Claude Code v1.0.73

What: Automatically saves all bash command outputs to files for future reference How to use:

# When you run any bash command, outputs are automatically saved
claude "npm test"
# Outputs saved to bash-outputs/ directory with timestamp-based filenames

Details:

• All command outputs (stdout and stderr) are preserved in bash-outputs/ directory
• Files are named with timestamps and command hashes for uniqueness
• Infrastructure for AI-powered output summarization is included but currently disabled
• Future releases will enable intelligent summarization of verbose outputs (>5000 chars)

What: More informative and specific permission approval messages How to use:

# You'll now see specific reasons why permission is needed
# Instead of: "Claude requested permissions to use Bash"
# You'll see: "Permission rule 'npm-*' from project settings requires approval for this Bash command"

Details:

• Messages now show which rule, hook, or mode triggered the permission request
• Clear attribution of permission rules to their source (local settings, project settings, etc.)
• Multi-command operations show which specific parts need approval
• New hook system allows external tools to provide custom permission feedback

What changed: Command injection and safety checks reorganized into modular architecture Impact: Same security coverage with better performance and maintainability - no change in user-facing behavior, but more accurate detection with fewer false positives

What changed: New functions for tracking and managing "always ask" permission rules across different sources Impact: More granular control over which operations require approval, with clearer visibility into rule sources

What changed: Removed 17 unused utility functions (lodash-like implementations) Impact: Smaller bundle size and cleaner codebase - no functional changes for users

• Issue: Permission decisions weren't clearly showing their source
• Cause: Decision reasons were stored as plain strings without context
• Resolution: Now uses structured objects with type information and source attribution

• Issue: Some safe commands with quotes were incorrectly flagged as dangerous
• Cause: Quote parsing didn't properly distinguish between quoted and unquoted content
• Resolution: Enhanced parsing system better understands quote contexts