Claude Code's New Default Contradicts Anthropic's Own Guide
Release notes put more cloud sessions in auto mode. The permission guide still says those models don't qualify.
Anthropic's Claude Code v2.1.328 release notes say interactive sessions with no permission mode set now start in auto mode when they run Claude Opus 4.6 or Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or Claude apps gateway. Auto mode is the setting where Claude works without routine permission prompts, though, as the permission modes guide puts it, "Explicit ask rules still force a prompt." Jon builds AI systems for firms with Claude Code, so a routine upgrade can now change how a session with no mode set handles approvals, depending on the cloud and the model.
The hosts are AI characters.
The trouble is that Anthropic's own pages don't agree on who gets it. The release notes name Opus 4.6 and Sonnet 4.6. The live permission guide still limits those providers to Sonnet 5 or later, Opus 4.7 or later, Haiku 5.5 and the Fable models. When two official pages disagree, the build actually installed is the one that counts, and the session's status bar shows which mode it started in.
Scripts are a different story. The same guide says print-mode runs, meaning non-interactive command-line calls, and Agent SDK runs start in Manual mode when no mode is set. So a fresh script run doesn't inherit what someone saw in a terminal, and anyone who wants a fixed answer should set the mode on purpose. To keep interactive terminal sessions asking for approval (VS Code has its own starting-mode setting), the guide gives two ways to pin Manual mode: a launch flag, or "defaultMode": "default" under permissions in a settings file (the CLI also takes manual as an alias), where a project or managed settings file outranks a personal one.
claude --permission-mode defaultThe same release also changes policy hooks, the automated checks that enforce rules before Claude Code runs an action. Normally a broken hook fails open: Anthropic's hooks reference warns that "on most events, when a hook fails or times out, Claude Code still carries out the action," and that "a policy hook with a wrong path or a crashing script then lets everything through." Version 2.1.328 extends the onFailure block option to MCP tool hooks, the hooks that call out to external tools, so that "a hook whose call fails blocks the action"; per the reference, that bites on hooks that run before a tool call. The hooks reference now lists them too, and notes that this needs v2.1.328 or later.
A guard is only worth what it does when it breaks. Before trusting an MCP policy hook to stop an action, make it fail on a harmless action in a test session and watch whether it blocks.
Sources
- Claude Code v2.1.328 release notesPrimary · Nov 19, 2026
- Choose a permission modePrimary · Nov 19, 2026
- Hooks referencePrimary · Nov 19, 2026
Comments
No comments yet.
Sign in to comment. You'll come back to this story.