SuperAPI AI Directory
Partner中文AI newsFavorites
Skillspeach-team-3a

peach-team-3a

Agents & models

View repository

Updated 10-15 21:36

About this skill

Peach Team 3A

Overview

Architect → Builder → Reviewer → Architect It is a 3-agent team skill that completes a single feature through a loop.

Agent Role Core Principles
Architect Establish design plan + final decision authority Deliver BRIEF via SendMessage → Accept/reject Reviewer verdict
Builder Implement code based on BRIEF Do not touch outside scope, do not modify files after completion, self-review + submit verification evidence
Reviewer Independent verification + 3-stage verdict qa-gate + code review integrated, read-only, worktree isolation, no abuse of CONDITIONAL

Differences from peach-team-dev

peach-team-dev peach-team-3a
Number of agents Up to 5 Fixed at 3
Suitable cases Large-scale fullstack Single feature, small-scale modifications
Verdict method 2-stage pass/fail 3-stage APPROVED / CONDITIONAL / REJECTED
Ambiguous cases Unconditional rework Architect makes the final judgment

Inputs

/peach-team-3a [작업 설명]

# 옵션
# model=sonnet|opus|haiku  (서브에이전트 모델 override, 기본값: sonnet)
# layer=backend|frontend|fullstack  (구현 범위, 미지정 시 Architect가 판단)

Orchestration

0. Input validation

Check that the agent team feature is enabled

cat ~/.claude/settings.json | grep -i "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS"

If the setting is absent or "1"is not stop immediatelyand output guidance:

⚠️  에이전트 팀 기능이 비활성화되어 있습니다.

~/.claude/settings.json에 아래 내용을 추가한 후 Claude Code를 재시작하세요:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

If the task description is missing, ask:

어떤 기능을 구현할까요? (예: "공지사항 목록 API 추가", "회원 상태 필드 수정")

model option:

  • Unspecified: run subagents with the default sonnet
  • If specified: override all subagents to that model
  • Allowed values: sonnet, opus, haiku

1. Environment check

# 프로젝트 구조 파악
ls api/src/modules/ 2>/dev/null
ls front/src/modules/ 2>/dev/null

# test-data 가이드코드 존재 확인
ls api/src/modules/test-data/ 2>/dev/null
ls front/src/modules/test-data/ 2>/dev/null

# DAO 라이브러리 감지
head -5 api/src/modules/test-data/dao/test-data.dao.ts 2>/dev/null
# from 'bunqldb' → 재할당 방식
# from 'sql-template-strings' → append 방식

# Controller 프레임워크 감지
head -3 api/src/modules/test-data/controller/test-data.controller.ts 2>/dev/null

# _common 구성 감지
ls api/src/modules/_common/ 2>/dev/null
ls api/src/modules/_common/constants/ 2>/dev/null

# 스킬 references 경로 감지
BACKEND_REFS=$(find ~/.claude ~/.agents -path "*/peach-gen-backend/references" -type d 2>/dev/null | head -1)
UI_REFS=$(find ~/.claude ~/.agents -path "*/peach-gen-ui/references" -type d 2>/dev/null | head -1)

2. Team composition diagram

Architect
   │
   │ SendMessage: BRIEF 전달
   ▼
Builder ──────────────────────────────────┐
   │                                       │ REJECTED → Ralph Loop
   │ SendMessage: 구현 완료 + 자기 검토     │
   ▼                                       │
Reviewer (worktree 격리)                  │
   │                                       │
   ├── APPROVED ──→ SendMessage → Architect│
   ├── CONDITIONAL → SendMessage → Architect│
   └── REJECTED ──→ SendMessage → Builder ─┘

3. Team creation and task registration

TeamCreate: team_name="3a-[작업축약명]-team"

TaskCreate:
1. "설계 계획 수립" (owner: architect)
2. "코드 구현" (blockedBy: Task1, owner: builder)
3. "검증 및 판정" (blockedBy: Task2, owner: reviewer)

4. Role-specific instructions

The full definition of each role is in references/. When creating a subagent, include the entire contents of that file in the prompt. model= If the option is specified, pass it as the model parameter to override the frontmatter default.

Role Reference file Execution environment
architect references/architect-agent.md General
builder references/builder-agent.md General
reviewer references/reviewer-agent.md worktree isolation

Architect instruction points

Context the orchestrator should include in the prompt when creating the Architect agent:

  • Task description: [사용자 입력 그대로]
  • Environment information: DAO library, Controller framework, modules path, _common constants list
  • Guide code location: api/src/modules/test-data/, front/src/modules/test-data/
  • references path: ${BACKEND_REFS}, ${UI_REFS}
  • Details: references/architect-agent.md reference

Builder instruction points

After the Architect delivers the BRIEF via SendMessage, spin up the Builder. The orchestrator includes the following when creating the Builder agent:

  • Architect's BRIEF content (full text received via SendMessage)
  • Environment information: same as Architect instructions
  • Details: references/builder-agent.md reference

Reviewer instruction points

After the Builder reports implementation completion via SendMessage, spin up the Reviewer. When creating the Reviewer agent, the orchestrator includes the following:

  • Builder's implementation report content (list of implemented files + self-review results)
  • Architect's BRIEF content (judgment criteria)
  • Details: references/reviewer-agent.md reference

5. Reviewer verdict handling

When the Reviewer reports the verdict result to the orchestrator via SendMessage, the orchestrator handles it.

✅ APPROVED

Reviewer → SendMessage → 오케스트레이터: "APPROVED" + 검증 결과
오케스트레이터 → SendMessage → Architect: 완료 보고
오케스트레이터 → /peach-qa-gate 자동 실행 (최종 증거 수집)
오케스트레이터 → 팀 정리

⚠️ CONDITIONAL

Reviewer → SendMessage → 오케스트레이터: "CONDITIONAL" + 조건 내용
오케스트레이터 → SendMessage → Architect: 조건 전달 + 판단 요청
Architect → SendMessage → 오케스트레이터:
    "조건 수용" → 오케스트레이터 → SendMessage → Builder: 수정 지시 → Reviewer 재검증
    "조건 무시" → 판단 근거 기록 후 APPROVED로 처리

Rules:

  • CONDITIONAL is not a Ralph Loop. Completion must not be processed until the Architect makes a judgment.
  • Reviewer는 Conditional items + the reason why it is not REJECTEDmust be reported together.
  • If the Architect chooses "ignore", the orchestrator includes the rationale in the final completion report.

❌ REJECTED — Ralph Loop activated

Reviewer → SendMessage → 오케스트레이터: "REJECTED" + 수정 필요 항목
오케스트레이터 → Ralph Loop 단계 판단
오케스트레이터 → SendMessage → Builder: 피드백 + 수정 지시
Builder 수정 완료 → SendMessage → 오케스트레이터 → Reviewer 재검증

Ralph Loop (iterative verification mechanism)

On REJECTED, rather than a simple retry, iterate with structured feedback injection.

Escalation stages

Iteration count Stage Action
1~3회 Autonomous fix Builder fixes based only on Reviewer feedback
4회 Re-reference guide Re-read the entire test-data baseline skeleton, then fix
5~7회 Codex diagnosis codex:codex-rescueIndependently diagnose the failure cause + re-reference the guide
8~10회 Minimal fix Focus only on Must Follow items, defer the rest
11+ Abort Escalate to user

Codex deployment conditions (5–7 iterations or 2 consecutive REJECTED)

  • CODEX_AVAILABLE=true deployed only when "codex@openai-codex": true detected in settings.json)
  • On 2 consecutive REJECTED: deploy codex:codex-rescue immediately without waiting 5 Ralph Loop iterations (3A induces early escape due to the small-loop nature)
  • CODEX_AVAILABLE=false: continue the existing Ralph Loop without Codex (re-reference the guide)
  • Codex no response/timeout: wait 60 seconds → if no response, skip and proceed with existing feedback
  • Codex diagnosis results insufficient: the orchestrator ignores them and proceeds with existing feedback

Escalation report

## Ralph Loop 에스컬레이션
- 작업: [작업 설명]
- 반복: N/10회
- 단계: [현재 단계]
- 미해결: [위반 항목]
- 권장: [수동 개입 사항]

Inter-agent communication structure

Same as the peach-team-dev series, communicate via SendMessage + TaskUpdate.

Architect → 오케스트레이터 (SendMessage): BRIEF 내용
오케스트레이터 → Builder 프롬프트에 BRIEF 포함하여 스핀업
Builder → 오케스트레이터 (SendMessage): 구현 완료 + 자기 검토 + 파일 목록
오케스트레이터 → Reviewer 프롬프트에 BRIEF + 구현 보고 포함하여 스핀업
Reviewer → 오케스트레이터 (SendMessage): 판정 + 검증 결과

On completion of each stage, update the status via TaskUpdate:

  • Task #1 (design): Architect complete → completed
  • Task #2 (implementation): Builder complete → completed
  • Task #3 (verification): Reviewer APPROVED → completed / REJECTED → Builder rework

Completion

1. Evidence collection

The orchestrator automatically runs /peach-qa-gate→ generate evidence report

  • If the verdict is ❌, fix the relevant item and rerun
  • If the verdict is ✅, proceed to the next step

2. Team cleanup

SendMessage(shutdown_request) → 모든 팀원에게
TeamDelete → 팀 정리

Completion report format

✅ 3A 팀 작업 완료

작업: [작업 설명]
반복: [총 Reviewer 검증 횟수]회

결과:
✅ Architect: BRIEF 수립 완료
✅ Builder: 구현 완료
✅ Reviewer: APPROVED (또는 CONDITIONAL → Architect 수용)
✅ qa-gate: 증거 보고서 생성 + 완료 가능 판정

생성/수정된 파일:
[파일 목록]

다음 단계:
[실행 방법 안내]

Examples

# 단일 API 추가
/peach-team-3a 공지사항 목록 조회 API 추가

# 기존 기능 수정
/peach-team-3a 회원 상태 필드에 '정지' 값 추가 layer=backend

# UI 컴포넌트 수정
/peach-team-3a 상품 목록 페이지에 엑셀 다운로드 버튼 추가 layer=frontend

# opus 모델로 실행
/peach-team-3a 결제 상태 전이 로직 구현 model=opus