the pending action, which adds a round trip before the action runs.
The classifier does not always run on your session's model. By default it uses Claude Sonnet 5. A session on Sonnet 4.6 uses Sonnet 4.6 for the classifier. If Sonnet 5 is not available from your provider or your organization's model policy excludes it, Claude Code falls back to the session's model. If your cloud account only has Opus 4.6 enabled, expect the classifier to run on Opus 4.6 too, and budget for it. CLAUDE_CODE_AUTO_MODE_MODEL (v2.1.302 or later) sets the classifier model to a full model ID, including your provider's own IDs.
Pick the mode on purpose
The fix is the same as for scripts: stop relying on the built-in default. A session takes its mode from the first of these that applies: the --permission-mode flag, then permissions.defaultMode in a settings file, then the built-in default.
To keep asking on your own machine, put this in ~/.claude/settings.json:
{
"permissions": {
"defaultMode": "default"
}
}
The next session shows ⏸ manual mode on in the status bar. A project's .claude/settings.json can set default too. It cannot set auto or bypassPermissions; the docs say those two values do not take effect from project files.
For a whole organization there are two levers in managed settings. permissions.defaultMode sets the starting mode, and people can still switch to auto with Shift+Tab. permissions.disableAutoMode set to "disable" removes auto from the cycle, and a session started with --permission-mode auto starts in Manual instead.
If you want to keep auto mode, add the checkpoints the docs suggest before you rely on it:
{
"permissions": {
"ask": ["Bash(git push *)", "Bash(gh pr create *)"]
}
}
The docs point out that these rules match commands that begin with git push or gh pr create, so git -C <dir> push slips past. A PreToolUse hook that reads the whole command is the stricter option. Then fill in autoMode.environment with your source-control organization, buckets and internal domains, or run /auto-mode-setup to generate entries, so the classifier stops blocking routine internal work. Put those entries in ~/.claude/settings.json or managed settings. The classifier does not read autoMode from a project's .claude/settings.json, so a cloned repository cannot add its own.
One more change in the same week points the same way. Version 2.1.330 makes more rm commands aimed at the root, the home folder, a top-level folder or the working folder ask for approval, and refuses them where nobody can answer.
Where Sessionboxer fits
We build Sessionboxer, so read this part with that in mind. This change does not touch Claude Code sessions in Sessionboxer. Each session runs in its own box, a Docker container or VM, and the agent runs there with all permissions granted, which is what agents that don't ask means. The box's Claude settings set defaultMode to bypassPermissions, so there is no prompt and no classifier. The container is the boundary.
That is a different trade from auto mode, and it has limits of its own. Nothing reviews each command, so the agent can do anything the credentials in its box allow, including a push to any repository they can write to. The box keeps your laptop's files and keys out of reach. It does not judge intent.
If your company sends Claude Code through its own gateway, Sessionboxer can do that from Global settings, including a rotating token through an apiKeyHelper command. Works behind a corporate proxy and the corporate proxy chapter of the guide cover it, and Claude Code through an LLM gateway in Docker explains why the helper has to run on the host. The guide does not describe Bedrock, Agent Platform or Foundry credentials, so if your company uses those directly rather than through a gateway, check the Connect your agent chapter before you plan around it.