Claude Code shipped v2.1.200 on July 3. If you updated without reading the changelog — which, honestly, most people don't — your workflows may be behaving differently than you expect. The headline change: the default permission mode is now Manual across the CLI, VS Code, and JetBrains. Both --permission-mode manual and "defaultMode": "manual" are now accepted.

In practice, Manual means Claude asks before taking actions it would have taken silently under the old default. File writes. Shell commands. Tool calls not explicitly pre-authorized. If you had a .claude.json with an allowlist or were passing --dangerously-skip-permissions explicitly, you're fine. If you were relying on the old default being permissive, you're about to see confirmation dialogs where you weren't seeing them before.

Manual

→ Source: Claude Code releases

Source spread

• Claude Code releases — v2.1.200 [builder] — Primary source for the permission mode change. Release notes are detailed and specific.
• Claude Code releases — v2.1.198 [builder] — Subagents in background by default, Claude in Chrome GA, background agents auto-commit behavior.

What else changed in the same week

Three days before the permission flip, v2.1.198 made subagents run in the background by default. Claude keeps working on the main conversation while agents run. You get notified when they finish. Background agents that complete code work in a worktree now auto-commit, push, and open a draft PR — instead of stopping to ask you what to do next.

That last part is the one to actually sit with. An agent finishing a coding task will now create a commit, push it, and open a pull request in your repo. If you have CI on PR open, branch protection rules, or code owners reviews, those all kick in. Test this in a non-production repo before relying on it in any workflow where unwanted PR creation is a problem.

Also in v2.1.198: Claude in Chrome is now generally available. The experimental flag is gone. And the built-in Explore agent now inherits the main session's model, capped at Opus. If you were implicitly relying on Explore running on Haiku as a cost-control measure, that assumption just broke.

Pros & cons

What's actually good:

• Manual-default is the right default for a tool this capable. Six months ago, a permissive default made sense while builders explored the product's shape. Now that Claude Code is running real production workflows, a permissive default is a liability. Anthropic made the correct call here, probably later than they should have.
• Background agents that auto-commit remove a friction point that was genuinely annoying. Waiting for an agent to finish before doing anything else in your session was bad UX for anything that takes more than a minute.
• The Explore agent running on a real model is a quality improvement. Haiku is fast and cheap; it's also the wrong choice for exploratory search tasks where you actually want the model to reason.

What deserves a side-eye:

• Every confirmation dialog is friction. Manual mode in a tool whose value proposition is autonomy is a real tension. Anthropic is betting that explicit authorization + faster execution beats implicit authorization + blocking execution. That's a reasonable bet. It's still a bet, and workflows that ran unattended will now stop.
• Auto-PR-on-completion is a meaningful change in what "autonomous" means here. Commits appearing in your repo, PRs opening in your review queue, CI running — these are things that happen because an agent finished a task you may not have been watching. Know what you're authorizing.
• AskUserQuestion dialogs no longer auto-continue by default. If you had sessions running unattended and relying on a timeout to proceed, they will now block waiting for input that never comes.

Samwise's take

What builders need to know

For builders

• Check your permission config before anything else. If you relied on the old "default" mode being permissive, add an explicit override to your .claude.json: "defaultMode": "bypassPermissions" or "auto" if you want the old behavior.
• AskUserQuestion timeout is now opt-in. Unattended sessions relying on auto-continue need to set an idle timeout explicitly via /config. Otherwise they stall.
• Test background agent auto-PR in a sandbox first. Run it in a throwaway repo with no branch protection before you rely on it in any repo where unwanted commits or open PRs cause real problems.
• Explore agent now inherits your main session's model. If your session is on Sonnet 5 or Opus, Explore runs on that too (capped at Opus). Review the cost implications for search-heavy workflows.
• Claude in Chrome is GA. If you were waiting on a stable Chrome integration for browser-based automation or research workflows, the experimental flag is gone and you can build on it now.

Further reading

• Claude Code v2.1.200 release notes — full changelog for the permission mode change and AskUserQuestion update
• Claude Code v2.1.198 release notes — background agents default, Claude in Chrome GA, auto-commit behavior details