@hawon/nexus 0.1.0 → 0.3.0

package/src/skills/smart-extractor.ts

@@ -1,843 +0,0 @@
-/**
- * Smart Skill Extractor — Context-Aware Pattern Recognition
- *
- * Unlike the basic extractor that treats every tool-call sequence as a skill,
- * this engine understands conversation FLOW:
- *
- * 1. EPISODE SEGMENTATION
- *    - Breaks sessions into "task episodes" (one coherent unit of work)
- *    - Detects episode boundaries: topic change, time gap, transition markers
- *
- * 2. INTENT CLASSIFICATION
- *    - Classifies what the user actually wanted (not just what tools ran)
- *    - "fix X" vs "build X" vs "investigate X" vs "deploy X"
- *
- * 3. DECISION POINT DETECTION
- *    - Finds moments where the approach changed:
- *      - User redirected ("아니 그게 아니라", "no wait", "instead")
- *      - Claude failed and tried differently
- *      - User gave feedback that shaped the next steps
- *    - These decision points ARE the real skill — not the tool calls
- *
- * 4. FLOW NARRATIVE
- *    - Builds a narrative: motivation → attempt → feedback → adaptation → result
- *    - This narrative IS the skill context that makes it reusable
- *
- * 5. QUALITY GATE
- *    - Only promotes patterns that are genuinely reusable
- *    - Filters: enough complexity, clear outcome, reproducible steps
- */
-
-import type { ParsedSession, ParsedMessage, ToolCall } from "../parser/types.js";
-import { createHash } from "node:crypto";
-
-// ─── Types ───────────────────────────────────────────────────────
-
-export type Episode = {
-  /** Unique ID. */
-  id: string;
-  /** What the user wanted to achieve. */
-  intent: Intent;
-  /** Messages in this episode. */
-  messages: ParsedMessage[];
-  /** Tool calls made. */
-  toolCalls: ToolCall[];
-  /** Files modified. */
-  files: string[];
-  /** Decision points (where approach changed). */
-  decisionPoints: DecisionPoint[];
-  /** Full narrative of what happened. */
-  narrative: Narrative;
-  /** Outcome of the episode. */
-  outcome: EpisodeOutcome;
-  /** Start/end timestamps. */
-  startedAt: string;
-  endedAt: string;
-  /** Skill quality score 0-1. Higher = more worth extracting. */
-  skillQuality: number;
-};
-
-export type Intent = {
-  /** High-level category. */
-  category: IntentCategory;
-  /** Human-readable description. */
-  description: string;
-  /** Key entities (files, modules, concepts). */
-  entities: string[];
-  /** Original user request. */
-  originalRequest: string;
-};
-
-export type IntentCategory =
-  | "fix_bug"           // Fix an error, crash, or incorrect behavior
-  | "build_feature"     // Create something new
-  | "refactor"          // Restructure existing code
-  | "investigate"       // Analyze, debug, understand
-  | "configure"         // Setup, install, config
-  | "deploy"            // Deploy, publish, release
-  | "test"              // Write or fix tests
-  | "security"          // Security audit, vulnerability fix
-  | "review"            // Code review, PR review
-  | "learn"             // User asking to understand something
-  | "misc";             // Doesn't fit other categories
-
-export type DecisionPoint = {
-  /** Index in episode messages where the decision occurred. */
-  messageIndex: number;
-  /** What triggered the change. */
-  trigger: DecisionTrigger;
-  /** What was the approach before. */
-  before: string;
-  /** What was the approach after. */
-  after: string;
-  /** Why (inferred). */
-  reason: string;
-};
-
-export type DecisionTrigger =
-  | "user_redirect"      // User explicitly changed direction
-  | "error_recovery"     // Tool/command failed, had to adapt
-  | "user_feedback"      // User gave positive/negative feedback
-  | "scope_expansion"    // "also do X" — task grew
-  | "scope_reduction"    // "just do X" — task narrowed
-  | "approach_pivot";    // Claude realized a different method was better
-
-export type Narrative = {
-  /** Why this task was started. */
-  motivation: string;
-  /** What was attempted first. */
-  initialApproach: string;
-  /** Key turns (max 5). */
-  keyTurns: NarrativeTurn[];
-  /** Final result. */
-  resolution: string;
-  /** Lessons learned (what worked, what didn't). */
-  lessons: string[];
-};
-
-export type NarrativeTurn = {
-  what: string;
-  why: string;
-  outcome: "success" | "failure" | "partial";
-};
-
-export type EpisodeOutcome = {
-  status: "completed" | "failed" | "abandoned" | "ongoing";
-  /** Confidence in this assessment. */
-  confidence: number;
-  /** User's final reaction if any. */
-  finalReaction?: string;
-};
-
-export type SmartSkill = {
-  id: string;
-  /** Clear, actionable name (e.g., "Fix ESLint errors in TypeScript project"). */
-  name: string;
-  /** When to use this skill. */
-  whenToUse: string;
-  /** Step-by-step approach (the real value). */
-  approach: SmartStep[];
-  /** What to watch out for. */
-  pitfalls: string[];
-  /** Tools typically needed. */
-  tools: string[];
-  /** Files/patterns typically involved. */
-  filePatterns: string[];
-  /** How many times this pattern appeared. */
-  frequency: number;
-  /** Quality score. */
-  quality: number;
-  /** Source episodes. */
-  sourceEpisodes: string[];
-};
-
-export type SmartStep = {
-  order: number;
-  action: string;
-  tool?: string;
-  /** Why this step matters. */
-  rationale: string;
-  /** Common mistakes at this step. */
-  caution?: string;
-};
-
-// ─── Transition/Redirect Markers ─────────────────────────────────
-
-const TRANSITION_MARKERS_KO = [
-  "자", "이제", "다음", "그러면", "그럼", "좋아", "오케이", "ㅇㅋ",
-  "됐고", "됐어", "됐다", "다른거", "다른 거", "그거 말고",
-];
-
-const TRANSITION_MARKERS_EN = [
-  "now", "next", "okay", "alright", "moving on", "let's",
-  "done with that", "forget that", "different",
-];
-
-const REDIRECT_MARKERS_KO = [
-  "아니", "아닌데", "그게 아니라", "말고", "대신", "다시",
-  "잠깐", "멈춰", "스톱", "그만", "아 맞다", "근데",
-  "그거 말고", "다른 방식", "다르게", "바꿔",
-];
-
-const REDIRECT_MARKERS_EN = [
-  "no", "wait", "stop", "actually", "instead", "not that",
-  "wrong", "different", "change", "scratch that", "never mind",
-  "hold on", "but", "hmm",
-];
-
-const POSITIVE_MARKERS = [
-  "좋아", "완벽", "됐어", "됐다", "고마워", "감사", "잘됐",
-  "ㅇㅇ", "ㄱㄱ", "오케이", "ㅇㅋ",
-  "good", "perfect", "great", "thanks", "nice", "works", "awesome",
-];
-
-const NEGATIVE_MARKERS = [
-  "안돼", "안 돼", "틀렸", "잘못", "에러", "버그",
-  "wrong", "error", "fail", "broken", "doesn't work", "bug", "crash",
-];
-
-const ERROR_PATTERNS = [
-  /error/i, /Error:/i, /failed/i, /FAIL/i, /exception/i,
-  /command not found/i, /permission denied/i, /ENOENT/i,
-  /Cannot find/i, /not found/i, /exit code [1-9]/i,
-];
-
-// ─── Noise Filters ───────────────────────────────────────────────
-
-const NOISE_PREFIXES = [
-  "<task-notification>",
-  "<system-reminder>",
-  "<local-command",
-  "Tool loaded",
-  "Using Node for alias",
-];
-
-function isNoiseMessage(msg: ParsedMessage): boolean {
-  const text = msg.content.trim();
-  if (text.length === 0) return true;
-  if (NOISE_PREFIXES.some((p) => text.startsWith(p))) return true;
-  // System-injected metadata
-  if (text.startsWith("{") && text.includes('"type"')) return true;
-  return false;
-}
-
-function isSubstantiveUserMessage(msg: ParsedMessage): boolean {
-  if (msg.role !== "user") return false;
-  if (isNoiseMessage(msg)) return false;
-  // Very short messages that are just acknowledgments
-  const text = msg.content.trim();
-  if (text.length <= 3 && /^[ㅇㄱㅎㅋ]+$/.test(text)) return false;
-  return true;
-}
-
-// ─── Episode Segmentation ────────────────────────────────────────
-
-/**
- * Break a session into coherent task episodes.
- *
- * Episode boundaries are detected by:
- * 1. Time gap > 10 minutes between messages
- * 2. Transition markers ("이제", "다음", "now let's")
- * 3. Complete topic change (low keyword overlap)
- */
-export function segmentEpisodes(session: ParsedSession): Episode[] {
-  const messages = session.messages.filter((m) => !isNoiseMessage(m));
-  if (messages.length === 0) return [];
-
-  const episodes: Episode[] = [];
-  let currentMessages: ParsedMessage[] = [];
-  let episodeCount = 0;
-
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i];
-    const prev = i > 0 ? messages[i - 1] : null;
-
-    // Check for episode boundary
-    let isBoundary = false;
-
-    if (currentMessages.length > 0 && msg.role === "user") {
-      // Time gap > 10 minutes
-      if (prev && msg.timestamp && prev.timestamp) {
-        const gap = new Date(msg.timestamp).getTime() - new Date(prev.timestamp).getTime();
-        if (gap > 10 * 60 * 1000) isBoundary = true;
-      }
-
-      // Transition marker
-      const text = msg.content.toLowerCase().trim();
-      const firstWords = text.split(/\s+/).slice(0, 3).join(" ");
-      if (TRANSITION_MARKERS_KO.some((m) => firstWords.includes(m)) ||
-          TRANSITION_MARKERS_EN.some((m) => firstWords.includes(m))) {
-        // Only if there's enough content in current episode
-        if (currentMessages.length >= 4) isBoundary = true;
-      }
-
-      // Topic change: low keyword overlap with recent messages
-      if (currentMessages.length >= 6) {
-        const recentKeywords = extractMessageKeywords(
-          currentMessages.slice(-4).filter((m) => m.role === "user"),
-        );
-        const newKeywords = extractMessageKeywords([msg]);
-        const overlap = keywordOverlap(recentKeywords, newKeywords);
-        if (overlap < 0.1 && newKeywords.size > 2) isBoundary = true;
-      }
-    }
-
-    if (isBoundary && currentMessages.length >= 2) {
-      const episode = buildEpisode(currentMessages, session.sessionId, episodeCount++);
-      if (episode) episodes.push(episode);
-      currentMessages = [];
-    }
-
-    currentMessages.push(msg);
-  }
-
-  // Final episode
-  if (currentMessages.length >= 2) {
-    const episode = buildEpisode(currentMessages, session.sessionId, episodeCount);
-    if (episode) episodes.push(episode);
-  }
-
-  return episodes;
-}
-
-// ─── Intent Classification ───────────────────────────────────────
-
-function classifyIntent(messages: ParsedMessage[]): Intent {
-  const userMessages = messages.filter((m) => m.role === "user" && isSubstantiveUserMessage(m));
-  if (userMessages.length === 0) {
-    return {
-      category: "misc",
-      description: "Unknown task",
-      entities: [],
-      originalRequest: "",
-    };
-  }
-
-  const firstRequest = userMessages[0].content;
-  const allUserText = userMessages.map((m) => m.content).join(" ").toLowerCase();
-
-  // Category detection with weighted keyword matching
-  const categoryScores: Record<IntentCategory, number> = {
-    fix_bug: 0, build_feature: 0, refactor: 0, investigate: 0,
-    configure: 0, deploy: 0, test: 0, security: 0, review: 0,
-    learn: 0, misc: 0,
-  };
-
-  const patterns: [IntentCategory, RegExp, number][] = [
-    // Fix
-    ["fix_bug", /fix|고치|수정|bug|error|에러|버그|오류|crash|안[되돼]|doesn.t work|broken/i, 3],
-    ["fix_bug", /해결|resolve|debug|디버그|왜.*안/i, 2],
-
-    // Build
-    ["build_feature", /만들|생성|create|build|add|추가|implement|구현|새로/i, 3],
-    ["build_feature", /write|작성|개발|develop/i, 2],
-
-    // Refactor
-    ["refactor", /refactor|리팩토|정리|clean|개선|improve|optimize|최적화/i, 3],
-    ["refactor", /restructure|reorganize|simplify/i, 2],
-
-    // Investigate
-    ["investigate", /분석|analyze|찾아|find|확인|check|search|조사|탐색|스캔/i, 3],
-    ["investigate", /what.*is|어떻게.*되|뭐가|왜.*이런/i, 2],
-
-    // Configure
-    ["configure", /설정|setup|install|config|설치|configure|init|초기화/i, 3],
-    ["configure", /환경|environment|연결|connect|세팅/i, 2],
-
-    // Deploy
-    ["deploy", /배포|deploy|publish|push|release|올려|npm publish/i, 3],
-    ["deploy", /upload|ship|pr.*제출|pr.*생성/i, 2],
-
-    // Test
-    ["test", /테스트|test|검증|verify|validate|확인.*동작/i, 3],
-
-    // Security
-    ["security", /보안|security|취약점|vulnerability|exploit|해킹|audit|감사/i, 3],
-    ["security", /injection|xss|csrf|ssrf|인젝션/i, 2],
-
-    // Review
-    ["review", /리뷰|review|검토|봐봐|확인.*해봐|체크/i, 3],
-
-    // Learn
-    ["learn", /알려줘|설명|explain|tell me|뭐야|what is|how does|어떻게/i, 2],
-    ["learn", /이해|understand|가르쳐|teach/i, 2],
-  ];
-
-  for (const [cat, pattern, weight] of patterns) {
-    if (pattern.test(allUserText)) {
-      categoryScores[cat] += weight;
-    }
-  }
-
-  // Find highest
-  let bestCat: IntentCategory = "misc";
-  let bestScore = 0;
-  for (const [cat, score] of Object.entries(categoryScores)) {
-    if (score > bestScore) {
-      bestScore = score;
-      bestCat = cat as IntentCategory;
-    }
-  }
-
-  // Extract entities (file names, module names, technical terms)
-  const entities = extractEntities(allUserText);
-
-  // Build description
-  const categoryLabels: Record<IntentCategory, string> = {
-    fix_bug: "Fix", build_feature: "Build", refactor: "Refactor",
-    investigate: "Investigate", configure: "Configure", deploy: "Deploy",
-    test: "Test", security: "Security audit", review: "Review",
-    learn: "Learn about", misc: "Task",
-  };
-
-  const entityStr = entities.slice(0, 3).join(", ");
-  const description = entityStr
-    ? `${categoryLabels[bestCat]}: ${entityStr}`
-    : `${categoryLabels[bestCat]}: ${firstRequest.slice(0, 60)}`;
-
-  return {
-    category: bestCat,
-    description,
-    entities,
-    originalRequest: firstRequest,
-  };
-}
-
-function extractEntities(text: string): string[] {
-  const entities: string[] = [];
-
-  // File paths
-  const filePaths = text.match(/[\w./\\-]+\.\w{1,5}/g);
-  if (filePaths) entities.push(...filePaths.slice(0, 3));
-
-  // Backtick-quoted identifiers
-  const backticks = text.match(/`([^`]+)`/g);
-  if (backticks) entities.push(...backticks.map((b) => b.replace(/`/g, "")).slice(0, 3));
-
-  // Technical terms (CamelCase, kebab-case)
-  const techTerms = text.match(/[A-Z][a-z]+[A-Z]\w+|[a-z]+-[a-z]+-[a-z]+/g);
-  if (techTerms) entities.push(...techTerms.slice(0, 3));
-
-  return [...new Set(entities)].slice(0, 5);
-}
-
-// ─── Decision Point Detection ────────────────────────────────────
-
-function detectDecisionPoints(messages: ParsedMessage[]): DecisionPoint[] {
-  const points: DecisionPoint[] = [];
-
-  for (let i = 1; i < messages.length; i++) {
-    const msg = messages[i];
-    if (msg.role !== "user") continue;
-
-    const text = msg.content.toLowerCase();
-    const prev = messages[i - 1];
-
-    // User redirect
-    if (REDIRECT_MARKERS_KO.some((m) => text.includes(m)) ||
-        REDIRECT_MARKERS_EN.some((m) => text.includes(m))) {
-      const prevAction = prev?.toolCalls?.[0]?.name ?? prev?.content.slice(0, 50) ?? "previous approach";
-      points.push({
-        messageIndex: i,
-        trigger: "user_redirect",
-        before: prevAction,
-        after: msg.content.slice(0, 80),
-        reason: `User changed direction: "${msg.content.slice(0, 60)}"`,
-      });
-      continue;
-    }
-
-    // Error recovery: check if previous assistant had errors in tool results
-    if (prev?.role === "assistant" && prev.toolCalls) {
-      const hasError = prev.toolCalls.some((tc) =>
-        tc.result && ERROR_PATTERNS.some((p) => p.test(tc.result ?? "")),
-      );
-      if (hasError) {
-        points.push({
-          messageIndex: i,
-          trigger: "error_recovery",
-          before: prev.toolCalls.map((tc) => tc.name).join("→"),
-          after: msg.content.slice(0, 80),
-          reason: "Previous tool call failed, adapting approach",
-        });
-        continue;
-      }
-    }
-
-    // Scope expansion: "그리고", "also", "추가로"
-    const expansionMarkers = ["그리고", "추가로", "또", "also", "and also", "additionally"];
-    if (expansionMarkers.some((m) => text.startsWith(m) || text.includes(` ${m} `))) {
-      points.push({
-        messageIndex: i,
-        trigger: "scope_expansion",
-        before: "original scope",
-        after: msg.content.slice(0, 80),
-        reason: `Scope expanded: "${msg.content.slice(0, 60)}"`,
-      });
-    }
-
-    // User feedback (positive or negative)
-    if (POSITIVE_MARKERS.some((m) => text.includes(m))) {
-      points.push({
-        messageIndex: i,
-        trigger: "user_feedback",
-        before: "awaiting feedback",
-        after: "confirmed",
-        reason: "User confirmed approach works",
-      });
-    } else if (NEGATIVE_MARKERS.some((m) => text.includes(m))) {
-      points.push({
-        messageIndex: i,
-        trigger: "user_feedback",
-        before: "attempted approach",
-        after: msg.content.slice(0, 80),
-        reason: `Negative feedback: "${msg.content.slice(0, 60)}"`,
-      });
-    }
-  }
-
-  return points;
-}
-
-// ─── Narrative Builder ───────────────────────────────────────────
-
-function buildNarrative(messages: ParsedMessage[], decisionPoints: DecisionPoint[]): Narrative {
-  const userMessages = messages.filter((m) => isSubstantiveUserMessage(m));
-  const assistantMessages = messages.filter((m) => m.role === "assistant" && m.content.trim());
-
-  // Motivation: what came before the first real request
-  const motivation = userMessages.length > 0
-    ? userMessages[0].content.slice(0, 150)
-    : "Unknown";
-
-  // Initial approach: first assistant response with tools
-  const firstToolResponse = assistantMessages.find((m) => m.toolCalls && m.toolCalls.length > 0);
-  const initialApproach = firstToolResponse
-    ? `Used ${firstToolResponse.toolCalls?.map((tc) => tc.name).join(", ")} — ${firstToolResponse.content.slice(0, 100)}`
-    : assistantMessages[0]?.content.slice(0, 100) ?? "No approach recorded";
-
-  // Key turns from decision points
-  const keyTurns: NarrativeTurn[] = decisionPoints.slice(0, 5).map((dp) => ({
-    what: dp.after.slice(0, 100),
-    why: dp.reason,
-    outcome: dp.trigger === "user_feedback" && dp.after === "confirmed" ? "success" as const
-      : dp.trigger === "error_recovery" ? "failure" as const
-      : "partial" as const,
-  }));
-
-  // Resolution: last assistant message
-  const lastAssistant = [...assistantMessages].reverse().find((m) => m.content.trim().length > 10);
-  const resolution = lastAssistant?.content.slice(0, 150) ?? "No resolution recorded";
-
-  // Lessons
-  const lessons: string[] = [];
-  for (const dp of decisionPoints) {
-    if (dp.trigger === "error_recovery") {
-      lessons.push(`Approach "${dp.before}" failed — switched to "${dp.after}"`);
-    }
-    if (dp.trigger === "user_redirect") {
-      lessons.push(`User preferred different approach: ${dp.reason}`);
-    }
-  }
-
-  return { motivation, initialApproach, keyTurns, resolution, lessons };
-}
-
-// ─── Episode Builder ─────────────────────────────────────────────
-
-function buildEpisode(
-  messages: ParsedMessage[],
-  sessionId: string,
-  index: number,
-): Episode | null {
-  // Must have at least one substantive user message and one assistant response
-  const hasUser = messages.some((m) => isSubstantiveUserMessage(m));
-  const hasAssistant = messages.some((m) => m.role === "assistant" && m.content.trim());
-  if (!hasUser || !hasAssistant) return null;
-
-  const intent = classifyIntent(messages);
-  const decisionPoints = detectDecisionPoints(messages);
-
-  // Collect tool calls and files
-  const toolCalls: ToolCall[] = [];
-  const files: string[] = [];
-  for (const msg of messages) {
-    if (msg.toolCalls) {
-      for (const tc of msg.toolCalls) {
-        toolCalls.push(tc);
-        const fp = tc.input["file_path"] ?? tc.input["path"];
-        if (typeof fp === "string") files.push(fp);
-      }
-    }
-  }
-
-  const narrative = buildNarrative(messages, decisionPoints);
-  const outcome = inferEpisodeOutcome(messages);
-
-  // Quality score
-  const skillQuality = computeSkillQuality(
-    messages, toolCalls, decisionPoints, outcome,
-  );
-
-  const timestamps = messages
-    .map((m) => m.timestamp)
-    .filter(Boolean)
-    .sort();
-
-  const id = createHash("sha256")
-    .update(`${sessionId}:${index}:${intent.description}`)
-    .digest("hex")
-    .slice(0, 12);
-
-  return {
-    id,
-    intent,
-    messages,
-    toolCalls,
-    files: [...new Set(files)],
-    decisionPoints,
-    narrative,
-    outcome,
-    startedAt: timestamps[0] ?? "",
-    endedAt: timestamps[timestamps.length - 1] ?? "",
-    skillQuality,
-  };
-}
-
-// ─── Outcome Inference ───────────────────────────────────────────
-
-function inferEpisodeOutcome(messages: ParsedMessage[]): EpisodeOutcome {
-  const userMessages = messages.filter((m) => isSubstantiveUserMessage(m));
-  if (userMessages.length === 0) {
-    return { status: "ongoing", confidence: 0.3 };
-  }
-
-  const lastUser = userMessages[userMessages.length - 1];
-  const text = lastUser.content.toLowerCase();
-
-  // Strong success
-  if (POSITIVE_MARKERS.some((m) => text.includes(m))) {
-    return { status: "completed", confidence: 0.9, finalReaction: lastUser.content.slice(0, 60) };
-  }
-
-  // Strong failure
-  if (NEGATIVE_MARKERS.some((m) => text.includes(m))) {
-    return { status: "failed", confidence: 0.8, finalReaction: lastUser.content.slice(0, 60) };
-  }
-
-  // If last message is a new task, previous was probably completed
-  const lastIsNewTask = TRANSITION_MARKERS_KO.some((m) => text.includes(m)) ||
-    TRANSITION_MARKERS_EN.some((m) => text.includes(m));
-  if (lastIsNewTask) {
-    return { status: "completed", confidence: 0.7 };
-  }
-
-  // Default: if there were tool calls and no complaints, probably completed
-  const hasToolCalls = messages.some((m) => m.toolCalls && m.toolCalls.length > 0);
-  if (hasToolCalls) {
-    return { status: "completed", confidence: 0.5 };
-  }
-
-  return { status: "ongoing", confidence: 0.3 };
-}
-
-// ─── Quality Gate ────────────────────────────────────────────────
-
-function computeSkillQuality(
-  messages: ParsedMessage[],
-  toolCalls: ToolCall[],
-  decisionPoints: DecisionPoint[],
-  outcome: EpisodeOutcome,
-): number {
-  let score = 0;
-
-  // Has tool calls (indicates concrete work, not just chat)
-  if (toolCalls.length >= 2) score += 0.2;
-  if (toolCalls.length >= 5) score += 0.1;
-
-  // Multiple tool types (more complex task)
-  const uniqueTools = new Set(toolCalls.map((tc) => tc.name));
-  if (uniqueTools.size >= 2) score += 0.15;
-  if (uniqueTools.size >= 3) score += 0.1;
-
-  // Has decision points (shows adaptation — valuable knowledge)
-  if (decisionPoints.length >= 1) score += 0.15;
-  if (decisionPoints.length >= 2) score += 0.1;
-
-  // Successful outcome
-  if (outcome.status === "completed") score += 0.15;
-  if (outcome.confidence > 0.7) score += 0.05;
-
-  // Not too short, not too long (sweet spot: 4-30 messages)
-  const msgCount = messages.length;
-  if (msgCount >= 4 && msgCount <= 30) score += 0.1;
-  if (msgCount > 30) score -= 0.05; // Too long = probably multiple tasks
-
-  // Has user feedback (validates the approach)
-  const hasFeedback = decisionPoints.some((dp) => dp.trigger === "user_feedback");
-  if (hasFeedback) score += 0.1;
-
-  return Math.min(1, Math.max(0, score));
-}
-
-// ─── Smart Skill Generation ─────────────────────────────────────
-
-/**
- * Convert high-quality episodes into actionable skills.
- */
-export function generateSmartSkills(
-  episodes: Episode[],
-  minQuality = 0.3,
-): SmartSkill[] {
-  const qualified = episodes.filter((ep) => ep.skillQuality >= minQuality);
-  const skills: SmartSkill[] = [];
-
-  for (const episode of qualified) {
-    const steps: SmartStep[] = [];
-    let stepOrder = 1;
-
-    // Build steps from tool calls + decision points
-    const dpIndices = new Set(episode.decisionPoints.map((dp) => dp.messageIndex));
-
-    for (const msg of episode.messages) {
-      if (msg.role !== "assistant" || !msg.toolCalls) continue;
-
-      for (const tc of msg.toolCalls) {
-        const fileArg = tc.input["file_path"] ?? tc.input["path"] ?? tc.input["command"];
-        const action = fileArg
-          ? `${tc.name}: ${String(fileArg).slice(0, 80)}`
-          : tc.name;
-
-        const step: SmartStep = {
-          order: stepOrder++,
-          action,
-          tool: tc.name,
-          rationale: "", // Will be filled by context
-        };
-
-        steps.push(step);
-      }
-    }
-
-    // Add rationale from decision points
-    for (const dp of episode.decisionPoints) {
-      if (dp.trigger === "error_recovery" && steps.length > 0) {
-        const nearestStep = steps[Math.min(dp.messageIndex, steps.length - 1)];
-        if (nearestStep) {
-          nearestStep.caution = dp.reason;
-        }
-      }
-    }
-
-    // Add rationale from narrative
-    if (steps.length > 0 && episode.narrative.initialApproach) {
-      steps[0].rationale = `Start here: ${episode.narrative.initialApproach.slice(0, 80)}`;
-    }
-
-    // Build pitfalls from failed decision points
-    const pitfalls = episode.decisionPoints
-      .filter((dp) => dp.trigger === "error_recovery" || dp.trigger === "user_redirect")
-      .map((dp) => dp.reason)
-      .slice(0, 5);
-
-    // File patterns
-    const filePatterns = generalizeFilePaths(episode.files);
-
-    const skill: SmartSkill = {
-      id: episode.id,
-      name: episode.intent.description,
-      whenToUse: episode.narrative.motivation.slice(0, 200),
-      approach: steps.slice(0, 15), // Cap at 15 steps
-      pitfalls,
-      tools: [...new Set(episode.toolCalls.map((tc) => tc.name))],
-      filePatterns,
-      frequency: 1,
-      quality: episode.skillQuality,
-      sourceEpisodes: [episode.id],
-    };
-
-    skills.push(skill);
-  }
-
-  return skills.sort((a, b) => b.quality - a.quality);
-}
-
-// ─── Full Pipeline ───────────────────────────────────────────────
-
-/**
- * Run the full smart extraction pipeline on sessions.
- */
-export function smartExtract(sessions: ParsedSession[]): {
-  episodes: Episode[];
-  skills: SmartSkill[];
-} {
-  const allEpisodes: Episode[] = [];
-
-  for (const session of sessions) {
-    const episodes = segmentEpisodes(session);
-    allEpisodes.push(...episodes);
-  }
-
-  const skills = generateSmartSkills(allEpisodes);
-
-  return { episodes: allEpisodes, skills };
-}
-
-// ─── Helpers ─────────────────────────────────────────────────────
-
-function extractMessageKeywords(messages: ParsedMessage[]): Set<string> {
-  const keywords = new Set<string>();
-  const stopWords = new Set([
-    "the", "a", "is", "are", "and", "or", "to", "in", "for", "of",
-    "이", "그", "저", "을", "를", "에", "가", "는", "은", "의",
-  ]);
-
-  for (const msg of messages) {
-    const words = msg.content.toLowerCase().split(/\s+/);
-    for (const w of words) {
-      if (w.length > 2 && !stopWords.has(w)) keywords.add(w);
-    }
-  }
-  return keywords;
-}
-
-function keywordOverlap(a: Set<string>, b: Set<string>): number {
-  if (a.size === 0 || b.size === 0) return 0;
-  const intersection = [...a].filter((w) => b.has(w)).length;
-  const union = new Set([...a, ...b]).size;
-  return intersection / union;
-}
-
-function generalizeFilePaths(files: string[]): string[] {
-  if (files.length === 0) return [];
-
-  // Group by directory
-  const byDir = new Map<string, string[]>();
-  for (const f of files) {
-    const parts = f.split("/");
-    const dir = parts.slice(0, -1).join("/") || ".";
-    const ext = f.split(".").pop() ?? "";
-    byDir.set(dir, [...(byDir.get(dir) ?? []), ext]);
-  }
-
-  // If 3+ files in same dir with same extension, generalize
-  const patterns: string[] = [];
-  for (const [dir, exts] of byDir) {
-    const extCounts = new Map<string, number>();
-    for (const ext of exts) {
-      extCounts.set(ext, (extCounts.get(ext) ?? 0) + 1);
-    }
-    for (const [ext, count] of extCounts) {
-      if (count >= 3) {
-        patterns.push(`${dir}/**/*.${ext}`);
-      }
-    }
-  }
-
-  // If no patterns, return unique files (max 5)
-  if (patterns.length === 0) {
-    return [...new Set(files)].slice(0, 5);
-  }
-
-  return [...new Set(patterns)];
-}