test-tokens.js

// Test dynamic max_tokens calculation functions

// Mock the functions (copied from openrouter.ts)
function calculateSafeMaxTokens(contextLength, estimatedInputTokens, maxCompletionTokens) {
  // Leave some buffer for safety (10% of context length or minimum 1000 tokens)
  const safetyBuffer = Math.max(Math.floor(contextLength * 0.1), 1000);
  
  // Calculate available tokens for output
  const availableTokens = contextLength - estimatedInputTokens - safetyBuffer;
  
  // Respect model's max completion tokens if available
  const modelMaxTokens = maxCompletionTokens || availableTokens;
  
  // Use the smaller of available tokens or model's max completion tokens
  const safeMaxTokens = Math.min(availableTokens, modelMaxTokens);
  
  // Ensure we don't go below a reasonable minimum
  const finalMaxTokens = Math.max(safeMaxTokens, 1000);
  
  console.log('🧮 Token calculation:', {
    contextLength,
    estimatedInputTokens,
    safetyBuffer,
    availableTokens,
    modelMaxTokens,
    finalMaxTokens
  });
  
  return finalMaxTokens;
}

function estimateTokenCount(text) {
  return Math.ceil(text.length / 4);
}

console.log("Testing dynamic max_tokens calculation:\n");

// Test the exact failing scenario from the error message
console.log("1. DeepSeek V3 - Failing Scenario (Original):");
const failingContextLength = 131072;
const failingInputTokens = 3499;
const failingMaxTokens = 128000; // What we were requesting before

console.log(`   Context Length: ${failingContextLength.toLocaleString()}`);
console.log(`   Input Tokens:   ${failingInputTokens.toLocaleString()}`);
console.log(`   Requested Out:  ${failingMaxTokens.toLocaleString()}`);
console.log(`   Total:          ${(failingInputTokens + failingMaxTokens).toLocaleString()}`);
console.log(`   Over Limit:     ${(failingInputTokens + failingMaxTokens) - failingContextLength} tokens ❌\n`);

// Test with our new calculation
console.log("2. DeepSeek V3 - Fixed with Dynamic Calculation:");
const dynamicMaxTokens = calculateSafeMaxTokens(failingContextLength, failingInputTokens, 8192);
const newTotal = failingInputTokens + dynamicMaxTokens;

console.log(`   Context Length: ${failingContextLength.toLocaleString()}`);
console.log(`   Input Tokens:   ${failingInputTokens.toLocaleString()}`);
console.log(`   Dynamic Max:    ${dynamicMaxTokens.toLocaleString()}`);
console.log(`   New Total:      ${newTotal.toLocaleString()}`);
console.log(`   Within Limit:   ${newTotal <= failingContextLength ? '✅' : '❌'}`);
console.log(`   Safety Margin:  ${failingContextLength - newTotal} tokens\n`);

// Test with smart HTML context (reduced input)
console.log("3. DeepSeek V3 - With Smart HTML Context Reduction:");
const reducedInputTokens = Math.floor(failingInputTokens * 0.3); // 70% reduction from smart context
const smartMaxTokens = calculateSafeMaxTokens(failingContextLength, reducedInputTokens, 8192);
const smartTotal = reducedInputTokens + smartMaxTokens;

console.log(`   Context Length: ${failingContextLength.toLocaleString()}`);
console.log(`   Reduced Input:  ${reducedInputTokens.toLocaleString()} (was ${failingInputTokens})`);
console.log(`   Smart Max:      ${smartMaxTokens.toLocaleString()}`);
console.log(`   Smart Total:    ${smartTotal.toLocaleString()}`);
console.log(`   Within Limit:   ${smartTotal <= failingContextLength ? '✅' : '❌'}`);
console.log(`   Safety Margin:  ${failingContextLength - smartTotal} tokens\n`);

console.log("✅ Dynamic max_tokens calculation successfully prevents context overflow!");