💡 Stop Digging Through Menus: /config key=value Sets Options Inline

Claude Code's /config command now accepts a direct key=value form, letting you change a setting from the prompt in one line instead of opening the interactive config menu and navigating to the right field. It's a small syntax addition, but it removes a real amount of friction for anyone who tweaks the same handful of settings repeatedly across sessions — model selection, permission mode, output verbosity — since each one used to cost several keystrokes of menu navigation.

Where this earns its keep

• Scripted session starts: pipe a /config line into a session-init routine to guarantee consistent settings without manual steps.
• Mid-session pivots: switch permission mode or model tier on the fly when a task's risk profile changes, without losing your place in the conversation.
• Muscle memory: if you already know the setting's key name from documentation or settings.json, you can now use that same name at the prompt — no separate menu vocabulary to learn.

/config permission-mode=plan
/config model=claude-sonnet-5

💡 Auto Mode Now Blocks Destructive Git Commands by Default

Claude Code's auto mode — where the agent executes commands without asking for per-step confirmation — now refuses destructive git operations (reset --hard, force pushes, clean -f, branch deletion) unless you've explicitly permitted them for that session. Previously, a permissive auto-mode configuration could let an agent run any git command it decided was necessary; this update carves out the genuinely irreversible ones as a standing exception, regardless of your broader permission settings.

Why a carve-out rather than a setting

The point of auto mode is to remove friction for the vast majority of safe, reversible actions — edits, reads, most git operations, test runs. Destructive git commands are a different category: they can discard work with no local recovery path. Anthropic's framing here matches general good agent-safety practice — reversible-by-default automation, with an explicit, narrower permission gate for the handful of actions that aren't reversible.