Claude Code v2.0.20

What: Four new commands for managing plugins directly from the command line

How to use:

# Install a plugin from a marketplace
claude plugin install <plugin-name>
claude plugin install <plugin-name>@<marketplace>

# Uninstall a plugin
claude plugin uninstall <plugin-name>

# Enable a previously disabled plugin
claude plugin enable <plugin-name>

# Disable an installed plugin
claude plugin disable <plugin-name>

Details:

• Commands support aliases: install → i, uninstall → remove
• Installation automatically searches configured marketplaces
• Telemetry tracks CLI plugin operations separately from UI operations
• Evidence: zSQ() at line 465305, CSQ() at line 465337, USQ() at line 465355, $SQ() at line 465370

What: Complete implementation of the Skill tool, enabling Claude to execute skills and commands directly during conversations (previously existed only as a stub)

How to use: Skills are invoked automatically by Claude when appropriate, or you can configure permission rules:

# In .claude/settings.local.json
{
  "permissions": {
    "allow": {
      "Skill": ["pdf", "xlsx", "custom-skill:*"]
    }
  }
}

Details:

• Supports both mode commands (high-priority) and regular skills
• Includes intelligent token budgeting (15,000 character default via SLASH_COMMAND_TOOL_CHAR_BUDGET)
• Permission system with wildcard patterns (e.g., skill-name:*)
• Skills are categorized by source: project, user, plugin, or managed
• Mode commands receive priority and full token allocation
• Evidence: Mo1 initialization at line 429046, W$Q() prompt generator at line 428983, nw variable at line 428903

What: Real-time visual feedback showing hook execution progress during tool operations

How to use: Progress appears automatically when hooks are running:

Running PreToolUse hooks… (3/5 done)
  · pre-validate: npm run lint
  · security-check: ./scripts/check-secrets.sh
  · backup: git stash

Details:

• Shows hook event name (e.g., "PreToolUse", "PostToolUse")
• Displays progress counter when multiple hooks run
• Verbose mode shows individual hook names and commands being executed
• Progress UI disappears when all hooks complete
• Evidence: mi1() at line 366839, LR5() at line 366871, CoB() at line 455699, UoB() at line 455708, n48() at line 384434

What: Export detailed context usage statistics as markdown tables for non-interactive sessions

How to use:

# Run context command in non-interactive mode
claude --output text context

# Output includes markdown tables:
# - Categories (with tokens and percentages)
# - MCP Tools
# - Custom Agents
# - Memory Files
# - SlashCommand Tool statistics

Details:

• Token counts formatted for readability (e.g., "1.5k" instead of "1500")
• Includes model information and total usage percentage
• Shows free space and autocompact buffer separately
• Agent sources identified (Built-in, Project, User, Local, Flag, Policy, Plugin)
• Evidence: Ma6() at line 301494, rk() token formatter at line 301491

What: Refactored project MCP server configuration loading into a dedicated function with enhanced scope handling

How to use: Create .mcp.json in your project root:

{
  "mcpServers": {
    "project-server": {
      "command": "node",
      "args": ["./server.js"]
    }
  }
}

Details:

• Dedicated function for loading project MCP configuration
• Proper error collection and reporting
• Supports environment variable expansion
• Explicitly scoped as "project" configuration
• Evidence: dR0() at line 280421

What: Override the default autocompact threshold using an environment variable

How to use:

# Set autocompact to trigger at 75% of context capacity
export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=75
claude

# Or inline
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=90 claude

Details:

• Accepts percentage values from 1-100
• Falls back to default if invalid value provided
• Provides more control over memory management
• Useful for balancing context usage vs. compaction frequency
• Evidence: Gn6() at line 296137

What: Automatic detection when Claude Code runs as a GitHub Action

How to use: Set the environment variable in your GitHub workflow:

- name: Run Claude Code
  env:
    CLAUDE_CODE_ACTION: true
  run: claude -p "Review this PR"

Details:

• Accepts flexible values: "1", "true", "yes", "on" (case-insensitive)
• Sets entrypoint to "claude-code-github-action" for proper telemetry
• Enables GitHub Action-specific behavior and logging
• Evidence: yT8() at line 469210, jA() validation at line 3336

What: Improved detection of backtick command substitution with escape-aware parsing

Details:

• Dedicated CV6() function properly handles escaped characters
• Backtick checks now execute before other pattern matching
• More accurate detection prevents false positives from escaped backticks
• Evidence: CV6() at line 205030, RV6() at line 205296

What: Optimized terminal rendering with synchronized updates

Details:

• Reduced function complexity in d50() at line 71278 (previously U50())
• Removed separate stderr handling for improved performance
• Simplified output buffering logic
• All output now uses synchronized update sequences

What: Simplified autocompact summary function parameters

Details:

• Added hasAppendSystemPrompt parameter to pi0() at line 434307 (previously Ei0())
• Better context awareness during compaction operations
• More precise token counting for system prompts

What: Enhanced tool permission checking with additional context parameter

Details:

• Updated _M() permission checker at line 458051 (previously wM())
• Added fourth parameter for improved context handling
• More robust permission decision-making

• Removed stub Skill tool implementation (ld0 at v2.0.19:410872)
• Removed deprecated hook progress renderer (aO5 at v2.0.19:366835, sO5 at v2.0.19:366851)
• Removed unused clear terminal function (ji2 at v2.0.19:208933)
• Removed empty placeholder functions (RLQ, LRQ, eOQ)
• Added PassThrough stream import for improved stream handling
• Added execSync and spawn imports for child process management
• Refactored Skills loading with enabled state check (function tw8() at line 450050)
• Added helper functions for plugin name parsing and matching (HSQ() at line 465275, ms0() at line 465282, us0() at line 465264)
• Added error handler Ue1() at line 465298 for plugin CLI operations
• Multiple variable initializations refactored for improved lazy loading