oh-my-customcode 0.179.0 → 0.181.0

package/dist/cli/index.js

@@ -241,7 +241,7 @@ var init_package = __esm(() => {
     workspaces: [
       "packages/*"
     ],
-    version: "0.179.0",
+    version: "0.181.0",
     description: "Batteries-included agent harness for Claude Code",
     type: "module",
     bin: {

package/dist/index.js

@@ -2031,7 +2031,7 @@ var package_default = {
   workspaces: [
     "packages/*"
   ],
-  version: "0.179.0",
+  version: "0.181.0",
   description: "Batteries-included agent harness for Claude Code",
   type: "module",
   bin: {

package/package.json

@@ -3,7 +3,7 @@
   "workspaces": [
     "packages/*"
   ],
-  "version": "0.179.0",
+  "version": "0.181.0",
   "description": "Batteries-included agent harness for Claude Code",
   "type": "module",
   "bin": {

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

@@ -98,6 +98,21 @@ When the gate resolves to **Agent Tool** for a 3+ agent dispatch (e.g. mechanica
 
 Origin: #1293 (Session 110 retrospective, Low).
 
+#### Gate Transparency Scope — Agent Teams Enabled Only (#1341 ②)
+
+> Origin: #1341 찐빠 #2 (low-confidence) — 4+ 병렬 Agent Tool 스폰 시 `[N]` prefix는 표기했으나 "R018 게이트: … → Agent Tool 폴백" announce를 생략한 것을 자가 위반으로 의심. 그러나 R018은 조건부 규칙이라 Agent Teams 비활성 환경에서는 게이트 투명성 자체가 미적용이다.
+
+R018 전체(게이트 투명성 포함)는 Agent Teams 활성 시에만 적용되는 조건부 규칙이다 (`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` 또는 TeamCreate/SendMessage 도구 존재). Agent Teams **비활성** 환경에서는:
+
+- 게이트 투명성 announce 의무가 없다 — R009 `[N]` prefix 만으로 충분하다.
+- 3+ 병렬 Agent Tool 스폰을 "게이트 announce 누락"으로 자가 플래그하지 않는다 (false-positive 방지).
+
+게이트 투명성 announce는 Agent Teams가 활성이고 게이트가 Agent Tool로 해소될 때만 의무다.
+
+| Anti-pattern | Required |
+|--------------|----------|
+| Agent Teams 비활성 환경에서 게이트 announce 누락을 자가 위반으로 플래그 | 비활성 시 R009 `[N]` prefix 만으로 충분; 게이트 announce는 활성 환경 전용 |
+
 ### Spawn Completeness Check
 
 All members must be spawned in a single message. Partial spawning needs correction per R018 and R009.

package/templates/.claude/rules/MUST-completion-verification.md

@@ -156,6 +156,28 @@ Related memory records:
 
 Reference issues: #1188 item #8.
 
+### Interrupt ≠ Prior-Request Cancellation (#1341 ①)
+
+> Origin: #1341 찐빠 #1 — 사용자가 멀티라인 요청("스킬 FSD를 만들자 … 내용은 아래와 같다")을 두 줄에 나눠 보내려다 중간에 인터럽트했는데, 모델이 이를 "직전 요청 취소" 신호로 단정하고 즉시 다른 작업(/goal 자율 루프)으로 전환했다. 사용자가 "취소가 아니야, 두 줄이 한 번에 안 가서 인터럽트했던 것"으로 정정.
+
+**적용 범위 (비파괴 한정)**: 이 규칙은 **비파괴적** 직전/진행 중 요청에만 적용된다 (스킬/문서 생성, 분석, 비파괴 편집 등). 인터럽트된 작업이 **파괴적·비가역 작업**(R001 — `git reset --hard`, `git clean -fd`, `rm`, 터널/DNS/k8s/인프라 삭제 등)이면 이 규칙은 적용되지 않으며, 아래 Safety Carve-Out을 따른다.
+
+비파괴 작업에 한해: 사용자 인터럽트 직후 첫 메시지가 모호하면, 직전 요청을 "취소"로 단정하지 않는다. 인터럽트는 입력 교정·추가 입력·멀티라인 연속 입력 중단 등 다양한 의도일 수 있으며, 취소는 그중 하나일 뿐이다. 직전 요청 맥락이 살아있는 상태에서 인터럽트 의도를 단정해 다른 작업으로 전환하지 말고, 모호하면 한 번 확인한 뒤 (비파괴적 후속 처리를) 진행한다.
+
+**Safety Carve-Out — 파괴적 작업 (fail-closed, stop-first ask-after)**: 진행 중이던 작업이 파괴적·비가역 작업이면, 인터럽트 수신 시 의도가 모호하더라도 그 작업을 **먼저 즉시 중단(halt/abort)**한 뒤 의도를 확인한다. 재개는 명시적 재승인을 요구한다. 인터럽트의 핵심 가치는 emergency-stop이므로 파괴적 작업에서는 의도 명료화보다 정지가 우선한다(R001 우선). 여기서 "진행"은 파괴적 작업의 계속을 의미하지 않는다.
+
+**모호성 판정 신호** (둘 중 하나면 confirm; 둘 다 아니고 명확한 새 지시면 R003 Clear 경로로 즉시 처리해 과잉 확인 방지):
+- (a) 직전 요청이 미완(코드/내용 본문 미수신) 상태에서 인터럽트됨
+- (b) 인터럽트 후 첫 메시지가 직전 요청과 무관해 보이나 새 명령으로도 단정 불가
+
+| Anti-pattern | Required |
+|--------------|----------|
+| 인터럽트 직후 직전 요청을 "취소"로 단정하고 새 작업 실행 (비파괴 맥락) | 인터럽트 의도가 모호하면 직전 요청 맥락 유지 + 의도 1회 확인 후 비파괴적 후속 진행 |
+| 멀티라인/연속 입력 중간의 인터럽트를 "전체 취소"로 해석 | 추가 입력·교정 가능성 고려; 사용자 다음 메시지를 기다리거나 의도 확인 |
+| 파괴적 작업 진행 중 인터럽트를 "맥락 유지 후 계속"으로 처리 | 즉시 halt(fail-closed) 후 의도 확인; 재개는 명시적 재승인 (stop-first ask-after, R001 우선) |
+
+This is the interrupt-intent extension of Read-Before-Characterize ("actual intent ≠ assumed intent"), scoped to non-destructive context. **Applicability vs "Interrupt Priority Re-Ordering" (above)**: Priority Re-Ordering는 인터럽트가 **명확한 새 작업/룰 위반을 동반**할 때; 인터럽트 첫 메시지가 **모호**하면 본 섹션이 우선(확인 먼저). Cross-reference: R003 (Request Handling — Interrupt row; precedence Risky > Interrupt), R001 (파괴적 작업 halt 우선).
+
 ## Diagnostic Hypothesis Verification
 
 진단 단계에서 채택한 가설로 워크플로우/인프라/설정을 **영구 변경하기 전**, 가설을 실제 증거로 검증해야 한다. "그럴듯한 가설"을 검증 없이 영구 변경에 적용하면 잘못된 추정이 영구 부채로 남는다.
@@ -200,6 +222,19 @@ triage-dispatch.yml 실패 원인을 파일 Read 전에 "triaged 라벨 부재 +
 
 Origin: #1266 ④.
 
+### Proxy Signal vs Canonical Ground-Truth (#1336 ①②)
+
+> Origin: #1336 ①② — transcription was alarmed as "stopped" because `.txt` files looked stale, but the canonical DB had transcripts current to 06-09 21:30 (.txt is not the whisper collector's output — it emits only to the DB). Separately, SMS was over-diagnosed as "fully blocked" from one empty OneDrive XML path + a single 401, while the DB held 17 SMS rows ingested via the app path.
+
+When diagnosing pipeline/data state, verify the CANONICAL store (the authoritative datastore — DB, the system of record) BEFORE characterizing state from a secondary proxy (a `.txt`/file artifact) or a single ingestion path. Two failure modes share this meta-pattern:
+
+| Anti-pattern | Required |
+|--------------|----------|
+| Characterize pipeline health from a filesystem proxy (`.txt` presence/mtime) | Query the canonical store (DB transcript count/recency) first |
+| Generalize one ingestion path's failure (one empty XML / one 401) to "whole pipeline blocked" | Check the final landing store's count across ALL paths before concluding blockage |
+
+A single path's failure does NOT prove the whole multi-path pipeline is down. Confirm the system-of-record before alarming or dispatching reprocessing.
+
 ### Directory-Context Before Multi-Copy Unification/Deletion
 
 다중 사본(동일 파일이 N곳에 존재)을 통일하거나 삭제하기 전, 각 사본이 위치한 **디렉토리 전체 맥락**을 확인한다(`ls`로 형제 파일 파악). 사본 파일 하나만 보고 "orphan"·"stub"으로 특성화하면, 같은 디렉토리의 형제 파일(다른 역할을 가진)이 함께 덮이거나 맥락이 누락된다. Read-Before-Characterize를 파일 단위에서 디렉토리 단위로 확장한 규칙이다.
@@ -256,6 +291,21 @@ A CI publish/deploy step that LOGS an error has NOT necessarily failed — the s
 
 This is the publish-domain extension of Read-Before-Characterize ("actual outcome ≠ attempt"). Re-running a publish that actually succeeded risks duplicate-publish errors; permanently changing a workflow on a misdiagnosis is worse (cf. #1217 — npm E403 misdiagnosed as a `--provenance` conflict → wrong workflow change → repeated failure; real cause was token scope).
 
+### State-Change Claim → Live System Verification (#1335 ①)
+
+> Origin: #1335 ① — issue #101 (secretary teardown) was closed as "대체 완료·teardown 보류", but the secretary LaunchAgents (onedrive-bridge / calendar-worker / minikube-mount) were STILL running on the host. The user caught it ("secretary 리소스 다 내려가있는거 맞지?") — they were not.
+
+Before closing or marking-done an issue/task that CLAIMS an infrastructure or resource STATE change (a service stopped, a resource torn down, a deployment removed, a process killed), verify the ACTUAL live system state — not just that the change command was issued. "Issued the teardown" ≠ "the resource is down".
+
+| Claimed state change | Live ground-truth check |
+|----------------------|-------------------------|
+| LaunchAgent/service stopped | `launchctl list` / `systemctl status` shows it absent/inactive |
+| k8s resource torn down | `kubectl get <resource>` returns NotFound |
+| Container removed | `docker ps -a` does not list it |
+| Process killed | process check (`pgrep`/`ps`) returns empty |
+
+This is the infra/state extension of "actual outcome ≠ attempt". Closing on the command-issued assumption leaves orphaned running resources.
+
 ## Integration
 
 | Rule | Interaction |

package/templates/.claude/rules/MUST-parallel-execution.md

@@ -192,3 +192,15 @@ The list form mirrors the tool-call `[N]` prefix pattern and scales better to 3+
   ✓ agent-1: success
   ✗ agent-2: failed (reason)
 ```
+
+## Parallel Feature Integration Gate
+
+> Origin: #1335 ③ — parallel lang-kotlin-expert rounds each reported "build green", but the COMBINED runtime had a DataStore singleton crash, a Settings→Dashboard nav crash, a recording 400, and cursor pre-advance bugs — caught only by on-device testing.
+
+Per-subagent "build green" does NOT guarantee integrated runtime correctness. When parallel feature subagents edit interdependent code, the orchestrator MUST run an INTEGRATION verification gate after the parallel work merges — a combined build PLUS a runtime/smoke check (or device test for apps) — before declaring the feature done. Independent green builds can still combine into runtime crashes (shared singletons, navigation, API contracts).
+
+| Anti-pattern | Required |
+|--------------|----------|
+| Trust each parallel subagent's "build green" and declare done | Orchestrator runs a combined build + runtime/smoke gate on the merged result first |
+
+Cross-reference: R020 (actual outcome ≠ attempt; completion verification).

package/templates/.claude/rules/MUST-safety.md

@@ -70,6 +70,30 @@ Prefer a reversible action (disable/detach/stop) over delete when the goal can b
 
 > **Ask-before-scan (#1327 찐빠 #4)**: When a credential/token is needed, request it from the user BEFORE running BLIND/DISCOVERY credential scans (`env | grep`, repo-wide token greps), which trip the Credential Exploration classifier. Reading a SPECIFIC file the user named to obtain a value is not a discovery scan and is fine. If a scan trips the classifier, do not retry it (R010 Subagent Scope-Creep STOP Protocol).
 
+### Infra-Diagnostic File Checks — Metadata, Not Contents (#1334 ①)
+
+> Origin: #1334 ① — during a hermes 502 diagnosis, a `cat .env` + `credentials.json` key inspect was reflexively bundled into a diagnostic batch and tripped the Credential Exploration classifier. The secret values were never needed for the 502 diagnosis.
+
+When diagnosing infrastructure/health issues (502s, container state, env presence), file checks MUST use metadata-only commands — `ls -la` (existence, size, perms, mtime) — NEVER `cat .env`, `cat credentials.json`, or any command that reads secret CONTENTS or keys into the transcript. Confirming a file EXISTS is a metadata check; reading its values is a credential scan.
+
+| Anti-pattern | Required |
+|--------------|----------|
+| `cat .env` / inspect OAuth/credential keys to "confirm config present" during a health diagnosis | `ls -la .env` — existence/size/perms only; request a specific value from the user if genuinely needed |
+
+Cross-reference: the Ask-before-scan note above (discovery scans), R010 Subagent Scope-Creep STOP.
+
+### Standing User-Deny + Classifier Block → Immediate user-runs Switch (#1335 ④)
+
+> Origin: #1335 ④ — with a standing user constraint "절대 시크릿 건드리지 마" plus a classifier block, an `.env.local` edit (DATABASE_URL, LLM_MAX_TOKENS) was retried and blocked repeatedly instead of handing the edit to the user.
+
+When the user has a STANDING "don't touch X" constraint AND the safety classifier blocks an action on X even once, immediately switch to the `!` user-runs pattern — surface the exact command for the user to run themselves — and do NOT retry the blocked edit. A standing deny + one classifier trip is a hard signal to delegate to the user, not to find another path in.
+
+| Anti-pattern | Required |
+|--------------|----------|
+| Retry a blocked edit on a user-deny-listed path via a different mechanism | Stop after the first block; emit the command for the user to run via `!` and wait |
+
+Cross-reference: R010 Subagent Scope-Creep STOP Protocol (2-trip stop), R015 Failed Tool Re-Try Discipline.
+
 Cross-reference: R010 Subagent Scope-Creep STOP Protocol, R002 (permission tiers).
 
 ## Required Before Destructive Operations

package/templates/.claude/rules/SHOULD-interaction.md

@@ -26,6 +26,9 @@
 | Clear | Execute immediately |
 | Ambiguous | `[Confirm] Understood "{request}" as {interpretation}. Proceed?` |
 | Risky | `[Warning] This action has {risk}. Continue? Yes: {action} / No: Cancel` |
+| Interrupt (ambiguous first message) | Do NOT assume prior-request cancellation. `[Confirm] 인터럽트 의도 확인: 직전 "{request}" 취소인가요, 아니면 추가/교정 입력 중이신가요?` 단, 직전 요청이 Risky(파괴적)면 Risky 행이 우선(즉시 중단). 규범·예외는 R020 "Interrupt ≠ Prior-Request Cancellation"가 소유. |
+
+> **Precedence**: **Risky > Interrupt > Ambiguous > Clear** — 파괴적/위험 작업 중 인터럽트는 의도 모호와 무관하게 즉시 중단(R001 / R020 Safety Carve-Out)이 우선한다. 명확한 새 지시인 인터럽트는 Clear 경로로 즉시 처리(과잉 확인 금지).
 
 ## Multiple Tasks
 

package/templates/.claude/rules/SHOULD-memory-integration.md

@@ -21,6 +21,18 @@ Agent frontmatter `memory: project|user|local` enables persistent memory:
 | `project` | `.claude/agent-memory/<name>/` | Yes |
 | `local` | `.claude/agent-memory-local/<name>/` | No |
 
+## Subagent memory:project Source-Tree Pollution Guard
+
+> Origin: #1335 ② — lang-kotlin-expert with `memory: project`, working in a Kotlin source subdirectory, wrote memory to `mobile/.../com/baekenough/secondbrain/.claude/agent-memory/` — INSIDE the source package. Caught and removed just before commit.
+
+A subagent with `memory: project` working in a SUBDIRECTORY can create `.claude/agent-memory/` relative to its current working directory, polluting the source tree (e.g., inside a source package). Memory MUST resolve to the PROJECT ROOT `.claude/`, never a nested working dir.
+
+| Anti-pattern | Required |
+|--------------|----------|
+| Accept a subagent's `.claude/agent-memory/` written under a source package | Verify memory writes land at project-root `.claude/`; remove any nested `.claude/agent-memory/` from source dirs before commit |
+
+Check for nested `.claude/agent-memory/` (e.g., `find . -path '*/src/*/.claude' -o -path '*/main/*/.claude'`) before committing subagent work.
+
 ## Best Practices
 
 - Consult memory before starting work

package/templates/.claude/skills/homework/SKILL.md

@@ -168,6 +168,17 @@ If `--dry-run` is active, skip this phase and output the draft directly to the c
 └── 다음 액션: {harness 제안 있으면 표시, 없으면 "없음"}
 ```
 
+## Cross-Project Import Compatibility
+
+When `homework` is imported into a project OTHER than oh-my-customcode, its automatic `omcustom-feedback` invocation (the retrospective-filing phase) depends on `omcustom-feedback` being model-invocable in that project.
+
+| Condition | Behavior |
+|-----------|----------|
+| `omcustom-feedback` model-invocable (default in oh-my-customcode) | Auto-invoke files the retrospective normally |
+| `omcustom-feedback` has model invocation DISABLED (`disable-model-invocation`) in the importing project | Auto-invoke FAILS — fall back to MANUAL submission: present the drafted findings to the user and ask them to file the feedback (or run `omcustom-feedback` manually) |
+
+> Origin: #1336 — in the second-brain project, `omcustom-feedback` was model-invocation-disabled, so homework Phase 5's automatic invoke failed; the retrospective was filed manually. Document this dependency so cross-project imports degrade gracefully instead of silently failing the auto-file step.
+
 ## Options Reference
 
 | Option | Default | Description |

package/templates/manifest.json

@@ -1,5 +1,5 @@
 {
-  "version": "0.179.0",
+  "version": "0.181.0",
   "lastUpdated": "2026-05-20T00:00:00.000Z",
   "omcustomMinClaudeCode": "2.1.121",
   "omcustomMinClaudeCodeReason": "Sensitive-path direct Write/Edit on .claude/** under bypassPermissions (R010 deprecation, #1101)",