Files

T

Antigravity Agent 504fd5fb42 [G1-Sync] Manual knowledge update

2026-05-10 22:08:15 +09:00

7.0 KiB

Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

title

Predictive Refactoring

매 한 줄

"매 LLM + 매 static-analysis + 매 git-history 를 결합해 '여기를 곧 고쳐야 한다'를 예측하고 매 PR 으로 제안". 매 reactive code review 가 매 proactive 로 이동. 매 2026 의 Claude Opus 4.7 / GPT-5 + 매 codebase RAG 의 핵심 use-case.

매 핵심

매 어디까지가 PR 인가

Reactive: 매 code 작성 후 IDE 의 quick-fix.
Predictive: 매 commit 패턴 / churn / 매 complexity drift 의 분석으로 매 미래 hotspot 예측 → 매 사전 PR.
Continuous: 매 main 에 push 마다 매 background agent 가 분석 → 매 weekly digest.

매 신호

Code churn: 매 같은 file 의 매 commit 빈도.
Complexity drift: 매 cyclomatic / cognitive complexity 가 매 임계 초과 향하는 추세.
Test coverage erosion: 매 module의 매 coverage 하락.
Code-smell ML model: 매 long method, large class, feature envy 의 학습된 detector.
Issue / bug correlation: 매 bug 가 자주 나는 file.
AST embedding similarity drift: 매 module 이 매 codebase 의 다른 module 패턴에서 멀어짐.

매 ML / LLM 결합

Static feature extraction (tree-sitter AST) → 매 vector.
History feature (git log) → 매 churn time-series.
LLM 매 candidate refactoring 생성 + 매 risk 평가 + 매 explanation.
Validation 매 test run + 매 type-check + 매 mutation testing.

매 trust gate

매 LLM 제안의 자동 merge X.
매 small / mechanical (rename, extract) — 매 high-confidence auto-PR.
매 architectural (split module, change interface) — 매 RFC 형식 review request.

💻 패턴

Hotspot detection (churn × complexity)

type FileMetric = { path: string; churn30d: number; cyclomatic: number; bugs90d: number };

function rankHotspots(metrics: FileMetric[]) {
  const max = (k: keyof FileMetric) =>
    Math.max(...metrics.map(m => m[k] as number)) || 1;
  const Mc = max("churn30d"), Mx = max("cyclomatic"), Mb = max("bugs90d");
  return metrics
    .map(m => ({
      ...m,
      score: (m.churn30d / Mc) * 0.4 +
             (m.cyclomatic / Mx) * 0.4 +
             (m.bugs90d / Mb) * 0.2,
    }))
    .sort((a, b) => b.score - a.score)
    .slice(0, 20);
}

LLM refactoring suggestion (Claude Code SDK)

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

async function suggestRefactor(filePath: string, source: string) {
  const res = await client.messages.create({
    model: "claude-opus-4-7",
    max_tokens: 4096,
    system: [{
      type: "text",
      text: "매 senior staff engineer. 매 unsafe / large refactor 의 reject. 매 small testable steps.",
      cache_control: { type: "ephemeral" },
    }],
    messages: [{
      role: "user",
      content: `파일: ${filePath}\n\n\`\`\`\n${source}\n\`\`\`\n\n` +
               "매 1) 가장 시급한 refactoring 1개 + 매 2) diff (unified) + 매 3) risk 평가 + 매 4) test 추가 제안.",
    }],
  });
  return parseStructured(res);
}

Auto-PR generation (mechanical refactor)

async function autoRefactorPR(repo: string, suggestion: Suggestion) {
  if (suggestion.risk !== "low" || suggestion.kind !== "mechanical") return;

  const branch = `refactor/predict-${suggestion.id}`;
  await git.createBranch(repo, branch);
  await git.applyDiff(repo, branch, suggestion.diff);

  const tests = await ci.run(repo, branch);
  if (!tests.green) return; // 매 abort

  await github.openPR(repo, {
    base: "main", head: branch,
    title: `refactor: ${suggestion.title}`,
    body: prTemplate(suggestion),
    labels: ["predictive-refactor", "auto"],
    reviewers: [suggestion.codeOwner],
  });
}

Risk classifier

function classifyRisk(s: Suggestion): "low" | "med" | "high" {
  if (s.linesChanged > 200 || s.filesTouched > 5) return "high";
  if (s.touchesPublicAPI || s.touchesMigration) return "high";
  if (s.changesBehavior) return "med";
  return "low"; // 매 rename, extract local fn, dead code, type tightening
}

Continuous monitor (GitHub Action)

name: Predictive Refactoring
on:
  schedule: [{ cron: "0 4 * * 1" }]   # 매 주 월요일 새벽
  workflow_dispatch: {}
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - run: npx tree-sitter parse . > ast.json
      - run: node scripts/hotspots.js > hotspots.json
      - uses: anthropics/claude-code-action@v1
        with:
          mode: refactor-digest
          input: hotspots.json
          api-key: ${{ secrets.ANTHROPIC_API_KEY }}
      - uses: peter-evans/create-pull-request@v6
        with:
          title: "weekly: 매 predictive refactoring digest"
          body-path: digest.md

Mutation-test gated merge

# 매 refactor 가 behavior 보존인지 확인
npx stryker run --reporters dashboard,json
# 매 mutation score ≥ pre-refactor 일 때만 merge 허용

매 결정 기준

상황	Approach
Mechanical (rename, extract)	LLM auto-PR + CI green
Logic 보존 needed	Mutation testing gate
API change	RFC + human review
Cross-module	RFC + ADR + staged plan
Legacy module no tests	매 test scaffold 먼저, refactor 후
Hotspot 미발견	churn/complexity threshold 낮춤

기본값: 매 weekly digest + 매 low-risk auto-PR + 매 high-risk RFC.

🔗 Graph

부모: Refactoring · AI-Assisted Development
변형: AI-Assisted Refactoring · Automated Refactoring Tools
응용: Claude Code · GitHub Copilot · Sourcegraph Cody
Adjacent: Code Smells · Code Churn · Mutation Testing · Static Analysis

🤖 LLM 활용

언제: 매 codebase >50k LoC, 매 ongoing maintenance, 매 hot-spot 의 패턴화 가능. 언제 X: 매 prototype, 매 throwaway, 매 매우 작은 codebase — 매 overhead.

❌ 안티패턴

자동 merge of LLM diff: 매 hallucination / behavior break. → 매 항상 CI + 매 PR review.
Refactor without tests: 매 mutation/regression 무방어.
Big-bang refactor: 매 small steps 무시.
Score 만 보고 hotspot 따라가기: 매 churn 이 곧 file 의 healthy domain change 일 수도.
No code-owner routing: 매 PR floods.

🧪 검증 / 중복

Verified (Microsoft Research churn-bug correlation, "Your Code as a Crime Scene" Adam Tornhill, Anthropic Claude Code SDK docs, Stryker mutation testing).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — hotspot signals + LLM-assisted PR pipeline

7.0 KiB Raw Blame History Unescape Escape