Version 2.1.133

You can now choose whether new worktrees start from a clean remote default branch or from your current local HEAD.

{
  "worktree": {
    "baseRef": "head"
  }
}

• fresh remains the default and branches from origin/<default-branch>.
• head branches from the current local HEAD, so unpushed commits and feature-branch state are present.
• The setting applies to --worktree, EnterWorktree, and agent isolation.
• The setting is also exposed in the config UI as “Worktree base ref”.

Worktree setting schema and UI were added (search for "Which ref new worktrees branch from" and "Worktree base ref").

Users with the agents interface enabled can make Claude Code open the agents view by default.

{
  "defaultToAgentsView": true
}

• A new config UI toggle appears as “Open agents view by default”.
• When enabled, normal interactive startup can enter the agents view without requiring claude agents.
• This depends on the existing agents-view gate, so not every user will necessarily see it.

Config option and startup behavior were added (search for "Open agents view by default" and "defaultToAgentsView").

Claude Code now warns when a plugin disabled in ~/.claude/settings.json is re-enabled by a higher-precedence source such as project, local, flag, or policy settings. The warning explains exactly where to disable it instead.

{
  "enabledPlugins": {
    "formatter@anthropic-tools": false
  }
}

• Project-enabled plugins should be disabled in .claude/settings.local.json.
• Plugins enabled through --settings must be removed from the --settings value.
• Plugins enabled by managed policy cannot be overridden locally.

New warning UI and remediation text (search for "Plugin settings overridden" and "These plugins are disabled in ~/.claude/settings.json, but a higher-precedence source re-enables them:").

The enabledPlugins setting description now documents the precedence order directly: user < project < local < flag < policy.

• This makes it clearer why disabling a plugin in user settings may not work when the project re-enables it.
• The schema description now points users to .claude/settings.local.json for project-specific opt-outs.

Updated setting description (search for "Settings precedence is user < project < local < flag < policy").

Hook authors get structured reasoning-effort information for tool-use lifecycle hooks, and Bash receives the active value as CLAUDE_EFFORT when the model supports effort.

echo "$CLAUDE_EFFORT"

• Hook input now includes an optional effort.level field.
• The value reflects the active effort level after any model-specific downgrade.
• The field is present for tool-use contexts such as PreToolUse, PostToolUse, Stop, and SubagentStop, but absent for session-lifecycle hooks and models without effort support.

Hook schema and Bash environment wiring (search for "Reasoning effort applied to the current turn" and "Active effort level for the current turn").

Administrators can now control whether SDK-provided parent managed settings are dropped or merged under an admin tier.

{
  "parentSettingsBehavior": "merge"
}

• first-wins keeps the previous admin-first behavior.
• merge layers a restrictive-only slice from the parent settings under the admin winner.
• This matters for managed-policy deployments where an SDK parent wants to add restrictions without overriding the administrator’s policy.

New managed-settings schema and merge path (search for "parentSettingsBehavior" and "Controls whether the SDK parent tier").

Managed settings can now provide absolute paths for Linux/WSL sandbox helper binaries.

{
  "sandbox": {
    "bwrapPath": "/usr/bin/bwrap",
    "socatPath": "/usr/bin/socat"
  }
}

• sandbox.bwrapPath overrides bubblewrap auto-detection.
• sandbox.socatPath controls the sandbox network proxy helper.
• These paths are only honored from admin-controlled managed settings.
• Error messages now distinguish “bubblewrap missing” from “configured bwrap path is not executable”.

Managed sandbox path descriptions and validation errors (search for "Linux/WSL only: Absolute path to the bwrap" and "sandbox.bwrapPath is set to").

Claude Code now tries to spawn upgraded background daemons on Windows through WMI/CIM, falling back to direct spawn if WMI fails.

• The WMI path uses Win32_Process.Create with a hidden process startup configuration.
• If WMI fails or times out, Claude Code logs a fallback warning and starts directly.
• The fallback note says the daemon may not survive SSH or terminal close.

Windows daemon spawn path (search for "daemon: WMI spawn failed" and "Win32_Process.Create rc=").

The allowed-tools schema now explicitly deprecates passing Skill through the generic tools field and points authors to the dedicated skills field.

• This helps agent and skill authors avoid older tool-allowlist patterns.
• Existing allowed-tools inheritance behavior is unchanged.

Updated schema description (search for "Note: passing 'Skill' here is deprecated").

• MCP session-expiry errors are now recognized when a 404 response includes "mcp_session_terminated", allowing Claude Code to mark the server as needing auth instead of treating it as a generic failure. Evidence: MCP error classifier (search for "mcp_session_terminated").

• Compact-cache fallback now preserves abort behavior instead of continuing through fallback paths after cancellation. Evidence: abort checks added around compact cache sharing (search for "Compact cache sharing: no text in response, falling back").

• Windows command/path handling gained additional normalization and quoting safeguards for daemon spawning, including rejection of unsupported Unicode single-quote variants in generated PowerShell command lines. Evidence: PowerShell quoting guard (search for "unsupported Unicode single-quote in command line").