Version 2.1.64

The claude remote-control command gains a server subcommand that runs as a persistent server accepting multiple concurrent sessions, with isolation and session management options.

# Each session gets its own git worktree (up to N concurrent)
claude remote-control server --spawn-worktree-sessions [<N>]

# All sessions share the current directory (up to N concurrent)
claude remote-control server --spawn-same-dir-sessions [<N>]

• --idle-timeout <ms> controls how long detached sessions remain alive (0 = never)
• --max-sessions <n> caps concurrent sessions (0 = unlimited)
• --auth-token <token> adds bearer token authentication; auto-generates tokens with sk-ant-cc- prefix if omitted
• --workspace <dir> sets the default working directory for sessions
• --host <string> and --port <number> control bind address
• Sessions can connect via unix domain sockets using the cc+unix:// protocol
• Session lifecycle is tracked in server-sessions.json
• The previous claude remote-control (without server) for single-session use still works

Server session manager (search for "--spawn-worktree-sessions", "server-sessions.json", "sk-ant-cc-")

A new planning mode that delegates plan creation to a remote session for more thorough, multi-step planning before implementation begins.

• When triggered, Claude enters "ultraplan" mode and generates a comprehensive plan via a dedicated remote session
• The completed plan is written to the plan file and presented to the user for approval
• Users then review the ultraplan and Claude responds with a summary before proceeding
• Gated as a prePlanMode option alongside the existing plan mode

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

Skills can now be scheduled to run on a cron-like schedule, executing automatically at specified intervals or times.

• Supports cron expression syntax with human-readable formatting (e.g., "Every hour", "Every day at 3:00 PM EST", "Weekdays at 9:00 AM EST")
• Schedule types include load (on startup), every (interval-based), and cron expressions
• The UI displays "1 scheduled skill" / "N scheduled skills" in the status area
• Users can view scheduled skill details and dismiss the notification

Scheduling engine (search for "scheduled skill", "Every hour", "Weekdays at")

A new slash command that hot-reloads plugin changes in the current session without requiring a restart.

/reload-plugins

• Activates pending plugin installations, configuration changes, and enable/disable toggling
• Throughout the plugin system, messages now say "Run /reload-plugins to activate" instead of the previous "Restart Claude Code to apply changes"
• Clears all plugin caches and reloads hooks, MCP servers, skills, commands, and agents

Plugin hot-reload (search for "/reload-plugins", "Activate pending plugin changes")

Terminal UI now supports mouse-based text selection with click-and-drag, and copying selected text to the clipboard.

• Click and drag to select text regions in the terminal output
• Selection is rendered with inverse colors to highlight the selected area
• Provides copySelection(), clearSelection(), and hasSelection() APIs
• Uses OSC 52 escape sequences to copy to the system clipboard

Selection engine (search for "copySelection", "clearTextSelection", "hasTextSelection")

A new hook event type that fires whenever an instruction file (CLAUDE.md or a rule file) is loaded during a session.

• Hook receives context about which instruction file was loaded
• Can be used for auditing, logging, or dynamically responding to instruction changes
• Follows the same async/sync hook response patterns as other hook events

Hook registration (search for "InstructionsLoaded", "When an instruction file (CLAUDE.md or rule) is loaded")

Introduces the concept of built-in plugins that ship with Claude Code and cannot be uninstalled or updated by users.

• Built-in plugins are identified by the @builtin source suffix
• They can be enabled or disabled via user settings but not removed
• Attempting to update or uninstall shows: "Built-in plugins cannot be updated or uninstalled."
• Built-in plugins can provide skills, hooks, and MCP servers just like regular plugins

Built-in plugin system (search for "Built-in plugin", "@builtin", "Built-in plugins cannot be updated or uninstalled")

New UI elements promote and integrate with Claude Code Desktop, a companion application with visual diffs, live app preview, and parallel sessions.

• "Try Claude Code Desktop" dialog appears as a suggestion
• "Open in Claude Code Desktop" option available in menus
• "Continue your session in Claude Code Desktop" enables session handoff
• Described as: "Same Claude Code with visual diffs, live app preview, parallel sessions, and more."

Desktop integration (search for "Claude Code Desktop", "Open in Claude Code Desktop")

A new git-subdir plugin source type that allows installing plugins from a subdirectory within a larger git repository (monorepo).

• Uses git sparse-checkout with --filter=tree:0 to minimize bandwidth
• Requires git version 2.25 or later for cone mode support
• Supports pinning to a specific branch or SHA
• Falls back to unshallow fetch if shallow SHA fetch fails
• Accepts GitHub owner/repo shorthand, HTTPS, or SSH URLs

Monorepo plugin source (search for "git-subdir", "sparse-checkout", "--filter=tree:0")

A new hook configuration option that allows hooks to run in the background and wake the model when they encounter a blocking error (exit code 2).

• When asyncRewake: true, the hook runs asynchronously like async: true
• If the hook exits with code 2, it signals a blocking error that wakes the model
• The blocking error message from the hook command is surfaced to the user
• New "Stop hook blocking error" notification type when this occurs

Async rewake hooks (search for "asyncRewake", "If true, hook runs in background and wakes the model on exit code 2")

A new boolean setting that controls whether Claude's system prompt includes built-in commit and PR workflow instructions.

{
  "commitWorkflowInstructions": false
}

• Default: true (instructions are included)
• When disabled, Claude will not receive the built-in guidance for git commit messages and PR creation
• Useful for teams that have their own commit/PR conventions via CLAUDE.md

Setting definition (search for "Include built-in commit and PR workflow instructions")

A new boolean setting that controls whether thinking summaries are displayed in the transcript view.

• Default: false
• When enabled, thinking process summaries appear in the transcript view (ctrl+o)
• Also gated behind the tengu_quiet_hollow feature flag

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

New security checks detect and flag potentially obfuscated bash commands that could bypass permission checks.

• Detects commands containing quoted newlines followed by #-prefixed lines, which can hide arguments from line-based permission scanning
• Detects consecutive quote characters at word start (potential obfuscation)
• Detects empty quote pairs adjacent to quoted dashes (flag obfuscation)
• Sanitizes redirects to Windows NUL device by replacing with /dev/null
• Flagged commands trigger an "ask" behavior requiring explicit user approval

Security checks (search for "potential obfuscation", "quoted newline", "$1/dev/null")

The model promotion banner now reads "Model updated to Opus 4.6" (previously "Opus 4.5"), reflecting the new default model.

Model banner (search for "Opus 4.6")

The previous three-tier brevity system (tengu_swann_brevity with strict/focused/polished levels) has been replaced by a single tengu_sotto_voce feature flag that, when enabled, injects a streamlined output efficiency instruction. This simplifies server-side control of model verbosity.

Brevity flag (search for "tengu_sotto_voce")

Plugin commands, skills, and agents are now loaded using Promise.all for concurrent processing, improving startup time when multiple plugins are installed. Previously, plugins were loaded sequentially in for loops.

Parallel loading (search for "Promise.all" in the plugin loading functions)

The hook response system now uses Promise.allSettled to check all pending async hooks concurrently rather than sequentially. Failed hook callbacks are caught and logged individually without blocking other hooks.

Hook processing (search for "checkForAsyncHookResponses callback rejected")

When a parallel tool call fails, sibling tool calls now receive a clearer cancellation message: "Cancelled: parallel tool call \<name\> errored" — replacing the previous generic "Sibling tool call errored" message. This helps identify which specific tool caused the failure.

Error message (search for "Cancelled: parallel tool call")

Hook event payloads now include agent_type and agent_id fields, providing context about whether the hook fired from a subagent and which agent type triggered it. This allows hook scripts to behave differently based on agent context.

Hook schema (search for "Agent type name", "Subagent identifier")

The TaskCompleted and TeammateIdle hook events can now prevent the session from continuing when they return a blocking response, surfacing messages like "TaskCompleted hook prevented continuation".

Hook blocking (search for "TaskCompleted hook prevented continuation", "TeammateIdle hook prevented continuation")

The ToolSearch tool now supports comma-separated tool names for loading multiple tools in a single call (e.g., select:Read,Edit,Grep), reducing the number of round-trips needed to load tools.

ToolSearch docs (search for "select:Read,Edit,Grep")

Binary content from MCP tool results is now saved to disk with proper MIME type detection and file extension mapping, rather than being displayed inline. Supports a comprehensive set of MIME types including Office documents, audio, video, and archives.

Binary content handling (search for "Binary content (", "application/vnd.openxmlformats")

Enterprise administrators can now configure a custom pluginTrustMessage in policy settings (managed-settings.json / MDM) that is appended to the plugin trust warning shown before installation.

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

The plugin error view has been completely rewritten with actionable guidance for each error type. Errors now show the source scope (user/project/local/managed), provide specific remediation actions (navigate to settings, remove marketplace, contact admin), and allow direct resolution from the error list.

Error UI (search for "Managed by your organization — contact your admin", "Restart to retry loading plugins")

A new pathPattern field in strictKnownMarketplaces settings allows regex-based matching against file and directory plugin sources, complementing the existing hostPattern for network sources.

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

A new notification banner "Voice mode is now available · /voice to enable" appears when voice mode is detected as available, improving discoverability of the feature.

Voice notification (search for "Voice mode is now available")

The .worktreeinclude file processing for worktree sessions now uses ignore pattern matching (similar to .gitignore) instead of simple file list comparison, supporting glob patterns and directory matching for more flexible file copying into worktrees.

Worktree include (search for ".worktreeinclude")

When fast mode status cannot be determined due to network connectivity issues, it now defaults to disabled (network_error) with the message "Fast mode unavailable due to network connectivity issues" instead of silently failing.

Fast mode fallback (search for "Fast mode unavailable due to network connectivity issues")

• Fixed stale lock file handling: lock files can now be force-removed when they become stale, with logging of the cleanup (search for "Force-removed lock file at")
• LSP diagnostics state is now properly reset when clearing state, preventing stale diagnostic data from persisting (search for "LSP Diagnostics: Resetting all state")
• MCP config validation now produces clearer error messages: "--mcp-config validation failed" and "MCP config is not valid JSON" replace generic parse errors (search for "--mcp-config validation failed")
• Symlink-aware path resolution added for filesystem operations, properly following symlinks through intermediate path components (search for "isSymbolicLink" in function Oiq)

Tool names in the deferred tool list can display alongside a short description hint (e.g., "Read — read files, images, PDFs, notebooks"), making it easier to identify the right tool.

Feature-flagged behind tengu_tst_hint_m7r and overridable via CLAUDE_CODE_SEARCH_HINTS_IN_LIST env var.

• Every built-in tool now has a searchHint property with a short plain-English description
• When enabled, the tool list displays "ToolName — description" format
• Hints include descriptions like "execute shell commands", "search file contents with regex (ripgrep)", "delegate work to a subagent"
• Environment variable takes precedence over feature flag (set to 1 to enable, 0 to disable)

Search hints (gated by tengu_tst_hint_m7r, search for "searchHint", "CLAUDE_CODE_SEARCH_HINTS_IN_LIST")