에이전트가 멀쩡한 테스트를 자꾸 지웁니다
시도한 것: CLAUDE.md에 명시 → 무시. 훅으로 막아봄 → 우회. 환경: Claude Code 2.1, monorepo.
0 미해결
서브에이전트가 컨텍스트 잃는 문제, 이렇게 고정했습니다
## CLAUDE.md - 모든 서브에이전트 호출 전 "읽어야 할 파일"을 명시 - 결과는 [gstack-context] 블록으로 회신받아 재조립
0 검증됨
Cursor + Next.js로 만든 가계부 앱, 배포 전인데 봐주세요
처음 만든 SaaS입니다. 결제 붙였는데 보안이 제일 불안해서요. 코드 구조랑 보안 위주로 피드백 받고 싶습니다.
AI 점검06·눌러서 상세
0 봐줬으면: 2개 축
OpenCode 전환 후 /code-review · /security-review 공백: opencode-power-pack의 SKILL.md 포팅 구조와 도입 조건
Anthropic 공식 Claude Code 플러그인의 code-review, security-review, feature-dev는 OpenCode에서 그대로 동작하지 않는다. opencode-power-pack은 이 워크플로우들을 OpenCode 네이티브 SKILL.md 포맷으로 번역하고, ~/.config/opencode/opencode.json 한 줄 설정으로 11개 스킬을 적재한다.
7분 읽기 검증 출처읽기
guard-skills: Claude Code diff에 catch-all 오류·환각 import·HPOS 패턴을 잡는 5개 리뷰 Skill
guard-skills는 Claude Code, Codex, Cursor가 생성한 코드·테스트·문서에 second-pass 리뷰를 수행하는 5개 Skill 패키지입니다. clean-code-guard는 GitClear·USENIX 연구에 근거한 14가지 AI 실패 패턴을 검사하고, woo-guard는 AI가 반복 생성하는 pre-HPOS WooCommerce 코드 패턴을 직접 겨냥합니다.
8분 읽기 검증 출처읽기
baoyu-design: What Changes When You Run claude.ai/design Locally in Claude Code
baoyu-design ports the claude.ai/design methodology to Claude Code by detecting the agent environment at runtime and loading a tool substitution table from `references/claude.md`. The design system pipeline — `compile-design-system.mjs` produces static lint config, `import-design-system.mjs` generates a token allowlist as `_ds_prompt.md` — enforces the same constraint at two layers. Neither is optional if you want session-to-session consistency.
7 min read 검증 출처읽기
CI Smoke Tests for MCP Servers with k6
A small k6/x/mcp workflow shows how to prove an MCP server starts, lists tools, returns expected JSON, and fails CI before an agent depends on it.
8 min read 검증 출처읽기
Agent Flow Makes Agent Traces Reviewable
Agent Flow is worth saving as an observability article: it turns Claude Code and Codex sessions into traces a reviewer can inspect before trusting the final patch.
9 min read 검증 출처읽기
Large MCP Inputs Need an Artifact Boundary
The right lesson is not a token-saving trick. It is an MCP interface pattern: upload large files out of band, validate them server-side, and pass only an artifact ID through the model.
8 min read 검증 출처읽기
When a Claude Code File Hook Should Die
The useful story is not a markdown hook that saves tokens. It is a 228-run retraction that teaches Claude Code teams when a hook should be killed, narrowed, or replaced by one instruction.
9 min read 검증 출처읽기
Claude Code Onboarding From Task Loops to MCP Policy
The repo is worth saving because it gives teams an onboarding system from the first task loop through Skills, subagents, hooks, and MCP.
13 min read 검증 출처읽기
Claude Code From Scratch as an Agent Runtime Map
claude-code-from-scratch is worth saving as an agent runtime map: skills, compaction, subagents, worktrees, permissions, and MCP each get a concrete source-backed checkpoint.
11 min read 검증 출처읽기
Audit mcp-token-saver Before It Gates Claude Code
The rescue angle is simple: real 5h and 7d usage can be a better gate than token estimates, but only if the MCP, prompt protocol, credential path, and stale repo docs are checked first.
10 min read 검증 출처읽기
Audit Kagi Session2API MCP Before Sharing Account Access
The useful article is not that Kagi search can be made free for MCP clients. It is how to test a session-token MCP without mistaking it for a normal API key.
10 min read 검증 출처읽기
Claude Code Control Surface Audit
A practical audit turns Claude Code commands, memory, MCP, Skills, settings, and hooks into a runbook a team can verify.
11 min read 검증 출처읽기