Claude Code v2.1.51 / Changelog

Pick an entry · j / k steps through

14 entries

Changesopen

`/commit-push-pr` Command #

What

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.

Usage

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

Details

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)

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

`.worktreeinclude` File Support #

What

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.

Usage

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

Details

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Enterprise MDM Settings Support #

What

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

Details

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Expanded Safe Command Auto-Approval #

What

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

Details

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Brace Expansion Security Check #

What

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.

Details

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Remote Control Bridge Overhaul #

Feature flag

tengu_worktree_mode On for this account, and not off by default

The flag server returned on for the one account this site reads, and nothing in this release compiles it off by default. The compiled default is shown below, and says which it is when we cannot read one: a fifth of gates compile in a string or a number rather than on or off, and most published releases have no gate table behind them at all. No client can see what the server returns for your account.

This account: on · anonymous baseline: on · compiled default in v2.1.51: no gate table built for this version

These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.

tengu_ccr_bridge On for this account, and not off by default

This account: on · anonymous baseline: on · compiled default in v2.1.51: no gate table built for this version

These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.

Read once, for one account on one subscription tier, against v2.1.51. It isn't a statement about your account. What a flag value here can and cannot tell you

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

LLM-Based Memory Retrieval #

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

v2.1.53 Improved Memory Selection Prompt

MCP OAuth Improvements #

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

WebSocket Transport Enhancements #

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Model Selector Version Hints #

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Plugin Registry and Version Support #

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Workspace Trust Enforcement for Hooks #

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

v2.1.79 Statusline Trust Check

Vim Mode ESC Fix #

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Duplicate Permission Response Handling #

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

4 entries

Bug Fixesopen

#

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")

Smart Tool Deferral [Gradual Rollout]#

What

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.

Status

Feature-flagged behind tengu_coral_whistle (default: true)

Details

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

v2.1.61 Built-in tools no longer deferred by default

Multiple Stubbed Commands [In Development]#

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Built-in Claude API / Agent SDK Skill [In Development]#

What

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.

Status

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

v2.1.97 Managed Agents API Documentation

Claude Code v2.1.51

Changesopen

`/commit-push-pr` Command #

`.worktreeinclude` File Support #

Enterprise MDM Settings Support #

Expanded Safe Command Auto-Approval #

Brace Expansion Security Check #

Remote Control Bridge Overhaul #

LLM-Based Memory Retrieval #

Related

MCP OAuth Improvements #

WebSocket Transport Enhancements #

Model Selector Version Hints #

Plugin Registry and Version Support #

Workspace Trust Enforcement for Hooks #

Related

Vim Mode ESC Fix #

Duplicate Permission Response Handling #

Bug Fixesopen

#

Smart Tool Deferral [Gradual Rollout]#

Related

Multiple Stubbed Commands [In Development]#

Built-in Claude API / Agent SDK Skill [In Development]#

Related

Anthropic’s official release notes