Version 2.1.51

A new built-in slash command that creates a branch, commits staged changes, pushes to origin, and opens (or updates) a pull request — all in a single action.

# Inside a Claude Code session:
/commit-push-pr
/commit-push-pr add reviewer @teammate

• Automatically detects the base branch and creates a feature branch using your username prefix
• Drafts a commit message following your repository's commit style
• Creates or updates a PR with summary, test plan, and changelog sections
• Includes enhanced PR attribution: 🤖 Generated with Claude Code (X% N-shotted by claude-model-name, M memories recalled)
• If CLAUDE.md references Slack channels, optionally posts the PR URL via Slack MCP tools
• Auto-approves the git and GitHub CLI commands needed (no manual permission prompts)

Skill definition with name: "commit-push-pr" and description: "Commit, push, and open a PR" (search for "/commit-push-pr")

A new configuration file that specifies gitignored files to copy into git worktrees, solving the problem of missing config files or local secrets in worktree-based workflows.

# Create .worktreeinclude in your project root
# List gitignored files/patterns that should be copied to worktrees:
.env
.env.local
config/local.json

• Placed at the project root alongside .gitignore
• When a worktree is created, files matching .worktreeinclude patterns that are also gitignored are automatically copied from the main working directory into the worktree
• Blank lines and lines starting with # are treated as comments
• Uses git ls-files --others --ignored --exclude-from=.worktreeinclude to identify matching files

New function lc4 reads .worktreeinclude and copies files (search for ".worktreeinclude")

Claude Code can now read enterprise-managed configuration policies from platform-native MDM (Mobile Device Management) sources, enabling IT administrators to enforce organizational settings.

• macOS: Reads managed preferences from /Library/Managed Preferences/com.anthropic.claudecode.plist (device-level and per-user)
• Windows: Reads registry keys from HKLM\SOFTWARE\Policies\ClaudeCode and HKCU\SOFTWARE\Policies\ClaudeCode
• Falls back to file-based managed settings (managed-settings.json) on other platforms
• Polls for MDM settings changes every 30 minutes and applies updates live
• Settings are parsed through the same Zod schema validation as regular settings

Domain identifier "com.anthropic.claudecode", registry paths "HKLM\\SOFTWARE\\Policies\\ClaudeCode", function pE1() reads platform-specific MDM sources (search for "com.anthropic.claudecode")

The bash permission classifier now auto-approves significantly more read-only commands without prompting the user.

• Docker Compose: docker compose ps, docker compose logs, docker compose top, docker compose config
• Docker: docker logs, docker inspect (with safe flag validation)
• Ripgrep (rg): comprehensive flag support for search operations
• Pyright: type-checking with safe flags (blocks --watch mode)
• Package managers: npm list, pip list with safe flags
• Version checks: node -v, npm -v, python --version, python3 --version
• Python removed from the "suspicious command" blocklist — python and python3 are no longer flagged

New safe command definitions for "docker compose ps", "docker compose logs", rg, pyright, "npm list", "pip list" (search for "docker compose ps")

Commands containing shell brace expansion patterns (e.g., {a,b} or {1..5}) are now detected and flagged, preventing potential command injection through expansion tricks.

• Detects balanced brace pairs containing commas or range operators (..)
• Properly handles nested braces and escaped characters
• Blocks commands where brace expansion could alter command parsing

New function ve9() performs brace detection; error message "Command contains brace expansion that could alter command parsing" (search for "brace expansion")

The Remote Control feature (/remote-control) has been substantially rebuilt with a full bridge architecture for controlling CLI sessions from claude.ai/code. The feature flag was changed from tengu_worktree_mode to tengu_ccr_bridge, and the implementation now includes:

• Full work-polling loop with session spawn/teardown lifecycle
• WebSocket-based session ingress for real-time communication
• QR code display (togglable with spacebar) for quick mobile connection
• Automatic OAuth token refresh with JWT expiry tracking
• Environment registration, deregistration, and re-registration on failure
• Reconnection with exponential backoff and error budget
• Session activity tracking (tool starts, text output, results)
• remoteControlAtStartup setting to auto-enable Remote Control for all sessions
• allow_remote_sessions policy for organization-level control

Feature flag tengu_ccr_bridge, command definition name: "remote-control" with aliases ["rc"] (search for "tengu_ccr_bridge")

The memory system has been redesigned to use LLM-based retrieval instead of grep-based keyword search. Previously, memories were found by extracting search terms and grep-matching against markdown files. Now:

• Memories are listed with filenames and descriptions
• An LLM selects the most relevant memories (up to 5) based on the user's query
• Memories already read via the Read tool in the current conversation are filtered out to avoid duplication
• Memory files support a description frontmatter field for better retrieval

Selection prompt "You are selecting relevant memories for a user's query..." and function x8q for deduplication (search for "You are selecting relevant memories")

• Step-up scopes are now persisted across auth revocations, preventing users from having to re-authorize expanded permissions
• OAuth discovery state (authorization server URL, resource metadata) is cached to avoid repeated discovery requests
• New skipBrowserOpen option for headless environments — logs the authorization URL instead of opening a browser
• preserveStepUpState flag during MCP authentication prevents loss of elevated permissions

"Persisted step-up scope:" log message, "Saving discovery state" log message (search for "Persisted step-up scope")

• New autoReconnect option to disable automatic reconnection (useful for bridge sessions that manage their own lifecycle)
• onConnect callback fires when connection is established
• getStateLabel() exposes current connection state
• Close callbacks now receive the WebSocket close code for better error handling

autoReconnect parameter in constructor, new setOnConnect and getStateLabel methods (search for "autoReconnect")

When selecting a model in the CLI, if you have a pinned older model version (e.g., claude-sonnet-4-5), the selector now shows a helpful hint like "Newer version available · select Sonnet for claude-sonnet-4-6", guiding you to the latest version.

Function o_z returns description: "Newer version available · select ${alias} for ${currentVersionName}" (search for "Newer version available")

npm-sourced plugins now accept --registry and --version options, allowing installation from private registries or pinning to specific versions.

Plugin install function now passes { registry: A.registry, version: A.version } (search for "--registry")

StatusLine and FileSuggestion hook commands now verify workspace trust before executing. If the workspace trust dialog hasn't been accepted, these hooks are skipped with a warning, preventing untrusted workspaces from running arbitrary hook commands.

"Skipping StatusLine command execution - workspace trust not accepted" and "Skipping FileSuggestion command execution - workspace trust not accepted" (search for "workspace trust not accepted")

Pressing Escape in Vim NORMAL mode now properly resets the command state to idle, fixing cases where partial command sequences could leave the input in an unexpected state.

New condition if (M.escape && W.mode === "NORMAL") resets to { mode: "NORMAL", command: { type: "idle" } } (search for "NORMAL" in the Vim handler)

The structured transport now tracks resolved tool-use IDs and silently ignores duplicate control_response messages, preventing race conditions where a late-arriving permission response could cause errors.

"Ignoring duplicate control_response for already-resolved toolUseID=" (search for "Ignoring duplicate control_response")

• Fixed notification content rendering crash when content is not a string — now returns null gracefully instead of attempting to render non-string content (search for "typeof $ !== \"string\"")
• Removed the "Denied by auto mode classifier" banner that previously displayed when permissions were auto-denied, reducing UI noise
• Permission rule normalization now properly trims and normalizes stored rules when checking against new requests, preventing duplicate or stale rules (search for "r5(nX(j))")
• HTTP hooks now properly skip when conditions aren't met instead of potentially erroring (search for "Skipping HTTP hook")

An intelligent tool deferral system that tracks your tool usage patterns and only defers tools you haven't used recently, keeping frequently-used tools immediately available.

Feature-flagged behind tengu_coral_whistle (default: true)

• Tracks tool usage count and last-used timestamp in config (toolUsage)
• Uses a 7-day half-life decay function to compute a "recency-weighted usage score"
• Tools below the threshold are deferred (loaded on demand via ToolSearch)
• MCP tools are always deferred regardless of this flag
• When enabled, the shouldDefer flag on a tool triggers the usage-based check

tengu_coral_whistle feature flag, functions h84 (track usage), I84 (compute score with decay), b84 (check threshold) (search for "tengu_coral_whistle")

Several new slash command slots have been added as stubs (isEnabled: () => !1, isHidden: !0), indicating upcoming features. At least 15 new stub commands were added in this version, suggesting significant new functionality in the pipeline.

Multiple { isEnabled: () => !1, isHidden: !0, name: "stub" } definitions throughout the added code

A comprehensive built-in skill for helping users build LLM-powered applications with Claude, including code examples for Python, TypeScript, Java, Go, Ruby, C#, PHP, and cURL.

Added as part of the skills system, contains full documentation for:

• All current Claude models (Opus 4.6, Sonnet 4.6, Haiku 4.5)
• Adaptive thinking, effort parameter, compaction
• Tool use, streaming, batches, files API
• Agent SDK patterns
• Language-specific SDK examples including a new C# SDK section

Large string literal beginning with "# Building LLM-Powered Applications with Claude" (search for "Building LLM-Powered Applications with Claude")