Claude Code v2.0.9

What: Plugins can now include "skills" - reusable AI behaviors defined in markdown files that can be loaded and executed by Claude Code.

How to use:

# Create a skills directory in your plugin
~/.claude/plugins/my-plugin/skills/my-skill/SKILL.md

# Or specify custom paths in your plugin manifest
{
  "name": "my-plugin",
  "skills": ["custom-skills-dir", "another-skills-dir"]
}

Details:

• Skills are loaded from subdirectories containing SKILL.md files
• By default, looks for a skills/ directory in your plugin root
• Custom paths can be specified in the plugin manifest using the skills field (single path or array)
• Each skill directory should contain a SKILL.md file with frontmatter and content
• Skills are processed similarly to commands and agents but with isSkillMode: true
• Evidence: hMB() at line 422839, skillsPath property at line 377365, isSkillMode flag at line 422855

What: Internal enhancement allowing custom fetch implementations to be injected into the API client for testing or custom network handling.

Details:

• The API client factory function (fF()) now accepts an optional fetchOverride parameter
• Replaces previous hardcoded fetch selection logic
• Primarily used internally but enables advanced customization scenarios
• Evidence: fF() at line 376120, fetch override usage at lines 411360 and 432771

What: Feedback from post-tool-use hooks is now properly converted to user messages, making it visible to the AI model for improved context awareness.

How it works now:

• Post-tool-hook feedback creates messages with subtype "post_tool_hook_feedback"
• These messages are automatically converted to user messages during processing
• The AI model can now reference and build upon hook outputs in subsequent responses

Details:

• Previously, hook feedback was treated as generic "informational" system messages
• New dedicated functions: eb1() creates the feedback messages, fPB() identifies them
• Conversion happens in message processing pipeline via AD5() function
• Includes dedicated UI rendering function h2B() for displaying feedback
• Evidence: fPB() at line 436505, eb1() at line 437160, conversion logic at line 436513, h2B() rendering at line 405844

What: The Grove climate contribution notice display logic now respects a grace period flag, allowing more persistent display of important notices.

How it works:

• When notice_is_grace_period is false, notices are shown regardless of previous user dismissals
• Grace period defaults to true, maintaining existing behavior
• After grace period ends, users see notices even if previously dismissed

Details:

• Affects the timing of when Grove-related notices appear in the CLI
• Allows for more aggressive communication of important announcements after initial grace period
• Evidence: GBB() at line 407802 (previously m2B() at line 407357 in v2.0.8), new grace period check at line 407805

• Updated AWS Bedrock SDK client code with new type definitions
• Refactored helper functions including DFB() (replaces tXB()) for local command message checking
• Function renames: vF() → fF(), m2B() → GBB(), HH5() → AD5()