@hawon/nexus 0.2.0 → 0.3.0

package/src/skills/wisdom-extractor.ts

@@ -1,737 +0,0 @@
-/**
- * Wisdom Extractor — Extracts principles, not procedures
- *
- * Instead of: "Run `Edit` on `src/scanner.ts`"
- * Produces:   "When hardening a regex-based scanner, add dynamic property
- *              access detection because attackers use bracket notation to
- *              bypass literal pattern matching."
- *
- * Three types of wisdom:
- *
- * 1. PATTERN → PRINCIPLE
- *    "이럴 때는 이렇게 해라"
- *    When [situation], do [approach] because [reason].
- *
- * 2. ANTI-PATTERN → WARNING
- *    "이렇게 하면 안 된다"
- *    Don't [approach] when [situation] because [consequence].
- *
- * 3. DECISION → RATIONALE
- *    "왜 이 방법을 선택했는가"
- *    Chose [A] over [B] because [tradeoff].
- */
-
-import type { ParsedMessage, ToolCall } from "../parser/types.js";
-import { createHash } from "node:crypto";
-
-// ─── Types ───────────────────────────────────────────────────────
-
-export type Wisdom = {
-  id: string;
-  type: "principle" | "warning" | "rationale";
-  /** One-line title. */
-  title: string;
-  /** The wisdom itself — situation → approach → reason. */
-  body: WisdomBody;
-  /** Abstraction level: how general is this wisdom? */
-  abstraction: "specific" | "moderate" | "general";
-  /** Domain tags. */
-  domains: string[];
-  /** Confidence. */
-  confidence: number;
-  /** Source session. */
-  sourceSessionId: string;
-  /** Source message range. */
-  sourceRange: [number, number];
-  /** When extracted. */
-  extractedAt: string;
-};
-
-export type WisdomBody = {
-  /** When does this apply? */
-  situation: string;
-  /** What should you do (principle) or not do (warning)? */
-  approach: string;
-  /** Why? The underlying reason. */
-  reason: string;
-  /** What's the alternative that doesn't work? (for warnings/rationales) */
-  alternative?: string;
-  /** Concrete example from the session (brief). */
-  example?: string;
-};
-
-// ─── Detection Patterns ──────────────────────────────────────────
-
-/**
- * Detects "wisdom-worthy" moments in conversations.
- *
- * A moment is wisdom-worthy when:
- * 1. User asked → Claude tried approach A → failed → tried approach B → succeeded
- *    → PRINCIPLE: "When X, do B instead of A because..."
- *
- * 2. User corrected Claude's approach
- *    → WARNING: "Don't do A when X because..."
- *
- * 3. Claude explicitly chose between alternatives
- *    → RATIONALE: "Chose A over B because..."
- *
- * 4. User learned something surprising
- *    → PRINCIPLE: "When X, note that Y (counterintuitive)"
- *
- * 5. A complex multi-step task succeeded
- *    → PRINCIPLE: "When doing X, the key steps are..."
- */
-
-type ConversationWindow = {
-  messages: ParsedMessage[];
-  startIndex: number;
-  endIndex: number;
-};
-
-// ─── Main Extraction ─────────────────────────────────────────────
-
-export function extractWisdom(
-  messages: ParsedMessage[],
-  sessionId: string,
-): Wisdom[] {
-  const wisdoms: Wisdom[] = [];
-
-  // Pass 1: Find failure-recovery patterns (approach A failed → B worked)
-  wisdoms.push(...findFailureRecoveries(messages, sessionId));
-
-  // Pass 2: Find user corrections (user redirected Claude's approach)
-  wisdoms.push(...findUserCorrections(messages, sessionId));
-
-  // Pass 3: Find explicit decisions (Claude chose between alternatives)
-  wisdoms.push(...findExplicitDecisions(messages, sessionId));
-
-  // Pass 4: Find learning moments (user expressed surprise or understanding)
-  wisdoms.push(...findLearningMoments(messages, sessionId));
-
-  // Pass 5: Find successful complex tasks (multi-tool sequences that worked)
-  wisdoms.push(...findSuccessfulPatterns(messages, sessionId));
-
-  // Deduplicate by similarity
-  return deduplicateWisdom(wisdoms);
-}
-
-// ─── Pass 1: Failure → Recovery ──────────────────────────────────
-
-function findFailureRecoveries(messages: ParsedMessage[], sessionId: string): Wisdom[] {
-  const wisdoms: Wisdom[] = [];
-
-  for (let i = 0; i < messages.length - 2; i++) {
-    // Look for: assistant with error → user/assistant retry → success
-    const msg = messages[i];
-    if (msg.role !== "assistant") continue;
-    if (isNoiseMessage(msg)) continue;
-
-    // Check for failure in tool results
-    const hasFailure = msg.toolCalls?.some((tc) =>
-      tc.result && /error|fail|denied|not found|exit code [1-9]/i.test(tc.result),
-    );
-    if (!hasFailure) continue;
-
-    const failedTools = msg.toolCalls?.filter((tc) =>
-      tc.result && /error|fail/i.test(tc.result),
-    ) ?? [];
-
-    // Look ahead for recovery (successful tool use)
-    let recoveryMsg: ParsedMessage | null = null;
-    let recoveryIndex = -1;
-    for (let j = i + 1; j < Math.min(i + 6, messages.length); j++) {
-      const candidate = messages[j];
-      if (candidate.role === "assistant" && isNoiseMessage(candidate)) continue;
-      if (candidate.role === "assistant" && candidate.toolCalls?.length) {
-        const allSuccess = candidate.toolCalls.every((tc) =>
-          !tc.result || !/error|fail/i.test(tc.result),
-        );
-        if (allSuccess) {
-          recoveryMsg = candidate;
-          recoveryIndex = j;
-          break;
-        }
-      }
-    }
-
-    if (!recoveryMsg || recoveryIndex < 0) continue;
-
-    // Extract the situation, what failed, what worked
-    const failedApproach = describeToolAction(failedTools[0]);
-    const successApproach = describeToolAction(recoveryMsg.toolCalls?.[0]);
-
-    // Get user context (what were they trying to do?)
-    const userContext = findNearestUserMessage(messages, i);
-    const situation = abstractSituation(userContext?.content ?? "");
-
-    wisdoms.push({
-      id: hashWisdom(`recovery:${i}:${sessionId}`),
-      type: "principle",
-      title: `${situation.slice(0, 40)} — ${successApproach.method}가 더 효과적`,
-      body: {
-        situation: situation || "이 상황에서",
-        approach: successApproach.description,
-        reason: `처음 시도한 ${failedApproach.method}가 실패했기 때문 (${failedApproach.error})`,
-        alternative: failedApproach.description,
-        example: `${failedApproach.method} → 실패 → ${successApproach.method} → 성공`,
-      },
-      abstraction: "moderate",
-      domains: extractDomains(messages.slice(i, recoveryIndex + 1)),
-      confidence: 0.7,
-      sourceSessionId: sessionId,
-      sourceRange: [i, recoveryIndex],
-      extractedAt: new Date().toISOString(),
-    });
-
-    i = recoveryIndex; // Skip to after recovery
-  }
-
-  return wisdoms;
-}
-
-// ─── Pass 2: User Corrections ────────────────────────────────────
-
-function findUserCorrections(messages: ParsedMessage[], sessionId: string): Wisdom[] {
-  const wisdoms: Wisdom[] = [];
-
-  const redirectPatterns = [
-    /아니[야요]?\s/i, /그게\s*아니라/i, /말고/i, /대신/i, /다시/i,
-    /아닌데/i, /그거\s*말고/i, /다르게/i, /바꿔/i,
-    /no[,.]?\s*(?:not|don't|instead)/i, /wait/i, /actually/i,
-    /wrong/i, /scratch that/i, /instead/i,
-  ];
-
-  for (let i = 1; i < messages.length; i++) {
-    const msg = messages[i];
-    if (msg.role !== "user") continue;
-    if (isNoiseMessage(msg)) continue;
-
-    const isRedirect = redirectPatterns.some((p) => p.test(msg.content));
-    if (!isRedirect) continue;
-
-    // What was Claude doing before?
-    const prevAssistant = findPrevAssistant(messages, i);
-    if (!prevAssistant) continue;
-
-    // What did the user want instead?
-    const userWant = msg.content;
-
-    // What happened after the correction?
-    const nextAssistant = findNextAssistant(messages, i);
-
-    const wrongApproach = abstractApproach(prevAssistant);
-    const rightApproach = abstractApproach(nextAssistant);
-    const situation = abstractSituation(
-      findNearestUserMessage(messages, Math.max(0, i - 3))?.content ?? "",
-    );
-
-    wisdoms.push({
-      id: hashWisdom(`correction:${i}:${sessionId}`),
-      type: "warning",
-      title: `${situation.slice(0, 40)}에서 ${wrongApproach.slice(0, 30)}하지 말 것`,
-      body: {
-        situation,
-        approach: rightApproach || userWant.slice(0, 100),
-        reason: `사용자가 "${userWant.slice(0, 60)}"로 수정함`,
-        alternative: wrongApproach,
-      },
-      abstraction: "moderate",
-      domains: extractDomains(messages.slice(Math.max(0, i - 2), Math.min(i + 3, messages.length))),
-      confidence: 0.8,
-      sourceSessionId: sessionId,
-      sourceRange: [Math.max(0, i - 1), Math.min(i + 2, messages.length - 1)],
-      extractedAt: new Date().toISOString(),
-    });
-  }
-
-  return wisdoms;
-}
-
-// ─── Pass 3: Explicit Decisions ──────────────────────────────────
-
-function findExplicitDecisions(messages: ParsedMessage[], sessionId: string): Wisdom[] {
-  const wisdoms: Wisdom[] = [];
-
-  const decisionPatterns = [
-    /(?:chose|선택|picked|decided|결정)\s+(\S+)\s+(?:over|대신|instead of|말고)\s+(\S+)/i,
-    /(?:better|나은|preferred|적합)\s+(?:to|approach|방법|than)\s/i,
-    /(?:A|B|option|방법|방안)\s*(?:vs|versus|대|or)\s/i,
-    /(?:tradeoff|trade-off|트레이드오프|장단점)/i,
-  ];
-
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i];
-    if (msg.role !== "assistant") continue;
-
-    const text = msg.content;
-    const hasDecision = decisionPatterns.some((p) => p.test(text));
-    if (!hasDecision) continue;
-
-    // Extract the decision context
-    const userQ = findNearestUserMessage(messages, i);
-    const situation = abstractSituation(userQ?.content ?? "");
-
-    // Try to extract the alternatives from the text
-    const alternatives = extractAlternatives(text);
-
-    if (alternatives) {
-      wisdoms.push({
-        id: hashWisdom(`decision:${i}:${sessionId}`),
-        type: "rationale",
-        title: `${alternatives.chosen}을 선택한 이유`,
-        body: {
-          situation,
-          approach: alternatives.chosen,
-          reason: alternatives.reason || "구체적 이유는 문맥 참조",
-          alternative: alternatives.rejected,
-        },
-        abstraction: "moderate",
-        domains: extractDomains([msg]),
-        confidence: 0.6,
-        sourceSessionId: sessionId,
-        sourceRange: [i, i],
-        extractedAt: new Date().toISOString(),
-      });
-    }
-  }
-
-  return wisdoms;
-}
-
-// ─── Pass 4: Learning Moments ────────────────────────────────────
-
-function findLearningMoments(messages: ParsedMessage[], sessionId: string): Wisdom[] {
-  const wisdoms: Wisdom[] = [];
-
-  const surprisePatterns = [
-    /아\s*그렇구나/i, /몰랐/i, /처음\s*알았/i, /신기/i,
-    /oh\s*i\s*see/i, /didn.t\s*know/i, /interesting/i, /til\b/i,
-    /그런\s*거였/i, /이유가\s*있었/i,
-  ];
-
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i];
-    if (msg.role !== "user") continue;
-
-    const isSurprised = surprisePatterns.some((p) => p.test(msg.content));
-    if (!isSurprised) continue;
-
-    // What did they just learn? (previous assistant message)
-    const prevAssistant = findPrevAssistant(messages, i);
-    if (!prevAssistant || prevAssistant.content.length < 20) continue;
-
-    const learned = abstractLearning(prevAssistant.content);
-    const context = abstractSituation(
-      findNearestUserMessage(messages, Math.max(0, i - 3))?.content ?? "",
-    );
-
-    wisdoms.push({
-      id: hashWisdom(`learning:${i}:${sessionId}`),
-      type: "principle",
-      title: `${context}에서 알게 된 점`,
-      body: {
-        situation: context,
-        approach: learned,
-        reason: "실제 작업 과정에서 발견",
-      },
-      abstraction: "general",
-      domains: extractDomains(messages.slice(Math.max(0, i - 2), i + 1)),
-      confidence: 0.65,
-      sourceSessionId: sessionId,
-      sourceRange: [Math.max(0, i - 1), i],
-      extractedAt: new Date().toISOString(),
-    });
-  }
-
-  return wisdoms;
-}
-
-// ─── Pass 5: Successful Complex Patterns ─────────────────────────
-
-function findSuccessfulPatterns(messages: ParsedMessage[], sessionId: string): Wisdom[] {
-  const wisdoms: Wisdom[] = [];
-
-  // Find sequences of 3+ tool calls followed by positive user feedback
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i];
-    if (msg.role !== "assistant") continue;
-
-    const toolCount = msg.toolCalls?.length ?? 0;
-    if (toolCount < 3) continue;
-
-    // Check for positive feedback within next 3 messages
-    const feedback = findPositiveFeedback(messages, i);
-    if (!feedback) continue;
-
-    // Abstract the pattern
-    const toolTypes = [...new Set(msg.toolCalls?.map((tc) => tc.name) ?? [])];
-    const userContext = findNearestUserMessage(messages, i);
-    const situation = abstractSituation(userContext?.content ?? "");
-    const approach = abstractMultiStepApproach(msg.toolCalls ?? []);
-
-    wisdoms.push({
-      id: hashWisdom(`pattern:${i}:${sessionId}`),
-      type: "principle",
-      title: `${situation} — 효과적 접근법`,
-      body: {
-        situation,
-        approach,
-        reason: "이 순서로 했을 때 성공적이었음",
-        example: toolTypes.join(" → "),
-      },
-      abstraction: "moderate",
-      domains: extractDomains([msg]),
-      confidence: 0.7,
-      sourceSessionId: sessionId,
-      sourceRange: [i, feedback.index],
-      extractedAt: new Date().toISOString(),
-    });
-
-    i = feedback.index;
-  }
-
-  return wisdoms;
-}
-
-// ─── Abstraction Helpers ─────────────────────────────────────────
-
-/** Convert specific user request into a general situation description. */
-function abstractSituation(text: string): string {
-  if (!text || text.length < 3) return "";
-
-  // Remove specific file paths, variable names, etc.
-  let abstracted = text
-    .replace(/\/[\w./\\-]+/g, "")
-    .replace(/`[^`]+`/g, "")
-    .replace(/https?:\/\/\S+/g, "")
-    .replace(/\b[0-9a-f]{8,}\b/g, "")
-    .replace(/<[^>]+>/g, "")
-    .replace(/\s+/g, " ")
-    .trim();
-
-  if (abstracted.length < 5) return "";
-
-  // Take first meaningful sentence
-  const firstSentence = abstracted.split(/[.!?\n]/)[0]?.trim() ?? abstracted;
-  return firstSentence.slice(0, 80);
-}
-
-/** Abstract what an assistant was doing into a general approach. */
-function abstractApproach(msg: ParsedMessage | null): string {
-  if (!msg) return "이전 접근법";
-
-  if (msg.toolCalls && msg.toolCalls.length > 0) {
-    const toolNames = [...new Set(msg.toolCalls.map((tc) => tc.name))];
-    return `${toolNames.join(", ")} 사용`;
-  }
-
-  // Extract key action from text
-  const text = msg.content.slice(0, 200);
-  const actionMatch = text.match(/(?:먼저|우선|first)\s+(.{10,40})/i) ??
-    text.match(/(.{10,40})(?:하겠|할게|합니다|했습니다)/);
-  if (actionMatch) return actionMatch[1].trim();
-
-  return text.slice(0, 60);
-}
-
-/** Abstract a learning into a general principle. */
-function abstractLearning(assistantText: string): string {
-  // Extract the key insight — usually the first substantive sentence
-  const sentences = assistantText.split(/[.!?\n]/).filter((s) => s.trim().length > 15);
-  if (sentences.length === 0) return assistantText.slice(0, 100);
-
-  // Prefer sentences with explanation markers
-  const explanatory = sentences.find((s) =>
-    /때문|because|이유|reason|즉|essentially|사실|actually|중요한|important/.test(s),
-  );
-
-  return (explanatory ?? sentences[0]).trim().slice(0, 150);
-}
-
-/** Abstract a multi-step tool sequence into a general approach. */
-function abstractMultiStepApproach(toolCalls: ToolCall[]): string {
-  const steps: string[] = [];
-  const seen = new Set<string>();
-
-  for (const tc of toolCalls) {
-    const toolName = tc.name;
-    if (seen.has(toolName)) continue;
-    seen.add(toolName);
-
-    switch (toolName) {
-      case "Bash":
-        steps.push("명령어로 현재 상태를 확인");
-        break;
-      case "Read":
-        steps.push("관련 파일을 읽어서 구조를 파악");
-        break;
-      case "Grep":
-        steps.push("코드베이스에서 관련 패턴을 검색");
-        break;
-      case "Edit":
-        steps.push("문제가 있는 부분을 수정");
-        break;
-      case "Write":
-        steps.push("새로운 파일을 생성");
-        break;
-      case "Agent":
-        steps.push("복잡한 하위 작업을 에이전트에 위임");
-        break;
-      case "WebSearch":
-        steps.push("최신 정보나 문서를 검색");
-        break;
-      default:
-        steps.push(`${toolName} 사용`);
-    }
-  }
-
-  if (steps.length <= 1) return steps[0] ?? "단일 단계 접근";
-  return steps.map((s, i) => `${i + 1}. ${s}`).join("\n");
-}
-
-/** Describe a tool call at an abstract level. */
-function describeToolAction(tc: ToolCall | undefined): { method: string; description: string; error: string } {
-  if (!tc) return { method: "알 수 없음", description: "알 수 없는 접근", error: "" };
-
-  const method = tc.name;
-  let description = method;
-  let error = "";
-
-  if (tc.name === "Bash" && typeof tc.input["command"] === "string") {
-    const cmd = tc.input["command"] as string;
-    // Abstract the command
-    if (cmd.includes("npm") || cmd.includes("pnpm")) description = "패키지 매니저 실행";
-    else if (cmd.includes("git")) description = "git 작업";
-    else if (cmd.includes("tsc") || cmd.includes("typescript")) description = "타입 체크";
-    else if (cmd.includes("test")) description = "테스트 실행";
-    else description = "쉘 명령 실행";
-  } else if (tc.name === "Edit") {
-    description = "코드 수정";
-  } else if (tc.name === "Read") {
-    description = "파일 읽기";
-  }
-
-  if (tc.result && /error|fail/i.test(tc.result)) {
-    const errorMatch = tc.result.match(/(?:error|Error)[:\s]+(.{10,60})/i);
-    error = errorMatch ? errorMatch[1].trim() : "실행 실패";
-  }
-
-  return { method, description, error };
-}
-
-// ─── Utility Helpers ─────────────────────────────────────────────
-
-function isNoiseMessage(msg: ParsedMessage): boolean {
-  const t = msg.content.trim();
-  return t.startsWith("<task-notification>") ||
-    t.startsWith("<system-reminder>") ||
-    t.startsWith("<local-command") ||
-    t.startsWith("┌──") ||
-    t.startsWith("{") ||
-    /^<[a-z]/.test(t) ||
-    t.length < 3;
-}
-
-function findNearestUserMessage(messages: ParsedMessage[], index: number): ParsedMessage | null {
-  for (let i = index; i >= Math.max(0, index - 8); i--) {
-    if (messages[i].role === "user" && messages[i].content.trim().length > 5) {
-      if (isNoiseMessage(messages[i])) continue;
-      return messages[i];
-    }
-  }
-  return null;
-}
-
-function findPrevAssistant(messages: ParsedMessage[], index: number): ParsedMessage | null {
-  for (let i = index - 1; i >= Math.max(0, index - 3); i--) {
-    if (messages[i].role === "assistant" && messages[i].content.trim().length > 10) {
-      return messages[i];
-    }
-  }
-  return null;
-}
-
-function findNextAssistant(messages: ParsedMessage[], index: number): ParsedMessage | null {
-  for (let i = index + 1; i < Math.min(index + 4, messages.length); i++) {
-    if (messages[i].role === "assistant" && messages[i].content.trim().length > 10) {
-      return messages[i];
-    }
-  }
-  return null;
-}
-
-function findPositiveFeedback(
-  messages: ParsedMessage[],
-  index: number,
-): { message: ParsedMessage; index: number } | null {
-  const positivePatterns = [
-    /좋아|완벽|됐어|됐다|고마워|감사|잘됐|ㅇㅇ|ㄱㄱ|ㅇㅋ/,
-    /good|perfect|great|thanks|nice|works|awesome/i,
-    /다음|이제|next|now/i, // Moving on = implicit success
-  ];
-
-  for (let i = index + 1; i < Math.min(index + 5, messages.length); i++) {
-    const msg = messages[i];
-    if (msg.role !== "user") continue;
-    if (positivePatterns.some((p) => p.test(msg.content))) {
-      return { message: msg, index: i };
-    }
-  }
-  return null;
-}
-
-function extractDomains(messages: ParsedMessage[]): string[] {
-  const domains = new Set<string>();
-
-  const domainPatterns: [string, RegExp][] = [
-    ["security", /보안|security|취약|vulnerab|exploit|injection|audit/i],
-    ["testing", /테스트|test|spec|assert|coverage/i],
-    ["devops", /deploy|배포|docker|ci\/cd|github actions|npm publish/i],
-    ["frontend", /react|vue|angular|css|html|component|jsx|tsx/i],
-    ["backend", /server|api|database|sql|rest|graphql/i],
-    ["git", /git|commit|push|branch|merge|pr|pull request/i],
-    ["performance", /성능|optimize|performance|cache|speed|slow/i],
-    ["refactoring", /refactor|리팩토|clean|정리|restructure/i],
-    ["debugging", /debug|디버그|error|에러|log|stack trace/i],
-    ["documentation", /문서|docs|readme|comment|주석/i],
-    ["open-source", /오픈소스|open.?source|contribute|pr|issue/i],
-  ];
-
-  const allText = messages.map((m) => m.content).join(" ");
-  for (const [domain, pattern] of domainPatterns) {
-    if (pattern.test(allText)) domains.add(domain);
-  }
-
-  return [...domains];
-}
-
-function extractAlternatives(text: string): { chosen: string; rejected: string; reason: string } | null {
-  // Try to find "A vs B" or "chose A over B" patterns
-  const vsMatch = text.match(/(\S+)\s+(?:vs|versus|대)\s+(\S+)/i);
-  if (vsMatch) {
-    return { chosen: vsMatch[1], rejected: vsMatch[2], reason: "" };
-  }
-
-  const choseMatch = text.match(/(?:chose|선택|picked)\s+(.+?)\s+(?:over|대신|instead of)\s+(.+?)(?:\.|,|$)/i);
-  if (choseMatch) {
-    return { chosen: choseMatch[1].trim(), rejected: choseMatch[2].trim(), reason: "" };
-  }
-
-  return null;
-}
-
-function deduplicateWisdom(wisdoms: Wisdom[]): Wisdom[] {
-  const unique: Wisdom[] = [];
-  const seen = new Set<string>();
-
-  for (const w of wisdoms) {
-    // Simple dedup key: type + first 30 chars of situation + first 30 chars of approach
-    const key = `${w.type}:${w.body.situation.slice(0, 30)}:${w.body.approach.slice(0, 30)}`;
-    if (seen.has(key)) continue;
-    seen.add(key);
-    unique.push(w);
-  }
-
-  return unique.sort((a, b) => b.confidence - a.confidence);
-}
-
-function hashWisdom(input: string): string {
-  return createHash("sha256").update(input).digest("hex").slice(0, 10);
-}
-
-// ─── Obsidian Renderer ───────────────────────────────────────────
-
-export function renderWisdomMarkdown(wisdom: Wisdom): string {
-  const lines: string[] = [];
-
-  const typeEmoji = wisdom.type === "principle" ? "💡"
-    : wisdom.type === "warning" ? "⚠️" : "🤔";
-  const typeLabel = wisdom.type === "principle" ? "원칙"
-    : wisdom.type === "warning" ? "주의" : "판단 근거";
-
-  lines.push("---");
-  lines.push(`type: wisdom`);
-  lines.push(`wisdom_type: ${wisdom.type}`);
-  lines.push(`domains: [${wisdom.domains.map((d) => `"${d}"`).join(", ")}]`);
-  lines.push(`confidence: ${wisdom.confidence.toFixed(2)}`);
-  lines.push(`abstraction: ${wisdom.abstraction}`);
-  lines.push(`tags: [claude/wisdom, claude/${wisdom.type}]`);
-  lines.push("---");
-  lines.push("");
-  lines.push(`# ${typeEmoji} ${wisdom.title}`);
-  lines.push("");
-  lines.push(`> **${typeLabel}** | 확신도: ${(wisdom.confidence * 100).toFixed(0)}% | 추상화: ${wisdom.abstraction}`);
-  lines.push("");
-
-  // Situation
-  lines.push("## 상황");
-  lines.push("");
-  lines.push(wisdom.body.situation);
-  lines.push("");
-
-  // Approach
-  if (wisdom.type === "warning") {
-    lines.push("## 하지 말 것");
-    lines.push("");
-    lines.push(`~~${wisdom.body.alternative ?? wisdom.body.approach}~~`);
-    lines.push("");
-    lines.push("## 대신 이렇게");
-    lines.push("");
-    lines.push(wisdom.body.approach);
-  } else {
-    lines.push("## 접근법");
-    lines.push("");
-    lines.push(wisdom.body.approach);
-  }
-  lines.push("");
-
-  // Reason
-  lines.push("## 이유");
-  lines.push("");
-  lines.push(wisdom.body.reason);
-  lines.push("");
-
-  // Example
-  if (wisdom.body.example) {
-    lines.push("## 실제 사례");
-    lines.push("");
-    lines.push(`\`${wisdom.body.example}\``);
-    lines.push("");
-  }
-
-  // Alternative (for rationales)
-  if (wisdom.type === "rationale" && wisdom.body.alternative) {
-    lines.push("## 대안");
-    lines.push("");
-    lines.push(wisdom.body.alternative);
-    lines.push("");
-  }
-
-  return lines.join("\n");
-}
-
-/** Export wisdom collection to Obsidian vault. */
-export function exportWisdomToObsidian(
-  wisdoms: Wisdom[],
-  vaultPath: string,
-): string[] {
-  const { writeFileSync, mkdirSync, existsSync } = require("node:fs") as typeof import("node:fs");
-  const { join } = require("node:path") as typeof import("node:path");
-
-  const dir = join(vaultPath, "Wisdom");
-  if (!existsSync(dir)) {
-    mkdirSync(dir, { recursive: true });
-  }
-
-  const files: string[] = [];
-  for (const w of wisdoms) {
-    const safeName = w.title
-      .replace(/[<>:"/\\|?*#^[\]]/g, "-")
-      .replace(/-+/g, "-")
-      .slice(0, 60);
-    const filePath = join(dir, `${safeName}.md`);
-    writeFileSync(filePath, renderWisdomMarkdown(w), "utf-8");
-    files.push(filePath);
-  }
-  return files;
-}