2nd/10_Wiki/Topics/AI_and_ML/하이브리드 코드 리뷰.md

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

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
wiki-2026-0508-하이브리드-코드-리뷰	하이브리드 코드 리뷰	10_Wiki/Topics	verified	self	Hybrid Code Review AI-Assisted Code Review	none	A	0.9	applied	code-review ai-tools devops ci-cd claude-code		2026-05-10	pending	language framework typescript github-actions

하이브리드 코드 리뷰

매 한 줄

"매 LLM 의 first-pass mechanical review + human 의 second-pass design judgment 의 layered combine". 매 2026 standard PR workflow 의 Claude Code Review / GitHub Copilot review / CodeRabbit 의 의 자동 lint-style + security + style critique 의 emit, 매 human reviewer 의 의 architecture / business / API design 의 final call 의 reserve — 매 cycle time 의 50%+ 의 reduce 한 이후 quality 의 maintain.

매 핵심

매 division of labor

• AI 의 strength (first pass): lint, type-check, test coverage gap, security CVE pattern, naming consistency, docstring missing, obvious bug, regex bug, off-by-one, null deref candidate.
• Human 의 strength (second pass): architecture fit, business logic correctness, API design taste, performance trade-off, team convention, mentoring, ambiguity resolution.
• Overlap zone: 매 both 의 의 catch — security critical / public API change.

매 workflow stage

• Stage 1 (PR open): CI lint + AI bot 의 inline comment 의 emit.
• Stage 2 (author response): 매 AI suggestion 의 author 의 accept / reject / mark "see human review".
• Stage 3 (human review): 매 human 의 의 AI noise 의 skip — 매 design / intent 의 focus.
• Stage 4 (merge gate): 매 critical AI finding 의 unresolved → block; non-critical → warn only.

매 응용

1. Claude Code /review slash command 의 PR diff review.
2. GitHub Actions + Anthropic API 의 PR comment 자동.
3. CodeRabbit / Greptile 의 contextual review.
4. Pre-commit local AI lint (claude-code, cursor).
5. Security-focused AI scanner (Snyk + LLM, Semgrep + LLM).

💻 패턴

GitHub Action — Claude AI review

# .github/workflows/ai-review.yml
name: AI Code Review
on:
  pull_request:
    types: [opened, synchronize]

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          model: claude-opus-4-7
          mode: review
          base_ref: ${{ github.event.pull_request.base.ref }}

Custom review script (Anthropic SDK)

// scripts/ai-review.ts
import Anthropic from '@anthropic-ai/sdk';
import { execSync } from 'child_process';

const client = new Anthropic();
const diff = execSync('git diff origin/main...HEAD').toString();

const res = await client.messages.create({
  model: 'claude-opus-4-7',
  max_tokens: 4096,
  system: [{
    type: 'text',
    text: REVIEW_RUBRIC,
    cache_control: { type: 'ephemeral' }   // 매 prompt cache — 매 rubric 의 reuse
  }],
  messages: [{
    role: 'user',
    content: `Review this diff. Output JSON array of findings.\n\n${diff}`
  }]
});

const findings = JSON.parse(extractJson(res.content[0].text));
postPRComments(findings);

Review rubric (cached system prompt)

const REVIEW_RUBRIC = `
You are a senior code reviewer. For each issue output:
{
  "file": string,
  "line": number,
  "severity": "critical" | "major" | "minor" | "nit",
  "category": "security" | "bug" | "perf" | "style" | "test",
  "message": string,
  "suggestion": string  // optional patch
}

Critical (block merge):
- SQL injection, XSS, path traversal, secret leak
- Null deref on user-reachable path
- Missing auth/authz check

Skip (do NOT comment):
- Style preferences without lint rule
- Architectural opinions (human's job)
- Speculative perf without measurement
`;

Inline PR comment posting

import { Octokit } from '@octokit/rest';
const gh = new Octokit({ auth: process.env.GITHUB_TOKEN });

for (const f of findings) {
  if (f.severity === 'nit') continue; // 매 noise reduction
  await gh.pulls.createReviewComment({
    owner, repo, pull_number,
    body: `**[${f.severity}] ${f.category}**: ${f.message}\n\n\`\`\`suggestion\n${f.suggestion}\n\`\`\``,
    commit_id: headSha,
    path: f.file,
    line: f.line
  });
}

Severity-gated merge

# require-human-on-critical.yml
- name: Check AI critical findings
  run: |
    CRITICAL=$(jq '[.[] | select(.severity=="critical")] | length' findings.json)
    if [ "$CRITICAL" -gt 0 ]; then
      gh pr edit $PR --add-label "needs-human-review"
      exit 1
    fi

Local pre-commit hook

#!/bin/bash
# .git/hooks/pre-commit — 매 staged diff 의 local AI quick-check
DIFF=$(git diff --cached)
[ -z "$DIFF" ] && exit 0

claude --print --model claude-haiku-4-5 \
  "Review this staged diff for obvious bugs. Reply DONE if clean, else list issues:\n$DIFF" \
  | tee /tmp/precommit-review.txt

grep -qi "^DONE" /tmp/precommit-review.txt

Reviewer dashboard (signal vs noise)

// 매 human 의 의 AI suggestion 의 accept rate 의 track — 매 rubric tuning loop
interface ReviewMetric {
  prNumber: number;
  aiFindings: number;
  humanAccepted: number;     // resolved as "good catch"
  humanDismissed: number;    // marked "noise"
  humanMissed: number;       // human found, AI didn't
}
// 매 acceptRate < 30% 의 rubric 의 too noisy — tighten.
// 매 humanMissed > 0 의 rubric 의 too narrow — broaden.

매 결정 기준

상황	Approach
Small team / OSS	Claude Code action (zero-config)
Enterprise / private	Self-hosted Anthropic API + GH Actions
Latency-critical	Pre-commit Haiku quick-check
Security-heavy	Semgrep + LLM context layer
Design-heavy review	Skip AI, pure human

기본값: Claude Code Action 의 PR + human reviewer 의 of design.

🔗 Graph

• 부모: Code-Review · CI-CD
• 변형: Static-Analysis
• Adjacent: Claude-Code · GitHub-Actions · Prompt-Cache

🤖 LLM 활용

언제: 매 mechanical pass (lint, security, naming, docstring, test gap); 매 PR description 자동 generate; 매 commit message rewrite. 언제 X: 매 architecture decision; 매 API design; 매 business logic correctness; 매 team mentoring — 매 human 의 final call.

❌ 안티패턴

• Rubber stamp: 매 AI suggestion 의 의 blind accept — 매 false positive 의 ship.
• AI noise flood: 매 every nit 의 comment — 매 reviewer fatigue.
• Bypass human: 매 AI green = merge — 매 design rot.
• No prompt cache: 매 매 PR 의 의 large rubric 의 re-send — cost 10×.
• Public diff leak: 매 private code 의 의 unconfigured 3rd-party AI 의 send — 매 secrets policy 의 violate.

🧪 검증 / 중복

• Verified (Anthropic Claude Code Action docs; GitHub blog "Copilot for PRs" 2025; CodeRabbit case studies).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Claude Code Action + cached rubric + severity gating patterns 추가