Claude Code v2.0.34

What: A new command-line interface for interacting with MCP (Model Context Protocol) servers directly from bash commands within Claude Code sessions.

How to use:

# Discover available MCP servers and tools
mcp-cli servers                        # List all connected MCP servers
mcp-cli tools [server]                 # List available tools
mcp-cli grep "weather"                 # Search tools by keyword

# Get tool schema (REQUIRED before calling any tool)
mcp-cli info slack/search_private      # View JSON schema for parameters

# Invoke MCP tools
mcp-cli call slack/search_private '{"query": "mentions", "max_results": 10}'

# Use stdin for complex JSON
mcp-cli call api/send_request - <<'EOF'
{
  "endpoint": "/data",
  "headers": {"Authorization": "Bearer token"}
}
EOF

Details:

• Enabled via ENABLE_EXPERIMENTAL_MCP_CLI environment variable
• Claude must call mcp-cli info before any mcp-cli call to check parameter schemas
• Includes full Commander.js CLI with multiple subcommands
• Automatically tracked with telemetry event tengu_mcp_cli_command_executed
• Evidence: Command parser Re() at line 3935, handler so2() at line 493008, entry point check at line 493024 in pretty-v2.0.34.js

What: New API client for managing Claude skills through the beta API.

How to use:

// Skills are managed through the beta client
const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });

// Create a skill
await client.beta.skills.create({
  name: "my-skill",
  // ... skill configuration
}, { betas: ["skills-2025-10-02"] });

// Create a skill version
await client.beta.skills.versions.create(skillId, {
  // ... version configuration
}, { betas: ["skills-2025-10-02"] });

// List, retrieve, delete skills and versions
await client.beta.skills.list();
await client.beta.skills.retrieve(skillId);
await client.beta.skills.delete(skillId);

Details:

• Requires skills-2025-10-02 beta header
• Full CRUD operations for skills and skill versions
• Accessed via client.beta.skills property
• API endpoints: /v1/skills and /v1/skills/{id}/versions
• Evidence: Class BFA at line 146202, class Q2A at line 146272, beta client integration at line 146346 in pretty-v2.0.34.js

What: Automatic collection of diagnostics (errors, warnings, hints) from LSP servers and delivery as message attachments.

Details:

• LSP servers now send textDocument/publishDiagnostics notifications automatically
• Diagnostics are queued and attached to the next API request as lsp_diagnostics attachments
• Includes file URI, line ranges, severity levels (Error/Warning/Info/Hint), and diagnostic messages
• Handles multiple LSP servers with individual error tracking and retry logic
• New saveFile method added to LSP manager for textDocument/didSave notifications
• Evidence: Handler registration lUQ() at line 276542, diagnostic registration dUQ() at line 276467, severity mapper WR6() at line 276504, transformer FR6() at line 276518 in pretty-v2.0.34.js

What: New settings UI to control whether commands can fall back to running outside the sandbox when sandbox restrictions prevent execution.

How to use:

# Access via /sandbox command in interactive mode
# Choose from:
# 1. "Allow unsandboxed fallback" - Commands retry outside sandbox if blocked
# 2. "Strict sandbox mode" - All commands must run in sandbox or be excluded

Details:

• Separate from the main sandbox enable/disable setting
• Can be locked by policy settings from higher-priority configuration sources
• Status displayed in /sandbox command output (e.g., "sandbox enabled, fallback allowed")
• New API methods: areUnsandboxedCommandsAllowed(), areSandboxSettingsLockedByPolicy(), setSandboxSettings()
• Documentation link: docs.claude.com/en/docs/claude-code/sandboxing#configure-sandboxing
• Evidence: Component bn2() at line 465885, policy check Fs2() at line 472000 in pretty-v2.0.34.js

What: Improved handling of emoji and CJK characters in terminal output through explicit width tracking in the screen buffer.

Details:

• Cell objects now include explicit width property (0 for normal, 1 for wide, 2/3 for placeholder)
• Automatic insertion of placeholder cells when writing wide characters
• Renderer skips placeholder cells to prevent double-rendering artifacts
• Width-aware cursor movement (advances 2 positions for wide characters)
• Fixes rendering bugs where placeholder cells were processed incorrectly
• Evidence: Enhanced cell type pf9 at line 67236, width-aware write function $P0() at line 67208, placeholder skipping logic at lines 67852-67853 in pretty-v2.0.34.js

What: Telemetry events now log to both Statsig and Segment (when enabled by feature gate).

Details:

• New feature gate tengu_log_segment_events controls Segment logging
• Synchronous logging via ZA() function, async via uw() function
• Separate formatters for each system: sc2() for Statsig (flattened strings), rc2() for Segment (structured objects)
• Parallel execution when both systems enabled for optimal performance
• Shared context builder t01() gathers environment, model, and session data
• Gate value cached to avoid repeated lookups
• Evidence: Dual logger ZA() at line 477386, async logger uw() at line 477389, context builder t01() at line 451726, gate checker Lr2() at line 477375 in pretty-v2.0.34.js

What: Terminal rendering code reorganized for better maintainability and to support ink2 mode.

Details:

• Split render() into render_v1() and render_v2() methods
• Cleaner separation between legacy and new rendering paths
• Improved parameter passing (prevFrame explicitly passed instead of using instance state)
• Better handling of non-TTY output modes
• Evidence: Class RG1 with split rendering methods at line 67663 in pretty-v2.0.34.js

What: Enhanced telemetry with optional process-level metrics.

Details:

• New processMetrics field in telemetry context
• Included in both Statsig and Segment events when available
• Retrieved via Lt5() function
• Adds visibility into resource usage patterns
• Evidence: Process metrics in t01() at line 451726, formatter handling in sc2() at line 451751 and rc2() at line 451762 in pretty-v2.0.34.js

What: Improved permission checking for MCP tools invoked via bash commands.

Details:

• New command parser Re() detects mcp-cli call syntax in bash commands
• Permission check function Bg8() validates MCP tool access before execution
• Prevents bash command splitting from bypassing MCP permissions
• Returns appropriate deny/ask/allow behavior with suggestions
• Evidence: Parser Re() at line 3935, permission check Bg8() at line 212208 in pretty-v2.0.34.js

What: Telemetry environment context gathering extracted into reusable function.

Details:

• New Nt5() async function collects platform, runtime, and environment data
• Includes: platform, arch, Node version, terminal type, package managers, runtimes, deployment environment, CI detection, WSL version
• Shared between Statsig and Segment telemetry
• Eliminates code duplication from previous inline implementation
• Evidence: Function Nt5() at line 451680 in pretty-v2.0.34.js

What: Session ID now uses environment variable when available for better tracking across processes.

Details:

• New SQI() function checks CLAUDE_CODE_SESSION_ID environment variable
• Falls back to generated session ID if not set
• Enables correlation of events across related processes
• Evidence: Function SQI() at line 483098 in pretty-v2.0.34.js

What: New telemetry event to track forked conversation branches in chat history.

Details:

• Event tengu_session_forked_branches_fetched tracks branch statistics
• Metrics: total sessions, sessions with branches, max branches per session, average branches
• Only fires when at least one session has multiple branches (max > 1)
• Helps understand user exploration patterns
• Evidence: Function f1I() at line 474125 in pretty-v2.0.34.js

• Old j1() function replaced with ZA() (sync) and uw() (async)
• Old Rs2() Statsig logger replaced with f00()
• Old Os2() Segment logger renamed to sA0() with updated helpers
• Context gathering, formatting, and logging now properly separated
• Feature gate system for conditional Segment logging

• Added saveFile method to LSP manager for textDocument/didSave notifications
• Improved server initialization with automatic startup on first LSP manager initialization
• Better error handling and logging throughout diagnostic flow

• Updated from v2.0.33 to v2.0.34
• Evidence: Version string at line 451680 in pretty-v2.0.34.js