Claude Code v2.0.70

What: A new interactive feature discovery interface that helps users explore and track their usage of Claude Code's capabilities across organized categories.

How to use:

# Open the feature discovery interface
/discover

Details:

• Organizes features into 6 categories: Quick Wins, Speed, Code, Collaborate, Customize, and Power User
• Tracks which features you've explored with progress indicators
• Shows completion percentage and visual progress bar
• Each feature includes a "Try it" prompt for hands-on learning
• Navigate with arrow keys, Enter to explore, Esc to close
• Evidence: Feature categories defined at lines 168232-168267 in pretty-v2.0.70.js; /discover command at PpZ variable around line 463990

What: A new tool that allows Claude to search and load MCP (Model Context Protocol) tools before using them, improving reliability of MCP integrations.

How to use: MCP tools are now deferred and must be loaded via MCPSearch before use. Claude will automatically use this tool when you request MCP functionality.

Details:

• Supports two query modes: direct selection (select:mcp__slack__read_channel) or keyword search
• Returns up to 5 matching tools ranked by relevance
• Caches tool descriptions for faster subsequent searches
• Returns tool_reference content blocks that make selected tools available
• Evidence: MCPSearchTool implementation at Ey2 variable, line 454185; tool name Gy2 = "MCPSearch" at line 454099

What: Visual progress indicator when teleporting into a remote session, showing each step of the process.

How to use: When teleporting to a session, you'll now see a multi-step progress display:

◐ Teleporting session…
  ✓ Validating session
  ◐ Fetching session logs
  ○ Getting branch info
  ○ Checking out branch

Details:

• Shows animated spinner for current step
• Displays checkmarks for completed steps
• Four stages: validating, fetching_logs, fetching_branch, checking_out
• Evidence: di2 steps array at line 503776; SU5 progress component at line 503695

What: Internal tracking of which features you've used to power the discovery system's progress indicators.

Details:

• Tracks feature usage in local config
• Powers the "explored X of Y features" progress display
• Per-category completion tracking
• Evidence: featureUsage state property referenced at lines 168212, 168631, 168637

What: New setting for controlling in-session coaching tips and guidance.

Details:

• Configurable coaching tip behavior
• Tracks coaching tips shown per session via coachingTipsThisSession
• Controlled via coachingMode setting (can be set to "off")
• Evidence: coachingMode getter at line 368908; coachingTipsThisSession at line 368906

The official Anthropic marketplace auto-installation now includes improved retry logic with exponential backoff:

• Tracks retry count, last attempt time, and next retry time
• Maximum 10 retry attempts with delays from 1 hour to 7 days
• Shows user notifications for installation status (success, failure, git unavailable)
• Evidence: Retry constants at o51 variable, line 445893 (MAX_ATTEMPTS: 10, INITIAL_DELAY_MS: 3600000, MAX_DELAY_MS: 604800000)

New support for dynamically adding and removing MCP servers at runtime:

• RU5() function handles hot-reloading of MCP configurations
• Properly cleans up removed servers and initializes new ones
• Returns detailed results including which servers were added, removed, and any errors
• Evidence: RU5() function at line 502788

New prompt coaching system that can display contextual tips during tool use:

• Tracks tips shown per session
• Integrates with the coaching mode setting
• Evidence: promptCoaching state with tip and shownAt at line 487684

The /security-review command now includes expanded false positive filtering guidelines:

• Added exclusions for React/Angular XSS (frameworks handle escaping)
• Added exclusions for client-side permission checking (backend handles this)
• Added exclusions for notebook vulnerabilities without concrete attack paths
• Evidence: Enhanced filtering rules in ZV5 template starting at line 471235