Claude Code v2.0.24

What: When sandboxing is enabled and you're in "accept edits" mode, safe bash commands now execute automatically without permission prompts.

How to use:

# Enable the sandbox with auto-allow feature
claude --configure
# Select "Sandbox BashTool, with auto-allow in accept edits mode"

# Now when you ask Claude to create files/directories, commands like
# mkdir, touch, mv, cp execute automatically when sandboxed

Details:

• Only activates when all conditions are met: sandbox enabled, auto-allow feature enabled, command is sandboxable, and user is in "acceptEdits" mode
• Safe commands like mkdir, touch, rm, rmdir, mv, cp, sed are pre-approved in acceptEdits mode
• User-defined deny/ask rules are still respected
• Commands can be excluded via sandbox.excludedCommands configuration
• Falls back to standard permission prompting if conditions aren't met
• Evidence: hL0() function at lines 206586-206604 and TN6() decision logic at line 206562 in pretty-v2.0.24.js

What: Keyboard event handlers can now stop event propagation to other handlers using event.stopImmediatePropagation().

How to use:

// For developers building on Claude Code's UI framework
useInput((input, key, event) => {
  if (input === 'q' && shouldQuit) {
    event.stopImmediatePropagation();  // Prevent other handlers from running
    handleQuit();
  }
});

Details:

• Third parameter added to keyboard event handlers containing the event object
• Enables priority-based event handling for modal dialogs, focus management, and vim-mode implementations
• Backward compatible - existing handlers with 2 parameters continue to work
• Implements DOM-style event propagation patterns
• Evidence: New VO1 event class at line 71066, B11 base class with stopImmediatePropagation() at line 69922, and updated callback signature in QT9() at line 71693 in pretty-v2.0.24.js

What: When using AWS Bedrock, credentials are now validated and refreshed automatically at startup.

How to use:

# Configure custom AWS auth commands in ~/.claude.json
{
  "awsAuthRefresh": "aws sso login --profile my-profile",
  "awsCredentialExport": "aws configure export-credentials --profile my-profile"
}

# Or skip auto-initialization if not needed
export CLAUDE_CODE_SKIP_BEDROCK_AUTH=1
claude

Details:

• Runs at startup when CLAUDE_CODE_USE_BEDROCK is set and CLAUDE_CODE_SKIP_BEDROCK_AUTH is not set
• First attempts standard AWS SDK authentication via STS GetCallerIdentity
• Falls back to custom awsAuthRefresh or awsCredentialExport commands if needed
• Includes workspace trust checks before executing custom commands
• Provides visual authentication feedback
• Evidence: New uhQ() initialization function at line 465179, called during startup at line 475285 in pretty-v2.0.24.js

Haiku 4.5 (claude-haiku-4-5-20251001) can now generate responses up to 32,000 tokens instead of being capped at 8,192 tokens. Claude 3.5 Haiku remains at 8,192 tokens.

Changelog fetching is now skipped in non-interactive sessions (SDK usage, --print mode, non-TTY environments), reducing unnecessary network traffic during programmatic usage.

Path resolution now correctly handles glob patterns (*, ?, [, ]) in sandbox file access rules, preventing errors when specifying pattern-based access controls.

Pro plan users may be assigned to experiments that change default model selection:

• Haiku default: Some Pro users may see Haiku 4.5 as default instead of Sonnet 4.5 (cost optimization)
• Sonnet 1M default: Users with 1M context access may default to Sonnet 4.5 [1M] instead of [200K]

These are A/B tests controlled by feature flags and not visible to all users.

Added telemetry for tool_use/tool_result mismatch errors to help engineers debug message normalization issues. This internal tracking does not affect user experience but helps improve reliability.