Version 1.0.114

What: Fine-grained filesystem access controls using macOS sandbox-exec, allowing you to restrict which files and directories bash commands can read or write.

How to use: Configure in .claude/settings.json:

{
  "sandbox": {
    "filesystem": {
      "read": {
        "allow": ["~/projects/*", "~/data"],
        "deny": ["/etc/passwd"]
      },
      "write": {
        "allow": ["~/projects/output/*"],
        "deny": ["~/projects/output/protected"]
      }
    }
  }
}

Details:

• Complements existing network sandboxing (v1.0.113 only had network restrictions)
• Read allow-list mode denies all reads except specified paths
• Write deny-list blocks specific paths while allowing essential system paths
• Automatically allows stdout/stderr, temp directories, and Claude config paths
• Supports path expansion (~, ./, ../) and wildcards (/* suffix only)
• Path validation prevents shell metacharacters (&&, ||, ;, backticks, $)
• Evidence: QBB() at line 374281, y41() at line 353864

What: Real-time detection and visual highlighting of thinking trigger words in 7 languages (English, Japanese, Chinese, Spanish, French, German, Korean) with three intensity levels.

How to use: Type trigger words in your prompt to activate extended thinking:

# Basic (4,000 tokens): "think", "考えて", "想"
claude "think about the best algorithm for this"

# Medium (10,000 tokens): "think deeply", "megathink", "もっと考えて"
claude "think deeply about the edge cases"

# High (31,999 tokens): "think harder", "ultrathink", "熟考"
claude "think harder about the security implications"

Details:

• Trigger words are highlighted with shimmer effects as you type
• Status bar shows: "Thinking on • [level] • /t to disable"
• Use /t command to toggle thinking mode on/off
• Intelligently ignores conversational phrases like "I think" and "we think"
• Color-coded by intensity: basic (low), medium (medium), high (max)
• Evidence: TTA() at line 366061, PTA() at line 366112, fA0() at line 366083, o1Q() at line 436378

What: Improved clear screen handling for Windows Terminal and VSCode integrated terminal on Windows.

Details:

• Detects Windows Terminal via WT_SESSION environment variable
• Detects VSCode terminal via TERM_PROGRAM environment variable
• Modern Windows terminals now use full clear sequence (\x1B[2J\x1B[3J\x1B[H) that clears scrollback buffer
• Legacy Windows consoles continue using compatible sequence (\x1B[2J\x1B[0f)
• Provides cleaner screen clearing experience in modern Windows environments
• Evidence: JN9() at line 362129, XN9() at line 362132, FN9() at line 362142

What: Automatic truncation of bash command output exceeding 64MB to prevent memory issues.

Details:

• Default limit: 64MB (67,108,736 bytes)
• Truncated output shows: ... [output truncated - ${KB}KB removed]
• Tracks total bytes received even when truncated
• Prevents crashes from commands generating massive output
• Evidence: class ga at line 362733

What: Restore multiple files from file history backups in a single operation.

Details:

• Restores files to their pre-session state from backups
• Handles three scenarios: restore from backup, delete new files, skip missing files
• Supports dry-run mode to preview changes
• Returns count of files affected
• Continues on errors with detailed logging
• Evidence: _3B() at line 393833, k3B() at line 393866

What: Intelligent editor selection that verifies editors are actually installed before attempting to use them.

How it works:

# Priority order: $VISUAL → $EDITOR → code → vi → nano
# Only selects editors that exist on the system
claude /memory

Details:

• Uses which (Unix) or where (Windows) to check editor availability
• Tries VS Code (code), then vi, then nano in order
• Trims whitespace from environment variables to prevent issues
• Throws clear error if no editor is available
• Improves reliability when opening memory files
• Evidence: If1() at line 404558, V95() at line 404550

What: Specialized system message types for hook execution outcomes instead of generic informational messages.

Details:

• New message type: post_tool_hook_success for successful hook completion
• New message type: post_tool_hook_cancelled for user-cancelled hooks
• Clearer semantic representation in message history
• Better structured for UI rendering and debugging
• Note: PostToolUse hooks existed in v1.0.113, only the message types are new
• Evidence: PIB() at line 398052, jq0() at line 398064

What: More informative interrupt messages that guide users on what to do next.

Details:

• Changed from "Interrupted by user" to "Interrupted · What should Claude do instead?"
• Clearer call-to-action for users after interrupting
• Better UX for mid-task cancellations
• Evidence: lM() at line 373333, t8() at line 373341

What: Configure default Haiku model via ANTHROPIC_DEFAULT_HAIKU_MODEL environment variable.

Details:

• Allows overriding the default Haiku model for specific use cases
• Falls back to standard Haiku 3.5 if not set
• Useful for testing or using alternative Haiku versions
• Evidence: FX2() at line 371423

What: Automatically removes empty directories after file operations.

Details:

• Cleans up empty directories left after file deletions
• Prevents directory clutter in file history operations
• Silently handles cases where directory removal fails
• Evidence: U65() at line 411458

What: Proper reset of session cost and usage metrics.

Details:

• Resets total cost, API duration, tool duration, line counts
• Ensures clean state for new sessions
• Prevents metric accumulation across sessions
• Evidence: Fo1() at line 340029

• Added hOA() at line 364421 for processing and optimizing base64-encoded images

• Added u00() at line 362718 for joining and truncating string arrays with size limits

• Added EG1() at line 374270 for expanding filesystem paths (~, ./, relative to absolute)
• Added NG1() at line 374278 for escaping paths in sandbox profiles

• Added $j9() at line 366080 for checking if text contains thinking triggers
• Added Nw1() at line 366092 for extracting trigger word positions
• Added Ew1() at line 365889 for splitting text into highlighted/normal segments