oh-my-customcode 0.105.0 → 0.106.0

package/README.md

@@ -222,7 +222,7 @@ Key rules: R010 (orchestrator never writes files), R009 (parallel execution mand
 
 ---
 
-### Guides (42)
+### Guides (43)
 
 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.
 
@@ -279,7 +279,7 @@ your-project/
 │   ├── specs/                  # Extracted canonical specs
 │   ├── contexts/               # 4 shared context files
 │   └── ontology/               # Knowledge graph for RAG
-└── guides/                     # 42 reference documents
+└── guides/                     # 43 reference documents
 ```
 
 ---

package/dist/cli/index.js

@@ -2334,7 +2334,7 @@ var init_package = __esm(() => {
     workspaces: [
       "packages/*"
     ],
-    version: "0.105.0",
+    version: "0.106.0",
     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: "0.105.0",
+  version: "0.106.0",
   description: "Batteries-included agent harness for Claude Code",
   type: "module",
   bin: {

package/package.json

@@ -3,7 +3,7 @@
   "workspaces": [
     "packages/*"
   ],
-  "version": "0.105.0",
+  "version": "0.106.0",
   "description": "Batteries-included agent harness for Claude Code",
   "type": "module",
   "bin": {

package/templates/.claude/rules/MUST-agent-design.md

@@ -232,7 +232,35 @@ Skills persist output to `.claude/outputs/sessions/{YYYY-MM-DD}/{skill-name}-{HH
 
 ### Sensitive Path Handling
 
-CC treats `.claude/` as a sensitive directory. The sensitive-path check runs **above** `bypassPermissions` and `Bash(*)` allow rules, so Bash-based file operations (`cp`, `mkdir`) on `.claude/` paths trigger permission prompts regardless of settings. Subagents and skills that modify `.claude/` files SHOULD prefer `Write`/`Edit` tools over `Bash(cp)`/`Bash(mkdir)`, and projects SHOULD add `Write(.claude/**)` and `Edit(.claude/**)` to `.claude/settings.local.json` allow list.
+CC treats `.claude/` as a sensitive directory, enforced across **all tool categories** — Bash, Write, and Edit. The sensitive-path check runs **above** `bypassPermissions` and explicit allow rules (e.g., `Write(.claude/**)`), so operations on sensitive paths may trigger permission prompts regardless of settings.
+
+#### Scope
+
+| Path pattern | Sensitive? | Applies to |
+|--------------|-----------|-----------|
+| `.claude/**` | Yes | Bash (`cp`, `mkdir`, `rm`), Write, Edit |
+| `templates/.claude/**` | Yes | Bash, Write, Edit (confirmed v2.1.116+, 3x repeat v0.105.0 session) |
+| `.claude/outputs/**` | No (artifact convention) | — |
+
+#### Behavior
+
+| Tool | Allow rule | Result |
+|------|-----------|--------|
+| `Bash(cp ...)` on `.claude/` | `Bash(*)` allowed | Prompt (sensitive-path overrides) |
+| `Write(.claude/**)` | `Write(.claude/**)` allowed | Prompt (sensitive-path overrides) |
+| `Edit(templates/.claude/**)` | `Edit(templates/.claude/**)` allowed | Prompt (sensitive-path overrides) |
+
+#### Recommended practice
+
+1. **Prefer `Write`/`Edit` over `Bash(cp)`/`Bash(mkdir)`** — even though both trigger prompts, `Write`/`Edit` provide better auditability and avoid shell injection risk
+2. **Add allow rules defensively** — `Write(.claude/**)`, `Edit(.claude/**)`, `Write(templates/.claude/**)`, `Edit(templates/.claude/**)` in `.claude/settings.local.json`. Rules may not bypass sensitive-path check but document intent and aid future CC behavior changes
+3. **Accept interactive prompts as a release-pipeline constraint** — `templates/.claude/` sync during release automation requires human approval; plan release windows accordingly
+4. **This is CC design behavior, not a bug** — sensitive-path check is a defense-in-depth layer. File upstream as a documentation request (not bug report) if behavior is unclear
+
+#### Cross-references
+
+- `feedback_sensitive_path.md` — session memory with Bash + Write scope (#960, #961, #981)
+- `feedback_templates_claude_glob.md` — `.claude/**` glob does not cover `templates/.claude/**`, separate allow rules required
 
 ### Artifact Channel Protocol
 

package/templates/CLAUDE.md

@@ -103,7 +103,7 @@ oh-my-customcode로 구동됩니다.
 | 릴리즈 | `/pipeline auto-dev`, `/omcustom-release-notes`, `/release-plan` | 자동 개발, 릴리즈 노트 |
 | 리서치 | `/research`, `/scout`, `/deep-plan`, `/omcustom:agora` | 병렬 분석, URL 평가, 연구 계획 |
 | 메모리 | `/memory-save`, `/memory-recall` | 세션 메모리 관리 |
-| 최적화 | `/token-efficiency-audit` | 토큰 효율 감사 (3계층 방어 스택) |
+| 최적화 | `/token-efficiency-audit` | 토큰 효율 감사 (5계층 방어 스택) |
 | 시스템 | `/omcustom:lists`, `/omcustom:status`, `/omcustom:help` | 전체 목록, 상태, 도움말 |
 
 > 전체 커맨드 목록 (60+ 커맨드): `/omcustom:lists` 실행
@@ -119,7 +119,7 @@ project/
 |   +-- rules/                   # 전역 규칙 (R000-R022)
 |   +-- hooks/                   # 훅 스크립트 (보안, 검증, HUD)
 |   +-- contexts/                # 컨텍스트 파일 (ecomode)
-+-- guides/                      # 레퍼런스 문서 (42 토픽)
++-- guides/                      # 레퍼런스 문서 (43 토픽)
 ```
 
 ## 오케스트레이션
@@ -236,6 +236,7 @@ Claude Code의 Agent Teams 기능이 활성화되어 있으면 (`CLAUDE_CODE_EXP
 | 플러그인 | 소스 | 용도 |
 |----------|------|------|
 | cc-token-saver | ww-w-ai/cc-token-saver | 토큰/비용 최적화 (Token Guardian, /continue, /usage-view) |
+| caveman | JuliusBrussee/caveman | 영어 출력 토큰 압축 (코드 리뷰·커밋 메시지). R013 ecomode 미적용 구간 보완. 한국어 컨텍스트(R000)에는 ecomode 우선. `lite`/`full` 권장, `ultra`/`文言文` 가독성 저하 주의 |
 
 ### 권장 MCP 서버
 

package/templates/guides/agent-design/ralph-loop-pattern.md

@@ -0,0 +1,32 @@
+# Ralph Loop Pattern — 세션 경계 지속 진화
+
+> 출처: Q00/ouroboros Ralph Loop 패턴 (INTEGRATE #966)
+> 도입일: v0.106.0
+
+## Context
+
+LLM 에이전트는 세션 종료 시 누적된 맥락(tacit knowledge, 실패 교훈, 선호 패턴)을 잃는다. Ralph Loop는 세션 간 경계를 넘는 지속 진화 메커니즘으로, 각 세션의 실패/성공 sticky patterns를 다음 세션의 초기 컨텍스트로 주입한다.
+
+## Claude Code 통합
+
+- **claude-mem MCP**: feedback memories + session archives를 영속 저장소로
+- **sys-memory-keeper 에이전트**: 세션 종료 시 Ralph Loop 요약 수행
+- **R011 Memory Integration**: native auto-memory를 Ralph Loop의 primary writer로
+
+## Pattern
+
+1. **Bootstrap**: 세션 시작 시 MEMORY.md에서 이전 Ralph Loop 요약 로드
+2. **Evolve**: 세션 진행 중 발견한 새 패턴/실패를 feedback memory에 기록
+3. **Compact**: 세션 종료 시 sys-memory-keeper가 Ralph Loop 요약을 MEMORY.md 업데이트
+4. **Persist**: claude-mem MCP에 long-term save (cross-session search)
+
+## Anti-patterns
+
+- 세션마다 처음부터 다시 시작하는 sessionless 모드 (R011 native memory 사용)
+- 모든 세션 로그를 저장하는 brute-force persistence (selective feedback만)
+
+## References
+
+- R011 SHOULD-memory-integration.md
+- sys-memory-keeper 에이전트
+- Q00/ouroboros `ooo interview/run/ralph` CLI

package/templates/guides/claude-code/14-token-efficiency.md

@@ -1,4 +1,4 @@
-# Token Efficiency — Three-Layer Defense Stack
+# Token Efficiency — Five-Layer Defense Stack
 
 > **Source**: [Claude Code & Codex token efficiency by settings adjustment](https://www.stdy.blog/increasing-token-efficiency-by-setting-adjustment-in-claude-and-codex/)
 > **Reference baseline**: Claude Code v2.1.114+ / Codex v0.121.0+
@@ -7,7 +7,7 @@
 
 Token spend in Claude Code is not purely a function of task complexity. A significant portion of token consumption occurs through structural overhead: auto-injected git instructions, IDE file listings, tool output that exceeds what the model actually needs, and session state reloaded unnecessarily across turns.
 
-The three-layer defense stack addresses this overhead at three distinct points in the session lifecycle:
+The five-layer defense stack addresses this overhead at distinct points in the session lifecycle:
 
 ```
 ┌─────────────────────────────────────────────────────────────────┐
@@ -22,6 +22,10 @@ The three-layer defense stack addresses this overhead at three distinct points i
 ├─────────────────────────────────────────────────────────────────┤
 │  Layer 4: playwright-compress (MCP OUTPUT INTELLIGENCE)         │
 │  PostToolUse hook — Haiku summarization of browser MCP output   │
+├─────────────────────────────────────────────────────────────────┤
+│  Layer 5: caveman (OUTPUT STYLE COMPRESSION)                    │
+│  SessionStart hook — primitive-speech output reduces token spend│
+│  English scenarios only — R000 Korean-first contexts use Layer 2│
 └─────────────────────────────────────────────────────────────────┘
 ```
 
@@ -153,6 +157,7 @@ Disabling `includeGitInstructions` removes git workflow guidance from the contex
 | R001 Safety | `apply-ci` mode is Risk Level High — requires user confirmation before applying. |
 | R012 HUD Statusline | Statusline shows CTX% — effective measure of Layer 2+3 combined impact. |
 | playwright-compress | Layer 4 hook — complements Layer 3 MAX_MCP_OUTPUT_TOKENS with intelligent lossless compression. |
+| caveman | Layer 5 style compression — orthogonal to Layer 2 (volume) and Layer 4 (MCP). R000 Korean-first: use Layer 2 as primary. |
 
 ---
 
@@ -175,6 +180,43 @@ Hook trigger: `mcp__playwright__.*` and `mcp__claude-in-chrome__.*`
 
 ---
 
+## Layer 5: caveman (Output Style Compression)
+
+External plugin (JuliusBrussee/caveman, 41.6k stars, MIT) that rewrites Claude responses into compressed "primitive speech" via a SessionStart hook.
+
+**Measured savings:**
+- ~75% output token reduction
+- ~46% input token reduction (context accumulation effect)
+
+**Intensity levels:**
+
+| Level | Description | Recommendation |
+|-------|-------------|----------------|
+| `lite` | Light compression, high readability | Recommended |
+| `full` | Moderate compression | Recommended |
+| `ultra` | Heavy compression, reduced readability | Caution |
+| `文言文` | Classical Chinese style | Avoid for shared sessions |
+
+**Installation:** `/plugin install caveman` (requires marketplace access)
+
+**When to use:**
+
+| Scenario | Layer to apply |
+|----------|---------------|
+| English-only output sessions (code review, commit messages) | Layer 5 (caveman) |
+| Korean-first sessions (R000 contexts) | Layer 2 (R013 ecomode) — caveman articles/prepositions not present in Korean |
+| Mixed sessions | Layer 2 primary, Layer 5 optional |
+
+**Coexistence with other layers:**
+
+- Layer 5 compresses output style; Layer 2 (ecomode) compresses output volume. These are orthogonal — both can be active simultaneously.
+- Layer 5 applies globally to all responses; Layer 2 activates conditionally (4+ tasks threshold).
+- For Korean-first projects following R000, Layer 2 provides the primary compression benefit. Layer 5's article-removal compression is largely ineffective against Korean (article-free language).
+
+**R000 compatibility note:** caveman's compression relies on English grammatical structure (removing articles, prepositions). Korean lacks these structures, so compression rates are significantly lower in Korean-dominant sessions. Use `lite` or `full` mode only; avoid `ultra`/`文言文` in shared or multilingual sessions.
+
+---
+
 ## Cross-References
 
 - `guides/claude-code/13-cli-flags.md` — CLI flags for non-interactive/CI invocation
@@ -185,3 +227,4 @@ Hook trigger: `mcp__playwright__.*` and `mcp__claude-in-chrome__.*`
 - `.claude/skills/update-config/` — Generic settings.json manipulation (broader scope)
 - `.claude/skills/playwright-compress/SKILL.md` — Layer 4 MCP output compression hook
 - `.claude/hooks/scripts/playwright-compress.sh` — Layer 4 hook script
+- `https://github.com/JuliusBrussee/caveman` — Layer 5 caveman plugin (external)

package/templates/manifest.json

@@ -1,6 +1,6 @@
 {
-  "version": "0.105.0",
-  "lastUpdated": "2026-04-18T00:00:00.000Z",
+  "version": "0.106.0",
+  "lastUpdated": "2026-04-24T07:30:00.000Z",
   "components": [
     {
       "name": "rules",
@@ -24,7 +24,7 @@
       "name": "guides",
       "path": "guides",
       "description": "Reference documentation",
-      "files": 42
+      "files": 43
     },
     {
       "name": "hooks",