2nd/10_Wiki/Topic_Programming/Architecture/Modern Review Workflow.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-modern-review-workflow	Modern Review Workflow	10_Wiki/Topics	verified	self	AI Code Review PR Review 2026 Augmented Review	none	A	0.9	applied	code-review ci-cd ai-augmented pr-workflow		2026-05-10	pending	language framework typescript github-actions

Modern Review Workflow

매 한 줄

"매 review = human judgment + AI scaffolding". 2026년 PR review 매 single-reviewer linting → multi-agent triage 의 진화. Claude Opus 4.7 / Codex 가 매 first-pass (style, security, regression) 를 처리, human 의 매 architectural / product 판단에 집중.

매 핵심

매 단계

• Pre-PR: 매 author-side claude review local 의 self-check.
• CI gate: 매 automated agent 의 매 diff scan — security, perf, test coverage.
• Human review: 매 design intent / API contract / UX trade-off 의 deep dive.
• Post-merge: 매 deploy preview + canary metrics 의 watch.

매 agent layer

• Linter agent: style, type, dead code.
• Security agent: secret scan, OWASP, dependency CVE.
• Test agent: coverage delta, flaky detect, mutation score.
• Review agent: 매 prose summary + risk flag (Claude Opus 4.7).

매 응용

1. Solo dev: 매 CI agent 만 = 매 reviewer 효과.
2. 팀 (10+): tiered — agent gate → senior architect.
3. OSS: 매 maintainer triage 의 cost 감소.

💻 패턴

GitHub Actions: Claude review hook

name: ai-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:
          api-key: ${{ secrets.ANTHROPIC_API_KEY }}
          model: claude-opus-4-7
          mode: review
          target-base: ${{ github.base_ref }}
          comment-style: inline

Inline comment poster

import Anthropic from "@anthropic-ai/sdk";
import { Octokit } from "@octokit/rest";

const a = new Anthropic();
const gh = new Octokit({ auth: process.env.GH_TOKEN });

export async function reviewDiff(owner: string, repo: string, pr: number) {
  const { data: files } = await gh.pulls.listFiles({ owner, repo, pull_number: pr });
  const diff = files.map((f) => `### ${f.filename}\n${f.patch ?? ""}`).join("\n\n");

  const res = await a.messages.create({
    model: "claude-opus-4-7",
    max_tokens: 4096,
    system: "You are a senior reviewer. Output JSON: {comments: [{path, line, body, severity}]}.",
    messages: [{ role: "user", content: diff }],
  });

  const { comments } = JSON.parse((res.content[0] as any).text);
  for (const c of comments) {
    await gh.pulls.createReviewComment({
      owner, repo, pull_number: pr, ...c,
      commit_id: process.env.HEAD_SHA!,
    });
  }
}

Local pre-PR self-check

# .git/hooks/pre-push
#!/usr/bin/env bash
set -e
git diff origin/main...HEAD | claude -p "Review this diff. Flag bugs, security, perf only. No style." --model claude-opus-4-7

Risk-tiered routing

type Risk = "low" | "medium" | "high";

function classify(diff: string): Risk {
  if (/migrations\/|schema\./.test(diff)) return "high";
  if (/auth|payment|crypto/i.test(diff)) return "high";
  if (diff.split("\n").length > 500) return "medium";
  return "low";
}

function reviewers(r: Risk): string[] {
  return {
    low: ["ai-bot"],
    medium: ["ai-bot", "@team-lead"],
    high: ["ai-bot", "@security", "@architect"],
  }[r];
}

Mutation-test gate

- name: stryker
  run: npx stryker run --threshold.break 70

Coverage delta comment

const before = await coverage("main");
const after = await coverage("HEAD");
const delta = after.lines - before.lines;
if (delta < -1) await gh.issues.createComment({
  ...ctx, body: `⚠️ Coverage dropped ${delta.toFixed(1)}%`,
});

Auto-merge on green + AI ack

- if: ${{ steps.ai-review.outputs.severity == 'none' && steps.tests.outcome == 'success' }}
  run: gh pr merge ${{ github.event.pull_request.number }} --squash --auto

매 결정 기준

상황	Approach
Solo / OSS	AI agent only, human spot-check
Small team	AI gate + 1 human (rotating)
Regulated (fin/health)	AI + 2 humans + audit log
Hot path / migrations	Mandatory architect review

기본값: AI first-pass + 1 human reviewer + risk-tiered escalation.

🔗 Graph

• 부모: CI CD
• 변형: Pair-Programming · Mob-Programming
• 응용: Trunk-Based-Development
• Adjacent: Static-Analysis · Mutation-Testing

🤖 LLM 활용

언제: 매 diff scan, security triage, coverage summary, prose explanation 의 PR description. 언제 X: 매 architectural decision, API contract negotiation, domain-specific business rule — human 의 judgment.

❌ 안티패턴

• AI rubber-stamp: 매 agent approve = 매 human 의 skip. 매 critical path 의 review 누락.
• Comment flood: agent 의 매 nitpick → noise. Severity threshold 의 setting.
• No risk tiering: schema migration 매 typo fix 와 동급 review → bottleneck.
• Secrets in prompt: diff 의 secret 의 leak. Pre-scan + redact.

🧪 검증 / 중복

• Verified (GitHub blog 2025, Anthropic Claude Code docs 2026, Google Eng Practices).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — AI-augmented PR review workflow with Claude Opus 4.7 patterns