---

name: optimize-readability-system description: "Interactive guide for tuning the unmangleJS Readability Scoring System(readability.ts) and Semantic Scoring System (semantic.ts). Use when Golden Set validation fails, when scoring heuristics need adjustment, or when improving obfuscation detection accuracy. Emphasizes experiment-driven development: run tests, analyze data, draw conclusions - never guess blindly."

Optimize Readability/Semantic Scoring System

Interactive guide for tuning the unmangleJS scoring systems.

Core Philosophy

"Do experiments, get data, draw conclusions. Don't guess blindly."

This is the golden rule for all optimization work.

---

System Overview

1. Semantic Scoring (src/utils/semantic.ts)

Analyzes names and structures for semantic meaning.

Key Functions:

• scoreKeySemantic(key) - Score a single identifier (0-2.5 average)
• getSemanticThreshold(keyLength) - Get length-based threshold (1.0-1.5)
• isObjectSemantic() / isSubjectSemantic() - Judge semantic content

2. Readability Scoring (src/utils/readability.ts)

Decides whether to simplify AST subtrees.

7 Signals:

• semantic-* (5 signals, veto: true) - Protect semantic content
• cognitive-density - Cognitive load (0-1)
• linearity-violation - Linearity (0-1)

Decision Rules:

1. Semantic Veto - One-vote veto power
2. Dual Confirmation - cognitive < 0.4 AND linearity < 0.4
3. Average Fallback - threshold: 0.5

3. Golden Set (tests/readability/golden/)

Constitutional test suite defining system values.

• should-simplify.json - Obfuscated samples (must simplify)
• should-not-simplify.json - Semantic samples (must preserve)
• Success metric: 100% pass rate (excluding undecided)

---

When to Use This Skill

Scenario A: Golden Set Already Failing

CopyCurrent state: Golden Set validation fails
Action: Use this skill directly to diagnose and fix

Scenario B: Adding New Samples

CopyCurrent state: Golden Set passes, but found edge case
Action: Follow "Adding New Samples" flow below

Scenario C: New Sample Causes Failure

CopyCurrent state: Golden Set passes → Add sample → Fails
Action: Use this skill starting from Step 2 (Analyze Failures)

---

Optimization Workflow

Step 1: Run Golden Set

Copynpm test -- tests/readability/golden.test.ts

Expected: 50/50 passed (100%)

If failing → Continue to Step 2

Step 2: Analyze Failures

Use professional analysis tools (DO NOT write temporary scripts):

Tool 1: readability-analyzer.js

Purpose: Analyze code and output detailed readability scoring reports

Usage:

Copy# Build first
npm run build

# Analyze a file
node tools/readability-analyzer.js <file>

# From stdin
echo 'const x = { a: 1 };' | node tools/readability-analyzer.js

# NEW: Filter by node type
node tools/readability-analyzer.js <file> --types SequenceExpression
node tools/readability-analyzer.js <file> -t seq,obj,cond  # Multiple types
node tools/readability-analyzer.js --list-types  # Show available types

Output:

• Part 1: Decision Summary (should simplify? yes/no, threshold check)
• Part 2: All Signals (cognitive-density, linearity-violation, semantic-*)
• Part 3: Veto Check (which signal triggered veto)
• Part 4: Decision Logic (which rules passed/failed)

When to use:

• Any Golden Set test failure
• Debugging shouldSimplify() behavior
• Understanding why veto was triggered

Tool 2: semantic-analyzer.js

Purpose: Analyze semantic structures (Object, Array, Function, Variable)

Usage:

Copynpm run build
node tools/semantic-analyzer.js <file>
echo 'const obj = { a: 1 };' | node tools/semantic-analyzer.js

# NEW: Filter by node type
node tools/semantic-analyzer.js <file> --types ObjectExpression
node tools/semantic-analyzer.js <file> -t obj,arr,var  # Multiple types
node tools/semantic-analyzer.js --list-types  # Show available types

Supported Node Types:

• ObjectExpression - 4-part analysis
• ArrayExpression - 2-part analysis
• Function (Declaration/Expression/Arrow) - 2-part analysis
• VariableDeclarator - 2-part analysis (simple + destructuring)

When to use:

• Object/array/function/variable semantic analysis
• Debugging isSubjectSemantic() behavior
• Checking individual scores with dynamic thresholds

Step 3: Answer Key Questions

For each failing sample:

1. Which signals are causing the wrong decision?
2. Are semantic signals vetoing correctly?
3. Is cognitive/linearity score appropriate?
4. For semantic nodes: Are individual name scores correct?

Step 4: Form Hypothesis

Based on data, form a specific hypothesis.

Examples:

• "Short obfuscated names score too high because entropy threshold is too lenient"
• "Object linearity scores (0.9) prevent deletion, should be 0.3"

Step 5: Test with CLI

CRITICAL: Always test with real CLI behavior.

Copy# Test obfuscated pattern
cat > tmp/test.js << 'EOF'
const _0x1 = { UoAwJ: "x", YXLlx: "y" };
EOF
./bin/unmanglejs.js tmp/test.js -p
# Expected: Object should be inlined

# Test semantic pattern
cat > tmp/test2.js << 'EOF'
const config = { UNKNOWN: 0, OK: 1 };
EOF
./bin/unmanglejs.js tmp/test2.js -p
# Expected: Object should be preserved

Interpret Results:

• Obfuscated code NOT transformed → False negative (too conservative)
• Semantic code IS transformed → False positive (too aggressive)

Step 6: Apply Minimal Fix

Change one thing at a time. Validate after each change.

Example fixes:

Copy// Fix 1: Adjust entropy threshold
let threshold: number;
if (key.length <= 4) threshold = 2.5;  // Very strict
else if (key.length <= 6) threshold = 3.0;  // Strict
else threshold = 3.5;  // Normal

// Fix 2: Add semantic shield check
if (t.isCallExpression(expr) && t.isIdentifier(expr.callee)) {
  const score = scoreKeySemantic(expr.callee.name);
  if (score.average >= 1.5) semanticCount++;
}

// Fix 3: Add obfuscation penalty
if (metrics.semanticRatio < 0.3 && metrics.distinctIdentifiers >= 2) {
  score -= 0.3;
}

Step 7: Validate Full Suite

Copynpm test
# Expected: All tests passing

---

Common Issues

Issue: Golden Set < 100% Pass Rate

Diagnosis:

1. Run tools on failing samples
2. Categorize: too conservative OR too aggressive
3. Identify root cause

Common Fixes:

• Too conservative → Raise threshold, add penalties
• Too aggressive → Lower threshold, strengthen veto

Issue: Semantic Names Not Protected

Symptom: if(initialize(), validate()) gets flattened

Root Cause: Semantic shield missing pattern OR names score below threshold

Fix: Add missing pattern check (see Step 6 examples)

Issue: Short Obfuscated Names Score High

Symptom: { UoAwJ: "x" } scores 1.25 (semantic!)

Root Cause: Short names have naturally low entropy

Fix: Use stricter thresholds for short keys (see Step 6 examples)

Issue: Object Linearity Too High

Symptom: Objects score 0.9 linearity, preventing deletion

Fix: Object linearity shouldn't affect deletion

Copyif (path.isObjectExpression()) {
  return 0.3;  // Low score = "linearity doesn't apply"
}

Issue: Pass Behavior Differs by Context

Symptom: for(a=1, b=2) behaves differently than if(a(), b())

Explanation: Some passes don't extract from certain parent contexts by design

Fix: Understand pass behavior first, don't assume uniform behavior

---

Golden Set Management

Adding New Samples

When you encounter a new edge case:

Step 1: Validate current state (100% pass)

Step 2: Test new sample with tools

Copynpm run build
node tools/readability-analyzer.js tmp/new-sample.js

Step 3: Categorize

• Clearly obfuscated → should-simplify.json
• Clearly semantic → should-not-simplify.json
• Ambiguous → Mark as "expected": "undecided"

Step 4: Add to appropriate file

Copycat >> tests/readability/golden/should-simplify.json << 'EOF'
{
  "id": "descriptive-id",
  "code": "const x = <obfuscated code>;",
  "expected": true,
  "invariant": false,
  "reason": "obfuscated pattern description"
}
EOF

Step 5: Validate

Copynpm test -- tests/readability/golden.test.ts

If tests pass: Done ✅

If tests fail: New sample revealed system defect → Use optimization workflow to fix

Invariant Levels

• invariant: true - System-defining values, should never change
• invariant: false - Heuristic cases, can evolve
• expected: "undecided" - Ambiguous boundary cases

---

Best Practices

1. Experiment-Driven Development

ALWAYS:

1. Run tests first
2. Use professional tools for analysis
3. Identify specific failures
4. Form hypothesis
5. Test with CLI
6. Apply minimal fix
7. Validate all tests

NEVER:

• ❌ Guess what the issue might be
• ❌ Tune thresholds without data
• ❌ Write temporary scripts when tools are available
• ❌ Make broad changes for specific failures

2. Keep Tools in Sync

When modifying semantic.ts or readability.ts, update analysis tools:

1. Adding new signals → Update readability-analyzer.js
2. Adding new semantic functions → Update semantic-analyzer.js
3. Changing thresholds → Update tool output and docs

Test tools after changes:

Copynpm run build
echo 'const x = { a: 1 };' | node tools/semantic-analyzer.js

3. Minimal Diff Principle

• Change one thing at a time
• Validate after each change
• Keep changelog of what worked

---

Quick Reference

Key Files

Source:

• src/utils/semantic.ts - Semantic scoring
• src/utils/readability.ts - Readability decision

Tests:

• tests/readability/golden/*.json - Golden Set samples
• tests/readability/golden.test.ts - Validator

Tools:

• tools/readability-analyzer.js - Readability analysis
• tools/semantic-analyzer.js - Semantic analysis

Key Constants

Copy// readability.ts
const DEFAULT_THRESHOLD = 0.5;
const COGNITIVE_CONFIRM = 0.4;
const LINEARITY_CONFIRM = 0.4;

// semantic.ts - Dynamic thresholds
function getSemanticThreshold(keyLength: number): number {
  if (keyLength <= 3) return 1.0;   // Very short
  if (keyLength <= 5) return 1.25;  // Short
  if (keyLength <= 7) return 1.4;   // Medium
  return 1.5;                       // Standard
}

// semantic.ts - Fixed thresholds
const OBJECT_SEMANTIC_THRESHOLD = 1.4;

Common Commands

Copy# Build (required before using tools)
npm run build

# Run Golden Set
npm test -- tests/readability/golden.test.ts

# Run all tests
npm test

# Analysis tools
node tools/readability-analyzer.js <file>
node tools/semantic-analyzer.js <file>

# With type filtering
node tools/readability-analyzer.js <file> -t seq,obj
node tools/semantic-analyzer.js <file> -t obj,arr,var

# List available types
node tools/readability-analyzer.js --list-types
node tools/semantic-analyzer.js --list-types

Tool Type Filters

readability-analyzer:

• SequenceExpression (seq, sequence)
• ObjectExpression (obj, object)
• ConditionalExpression (cond, conditional)
• Identifier (id, identifier)

semantic-analyzer:

• ObjectExpression (obj, object)
• ArrayExpression (arr, array)
• FunctionDeclaration (func-decl)
• FunctionExpression (func-expr)
• ArrowFunctionExpression (arrow)
• VariableDeclarator (var, variable)

Success Metrics

• Golden Set: 50/50 passed (100%)
• All tests: 976/976 passing
• Zero false positives (semantic protected)
• Minimal false negatives (obfuscated caught)

---

Summary

Golden Rule: Do experiments, get data, draw conclusions.

Key Principles:

1. Always start with Golden Set validation
2. Use professional tools for analysis
3. Apply minimal, targeted fixes
4. Validate with full test suite
5. Keep tools in sync with implementation

Remember:

• The system is heuristic, not a correctness engine
• Golden Set is the constitution
• Veto provides precision, scoring provides recall
• Context matters - test in real scenarios