[deploy] admin-v1.1.6

.agents/skills/debug/SKILL.md

@@ -157,10 +157,23 @@ After fixing a bug:
 - [ ] Test doubles updated if interfaces changed
 - [ ] Original failure scenario verified end-to-end
 
+## Runtime Boundary — HARD STOP
+
+This skill ENDS after the Verification checklist and final report are completed.
+
+For codex and any runtime without an enforced skill-return boundary:
+- MUST stop the assistant turn here.
+- MUST NOT invoke, load, or execute any next GSL skill in the same response turn.
+- MUST NOT continue into `/next-flow`, `/define`, `/plan`, `/implement`, `/test`, `/verify`, `/debug`, or `/self-review`.
+- MAY print exactly one suggested next command as plain text.
+- MUST wait for the user's next message before running any next skill.
+
+If the user says only "continue", treat that as permission to report the next recommended command, not permission to execute it.
+
 ---
 
 ## Lifecycle Integration
 
 **Before this skill:** if `plan/conventions.md` does not exist, run `/scan-conventions` first — analysis relies on knowing the project's actual conventions (naming, layering, test patterns, build system).
 
-**After this skill:** invoke `/next-flow` to diagnose lifecycle state and propose the next command. `/next-flow` auto-progresses read-only verification only and never writes files. Note: `/plan` is a Claude Code UI command and cannot be auto-invoked — the user must type it themselves; `/next-flow` will print a notice in that case.
+**After this skill:** the next GSL skill is started by the user, not by this skill — see the Runtime Boundary section above. `/next-flow` may be suggested for lifecycle diagnosis but is not auto-invoked. Runtime note: some environments expose slash commands as UI commands; codex loads GSL skills from `.agents/skills/`. In both cases, the next GSL skill requires a new explicit user message.

.agents/skills/define/SKILL.md

@@ -165,10 +165,23 @@ Before proceeding to `/plan`:
 - [ ] Impact scope identifies affected modules / surfaces / tests
 - [ ] User has explicitly approved the Overview
 
+## Runtime Boundary — HARD STOP
+
+This skill ENDS after the Verification checklist and final report are completed.
+
+For codex and any runtime without an enforced skill-return boundary:
+- MUST stop the assistant turn here.
+- MUST NOT invoke, load, or execute any next GSL skill in the same response turn.
+- MUST NOT continue into `/next-flow`, `/define`, `/plan`, `/implement`, `/test`, `/verify`, `/debug`, or `/self-review`.
+- MAY print exactly one suggested next command as plain text.
+- MUST wait for the user's next message before running any next skill.
+
+If the user says only "continue", treat that as permission to report the next recommended command, not permission to execute it.
+
 ---
 
 ## Lifecycle Integration
 
 **Before this skill:** if `plan/conventions.md` does not exist, run `/scan-conventions` first — analysis relies on knowing the project's actual conventions (naming, layering, test patterns, build system).
 
-**After this skill:** invoke `/next-flow` to diagnose lifecycle state and propose the next command. `/next-flow` auto-progresses read-only verification only and never writes files. Note: `/plan` is a Claude Code UI command and cannot be auto-invoked — the user must type it themselves; `/next-flow` will print a notice in that case.
+**After this skill:** the next GSL skill is started by the user, not by this skill — see the Runtime Boundary section above. `/next-flow` may be suggested for lifecycle diagnosis but is not auto-invoked. Runtime note: some environments expose slash commands as UI commands; codex loads GSL skills from `.agents/skills/`. In both cases, the next GSL skill requires a new explicit user message.

.agents/skills/implement/SKILL.md

@@ -125,11 +125,12 @@ feat: rating statistics service
 **Cycle per Task:**
 1. Implement Slice → compile / type-check → pass? continue : fix
 2. Repeat until all Slices in the Task are done
-3. Run `/self-review` on the Task's changes
-4. Run `/verify standard` (unit + arch rules)
+3. Run a self-review pass on the Task's changes (the 5-axis check). For a full independent review, stop and recommend `/self-review` as a separate invocation — do not run its full skill body inside `/implement`.
+4. Run the compile/type-check and unit-test commands needed for this Task. For full `/verify standard` or `/verify full`, stop and recommend `/verify` as a separate invocation.
 5. Commit with descriptive message
 6. Update plan document (check off Task, add to Progress Log)
-7. Move to next Task
+7. **HARD STOP after this Task.** Do NOT start the next Task in the same response turn. Report: completed Task number/title, verification evidence, changed files, next recommended Task. Then wait for the user's next message.
+   - Exception: proceed to the next Task in the same turn ONLY IF the user explicitly named multiple Tasks for continuous execution in their request (e.g. "do Tasks 1 through 3"). An ambiguous "continue" / "진행하자" is NOT such permission.
 
 ### Phase 4: Final Verification
 
@@ -211,6 +212,9 @@ During Slice execution, load ONLY the references for the current Slice's module.
 - Skipping Phase 0 (Explore) and jumping straight to coding
 - Multiple unrelated changes in a single Task
 - Task completed without running `/self-review`
+- Starting Task N+1 after completing Task N without explicit user permission for continuous execution
+- Treating an ambiguous "continue" as permission to finish all remaining Tasks
+- Running `/test`, `/verify`, or `/self-review` as full skill bodies inside `/implement` instead of stopping at that skill boundary
 
 ## Verification
 
@@ -225,10 +229,23 @@ After completing all Tasks for a feature:
 - [ ] Build / package succeeds
 - [ ] Plan document updated (all Tasks checked, Progress Log filled)
 
+## Runtime Boundary — HARD STOP
+
+This skill ENDS after the Verification checklist and final report are completed.
+
+For codex and any runtime without an enforced skill-return boundary:
+- MUST stop the assistant turn here.
+- MUST NOT invoke, load, or execute any next GSL skill in the same response turn.
+- MUST NOT continue into `/next-flow`, `/define`, `/plan`, `/implement`, `/test`, `/verify`, `/debug`, or `/self-review`.
+- MAY print exactly one suggested next command as plain text.
+- MUST wait for the user's next message before running any next skill.
+
+If the user says only "continue", treat that as permission to report the next recommended command, not permission to execute it.
+
 ---
 
 ## Lifecycle Integration
 
 **Before this skill:** if `plan/conventions.md` does not exist, run `/scan-conventions` first — analysis relies on knowing the project's actual conventions (naming, layering, test patterns, build system).
 
-**After this skill:** invoke `/next-flow` to diagnose lifecycle state and propose the next command. `/next-flow` auto-progresses read-only verification only and never writes files. Note: `/plan` is a Claude Code UI command and cannot be auto-invoked — the user must type it themselves; `/next-flow` will print a notice in that case.
+**After this skill:** the next GSL skill is started by the user, not by this skill — see the Runtime Boundary section above. `/next-flow` may be suggested for lifecycle diagnosis but is not auto-invoked. Runtime note: some environments expose slash commands as UI commands; codex loads GSL skills from `.agents/skills/`. In both cases, the next GSL skill requires a new explicit user message.

.agents/skills/implement/references/languages/bottlenote-patterns.md

@@ -0,0 +1,114 @@
+# bottlenote-patterns
+
+> bottle-note 한정 implement 함정 / 실수 방지 부록.
+> 일반 컨벤션은 `plan/conventions.md`, 일반 Java/Spring 패턴은 `java-spring.md` 참조.
+> 본 파일은 GSL sync 와 무관한 프로젝트 특화 reference.
+
+## When to load
+
+`/implement` Phase 0 에서 `plan/conventions.md` 와 함께 자동 참조. Repository / RestDocs / 이벤트 발행 관련 작업 시 우선 확인.
+
+---
+
+## P4. InMemory 구현체 갱신 누락 (Repository interface 변경 시)
+
+### 함정
+
+Repository interface 에 메서드 추가·시그니처 변경 시, JPA 구현체만 갱신하고 **InMemory 테스트 구현체 갱신을 잊는 경우** — PR #578 rebase 등에서 반복 발생.
+
+### 점검 경로 (필수)
+
+```text
+bottlenote-mono/src/test/java/app/bottlenote/{domain}/fixture/InMemory{Domain}Repository.java
+bottlenote-product-api/src/test/java/app/bottlenote/{domain}/fixture/InMemory{Domain}Repository.java
+```
+
+Admin / 외부 모듈에서도 동일 패턴이 추가될 수 있음:
+```text
+bottlenote-admin-api/src/test/kotlin/.../fixture/InMemory{Domain}Repository.kt
+```
+
+### 절차 (`/implement` 또는 `/self-review` 단계)
+
+1. Repository interface diff 의 변경 메서드 시그니처 확인
+2. 위 두 경로에서 `InMemory{Domain}Repository` 검색
+3. 동일 시그니처 구현 추가 (도메인 객체 상태 변경의 단순 모방으로 충분)
+4. 해당 fake 를 사용하는 단위 테스트가 컴파일·통과하는지 확인
+
+---
+
+## P3 흡수. RestDocs / asciidoctor 워크플로우
+
+원래 `/docs` 신규 스킬 후보였던 흐름. GSL 9 스킬로 분배:
+
+### 책임 분리
+
+| 단계 | GSL 스킬 | bottle-note 구체 |
+|------|---------|------------------|
+| RestDocs 테스트 작성 | `/test` | `Rest{Domain}ControllerDocsTest` (product), `Admin{Domain}ControllerDocsTest` (admin) |
+| `.adoc` 인덱스·조각 작성 | `/implement` Slice | `bottlenote-{product\|admin}-api/src/docs/asciidoc/**.adoc` |
+| asciidoctor 빌드 검증 | `/verify` L3 | `./gradlew asciidoctor` |
+| admin default test 검증 | `/verify` L3 | `./gradlew :bottlenote-admin-api:test` |
+
+### admin default test 케이스 (PR #578 사례)
+
+- product 의 `Rest*DocsTest` 는 `@Tag("integration")` 로 분리돼 root `test` 에서 제외됨
+- 그러나 **admin 의 `Admin*DocsTest` 는 default `:bottlenote-admin-api:test` 에서 함께 도는 경우 있음**
+- 따라서 admin RestDocs 수정 시 `./gradlew :bottlenote-admin-api:test` 도 반드시 확인 (누락 시 PR drift 발생)
+
+### `/verify full` 권장 명령 시퀀스
+
+```bash
+./gradlew check_rule_test                  # baseline rule check
+./gradlew unit_test integration_test       # 표준 단위·통합
+./gradlew admin_integration_test           # admin 통합
+./gradlew :bottlenote-admin-api:test       # admin default test (DocsTest 누락 방지)
+./gradlew asciidoctor                      # RestDocs HTML 빌드 검증
+```
+
+---
+
+## P5. (GSL 참조) `Projections.constructor` 와 local record
+
+→ 이미 GSL `java-spring.md` Tier 3 trap 에 명시됨. 본 plan 의 SC3 는 GSL 표준으로 충족.
+요약: `Projections.constructor()` 인자에 쓰는 record 는 **메서드 본문 안 local record 금지**, 반드시 클래스/인터페이스 레벨 정의.
+
+---
+
+## 추가 함정
+
+### publishEvent drift
+
+- `ApplicationEventPublisher.publishEvent(...)` 호출이 도메인 로직 수정·리팩토링 중 **소리 없이 누락** 되는 경우
+- 검출: `/self-review` Correctness 축에서 "이 변경이 기존 publishEvent 호출을 제거·우회하지 않았는가?"
+- 보조: `git diff` 시 `publishEvent` 키워드 grep 으로 누락 여부 확인
+
+### Facade ↔ Service 경계 drift
+
+- Cross-domain 접근은 `{Domain}Facade` 경유 원칙
+- 일부 service 가 facade 도 구현하는 기존 drift 존재 (`conventions.md` Comparison 표 참조)
+- 새 작업: facade 분리 권장. 기존 drift 는 plan 에 명시되지 않은 한 정리 대상 아님
+
+### `@ThirdPartyService` 사용 시점
+
+→ GSL `java-spring.md` External Integration Layer 섹션 참조. bottle-note 에서는 `app.external` 또는 도메인 외부 패키지에 위치.
+- 테스트 격리: 외부 client 의 fake (예: `FakeProfanityClient`, `FakeWebhookRestTemplate`) 사용
+- 트랜잭션 불필요, `@Service` 포함하지만 도메인 service 와 의도 명확히 분리
+
+### Batch 모듈 특이사항 (`/define` / `/plan` 시 주의)
+
+`conventions.md` 의 "Batch-Specific Current Conventions" 참조. 핵심만:
+- `bottlenote-batch` 는 `testFixtures` 대신 `mono` test output 직접 사용 (drift)
+- `git.environment-variables` 가 main+test 양쪽 resources 에 포함됨
+- main resource 에 하드코딩된 JWT secret / nonce salt default 존재
+- batch source test 가 0 (테스트 공백 인지 필요)
+
+위 항목들은 **알려진 drift** — 새 plan 이 명시적으로 정리하지 않는 한 보존.
+
+---
+
+## Maintenance
+
+- 항목이 GSL 표준 `java-spring.md` 에 흡수되면 본 파일에서 제거 → 중복 방지
+- 새 함정 발견 시 추가 (PR 사례 인용 권장)
+- conventions.md 가 자체 갱신될 때 본 파일과 충돌하는 부분은 conventions.md 우선

.agents/skills/next-flow/SKILL.md

@@ -28,7 +28,7 @@ Hard policy (from codex consultation): *auto-progression is restricted to read-o
 - Before `/define` (no plan document yet — start with `/define` directly)
 - When a `/debug` recovery is in progress (let `/debug` finish first)
 - When the user has already named the next step explicitly
-- For `/plan` transition — `/plan` is a Claude Code UI command and cannot be auto-invoked from a skill; this skill will only print a notice telling the user to type `/plan` themselves
+- For any next GSL command transition — print the command name and stop. In slash-command UI runtimes, tell the user to type that command directly; in codex, tell the user to send a new message containing that skill name.
 
 ## Process
 
@@ -45,7 +45,7 @@ Read Overview, Tasks, Progress Log. Classify the work into exactly one of:
 
 | Signal | Current step | Next step |
 |--------|--------------|-----------|
-| Plan doc has Overview only, no Tasks | post-`/define` | `/plan` (UI command — print notice) |
+| Plan doc has Overview only, no Tasks | post-`/define` | `/plan` (print command and stop) |
 | Tasks exist, Progress Log empty | post-`/plan` | `/implement` |
 | Some Tasks done, more remain | mid-`/implement` | continue `/implement` with next Task |
 | All Tasks done, no integration tests written | post-`/implement` | `/test` (integration tests) |
@@ -97,10 +97,10 @@ Preconditions: PASS
 [Run it? You can type /<command> yourself, or reply "yes" to have me run it only if it is read-only.]
 ```
 
-For `/plan` and other UI commands, replace the last line with:
+For any next GSL command, do not run it in this turn. Replace the last line with runtime-neutral guidance:
 
 ```
-/plan is a Claude Code UI command — please type it in the input box yourself; I cannot trigger it from a skill.
+To continue, send a new message containing this command. In slash-command UI runtimes, type the command directly; in codex, send the skill name in a new message.
 ```
 
 ## Common Rationalizations
@@ -111,16 +111,16 @@ For `/plan` and other UI commands, replace the last line with:
 | "User said 'continue', so I can just push through commit" | Commit is a write boundary and an externally-visible action. Propose, never auto-run. |
 | "It's only a small edit to plan doc Progress Log" | Plan document is the single source of truth. Only `/implement` updates it after a Task commit. |
 | "/verify full passed — let me start the next feature" | Closing the current feature (stamp + plan/complete/ move) is a manual decision. Propose, don't act. |
-| "I'll auto-call /plan since /implement needs it" | `/plan` is a UI command. You cannot. Print the notice and stop. |
+| "I'll auto-call the next GSL command since it is obvious" | Next GSL commands are runtime boundaries. Print the command name and stop. |
 
 ## Red Flags
 
 - Auto-progression has caused any file to be created, edited, or deleted (immediate STOP — this skill should never do that)
 - Multiple next-step candidates proposed (the table in Step 2 should yield exactly one)
 - Preconditions skipped because "the user implied it was fine"
-- Proposing `/plan` for auto-invocation (impossible — print notice instead)
-- **Bypassing the UI-command notice by invoking `/plan` (or any other Claude Code UI command) through Bash, shell scripts, `subprocess`, MCP tools, or other wrappers — anti-pattern, STOP.** The notice is the only correct output for UI-command transitions.
-- The next command requires user approval (e.g., post-`/define` → `/plan`) and you tried to invoke it anyway
+- Proposing any next GSL command for same-turn execution instead of printing the command name and stopping
+- **Bypassing the runtime-boundary notice by invoking the next GSL command through Bash, shell scripts, `subprocess`, MCP tools, or other wrappers — anti-pattern, STOP.** The notice is the only correct output for GSL command transitions.
+- The next command requires user approval (e.g., post-`/define` → `/plan`) and you tried to execute it anyway
 
 ## Verification
 
@@ -130,5 +130,18 @@ After running:
 - [ ] All preconditions were checked in a read-only way
 - [ ] If any auto-action was run, it was read-only only (verify / status / log / diff)
 - [ ] No file was created, edited, or deleted by this skill
-- [ ] If the next step is `/plan`, the UI-command notice was printed instead of attempting auto-invoke
+- [ ] If the next step is a GSL command, the runtime-boundary notice was printed instead of attempting same-turn execution
 - [ ] The proposal includes the reason and the user's explicit invocation prompt
+
+## Runtime Boundary — HARD STOP
+
+This skill ENDS after the Verification checklist and final report are completed.
+
+For codex and any runtime without an enforced skill-return boundary:
+- MUST stop the assistant turn here.
+- MUST NOT invoke, load, or execute any next GSL skill in the same response turn.
+- MUST NOT continue into `/next-flow`, `/define`, `/plan`, `/implement`, `/test`, `/verify`, `/debug`, or `/self-review`.
+- MAY print exactly one suggested next command as plain text.
+- MUST wait for the user's next message before running any next skill.
+
+If the user says only "continue", treat that as permission to report the next recommended command, not permission to execute it.

.agents/skills/plan/SKILL.md

@@ -180,10 +180,23 @@ Before starting `/implement`, confirm:
 - [ ] Checkpoints exist between major groups of Tasks
 - [ ] User has explicitly approved the task list
 
+## Runtime Boundary — HARD STOP
+
+This skill ENDS after the Verification checklist and final report are completed.
+
+For codex and any runtime without an enforced skill-return boundary:
+- MUST stop the assistant turn here.
+- MUST NOT invoke, load, or execute any next GSL skill in the same response turn.
+- MUST NOT continue into `/next-flow`, `/define`, `/plan`, `/implement`, `/test`, `/verify`, `/debug`, or `/self-review`.
+- MAY print exactly one suggested next command as plain text.
+- MUST wait for the user's next message before running any next skill.
+
+If the user says only "continue", treat that as permission to report the next recommended command, not permission to execute it.
+
 ---
 
 ## Lifecycle Integration
 
 **Before this skill:** if `plan/conventions.md` does not exist, run `/scan-conventions` first — analysis relies on knowing the project's actual conventions (naming, layering, test patterns, build system).
 
-**After this skill:** invoke `/next-flow` to diagnose lifecycle state and propose the next command. `/next-flow` auto-progresses read-only verification only and never writes files. Note: `/plan` is a Claude Code UI command and cannot be auto-invoked — the user must type it themselves; `/next-flow` will print a notice in that case.
+**After this skill:** the next GSL skill is started by the user, not by this skill — see the Runtime Boundary section above. `/next-flow` may be suggested for lifecycle diagnosis but is not auto-invoked. Runtime note: some environments expose slash commands as UI commands; codex loads GSL skills from `.agents/skills/`. In both cases, the next GSL skill requires a new explicit user message.