Claude Code 2.1.0 Major Update - From Skill Hot-Reload to Enhanced Vim Motions

Key New Features

1. Skill Hot-Reload

1. Create ~/.claude/skills/my-skill/ directory
2. Write and save SKILL.md file
3. Immediately usable in Claude Code (no restart required)
4. Changes are reflected in real-time

• Skill loading fails if YAML frontmatter has errors
• Filename must be exactly SKILL.md (case-sensitive)
• First line must start with --- (no blank lines before it)

2. Forked Sub-Agent Context

---
name: code-analysis
description: Analyze code quality in isolated context
context: fork
agent: Explore
---

• Explore: Codebase exploration and analysis
• Plan: Planning and architecture design
• general-purpose: General-purpose tasks (default)
• Custom: Agents defined in .claude/agents/

• Execute complex multi-step tasks without mixing into main conversation
• Perform specialized analysis in a separate context
• Maintain independent conversation history per sub-agent

3. Language Configuration

{
  "language": "Korean"
}

• Option+Arrow word navigation treats CJK text as word boundaries rather than entire sequences
• Removed extra blank lines in multiline prompts containing CJK characters
• IME (Input Method Editor) support improved: composition window displays accurately at cursor position
• Word deletion (opt+delete) and word navigation work correctly with non-Latin text including Cyrillic, Greek, Arabic, Hebrew, Thai, and Chinese

4. Enhanced Vim Motions

5. Unified Backgrounding

1. Request background execution to Claude Code or press Ctrl+B
2. Background tasks run in separate shells with unique IDs
3. You can monitor output, check status, and terminate tasks
4. Tmux users need to press Ctrl+B twice (due to tmux prefix key)

• Continue development while monitoring servers, builds, and other processes
• Send long-running tests to background and work on other features
• Run agents in background while performing other tasks

• Clean completion messages shown when background tasks finish (noisy output removed)
• Automatic truncation to 30K characters when producing large output to prevent API context overflow

6. MCP list_changed Notifications

• notifications/tools/list_changed: Tool list change notification
• notifications/prompts/list_changed: Prompt list change notification
• notifications/resources/list_changed: Resource list change notification

1. MCP server sends tool change notification
2. Claude Code automatically refreshes tool list from that server
3. Changes are immediately reflected in /mcp menu

• Claude Code didn't respond even when notifications were sent
• Dynamic prompt updates didn't appear in UI, requiring restart
• Custom MCP commands didn't appear when added to .claude/commands/

7. /plan Command Shortcut

8. Slash Command Improvements

claude --tools Read,Grep,Bash

9. Significantly Enhanced Hooks

• once: true setting to specify Hooks that run only once per session
• YAML-style list support in allowed-tools field in frontmatter
• Support for prompt and agent hook types in plugins (previously only command hooks were supported)
• PreToolUse, PostToolUse, Stop Hook support in agent frontmatter
• Hook support in Skill and slash command frontmatter

---
name: secure-operations
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/security-check.sh"
---

hooks:
  - type: command
    command: "echo 'Session initialized'"
    once: true

• Notifications: Custom alerts when Claude Code is waiting for input
• Auto-formatting: Run prettier, gofmt, etc. after file edits
• Logging: Track executed commands for compliance/debugging
• Feedback: Provide automatic feedback on code rules
• Custom permissions: Block modifications to production files or sensitive directories

10. Agent Control

{
  "permissions": {
    "deny": ["Task(Plan)", "Task(Explore)"]
  }
}

11. Bash Wildcard Patterns

Bash(npm *)      # All commands starting with npm
Bash(* install)  # All commands ending with install
Bash(git * main) # Any command between git and main

{
  "permissions": {
    "allow": [
      "Bash(npm:*)",
      "Bash(git status)",
      "Bash(git diff)",
      "Bash(git add *)",
      "Bash(git commit *)"
    ],
    "ask": ["Bash(npm install *)"]
  }
}

1. Deny rules: Completely block tool usage
2. Ask rules: Take precedence over Allow rules
3. Allow rules: Automatically permit

• Start restrictive and expand permissions as needed
• Configure project-specific tools in .claude/settings.local.json
• Configure common tools (npm, git, python, etc.) in ~/.claude/settings.json

12. Remote Session Support

• Seamless movement between browser version and CLI tool
• Chat history and context are copied together
• --teleport <session-id> flag available in CLI

1. Start work on the web
2. Move task to background
3. Teleport session in CLI to continue work

• Configure remote session environment
• Hooks can be configured in repository's .claude/settings.json
• By default, all Hooks run in both local and remote environments
• Check CLAUDE_CODE_REMOTE environment variable to run only in specific environments

• Remote environments run behind HTTP/HTTPS network proxy for security
• Network access is limited to allowlisted domains by default

• Session teleport is currently one-directional only: Claude Code on the web → local CLI
• No way to continue local sessions to remote machines (e.g., AWS instances)

Terminal Improvements

Shift+Enter Support

Keyboard Fixes

• Alt+B and Alt+F (word movement) now work correctly in iTerm2, Ghostty, Kitty, and WezTerm
• Terminal keyboard mode resets correctly on exit in Ghostty, iTerm2, Kitty, and WezTerm
• Removed unnecessary blank lines in multiline prompts containing CJK characters (Korean, Japanese, Chinese)

Image Pasting

Security Fixes

Performance Improvements

Startup Performance

Terminal Rendering

Jupyter Notebooks

Stability Improvements

• Improved pipe input stability (cat refactor.md | claude)
• Improved AskQuestion tool stability
• Improved compaction stability
• sed in-place edit commands now render as file edits with diff preview
• Automatically continues when response is truncated by output token limit (instead of error message)
• Sub-agents (Task tool) continue working after permission denial, enabling alternative approach attempts

Major Bug Fixes

Session Management

• Files and Skills are correctly discovered when resuming sessions with -c or --resume
• Fixed pasted content loss when replaying prompts from history with up arrow or Ctrl+R search
• OAuth token refresh triggers even when server reports and local expiration checks don't match
• Fixed session persistence race conditions and 409 conflicts

File Operations

• Write tool now respects system umask instead of hardcoded 0o600 permissions
• Skip symlink resolution when reading FIFO files to prevent hangs

Command Execution

• $() command substitution executes without parsing errors
• Multiline bash commands with backslash continuation are processed correctly
• Subcommands after global options are correctly identified (git -C /path log, etc.)

Image Processing

• Large pasted images are processed without "Image was too large" error

AWS/Cloud

• AWS Bedrock sub-agents inherit EU/APAC cross-region inference model configuration, resolving 403 errors
• Large output from background tasks is truncated to 30K characters with file path reference provided

Other

• Fixed background tasks failing with 'git repository not found' error in repositories with dots in names
• Claude Code support for Chrome in WSL environments
• Fixed ${CLAUDE_PLUGIN_ROOT} substitution in plugin allowed-tools frontmatter

SDK and Extension Changes

SDK

• Minimum zod peer dependency changed to ^4.0.0 (Breaking change)

VSCode Extension

• Current selected model name displayed in context menu
• Added descriptive labels to auto-accept permission buttons (e.g., "Yes, allow npm for this project")
• Paragraph breaks render correctly in markdown content
• Fixed parent iframe scrolling when scrolling in extension

Windows

• Fixed improper rendering issues
• Fixed native installer silently failing when executable creation fails

UI/UX Improvements

• Progress display shown during Skill execution (shown whenever tool use occurs)
• Skills in /skills/ directory are shown by default in slash command menu (user-invocable: false to exclude)
• Skill suggestions prioritize recent and frequently used Skills
• Improved spinner feedback when waiting for first response token
• Background agent tokens included in spinner token count
• Improved permission prompt UX (Tab hint moved to footer, clean Yes/No labels)
• Removed permission prompt when entering plan mode
• "Interrupted" message color changed to be less alarming
• Skills displayed as separate category in context visualization

New Environment Variables and Settings

Breaking Changes and Removals

• Minimum zod peer dependency changed to ^4.0.0 (SDK)
• Removed permission prompt when entering plan mode
• Removed # shortcut for quick memory input (use /edit CLAUDE.md instead)

Impact and Expected Effects

How to Update

claude update

# Or re-run the installation script
curl -fsSL https://claude.ai/install.sh | sh

References