@hawon/nexus 0.2.0 → 0.3.0

package/src/skills/context-engine.ts

@@ -1,863 +0,0 @@
-/**
- * Context Engine — Algorithmic Language Understanding
- *
- * A mini-LLM built with pure algorithms. No neural networks, no API calls.
- * Understands conversation context, intent, and flow through:
- *
- * 1. CONTEXTUAL STATE MACHINE
- *    Tracks conversation phase: greeting → problem_statement → exploration →
- *    execution → verification → completion
- *
- * 2. BAYESIAN INTENT CLASSIFIER
- *    P(intent | words, state, history) using Bayes theorem
- *    Prior: conversation state, recent intents
- *    Likelihood: word probabilities per intent class
- *    Evidence: full message + surrounding context
- *
- * 3. SLIDING ATTENTION WINDOW
- *    Dynamically weights which past messages matter most for current understanding
- *    Recency bias + relevance scoring + topic coherence
- *
- * 4. CONCEPT GRAPH
- *    Builds a graph of entities and relationships mentioned in conversation
- *    Tracks: what files, what errors, what goals, what tools, what decisions
- *
- * 5. AUTO-TRIGGER
- *    Automatically detects when a skill-worthy pattern has completed
- *    No explicit commands needed — watches the conversation flow
- */
-
-import type { ParsedMessage, ToolCall } from "../parser/types.js";
-
-// ═══════════════════════════════════════════════════════════════════
-// TYPES
-// ═══════════════════════════════════════════════════════════════════
-
-/** Conversation phase — tracked by the state machine. */
-export type ConversationPhase =
-  | "idle"              // No active task
-  | "problem_stated"    // User described a problem/goal
-  | "exploring"         // Discussing approaches, asking questions
-  | "executing"         // Claude is running tools, making changes
-  | "verifying"         // Checking if the result works
-  | "completed"         // Task done, user acknowledged
-  | "failed"            // Task didn't work
-  | "pivoting";         // Changing approach mid-task
-
-/** Fine-grained intent — what the user wants RIGHT NOW. */
-export type GranularIntent = {
-  primary: IntentType;
-  confidence: number;
-  /** Secondary intents (e.g., fix_bug + security). */
-  secondary: IntentType[];
-  /** What specifically (extracted from context). */
-  target: string;
-  /** Why (inferred from conversation history). */
-  motivation: string;
-  /** How urgent/important (from language intensity). */
-  urgency: "low" | "medium" | "high" | "critical";
-};
-
-export type IntentType =
-  | "fix_bug" | "build_feature" | "refactor" | "investigate"
-  | "configure" | "deploy" | "test" | "security_audit"
-  | "code_review" | "explain" | "optimize" | "migrate"
-  | "debug" | "cleanup" | "document" | "monitor"
-  | "acknowledge" | "redirect" | "clarify" | "continue";
-
-/** A node in the concept graph. */
-export type ConceptNode = {
-  id: string;
-  type: "file" | "error" | "tool" | "goal" | "decision" | "entity" | "concept";
-  label: string;
-  /** How many times mentioned. */
-  weight: number;
-  /** When first/last mentioned. */
-  firstSeen: number;
-  lastSeen: number;
-  /** Sentiment: positive/negative/neutral. */
-  sentiment: number; // -1 to 1
-};
-
-export type ConceptEdge = {
-  from: string;
-  to: string;
-  relation: "causes" | "fixes" | "modifies" | "requires" | "produces" | "related";
-  weight: number;
-};
-
-export type ConceptGraph = {
-  nodes: Map<string, ConceptNode>;
-  edges: ConceptEdge[];
-};
-
-/** Attention-weighted message. */
-export type AttentionMessage = {
-  message: ParsedMessage;
-  /** How relevant this message is to the current context. 0-1. */
-  attention: number;
-  /** What phase the conversation was in when this was said. */
-  phase: ConversationPhase;
-  /** Intent at this point. */
-  intent: GranularIntent;
-};
-
-/** Auto-detected skill trigger event. */
-export type SkillTrigger = {
-  type: "task_completed" | "pattern_repeated" | "lesson_learned" | "pivot_point";
-  /** Episode messages that form this skill. */
-  messages: AttentionMessage[];
-  /** Why this was triggered. */
-  reason: string;
-  /** Quality estimate. */
-  quality: number;
-};
-
-/** Full engine state. */
-export type EngineState = {
-  phase: ConversationPhase;
-  phaseHistory: ConversationPhase[];
-  currentIntent: GranularIntent;
-  intentHistory: GranularIntent[];
-  attentionWindow: AttentionMessage[];
-  conceptGraph: ConceptGraph;
-  triggers: SkillTrigger[];
-  /** Running topic summary. */
-  topicStack: string[];
-  /** Turn counter. */
-  turnCount: number;
-};
-
-// ═══════════════════════════════════════════════════════════════════
-// BAYESIAN INTENT CLASSIFIER
-// ═══════════════════════════════════════════════════════════════════
-
-/**
- * Word-intent probability table.
- * P(word | intent) — how likely each word is given an intent.
- * Built from domain knowledge, not training data.
- */
-const INTENT_WORD_PROBS: Record<IntentType, Record<string, number>> = {
-  fix_bug: {
-    "fix": 0.9, "bug": 0.9, "error": 0.85, "broken": 0.85, "crash": 0.8,
-    "고치": 0.9, "수정": 0.85, "에러": 0.85, "버그": 0.9, "오류": 0.8,
-    "안돼": 0.7, "안되": 0.7, "doesn": 0.6, "work": 0.3, "fail": 0.8,
-    "wrong": 0.7, "issue": 0.6, "problem": 0.6, "문제": 0.7, "해결": 0.7,
-    "resolve": 0.7, "debug": 0.7, "디버그": 0.7, "왜": 0.3,
-  },
-  build_feature: {
-    "만들": 0.9, "생성": 0.85, "create": 0.9, "build": 0.85, "add": 0.7,
-    "추가": 0.8, "implement": 0.9, "구현": 0.85, "new": 0.5, "새로": 0.7,
-    "write": 0.6, "작성": 0.7, "개발": 0.8, "develop": 0.8, "feature": 0.8,
-    "기능": 0.8, "만들어": 0.9, "setup": 0.5, "init": 0.5,
-  },
-  refactor: {
-    "refactor": 0.95, "리팩토": 0.95, "정리": 0.7, "clean": 0.6,
-    "improve": 0.6, "개선": 0.7, "restructure": 0.9, "simplify": 0.8,
-    "organize": 0.7, "reduce": 0.5, "optimize": 0.4, "rename": 0.6,
-  },
-  investigate: {
-    "찾아": 0.7, "find": 0.5, "search": 0.6, "조사": 0.8, "분석": 0.8,
-    "analyze": 0.8, "check": 0.5, "확인": 0.6, "look": 0.4, "examine": 0.8,
-    "inspect": 0.8, "어디": 0.5, "뭐가": 0.5, "what": 0.3, "scan": 0.7,
-  },
-  configure: {
-    "설정": 0.9, "setup": 0.85, "install": 0.8, "설치": 0.85,
-    "config": 0.9, "configure": 0.9, "init": 0.7, "초기화": 0.7,
-    "환경": 0.6, "environment": 0.5, "연결": 0.5, "connect": 0.5,
-  },
-  deploy: {
-    "배포": 0.95, "deploy": 0.95, "publish": 0.9, "push": 0.5,
-    "release": 0.85, "올려": 0.7, "npm": 0.5, "ship": 0.7,
-  },
-  test: {
-    "테스트": 0.95, "test": 0.9, "검증": 0.8, "verify": 0.8,
-    "assert": 0.8, "spec": 0.7, "coverage": 0.7, "확인": 0.4,
-  },
-  security_audit: {
-    "보안": 0.95, "security": 0.95, "취약점": 0.95, "vulnerability": 0.9,
-    "exploit": 0.9, "hack": 0.7, "audit": 0.8, "감사": 0.5,
-    "injection": 0.85, "xss": 0.9, "csrf": 0.9, "ssrf": 0.9,
-  },
-  code_review: {
-    "리뷰": 0.9, "review": 0.85, "검토": 0.8, "봐봐": 0.6,
-    "봐줘": 0.6, "체크": 0.5, "pr": 0.6, "diff": 0.5,
-  },
-  explain: {
-    "설명": 0.9, "explain": 0.9, "알려": 0.8, "뭐야": 0.7,
-    "what": 0.4, "how": 0.5, "어떻게": 0.6, "why": 0.5, "왜": 0.5,
-    "이해": 0.8, "understand": 0.7, "mean": 0.5, "뜻": 0.7,
-  },
-  optimize: {
-    "최적화": 0.95, "optimize": 0.95, "빠르게": 0.7, "faster": 0.7,
-    "performance": 0.8, "성능": 0.8, "speed": 0.7, "slow": 0.6,
-    "느려": 0.6, "memory": 0.4, "efficient": 0.7,
-  },
-  migrate: {
-    "마이그레이션": 0.95, "migrate": 0.95, "이전": 0.5, "변환": 0.7,
-    "convert": 0.7, "upgrade": 0.7, "업그레이드": 0.7, "port": 0.6,
-  },
-  debug: {
-    "디버그": 0.95, "debug": 0.95, "로그": 0.5, "log": 0.4,
-    "trace": 0.7, "breakpoint": 0.8, "원인": 0.5, "cause": 0.5,
-  },
-  cleanup: {
-    "정리": 0.7, "cleanup": 0.8, "삭제": 0.6, "remove": 0.5,
-    "delete": 0.5, "unused": 0.7, "불필요": 0.7, "쓸데없": 0.6,
-  },
-  document: {
-    "문서": 0.85, "document": 0.8, "readme": 0.9, "docs": 0.8,
-    "comment": 0.5, "주석": 0.6, "api": 0.3, "설명": 0.4,
-  },
-  monitor: {
-    "모니터": 0.9, "monitor": 0.9, "watch": 0.6, "감시": 0.8,
-    "alert": 0.7, "status": 0.5, "health": 0.6, "확인": 0.3,
-  },
-  acknowledge: {
-    "ㅇㅇ": 0.95, "ㅇㅋ": 0.95, "ㄱㄱ": 0.9, "응": 0.8,
-    "ok": 0.7, "okay": 0.7, "sure": 0.7, "알겠": 0.8, "good": 0.6,
-    "네": 0.7, "yes": 0.6, "그래": 0.7,
-  },
-  redirect: {
-    "아니": 0.8, "말고": 0.9, "대신": 0.85, "아닌데": 0.9,
-    "no": 0.5, "wait": 0.7, "stop": 0.7, "instead": 0.85,
-    "잠깐": 0.8, "다시": 0.6, "바꿔": 0.8, "actually": 0.7,
-    "근데": 0.5, "그거": 0.3, "scratch": 0.7,
-  },
-  clarify: {
-    "뭐": 0.5, "무슨": 0.6, "어떤": 0.5, "which": 0.5,
-    "what": 0.4, "?": 0.3, "정확히": 0.7, "specifically": 0.7,
-    "구체적": 0.7, "어떻게": 0.4,
-  },
-  continue: {
-    "계속": 0.9, "continue": 0.85, "go": 0.4, "ㄱㄱ": 0.7,
-    "next": 0.6, "다음": 0.6, "진행": 0.8, "이어서": 0.85,
-    "해줘": 0.4, "하자": 0.5,
-  },
-};
-
-// Prior probabilities based on conversation phase
-const PHASE_INTENT_PRIOR: Record<ConversationPhase, Partial<Record<IntentType, number>>> = {
-  idle: { build_feature: 0.2, investigate: 0.15, explain: 0.15, configure: 0.1 },
-  problem_stated: { fix_bug: 0.3, investigate: 0.2, debug: 0.15, explain: 0.1 },
-  exploring: { investigate: 0.2, clarify: 0.2, explain: 0.15, redirect: 0.1 },
-  executing: { continue: 0.2, acknowledge: 0.15, redirect: 0.15, clarify: 0.1 },
-  verifying: { acknowledge: 0.2, fix_bug: 0.15, redirect: 0.15, continue: 0.1 },
-  completed: { build_feature: 0.15, continue: 0.15, acknowledge: 0.15, deploy: 0.1 },
-  failed: { fix_bug: 0.25, redirect: 0.2, debug: 0.15, investigate: 0.1 },
-  pivoting: { redirect: 0.2, build_feature: 0.15, fix_bug: 0.15, investigate: 0.1 },
-};
-
-const DEFAULT_PRIOR = 0.02;
-
-function classifyIntentBayesian(
-  text: string,
-  phase: ConversationPhase,
-  recentIntents: IntentType[],
-  conceptGraph: ConceptGraph,
-): GranularIntent {
-  const words = tokenize(text);
-  const scores: Record<string, number> = {};
-  const phasePrior = PHASE_INTENT_PRIOR[phase] ?? {};
-
-  for (const [intent, wordProbs] of Object.entries(INTENT_WORD_PROBS)) {
-    // Prior: from phase + recent intent momentum
-    let prior = phasePrior[intent as IntentType] ?? DEFAULT_PRIOR;
-
-    // Intent momentum: if same intent appeared recently, boost slightly
-    const recentCount = recentIntents.slice(-3).filter((i) => i === intent).length;
-    prior += recentCount * 0.05;
-
-    // Likelihood: P(words | intent)
-    let logLikelihood = 0;
-    let matchedWords = 0;
-    for (const word of words) {
-      const prob = wordProbs[word];
-      if (prob !== undefined) {
-        logLikelihood += Math.log(prob);
-        matchedWords++;
-      } else {
-        logLikelihood += Math.log(0.01); // Smoothing for unknown words
-      }
-    }
-
-    // Context boost: if concept graph has related entities
-    let contextBoost = 0;
-    if (intent === "fix_bug" && hasConceptType(conceptGraph, "error")) contextBoost += 0.3;
-    if (intent === "code_review" && hasConceptType(conceptGraph, "file")) contextBoost += 0.2;
-    if (intent === "security_audit" && hasConceptSentiment(conceptGraph, -0.5)) contextBoost += 0.3;
-
-    // Message length heuristic
-    let lengthBoost = 0;
-    if (intent === "acknowledge" && text.length < 10) lengthBoost += 0.5;
-    if (intent === "redirect" && text.length < 30) lengthBoost += 0.2;
-    if (intent === "explain" && text.includes("?")) lengthBoost += 0.3;
-
-    // Combine: log(prior) + logLikelihood + boosts
-    scores[intent] = Math.log(Math.max(prior, 0.001)) + logLikelihood +
-      (matchedWords > 0 ? contextBoost + lengthBoost : 0);
-  }
-
-  // Softmax-like normalization
-  const maxScore = Math.max(...Object.values(scores));
-  const expScores: Record<string, number> = {};
-  let sumExp = 0;
-  for (const [intent, score] of Object.entries(scores)) {
-    const exp = Math.exp(score - maxScore);
-    expScores[intent] = exp;
-    sumExp += exp;
-  }
-  if (sumExp === 0) sumExp = 1; // Prevent division by zero
-
-  // Sort by probability
-  const ranked = Object.entries(expScores)
-    .map(([intent, exp]) => ({ intent: intent as IntentType, prob: exp / sumExp }))
-    .sort((a, b) => b.prob - a.prob);
-
-  const primary = ranked[0];
-  const secondary = ranked.slice(1, 4).filter((r) => r.prob > 0.1).map((r) => r.intent);
-
-  // Extract target from context
-  const target = extractTarget(text, primary.intent, conceptGraph);
-  const motivation = inferMotivation(phase, recentIntents, conceptGraph);
-  const urgency = detectUrgency(text);
-
-  return {
-    primary: primary.intent,
-    confidence: primary.prob,
-    secondary,
-    target,
-    motivation,
-    urgency,
-  };
-}
-
-// ═══════════════════════════════════════════════════════════════════
-// CONTEXTUAL STATE MACHINE
-// ═══════════════════════════════════════════════════════════════════
-
-function transitionPhase(
-  current: ConversationPhase,
-  message: ParsedMessage,
-  intent: GranularIntent,
-): ConversationPhase {
-  const hasTools = (message.toolCalls?.length ?? 0) > 0;
-  const isUser = message.role === "user";
-
-  // Transition rules (state × signal → new state)
-  if (isUser) {
-    switch (current) {
-      case "idle":
-        if (["fix_bug", "build_feature", "investigate", "security_audit", "debug"].includes(intent.primary)) {
-          return "problem_stated";
-        }
-        if (intent.primary === "explain" || intent.primary === "clarify") return "exploring";
-        return "idle";
-
-      case "problem_stated":
-        if (intent.primary === "redirect") return "pivoting";
-        if (intent.primary === "clarify") return "exploring";
-        if (intent.primary === "acknowledge" || intent.primary === "continue") return "executing";
-        return "problem_stated";
-
-      case "exploring":
-        if (intent.primary === "acknowledge" || intent.primary === "continue") return "executing";
-        if (intent.primary === "redirect") return "pivoting";
-        return "exploring";
-
-      case "executing":
-        if (intent.primary === "redirect") return "pivoting";
-        if (intent.primary === "acknowledge") return "verifying";
-        if (["fix_bug", "build_feature"].includes(intent.primary)) return "problem_stated";
-        return "executing";
-
-      case "verifying":
-        if (intent.primary === "acknowledge") return "completed";
-        if (intent.primary === "redirect" || intent.primary === "fix_bug") return "failed";
-        return "verifying";
-
-      case "completed":
-        if (["fix_bug", "build_feature", "investigate", "configure", "deploy"].includes(intent.primary)) {
-          return "problem_stated"; // New task
-        }
-        return "idle";
-
-      case "failed":
-        if (intent.primary === "redirect") return "pivoting";
-        if (intent.primary === "fix_bug") return "problem_stated";
-        if (intent.primary === "acknowledge") return "executing"; // Try again
-        return "failed";
-
-      case "pivoting":
-        if (["fix_bug", "build_feature", "investigate"].includes(intent.primary)) return "problem_stated";
-        if (intent.primary === "acknowledge") return "executing";
-        return "exploring";
-    }
-  } else {
-    // Assistant messages
-    if (hasTools) return "executing";
-    if (current === "executing" && !hasTools) return "verifying";
-  }
-
-  return current;
-}
-
-// ═══════════════════════════════════════════════════════════════════
-// SLIDING ATTENTION WINDOW
-// ═══════════════════════════════════════════════════════════════════
-
-function computeAttention(
-  messages: AttentionMessage[],
-  currentMessage: ParsedMessage,
-  conceptGraph: ConceptGraph,
-): void {
-  if (messages.length === 0) return;
-
-  const currentKeywords = new Set(tokenize(currentMessage.content));
-  const now = messages.length;
-
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i];
-    const age = now - i;
-
-    // Recency: exponential decay
-    const recency = Math.exp(-age * 0.15);
-
-    // Relevance: keyword overlap with current message
-    const msgKeywords = tokenize(msg.message.content);
-    const overlap = msgKeywords.filter((w) => currentKeywords.has(w)).length;
-    const relevance = msgKeywords.length > 0 ? overlap / Math.sqrt(msgKeywords.length) : 0;
-
-    // Phase importance: decision points and pivots are always important
-    const phaseBoost = (msg.phase === "pivoting" || msg.phase === "failed") ? 0.3 : 0;
-
-    // User messages are slightly more important than assistant
-    const roleBoost = msg.message.role === "user" ? 0.1 : 0;
-
-    // Tool calls are important context
-    const toolBoost = (msg.message.toolCalls?.length ?? 0) > 0 ? 0.15 : 0;
-
-    msg.attention = Math.min(1,
-      recency * 0.4 +
-      relevance * 0.3 +
-      phaseBoost +
-      roleBoost +
-      toolBoost,
-    );
-  }
-}
-
-// ═══════════════════════════════════════════════════════════════════
-// CONCEPT GRAPH
-// ═══════════════════════════════════════════════════════════════════
-
-function updateConceptGraph(
-  graph: ConceptGraph,
-  message: ParsedMessage,
-  turnCount: number,
-): void {
-  const text = message.content;
-
-  // Extract file paths
-  const filePaths = text.match(/[\w./\\-]+\.\w{1,5}/g) ?? [];
-  for (const fp of filePaths) {
-    upsertConcept(graph, fp, "file", turnCount);
-  }
-
-  // Extract error patterns
-  const errorPatterns = text.match(/(?:error|Error|ERROR|에러|오류)[:\s].*?(?:\n|$)/g) ?? [];
-  for (const err of errorPatterns) {
-    const label = err.slice(0, 60).trim();
-    upsertConcept(graph, label, "error", turnCount, -0.7);
-  }
-
-  // Extract backtick entities
-  const backticks = text.match(/`([^`]+)`/g) ?? [];
-  for (const bt of backticks) {
-    const label = bt.replace(/`/g, "");
-    if (label.length > 1 && label.length < 50) {
-      upsertConcept(graph, label, "entity", turnCount);
-    }
-  }
-
-  // Extract tool calls as concepts
-  if (message.toolCalls) {
-    for (const tc of message.toolCalls) {
-      upsertConcept(graph, tc.name, "tool", turnCount);
-
-      // Link tool to files it operates on
-      const fp = tc.input["file_path"] ?? tc.input["path"];
-      if (typeof fp === "string") {
-        upsertConcept(graph, fp, "file", turnCount);
-        addEdge(graph, tc.name, fp, "modifies");
-      }
-
-      // Check tool results for errors
-      if (tc.result && /error|fail|denied|not found/i.test(tc.result)) {
-        const errLabel = `${tc.name} failed`;
-        upsertConcept(graph, errLabel, "error", turnCount, -0.8);
-        addEdge(graph, tc.name, errLabel, "causes");
-      }
-    }
-  }
-
-  // Extract goals from user messages
-  if (message.role === "user") {
-    const goalPatterns = [
-      /(?:want|need|should|must|해줘|하자|해봐|만들|고쳐)\s+(.{5,40})/i,
-    ];
-    for (const pattern of goalPatterns) {
-      const match = pattern.exec(text);
-      if (match) {
-        upsertConcept(graph, match[1].trim(), "goal", turnCount, 0.5);
-      }
-    }
-  }
-}
-
-function upsertConcept(
-  graph: ConceptGraph,
-  label: string,
-  type: ConceptNode["type"],
-  turn: number,
-  sentiment = 0,
-): void {
-  const id = `${type}:${label}`;
-  const existing = graph.nodes.get(id);
-  if (existing) {
-    existing.weight++;
-    existing.lastSeen = turn;
-    existing.sentiment = (existing.sentiment + sentiment) / 2;
-  } else {
-    graph.nodes.set(id, {
-      id,
-      type,
-      label,
-      weight: 1,
-      firstSeen: turn,
-      lastSeen: turn,
-      sentiment,
-    });
-  }
-}
-
-function addEdge(graph: ConceptGraph, from: string, to: string, relation: ConceptEdge["relation"]): void {
-  const fromId = findNodeId(graph, from);
-  const toId = findNodeId(graph, to);
-  if (!fromId || !toId) return;
-
-  const existing = graph.edges.find((e) => e.from === fromId && e.to === toId);
-  if (existing) {
-    existing.weight++;
-  } else {
-    graph.edges.push({ from: fromId, to: toId, relation, weight: 1 });
-  }
-}
-
-function findNodeId(graph: ConceptGraph, label: string): string | undefined {
-  for (const [id, node] of graph.nodes) {
-    if (node.label === label) return id;
-  }
-  return undefined;
-}
-
-function hasConceptType(graph: ConceptGraph, type: ConceptNode["type"]): boolean {
-  for (const node of graph.nodes.values()) {
-    if (node.type === type && node.lastSeen >= 0) return true;
-  }
-  return false;
-}
-
-function hasConceptSentiment(graph: ConceptGraph, threshold: number): boolean {
-  for (const node of graph.nodes.values()) {
-    if (node.sentiment <= threshold) return true;
-  }
-  return false;
-}
-
-// ═══════════════════════════════════════════════════════════════════
-// AUTO-TRIGGER — Detects when to extract a skill
-// ═══════════════════════════════════════════════════════════════════
-
-function checkAutoTrigger(state: EngineState): SkillTrigger | null {
-  const { phase, phaseHistory, attentionWindow, conceptGraph } = state;
-
-  // Trigger 1: Task completed (phase went problem_stated → ... → completed)
-  if (phase === "completed" && phaseHistory.length >= 3) {
-    const recentPhases = phaseHistory.slice(-6);
-    const hadExecution = recentPhases.includes("executing");
-    const hadProblem = recentPhases.includes("problem_stated");
-
-    if (hadExecution && hadProblem) {
-      const relevantMessages = attentionWindow
-        .filter((m) => m.attention > 0.2)
-        .sort((a, b) => b.attention - a.attention);
-
-      if (relevantMessages.length >= 3) {
-        return {
-          type: "task_completed",
-          messages: relevantMessages,
-          reason: `Task completed: ${state.currentIntent.target}`,
-          quality: computeTriggerQuality(relevantMessages, conceptGraph),
-        };
-      }
-    }
-  }
-
-  // Trigger 2: Pivot point (user changed approach — valuable lesson)
-  if (phase === "pivoting") {
-    const pivotMessages = attentionWindow.slice(-5);
-    if (pivotMessages.length >= 2) {
-      return {
-        type: "pivot_point",
-        messages: pivotMessages,
-        reason: `Approach changed: ${state.currentIntent.motivation}`,
-        quality: 0.6,
-      };
-    }
-  }
-
-  // Trigger 3: Lesson learned (failure → recovery → success)
-  if (phase === "completed" && phaseHistory.includes("failed")) {
-    const journeyMessages = attentionWindow.filter((m) => m.attention > 0.15);
-    return {
-      type: "lesson_learned",
-      messages: journeyMessages,
-      reason: "Recovered from failure — approach evolution captured",
-      quality: 0.8,
-    };
-  }
-
-  return null;
-}
-
-function computeTriggerQuality(
-  messages: AttentionMessage[],
-  graph: ConceptGraph,
-): number {
-  let quality = 0;
-
-  // More high-attention messages = higher quality
-  const highAttention = messages.filter((m) => m.attention > 0.4).length;
-  quality += highAttention * 0.1;
-
-  // More concept nodes = richer context
-  quality += Math.min(0.3, graph.nodes.size * 0.03);
-
-  // Has files = concrete work
-  const fileNodes = [...graph.nodes.values()].filter((n) => n.type === "file");
-  if (fileNodes.length > 0) quality += 0.15;
-
-  // Has tools = actual execution
-  const toolNodes = [...graph.nodes.values()].filter((n) => n.type === "tool");
-  if (toolNodes.length > 0) quality += 0.15;
-
-  return Math.min(1, quality);
-}
-
-// ═══════════════════════════════════════════════════════════════════
-// MAIN ENGINE
-// ═══════════════════════════════════════════════════════════════════
-
-export function createContextEngine(): {
-  /** Process a new message. Returns any triggered skills. */
-  process: (message: ParsedMessage) => SkillTrigger | null;
-  /** Get current engine state. */
-  getState: () => EngineState;
-  /** Get attention-weighted context summary. */
-  getContextSummary: () => string;
-  /** Get the concept graph. */
-  getConceptGraph: () => ConceptGraph;
-} {
-  const state: EngineState = {
-    phase: "idle",
-    phaseHistory: ["idle"],
-    currentIntent: {
-      primary: "continue",
-      confidence: 0,
-      secondary: [],
-      target: "",
-      motivation: "",
-      urgency: "low",
-    },
-    intentHistory: [],
-    attentionWindow: [],
-    conceptGraph: { nodes: new Map(), edges: [] },
-    triggers: [],
-    topicStack: [],
-    turnCount: 0,
-  };
-
-  return {
-    process(message: ParsedMessage): SkillTrigger | null {
-      state.turnCount++;
-
-      // 1. Classify intent
-      const intent = message.role === "user"
-        ? classifyIntentBayesian(
-            message.content,
-            state.phase,
-            state.intentHistory.map((i) => i.primary),
-            state.conceptGraph,
-          )
-        : state.currentIntent; // Keep current intent for assistant messages
-
-      state.currentIntent = intent;
-      state.intentHistory.push(intent);
-
-      // 2. Transition phase
-      const newPhase = transitionPhase(state.phase, message, intent);
-      if (newPhase !== state.phase) {
-        state.phaseHistory.push(newPhase);
-      }
-      state.phase = newPhase;
-
-      // 3. Update concept graph
-      updateConceptGraph(state.conceptGraph, message, state.turnCount);
-
-      // 4. Add to attention window
-      const attMsg: AttentionMessage = {
-        message,
-        attention: 1.0, // Will be recalculated
-        phase: state.phase,
-        intent,
-      };
-      state.attentionWindow.push(attMsg);
-
-      // 5. Recompute attention weights
-      computeAttention(state.attentionWindow, message, state.conceptGraph);
-
-      // 6. Prune old messages (keep max 50)
-      if (state.attentionWindow.length > 50) {
-        // Remove lowest-attention messages beyond window
-        state.attentionWindow.sort((a, b) => b.attention - a.attention);
-        state.attentionWindow = state.attentionWindow.slice(0, 50);
-      }
-
-      // 7. Update topic stack
-      if (message.role === "user" && intent.target) {
-        if (state.topicStack[state.topicStack.length - 1] !== intent.target) {
-          state.topicStack.push(intent.target);
-          if (state.topicStack.length > 10) state.topicStack.shift();
-        }
-      }
-
-      // 8. Check auto-trigger
-      const trigger = checkAutoTrigger(state);
-      if (trigger) {
-        state.triggers.push(trigger);
-        // Reset phase after trigger
-        state.phase = "idle";
-        state.phaseHistory.push("idle");
-      }
-
-      return trigger;
-    },
-
-    getState(): EngineState {
-      return state;
-    },
-
-    getContextSummary(): string {
-      const topConcepts = [...state.conceptGraph.nodes.values()]
-        .sort((a, b) => b.weight - a.weight)
-        .slice(0, 10)
-        .map((n) => `${n.type}:${n.label}(${n.weight})`);
-
-      return [
-        `Phase: ${state.phase}`,
-        `Intent: ${state.currentIntent.primary} (${(state.currentIntent.confidence * 100).toFixed(0)}%)`,
-        `Target: ${state.currentIntent.target}`,
-        `Topics: ${state.topicStack.slice(-5).join(" → ")}`,
-        `Concepts: ${topConcepts.join(", ")}`,
-        `Triggers: ${state.triggers.length}`,
-      ].join("\n");
-    },
-
-    getConceptGraph(): ConceptGraph {
-      return state.conceptGraph;
-    },
-  };
-}
-
-// ═══════════════════════════════════════════════════════════════════
-// HELPERS
-// ═══════════════════════════════════════════════════════════════════
-
-function tokenize(text: string): string[] {
-  return text
-    .toLowerCase()
-    .replace(/[^a-z가-힣0-9\s]/g, " ")
-    .split(/\s+/)
-    .filter((w) => w.length > 1);
-}
-
-function extractTarget(
-  text: string,
-  intent: IntentType,
-  graph: ConceptGraph,
-): string {
-  // Try to find the most relevant concept for this intent
-  const recentFiles = [...graph.nodes.values()]
-    .filter((n) => n.type === "file")
-    .sort((a, b) => b.lastSeen - a.lastSeen);
-
-  const recentErrors = [...graph.nodes.values()]
-    .filter((n) => n.type === "error")
-    .sort((a, b) => b.lastSeen - a.lastSeen);
-
-  if (intent === "fix_bug" && recentErrors.length > 0) {
-    return recentErrors[0].label;
-  }
-  if (["code_review", "refactor", "optimize"].includes(intent) && recentFiles.length > 0) {
-    return recentFiles[0].label;
-  }
-
-  // Fall back to extracting from the message itself
-  const backtick = text.match(/`([^`]+)`/);
-  if (backtick) return backtick[1];
-
-  const quoted = text.match(/"([^"]+)"/);
-  if (quoted) return quoted[1];
-
-  // First noun-like phrase after the intent verb
-  const words = text.split(/\s+/).slice(0, 8);
-  return words.slice(1, 4).join(" ").slice(0, 40) || "unknown";
-}
-
-function inferMotivation(
-  phase: ConversationPhase,
-  recentIntents: IntentType[],
-  graph: ConceptGraph,
-): string {
-  if (phase === "failed") return "Previous approach failed, trying alternative";
-  if (phase === "pivoting") return "User redirected to a different approach";
-
-  const errors = [...graph.nodes.values()].filter((n) => n.type === "error");
-  if (errors.length > 0) return `Responding to error: ${errors[errors.length - 1].label.slice(0, 50)}`;
-
-  const goals = [...graph.nodes.values()].filter((n) => n.type === "goal");
-  if (goals.length > 0) return goals[goals.length - 1].label;
-
-  if (recentIntents.length > 0) {
-    const last = recentIntents[recentIntents.length - 1];
-    return `Continuation of ${last} task`;
-  }
-
-  return "User initiated new task";
-}
-
-function detectUrgency(text: string): "low" | "medium" | "high" | "critical" {
-  const lower = text.toLowerCase();
-
-  // Critical: exclamation, urgent words
-  if (/급해|urgent|critical|asap|지금 당장|immediately|!{2,}/.test(lower)) return "critical";
-
-  // High: strong language, errors mentioned
-  if (/빨리|quickly|error|crash|broken|안돼|망했/.test(lower)) return "high";
-
-  // Medium: requests with some emphasis
-  if (/please|좀|부탁|important|need/.test(lower)) return "medium";
-
-  return "low";
-}