Version 2.1.47

Access your CLI session from the web (claude.ai/code) or the Claude app, continuing conversations across devices.

/remote-control

• Opens a secure connection to claude.ai
• Shows "Remote Control active" status in the CLI
• Reconnects automatically if connection drops
• Run /remote-control again to disconnect remote access
• Can be enabled for all sessions or just the current session

Remote Control UI and SSETransport infrastructure (search for "/remote-control", "Remote Control lets you access")

New --thinking flag provides semantic control over Claude's thinking/reasoning behavior, replacing the deprecated --max-thinking-tokens.

claude --thinking enabled   # Claude decides when and how much to think
claude --thinking disabled  # No extended thinking
claude --thinking 10000     # Fixed thinking token budget (older models)

• For Opus 4.6+, adaptive mode lets Claude autonomously decide thinking depth
• Takes precedence over the deprecated maxThinkingTokens setting
• The old --max-thinking-tokens flag is now deprecated with guidance to use --thinking instead

Thinking mode options (search for "--thinking <mode>", "Claude decides when and how much to think (Opus 4.6+)")

Quickly terminate all running background agents with a keyboard shortcut.

• Press ctrl+f once to open search
• Press ctrl+f again to kill all background agents

• New chat:killAgents action bound to double ctrl+f
• Shows "All background agents killed" confirmation
• Also available via text commands: "kill agents", "kill all agents"

Agent termination system (search for "All background agents killed", "chat:killAgents")

Pre-configured SSH connections for remote environments, typically set by enterprise administrators.

• New sshConfigs setting for managed settings
• Supports configuration of:
• SSH host (user@hostname or alias from ~/.ssh/config)
• SSH port (default: 22)
• Display name for the connection
• Path to SSH identity file (private key)
• Unique identifier for matching configs across settings sources

SSH config schema (search for "SSH connection configurations for remote environments")

Bash commands that may have destructive effects now show detailed warning notes explaining the potential impact.

• New warning notes for potentially dangerous commands:
• Note: may delete Kubernetes resources
• Note: may delete all rows from a database table
• Note: may destroy Terraform infrastructure
• Note: may discard all working tree changes
• Note: may discard uncommitted changes
• Note: may drop or truncate database objects
• Note: may force-delete a branch
• Note: may force-remove files
• Note: may overwrite remote history
• Note: may permanently delete untracked files
• Note: may permanently remove stashed changes
• Note: may recursively force-remove files
• Note: may rewrite the last commit
• Note: may skip safety hooks
• Blocked commands show "Permission for this action was denied by the dangerous action safety classifier"

Warning note strings (search for "Note: may delete", "dangerous action safety classifier")

Automatic detection and handling of security-flagged plugins from the official marketplace.

• Fetches security messages from https://raw.githubusercontent.com/anthropics/claude-plugins-official/refs/heads/security/security.json
• Auto-uninstalls delisted plugins when autoUninstallDelisted is enabled
• Shows "Plugins flagged. Check /plugins" notification
• New blocklist.json and flagged-plugins.json files for tracking
• "Removed from marketplace" indicator with optional reason

Plugin security infrastructure (search for "Fetching plugin security messages", "flagged-plugins.json")

Automatic detection and logging of asdf and mise version manager installations.

• Logs "Detected asdf installation:" when asdf is found
• Logs "Detected mise installation:" when mise is found
• Helps diagnose shell environment issues

Version manager detection (search for "Detected asdf installation")

Plugins can now load settings from a settings.json file in addition to manifest settings.

• Looks for settings.json in plugin directory
• Falls back to manifest settings if not found
• Settings can specify agent configuration to merge when plugin is enabled
• Logs which settings source was used

Plugin settings loader (search for "Loaded settings from settings.json for plugin")

Better messaging when fast mode encounters issues.

• "Fast mode overloaded and is temporarily unavailable" message
• Shows reset time: "resets in X"
• "· Fast mode ON" indicator

Fast mode messages (search for "Fast mode overloaded")

Added BMP format support for clipboard image operations on Linux.

• Now detects image/bmp in clipboard alongside png, jpeg, gif, webp
• Uses wl-paste --type image/bmp or xclip -selection clipboard -t image/bmp

BMP handling (search for "image/bmp")

Warp now supports Shift+Enter natively, removing previous workaround messaging.

• Updated tip: "Note: iTerm2, WezTerm, Ghostty, Kitty, and Warp support Shift+Enter natively."
• Removed previous warnings about Warp not supporting custom Shift+Enter keybindings
• Warp removed from the list of terminals requiring manual configuration

Terminal compatibility notes (search for "Warp support Shift+Enter natively")

New setting to control whether /rename updates the terminal tab title.

• Setting: "Whether /rename updates the terminal tab title (defaults to true)"
• Set to false to keep auto-generated topic titles
• Replaces the previous inverse setting behavior

Rename setting (search for "Whether /rename updates the terminal tab title")

Invalid permission rules in configuration files are now validated and skipped with warnings.

• Non-string values in permission arrays are removed with warning
• Invalid permission rules show specific error messages
• Logs which file and path contained the invalid rule

Permission validation (search for "Invalid permission rule", "was skipped")

Better management and cleanup of git worktrees created by agents.

• Detects when agent worktree has uncommitted changes: "Agent worktree has changes, keeping:"
• Logs worktree removal: "Removed agent worktree at:"
• Handles branch cleanup failures gracefully

Worktree management (search for "Agent worktree has changes", "worktreePath:")

Chrome browser automation MCP server can now run in-process for better performance.

• Logs "In-process Chrome MCP server started" when initialized
• Uses new SSE transport infrastructure
• Cleaner shutdown handling

Chrome MCP server (search for "In-process Chrome MCP server started")

Windows path conversion (Unix ↔ Windows) is now done in pure JavaScript, removing the dependency on cygpath.

• Handles /cygdrive/ prefix conversion
• Converts forward/backslashes appropriately
• Works with UNC paths (starting with // or \\)

Path conversion logic (search for "/cygdrive/")

• Fixed backslash-escaped whitespace warning for commands that could alter command parsing (search for "Command contains backslash-escaped whitespace")
• Fixed plugin hooks reload to only trigger on actual enabledPlugins changes, not all policySettings changes (search for "Plugin hooks: skipping reload, enabledPlugins unchanged")
• Fixed hook file reading errors to log properly: "Failed to read hook file" (search for "Failed to read hook file")
• Added validation for parent directory existence: "Parent directory does not exist or is not accessible:" (search for "Parent directory does not exist")

Deprecation: The --max-thinking-tokens flag is deprecated. Use --thinking <mode> instead for newer models. The flag still works for backward compatibility.

Removed Features: Delegate Mode infrastructure has been removed. References to TeammateTool, delegate mode, and team coordination are no longer present in the codebase.