Version 2.0.32

What: A new toolRunner() method in the Messages API that automatically executes tools in an agentic loop, eliminating the need to manually handle tool_use/tool_result message cycles.

How to use:

const runner = client.beta.messages.toolRunner({
  model: "claude-sonnet-4-20250514",
  max_tokens: 1024,
  max_iterations: 10, // Limit agentic loops
  tools: [
    {
      name: "get_weather",
      run: async (input) => {
        // Your tool implementation
        return "Sunny, 72°F";
      }
    }
  ],
  messages: [{ role: "user", content: "What's the weather?" }]
});

// Stream each iteration
for await (const messageStream of runner) {
  // Handle streaming response
}

// Or wait for completion
const finalMessage = await runner.done();

Details:

• Automatically detects tool_use blocks in assistant responses and executes matching tools
• Supports both streaming and non-streaming modes
• Configurable iteration limits via max_iterations parameter
• Tools must implement a run() method that returns results
• Optional parse() method for input validation/transformation
• Returns properly formatted tool_result messages with error handling
• Can be controlled manually via generateToolResponse() and pushMessages() methods
• Evidence: Class TYA at line 82603, function bh9() at line 82564, variable uZ1 at line 82604, new toolRunner() method at line 82828

What: New slash command to order official Claude Code stickers

How to use:

/stickers

Details:

• Opens browser to https://www.stickermule.com/claudecode
• Falls back to displaying URL if browser open fails
• Available in all contexts (always enabled, not hidden)
• Does not support non-interactive mode
• Evidence: Variable $d2 at line 463792

What: Automatic notifications when users with Sonnet 1M access approach context limits

How to use: These hints appear automatically when relevant:

# Warning when context is full:
Run /compact or /model sonnet[1m] to continue

# Tip after compaction:
Tip: You have access to Sonnet 1M with 5x more context

Details:

• Only shown to users with 1M context access (controlled via access check)
• Highlights the 5x context multiplier benefit
• Integrated into warning and tip notification system
• Improves discoverability of the 1M context feature
• Evidence: Functions Wh6() at line 317408 and p8A() at line 317415

What: New screen diffing algorithm that only updates changed characters instead of clearing and re-rendering the entire terminal, eliminating flicker and improving performance.

How it works: The renderer now:

1. Compares previous and current screen buffers character-by-character
2. Only updates positions where content changed
3. Manages cursor movement efficiently during incremental updates
4. Falls back to full re-render only when screen dimensions change

Details:

• Significantly reduces terminal flicker during updates
• Improves performance for large terminal windows
• Better user experience during streaming responses
• Implements transaction-based cursor management for atomic updates
• Evidence: Functions qRA() at line 70135, uR0() at line 70140, mR0() at line 70379, class dR0 at line 70395

What changed: The in-app feedback survey mechanism was refactored for improved accuracy and maintainability

Key improvements:

• Switched from tracking user message counts to tracking actual submission counts, preventing inconsistencies from message editing or restoration
• Moved survey-specific state from mixed local/global storage to centralized app-level state management
• Simplified conditional logic with early returns for better maintainability
• Adjusted triggering configuration:
• Enabled for all models (was disabled by default)
• Reduced probability from 100% to 0.5% when conditions are met
• Reduced turn requirement between surveys from 15 to 10
• Effectively disabled global cooldown (increased from 1 hour to ~1157 days)
• Removed per-conversation cooldown (30 minutes)

Net effect: Surveys appear less frequently but with more accurate tracking across sessions

Function Ky2() at line 445844 (renamed from mS2), configuration ku5 at line 445950, state structure at line 298044

What changed: Consolidated and reorganized import statements for better tree-shaking and bundle size optimization

Changes:

• Replaced import We from "stream" with import { PassThrough as sr5 } from "stream" at line 482663
• Replaced import XI1 from "node:child_process" with import { execSync as l6Q, spawn as yd8 } from "node:child_process" at line 249497
• Replaced import _P9 from "node:os" with import { homedir as Wo5 } from "node:os" at line 484356
• Added import { cwd as XT0 } from "node:process" at line 70828
• Added import { join as Dn2, basename as Pr5 } from "path" at line 481476

Diff lines 10-21 (removed) and 534-537, 955-958, 1093-1096, 1105-1112 (added)

What changed: The SDK initialization function now sends authentication status after successful initialization when in headless mode

Details:

• After initialization completes, checks for ongoing authentication
• Sends auth_status event with authentication state, output, and error information
• Only triggers in headless/non-interactive contexts (parameter J must be true)
• Helps SDKs and automation tools track authentication progress
• Evidence: Function Go5() (renamed from kr5) at line 483837, lines 1288-1301 show the new auth status emission

Several new utility functions were added to improve code organization:

• yn1() at line 416368: Enhanced slice function with offset support (replaces Fn1())
• kn1() at line 416372: Relative path calculator that returns absolute path if result would escape working directory
• _n1() at line 416377: Formats limit/offset parameters for display
• Sb5() at line 430186: Reads CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY environment variable with default value of 10
• YuA() at line 207879: Strips <sandbox_violations> tags from content using regex
• dS2() at line 470949: Comparison function for checking if one value exceeds another in a specific context

What changed: Added feedbackSurvey object to app-level state structure

New state fields:

feedbackSurvey: {
  timeLastShown: null,
  submitCountAtLastAppearance: null
}

Details:

• Persists survey appearance timing and submission count
• Enables accurate rate limiting across sessions
• Synced to global settings via state change handler
• Evidence: State structure at line 298044, sync handler in Wb() (renamed from Bb) at line 482021

What: Fixed message selection index calculation after restoring previous conversations

Details:

• Removed logic that attempted to find the previously selected message by ID after restoration
• Now resets to first message (index 0) when message list changes
• Prevents out-of-bounds errors when message IDs don't match after restoration
• Evidence: Function w11() at line 439713 (simplified from I11() at line 439313)

What: Removed Halloween-themed ASCII art mascot variants

Details:

• Removed functions: do1(), cg5(), pg5(), ig5(), ag5()
• Removed feature flag check: tengu_halloween
• Removed environment variable check: CLAUDE_CODE_FORCE_HALLOWEEN
• Simplified mascot rendering to single implementation (function Ft1() at line 444608)
• The Halloween mascots featured witch hats with green decoration that were displayed during October/November 2024
• Evidence: Diff lines 222-416 (removed functions), new simplified function at line 444608

What: Consolidated prompt template definitions into reusable variable references

Details:

• TodoWrite tool description moved to variable HoB/DoB at line 205368
• ExitPlanMode tool description moved to variable DR2/KR2 at line 434031
• Reduces duplication and makes prompt updates easier to maintain
• Evidence: Variables removed from v2.0.31 lines 205049-220 and 433633-253, re-added as variables in v2.0.32

• SDK Version: Updated from 0.60.0 to 0.66.0
• Variable Dh at line 81689

• Mk0 at line 83970: "\n\nHuman:" prompt separator
• Ok0 at line 83971: "\n\nAssistant:" prompt separator

• b10 at line 481665: New initialization block for path-related utilities