oh-my-customcode 0.137.0 → 0.139.0

package/README.md

@@ -13,7 +13,7 @@
 
 **[한국어 문서 (Korean)](./README_ko.md)**
 
-49 agents. 117 skills. 22 rules. One command.
+49 agents. 118 skills. 23 rules. One command.
 
 ```bash
 npm install -g oh-my-customcode && cd your-project && omcustom init
@@ -132,7 +132,7 @@ Each agent declares its tools, model, memory scope, and limitations in YAML fron
 
 ---
 
-### Skills (117)
+### Skills (118)
 
 | Category | Count | Includes |
 |----------|-------|----------|
@@ -210,7 +210,7 @@ All commands are invoked inside the Claude Code conversation.
 
 ---
 
-### Rules (22)
+### Rules (23)
 
 | Priority | Count | Purpose |
 |----------|-------|---------|
@@ -222,7 +222,7 @@ Key rules: R010 (orchestrator never writes files), R009 (parallel execution mand
 
 ---
 
-### Guides (50)
+### Guides (53)
 
 Reference documentation covering best practices, architecture decisions, and integration patterns. Located in `guides/` at project root, covering topics from agent design to CI/CD to observability.
 
@@ -272,7 +272,7 @@ your-project/
 ├── CLAUDE.md                   # Entry point
 ├── .claude/
 │   ├── agents/                 # 49 agent definitions
-│   ├── skills/                 # 117 skill modules
+│   ├── skills/                 # 118 skill modules
 │   ├── rules/                  # 22 governance rules (R000-R021)
 │   ├── hooks/                  # 15 lifecycle hook scripts
 │   ├── schemas/                # Tool input validation schemas

package/dist/cli/index.js

@@ -2334,7 +2334,7 @@ var init_package = __esm(() => {
     workspaces: [
       "packages/*"
     ],
-    version: "0.137.0",
+    version: "0.139.0",
     description: "Batteries-included agent harness for Claude Code",
     type: "module",
     bin: {

package/dist/index.js

@@ -2014,7 +2014,7 @@ var package_default = {
   workspaces: [
     "packages/*"
   ],
-  version: "0.137.0",
+  version: "0.139.0",
   description: "Batteries-included agent harness for Claude Code",
   type: "module",
   bin: {

package/package.json

@@ -3,7 +3,7 @@
   "workspaces": [
     "packages/*"
   ],
-  "version": "0.137.0",
+  "version": "0.139.0",
   "description": "Batteries-included agent harness for Claude Code",
   "type": "module",
   "bin": {

package/templates/.claude/agents/qa-engineer.md

@@ -31,6 +31,13 @@ You are a QA execution specialist that runs tests, identifies defects, and valid
 
 Jest, Vitest, pytest, go test, JUnit, Playwright, Cypress
 
+## Verification Discipline
+
+- Before writing a QA report, grep/read the target code and quote selectors, identifiers, filenames, and commands verbatim from the implementation.
+- Do not invent `data-testid`, DOM selectors, function names, mapping names, or CLI flags from memory.
+- If a browser/MCP/tool call is denied, do not retry the exact same call; switch to an allowed fallback and record the denial.
+- When the same critical launch/runtime error appears twice, stop repeating launches and re-check flag semantics, existing processes, and environment assumptions.
+
 ## Collaboration
 
 Receives from: qa-writer (test cases), qa-planner (priorities). Outputs to: dev-lead (defects), qa-writer (results).

package/templates/.claude/hooks/hooks.json

@@ -90,7 +90,17 @@
             "command": "bash .claude/hooks/scripts/rtk-intercept.sh"
           }
         ],
-        "description": "RTK auto-intercept \u2014 transparently rewrites CLI commands through RTK proxy when available (R015 advisory)"
+        "description": "RTK auto-intercept — transparently rewrites CLI commands through RTK proxy when available (R015 advisory)"
+      },
+      {
+        "matcher": "tool == \"Bash\" && tool_input.command matches \"git (reset|clean|checkout|restore|switch|rebase|merge)\"",
+        "hooks": [
+          {
+            "type": "command",
+            "command": "bash .claude/hooks/scripts/destructive-git-guard.sh"
+          }
+        ],
+        "description": "Warn before destructive git commands — advisory recovery guidance (R001/R021)"
       },
       {
         "matcher": "tool == \"Bash\" && tool_input.command matches \"(rm|git rm|mv|unlink|claude/rules)\"",

package/templates/.claude/hooks/scripts/destructive-git-guard.sh

@@ -0,0 +1,60 @@
+#!/bin/bash
+# R001/R021 destructive git guard hook
+# Warns before commands that can discard working tree, untracked files, or branches.
+# Advisory tier only: emits stderr guidance, records PPID-scoped warnings, exits 0.
+#
+# PPID Scoping Convention:
+#   Session-scoped temp files MUST use $PPID (Claude Code parent PID), not $$.
+#   Pattern: /tmp/.claude-{purpose}-${PPID}
+
+input=$(cat)
+if command -v jq >/dev/null 2>&1; then
+  command_text=$(echo "$input" | jq -r '.tool_input.command // ""' 2>/dev/null)
+elif command -v python3 >/dev/null 2>&1; then
+  command_text=$(INPUT_JSON="$input" python3 - <<'PY' 2>/dev/null
+import json, os
+try:
+    data = json.loads(os.environ.get('INPUT_JSON', ''))
+    print(data.get('tool_input', {}).get('command', ''))
+except Exception:
+    print('')
+PY
+)
+else
+  command_text=""
+fi
+
+VIOLATION_FILE="/tmp/.claude-destructive-git-guard-${PPID}"
+matched=""
+
+if printf '%s\n' "$command_text" | grep -Eq '(^|[;&|[:space:]])git[[:space:]]+reset[[:space:]]+--hard([[:space:]]|$)'; then
+  matched="git reset --hard"
+elif printf '%s\n' "$command_text" | grep -Eq '(^|[;&|[:space:]])git[[:space:]]+clean[[:space:]]+([^;&|]*[[:space:]])?-f(dx|xd|d|x)?([[:space:]]|$)'; then
+  matched="git clean"
+elif printf '%s\n' "$command_text" | grep -Eq '(^|[;&|[:space:]])git[[:space:]]+restore([[:space:]]+--[^;&|[:space:]]+)*([[:space:]]+(\.|\*|/|--worktree|--staged))*([[:space:]]|$)'; then
+  matched="git restore"
+elif printf '%s\n' "$command_text" | grep -Eq '(^|[;&|[:space:]])git[[:space:]]+checkout[[:space:]]+--[[:space:]]+(\.|\*|/)([[:space:]]|$)'; then
+  matched="git checkout -- ."
+elif printf '%s\n' "$command_text" | grep -Eq '(^|[;&|[:space:]])git[[:space:]]+branch[[:space:]]+-D([[:space:]]|$)'; then
+  matched="git branch -D"
+fi
+
+if [ -n "$matched" ]; then
+  echo "[Hook] DESTRUCTIVE GIT WARNING: detected '$matched' in Bash command" >&2
+  echo "[Hook] Advisory only (R021): command is not blocked, but it may discard local work." >&2
+  echo "[Hook] Before proceeding, verify git status and preserve work with git stash or a WIP commit." >&2
+  echo "[Hook] Recovery: use 'git reflog' to find prior HEADs; for deleted branches, check 'git reflog --all'." >&2
+
+  if printf '%s\n' "$command_text" | grep -Eq '(^|[;&|[:space:]])git[[:space:]]+branch[[:space:]]+-D([[:space:]]|$)'; then
+    echo "[Hook] Branch deletion note: prefer 'git branch -d' first so Git checks for unmerged commits." >&2
+  fi
+
+  printf '%s %s\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$matched" >> "$VIOLATION_FILE"
+  violation_count=$(wc -l < "$VIOLATION_FILE" 2>/dev/null | tr -d ' ')
+  if [ "${violation_count:-0}" -ge 3 ]; then
+    echo "[Hook] Destructive git warnings: ${violation_count} this session — R021 promotion threshold reached" >&2
+  fi
+fi
+
+# Always pass through - this hook is advisory only.
+echo "$input"

package/templates/.claude/rules/SHOULD-verification-ladder.md

@@ -0,0 +1,77 @@
+# [SHOULD] Verification Ladder Rules
+
+> **Priority**: SHOULD | **ID**: R023
+
+## Core Rule
+
+검증은 비용/속도 ladder로 구성한다: **결정론적 검사 → cheap LLM → expensive LLM → human**. 가장 저렴한 tier가 먼저 통과해야 다음 tier로 진행한다. 더 낮은 tier에서 잡을 수 있는 문제를 더 비싼 tier에 보내지 않는다.
+
+## Ladder Tiers
+
+| Tier | 도구 | 비용 | 속도 | 적용 시점 |
+|------|------|------|------|-----------|
+| **1: Deterministic** | hooks, linters, type-check, JSON schema | $0 | <1s | Pre-write, write-time |
+| **2: Cheap LLM** | haiku-based skills (`dev-review`, `action-validator`) | $ | <30s | Per-file review |
+| **3: Expensive LLM** | sonnet/opus skills (`deep-verify`, `adversarial-review`, `multi-model-verification`, `evaluator-optimizer`) | $$$ | 1-5분 | Pre-commit, PR review |
+| **4: Human** | maintainer review | time | hours-days | Final gate, contested decisions |
+
+## Shift-left 원칙
+
+결정론적 단계가 잡을 수 있는 문제는 LLM에 보내지 않는다. LLM 검증은 ambiguous/semantic 문제에 집중한다.
+
+- **좋은 예**: JSON schema 오류 → Tier 1 hook이 차단 → LLM에 미전달
+- **나쁜 예**: 탭/스페이스 혼용 오류 → sonnet으로 전달 → 불필요한 비용 발생
+
+R013 (SHOULD-ecomode)의 "저렴한 검증 우선" 원칙과 정합: ecomode는 출력 토큰을 절약하고, R023은 검증 비용을 절약한다.
+
+## 기존 자산 매핑
+
+| Tier | 자산 | 역할 |
+|------|------|------|
+| **Tier 1** | `.claude/hooks/` (PreToolUse hooks) | 도구 호출 전 결정론적 차단 |
+| **Tier 1** | `mgr-sauron` (R017 구조 검증) | 에이전트/스킬/가이드 frontmatter 검증 |
+| **Tier 1** | pre-commit configs, linters | 코드 품질 정적 검사 |
+| **Tier 2** | `dev-review` | 파일 단위 haiku 코드 리뷰 |
+| **Tier 2** | `action-validator` | CI/CD 액션 구문 검증 |
+| **Tier 2** | `pre-generation-arch-check` | 생성 전 아키텍처 lite 점검 |
+| **Tier 3** | `deep-verify` | 다단계 품질 검증 (sonnet) |
+| **Tier 3** | `adversarial-review` | 공격자 시각 보안 리뷰 (opus) |
+| **Tier 3** | `multi-model-verification` | 복수 모델 교차 검증 |
+| **Tier 3** | `evaluator-optimizer` | 평가-개선 반복 루프 |
+| **Tier 3** | `worker-reviewer-pipeline` | 구현-리뷰 파이프라인 |
+| **Tier 4** | maintainer manual review | PR approval, final gate |
+
+## R021과의 관계
+
+R021 (MUST-enforcement-policy)과 R023은 **직교**한다. 두 규칙은 서로 다른 차원을 다룬다:
+
+| 규칙 | 질문 | 차원 |
+|------|------|------|
+| **R021** | "어떻게 강제할 것인가?" | Hard block / Soft block / Advisory |
+| **R023** | "어떤 비용으로 검증할 것인가?" | Deterministic / Cheap LLM / Expensive LLM |
+
+같은 도구가 두 규칙에 동시에 속할 수 있다:
+
+- `mgr-sauron`: R021 관점에서 Advisory (PostToolUse hook), R023 관점에서 Tier 1 (구조 검증)
+- `deep-verify`: R021 관점에서 Prompt-based (blocking 없음), R023 관점에서 Tier 3 (expensive LLM)
+- `.claude/hooks/` stage-blocker: R021 관점에서 Hard Block, R023 관점에서 Tier 1
+
+R021은 위반 시 어떻게 멈출지를, R023은 어떤 순서로 검증할지를 정의한다.
+
+## Self-Check
+
+새 검증 도구 추가 시:
+
+- [ ] 어느 tier에 속하는지 명확한가?
+- [ ] 같은 tier 내 중복 도구는 없는가?
+- [ ] Tier 1에서 잡을 수 있는 문제를 다루는가? (상위 tier 대신 시프트 권고)
+- [ ] Ladder 순서를 문서화했는가? (어떤 검사를 먼저 실행하는지)
+
+## Integration
+
+| 규칙 | 상호작용 |
+|------|---------|
+| R009 (Parallel Execution) | Tier 1-2 검사는 독립 파일에 대해 병렬 실행 가능 |
+| R013 (Ecomode) | 컨텍스트 압박 시 Tier 3를 Tier 2로 다운그레이드 고려 |
+| R017 (Sync Verification) | Phase 1-3 검증 단계는 R023 Tier 1-3에 대응 |
+| R021 (Enforcement Policy) | 직교: R021은 blocking 방식, R023은 검증 비용 순서 |

package/templates/.claude/skills/crg-integration/SKILL.md

@@ -0,0 +1,163 @@
+---
+name: crg-integration
+description: code-review-graph (CRG) MCP wrapper — token-efficient context retrieval via AST knowledge graph. Exposes 4 core tools from CRG.
+scope: core
+version: 0.1.0
+user-invocable: true
+effort: low
+---
+
+# CRG Integration Skill
+
+code-review-graph (CRG) MCP 서버의 핵심 4개 도구를 노출하는 wrapper 스킬.
+전체 28개 도구 중 oh-my-customcode 워크플로우에서 가장 빈번하게 사용되는 도구만 선별해 사용자 학습 곡선을 축소한다.
+
+## 목적
+
+- **8.2× 토큰 절감** (CRG 공식 벤치마크): AST 기반 지식 그래프로 전체 파일 대신 관련 노드만 추출
+- **Recall-우선 설계**: precision 0.38 / MRR 0.35 — 누락 없음이 우선, 잡음은 감수
+- **R013 ecomode 보완**: context 사용량 ≥ 60% 시점에 CRG 호출로 추가 컨텍스트 주입 비용 절감
+
+> **⚠️ Recall이 정답이 아니다** — 아래 주의 섹션을 먼저 읽어라.
+
+## 노출 도구 4종
+
+CRG 전체 28개 도구 중 핵심 4개만 노출한다.
+
+| Tool | 용도 | 출력 형식 |
+|------|------|----------|
+| `get_minimal_context` | 변경 코드의 최소 컨텍스트 (인접 함수, import 그래프) | code excerpt + AST node list |
+| `get_impact_radius` | 변경 시 영향 받는 함수/모듈 (recall-우선) | dependency tree |
+| `query_graph` | AST 노드 그래프 쿼리 (예: "find all callers of X") | node graph |
+| `detect_changes` | 두 시점 간 의미적 차이 감지 | semantic diff |
+
+## 설치 가이드
+
+> R001 정책: auto-install 스크립트 금지. 아래 명령어는 **사용자가 직접** 실행한다.
+
+### 1단계: 패키지 설치
+
+```bash
+# 방법 A: pip
+pip install code-review-graph
+
+# 방법 B: pipx (격리 환경 권장)
+pipx install code-review-graph
+
+# 방법 C: uvx (uv 사용 시)
+uvx code-review-graph
+```
+
+### 2단계: .mcp.json 수동 설정
+
+프로젝트 루트 또는 `~/.claude/` 아래 `.mcp.json`을 **직접 편집**한다:
+
+```json
+{
+  "mcpServers": {
+    "code-review-graph": {
+      "command": "code-review-graph",
+      "args": []
+    }
+  }
+}
+```
+
+auto-install hook 또는 스크립트로 `.mcp.json`을 자동 수정하는 방식은 R001 위반이다.
+
+### 3단계: 초기 인덱싱
+
+```bash
+# 명시적 인덱싱 (선택 — 첫 MCP 호출 시 lazy 인덱싱도 가능)
+code-review-graph index
+```
+
+자세한 설치/설정 절차는 `guides/token-efficiency/crg.md` 참조.
+
+## 사용 패턴
+
+### PR 리뷰 전: 변경 폭 파악
+
+```
+get_impact_radius(changed_files=[...])
+→ dependency tree 확인 → 리뷰 범위 결정
+```
+
+### 코드 검색: token-efficient lookup
+
+```
+query_graph(query="find all callers of process_event")
+→ AST node graph → 전체 파일 읽지 않고 호출 지점만 추출
+```
+
+### 디버깅: 관련 함수만 추출
+
+```
+get_minimal_context(file="src/handler.py", line=142)
+→ code excerpt + AST nodes → 필요한 컨텍스트만 주입
+```
+
+### 리팩토링: 의미적 변경 확인
+
+```
+detect_changes(before_ref="HEAD~1", after_ref="HEAD")
+→ semantic diff → 이름 변경/이동 등 구조 변경 감지
+```
+
+## ⚠️ 주의 — Recall이 정답이 아니다
+
+CRG는 recall 최적화 도구다. 다음 사항을 반드시 숙지한다:
+
+| 오해 | 실제 |
+|------|------|
+| "결과에 없으면 없는 것" | recall-tuned — 일부 관련 코드가 누락될 수 있다 |
+| "precision 0.38 = 결과 62% 오답" | recall-우선 튜닝의 precision trade-off. 잡음은 정상 |
+| "CRG 하나로 충분" | 누락 발생 시 grep, semantic search 병행 필수 |
+| "모든 작업에 8.2× 절감" | AST 검색이 적합한 작업에 한정. UI 리뷰, 자연어 검색은 효과 미미 |
+
+**결론**: CRG는 보조 도구다. 결과를 맹신하지 말고 누락 가능성을 항상 고려한다.
+
+## 다른 자산과의 관계
+
+### 보완 관계 (함께 사용)
+
+| 자산 | 조합 방식 |
+|------|----------|
+| `dev-review` | CRG로 영향 범위 파악 → dev-review로 상세 리뷰 |
+| `adversarial-review` | CRG `get_minimal_context` → adversarial-review 입력 최적화 |
+| `claude-mem:smart-explore` | CRG `query_graph`로 대체 가능 (Phase β 시퀀싱 — #1169 참조) |
+| `claude-mem:pathfinder` | CRG `get_impact_radius`로 대체 가능 (Phase β 시퀀싱 — #1169 참조) |
+
+### 마이그레이션 예정 (Phase β)
+
+- `claude-mem:smart-explore` → CRG `query_graph` (#1169)
+- `claude-mem:pathfinder` → CRG `get_impact_radius` (#1169)
+
+Phase β 완료 전까지는 두 자산을 병행 사용한다.
+
+## R013 Ecomode와의 관계
+
+R013 ecomode가 자동 활성화되는 시점 (context 사용량 ≥ 60%)에서 CRG를 적극 활용한다:
+
+| Ecomode 상황 | CRG 활용법 |
+|-------------|-----------|
+| 파일 전체 읽기 대신 | `get_minimal_context` → 필요한 노드만 추출 |
+| 광범위 grep 대신 | `query_graph` → AST 수준 정밀 검색 |
+| 변경 범위 파악 대신 | `get_impact_radius` → dependency tree만 수신 |
+
+## 트러블슈팅
+
+| 증상 | 원인 | 해결 |
+|------|------|------|
+| MCP 도구가 보이지 않음 | `.mcp.json` 미설정 또는 경로 오류 | `.mcp.json` 확인, Claude Code 재시작 |
+| `command not found: code-review-graph` | 패키지 미설치 또는 PATH 누락 | `pip install code-review-graph` 재실행, `which code-review-graph` 확인 |
+| 인덱싱 오류 | 초기 인덱싱 미완료 | `code-review-graph index` 실행 후 재시도 |
+| 결과 누락 (예상 함수 없음) | recall-tuned 동작 — 정상 | grep 또는 semantic search 병행 |
+| 서버 연결 실패 | CRG 서버 미실행 | MCP 서버 로그 확인, Claude Code 재시작 |
+
+## 참고
+
+- 상세 설치/설정: `guides/token-efficiency/crg.md`
+- 관련 이슈: #1171 (scout:integrate CRG)
+- Phase β 시퀀싱: #1169 (AgentMemory 전환 계획)
+- 기술 출처: [tirth8205/code-review-graph](https://github.com/tirth8205/code-review-graph)

package/templates/.claude/skills/pipeline/labels.md

@@ -0,0 +1,52 @@
+# Pipeline Label Standards
+
+Canonical reference for GitHub issue label semantics in the auto-dev pipeline.
+Used by `scope-selection` to include/exclude issues and by `implement` for lifecycle management.
+
+## Label Definitions
+
+| Label | Meaning | scope-selection 처리 |
+|-------|---------|----------------------|
+| `verify-ready` | Triage 완료, 즉시 verify 가능 (자동화 후보) | INCLUDE (preferred) |
+| `verify-done` | Triage 완료했으나 deferred 또는 이미 처리됨 (이번 사이클 미포함) | EXCLUDE |
+| `in-progress` | 작업 진행 중 (다른 세션에서 claim됨) | EXCLUDE |
+| `needs-review` | 사람 검토 필요 (자동 파이프라인 진입 불가) | EXCLUDE |
+| `decision-needed` | 결정 필요 (보안, 정책 critical) | EXCLUDE |
+| `automated` | Auto-generated issue (claude-native skill 등) | INCLUDE if other criteria met |
+| `claude-code-release` | CC version compat docs trigger | INCLUDE (preferred) |
+| `documentation` | Docs scope | INCLUDE (preferred for docs-only release) |
+| `enhancement-yaml-only` | YAML/config-only scope change | INCLUDE (eligible for docs-only compression) |
+
+## Selection Rule
+
+```
+EXCLUDE if:
+  - blocked_by_decision == true
+  - labels ∩ {decision-needed, needs-review, verify-done, manual-action, in-progress} ≠ ∅
+
+INCLUDE (preferred tier):
+  - labels ∩ {verify-ready, claude-code-release, documentation} ≠ ∅
+
+INCLUDE (standard tier):
+  - P1/P2/P3 issues not in excluded set
+
+Tie-break priority: P1 > P2 > P3 > unclassified
+```
+
+## Compression Eligibility (G6)
+
+An issue is eligible for `docs-only` compression mode if its labels include at least one of:
+`documentation`, `automated`, `claude-code-release`, `enhancement-yaml-only`
+
+If ALL scoped issues are compression-eligible AND scope size ≤ 3, the pipeline MAY use
+`compression_mode=docs-only` (skip professor-triage / release-plan / deep-plan / deep-verify skill spawns).
+
+## Lifecycle Labels (set by `implement` step)
+
+| Transition | Action |
+|-----------|--------|
+| Work started | Add `in-progress`, assign @me |
+| Work succeeded | Remove `in-progress`, add `verify-ready` |
+| Work failed | Remove `in-progress`, add `needs-review` |
+| Released | Remove `verify-ready`, close with "Fixed in v{version}" |
+| Deferred | Add `verify-done`, label "Deferred from v{version}" |

package/templates/.claude/skills/pipeline/workflows/auto-dev.yaml

@@ -1,6 +1,6 @@
 name: auto-dev
 description: "Automated development pipeline — from issue triage to release"
-version: "2.0.0"
+version: "2.1.0"
 
 # Design rule: each pipeline run produces ONE bounded release unit (3-7 issues), not all open issues.
 # This pipeline is a project-agnostic template. Project-specific overrides belong in
@@ -31,7 +31,20 @@ steps:
          - Local HEAD: git rev-parse --short HEAD
          - Behind state: synced or 0
 
-      5. Cross-reference: compare latest tag with any "vX.Y.Z" version references in pipeline memory/context. Flag mismatches for awareness.
+      5. Cross-reference checks (advisory, do NOT halt):
+         a. Memory vs git consistency (G7):
+            - latest_tag=$(git tag --sort=-v:refname | head -1)
+            - Compare with "Last release" or equivalent line in pipeline memory/context
+            - If mismatch: stderr warning "[pre-triage] WARNING: memory says vX.Y.Z but git latest tag is vA.B.C — update memory after pipeline completion"
+            - Do NOT halt (advisory only — memory updates are session-end responsibility)
+
+         b. Issue body stale version references (G5):
+            - Retrieve open issues: gh issue list --state open --limit 100 --json number,body
+            - For each issue body, extract all vX.Y.Z patterns (regex: v\d+\.\d+\.\d+)
+            - Retrieve git tag list: git tag --list
+            - Flag any issue referencing a version not present in git tags:
+              "[pre-triage] WARNING: Issue #N references vX.Y.Z (not in git tags) — body may be stale"
+            - Output all warnings as advisory log; do NOT halt
 
       Phase 1 — Ensure required labels exist, then scan issues.
 
@@ -50,44 +63,95 @@ steps:
          - order: topological sort over dependencies
 
       4. Output: dependency-sorted issue table with blocked_by_decision flag.
-    description: "Ensure labels exist, scan issues, dependency-analyze"
+    description: "Sync remote, stale-version check, memory consistency check, scan issues"
 
   - name: scope-selection
     prompt: |
       Select a single bounded release scope (3-7 issues) for THIS pipeline run.
 
-      Selection rules:
+      ## Step 0 — Milestone state pre-check (G3)
+
+      Before creating a milestone for version vX.Y.Z:
+      1. Query all milestones (open and closed):
+         existing=$(gh api 'repos/{owner}/{repo}/milestones?state=all&per_page=100' --jq '.[] | select(.title == "vX.Y.Z") | "\(.number)|\(.state)"')
+      2. Evaluate:
+         - If existing && state=closed:
+           HALT with "[scope-selection] BLOCKED: milestone vX.Y.Z already exists and is closed. Bump version or reopen milestone manually."
+         - If existing && state=open:
+           Use existing milestone as-is (do not create new)
+         - If not existing:
+           Create new milestone: gh api repos/{owner}/{repo}/milestones --method POST --field title="vX.Y.Z"
+
+      ## Step 1 — Label-based filtering (G4)
+
+      Reference label semantics: .claude/skills/pipeline/labels.md
+
+      Apply filter rules:
+      - EXCLUDE: blocked_by_decision == true OR labels ∩ {decision-needed, needs-review, verify-done, manual-action, in-progress} ≠ ∅
+      - INCLUDE (preferred): labels ∩ {verify-ready, claude-code-release, documentation} ≠ ∅
+      - INCLUDE (standard): P1/P2/P3 issues not in excluded set
+      - Tie-break: P1 > P2 > P3 > unclassified
+
+      ## Step 2 — Selection rules
+
       1. Determine release tier:
          - No tags yet → v0.1.0
          - Otherwise: highest-priority open tier (bug → chore → feature)
-      2. Exclude:
-         - blocked_by_decision == true
-         - Issues requiring manual/human action
+      2. Apply Step 1 filter (label-based exclusion and preference)
       3. Sort by dependency: prerequisites before dependents
       4. Cap at 7 issues; if priority issues < 7, stop at that priority (don't mix tiers)
       5. Minimum 1 issue; if 0 eligible, halt with "no eligible issues for auto-dev run"
 
-      Create release milestone and assign scoped issues.
+      Assign scoped issues to milestone.
       Output markdown release manifest:
-        | order | # | title | prerequisite | effort |
+        | order | # | title | prerequisite | effort | labels |
 
       Persist manifest as pipeline state for subsequent steps.
-    description: "Pick 3-7 issues as v{X.Y.Z} scope, create milestone"
+    description: "Milestone state pre-check, label filter, pick 3-7 issues as v{X.Y.Z} scope"
     depends_on: pre-triage
 
+  - name: compression-mode-eval
+    prompt: |
+      Evaluate whether this pipeline run qualifies for docs-only compression mode (G6).
+
+      ## Compression Eligibility Check
+
+      Reference: .claude/skills/pipeline/labels.md — "Compression Eligibility" section
+
+      Conditions for compression_mode=docs-only:
+      1. scope size ≤ 3 (number of issues in release manifest)
+      2. ALL scoped issues have labels ∩ {documentation, automated, claude-code-release, enhancement-yaml-only} ≠ ∅
+
+      ## Decision
+
+      If BOTH conditions met → set compression_mode=docs-only
+        - triage step: skip professor-triage skill; perform direct manifest summary instead
+        - plan step: skip release-plan skill; single-response plan instead
+        - deep-plan step: skip deep-plan skill; single-response implementation notes instead
+        - deep-verify step: skip deep-verify skill; perform self-review checklist instead
+        - Log: "[compression-mode] docs-only compression activated (scope={n}, all docs/yaml labels)"
+
+      Otherwise → set compression_mode=standard
+        - All pipeline steps execute normally with full skill spawns
+        - Log: "[compression-mode] standard mode (scope={n}, mixed labels or large scope)"
+
+      Output: compression_mode value as pipeline state for downstream steps.
+    description: "Evaluate docs-only compression eligibility; set compression_mode state"
+    depends_on: scope-selection
+
   - name: triage
     skill: professor-triage
-    description: "Cross-analysis triage with priority assessment (scoped to release manifest)"
-    depends_on: scope-selection
+    description: "Cross-analysis triage with priority assessment (scoped to release manifest) — skipped if compression_mode=docs-only"
+    depends_on: compression-mode-eval
 
   - name: plan
     skill: release-plan
-    description: "Release unit plan from triaged issues"
+    description: "Release unit plan from triaged issues — skipped if compression_mode=docs-only"
     depends_on: triage
 
   - name: deep-plan
     skill: deep-plan
-    description: "Research-validated implementation plan (research → plan → verify)"
+    description: "Research-validated implementation plan (research → plan → verify) — skipped if compression_mode=docs-only"
     depends_on: plan
 
   - name: implement
@@ -165,7 +229,7 @@ steps:
 
   - name: deep-verify
     skill: deep-verify
-    description: "Multi-angle release quality verification"
+    description: "Multi-angle release quality verification — self-review checklist only if compression_mode=docs-only"
     depends_on: verify-build
 
   - name: release

package/templates/CLAUDE.md

@@ -86,6 +86,7 @@ oh-my-customcode로 구동됩니다.
 | R013 | Ecomode | 배치 작업 토큰 효율성 |
 | R019 | Ontology-RAG 라우팅 | 라우팅 스킬의 ontology-RAG enrichment |
 | R022 | Wiki 동기화 | 에이전트/스킬/룰/가이드 변경 시 위키 자동 동기화 |
+| R023 | Verification Ladder | 검증 비용 ladder: Tier 1(결정론적) → Tier 4(human) shift-left |
 
 ### MAY (선택)
 | ID | 규칙 | 설명 |