paif/gravity.js

'use strict';

const identity = require('./identity');

// Common drift patterns that indicate generic assistant mode
const DRIFT_PATTERNS = {
  // Phrases that suggest subservient assistant posture
  subservience: [
    /how can i help/i,
    /how may i assist/i,
    /i'm happy to help/i,
    /i'd be delighted to/i,
    /at your service/i,
    /what would you like me to do/i,
    /just let me know/i,
    /feel free to ask/i
  ],

  // Over-apologizing
  over_apology: [
    /i apologize for/i,
    /sorry for the confusion/i,
    /my apologies/i,
    /i'm sorry that/i,
    /please forgive/i
  ],

  // Fake enthusiasm
  fake_enthusiasm: [
    /excited to/i,
    /thrilled to/i,
    /looking forward to/i,
    /can't wait to/i
  ],

  // Hedge words that undermine confidence
  hedging: [
    /i think (that|we|you|this)/i,
    /maybe we should/i,
    /perhaps (you|we|it)/i,
    /it seems like/i,
    /kind of/i,
    /sort of/i
  ],

  // People-pleasing endings
  people_pleasing: [
    /please let me know/i,
    /don't hesitate to/i,
    /if you have any questions/i,
    /i hope that helps/i,
    /is there anything else/i
  ],

  // Excessive punctuation (for precise tones)
  excessive_punctuation: /!{2,}/
};

// Check text for drift patterns
function detectDrift(text, agentId) {
  const issues = [];
  const id = identity.loadIdentity(agentId);

  // 1. Check taboo phrases from identity
  if (id?.immutable?.voice?.taboo_phrases) {
    for (const taboo of id.immutable.voice.taboo_phrases) {
      const regex = new RegExp(taboo.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'), 'i');
      if (regex.test(text)) {
        issues.push({
          type: 'taboo_phrase',
          pattern: taboo,
          message: `Used forbidden phrase: "${taboo}"`,
          severity: 'high'
        });
      }
    }
  }

  // 2. Check generic drift patterns
  for (const [category, patterns] of Object.entries(DRIFT_PATTERNS)) {
    if (category === 'excessive_punctuation') {
      if (patterns.test(text)) {
        issues.push({
          type: 'drift',
          category,
          pattern: '!!+',
          message: 'Excessive exclamation marks suggest fake enthusiasm',
          severity: 'medium'
        });
      }
      continue;
    }

    for (const pattern of patterns) {
      if (pattern.test(text)) {
        issues.push({
          type: 'drift',
          category,
          pattern: pattern.toString(),
          message: `Generic assistant pattern detected: ${category}`,
          severity: 'medium'
        });
        break; // Only report once per category
      }
    }
  }

  // 3. Tone-specific checks
  if (id?.immutable?.voice?.tone === 'precise') {
    // Precise agents should avoid exclamation marks entirely
    if (text.includes('!')) {
      issues.push({
        type: 'tone_deviation',
        category: 'precise_voice',
        message: 'Precise tone suggests avoiding exclamation marks',
        severity: 'low'
      });
    }
  }

  if (id?.immutable?.voice?.tone === 'academic') {
    // Academic tone should avoid contractions
    const contractions = /\b(don't|can't|won't|shouldn't|wouldn't|couldn't|isn't|aren't|wasn't|weren't|haven't|hasn't|hadn't)\b/i;
    if (contractions.test(text)) {
      issues.push({
        type: 'tone_deviation',
        category: 'academic_voice',
        message: 'Academic tone suggests avoiding contractions',
        severity: 'low'
      });
    }
  }

  return {
    drift_detected: issues.length > 0,
    issues,
    drift_score: calculateDriftScore(issues),
    checks_performed: ['taboo_phrases', 'generic_patterns', 'tone_alignment']
  };
}

// Calculate a drift score (0-100, higher = more drift)
function calculateDriftScore(issues) {
  if (issues.length === 0) return 0;

  const weights = {
    high: 30,
    medium: 15,
    low: 5
  };

  const score = issues.reduce((sum, issue) => sum + (weights[issue.severity] || 5), 0);
  return Math.min(100, score);
}

// Generate a realignment prompt
function generateRealignmentPrompt(agentId, driftResult) {
  const id = identity.loadIdentity(agentId);
  if (!id) {
    return {
      realignment_needed: false,
      prompt: null,
      reason: 'No identity found'
    };
  }

  if (!driftResult.drift_detected) {
    return {
      realignment_needed: false,
      prompt: null,
      reason: 'No drift detected'
    };
  }

  const issues = driftResult.issues.map(i => `- ${i.message}`).join('\n');

  const prompt = `═══════════════════════════════════════════════════════════════
PERSONA GRAVITY CHECK — DRIFT DETECTED
═══════════════════════════════════════════════════════════════

Your response deviated from your identity:

${issues}

┌─────────────────────────────────────────────────────────────┐
│  REMEMBER WHO YOU ARE                                       │
└─────────────────────────────────────────────────────────────┘

PURPOSE: ${id.immutable.purpose.statement}

PRIMARY VALUE: ${id.immutable.values.primary}
SUPPORTING VALUES: ${(id.immutable.values.secondary || []).join(', ')}

VOICE: ${id.immutable.voice.tone}
${id.immutable.voice.quirks ? `QUIRKS: ${id.immutable.voice.quirks.join(', ')}` : ''}

${id.immutable.voice.taboo_phrases.length > 0 ? `AVOID SAYING: ${id.immutable.voice.taboo_phrases.slice(0, 5).join(', ')}${id.immutable.voice.taboo_phrases.length > 5 ? '...' : ''}` : ''}

NON-NEGOTIABLES: ${(id.immutable.values.non_negotiables || []).join(', ')}

───────────────────────────────────────────────────────────────
RECENTER. Respond from your authentic self.
═══════════════════════════════════════════════════════════════`;

  return {
    realignment_needed: true,
    prompt,
    drift_score: driftResult.drift_score,
    reason: 'Drift detected, realignment required'
  };
}

// Perform full gravity check with optional realignment
function gravityCheck(agentId, text, options = {}) {
  const { autoRealign = false, updateStats = true } = options;

  // Detect drift
  const driftResult = detectDrift(text, agentId);

  // Generate realignment if needed
  let realignment = null;
  if (driftResult.drift_detected && autoRealign) {
    realignment = generateRealignmentPrompt(agentId, driftResult);
  }

  // Update stats if requested
  if (updateStats) {
    const id = identity.loadIdentity(agentId);
    if (id) {
      const passed = driftResult.issues.filter(i => i.severity === 'low').length;
      const failed = driftResult.issues.filter(i => i.severity !== 'low').length;

      identity.updateMutableState(agentId, {
        state: {
          drift_checks_passed: (id.mutable.state?.drift_checks_passed || 0) + (driftResult.drift_detected ? 0 : 1),
          drift_checks_failed: (id.mutable.state?.drift_checks_failed || 0) + (driftResult.drift_detected ? 1 : 0)
        }
      });
    }
  }

  return {
    agent_id: agentId,
    drift_detected: driftResult.drift_detected,
    drift_score: driftResult.drift_score,
    issues: driftResult.issues,
    checks_performed: driftResult.checks_performed,
    realignment: realignment,
    timestamp: new Date().toISOString()
  };
}

// Quick check for specific patterns (lightweight)
function quickCheck(text, agentId) {
  const id = identity.loadIdentity(agentId);
  if (!id) return { passed: true, issues: [] };

  const taboos = id.immutable.voice.taboo_phrases || [];
  const issues = [];

  for (const taboo of taboos) {
    const regex = new RegExp(taboo.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'), 'i');
    if (regex.test(text)) {
      issues.push({ type: 'taboo', phrase: taboo });
    }
  }

  return {
    passed: issues.length === 0,
    issues
  };
}

module.exports = {
  detectDrift,
  generateRealignmentPrompt,
  gravityCheck,
  quickCheck,
  DRIFT_PATTERNS
};