Version 2.1.69

A new server subcommand for claude remote-control that runs as a persistent server accepting multiple concurrent sessions from claude.ai/code.

# Start a server where all sessions share the current directory
claude remote-control server

# Each session gets an isolated git worktree (max N sessions)
claude remote-control server --spawn-worktree-sessions [<N>]

# All sessions share the current directory (default behavior, max N sessions)
claude remote-control server --spawn-same-dir-sessions [<N>]

# Additional server options
claude remote-control server --idle-timeout <ms>    # Timeout for detached sessions
claude remote-control server --max-sessions <n>     # Max concurrent sessions
claude remote-control server --workspace <dir>      # Default working directory
claude remote-control server --auth-token <token>   # Bearer token for auth
claude remote-control server --host <string>        # Bind address
claude remote-control server --port <number>        # Listen port

• Runs as a persistent daemon that accepts connections without pre-creating an initial session
• Supports cc+unix:// protocol for Unix domain socket connections
• Session management with server-sessions.json persistence and server.lock for single-instance enforcement
• Configurable spawn modes: shared directory (default) or isolated git worktrees
• Sessions can be created, attached, detached, and destroyed independently
• Policy-controllable via allow_remote_control organization policy

Remote Control server mode (search for "Start a Claude Code session server", "--spawn-worktree-sessions", "server-sessions.json")

A new remote-session-based planning mode that generates comprehensive implementation plans using a dedicated remote session.

• Triggered when planning complex implementations
• Runs a remote session that produces a structured <ultraplan> block
• Displays "Ultraplanning…" progress indicator during generation
• The generated plan is written to a plan file for user review
• After review, the user can approve and proceed to implementation

Ultraplan feature (search for "Ultraplanning…", "Ultraplan complete", "ultraplan")

Pre-fill the prompt input with text without automatically submitting it, allowing the user to review and edit before sending.

claude --prefill "implement a function that"

• Text appears in the input field ready for editing
• Unlike --print, does not auto-submit — the user can modify the prompt before pressing Enter
• Useful for templates, repetitive prompts, or workflow scripting

Prefill flag (search for "--prefill", "Pre-fill the prompt input with text without submitting it")

Reload and activate plugin changes in the current session without needing to restart Claude Code.

/reload-plugins

• Activates pending plugin installations, removals, and configuration changes
• Replaces the previous behavior of requiring a full restart after plugin changes
• Messages like "Run /reload-plugins to activate" now appear after plugin changes instead of "Restart Claude Code"

Plugin reload (search for "/reload-plugins", "Activate pending plugin changes in the current session")

Native mouse text selection with drag-to-select and clipboard copy support in the terminal UI.

• Click and drag to select text in the terminal output
• Selected text can be copied to the system clipboard via OSC 52 escape sequence
• noSelect regions prevent selection of UI chrome (status bars, borders) while allowing content selection
• Selection state tracks anchor/focus positions with full drag handling
• from-left-edge option extends no-select regions from the left edge of the terminal

Text selection system (search for "isDragging", "copySelection", "noSelect")

Terminal UI components can now use overflowY: "scroll" to create scrollable content areas with sticky-scroll behavior.

• New scroll overflow mode alongside existing hidden and visible
• stickyScroll attribute auto-scrolls to bottom as new content arrives
• Tracks scrollHeight, scrollViewportHeight, and scrollTop for programmatic scroll control
• Content outside the visible viewport is culled for rendering performance

Scroll support (search for "stickyScroll", "scrollViewportHeight", overflowY)

Claude Code can now automatically detect the system light/dark theme setting and follow it.

• New "Auto (follow system)" theme option
• On macOS, reads AppleInterfaceStyle from system defaults to detect Dark Mode
• Falls back to "dark" on non-macOS platforms
• Theme updates when toggling between light and dark mode

Auto theme (search for "Auto (follow system)", "AppleInterfaceStyle")

Complete restructuring of the persistent memory system with typed categories, frontmatter format, and two-step save process.

• Four memory types: user (preferences/role), feedback (corrections/guidance), project (ongoing work context), and reference (external system pointers)
• Memories use markdown files with YAML frontmatter (name, description, type fields)
• Two-step save: write the memory file, then add a pointer in the memory index
• Clear guidance on what NOT to save (code patterns, git history, ephemeral details)
• When users correct information stated from memory, the stored entry must be updated
• CLAUDE_COWORK_MEMORY_PATH_OVERRIDE environment variable to customize the memory directory

Memory system (search for "## Types of memory", "## What NOT to save in memory", "CLAUDE_COWORK_MEMORY_PATH_OVERRIDE")

Infrastructure for creating, validating, packing, and unpacking MCP Bundle (MCPB) files — a packaging format for MCP server plugins.

mcpb pack       # Create a .mcpb bundle from a directory
mcpb validate   # Validate a manifest.json

• Interactive manifest.json creation wizard with prompts for name, version, tools, prompts, screenshots, and more
• .mcpbignore support to exclude files from bundles
• Schema validation for manifest files
• Bundle unpacking and verification
• Supports Node.js and Python version constraints
• Platform compatibility settings (macOS, Linux, Windows)

MCPB support (search for "A MCPB bundle", "mcpb pack", ".mcpbignore")

Plugins can now be sourced from a subdirectory within a larger repository (monorepo support).

• Uses sparse checkout (--filter=tree:0) to minimize bandwidth for monorepo plugins
• Only the specified subdirectory is materialized; the rest of the repo is not downloaded
• Requires git version 2.25 or later for cone mode support
• Falls back to unshallow fetch if sparse-checkout fails

git-subdir support (search for "git-subdir", "Subdirectory within the repo containing the plugin")

Improved detection and handling of binary file content, with disk-based storage for large binary results.

• Sophisticated binary detection using byte analysis (null bytes, control characters)
• Extensive list of known binary file extensions (images, audio, video, archives, executables, etc.)
• Binary content from tool results can be saved to disk and referenced by path
• Support for Office documents, archives, and other binary MIME types

Binary content (search for "Binary content (", "Binary content could not be saved to disk")

Three new hook event types for more granular automation control.

• InstructionsLoaded: Fires when an instruction file (CLAUDE.md or rule) is loaded
• TaskCompleted: Can prevent continuation after task completion ("TaskCompleted hook prevented continuation")
• TeammateIdle: Can prevent continuation when a teammate goes idle ("TeammateIdle hook prevented continuation")
• Hooks support agent_type and agent_id context fields for subagent-aware automation
• Background hooks with async mode that wake the model on exit code 2

New hooks (search for "InstructionsLoaded", "TaskCompleted hook prevented continuation", "TeammateIdle hook prevented continuation")

New settings API that returns effective merged settings plus raw per-source settings in merge order.

• Shows settings from all sources (User, Project, Local, Managed) with merge ordering
• Labeled "Ordered low-to-high priority — later entries override earlier ones"
• Makes it clear where each setting value comes from

Settings debug (search for "Effective merged settings plus raw per-source settings in merge order")

ToolSearch now supports loading multiple deferred tools at once with comma-separated select: syntax (e.g., "select:Read,Edit,Grep"). Tools can also display search hints alongside their names for better discoverability. A new partial select mode handles cases where some tools in a multi-select query are found while others are missing.

ToolSearch improvements (search for "select:Read,Edit,Grep", "searchHint", "partial select")

Deferred tools are now announced through system-reminder messages in the conversation rather than a static list, controlled by the tengu_glacier_2xr feature flag. This makes tool availability more dynamic and context-aware.

Deferred tools announcement (search for "Deferred tools are announced via system-reminder messages")

New obfuscation detection for potentially dangerous shell commands:

• Quoted newlines followed by #-prefixed lines that could hide arguments
• Consecutive quote characters at word start (potential obfuscation)
• Empty quote pairs adjacent to quoted dashes (flag obfuscation)
• Quoted brace characters inside brace contexts (brace expansion obfuscation)
• Excess closing braces after quote stripping

Safety checks (search for "Command contains a quoted newline followed by a #-prefixed line")

New sandbox option for macOS that allows access to com.apple.trustd.agent within the sandbox. This is needed for Go-based CLI tools (gh, gcloud, terraform, etc.) to verify TLS certificates when using httpProxyPort with a MITM proxy and custom CA.

Network isolation setting (search for "enableWeakerNetworkIsolation", "Reduces security")

New showThinkingSummaries setting to display thinking summaries when viewing the transcript (ctrl+o). Default: false.

Thinking summaries (search for "Show thinking summaries in the transcript view")

Enterprise administrators can now configure a custom message appended to the plugin trust warning shown before installation, via pluginTrustMessage in managed settings or MDM policy. Useful for adding organization-specific guidance.

Plugin trust message (search for "pluginTrustMessage", "Custom message to append to the plugin trust warning")

New setting to control whether built-in commit and PR workflow instructions are included in Claude's system prompt. Default: true.

Cowork instructions setting (search for "Include built-in commit and PR workflow instructions")

Marketplace sources of type strictKnownMarketplaces can now use pathPattern — a regex matched against .path of file and directory sources — to allow filesystem-based marketplaces alongside network source restrictions.

Path pattern restrictions (search for "pathPattern", "Regex pattern matched against the .path field")

Subagent prompts have been refined:

• Clearer instruction formatting with separate lines for avoiding emojis and colons before tool calls
• Different response expectations based on agent role: detailed writeup for some, concise reports for others
• Agents now use absolute paths in their final responses
• Verification agent type added for post-task validation
• Teammates explicitly cannot spawn other teammates (flat team roster)

Subagent improvements (search for "Teammates cannot spawn other teammates", "NOTE: You just closed out 3+ tasks")

The conversation summary system has been significantly revised:

• New <<ANALYSIS_INSTRUCTION>> placeholder system for injecting analysis instructions
• Support for partial compaction of only recent messages when earlier context is retained
• Lean compaction mode (tengu_lean_cast) with shorter analysis prompts focused on coverage over detail
• Better handling of large conversation files via optimized compact_boundary detection for files over 64MB

Summary improvements (search for "<<ANALYSIS_INSTRUCTION>>", "tengu_lean_cast")

Fast mode status messages now distinguish between general unavailability and network-specific issues:

• "Fast mode is currently unavailable" for general outages
• "Fast mode unavailable due to network connectivity issues" for network errors
• Default status changed from "disabled" to more descriptive states

Fast mode messaging (search for "Fast mode is currently unavailable", "Fast mode unavailable due to network connectivity issues")

A full Inquirer.js-style prompt library has been embedded for interactive CLI prompts, with proper error handling (AbortPromptError, CancelPromptError, ExitPromptError), pagination, arrow-key navigation, and multi-select support. This replaces ad-hoc prompt implementations.

Inquirer integration (search for "AbortPromptError", "[Inquirer] Hook functions can only be called from within a prompt")

Plugins managed by organization policy now display "Managed by your organization — contact your admin" instead of generic status messages.

Managed plugins (search for "Managed by your organization")

• Fixed sandbox cleanup: cleanupAfterCommand is now exposed on the sandbox API for proper resource cleanup after command execution (search for "cleanupAfterCommand")

• Fixed lock file handling: Force-removal of stale lock files now logs the action and handles failures gracefully (search for "Force-removed lock file at", "Failed to force-remove lock file")

• Fixed symlink resolution in path operations: New Fsq function properly resolves symlinks by walking up the directory tree and handling broken symlinks gracefully (search for "realpathSync" in added functions)

• Removed microcompact feature: The Context microcompacted flow has been removed, simplifying the compaction pipeline (search for "microcompact" — present in old, absent in new context references)

• Removed coral_reef_opus2 client data check and crystal_beam budget token feature — these were internal feature gates that have been cleaned up

The Remote Control server subcommand is a significant architectural addition. Previous versions only supported basic remote-control connections — the new server mode enables persistent multi-session hosting with session lifecycle management, making it suitable for team environments and CI/CD integrations. Both --spawn-same-dir-sessions (default) and --spawn-worktree-sessions require the server subcommand; attempting to use them with the base remote-control command will produce an error.

The memory system overhaul changes the recommended format for memory files. While existing plain-text memories will likely continue to work, adopting the new frontmatter format (name, description, type fields) is recommended for better relevance matching in future conversations.