Skip to content

TECHNICAL.md

Latest commit

 

History

History
128 lines (91 loc) · 5.28 KB

TECHNICAL.md

File metadata and controls

128 lines (91 loc) · 5.28 KB

Claude Code (Tengu) Processing Words Analysis - Technical Documentation

A comprehensive analysis of all processing words used in the Claude Code CLI (@anthropic-ai/claude-code) across its version history.

Overview

This repository contains tools and data analyzing the whimsical "processing words" displayed by Claude Code while it's thinking. These gerund-form verbs (words ending in -ing) add personality to the CLI experience.

Key Findings

  • 139 versions analyzed (from v0.2.9 to v1.0.48)
  • 100% extraction success rate
  • 3 distinct phases of processing words:
    1. 56 words (v0.2.9 - v0.2.41): Original set missing "Pontificating"
    2. 57 words (v0.2.42 - v1.0.28): Standard set
    3. 90 words (v1.0.29 - current): Expanded set with 33 new whimsical additions

Files

Data Files

  • processing-words-all-versions.json - Comprehensive data for all versions
  • processing-words-history.json - Simplified word arrays per version
  • processing-words-analysis-report.json - Summary statistics and insights
  • known-processing-words.json - Dictionary of all 90 known processing words

Scripts

  • extract-processing-words-robust.js - Main extraction script using pattern matching
  • extract-batch.js - Batch processing script for handling large version ranges
  • combine-batches.js - Combines batch results into final datasets

Major Changes in Processing Words

Version 0.2.42

  • Added "Pontificating" to complete the standard set (56 → 57 words)

Version 1.0.29

  • Major expansion adding 33 new words (57 → 90 words)
  • New words include whimsical additions like:
    • Flibbertigibbeting
    • Discombobulating
    • Wizarding
    • Booping
    • And 29 more creative verbs

Technical Details

Processing Words Storage

  • Processing words are stored in minified JavaScript variables
  • Variable names change with each build (e.g., Lj6, pz5, zw5, jy6)
  • The 90 processing words are all positive/whimsical verbs - no concerning words like "Terminating" or "Crashing" appear in the dataset
  • All processing words are unique with no duplicates in the array

Dynamic Spinner Words System

The CLI includes a sophisticated dynamic configuration system that can fetch additional "spinner words" from Anthropic's servers at runtime:

Key Components:

  1. Static Processing Words: Hardcoded array of 90 words
  2. Dynamic Configuration: Fetches additional words via the tengu_spinner_words configuration key
  3. Merger System: Combines static and dynamic words: [...hardcodedWords, ...dynamicWords]
  4. Default Fallback: Uses empty array {words: []} if dynamic fetch fails

Implementation Details:

  • Uses React hooks for state management
  • Includes environment variable checks to disable in certain contexts (e.g., DISABLE_TELEMETRY)
  • Random word selection every 1000ms for display
  • Memoization prevents unnecessary re-fetching
  • 5-minute cache TTL for configuration data

"Tengu" - Internal Codename

"Tengu" appears to be the internal project codename for Claude Code CLI:

  • Configuration keys: tengu_spinner_words for dynamic words
  • Telemetry events: tengu_oauth_*, tengu_config_*, tengu_preflight_check_failed
  • Used consistently throughout the codebase for feature flags and analytics

The name likely references the Japanese mythological creature known for wisdom and mischief, fitting the CLI's whimsical personality.

Dynamic Configuration Benefits

The dynamic system powered by Statsig provides several advantages:

  1. Real-time Updates: Update spinner words without releasing new CLI versions
  2. A/B Testing: Test different word sets with user segments
  3. Seasonal Content: Add holiday or event-specific words dynamically
  4. User Personalization: Customize words based on user type (pro/max/enterprise)
  5. Gradual Rollout: Test new words with small user percentages first
  6. Analytics: Track which words appear most often and user engagement

This architecture explains why static analysis finds only 90 unique words - the system is designed to be extensible through server-side configuration.

Technical Architecture

The spinner word system follows this flow:

  1. Initialization: CLI starts and loads static words (90 items)
  2. Dynamic Fetch: Attempts to fetch additional words from Statsig
  3. Merge: Combines static and dynamic word arrays
  4. Selection: Randomly picks words from combined list
  5. Display: Shows selected word in spinner component
  6. Fallback: Uses only static words if dynamic fetch fails

This ensures reliable operation even without network connectivity while enabling real-time customization when connected.

Usage

To run the extraction on new versions:

# Extract all versions from 0.2.9 onwards
node extract-processing-words-robust.js

# Or process in batches
node extract-batch.js 1  # Process batch 1 (first 30 versions)
node extract-batch.js 2  # Process batch 2 (next 30 versions)
# etc...

# Combine batch results
node combine-batches.js

Variable Name Patterns

The analysis discovered 119 unique variable names used across versions, showing how the minified code changes with each build. Examples include: Aq6, At6, C31, CL3, Cu5, D31, DE5, Dt6, Ed5, F31, and many more.

License

This analysis is provided for educational and research purposes. The Claude Code CLI is property of Anthropic.