Version 2.0.50

What: View and resume conversations from any project directory, not just the current one. How to use:

# Use /resume command and toggle "Show all projects" 
claude
> /resume
# Press the toggle key to switch between current project and all projects

Details:

• When enabled, scans ~/.claude/projects/ for all project directories
• Shows project path in conversation list for cross-project items
• For conversations from different directories, copies the resume command to clipboard
• Provides clear instructions: cd /path/to/project && claude --resume <session-id>
• Evidence: FZ1() at line 487757, VZ1() at line 504532 in v2.0.50

What: HTTP REST API server for external tools to interact with Claude Code's MCP servers programmatically. How to use:

# Enable the endpoint (enabled by default when Claude Code is running)
# External tools can make requests to localhost with bearer token auth

# Example: List available MCP servers
curl -X POST http://127.0.0.1:<port>/mcp \
  -H "Authorization: Bearer <secret>" \
  -H "Content-Type: application/json" \
  -d '{"command": "servers"}'

# Example: Call an MCP tool
curl -X POST http://127.0.0.1:<port>/mcp \
  -H "Authorization: Bearer <secret>" \
  -d '{"command": "call", "params": {"server": "myserver", "tool": "mytool", "args": {}}}'

Details:

• Supports commands: servers, tools, info, call, grep, resources, read
• Bearer token authentication with 32-byte random secret
• Localhost-only binding for security
• Endpoint info stored in ~/.claude/<workspace>.endpoint
• Can be disabled via ENABLE_MCP_CLI_ENDPOINT=false environment variable
• Evidence: class hX0 at line 515880 in v2.0.50

What: Plan files now use memorable, human-readable names instead of UUIDs. How to use:

# Session plans are now named with friendly slugs like:
# ~/.claude/plans/fluffy-bouncing-bunny.md
# ~/.claude/plans/golden-dancing-butterfly.md

Details:

• 146 adjectives × 112 verbs × 300 nouns = 4.9 million unique combinations
• Word categories include nature terms, animals, and whimsical objects
• Names like "sparkly-juggling-penguin" or "cozy-dreaming-unicorn"
• Evidence: $A2() at line 300930, word arrays at lines 300938-301497 in v2.0.50

What: Real-time display of file modifications during a conversation showing files changed and lines added/removed. Details:

• Shows: 2 files +15 -8 format with color-coded counts
• Green for additions, red for deletions
• Tracks changes from tool use results
• Updates as files are modified during the session
• Evidence: Fo2() component at line 451329 in v2.0.50

What: Pass JSON schemas at SDK initialization time for automatic structured output validation. How to use:

# Python SDK example
from claude_code import ClaudeCode

client = ClaudeCode()
response = client.run(
    prompt="Generate user data",
    json_schema={
        "type": "object",
        "properties": {
            "name": {"type": "string"},
            "age": {"type": "integer"}
        },
        "required": ["name", "age"]
    }
)

Details:

• Schema is validated using Ajv validator
• Automatically creates a StructuredOutput validation tool
• Validates all outputs against the schema throughout the session
• Works via both SDK initialization and CLI --json-schema flag
• Evidence: $D0() at line 2327, $G1() at line 470974 in v2.0.50

What: Claude must now include a "Sources:" section with hyperlinked references when using web search. Details:

• Mandatory "Sources:" section at end of responses using web search
• Links formatted as markdown hyperlinks: [Title](URL)
• Returns now include "links as markdown hyperlinks" in results
• Evidence: k9B variable at line 187201 in v2.0.50, adds "CRITICAL REQUIREMENT" section

What: Significant improvements to Language Server Protocol client reliability and error handling. Details:

• New spawn event handling ensures process fully started before using stdio
• Added stdin error handler for write failures
• Added connection-level error and close handlers
• New cleanup flag prevents spurious errors during intentional shutdown
• Enhanced listener cleanup removes stdin/stderr handlers on stop
• Trace setup now has error handling
• Evidence: w02() function at line 304885 in v2.0.50

What: Added "Key results" section to session memory template for capturing concrete outputs. Details:

• New section between "Learnings" and "Worklog"
• Purpose: Store specific answers, tables, or documents user requested
• Template prompt: "_If the user asked a specific output such as an answer to a question, a table, or other document, repeat the exact result here_"
• Evidence: hP3 template at line 508626 in v2.0.50

What: Added retry mechanism for remote session log persistence. Details:

• Retries failed persistence attempts with exponential backoff
• Base delay of 1 second, max 8 seconds between attempts
• Better error messages including entry UUID in conflict errors
• Evidence: JS5() function at line 393547 in v2.0.50

What: Improved detection and display of IDE names from various formats. Details:

• Better handling of lowercase and trimmed IDE identifiers
• Fallback lookup for IDE display names
• Handles space-separated IDE names (uses first word)
• Evidence: CE() function at line 318709 in v2.0.50

What: Improved handling of oversized MCP tool results with structured output support. Details:

• Detects when results exceed maximum token limits
• Saves large outputs to temporary files
• Reports output format (plain text, JSON, JSON array with schema)
• Provides clear instructions for reading chunked content
• Evidence: HZ5() function at line 333257 in v2.0.50

What: Added detailed startup performance tracking for diagnostics. Details:

• Tracks import time, init time, settings load time, and total startup time
• Reports checkpoint counts and timing intervals
• Sampled at 0.1% of sessions for telemetry
• Evidence: wT3() function at line 506078, UT3 timings at line 506093 in v2.0.50