Claude Code v1.0.78

What: Authentication tokens can now be passed via file descriptors instead of environment variables How to use:

# Pass token through file descriptor (more secure)
export CLAUDE_CODE_WEBSOCKET_AUTH_FILE_DESCRIPTOR=3
exec 3< <(echo "your-token-here")
claude

# Traditional method still supported
export CLAUDE_CODE_SESSION_ACCESS_TOKEN="your-token-here"
claude

Details:

• Prevents tokens from appearing in process listings or environment dumps
• Reads token from /proc/self/fd/<number> on Linux systems
• Falls back to environment variable for backward compatibility
• Provides clear error messages for invalid file descriptors

What: New granular permission options for file operations How to use: When Claude requests file permissions, you now have three options:

• Accept once - Permission for this operation only
• Accept for session (Shift+S) - Permission for all operations this session
• Accept for project (Shift+A) - Persistent permission for this project

Details:

• Different options appear based on whether files are inside or outside your project
• Project-level permissions are saved to your local settings
• Read operations don't offer project-level write permissions
• Clearer distinction between temporary and persistent permissions

What: New high-context variant of Claude 3.5 Sonnet with 1 million token context window Details:

• Available as "Sonnet (1M context)" option
• Uses rate limits faster than standard models
• Access may be gated based on account type or feature flags
• Provides extended context for working with very large codebases

What: New streamlined onboarding for first-time users Details:

• Shows "new-user-warmup" flow for users with fewer than 10 app startups
• Controlled via cc_simple_onboarding experiment flag
• More focused introduction compared to standard onboarding
• Automatically selected based on usage patterns

What changed: The entire "output modes" feature is now called "output styles" Impact:

• Directory structure changed from ~/.claude/output-modes/ to ~/.claude/output-styles/
• Built-in options renamed: "Insights" → "Explanatory", "Learn By Doing" → "Learning"
• Clearer terminology that better reflects the feature's purpose
• Agent type renamed from output-mode-setup to output-style-setup

What changed: Models can now be configured remotely via Statsig Impact:

• Claude Code can receive model updates without requiring new releases
• Supports time-based feature gating for gradual rollouts
• Enables A/B testing of different model configurations
• External overrides respect user's first usage date to prevent unexpected changes

What changed: Agent definitions now include source tracking Impact:

• Agents now specify their source: built-in, localSettings, or projectSettings
• Better management and display of agent origins
• Statusline setup agent now recommends skipping optional git locks for better performance

What changed: Simplified and more reliable permission persistence Impact:

• Clear tracking of permission sources (session vs persistent)
• New helper functions for managing directory permissions
• Automatic addition of allowed directories to settings
• Better handling of permissions for files outside project directory

• Issue: Terminal raw mode wasn't properly cleaned up on exit
• Resolution: Added Ao4() function to properly reset stdin terminal mode and unref the stream

• Issue: Enterprise and team accounts weren't properly detected for feature gating
• Cause: Missing logic to check account type in conjunction with environment flags
• Resolution: New BL6() function correctly identifies enterprise/team accounts

• Issue: Model selection didn't properly handle all user access scenarios
• Cause: Complex conditional logic without proper fallback chain
• Resolution: Refactored into cleaner function chain with proper override points