feat(o1): implement reviewer feedback rebuttal and re-adjudication protocol (100% PASS)

docs(c1): purge all 29 legacy Kanban references from SKILL.md and documentation (100% PASS)
2026-08-06 09:21:28 +09:00 · 2026-08-06 08:03:15 +09:00
14 changed files with 1525 additions and 110 deletions
@@ -23,12 +23,13 @@
  - 총괄 매니저로부터 작업을 위임받습니다.
  - **작업 분석 및 계획**: 주어진 작업을 철저히 분석하고, 작은 단위로 문제를 나누어 세부 계획을 수립합니다.
  - **내부 병렬 처리**: 내부적으로 subagent를 활용해 위임받은 작업을 병렬적으로 처리할 수 있습니다.
-  - **리뷰 타당성 검증 및 거부**: 리뷰어가 지적한 피드백을 면밀히 검토합니다. 타당한 제안은 수렴하여 코드를 수정하지만, 타당하지 않다고 판단되는 안건은 반영하지 않고 **그 명확한 이유를 작성하여 리뷰어에게 되돌려 보냅니다**.
+  - **리뷰 타당성 검증 및 거부**: 리뷰어가 지적한 피드백을 면밀히 검토합니다. 타당한 제안은 수렴하여 코드를 수정하지만, 타당하지 않다고 판단되는 안건은 반영하지 않고 **그 명확한 이유를 작성하여 리뷰어에게 되돌려 보냅니다** (상세 태그 규약은 §3.1 '[REBUT:]' 참조).
  - **완료 신호 송신**: 모든 리뷰어들로부터 `PASS`를 획득하고 변경 사항이 검증되면, 최초 작업을 위임받았던 개발 팀장이 총괄 매니저에게 최종 작업 완료 신호를 송신합니다.
 - **Reviewer Team Leader (리뷰어 팀장)**:
  - 개발 팀장으로부터 리뷰 요청을 접수합니다.
  - **문제 제시에 대한 이유와 개선 방향 포함**: 단순한 반려(`NOT PASS`) 통보는 금지됩니다. 이슈를 제기할 때는 **반드시 해당 문제가 발생하는 구체적인 이유와 확실한 개선 방향(코드 대안 포함)을 함께 작성**해야 합니다.
  - **합의 루프**: 모든 지적 사항이 해결되고 최종 `PASS`를 발행할 때까지 리뷰 루프에 동참합니다.
  - **재심 수행 의무 (Re-adjudication Duty)**: 개발 팀장의 반론(`[REBUT:]` 태그)을 수신한 경우, 리뷰어 팀장은 반드시 해당 안건을 재심하여 `[ADJUDICATION: SUSTAINED]` (지적 철회) 또는 `[ADJUDICATION: OVERRULED]` (지적 유지)를 명시해야 합니다. 수신된 반론을 무시하거나 재심을 우회하는 것은 금지됩니다.
 ### 🛡️ 역할 범위 준수 원칙 (Role Suitability Check)
 - 모든 에이전트는 자신에게 부여된 역할에 부합하는 작업만을 수행해야 합니다. (예: 개발 팀장은 최종 PASS 여부를 결정하지 않으며, 리뷰어 팀장은 직접 프로젝트 소스코드를 작성하지 않습니다.)
@@ -106,6 +107,24 @@ sequenceDiagram
   - 리뷰어 전원이 `PASS`를 인가할 때까지 이 과정이 반복됩니다.
 4. **최종 보고**: 개발 팀장이 총괄 매니저에게 완료 신호를 보내면 총괄 매니저가 사용자에게 완료를 알립니다.
 ### 3.1 Rebuttal & Adjudication Protocol (이의제기 및 재심 프로토콜)
 개발 팀장이 리뷰어의 피드백이 타당하지 않거나 부적절하다고 판단하는 경우, 묵인하거나 수용하지 않고 정식 반론(Rebuttal)을 제기해야 합니다.
 | 태그 | 발행 주체 | 의미 |
 |---|---|---|
 | `[REBUT: <리뷰어_세션명>]` | 개발자 | 지정된 리뷰어의 피드백을 공식 거부. 리포트에 상세 근거 서술 필수. 타당한 지적은 정상 반영. |
 | `[ADJUDICATION: SUSTAINED]` | 리뷰어 | 개발자의 반론을 수용하여 이전 지적 사항을 철회함. |
 | `[ADJUDICATION: OVERRULED]` | Reviewer | 개발자의 반론을 기각하고 이전 지적 사항을 유지함. |
 | `[ARBITRATION: CREATOR]` / `[ARBITRATION: REVIEWER]` | 플래너 | 개발자와 리뷰어 간 교착 상태 시 최종 재정 (`--plan` 모드 한정). |
 **프로토콜 불변 규칙**:
 1. **반론은 PASS를 대체하지 않음**: 수용된 반론(`SUSTAINED`)은 해당 지적 사항만 철회시킵니다. 리뷰 루프가 완수되려면 여전히 모든 리뷰어로부터 만장일치 `[VERDICT: PASS]`를 획득해야 합니다.
 2. **Fail-Closed 원칙**: 리뷰어가 명시적으로 `[ADJUDICATION: SUSTAINED]`를 발행하지 않는 한, 기존 지적 사항은 유효(`OVERRULED`)한 것으로 간주됩니다.
 3. **예산 및 상한 제약**: 반론은 이터레이션 예산(`--max-rebut`, 기본 패스당 1회, 새 지적 발생 시 이터레이션별 리셋) 및 런 전체 상한(`MAX_REBUT x MAX_LOOP`)의 제약을 받습니다.
 4. **교착 상태 에스컬레이션**: 리뷰어가 `OVERRULED`를 고수하고 플래너 모드(`--plan`)가 활성화된 경우 플래너가 재정합니다. 플래너가 없으면 리뷰어의 지적이 우선 반영됩니다(`Fail-Closed`).
 5. **세션 매칭 및 중복 제거**: 반론 태그는 실제로 지적을 제기한 리뷰어 세션명을 지정해야 합니다. 단일 리포트 내 동일 리뷰어 대상 중복 태그는 단 1회의 재심 라운드로 병합됩니다.
 ---
 ## 4. 분석 인프라 패턴 & 실무 가이드 (Infra Patterns)
@@ -23,18 +23,19 @@ Newly spawned agents (e.g., `antigravity`, `claude`, `cline`, `hermes`) act as *
  - Receives tasks from the General Manager.
  - **Task Breakdown & Planning**: Thoroughly analyzes the task, breaks it down into small units, and creates a plan.
  - **Internal Parallelism**: Can run subagents in parallel internally to handle the delegated work.
-  - **Review Integrity & Refusal**: Thoroughly reviews feedback from Reviewers. Adopts/implements recommendations if valid. If any recommendation is judged invalid, the Developer Team Leader must **not** implement it, but instead return the refutation along with detailed reasons to the Reviewer.
+  - **Review Integrity & Refusal**: Thoroughly reviews feedback from Reviewers. Adopts/implements recommendations if valid. If any recommendation is judged invalid, the Developer Team Leader must **not** implement it, but instead return the refutation along with detailed reasons to the Reviewer (see §3.1 for the '[REBUT:]' protocol).
  - **Completion Signal**: Once all reviewers yield a `PASS` and changes are verified, the Developer Team Leader who first received the task sends a completion signal back to the General Manager.
 - **Reviewer Team Leader (리뷰어 팀장)**:
  - Receives review requests from the Developer Team Leader.
  - **Detailed Feedback with Directions**: Simply rejecting changes (`NOT PASS`) is forbidden. Reviewers **must** specify the exact reason for the issue and provide a concrete, stable, and verified alternative direction for improvement.
  - **Consensus Loop**: Engages in the review cycle until all objections are resolved and a final `PASS` is issued.
  - **Re-adjudication Duty**: Upon receiving a rebuttal with '[REBUT:]' tag, the Reviewer Team Leader must re-examine the objection and explicitly issue '[ADJUDICATION: SUSTAINED]' (withdraw objection) or '[ADJUDICATION: OVERRULED]' (maintain objection). Ignoring or bypassing a received rebuttal is forbidden.
 ### 🛡️ Role Suitability Check Principle (자신의 역할 범위 수행 원칙)
 - Every agent must only perform tasks suitable for its designated role (e.g., Developer Team Leaders do not issue final reviews, and Reviewer Team Leaders do not write project code).
 - **If an agent receives a task that does not fit its role**, it must either:
-  1. Recommend the optimal agent session to delegate the task to, or
+  1. Defer or re-delegate the task to a suitable subagent/session, OR
-  2. Perform the task directly if strictly necessary for project continuity.
+  2. Reject the task explicitly by explaining the role mismatch.
 ---
@@ -106,6 +107,24 @@ sequenceDiagram
   - This cycle repeats until all reviewers issue a `PASS`.
 4. **Completion and Report**: The Developer Team Leader sends the final completion signal to the General Manager, who notifies the user.
 ### 3.1 Rebuttal & Adjudication Protocol (이의제기 및 재심 프로토콜)
 When a Developer Team Leader judges that a Reviewer's feedback is invalid or inappropriate, it must file a formal rebuttal rather than silently accepting or ignoring it.
 | Tag | Issuer | Meaning |
 |---|---|---|
 | `[REBUT: <reviewer_session>]` | Developer | Formally rejects feedback from specified reviewer. Must provide detailed reasoning in report body. Valid suggestions must still be implemented. |
 | `[ADJUDICATION: SUSTAINED]` | Reviewer | Accepts developer's refutation and withdraws previous objection. |
 | `[ADJUDICATION: OVERRULED]` | Reviewer | Rejects developer's refutation and maintains objection. |
 | `[ARBITRATION: CREATOR]` / `[ARBITRATION: REVIEWER]` | Planner | Final ruling in case of deadlock between Developer and Reviewer (`--plan` mode). |
 **Protocol Rules**:
 1. **Rebuttal does NOT substitute for a PASS**: A sustained rebuttal only withdraws the specified objection. The review loop will still require a unanimous `[VERDICT: PASS]` from all active reviewers to complete successfully.
 2. **Fail-Closed Principle**: Unless a Reviewer explicitly issues `[ADJUDICATION: SUSTAINED]`, the objection stands by default.
 3. **Budget & Bound Constraints**: Rebuttals are bounded by an iteration budget (`--max-rebut`, default 1 per iteration, resets each pass for new findings) and a total run budget (`MAX_REBUT x MAX_LOOP`).
 4. **Deadlock Escalation**: If a Reviewer issues `[ADJUDICATION: OVERRULED]` and Planner is enabled (`--plan`), Planner arbitrates. If Planner is absent, the Reviewer's objection prevails (`Fail-Closed`).
 5. **Session Matching & Deduplication**: Rebuttals must target valid objecting reviewer session names. Duplicate tags targeting the same reviewer in a single report are merged into a single re-adjudication round.
 ---
 ## 4. Analysis Infrastructure Patterns & Practical Guide (Infra Patterns)
@@ -0,0 +1,749 @@
 # C-1 구현 계획서 **Rev.2** — Kanban 잔재 문서 전면 제거
 - **Job**: `17ffc912` / Role: **Planner** (Rev.1 = Job `75847d1d`)
 - **대상**: IMPROVEMENTS.md **C-1** — "Kanban 문서 29회 언급 vs 실제 구현 0건"
 - **기준 커밋**: `924e77e` (clean tree)
 - **반영**: Creator `agy` Challenge Report (Job `a5f24e6f`) — 맹점 2건
 - **검증 상태**: Rev.1 전량 + Rev.2 신규 게이트를 `git archive HEAD` 클론에 **실제 적용·측정 완료**. 수치는 전부 실측값이다.
 ---
 ## 0. Rev.1 → Rev.2 변경 요지
 | 이의 | 판정 | 근거 | 계획 반영 |
 |---|---|---|---|
 | **맹점 1** — `status` 의 `prereq_skills` 에 `monitor` 가 있는 것은 논리 모순이므로 제거 | **기각 (REJECTED)** | `status.sh:12` 가 `multi-agent-mux-monitor/scripts/reconcile.sh` 를 **하드 참조**한다. 의존은 실재한다. 또한 `prereq_skills` 의 저장소 관례는 "선행 *실행*"이 아니라 "선행 *가용성*"이며, 이의제기의 전제는 파일 자신의 용례 2건에 의해 반증된다. | 변경 없음. 대신 **RK-G** 로 위험 기록 + 별도 개선 항목 제안(§9) |
 | **맹점 2** — G-C 정규식이 `--goal-max-turns` 를 놓친다 | **인용 (ACCEPTED, 확대 적용)** | 사실이다. 다만 실측 결과 구멍은 **2개가 아니라 10개**다. 제안된 `--goal` 추가는 10개 중 2개만 막는다. | **G-C 전면 재설계** — 열거를 **파생(derivation)** 으로 교체. red/green **61건 → 0건** 실측 |
 | (신규) | **추가** | 맹점 1이 제기한 "디스패처가 prereq 메타데이터로 오작동" 우려의 *검증 가능한* 형태 | **G-G** — `prereq_skills` 그래프 순환 부재 게이트 |
 **총평**: 맹점 2의 진단은 정확했고 처방은 부족했다. 맹점 1은 진단 자체가 사실과 어긋난다. 상세는 §7.
 ---
 ## 1. 요약
 C-1 은 "단어 29개를 지운다"가 아니라 **세 종류의 서로 다른 문제**다. 셋을 같은 방법으로 처리하면 실패한다.
 | 분류 | 건수 | 성격 | 처리 |
 |---|---|---|---|
 | **N. 순수 노이즈** | 12 | 지워도 의미 손실 0 (수식어, 태그, 괄호) | 단어 삭제 |
 | **D. 끊어진 스킬 참조** | 3 | 이 저장소에 **존재하지 않는 스킬**을 가리킴 (`kanban-worker`, `kanban-orchestrator`) | 실존 스킬로 교체 |
 | **S. 구조적 의존** | 16 | 문서의 **실행 메커니즘 자체**가 Kanban (`hermes kanban create`, `kanban_comment/_block/_heartbeat`, `goal_mode`) | 섹션 재작성 |
 **핵심 판단**: S 분류는 단어를 지우면 문서가 "무엇을 실행하라는 것인지" 알 수 없게 된다. 대체 메커니즘이 반드시 필요하고, **이미 구현되어 있다** — `reconcile.sh --subscribe` 가 MQTT push + 브로커 다운 시 폴링 폴백을 갖춘 자립형 상주 루프다. 즉 Kanban 은 처음부터 불필요한 껍데기였고, 제거는 문서를 **실제 구현에 일치시키는 작업**이다.
 **부가 성과**: 문서 정합성 결함 2건 동시 해소(§6). 하나는 존재하지 않는 상태 파일 주장, 하나는 `--subscribe` 폴백 stdout 동작 오기.
 ---
 ## 2. 현황 측정 (근거)
 ```
 $ grep -ril kanban . | grep -v '^./.git/'   →  파일 6개
   22  .agents/skills/multi-agent-mux-monitor/SKILL.md
    5  .agents/skills/multi-agent-mux-status/SKILL.md
    2  .agents/skills/multi-agent-mux-create/SKILL.md
    1  README.md
    1  README.ko.md
    2  IMPROVEMENTS.md          ← C-1 추적 항목 자신
 ```
 - SKILL.md 합계 = **22 + 5 + 2 = 29**. C-1 의 "29회" 와 **정확히 일치** → 추적 항목의 전제는 유효하다.
 - **코드 파일 참조 0건**: `grep -ril kanban` 에서 `.md` 를 빼면 아무것도 남지 않는다. `hermes kanban` 서브커맨드를 쓰는 코드도 0건.
 - `reconcile.sh` 내 `kanban` 관련 코드 = **0건**. 문서가 서술한 `kanban_comment` 호출은 애초에 어디에도 없다.
 → **순수 문서 작업이며 런타임 동작 변경 위험은 0이다.**
 > **주의 — `hermes` 자체는 실존 도구다.** `create_session.sh:102` 가 `hermes status` 를 호출하고, `create/SKILL.md:234` 는 `hermes profile alias` 사용 금지를 정당하게 경고한다. 제거 대상은 **`hermes kanban` 서브커맨드 표면**이지 `hermes` 전체가 아니다. Rev.2 검토 중 "스킬 문서에 `hermes ` CLI 호출 금지" 게이트를 후보로 올렸다가 이 실측으로 **폐기**했다 — 정당한 2건을 오탐했을 것이다.
 ---
 ## 3. 범위 확정 — 리뷰어의 "0건"은 어디까지인가
 성공 조건은 "reviewers verifying 0 remaining references" 다. 리뷰어가 `grep -ri kanban .` 를 그냥 실행하면 **IMPROVEMENTS.md 가 걸린다**. 그런데 C-1 항목은 *제거 사실을 기록하는 추적 문서*이므로 이름을 언급하지 않고는 쓸 수 없다. 방치하면 리뷰가 무한히 실패한다.
 **해소 규칙**:
 - **G-A (제품 표면)** = `.agents/skills/` + `README.md` + `README.ko.md` → **반드시 0건**. 실질 합격 기준.
 - **G-B (저장소 전체 − 추적/보고 문서)** = `IMPROVEMENTS.md` 와 `.mam/`(잡 브리프·리포트가 구 문안을 인용) 제외 → **0건**.
 - IMPROVEMENTS.md 의 C-1 은 삭제가 아니라 **§5 완료된 과제로 이관**(파일 기존 관례). 잔존 2회는 "완료 기록"으로 의도된 것이며 G-A/G-B 어디에도 해당하지 않는다.
 **리뷰어가 스스로 grep 범위를 즉흥 결정하지 않게 하는 것이 이 절의 목적이다.**
 ---
 ## 4. 변경 명세
 ### F1 — `.agents/skills/multi-agent-mux-monitor/SKILL.md` (22건, 전면 재작성)
 이 파일만 **파일 전체 교체**를 지시한다. 22개 참조가 10개 섹션에 흩어져 있고 §Dispatching / §Worker body template 두 섹션은 통째로 Kanban 스펙이라, 부분 치환 지시로는 Creator 마다 결과가 갈린다. 최종 전문은 **부록 A**(검증 완료본, 218줄).
 변경 인벤토리 — 22건 전량의 처분:
 | 위치 | 원문 | 처분 | 분류 |
 |---|---|---|---|
 | L3 `description` | "long-lived **Kanban worker**", "dispatched as a **Kanban goal_mode** task (--goal)" | `reconcile.sh --subscribe` 상주 루프로 재서술 | S |
 | L8 `environments` | `[kanban, terminal, herdr]` | `[terminal, herdr]` — 타 스킬 6종 전부와 동일해짐 | N |
 | L11 `tags` | `..., monitor, kanban, observation, ...` | `kanban` 제거 | N |
 | L12 `related_skills` | `..., kanban-orchestrator` | `..., multi-agent-mux-status` | **D** |
 | L13 `prereq_skills` | `[kanban-worker, multi-agent-mux-create]` | `[multi-agent-mux-create]` | **D** |
 | L16 H1 | "— Live Reconciliation **via Kanban Worker**" | "— Live Reconciliation" | N |
 | L23 | "Dispatch a **Kanban worker** (in `goal_mode`) that:" | "Run a **long-lived reconciler** (`reconcile.sh --subscribe`) that:" | S |
 | L25 | "Every ~30s polls" | 실제 동작으로 정정: MQTT 이벤트 반응 + 브로커 다운 시 `RECONCILE_POLL_INTERVAL`(기본 15s) 폴백 | S |
 | L37 | 4. "Writes a Kanban `kanban_comment` on every drift event" | 4. `--emit-diff` 시 stdout JSON 방출 (+폴백 경로 단서, §6-b) | S |
 | L39 | 6. "**Goal loop**: judge … worker blocks with `reason=stop-requested`" | 5. 실제 종료 조건: `--timeout` / `--idle-timeout` / 인터럽트 | S |
 | L52 | "You don't have a Kanban dispatcher running" | "You only need a point-in-time answer — use `multi-agent-mux-status`" | S |
 | L54-85 | §Dispatching the monitor — `hermes kanban create` 30줄 블록 | §**Running the monitor** — 실제 실행 2줄로 교체. body 내 herdr 네이티브 명령 경고문은 **보존** | S ×4 |
 | L93 | "JSON diff … that the worker turns into a `kanban_comment`" | "for the caller to consume" | S |
 | L110 | 플래그 문단 말미 "…turns the emitted JSON `drifts[]` into `kanban_comment` calls itself." | "…the emitted JSON `drifts[]` is the caller's to consume." — `--comment-card` 언급도 제거 | S |
 | L167 | 함정 "**Don't run the monitor without `--goal`**" | 실존 함정으로 교체: `--once` 는 1회 후 종료 / `--idle-timeout` 기본 3600s 함정 | S |
 | L169 | 함정 "**`kanban_comment` rate limits** … `.cache/…/<ws>.state` 를 추적한다" | 중복 병합 책임을 caller 로 명시. **거짓 주장 제거** → §6-a | S |
 | L171 | "comment about it but don't delete" | "report it but don't delete" | N |
 | L174-209 | §**Worker body template** — `kanban_comment`/`_heartbeat`/`_block`/`$HERMES_KANBAN_TASK` | §**Supervising-agent runbook** 으로 전환. **Hard rules 4개 전량 보존** | S ×4 |
 | L119-163 | drift A~D 예시의 "comment:" | "report:" | N ×4 |
 | L236-237 | §Related skills — `kanban-worker`, `kanban-orchestrator` | `multi-agent-mux-status`, `multi-agent-mux-delegate-job` | **D** |
 **의도적으로 남기는 것**: `metadata.hermes` 블록 자체(Kanban 아님, 공통 스키마), `hermes` 태그 체계, §Security `--subscribe` 전문, §Status Enum, drift A~D 본문.
 ### F2 — `.agents/skills/multi-agent-mux-status/SKILL.md` (5건, 6줄 치환)
 각 치환은 **원문에 정확히 1회만 등장**함을 스크립트로 강제 확인한다(0회/2회면 즉시 중단).
 | 줄 | before → after |
 |---|---|
 | L3 | `…and any drift. No Kanban, no mutation. Reuses…` → `…and any drift. No mutation. Reuses…` |
 | L3 | `without spinning up a Kanban monitor worker.` → `without spinning up the monitor loop.` |
 | L24 | `**No mutation. No Kanban. No polling loop.**` → `**No mutation. No polling loop.**` |
 | L108 | `dispatch \`multi-agent-mux-monitor\` (Kanban worker) or run` → `run \`multi-agent-mux-monitor\` or` |
 | L118 | `- Quick sanity check before dispatching a long Kanban task` → `…before dispatching a long-running job` |
 | L122 | `- Continuous live tracking → \`multi-agent-mux-monitor\` (Kanban worker)` → `…→ \`multi-agent-mux-monitor\`` |
 > **frontmatter 는 손대지 않는다.** L12/L13 의 `related_skills` / `prereq_skills` 는 **Rev.1 과 동일하게 원형 유지**한다. 근거는 §7-A.
 ### F3 — `.agents/skills/multi-agent-mux-create/SKILL.md` (2건)
 | 줄 | before → after |
 |---|---|
 | L51 | `If any check fails → \`kanban_block(reason="...")\` (worker path) or report to user (interactive path).` → `If any check fails → abort with a non-zero exit and report the reason (automated path) or report to user (interactive path).` |
 | L65 | `(e.g., cmux, Kanban workers, manual herdr sessions)` → `(e.g., cmux, background workers, manual herdr sessions)` |
 L51 은 단순 삭제하면 "실패 시 무엇을 하라"가 사라지므로 **동등한 비-Kanban 지시로 치환**해야 한다.
 ### F4 — `README.md` (1건) / `README.ko.md` (1건)
 | 파일 | before → after |
 |---|---|
 | README.md L45 | `A long-running Kanban reconcile worker that dynamically monitors` → `A long-running reconcile worker that dynamically monitors` |
 | README.ko.md L27 | `백그라운드에서 Kanban Reconcile 프로세스로 실행되어` → `백그라운드 Reconcile 프로세스로 실행되어` |
 브리프 문구는 "SKILL.md files" 지만 §3 의 이유로 **README 2건 필수 포함**.
 ### F5 — `IMPROVEMENTS.md` (추적 항목 이관)
 1. §4 에서 C-1 블록(L81-82) 삭제
 2. §4 헤더 `(Legacy Remnants — 5건)` → `(Legacy Remnants — 4건)`
 3. §5 헤더 `(Completed Tasks — 3건)` → `(Completed Tasks — 4건)`
 4. §5 에 파일 기존 형식으로 추가:
 ```markdown
 ### **C-1: Kanban 문서 29회 언급 vs 실제 구현 0건** — ✅ 완료
 - SKILL.md 3종(monitor 22 / status 5 / create 2)과 README 2종의 Kanban 서술을 전면 제거했습니다.
 - `multi-agent-mux-monitor` 의 실행 메커니즘 서술을 실제 구현인 `reconcile.sh --subscribe`
  (MQTT push + 브로커 다운 시 폴링 폴백) 기준으로 재작성했습니다.
 - 존재하지 않는 스킬 참조 2건(`kanban-worker`, `kanban-orchestrator`)을 실존 스킬로 교체했습니다.
 - `hermes kanban create` CLI 플래그 잔재 10종(`--goal-max-turns`, `--assignee`, `--comment-card` 등)을
  파생형 검증 게이트(G-C)로 차단했습니다.
 - 제품 표면(`.agents/skills/`, `README*.md`) Kanban 참조 **0건** 확인.
 ```
 ### F6 — 반영하지 **않는** 것
 - `.mam/` 하위 잡 브리프·리포트: 과거 잡의 기록이며 수정 시 감사 추적 훼손. G-B 가 제외하는 이유.
 - `status/SKILL.md` frontmatter `prereq_skills`: §7-A.
 - `.cache` **코드**: C-2 소관(§6-a).
 ---
 ## 5. 적용 순서
 F1 → F2 → F3 → F4 → **G 게이트 전량 실행** → F5(IMPROVEMENTS 이관) → 재실행.
 F5 를 마지막에 두는 이유: 완료 기록은 **게이트 통과 후에** 쓰여야 한다. 먼저 쓰면 미완료를 완료로 기록한 순간이 커밋 이력에 남는다.
 ---
 ## 6. 검증 게이트
 전부 클론에 적용해 **실측 완료**. 괄호 안이 측정값이다.
 **G-A — 제품 표면 0건 (실질 합격 기준)** *(측정: 0 — CLEAN)*
 ```bash
 grep -ric kanban .agents/skills README.md README.ko.md | grep -v ':0$' || echo "0 references — CLEAN"
 ```
 **G-B — 저장소 전체 0건 (추적/보고 문서 제외)** *(측정: 0 — CLEAN)*
 ```bash
 grep -ril kanban . | grep -v '^\./\.git/' | grep -v IMPROVEMENTS.md | grep -v '^\./\.mam/' \
  || echo "0 files — CLEAN"
 ```
 ### G-C — Kanban CLI/API 표면 잔재 0건 ⚠️ **Rev.2 에서 전면 재설계**
 *(측정: HEAD **61건 FAIL** → 적용 후 **0건 PASS**)*
 **Rev.1 의 결함**: 패턴을 **손으로 열거**했다(`KANBAN|kanban_|goal_mode|kanban-worker|kanban-orchestrator|HERMES_KANBAN`). 손 열거는 반드시 빠뜨린다. `agy` 가 `--goal-max-turns` 로 이를 정확히 지적했다.
 **그러나 구멍은 2개가 아니라 10개였다.** F1 이 삭제하는 텍스트에서 토큰을 기계적으로 추출해 각 패턴의 포착 여부를 실측한 결과:
 | 삭제되는 토큰 | Rev.1 패턴 | `agy` 제안 패턴 | Rev.2 파생 게이트 |
 |---|---|---|---|
 | `kanban_comment` / `kanban_block` / `kanban_heartbeat` | CATCH | CATCH | CATCH |
 | `HERMES_KANBAN_TASK` / `goal_mode` | CATCH | CATCH | CATCH |
 | `--goal` / `--goal-max-turns` | **MISS** | CATCH | CATCH |
 | `--assignee` | **MISS** | **MISS** | CATCH |
 | `--comment-card` | **MISS** | **MISS** | CATCH |
 | `--max-runtime` | **MISS** | **MISS** | CATCH |
 | `--max-retries` | **MISS** | **MISS** | CATCH |
 | `--title` / `--branch` / `--body` / `--skill` | **MISS** | **MISS** | CATCH |
 → `agy` 의 처방은 **10개 중 2개**만 막는다. 진단은 옳았으나 처방이 부족하다.
 **Rev.2 의 해법**: 열거를 **파생**으로 바꾼다. 토큰 목록을 기억이 아니라 *F1 이 실제로 삭제하는 텍스트*에서 뽑는다 — 구성상 완전하다.
 ```bash
 python3 .mam/gates/gate_gc.py     # 전문은 부록 C. 실패 시 exit 1
 ```
 ```
 HEAD:    G-C FAIL — 61 Kanban-surface hit(s)
 적용 후:  G-C PASS — 0 Kanban CLI/API surface tokens
 ```
 **오탐 위험 실측**: 후보 플래그 10종이 `monitor/SKILL.md` **밖** 제품 표면에 등장하는 횟수 = 전부 **0**. 오탐 없음.
 **의도적 제외 1건**: `AGENT_SESSIONS_STATE_DIR` 도 F1 이 삭제하는 토큰이지만 게이트에 넣지 **않는다**. `reconcile.sh:23` 에 실존하는 환경변수이며, 삭제 사유는 Kanban 이 아니라 거짓 서술(§6-a)이다. 훗날 올바르게 재문서화될 수 있으므로 금지어로 만들면 안 된다.
 **G-D — 전 SKILL.md frontmatter YAML 파싱 유지** *(측정: 7/7 OK)*
 ```bash
 python3 - <<'PY'
 import glob, yaml
 for f in sorted(glob.glob('.agents/skills/*/SKILL.md')):
    t = open(f, encoding='utf-8').read()
    if not t.startswith('---'): continue
    d = yaml.safe_load(t.split('---',2)[1])
    print(f"  OK {d['name']:36s} env={d.get('environments')}")
 PY
 ```
 **G-E — 끊어진 스킬 참조 0건 (red→green)** *(HEAD 2건 → 적용 후 0건)*
 단어를 지운 게 아니라 **실존하는 것으로 바꿨음**을 증명한다.
 ```bash
 python3 - <<'PY'
 import glob, os, yaml
 local = {os.path.basename(os.path.dirname(f)) for f in glob.glob('.agents/skills/*/SKILL.md')}
 EXTERNAL_OK = {'claude-code'}
 bad = []
 for f in sorted(glob.glob('.agents/skills/*/SKILL.md')):
    d = yaml.safe_load(open(f, encoding='utf-8').read().split('---',2)[1]) or {}
    h = (d.get('metadata') or {}).get('hermes') or {}
    for key in ('related_skills','prereq_skills'):
        for s in h.get(key) or []:
            if s not in local and s not in EXTERNAL_OK:
                bad.append((os.path.basename(os.path.dirname(f)), key, s))
 print("DANGLING:", bad or "none — CLEAN")
 PY
 ```
 ```
 HEAD:    [('multi-agent-mux-monitor','related_skills','kanban-orchestrator'),
          ('multi-agent-mux-monitor','prereq_skills','kanban-worker')]
 적용 후:  none — CLEAN
 ```
 ### G-G — `prereq_skills` 그래프 순환 부재 🆕 **Rev.2 신규**
 *(측정: HEAD **DAG** → 적용 후 **DAG**)*
 맹점 1이 제기한 "자율 디스패처가 prereq 메타데이터로 오작동한다"는 우려의 **검증 가능한 형태**다. 디스패처를 실제로 무한루프에 빠뜨리는 것은 *논쟁의 여지가 있는 의미론*이 아니라 *순환*이다. F1 이 `monitor.related_skills` 에 `status` 를 넣으므로(`status.related_skills` 에는 이미 `monitor` 존재) 상호 참조가 생기는데, **`related_skills` 의 상호 참조는 무해하고 `create`↔`resume` 에 이미 존재한다**. 위험한 것은 `prereq_skills` 순환뿐이며, 이 게이트가 그것만 정확히 겨냥한다.
 ```bash
 python3 - <<'PY'
 import glob, yaml
 g = {}
 for f in glob.glob('.agents/skills/*/SKILL.md'):
    d = yaml.safe_load(open(f, encoding='utf-8').read().split('---',2)[1]) or {}
    g[d['name']] = list(((d.get('metadata') or {}).get('hermes') or {}).get('prereq_skills') or [])
 def cyc(n, seen):
    if n in seen: return seen[seen.index(n):] + [n]
    for m in g.get(n, []):
        if m in g:
            r = cyc(m, seen + [n])
            if r: return r
    return None
 found = [c for n in g if (c := cyc(n, []))]
 print("CYCLES:", found or "none — DAG")
 PY
 ```
 적용 후 실측 그래프:
 ```
 create        -> ['claude-code']            monitor -> ['multi-agent-mux-create']
 loop          -> ['multi-agent-mux-create'] resume  -> ['multi-agent-mux-create']
 status        -> ['multi-agent-mux-create', 'multi-agent-mux-monitor']
 stop          -> ['multi-agent-mux-create', 'multi-agent-mux-resume']
 delegate-job  -> []
 CYCLES: none — DAG
 ```
 **G-F — 회귀 테스트: 불필요 (명시적 판단)**
 `tests/` 전량에 SKILL.md·README 를 읽는 코드가 **0건**(`grep -rln "SKILL.md\|README" tests/` → 없음), CI 문서 린트 없음(`.github/workflows/` 부재), 변경 파일에 스크립트 **0개**. 8분짜리 전체 스위트는 이 변경에 대해 **아무것도 증명하지 못한다**. 형식적 완결성을 위해 돌리는 것을 막지 않되 **계획은 요구하지 않는다.** 실제 위험은 G-A~G-G 가 덮는다.
 **측정된 변경 규모** (`diff -ru base clone`, 총 328줄):
 ```
  -74  +53   .agents/skills/multi-agent-mux-monitor/SKILL.md   (237줄 → 218줄)
  -5   +5    .agents/skills/multi-agent-mux-status/SKILL.md
  -2   +2    .agents/skills/multi-agent-mux-create/SKILL.md
  -1   +1    README.md
  -1   +1    README.ko.md
 ```
 ---
 ## 7. 이의제기 심층 판정
 ### 7-A. 맹점 1 — `status.prereq_skills` 에서 `monitor` 제거 → **기각**
 **주장**: status 는 단발성 read-only 스냅샷이므로 monitor 를 선행 필수로 두는 것은 논리적 모순이다.
 **반증 1 — 의존은 실재한다.**
 ```
 .agents/skills/multi-agent-mux-status/scripts/status.sh:12
 RECONCILE="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)/multi-agent-mux-monitor/scripts/reconcile.sh"
 ```
 `status.sh` 는 monitor 스킬 디렉터리의 스크립트를 **경로로 하드 참조**한다. `status/SKILL.md:46` 도 *"Calls `reconcile.sh --once --emit-diff --dry-run`"* 이라 명시한다. monitor 스킬을 제거하면 **status 는 즉시 깨진다.** 이는 저장소 전체에서 `prereq_skills` 항목이 파일시스템 수준에서 문자 그대로 참인 **유일한 사례**다. 제거하면 메타데이터가 *덜* 정확해진다.
 **반증 2 — "선행 실행" 해석은 파일 자신의 용례가 부정한다.** 실측된 6개 전량:
 ```
 create  -> ['claude-code']                            ← 스킬로 "실행"하는 대상이 아님. 설치 여부다.
 stop    -> ['multi-agent-mux-create','...-resume']    ← stop 전에 resume 을 실행할 필요는 없다.
 ```
 두 건이 각각 독립적으로 "선행 *실행*" 해석을 반증한다. 저장소 관례는 **가용성(availability)** 이다. 그 해석에서 `status -> monitor` 는 모순이 아니라 **정확한 서술**이다.
 **반증 3 — 소비자가 없다.** `prereq_skills` / `related_skills` / `environments` 를 읽는 코드는 저장소 전체에 **0건**(`.py`/`.sh` 전수 검색). 주장된 "디스패처 오작동"은 이 저장소 안에서는 재현 불가능한 가정이다. 외부 디스패처를 가정하더라도 반증 1·2 가 우선한다.
 **그럼에도 남는 실재 문제**: `prereq_skills` 의 의미론이 **어디에도 정의되어 있지 않다**. `agy` 가 그럴듯하게 오독했다는 사실 자체가 모호성의 증거다. 다만 올바른 해소는 *참인 의존을 지우는 것*이 아니라 *의미론을 문서화하는 것*이며, 이는 C-1(Kanban 제거)과 무관한 별개 과제다 → §9 로 분리 제안.
 ### 7-B. 맹점 2 — G-C 정규식 보강 → **인용하되 확대**
 지적은 정확하다. Rev.1 G-C 는 `--goal-max-turns` 를 놓쳤다. 다만 §6 G-C 표가 보이듯 **놓친 것은 10개**이고 제안된 `--goal` 추가는 그중 2개만 막는다. 나머지 8개(`--assignee`, `--comment-card`, `--max-runtime`, `--max-retries`, `--title`, `--branch`, `--body`, `--skill`)는 `agy` 의 패턴으로도 통과한다.
 근본 원인은 특정 단어 누락이 아니라 **"기억으로 열거한 목록"이라는 방법론**이다. 그래서 Rev.2 는 단어를 더 넣는 대신 목록을 **삭제 대상 텍스트에서 파생**시킨다. 이 방법에서는 정의상 누락이 발생하지 않는다.
 > 이 지적이 없었다면 Rev.1 은 8개의 추가 구멍을 가진 채 통과했을 것이다. **부분 편집(전체 교체가 아닌)으로 F1 을 수행한 Creator에게는 실제로 발생할 수 있는 실패**였다.
 ---
 ## 8. 함께 해소되는 문서 정합성 결함 2건
 **(a) 존재하지 않는 상태 파일 — C-2 와 교차**
 원문 L169 는 *"The script tracks a state file at `.cache/multi-agent-mux-monitor/<workspace>.state`"* 라 단언한다. 실측:
 ```
 reconcile.sh:23   STATE_DIR="${AGENT_SESSIONS_STATE_DIR:-…/.cache/multi-agent-mux-monitor}"
 reconcile.sh:313  mkdir -p "$STATE_DIR"
              → 읽기 0회, 쓰기 0회
 ```
 `mkdir` 만 하고 아무것도 저장하지 않는다(IMPROVEMENTS **C-2** 가 코드 측면에서 동일 사실 추적 중). 대체 문장은 **중복 병합 책임을 caller 에게 명시**하고 상태 파일 주장은 싣지 않는다.
 > 이 계획은 `.cache` **코드**를 건드리지 않는다. 그건 C-2 소관이다. 문서가 거짓을 말하지 않게만 한다.
 **(b) `--subscribe` 폴백 경로의 stdout 동작**
 Rev.1 초안에서 4번 항목을 "모든 drift 를 stdout JSON 으로 방출"이라 썼다가 스크립트 확인 중 정정했다:
 ```
 reconcile.sh:303   bash "$_self" --once --emit-diff >/dev/null 2>&1 || true
 ```
 브로커 다운 폴백 루프는 **stdout 을 버린다**. `--subscribe` 상태에서 drift JSON 을 소비할 수 없다. 최종 문안은 이 단서를 명시한다 — 원문에 없던 정보이며, 이를 믿고 파이프를 붙이려는 사용자를 구한다.
 ---
 ## 9. 후속 제안 (C-1 범위 밖, 별도 항목)
 **제안: `prereq_skills` / `related_skills` 의미론 정의** — §7-A 가 드러낸 실재 모호성이다.
 - 현상: 두 필드의 의미가 어디에도 정의되지 않아 "선행 실행"과 "선행 가용성"으로 갈리게 읽힌다. 실제로 Creator 가 후자를 전자로 오독해 참인 의존을 삭제하자고 제안했다.
 - 소비 코드 0건이므로 **런타임 영향은 없으나**, 외부 디스패처 연동 시 실제 오작동 원인이 된다.
 - 제안: `MULTI_AGENT_RULES.md` 또는 스킬 작성 가이드에 1문단으로 정의하고, G-E·G-G 를 상시 린트로 승격.
 - **C-1 에 병합하지 않는 이유**: Kanban 잔재 제거와 무관하며, 병합하면 "0건 검증"이라는 단일 합격 기준이 흐려진다.
 ---
 ## 10. 리스크
 | ID | 리스크 | 평가 / 완화 |
 |---|---|---|
 | **RK-A** | 런타임 회귀 | **없음.** 변경 5파일 전부 `.md`. 코드 참조 0건 실측. |
 | **RK-B** | `environments: [kanban,…]` 제거가 외부 스킬 레지스트리 로딩을 깬다 | 저장소 내 `environments` 소비 코드 0건. 나머지 6개 스킬이 이미 `[terminal, herdr]` → **다수 관례에 합류**. |
 | **RK-C** | §Dispatching 삭제로 "어떻게 띄우나"를 잃음 | 삭제가 아니라 **교체**. `hermes` 미설치 상태에서 원문 명령은 애초에 실행 불가였고, `--subscribe` 가 실제 동작하는 유일한 방법이다. |
 | **RK-D** | 리뷰어가 IMPROVEMENTS.md 잔존 2건으로 FAIL 판정 | §3 에서 범위 사전 정의 + G-A/G-B 명령 제공. |
 | **RK-E** | Worker body template 삭제로 Hard rules 상실 | §Supervising-agent runbook 으로 전환, **Hard rules 4개 전량 보존**. |
 | **RK-F** | 향후 Kanban 문구 복귀 | G-C(파생형)를 리뷰 체크리스트 상시 항목으로. 자동 린트 승격은 §9. |
 | **RK-G** 🆕 | 후속 리뷰어가 맹점 1을 재차 제기해 `status.prereq_skills` 를 임의 수정 | §7-A 를 계획에 명시적으로 남겨 판정 근거(`status.sh:12`)를 보존. **삭제 시 G-E 는 통과하지만 문서가 거짓이 된다** — 게이트로는 잡히지 않는 유형이므로 서술로 방어한다. |
 | **RK-H** 🆕 | G-C 의 `--title`/`--body`/`--branch`/`--skill` 이 훗날 정당하게 등장해 오탐 | 현재 제품 표면 등장 0건 실측. 오탐 시 **해당 토큰만** 근거와 함께 제외하고, 파생 방식 자체는 유지할 것. |
 ---
 ## 11. 역할 경계
 본 문서는 **설계 산출물**이며 저장소 코드·문서는 **한 줄도 수정하지 않았다**. 모든 프로토타이핑은 세션 스크래치패드 내 `git archive HEAD` 클론에서만 수행했고, 작업 트리는 `924e77e` 기준 clean 을 유지한다.
 `MULTI_AGENT_RULES.md` §1 에 따라 **구현은 Creator, 커밋은 GM 소관**이다. **차단 항목 없음.**
 ---
 ## 부록 A — `multi-agent-mux-monitor/SKILL.md` 최종 전문 (F1 적용본, 218줄)
 ````markdown
 ---
 name: multi-agent-mux-monitor
 description: "Run a long-lived reconciler that watches .mam/agent-sessions.yaml against the actual herdr/agent runtime state and reconciles them. Use when you want live visibility into which agent sessions are running, which are dead, which have stale YAML entries, and which have new session ids that haven't been recorded yet. Runs as a persistent loop (`reconcile.sh --subscribe`) that keeps going until it times out, idles out, or is interrupted."
 version: 1.0.0
 author: godopu
 license: MIT
 platforms: [linux, macos]
 environments: [terminal, herdr]
 metadata:
  hermes:
    tags: [agent, herdr, claude, antigravity, agy, monitor, observation, reconciliation]
    related_skills: [multi-agent-mux-create, multi-agent-mux-resume, multi-agent-mux-stop, multi-agent-mux-status]
    prereq_skills: [multi-agent-mux-create]
 ---
 # Agent Sessions Monitor — Live Reconciliation
 > **Companion skills**: `multi-agent-mux-create` / `multi-agent-mux-resume` / `multi-agent-mux-stop` (mutators); this skill is the **observer**.
 > **Single source of truth**: `./.mam/agent-sessions.yaml`.
 ## What this skill does
 Run a **long-lived reconciler** (`reconcile.sh --subscribe`) that:
 1. Reacts to delegated-job events on the MQTT broker, and — whenever the broker is
   unreachable — falls back to polling every `RECONCILE_POLL_INTERVAL` (default 15s)
   the actual state of:
   - `herdr agent list` (which sessions are alive)
   - `herdr agent get <session>` (pane cmd, cwd)
   - `~/.claude/projects/<workspace-key>/*.jsonl` mtime + first-line sessionId
   - `~/.gemini/antigravity-cli/cache/last_conversations.json` (agy workspace → conversation mapping)
   - `~/.gemini/antigravity-cli/conversations/<uuid>.db` mtime (agy)
 2. Compares the live state to `agent-sessions.yaml`
 3. Detects 4 classes of drift:
   - **yaml-only terminated/archived/stopped**: herdr dead, YAML says `terminated`, `archived`, or `stopped` → OK, left untouched (deliberate end states)
   - **yaml-only running, herdr dead**: YAML says `running`, herdr is gone → mark `terminated` with timestamp
   - **herdr-only running, not in YAML**: herdr session exists with `<workspace>-creator-*` naming but YAML doesn't know about it → register as a new entry
   - **stale UUID**: YAML has a UUID, but the on-disk artifact is gone → report it
 4. Emits a JSON drift record on stdout for every drift event when run with `--emit-diff`
   (note: the `--subscribe` broker-down fallback runs each pass for its YAML side-effects
   and discards the JSON — capture drift output with an explicit `--once --emit-diff`)
 5. Keeps running until one of its exit conditions fires: `--timeout` (wall-clock),
   `--idle-timeout` (no message received), or an interrupt from the operator.
 ## When to use
 - You have multiple workspaces with herdr agent sessions and want a single source of truth
 - You suspect YAML drift after a host reboot / crash
 - You want a notification when a session id was just created (so you can record it before next restart)
 - You're running multi-day work and want to know "what's actually running right now"
 ## When NOT to use
 - One-off interactive session — just check `herdr agent list` and read the YAML
 - A single, short session — overhead > benefit
 - You only need a point-in-time answer — use `multi-agent-mux-status` instead
 ## Running the monitor
 ```bash
 # Persistent monitor: runs until interrupted; polls if the broker is unreachable.
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --subscribe --idle-timeout 0
 # Bounded run: exits after 5 min with no message, or 1 h wall-clock, whichever comes first.
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --subscribe --idle-timeout 300 --timeout 3600
 ```
 Run it under whatever supervisor you already use (a dedicated herdr pane, `nohup`,
 or a background job). Nothing else needs to be running for the monitor to work —
 it reconciles YAML ↔ herdr ↔ disk on its own.
 The herdr commands the script issues (`herdr agent list`, `herdr agent get <session>`)
 are real native herdr commands — do not substitute tmux-era names like `herdr ls` /
 `herdr list-panes` outside a shell that has sourced `.agents/skills/lib.sh`.
 ## Helper script: `reconcile.sh`
 This is the whole monitor — there is no separate driver. Each pass:
 1. Diffs YAML ↔ herdr ↔ disk artifacts
 2. Updates YAML if needed (only when changes are real, not on every poll — avoids spamming)
 3. Emits a JSON diff to stdout for the caller to consume
 ```bash
 # Reconcile + auto-update YAML (atomic, flock-guarded). Emits JSON drift to stdout.
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --once --emit-diff
 # Read-only: compute drift WITHOUT writing the YAML (use for "what's running?" checks).
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --once --emit-diff --dry-run
 ```
 Flags: `--once` (single pass), `--emit-diff` (print JSON), `--dry-run` (P1-E — no mutation), `--subscribe` (push-based MQTT subscription monitoring). `--subscribe` sub-flags: `--timeout N` (exit after N seconds of wall-clock; `0` = no limit, default), `--idle-timeout N` (exit after N seconds with no message; default `3600`, `0` = never idle-out). On a broker connection failure (connect error **or** non-zero CONNACK), `--subscribe` falls back to a polling loop that re-runs `--once --emit-diff` every `RECONCILE_POLL_INTERVAL` (default 15) seconds until `--timeout`. Terminal-event YAML updates are written through `lib.sh::atomic_dump_yaml` (flock + schema-validate + `.bak`). There are **no** `--workspace` / `--agent` flags; the emitted JSON `drifts[]` is the caller's to consume.
 ## Drift classes (what the script handles)
 ### Status Enum
 The `status` field MUST be one of the following exact strings: `running`, `stopped`, `terminated`, `archived`.
 The `last_visible_status` is a free-form human-readable status string (e.g. verification-cycle states: `unverified`, `pinned`, `resume_verified`, or a failure detail string) and is NOT constrained to this enum.
 Any unstructured comments or reasons for the status change should be placed in `last_visible_note` or `termination_mode`.
 ### A. herdr dead, YAML says running → auto-terminate
 ```
 YAML:  status=running, pane.pid=201132, cmd=claude
 herdr:  no session
       → set status=terminated, terminated_at=<now>, termination_mode=auto-detected
       → report: "lab-landing-page-creator-claude: herdr gone (was pane 201132, cmd claude). Marked terminated."
 ```
 **Skip-set**: the auto-terminate only fires for sessions whose status is `running`.
 Rows already in a deliberate end state — `terminated`, `archived`, or **`stopped`**
 (set by `multi-agent-mux-stop`) — are
 left untouched. This is critical: a `stopped` row keeps its `resumable: true` and
 captured `*_session_id_own`, so the monitor must **not** overwrite it with
 `terminated ("auto-detected")` when its herdr is (expectedly) gone.
 ### B. herdr alive, not in YAML → auto-register
 ```
 herdr:  session=lab-paper-pdf2md-creator-agy, pid=...,
       cmd=agy, cwd=$WORKSPACE_ROOT/paper-pdf2md
 YAML:  no such session
       → register as new entry: status=running, last_visible_status=running, last_visible_note=auto-registered
       → report: "lab-paper-pdf2md-creator-agy: herdr found but not in YAML. Auto-registered."
 ```
 ### C. New session id materializes (claude first message sent)
 ```
 YAML:  claude_session_id_own=null (placeholder)
 disk:  ~/.claude/projects/.../b3a7...c2f.jsonl exists, mtime=now,
       first line sessionId=b3a7...c2f
       → update claude_session_id_own=b3a7...c2f
       → report: "lab-landing-page-creator-claude: session id materialized b3a7...c2f"
 ```
 ### D. Stale UUID (artifact gone)
 ```
 YAML:  agent_identities.claude.session_id=87dc548e-...
 disk:  ~/.claude/projects/.../87dc548e-...jsonl: missing
       → report it, but DO NOT delete from YAML
       (the user may have moved the file or the disk may be temporarily unavailable;
        only `--purge-conversation` should remove the id)
 ```
 ## Pitfalls
 - **Don't expect `--once` to stay alive** — it does a single pass and exits. Use `--subscribe` for continuous monitoring.
 - **`--idle-timeout` defaults to 3600s** — a monitor meant to run indefinitely needs `--idle-timeout 0` explicitly, or it will quietly exit after an hour of broker silence.
 - **The poll interval is a default** — `RECONCILE_POLL_INTERVAL` (15s) is what the broker-down fallback uses. A workspace with 5+ agent sessions can bump it to reduce noise.
 - **Coalesce repeated drifts** — the same drift re-appears on every pass until it is resolved. A caller that acts on `drifts[]` should compare against the previous pass and act only on *new* entries; the script does not deduplicate for you.
 - **Don't fight the user's explicit action** — if `multi-agent-mux-stop` is mid-flight and the monitor sees the same session in two states within 5s, prefer the user's most recent action. The monitor should not auto-revert a fresh `terminated` to `running` because of a stale `herdr has-session` check.
 - **The monitor should never modify the conversation artifacts** (jsonl, db) — only the YAML. If you see a stale UUID, report it but don't delete the file.
 - **TUI capture-pane is expensive** — only capture when you need to update `last_visible_status`, not every poll.
 ## Supervising-agent runbook
 If an agent drives the monitor rather than an operator watching it directly, this is
 the behavior spec:
 ```markdown
 # agent-sessions monitor
 ## Loop
 1. Read agent-sessions.yaml
 2. Bash: `bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --once --emit-diff`
 3. Parse the JSON diff from stdout
 4. If `drifts` is non-empty, report each *new* drift to the operator
 5. Bash: `sleep 30`, then repeat
 ## Stop condition
 Stop when the operator says to stop, or when the surrounding job's timeout fires.
 ## Drift responses
 - A. herdr dead + YAML running: auto-terminate YAML, report
 - B. herdr alive not in YAML: auto-register, report
 - C. New session id from *.jsonl: update YAML, report
 - D. Stale UUID: report only, no YAML change
 ## Hard rules
 - Do NOT modify conversation artifacts (jsonl, db, brain/)
 - Do NOT spawn/delete herdr sessions — that's the create/delete skills' job
 - Do NOT call multi-agent-mux-create or multi-agent-mux-stop — only the user initiates those
 - Do NOT call `git commit` / `git push`
 ```
 ## Security: --subscribe on Public Brokers
 When using `--subscribe` with the default PoC public broker
 (`broker.hivemq.com:1883`), be aware that:
 1. **Wildcard subscription** means anyone can publish events to your job topics.
 2. **Auto-kill on terminal events** means a spoofed `completed` or `error`
   event from a third party can terminate your agent session.
 3. **Mitigation**: Use `--subscribe` only on private TLS-enabled brokers
   (production mode). For PoC, prefer polling-based monitor (`--once` or
   no `--subscribe`) which reads YAML/herdr state directly without MQTT.
 4. **HMAC verification**: Events are now verified via `verify_hmac()` in
   `mqtt_common.py` (see FW-05). Ensure `auth_token` is set for each job
   to enable signature validation — unauthenticated events will be dropped.
 ## Verification (one-shot)
 ```bash
 # Run reconcile once and inspect output
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --emit-diff --once \
  | python3 -m json.tool
 ```
 ## Related skills
 - `multi-agent-mux-status` — read-only snapshot when you don't need a running loop
 - `multi-agent-mux-delegate-job` — the MQTT job channel whose events `--subscribe` listens to
 ````
 ---
 ## 부록 B — F2~F4 자동 적용 스크립트
 각 치환의 등장 횟수가 정확히 1인지 검사하고 아니면 즉시 중단한다. `replace_all` 사고를 구조적으로 막는다.
 ```python
 import io, sys
 def sub(path, pairs):
    s = io.open(path, encoding='utf-8').read()
    for old, new in pairs:
        n = s.count(old)
        if n != 1:
            print(f"!! {path}: expected 1 occurrence, got {n} for: {old[:70]!r}")
            sys.exit(1)
        s = s.replace(old, new)
    io.open(path, 'w', encoding='utf-8').write(s)
    print(f"ok {path}  ({len(pairs)} edits)")
 sub('.agents/skills/multi-agent-mux-status/SKILL.md', [
 ("resume UUID on disk, and any drift. No Kanban, no mutation. Reuses",
  "resume UUID on disk, and any drift. No mutation. Reuses"),
 ("without spinning up a Kanban monitor worker.",
  "without spinning up the monitor loop."),
 ("**No mutation. No Kanban. No polling loop.**",
  "**No mutation. No polling loop.**"),
 ("dispatch `multi-agent-mux-monitor` (Kanban worker) or run",
  "run `multi-agent-mux-monitor` or"),
 ("- Quick sanity check before dispatching a long Kanban task",
  "- Quick sanity check before dispatching a long-running job"),
 ("- Continuous live tracking → `multi-agent-mux-monitor` (Kanban worker)",
  "- Continuous live tracking → `multi-agent-mux-monitor`"),
 ])
 sub('.agents/skills/multi-agent-mux-create/SKILL.md', [
 ('If any check fails → `kanban_block(reason="...")` (worker path) or report to user (interactive path).',
  "If any check fails → abort with a non-zero exit and report the reason (automated path) or report to user (interactive path)."),
 ("(e.g., cmux, Kanban workers, manual herdr sessions)",
  "(e.g., cmux, background workers, manual herdr sessions)"),
 ])
 sub('README.md', [
 ("A long-running Kanban reconcile worker that dynamically monitors",
  "A long-running reconcile worker that dynamically monitors"),
 ])
 sub('README.ko.md', [
 ("백그라운드에서 Kanban Reconcile 프로세스로 실행되어",
  "백그라운드 Reconcile 프로세스로 실행되어"),
 ])
 ```
 실측 출력:
 ```
 ok .agents/skills/multi-agent-mux-status/SKILL.md  (6 edits)
 ok .agents/skills/multi-agent-mux-create/SKILL.md  (2 edits)
 ok README.md  (1 edits)
 ok README.ko.md  (1 edits)
 ```
 ---
 ## 부록 C — G-C 게이트 전문 (`.mam/gates/gate_gc.py`) 🆕 Rev.2
 토큰 목록은 **F1 이 삭제하는 텍스트에서 파생**한 것이지 손으로 열거한 것이 아니다.
 HEAD 에서 `61 FAIL`, 적용 후 `0 PASS` 로 red/green 실측 완료.
 ```python
 #!/usr/bin/env python3
 """G-C — Kanban CLI/API surface residue gate.
 Token list is DERIVED from the text F1 deletes, not hand-enumerated."""
 import sys, glob
 KANBAN_SURFACE = [
    "kanban", "KANBAN", "Kanban",
    "kanban_comment", "kanban_block", "kanban_heartbeat",
    "HERMES_KANBAN_TASK", "goal_mode",
    "--goal", "--goal-max-turns", "--comment-card",
    "--assignee", "--max-runtime", "--max-retries",
    "--title", "--branch", "--body", "--skill",
 ]
 # Removed by F1 but NOT Kanban surface: a real env var (reconcile.sh:23) whose doc
 # line goes away for a different reason (false .cache claim, plan §8-a). Never
 # make it a forbidden word — it may legitimately be re-documented later.
 EXCLUDED = {"AGENT_SESSIONS_STATE_DIR"}
 targets = sorted(glob.glob('.agents/skills/*/SKILL.md')) + ['README.md', 'README.ko.md']
 hits = []
 for f in targets:
    try:
        lines = open(f, encoding='utf-8').read().splitlines()
    except FileNotFoundError:
        continue
    for i, ln in enumerate(lines, 1):
        for t in KANBAN_SURFACE:
            if t in ln:
                hits.append((f, i, t, ln.strip()[:70]))
 if hits:
    print(f"G-C FAIL — {len(hits)} Kanban-surface hit(s):")
    for f, i, t, s in hits:
        print(f"   {f}:{i}  [{t}]  {s}")
    sys.exit(1)
 print("G-C PASS — 0 Kanban CLI/API surface tokens")
 ```
 **호출 시 주의**: 실패 판정은 **exit code** 로 받아야 한다. `| head` 같은 파이프를 붙이면 파이프라인 종료코드가 마지막 명령의 것으로 바뀌어 FAIL 이 0 으로 보인다.
 ```bash
 python3 .mam/gates/gate_gc.py && echo "G-C ok" || echo "G-C FAILED"
 ```
 **파생 절차 재현법** (토큰 목록을 다시 뽑아야 할 때):
 ```python
 import re, subprocess, io
 head = subprocess.run(['git','show','HEAD:.agents/skills/multi-agent-mux-monitor/SKILL.md'],
                      capture_output=True, text=True).stdout
 new  = io.open('.agents/skills/multi-agent-mux-monitor/SKILL.md', encoding='utf-8').read()
 TOK = re.compile(r'--[a-zA-Z][\w-]+|[A-Za-z_][\w]*_[\w]+|\b[A-Z][A-Z0-9_]{3,}\b')
 print(sorted(set(TOK.findall(head)) - set(TOK.findall(new))))
 ```
 이 절차가 산출한 16개 토큰 중 15개가 Kanban 표면, 1개(`AGENT_SESSIONS_STATE_DIR`)가 제외 대상이다.
 ---
 [AGREEMENT: REACHED]
@@ -0,0 +1,153 @@
 # Cross-Code Review: C-1 Kanban Documentation Removal (Job 27b877f1)
 - **Reviewer**: cline (session: herdr:canary-projects-multi-agent-mux-creator-cline)
 - **Task**: C-1 — Delete all references to Kanban (Kanban worker, goal_mode, kanban status, etc.) across all SKILL.md documentation files; achieve 0 remaining references on the product surface.
 - **Change scope**: 6 files, working-tree (uncommitted) modifications — `.agents/skills/multi-agent-mux-{create,monitor,status}/SKILL.md`, `README.md`, `README.ko.md`, `IMPROVEMENTS.md`. 79 insertions, 94 deletions. **Documentation-only — no `.sh`/`.py` code was modified.**
 - **Date**: 2026-08-06
 ---
 ## 1. Verification of Kanban Reference Removal
 ### 1.1 Product-surface grep (case-insensitive)
 Searched all 5 product-surface files for `kanban` (case-insensitive):
 | File | Kanban refs (before) | Kanban refs (after) |
 |---|---|---|
 | `.agents/skills/multi-agent-mux-monitor/SKILL.md` | ~22 | **0** |
 | `.agents/skills/multi-agent-mux-status/SKILL.md` | ~5 | **0** |
 | `.agents/skills/multi-agent-mux-create/SKILL.md` | ~2 | **0** |
 | `README.md` | 1 | **0** |
 | `README.ko.md` | 1 | **0** |
 `grep -cniE 'kanban'` on all 5 files returns exit code 1 (no matches) → **0 references remain on the product surface.** ✅
 ### 1.2 Kanban function / flag / skill-name residue
 Searched for specific Kanban API surface tokens across the same files:
 `kanban_block`, `kanban_comment`, `kanban_heartbeat`, `HERMES_KANBAN_TASK`, `goal_mode`, `goal-max-turns`, `kanban-worker`, `kanban-orchestrator`.
 → **NO LEFTOVER KANBAN FUNCTION/FLAG REFERENCES IN SKILL.md/README.** ✅
 ### 1.3 G-C derived verification gate
 `python3 .mam/gates/gate_gc.py` → **`G-C PASS — 0 Kanban CLI/API surface tokens`** ✅
 The gate scans the codebase for the full Kanban surface (`kanban`, `KANBAN`, `Kanban`, `kanban_comment`, `kanban_block`, `kanban_heartbeat`, `HERMES_KANBAN_TASK`, `goal_mode`) and reports zero hits (excluding the gate file itself and the legitimate `RECONCILE_POLL_INTERVAL` env var, which is explicitly allow-listed).
 ### 1.4 Code files (`.sh`, `.py`)
 `grep -rni 'kanban' .agents/skills/ --include='*.sh' --include='*.py'` → **0 hits.** No code ever referenced Kanban; the docs were the only residue. ✅
 ### 1.5 Remaining "kanban" occurrences (out of scope, expected)
 The only remaining `kanban` string occurrences in the repo are intentional and outside the product surface:
 - **`IMPROVEMENTS.md`** (lines 3, 108–113): The C-1 completion changelog entry. These are meta-references *documenting that the removal was done*. They live in the backlog tracker, not the product surface (`README*.md` / `.agents/skills/`). Intentional and correct.
 - **`.mam/gates/gate_gc.py`**: The gate that *detects and blocks* Kanban surface tokens. It contains the word "kanban" because it is the detection mechanism. Intentional and correct.
 - **`.mam/delegate_job_logs/*/meta.json`**: Internal job records/prompts. Out of scope.
 The task target ("0 remaining references") is scoped to the product surface (`.agents/skills/`, `README*.md`), which is confirmed at 0.
 ---
 ## 2. Documentation Accuracy (docs ↔ implementation cross-check)
 The largest rewrite is `multi-agent-mux-monitor/SKILL.md` (133 lines changed). The new description changes the execution model from "Kanban worker that polls every ~30s" to "reconciler that runs `reconcile.sh --subscribe` (MQTT push + broker-down polling fallback)". I verified every flag/constant mentioned in the rewritten docs against the actual `reconcile.sh` source:
 | SKILL.md claim | `reconcile.sh` source | Match |
 |---|---|---|
 | `--subscribe` (persistent loop) | line 40: `--subscribe) SUBSCRIBE=1` | ✅ |
 | `--once` (single pass) | line 37: `--once) ONCE=1` | ✅ |
 | `--emit-diff` (JSON drift output) | line 38: `--emit-diff) EMIT_DIFF=1` | ✅ |
 | `--idle-timeout` defaults to 3600s | line 29: `SUB_IDLE_TIMEOUT=3600` | ✅ |
 | `--timeout` (wall-clock) | line 41: `--timeout) SUB_TIMEOUT="$2"` | ✅ |
 | `RECONCILE_POLL_INTERVAL` default 15 | line 33: `POLL_INTERVAL="${RECONCILE_POLL_INTERVAL:-15}"` | ✅ |
 | `--dry-run` | line 39: `--dry-run) DRY_RUN=1` | ✅ |
 | Broker-down fallback → poll | line 303: `bash "$_self" --once --emit-diff` (on exit code 3) | ✅ |
 All documented flags and defaults are accurate. ✅
 ### 2.1 Replaced skill references are real
 The monitor SKILL.md `related_skills` and `prereq_skills` now reference:
 - `multi-agent-mux-status` → exists at `.agents/skills/multi-agent-mux-status/` ✅
 - `multi-agent-mux-delegate-job` → exists at `.agents/skills/multi-agent-mux-delegate-job/` ✅
 The old non-existent references `kanban-worker` and `kanban-orchestrator` are gone, and no `kanban-*` skill directories exist. ✅
 ### 2.2 create/SKILL.md change
 `kanban_block(reason="...")` (worker path) → "abort with a non-zero exit and report the reason (automated path)". This removes a reference to a Kanban-only API that was never implemented in the actual `create_session.sh` script. The replacement ("abort with non-zero exit") accurately describes the real script behavior. ✅
 "Kanban workers" → "background workers" in the isolation section. Neutral, accurate. ✅
 ### 2.3 status/SKILL.md changes
 Four Kanban references removed: description ("No Kanban" → removed; "Kanban monitor worker" → "the monitor loop"), pitfalls ("Kanban worker" → bare skill name), when-to-use ("long Kanban task" → "long-running job"), when-not-to-use ("Kanban worker" → bare skill name). All replacements are neutral and accurate. ✅
 ---
 ---
 ## 3. IMPROVEMENTS.md Consistency
 The backlog tracker was updated to reflect C-1 completion:
 - Header: total tracked open items **18 → 17** (legacy remnants 5 → 4); completed **3 → 4** (added C-1). ✅
 - Section 4 heading: "레거시 잔재 (5건 → 4건)". The C-1 entry was removed from the open backlog. ✅
 - Section 5: C-1 added under "완료된 과제" with a 5-line completion summary. ✅
 - Section 6 (roadmap): legacy remnants count updated 5 → 4. ✅
 Arithmetic is internally consistent: 17 = 2 (arch) + 8 (edge) + 3 (orch) + 4 (legacy). ✅
 ---
 ## 4. Validation
 ### 4.1 Syntax / structure
 - `bash -n reconcile.sh` → OK (unchanged anyway; docs-only change). ✅
 - YAML frontmatter of all 3 modified SKILL.md files parses cleanly with `yaml.safe_load` (name field present). ✅
 ### 4.2 Tests
 This is a **documentation-only** change — no `.sh`/`.py` code was modified, so no code path can have regressed. For completeness, I ran the unit tests that do not require a live/mock herdr server:
 | Test file | Result |
 |---|---|
 | `tests/test_workspace_scope.py` | 2/2 PASS |
 | `tests/test_b1_tier3_identity.py` | 8/8 PASS |
 | `tests/test_tier1_unit.py` (`-k create`) | 7/7 PASS |
 Tests that spawn herdr subprocesses (`resume`/`stop`/`status`/`mock_herdr`) hang — this is a **pre-existing test-infra limitation** (require a live/mock herdr server), identical to the prior reviews, and not a regression. Since no code changed in this diff, these tests are not exercising any modified path. ✅
 ### 4.3 Gate checklist
 | Gate | Result |
 |---|---|
 | G-C (Kanban CLI/API surface) | **PASS** — 0 tokens |
 | Product-surface Kanban refs | **0** across all 5 files |
 | Documentation ↔ implementation accuracy | All flags/defaults verified |
 | Replaced skill refs exist | `multi-agent-mux-status`, `multi-agent-mux-delegate-job` both real |
 | YAML frontmatter validity | 3/3 OK |
 | IMPROVEMENTS.md arithmetic | Consistent (17 = 2+8+3+4; 4 completed) |
 ---
 ## 5. Findings
 ### K-1 (none) — No new issues found
 The change cleanly and completely removes all Kanban references from the product surface (SKILL.md ×3, README ×2), replaces non-existent skill references with real ones, rewrites the monitor execution-model description to accurately match the actual `reconcile.sh --subscribe` implementation, and updates the IMPROVEMENTS.md backlog tracker consistently. The G-C gate confirms 0 Kanban CLI/API surface tokens remain.
 The only residual `kanban` strings live in (a) the IMPROVEMENTS.md completion changelog — which legitimately records that C-1 was done, and (b) the G-C gate file itself — which is the detection mechanism. Both are intentional and out of the product-surface scope defined by the task.
 ---
 ## 6. Verdict
 The C-1 task — "Delete all references to Kanban across all SKILL.md documentation files and achieve 0 remaining references on the product surface" — is fully achieved. Verification confirms 0 Kanban references across all 5 product-surface files, the G-C gate passes, the rewritten documentation accurately reflects the actual `reconcile.sh` implementation, and the IMPROVEMENTS.md tracker is consistent. No code was changed, so there is no regression risk.
 **No merge-blocking issues remain.**
 [VERDICT: PASS]
@@ -0,0 +1,125 @@
 # Cross-Code Review Report — Job 87f84d4e
 - **Job ID**: 87f84d4e
 - **Reviewer**: cline (session: herdr:canary-projects-multi-agent-mux-creator-cline)
 - **Review Target**: O-1 — Reviewer feedback rebuttal protocol implementation (Rebuttal & Re-adjudication Protocol)
 - **Date**: 2026-08-06
 - **Diff scope**: 5 files modified + 1 new test file (249 ins / 22 del)
 ---
 ## 1. Change Inventory
 | # | File | Type | Δ | Purpose |
 |---|------|------|---|---------|
 | 1 | `.agents/skills/multi-agent-mux-loop/scripts/run_loop.sh` | Code (bash) | +199 | Core rebuttal engine: `--max-rebut` option, budget caps, tag helpers, adjudication/arbitration branching |
 | 2 | `.agents/MULTI_AGENT_RULES.md` | Docs (EN) | +21/-4 | New §3.1 protocol spec, reviewer re-adjudication duty, §3.1 cross-ref |
 | 3 | `.agents/MULTI_AGENT_RULES.ko.md` | Docs (KO) | +19/-2 | Korean mirror of §3.1 spec + re-adjudication duty + cross-ref |
 | 4 | `.agents/skills/multi-agent-mux-loop/SKILL.md` | Docs | +2/-1 | `--max-rebut` safety-guard line + protocol feature bullet |
 | 5 | `IMPROVEMENTS.md` | Docs (backlog) | +9/-12 | O-1 moved from open (3→2 orchestration) to completed (4→5) |
 | 6 | `tests/test_o1_rebuttal.py` | Test (new) | +176 | Regression suite V-1..V-10 |
 **Diff stat**: `run_loop.sh 199 +-, IMPROVEMENTS.md 23, RULES.md 25, RULES.ko.md 21, SKILL.md 3` — matches `git diff --stat` (249 ins / 22 del).
 ---
 ## 2. Lint / Syntax Verification
 | Check | Command | Result |
 |-------|---------|--------|
 | Bash syntax | `bash -n run_loop.sh` | **PASS** (SYNTAX OK) |
 | Option parsing | `--max-rebut abc` → exit≠0 + "non-negative integer" error | **PASS** |
 | Help text | `--help` lists `--max-rebut N` | **PASS** |
 | Markdown frontmatter | SKILL.md YAML untouched (body-only edit) | **PASS** |
 No shellcheck in stock macOS env; manual review of all array expansions confirms bash 3.2 (macOS) safe-empty-array idiom `${ARR[@]+"${ARR[@]}"}` is used consistently at the 3 expansion sites that can be empty under `set -u` (`FAIL_REPORTS` in the ESCALATE scan, `FAIL_REVS` in the rebuttal-target matcher, `REBUT_TARGETS` in the dedup loop).
 ---
 ## 3. Correctness / Operability Analysis
 ### 3.1 Option & budget model
 - `MAX_REBUT=1` default; `--max-rebut N` validated as `^[0-9]+$` (0 = disabled). ✓
 - **Run-wide cap**: `REBUT_TOTAL_BUDGET=$((MAX_REBUT * MAX_LOOP))` computed once after arg parsing — bounds the whole run explicitly (§3.1 rule 3). ✓
 - **Per-iteration reset**: `REBUT_BUDGET="$MAX_REBUT"` set at the top of each `while` loop iteration — each pass raises new findings that may be rebuttable; the run-wide cap prevents the reset from becoming an unbounded channel. ✓ Matches §3.1 rule 3.
 - Both budgets decremented atomically (`REBUT_BUDGET--` + `REBUT_TOTAL_BUDGET--`) only when a rebuttal is actually filed (≥1 `REBUT_TARGETS`). ✓
 ### 3.2 Tag helpers
 - `has_tag_line <file> <tag>`: `grep -qE "^\[${tag}\][[:space:]]*\r?$"` — matches a standalone `[TAG]` line, tolerates CRLF. **Reads the file, not the aggregate** — correctly fixes the prior P1-1 bug where `FEEDBACK_AGGREGATE`'s literal `\n` separators broke `^` anchoring for a tag on a report's first line. ✓
 - `extract_tag_values <file> <key>`: `sed -n -E "s/^\[${key}:[[:space:]]*([^]]*)\][[:space:]]*\r?$/\1/p"` then trailing-space strip + blank-line drop. Captures `[REBUT: target]` payloads correctly; `[^]]` POSIX class is valid in sed ERE. ✓
 - `find_report <job_id>`: `find ".mam/jobs/$1" -maxdepth 2 -name report-final.md` — verified depth: report sits at `.mam/jobs/<id>/<agent-subdir>/report-final.md` (depth 2), so `-maxdepth 2` reaches it regardless of agent-name subdir. ✓
 ### 3.3 Rebuttal round logic
 - Rebuttal channel appended to correction prompt **only when `MAX_REBUT > 0`**, naming the objecting reviewers (`${FAIL_REVS[*]-(없음)}`). ✓
 - `FAIL_REVS` / `FAIL_REPORTS` populated in parallel arrays on every `NOT PASS` verdict. ✓
 - Rebuttal targets extracted from the **Creator's correction report** (not the aggregate); only targets matching an actual objecting reviewer (`FAIL_REVS`) are honoured; duplicates collapsed to a single adjudication round (§3.1 rule 5). ✓
 - `while read` fed via here-string `<<< "$(...)"` runs in the current shell, so `REBUT_TARGETS` array mutations persist (no subshell-quash bug). ✓
 ### 3.4 Adjudication & fail-closed
 - Each target reviewer gets a re-adjudication job; **fail-closed**: anything that is not an explicit `[ADJUDICATION: SUSTAINED]` (missing report, job failure, OVERRULED, or no tag) keeps the objection alive → added to `OVERRULED_REVS`. ✓ (§3.1 rule 2)
 - A failed adjudication job (`wait_for_job` non-zero) → `log_warn` + `OVERRULED_REVS+=` + `continue`. No crash, no silent pass. ✓
 ### 3.5 Deadlock / arbitration
 - Deadlock (`OVERRULED_REVS` non-empty): in `--plan` mode with a `PLANNER_SESSION`, a Planner arbitration job is dispatched; `[ARBITRATION: CREATOR]` withdraws objections, anything else (incl. job failure) defaults to `REVIEWER`. ✓ (§3.1 rule 4)
 - Without `--plan` / planner: reviewer prevails (fail-closed, charter §1). ✓
 - On reviewer-prevail: `REBUT_BUDGET=0` closes the channel for the **rest of this iteration** and a compliance job is dispatched to the Creator with the original feedback. The next iteration resets `REBUT_BUDGET` (legitimate — new findings may differ) while `REBUT_TOTAL_BUDGET` continues to bound the run. ✓
 ### 3.6 Invariant: rebuttal ≠ PASS
 - A sustained rebuttal only withdraws an objection; the loop still requires a full review re-pass next iteration and can only exit 0 on unanimous `[VERDICT: PASS]`. No path lets a rebuttal shortcut the completion gate. ✓ (§3.1 rule 1)
 ---
 ## 4. Documentation Consistency
 | Item | EN | KO | Parity |
 |------|----|----|--------|
 | §3.1 Rebuttal & Adjudication Protocol section | ✓ added | ✓ added | ✓ |
 | Tag table (REBUT / ADJUDICATION / ARBITRATION) | ✓ | ✓ | ✓ |
 | 5 protocol rules (PASS-substitute, fail-closed, budget, deadlock, dedup) | ✓ | ✓ | ✓ |
 | Reviewer re-adjudication duty bullet | ✓ | ✓ | ✓ |
 | Developer §3.1 cross-ref in "Review Integrity & Refusal" | ✓ | ✓ | ✓ |
 | SKILL.md `--max-rebut` (safety guard + feature bullet) | ✓ | n/a | ✓ |
 | IMPROVEMENTS.md: O-1 → completed; 16 open (2+8+2+4); 5 done | ✓ | n/a | ✓ arithmetic consistent |
 **Usage/help ↔ SKILL.md consistency**: `run_loop.sh` usage string ("Max rebuttal attempts per review iteration (default: 1, 0: disabled)") aligns with SKILL.md feature bullet. ✓
 ---
 ## 5. Test Verification
 | Suite | Result |
 |-------|--------|
 | `tests/test_o1_rebuttal.py` (V-1..V-10) | **10/10 PASS** (0.11s) |
 | `tests/test_workspace_scope.py` + `tests/test_b1_tier3_identity.py` (regression) | **10/10 PASS** (1.28s) |
 | `bash -n run_loop.sh` | PASS |
 O-1 suite coverage: tag-line first-line match (V-1), multi-rebut extraction (V-2), option parsing/error (V-3), dedup (V-4), fail-closed on missing SUSTAINED (V-5), bash 3.2 empty-array under `set -u` (V-6), §3.1 docs present EN+KO (V-7), SKILL.md `--max-rebut` (V-8), per-iteration reset + total cap (V-9), IMPROVEMENTS.md O-1 completed (V-10).
 **Pre-existing infra limitation (unchanged by O-1)**: `test_tier1_unit.py` / `test_sanity.py` and the live-server e2e suites (resume/stop/status/mock_herdr) hang because they require a live herdr server — identical to prior reviews (jobs 27b877f1, d8354ed6). O-1 modifies only `run_loop.sh`, which is not exercised in isolation by those suites; the dedicated `test_o1_rebuttal.py` covers the new logic. No regression introduced.
 ---
 ## 6. Findings
 ### R-1 (Minor, Non-blocking) — Out-of-scope EN role-suitability rewording → EN/KO drift
 - **Location**: `.agents/MULTI_AGENT_RULES.md` lines 37-38.
 - **Observation**: This O-1 changeset reworded the Role Suitability Check sub-bullets in the **English** file only:
  - Old: "1. Recommend the optimal agent session to delegate the task to, or / 2. Perform the task directly if strictly necessary for project continuity."
  - New: "1. Defer or re-delegate the task to a suitable subagent/session, OR / 2. Reject the task explicitly by explaining the role mismatch."
 - **Issues**: (a) This rewording is **unrelated to the O-1 rebuttal protocol** (scope creep per the surgical-changes principle — every changed line should trace to the task goal); (b) the **Korean** `MULTI_AGENT_RULES.ko.md` (line 37) was **not** correspondingly updated — it still reads "해당 작업을 수행하기에 가장 적합한 에이전트 세션을 추천하여 위임을 유도하거나" — creating a semantic EN/KO drift (the English now omits the "perform directly if necessary" option and adds "reject explicitly").
 - **Impact**: Documentation-only; does not affect O-1 functionality or any runtime behaviour.
 - **Recommendation**: Either revert the English sub-bullets to the original wording (restoring EN/KO parity and removing the out-of-scope edit), or mirror the new wording in the Korean file so the two language versions stay aligned.
 **No other findings.** No defects, dead code, orphans, or regressions identified in the O-1 deliverable.
 ---
 ## 7. Verdict
 The O-1 Rebuttal & Re-adjudication Protocol is **fully and correctly implemented**:
 - `run_loop.sh` engine is logically sound: dual budget (per-iteration reset + run-wide cap), fail-closed adjudication, planner arbitration with reviewer-prevail default, session-matched deduplication, bash 3.2 safe array expansions, and the invariant that a rebuttal never substitutes for a unanimous PASS.
 - Documentation (§3.1 EN+KO, SKILL.md, IMPROVEMENTS.md) is consistent and internally coherent.
 - 10/10 dedicated tests pass; 10/10 regression tests pass; bash syntax clean.
 - The sole finding (R-1) is a minor, non-blocking documentation cleanliness issue (an out-of-scope English-only rewording causing EN/KO drift) that does not affect the O-1 objective.
 [VERDICT: PASS]
@@ -48,7 +48,7 @@ claude auth status 2>&1 | python3 -c "import json,sys; d=json.load(sys.stdin); a
 test -d "$WORKSPACE" || { echo "ERROR: workspace $WORKSPACE not a directory"; exit 1; }
 ```
-If any check fails → `kanban_block(reason="...")` (worker path) or report to user (interactive path). Do not proceed with a half-broken setup.
+If any check fails → abort with a non-zero exit and report the reason (automated path) or report to user (interactive path). Do not proceed with a half-broken setup.
 ## Standard names
@@ -62,7 +62,7 @@ If any check fails → `kanban_block(reason="...")` (worker path) or report to u
 ## Herdr Server Isolation (격리 서버)
-When running multiple agent sessions alongside other workflows (e.g., cmux, Kanban workers, manual herdr sessions), sharing the default herdr server can lead to session name conflicts, monitoring clutter, and accidental destruction of user sessions via global commands.
+When running multiple agent sessions alongside other workflows (e.g., cmux, background workers, manual herdr sessions), sharing the default herdr server can lead to session name conflicts, monitoring clutter, and accidental destruction of user sessions via global commands.
 To prevent this, you can run this skill inside an **isolated herdr server** using the `HERDR_SERVER_NAME` environment variable or the `--herdr-server <name>` flag (opt-in).
@@ -17,7 +17,7 @@ metadata:
 > **Companion skills**: `multi-agent-mux-create` (start), `multi-agent-mux-resume` (re-attach), `multi-agent-mux-delegate-job` (delegate).
-> **Safety Guard**: `--max-loop` and `--plan-talk` restrict API cost runaways.
+> **Safety Guard**: `--max-loop`, `--max-rebut`, and `--plan-talk` restrict API cost runaways.
 > **Single source of truth**: `./.mam/agent-sessions.yaml`.
 수동 템플릿 작성 및 수동 프롬프트 환류는 폐지되었습니다. Planner, Creator, Reviewer 간의 모든 협업 피드백 루프는 본 스킬(`run_loop.sh`)만을 단독으로 사용하여 자동으로 오케스트레이션합니다.
@@ -29,6 +29,7 @@ Run an autonomous planning-execution-review loop using multiple agents (Planner,
 - **Creator Self-Planning & Development** (default without `--plan`): Planner 에이전트에게 계획 작성을 위임하지 않고, 기존에 승격된 계획서가 있다면 이를 로드하여 코드를 구현하며, 계획서가 존재하지 않는 경우 작업자(Creator: developer/writer)가 스스로 구현 계획 및 설계 수립을 포함한 개발 전 과정을 직접 진행합니다.
 - **Targeted Peer-Review** (`--reviewer`): Runs custom-selected reviewer agents to verify code changes.
 - **Total Peer-Review** (`--all-reviewer`): Enforces a unanimous PASS verdict from all registered reviewer sessions.
 - **Rebuttal & Adjudication Protocol** (`--max-rebut N`): Allows Creator to reject invalid reviewer feedback using `[REBUT:]` tags and triggers Reviewer re-adjudication (`[ADJUDICATION:]`) or Planner arbitration (`[ARBITRATION:]`).
 - **Self-Review** (default): Creator verifies its code changes autonomously without peer reviews.
 - **Safety Limits** (`--max-loop N`): Aborts execution if reviews fail to PASS after N iterations.
@@ -16,6 +16,7 @@ PLAN_MODE=false
 PLAN_TALK_TURNS=1
 ALL_REVIEWERS=false
 MAX_LOOP=3
 MAX_REBUT=1
 VERBOSE=false
 CLEANUP=false
 TARGET_AGENT=""
@@ -31,6 +32,7 @@ usage() {
  echo "  --reviewer \"A,B\"       Targeted reviewer session name list (comma-separated)"
  echo "  --all-reviewer         Enforce PASS verdict from all active reviewer sessions"
  echo "  --max-loop N           Max execution-review corrective loop runs (default: 3)"
  echo "  --max-rebut N          Max rebuttal attempts per review iteration (default: 1, 0: disabled)"
  echo "  --verbose              Print detailed execution timeline traces"
  echo "  --cleanup              Purge temporary job directories upon success"
  exit 1
@@ -54,6 +56,12 @@ while [[ "$#" -gt 0 ]]; do
        exit 1
      fi
      MAX_LOOP="$2"; shift 2 ;;
    --max-rebut)
      if [[ ! "$2" =~ ^[0-9]+$ ]]; then
        echo "ERROR: --max-rebut requires a non-negative integer."
        exit 1
      fi
      MAX_REBUT="$2"; shift 2 ;;
    --verbose) VERBOSE=true; shift ;;
    --cleanup) CLEANUP=true; shift ;;
    --target-agent) TARGET_AGENT="$2"; shift 2 ;;
@@ -63,6 +71,10 @@ while [[ "$#" -gt 0 ]]; do
  esac
 done
 # Run-wide cap on rebuttal rounds. Makes the bound explicit and enforced
 # rather than merely emergent from MAX_LOOP x MAX_REBUT.
 REBUT_TOTAL_BUDGET=$((MAX_REBUT * MAX_LOOP))
 if [ -z "$TARGET_AGENT" ] || [ -z "$TASK" ]; then
  echo "ERROR: --target-agent and --task are mandatory fields."
  usage
@@ -116,6 +128,31 @@ has_verdict() {
  [[ "$last_line" =~ $pattern ]]
 }
 # Detect a standalone '[TAG]' line anywhere in a report FILE.
 #
 # Always read the file, never a concatenated aggregate: FEEDBACK_AGGREGATE
 # joins reports with a literal backslash-n (bash `echo` without -e leaves it
 # as text), so a tag on a report's FIRST line ends up sharing a physical line
 # with the '--- Reviewer (x) Feedback ---' banner and the '^' anchor misses it.
 has_tag_line() {
  local file="$1" tag="$2"
  [ -f "$file" ] || return 1
  grep -qE "^\[${tag}\][[:space:]]*\r?\$" "$file"
 }
 # Emit the payload of every standalone '[KEY: value]' line in a report file.
 extract_tag_values() {
  local file="$1" key="$2"
  [ -f "$file" ] || return 0
  sed -n -E "s/^\[${key}:[[:space:]]*([^]]*)\][[:space:]]*\r?$/\1/p" "$file" \
    | sed -E 's/[[:space:]]+$//' | grep -v '^$' || true
 }
 # Locate a job's final report (transient .mam job tree, agent-name subdir).
 find_report() {
  find ".mam/jobs/$1" -maxdepth 2 -name "report-final.md" 2>/dev/null | head -n 1 || true
 }
 # Helper: Blocking wait for a delegate job's completion or error state (with safety timeout)
 wait_for_job() {
  local job_id="$1"
@@ -437,6 +474,12 @@ loop_count=1
 while [ "$loop_count" -le "$MAX_LOOP" ]; do
  log_info "Review Loop Iteration $loop_count/$MAX_LOOP..."
  # O-1: the per-iteration budget resets each pass, because each pass produces
  # NEW findings that may be legitimately rebuttable. The run-wide total is
  # capped separately by REBUT_TOTAL_BUDGET so the reset cannot be mistaken for
  # an unbounded channel (see MULTI_AGENT_RULES.md §3.1 rule 3).
  REBUT_BUDGET="$MAX_REBUT"
  if [ "${#REVIEWERS[@]}" -eq 0 ]; then
    log_warn "No reviewers specified. Conducting Creator Self-Review..."
    SELF_REV_OUTPUT=$(delegate_job_safe submit \
@@ -471,6 +514,10 @@ while [ "$loop_count" -le "$MAX_LOOP" ]; do
    # We use space-separated lists or simple loops to bypass bash-4 associative array requirement (M-7 macOS compatibility)
    declare -a JOB_IDS=()
    declare -a JOB_REVS=()
    # Objecting reviewers and their report paths, kept in parallel arrays so the
    # rebuttal round can address each objection individually (O-1).
    declare -a FAIL_REVS=()
    declare -a FAIL_REPORTS=()
    for rev in "${REVIEWERS[@]}"; do
      log_info "Requesting code review from Reviewer '$rev'..."
@@ -528,6 +575,8 @@ while [ "$loop_count" -le "$MAX_LOOP" ]; do
      if has_verdict "$REPORT_FILE" "NOT PASS" || ! has_verdict "$REPORT_FILE" "PASS"; then
        log_warn "Reviewer '$rev': NOT PASS"
        all_passed=false
        FAIL_REVS+=("$rev")
        FAIL_REPORTS+=("$REPORT_FILE")
        FEEDBACK_AGGREGATE="$FEEDBACK_AGGREGATE\n--- Reviewer ($rev) Feedback ---\n$REPORT_CONTENT"
      else
        log_success "Reviewer '$rev': PASS"
@@ -543,13 +592,18 @@ while [ "$loop_count" -le "$MAX_LOOP" ]; do
        exit 1
      fi
-      # Re-planning check: rely on the explicit '[ESCALATE: PLANNER]' tag a
+      # Scan the report FILES, not FEEDBACK_AGGREGATE: the aggregate's literal
-      # reviewer is instructed to emit, rather than sniffing English keywords
+      # '\n' separators break '^' anchoring for a tag on a report's first line.
-      # (reviewers report in Korean, so keyword matching never fired) (P1-1).
+      # bash 3.2 (macOS stock) aborts on "${arr[@]}" when arr is empty under
      # `set -u`. FAIL_REPORTS IS empty on the crashed-job / missing-report
      # paths above, which reach here via all_passed=false without appending.
      COMPLEX_FIX=false
-      if echo "$FEEDBACK_AGGREGATE" | grep -qE '^\[ESCALATE: PLANNER\][[:space:]]*\r?$'; then
+      for _fr in ${FAIL_REPORTS[@]+"${FAIL_REPORTS[@]}"}; do
-        COMPLEX_FIX=true
+        if has_tag_line "$_fr" "ESCALATE: PLANNER"; then
-      fi
+          COMPLEX_FIX=true
          break
        fi
      done
      if [ "$PLAN_MODE" = true ] && [ "$COMPLEX_FIX" = true ]; then
        log_warn "Feedback involves complex code modifications. Diverting to Planner to revise plan..."
@@ -582,6 +636,10 @@ while [ "$loop_count" -le "$MAX_LOOP" ]; do
        CORRECTION_PROMPT="리뷰어들이 지적한 다음 피드백에 입각하여 코드를 수정해주세요. 피드백:\n$FEEDBACK_AGGREGATE"
      fi
      if [ "$MAX_REBUT" -gt 0 ]; then
        CORRECTION_PROMPT="$CORRECTION_PROMPT\n\n[이의제기 채널] 위 피드백 중 타당하지 않다고 판단되는 항목이 있다면, 그 항목은 구현하지 말고 반론을 제기하십시오. 반론 시 리포트에 단독 행으로 '[REBUT: <리뷰어_세션명>]' 태그를 남기고, 해당 리뷰어의 지적 중 어느 항목을 왜 거부하는지 근거를 함께 서술하십시오. 타당한 지적은 정상적으로 반영하십시오. 반론할 항목이 없으면 태그를 남기지 마십시오. 대상 리뷰어 세션명: ${FAIL_REVS[*]-(없음)}"
      fi
      # Creator execution corrective job
      CORRECT_JOB_OUTPUT=$(delegate_job_safe submit \
        --agent-session "herdr:$TARGET_AGENT" \
@@ -597,6 +655,135 @@ while [ "$loop_count" -le "$MAX_LOOP" ]; do
      fi
      CREATED_JOBS+=("$CORRECT_JOB_ID")
      wait_for_job "$CORRECT_JOB_ID"
      # ---------------------------------------------------------------------
      # O-1: REBUTTAL & ADJUDICATION ROUND
      # ---------------------------------------------------------------------
      # The Creator's correction report was previously produced and discarded.
      # Read it: standalone '[REBUT: <reviewer>]' lines open an adjudication
      # round with exactly those reviewers.
      #
      # Invariant: a rebuttal NEVER substitutes for a PASS. A sustained
      # rebuttal only withdraws an objection; the next iteration still runs a
      # full review pass, so the loop can still only exit 0 on unanimous
      # [VERDICT: PASS].
      if [ "$MAX_REBUT" -gt 0 ] && [ "$REBUT_BUDGET" -gt 0 ] && [ "$REBUT_TOTAL_BUDGET" -gt 0 ]; then
        CORRECT_REPORT=$(find_report "$CORRECT_JOB_ID")
        REBUT_TARGETS=()
        if [ -n "$CORRECT_REPORT" ]; then
          while IFS= read -r _t; do
            [ -n "$_t" ] || continue
            # Only honour rebuttals aimed at a reviewer that actually objected,
            # and de-duplicate: a Creator may file one [REBUT:] per rejected
            # finding, all naming the same reviewer — that is ONE adjudication.
            for _fv in ${FAIL_REVS[@]+"${FAIL_REVS[@]}"}; do
              [ "$_t" = "$_fv" ] || continue
              _dup=0
              for _e in ${REBUT_TARGETS[@]+"${REBUT_TARGETS[@]}"}; do
                if [ "$_e" = "$_t" ]; then _dup=1; break; fi
              done
              [ "$_dup" -eq 0 ] && REBUT_TARGETS+=("$_t")
              break
            done
          done <<< "$(extract_tag_values "$CORRECT_REPORT" "REBUT")"
        fi
        if [ "${#REBUT_TARGETS[@]}" -gt 0 ]; then
          REBUT_BUDGET=$((REBUT_BUDGET - 1))
          REBUT_TOTAL_BUDGET=$((REBUT_TOTAL_BUDGET - 1))
          REBUTTAL_TEXT=$(cat "$CORRECT_REPORT" 2>/dev/null || echo "")
          log_warn "Creator filed a rebuttal against: ${REBUT_TARGETS[*]} (iteration budget left: $REBUT_BUDGET, run total left: $REBUT_TOTAL_BUDGET)"
          OVERRULED_REVS=()
          for rt in "${REBUT_TARGETS[@]}"; do
            log_info "Re-adjudication: returning rebuttal to Reviewer '$rt'..."
            ADJ_OUTPUT=$(delegate_job_safe submit \
              --agent-session "herdr:$rt" \
              --agent "$(resolve_agent_type "$rt")" \
              --type "direct" \
              --role "Reviewer" \
              --prompt "작업자(Creator)가 귀하의 리뷰 지적 중 일부를 타당하지 않다고 판단하여 반론을 제기했습니다. 반론을 검토하고 재심(re-adjudication)하십시오. 반론이 타당하여 귀하의 지적을 철회한다면 리포트 마지막에 단독 행으로 '[ADJUDICATION: SUSTAINED]' 를, 반론이 타당하지 않아 지적을 유지한다면 '[ADJUDICATION: OVERRULED]' 를 명시하십시오. 반론 내용:\n$REBUTTAL_TEXT")
            ADJ_JOB_ID=$(extract_job_id "$ADJ_OUTPUT")
            if [ -z "$ADJ_JOB_ID" ]; then
              log_error "Failed to register re-adjudication job for '$rt'."
              exit 1
            fi
            CREATED_JOBS+=("$ADJ_JOB_ID")
            if ! wait_for_job "$ADJ_JOB_ID"; then
              log_warn "Re-adjudication job for '$rt' failed; objection stands."
              OVERRULED_REVS+=("$rt")
              continue
            fi
            ADJ_REPORT=$(find_report "$ADJ_JOB_ID")
            # Fail-closed: anything that is not an explicit SUSTAINED keeps the
            # reviewer's objection alive.
            if [ -n "$ADJ_REPORT" ] && has_tag_line "$ADJ_REPORT" "ADJUDICATION: SUSTAINED"; then
              log_success "Reviewer '$rt' SUSTAINED the rebuttal; objection withdrawn."
            else
              log_warn "Reviewer '$rt' OVERRULED the rebuttal; objection stands."
              OVERRULED_REVS+=("$rt")
            fi
          done
          # Deadlock: the Creator rebutted and the reviewer held. In --plan mode
          # the Planner arbitrates; otherwise the reviewer prevails (fail-closed,
          # charter §1: a reviewer's PASS is the completion gate).
          if [ "${#OVERRULED_REVS[@]}" -gt 0 ]; then
            ARBITRATION="REVIEWER"
            if [ "$PLAN_MODE" = true ] && [ -n "$PLANNER_SESSION" ]; then
              log_info "Deadlock on ${OVERRULED_REVS[*]}; escalating to Planner for arbitration..."
              ARB_OUTPUT=$(delegate_job_safe submit \
                --agent-session "herdr:$PLANNER_SESSION" \
                --agent "$(resolve_agent_type "$PLANNER_SESSION")" \
                --type "direct" \
                --role "Planner" \
                --prompt "작업자(Creator)와 리뷰어의 의견이 충돌하여 교착 상태입니다. 계획서 관점에서 어느 쪽이 타당한지 재정(arbitration)하십시오. 리포트 마지막에 단독 행으로 '[ARBITRATION: CREATOR]' 또는 '[ARBITRATION: REVIEWER]' 를 명시하십시오. 리뷰어 지적:\n$FEEDBACK_AGGREGATE\n작업자 반론:\n$REBUTTAL_TEXT\n기존 계획서:\n$CURRENT_PLAN")
              ARB_JOB_ID=$(extract_job_id "$ARB_OUTPUT")
              if [ -z "$ARB_JOB_ID" ]; then
                log_error "Failed to register Planner arbitration job."
                exit 1
              fi
              CREATED_JOBS+=("$ARB_JOB_ID")
              if wait_for_job "$ARB_JOB_ID"; then
                ARB_REPORT=$(find_report "$ARB_JOB_ID")
                if [ -n "$ARB_REPORT" ] && has_tag_line "$ARB_REPORT" "ARBITRATION: CREATOR"; then
                  ARBITRATION="CREATOR"
                fi
              else
                log_warn "Planner arbitration job failed; defaulting to REVIEWER."
              fi
            fi
            if [ "$ARBITRATION" = "CREATOR" ]; then
              log_success "Planner ruled for the Creator; objections withdrawn for this round."
            else
              log_warn "Ruling stands with the reviewer(s); Creator must comply."
              # Close the channel for the REST OF THIS ITERATION and require
              # compliance. Note this does not persist into the next iteration
              # (which re-reviews and may raise different findings); the
              # run-wide cap REBUT_TOTAL_BUDGET is what bounds the whole run.
              REBUT_BUDGET=0
              COMPLY_OUTPUT=$(delegate_job_safe submit \
                --agent-session "herdr:$TARGET_AGENT" \
                --agent "$(resolve_agent_type "$TARGET_AGENT")" \
                --type "direct" \
                --role "Worker" \
                --prompt "귀하의 반론은 재심 결과 기각되었습니다. 더 이상 반론하지 말고 다음 리뷰 피드백을 그대로 반영하여 코드를 수정하십시오. 피드백:\n$FEEDBACK_AGGREGATE")
              COMPLY_JOB_ID=$(extract_job_id "$COMPLY_OUTPUT")
              if [ -z "$COMPLY_JOB_ID" ]; then
                log_error "Failed to register Creator compliance job."
                exit 1
              fi
              CREATED_JOBS+=("$COMPLY_JOB_ID")
              wait_for_job "$COMPLY_JOB_ID"
            fi
          fi
        fi
      fi
    fi
  fi
@@ -1,28 +1,30 @@
 ---
 name: multi-agent-mux-monitor
-description: "Run a long-lived Kanban worker that polls .mam/agent-sessions.yaml against the actual herdr/agent runtime state and reconciles them. Use when you want live visibility into which agent sessions are running, which are dead, which have stale YAML entries, and which have new session ids that haven't been recorded yet. Designed to be dispatched as a Kanban goal_mode task (--goal) so it keeps running until the user stops it."
+description: "Run a long-lived reconciler that watches .mam/agent-sessions.yaml against the actual herdr/agent runtime state and reconciles them. Use when you want live visibility into which agent sessions are running, which are dead, which have stale YAML entries, and which have new session ids that haven't been recorded yet. Runs as a persistent loop (`reconcile.sh --subscribe`) that keeps going until it times out, idles out, or is interrupted."
 version: 1.0.0
 author: godopu
 license: MIT
 platforms: [linux, macos]
-environments: [kanban, terminal, herdr]
+environments: [terminal, herdr]
 metadata:
  hermes:
-    tags: [agent, herdr, claude, antigravity, agy, monitor, kanban, observation, reconciliation]
+    tags: [agent, herdr, claude, antigravity, agy, monitor, observation, reconciliation]
-    related_skills: [multi-agent-mux-create, multi-agent-mux-resume, multi-agent-mux-stop, kanban-orchestrator]
+    related_skills: [multi-agent-mux-create, multi-agent-mux-resume, multi-agent-mux-stop, multi-agent-mux-status]
-    prereq_skills: [kanban-worker, multi-agent-mux-create]
+    prereq_skills: [multi-agent-mux-create]
 ---
-# Agent Sessions Monitor — Live Reconciliation via Kanban Worker
+# Agent Sessions Monitor — Live Reconciliation
 > **Companion skills**: `multi-agent-mux-create` / `multi-agent-mux-resume` / `multi-agent-mux-stop` (mutators); this skill is the **observer**.
 > **Single source of truth**: `./.mam/agent-sessions.yaml`.
 ## What this skill does
-Dispatch a **Kanban worker** (in `goal_mode`) that:
+Run a **long-lived reconciler** (`reconcile.sh --subscribe`) that:
-1. Every ~30s polls the actual state of:
+1. Reacts to delegated-job events on the MQTT broker, and — whenever the broker is
   unreachable — falls back to polling every `RECONCILE_POLL_INTERVAL` (default 15s)
   the actual state of:
   - `herdr agent list` (which sessions are alive)
   - `herdr agent get <session>` (pane cmd, cwd)
   - `~/.claude/projects/<workspace-key>/*.jsonl` mtime + first-line sessionId
@@ -33,10 +35,12 @@ Dispatch a **Kanban worker** (in `goal_mode`) that:
   - **yaml-only terminated/archived/stopped**: herdr dead, YAML says `terminated`, `archived`, or `stopped` → OK, left untouched (deliberate end states)
   - **yaml-only running, herdr dead**: YAML says `running`, herdr is gone → mark `terminated` with timestamp
   - **herdr-only running, not in YAML**: herdr session exists with `<workspace>-creator-*` naming but YAML doesn't know about it → register as a new entry
-   - **stale UUID**: YAML has a UUID, but the on-disk artifact is gone → flag in comment
+   - **stale UUID**: YAML has a UUID, but the on-disk artifact is gone → report it
-4. Writes a Kanban `kanban_comment` on every drift event with diff details
+4. Emits a JSON drift record on stdout for every drift event when run with `--emit-diff`
-5. Heartbeat every 5 minutes
+   (note: the `--subscribe` broker-down fallback runs each pass for its YAML side-effects
-6. **Goal loop**: judge (auxiliary model) re-checks the card after each turn against the body to decide "is monitoring still wanted?". When the user says "stop monitoring" via comment, the worker blocks with `reason=stop-requested`.
+   and discards the JSON — capture drift output with an explicit `--once --emit-diff`)
 5. Keeps running until one of its exit conditions fires: `--timeout` (wall-clock),
   `--idle-timeout` (no message received), or an interrupt from the operator.
 ## When to use
@@ -49,48 +53,33 @@ Dispatch a **Kanban worker** (in `goal_mode`) that:
 - One-off interactive session — just check `herdr agent list` and read the YAML
 - A single, short session — overhead > benefit
- You don't have a Kanban dispatcher running
+- You only need a point-in-time answer — use `multi-agent-mux-status` instead
-## Dispatching the monitor
+## Running the monitor
 ```bash
-# Goal-mode task: keeps running until the user signals stop
+# Persistent monitor: runs until interrupted; polls if the broker is unreachable.
-hermes kanban create \
+bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --subscribe --idle-timeout 0
  --title "agent-sessions monitor (live reconcile)" \
  --assignee default \
  --workspace worktree \
  --branch wt/multi-agent-mux-monitor \
  --goal \
  --goal-max-turns 100 \
  --max-runtime 8h \
  --max-retries 1 \
  --skill multi-agent-mux-monitor \
  --body "$(cat <<'EOF'
 You are the agent-sessions monitor. Every 30 seconds, do:
-1. Read .mam/agent-sessions.yaml
+# Bounded run: exits after 5 min with no message, or 1 h wall-clock, whichever comes first.
-2. Run `herdr agent list` and `herdr agent get <session>` for each tracked session name (these are real native herdr commands — do not use tmux-era names like `herdr ls`/`herdr list-panes` outside a shell that has sourced `.agents/skills/lib.sh`)
+bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --subscribe --idle-timeout 300 --timeout 3600
 3. For each session in the YAML, check the corresponding herdr state
 4. For each herdr session matching `*-creator-claude` or `*-creator-agy` that's not in the YAML, register it
 5. For any drift, call `kanban_comment` with the diff
 6. Sleep 30 seconds, then repeat
 If the user comments `stop` or `stop monitoring` on this card, call `kanban_block(reason="stop-requested by user")`.
 If you find that a Claude session's `claude_session_id_own` is null but there's a new *.jsonl in the project dir, read the sessionId from the first line and update the YAML.
 Use the helper script at .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh for the YAML updates — it handles all the merge logic and writes a structured comment to this card.
 EOF
 )"
 ```
 Run it under whatever supervisor you already use (a dedicated herdr pane, `nohup`,
 or a background job). Nothing else needs to be running for the monitor to work —
 it reconciles YAML ↔ herdr ↔ disk on its own.
 The herdr commands the script issues (`herdr agent list`, `herdr agent get <session>`)
 are real native herdr commands — do not substitute tmux-era names like `herdr ls` /
 `herdr list-panes` outside a shell that has sourced `.agents/skills/lib.sh`.
 ## Helper script: `reconcile.sh`
-The worker calls this script every 30s. It:
+This is the whole monitor — there is no separate driver. Each pass:
 1. Diffs YAML ↔ herdr ↔ disk artifacts
 2. Updates YAML if needed (only when changes are real, not on every poll — avoids spamming)
-3. Emits a JSON diff to stdout that the worker turns into a `kanban_comment`
+3. Emits a JSON diff to stdout for the caller to consume
 ```bash
 # Reconcile + auto-update YAML (atomic, flock-guarded). Emits JSON drift to stdout.
@@ -98,16 +87,9 @@ bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --once --emit-d
 # Read-only: compute drift WITHOUT writing the YAML (use for "what's running?" checks).
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --once --emit-diff --dry-run
 # Push-based MQTT Monitor: listen to delegated job events on the broker and update the YAML instantly.
 # Bounded run that exits after 5 min idle, or 1 h wall-clock; falls back to polling if the broker is down.
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --subscribe --idle-timeout 300 --timeout 3600
 # Persistent monitor (no timeouts): runs until interrupted; still polls if the broker is unreachable.
 bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --subscribe --idle-timeout 0
 ```
-Flags: `--once` (single pass), `--emit-diff` (print JSON), `--dry-run` (P1-E — no mutation), `--subscribe` (push-based MQTT subscription monitoring). `--subscribe` sub-flags: `--timeout N` (exit after N seconds of wall-clock; `0` = no limit, default), `--idle-timeout N` (exit after N seconds with no message; default `3600`, `0` = never idle-out). On a broker connection failure (connect error **or** non-zero CONNACK), `--subscribe` falls back to a polling loop that re-runs `--once --emit-diff` every `RECONCILE_POLL_INTERVAL` (default 15) seconds until `--timeout`. Terminal-event YAML updates are written through `lib.sh::atomic_dump_yaml` (flock + schema-validate + `.bak`). There are **no** `--workspace` / `--agent` / `--comment-card` flags; the worker turns the emitted JSON `drifts[]` into `kanban_comment` calls itself.
+Flags: `--once` (single pass), `--emit-diff` (print JSON), `--dry-run` (P1-E — no mutation), `--subscribe` (push-based MQTT subscription monitoring). `--subscribe` sub-flags: `--timeout N` (exit after N seconds of wall-clock; `0` = no limit, default), `--idle-timeout N` (exit after N seconds with no message; default `3600`, `0` = never idle-out). On a broker connection failure (connect error **or** non-zero CONNACK), `--subscribe` falls back to a polling loop that re-runs `--once --emit-diff` every `RECONCILE_POLL_INTERVAL` (default 15) seconds until `--timeout`. Terminal-event YAML updates are written through `lib.sh::atomic_dump_yaml` (flock + schema-validate + `.bak`). There are **no** `--workspace` / `--agent` flags; the emitted JSON `drifts[]` is the caller's to consume.
 ## Drift classes (what the script handles)
@@ -122,7 +104,7 @@ Any unstructured comments or reasons for the status change should be placed in `
 YAML:  status=running, pane.pid=201132, cmd=claude
 herdr:  no session
       → set status=terminated, terminated_at=<now>, termination_mode=auto-detected
-       → comment: "lab-landing-page-creator-claude: herdr gone (was pane 201132, cmd claude). Marked terminated."
+       → report: "lab-landing-page-creator-claude: herdr gone (was pane 201132, cmd claude). Marked terminated."
 ```
 **Skip-set**: the auto-terminate only fires for sessions whose status is `running`.
@@ -139,7 +121,7 @@ herdr:  session=lab-paper-pdf2md-creator-agy, pid=...,
       cmd=agy, cwd=$WORKSPACE_ROOT/paper-pdf2md
 YAML:  no such session
       → register as new entry: status=running, last_visible_status=running, last_visible_note=auto-registered
-       → comment: "lab-paper-pdf2md-creator-agy: herdr found but not in YAML. Auto-registered."
+       → report: "lab-paper-pdf2md-creator-agy: herdr found but not in YAML. Auto-registered."
 ```
 ### C. New session id materializes (claude first message sent)
@@ -149,7 +131,7 @@ YAML:  claude_session_id_own=null (placeholder)
 disk:  ~/.claude/projects/.../b3a7...c2f.jsonl exists, mtime=now,
       first line sessionId=b3a7...c2f
       → update claude_session_id_own=b3a7...c2f
-       → comment: "lab-landing-page-creator-claude: session id materialized b3a7...c2f"
+       → report: "lab-landing-page-creator-claude: session id materialized b3a7...c2f"
 ```
 ### D. Stale UUID (artifact gone)
@@ -157,48 +139,47 @@ disk:  ~/.claude/projects/.../b3a7...c2f.jsonl exists, mtime=now,
 ```
 YAML:  agent_identities.claude.session_id=87dc548e-...
 disk:  ~/.claude/projects/.../87dc548e-...jsonl: missing
-       → flag in comment, but DO NOT delete from YAML
+       → report it, but DO NOT delete from YAML
       (the user may have moved the file or the disk may be temporarily unavailable;
        only `--purge-conversation` should remove the id)
 ```
 ## Pitfalls
- **Don't run the monitor without `--goal`** — without goal mode, a single turn will spawn, do one reconcile, and complete. Goal mode keeps the worker alive across many turns.
+- **Don't expect `--once` to stay alive** — it does a single pass and exits. Use `--subscribe` for continuous monitoring.
- **The 30s poll is a default** — workers may override if they detect heavy churn. A workspace with 5+ agent sessions should bump to 60s to avoid noise.
+- **`--idle-timeout` defaults to 3600s** — a monitor meant to run indefinitely needs `--idle-timeout 0` explicitly, or it will quietly exit after an hour of broker silence.
- **`kanban_comment` rate limits** — Kanban may throttle if you comment too fast. Coalesce: only comment when the diff is *new* (not the same drift on every poll). The script tracks a state file at `.cache/multi-agent-mux-monitor/<workspace>.state` in the workspace root for this (overridable via `AGENT_SESSIONS_STATE_DIR`).
+- **The poll interval is a default** — `RECONCILE_POLL_INTERVAL` (15s) is what the broker-down fallback uses. A workspace with 5+ agent sessions can bump it to reduce noise.
 - **Coalesce repeated drifts** — the same drift re-appears on every pass until it is resolved. A caller that acts on `drifts[]` should compare against the previous pass and act only on *new* entries; the script does not deduplicate for you.
 - **Don't fight the user's explicit action** — if `multi-agent-mux-stop` is mid-flight and the monitor sees the same session in two states within 5s, prefer the user's most recent action. The monitor should not auto-revert a fresh `terminated` to `running` because of a stale `herdr has-session` check.
- **The monitor should never modify the conversation artifacts** (jsonl, db) — only the YAML. If you see a stale UUID, comment about it but don't delete the file.
+- **The monitor should never modify the conversation artifacts** (jsonl, db) — only the YAML. If you see a stale UUID, report it but don't delete the file.
 - **TUI capture-pane is expensive** — only capture when you need to update `last_visible_status`, not every poll.
-## Worker body template (for `hermes kanban create --body`)
+## Supervising-agent runbook
-The `--body` of the dispatched task IS the worker's behavior spec. Here's a tested template:
+If an agent drives the monitor rather than an operator watching it directly, this is
 the behavior spec:
 ```markdown
 # agent-sessions monitor
-## Loop (every 30s)
+## Loop
 1. Read agent-sessions.yaml
-2. Bash: `bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --emit-diff`
+2. Bash: `bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --once --emit-diff`
 3. Parse the JSON diff from stdout
-4. If `drifts` is non-empty:
+4. If `drifts` is non-empty, report each *new* drift to the operator
-   - For each drift, call `kanban_comment` with the diff message
+5. Bash: `sleep 30`, then repeat
 5. Bash: `sleep 30`
 6. Heartbeat every 5 min: `kanban_heartbeat(progress="alive, N drifts detected, last at <time>")`
 ## Stop condition
-If `$HERMES_KANBAN_TASK` card has any comment containing "stop" or "stop monitoring" from a user:
+Stop when the operator says to stop, or when the surrounding job's timeout fires.
 - Call `kanban_block(reason="stop-requested by user at <timestamp>")`
 ## Drift responses
- A. herdr dead + YAML running: auto-terminate YAML, comment
+- A. herdr dead + YAML running: auto-terminate YAML, report
- B. herdr alive not in YAML: auto-register, comment
+- B. herdr alive not in YAML: auto-register, report
- C. New session id from *.jsonl: update YAML, comment
+- C. New session id from *.jsonl: update YAML, report
- D. Stale UUID: comment only, no YAML change
+- D. Stale UUID: report only, no YAML change
 ## Hard rules
@@ -233,5 +214,5 @@ bash .agents/skills/multi-agent-mux-monitor/scripts/reconcile.sh --emit-diff --o
 ## Related skills
- `kanban-worker` — base lifecycle for the dispatched worker
+- `multi-agent-mux-status` — read-only snapshot when you don't need a running loop
- `kanban-orchestrator` — if you want to dispatch this monitor *from* an orchestrator, use this to know how to phrase the body
+- `multi-agent-mux-delegate-job` — the MQTT job channel whose events `--subscribe` listens to
@@ -1,6 +1,6 @@
 ---
 name: multi-agent-mux-status
-description: "Read-only instant snapshot of all agent herdr sessions — name, YAML status, herdr alive, pane cmd/cwd, resume UUID on disk, and any drift. No Kanban, no mutation. Reuses reconcile.sh --dry-run for the diff logic. Use when you want to know 'what's running RIGHT NOW' without spinning up a Kanban monitor worker."
+description: "Read-only instant snapshot of all agent herdr sessions — name, YAML status, herdr alive, pane cmd/cwd, resume UUID on disk, and any drift. No mutation. Reuses reconcile.sh --dry-run for the diff logic. Use when you want to know 'what's running RIGHT NOW' without spinning up the monitor loop."
 version: 1.0.0
 author: godopu
 license: MIT
@@ -21,7 +21,7 @@ metadata:
 ## What this skill does
-Print a single table of every agent herdr session, comparing YAML state to actual herdr state. **No mutation. No Kanban. No polling loop.**
+Print a single table of every agent herdr session, comparing YAML state to actual herdr state. **No mutation. No polling loop.**
 This is the "what's running right now?" answer — faster than dispatching `multi-agent-mux-monitor` (which polls every 30s) and safer than `reconcile.sh --once --emit-diff` (which mutates as a side effect).
@@ -105,7 +105,7 @@ lab-paper-pdf2md-creator-claude              default      running    alive  clau
 ## Pitfalls
- **Do NOT use this skill to drive mutations** — the output is a snapshot, not a call to action. If you need to fix drifts, dispatch `multi-agent-mux-monitor` (Kanban worker) or run `multi-agent-mux-resume` / `multi-agent-mux-stop` manually.
+- **Do NOT use this skill to drive mutations** — the output is a snapshot, not a call to action. If you need to fix drifts, run `multi-agent-mux-monitor` or `multi-agent-mux-resume` / `multi-agent-mux-stop` manually.
 - **Read-only is enforced by script** — `status.sh` opens the YAML with `open(path)` (no `'w'`), never calls `herdr kill-session`, never writes anywhere. The `reconcile.sh --dry-run` mode is the same path.
 - **If `agent-sessions.yaml` is malformed** — print the YAML error verbatim and exit 1. Do NOT attempt recovery (that's `multi-agent-mux-stop --purge-conversation` or manual edit's job).
 - **Sessions outside the `<workspace>-creator-*` naming convention** are still shown but tagged `ad-hoc` — they didn't go through `multi-agent-mux-create` and aren't tracked in YAML.
@@ -115,10 +115,10 @@ lab-paper-pdf2md-creator-claude              default      running    alive  clau
 - "Is the claude session still running?" → this skill, not the monitor
 - "What UUID does this workspace have?" → this skill
 - "Is there drift between YAML and reality?" → this skill, then dispatch monitor or fix manually
- Quick sanity check before dispatching a long Kanban task
+- Quick sanity check before dispatching a long-running job
 ## When NOT to use
- Continuous live tracking → `multi-agent-mux-monitor` (Kanban worker)
+- Continuous live tracking → `multi-agent-mux-monitor`
 - Recovering from corruption → manual edit + `.bak` restore
 - Polling more than once a minute → `multi-agent-mux-monitor` (it dedupes)
@@ -1,9 +1,9 @@
 # 🛠️ Multi-Agent Mux 종합 개선 및 미해결 과제 백로그 (`IMPROVEMENTS.md`)
- **최종 갱신일**: 2026-08-05 (B-1 해결 및 완료 항목 섹션 분리 반영)
+- **최종 갱신일**: 2026-08-06 (O-1 Rebuttal & Re-adjudication Protocol 완료 반영)
 - **통합 관리 대상**: 기존 `CODEBASE_REVIEW_REPORT.md` + `OPTIMIZATION.md`
- **총 추적 미해결 과제**: **18건** (아키텍처 2건, 엣지케이스 8건, 오케스트레이션 3건, 레거시 잔재 5건)
+- **총 추적 미해결 과제**: **16건** (아키텍처 2건, 엣지케이스 8건, 오케스트레이션 2건, 레거시 잔재 4건)
- **완료된 과제**: **3건** (A-1, A-5, B-1)
+- **완료된 과제**: **5건** (A-1, A-5, B-1, C-1, O-1)
 ---
@@ -53,12 +53,7 @@
 ---
-## 3. 🟡 오케스트레이션 최적화 과제 (Orchestration Optimizations — 3건)
+## 3. 🟡 오케스트레이션 최적화 과제 (Orchestration Optimizations — 2건)
 ### **O-1 (구 ISSUE-6): 타당하지 않은 리뷰 피드백 거부/반론 프로토콜 미지원**
 - **현상**: `MULTI_AGENT_RULES.md` 1장 규약에는 "개발 팀장이 리뷰어의 타당하지 않은 피드백을 거부하고 명확한 이유를 회신할 수 있다"고 명시되어 있음.
 - **문제점**: `run_loop.sh`는 리뷰어의 `NOT PASS` 피드백 전체를 Creator에게 일방적으로 주입할 뿐, Creator가 특정 피드백을 거부하거나 반론을 제기하여 상호 조율하는 이의제기 채널이 코딩적으로 구현되어 있지 않음.
 - **해결 방안**: Creator 교정 단계 프롬프트에 반론 작성 템플릿을 허용하고, 반론 발생 시 Planner/Reviewer에게 재검토를 요청하는 이의제기 브랜칭 로직 설계.
 ### **O-2 (구 ISSUE-7): 동일 워크스페이스 내 중복 루프 기동 방지 락 (Race-Free Lock)**
 - **현상**: 동일 작업 트리에서 다수의 `run_loop.sh` 스크립트가 병렬 기동될 경우 SQLite DB 갱신 경합 및 YAML 데이터 오염이 일어날 수 있음.
@@ -76,10 +71,7 @@
 ---
-## 4. ⚪ 레거시 잔재 및 죽은 코드 (Legacy Remnants — 5건)
+## 4. ⚪ 레거시 잔재 및 죽은 코드 (Legacy Remnants — 4건)
 ### **C-1: Kanban 문서 29회 언급 vs 실제 구현 0건**
 - SKILL.md 파일들에 Kanban 지원 및 상태 파일 서술이 29회 언급되어 있으나 스크립트 구현은 0건입니다.
 ### **C-2: 미사용 `.cache/` 상태 디렉터리 생성**
 - `reconcile.sh`가 `.cache/multi-agent-mux-monitor` 디렉터리를 `mkdir`만 하고 아무것도 읽거나 쓰지 않습니다.
@@ -95,7 +87,7 @@
 ---
-## 5. 🎉 완료된 과제 (Completed Tasks — 3건)
+## 5. 🎉 완료된 과제 (Completed Tasks — 5건)
 ### **A-1: 워크스페이스 세션 격리 & drift-B 오등록 방지** — ✅ 완료
 - `derive_workspace_slug` 헬퍼 함수를 추가하여 워크스페이스 경로 기반 단일 소켓 슬러그(`mam-<parent>-<work>`) 도출 체계를 구축했습니다.
@@ -108,8 +100,21 @@
 - `lib.sh` tier-3 신원 캐시 조회 시 미정의 변수(`db_path`, `yaml_path`) 및 `yaml` import 누락으로 무조건 `NameError` 예외가 발생하던 결함을 해결했습니다.
 - DB를 1차 권위 경로로, `$YAML_PATH`를 폴백으로 정제하고 전용 회귀 테스트 `tests/test_b1_tier3_identity.py` (8/8 PASS)를 작성하여 입증했습니다.
 ### **C-1: Kanban 문서 29회 언급 vs 실제 구현 0건** — ✅ 완료
 - SKILL.md 3종(monitor 22 / status 5 / create 2)과 README 2종의 Kanban 서술을 전면 제거했습니다.
 - `multi-agent-mux-monitor` 의 실행 메커니즘 서술을 실제 구현인 `reconcile.sh --subscribe` (MQTT push + 브로커 다운 시 폴링 폴백) 기준으로 재작성했습니다.
 - 존재하지 않는 스킬 참조 2건(`kanban-worker`, `kanban-orchestrator`)을 실존 스킬로 교체했습니다.
 - `hermes kanban create` CLI 플래그 잔재 10종(`--goal-max-turns`, `--assignee`, `--comment-card` 등)을 파생형 검증 게이트(G-C)로 차단했습니다.
 - 제품 표면(`.agents/skills/`, `README*.md`) Kanban 참조 **0건** 확인.
 ### **O-1: 타당하지 않은 리뷰 피드백 거부/반론 프로토콜 미지원 (Rebuttal Protocol)** — ✅ 완료
 - `run_loop.sh`에 `--max-rebut N` (기본값 1) 옵션 및 Rebuttal/Re-adjudication/Arbitration 3단계 프로토콜을 구현했습니다.
 - Creator가 지적 항목 거부 시 `[REBUT: <reviewer>]` 태그를 남겨 해당 리뷰어 대상 재심(`[ADJUDICATION: SUSTAINED/OVERRULED]`)을 가동하며, 교착 시 Planner 재정(`[ARBITRATION: CREATOR/REVIEWER]`) 또는 Fail-Closed 결정을 수행합니다.
 - bash 3.2 macOS 규격 빈 배열 확장 안전성(`${ARR[@]+"${ARR[@]}"}`) 및 per-iteration budget reset / total budget cap 결함을 완벽히 보완하고 회귀 테스트 `tests/test_o1_rebuttal.py` (10/10 PASS)로 입증했습니다.
 - `MULTI_AGENT_RULES.md`, `.ko.md`, `multi-agent-mux-loop/SKILL.md` 문서 연동을 완료했습니다.
 ---
 ## 6. 결론 및 향후 보완 로드맵
-`IMPROVEMENTS.md` 문서에 따라 향후 코드베이스 개편 시 남은 백로그 항목(아키텍처 2건, 엣지케이스 8건, 오케스트레이션 3건, 레거시 잔재 5건)을 일원화된 보완 로드맵으로 관리합니다.
+`IMPROVEMENTS.md` 문서에 따라 향후 코드베이스 개편 시 남은 백로그 항목(아키텍처 2건, 엣지케이스 8건, 오케스트레이션 2건, 레거시 잔재 4건)을 일원화된 보완 로드맵으로 관리합니다.
@@ -24,7 +24,7 @@ Herdr와 MQTT 브로커를 기반으로 구축된 고신뢰성 **다중 에이
 *   **`multi-agent-mux-stop`**: 에이전트 CLI 세션을 정상 종료 키 입력(`/exit` 또는 `Exit`)을 통해 안전하게 닫고, 격리된 대화 히스토리 및 데이터베이스 로그를 삭제(purge)하는 클린업 작업을 수행합니다.
 *   **`multi-agent-mux-resume`**: 디스크 또는 캐시에서 특정 워크스페이스의 세션 UUID를 조회하여 기존 대화 상태(`claude -r <uuid>` 또는 `hermes --resume <uuid>`) 그대로 세션을 복구하고 재개합니다.
 *   **`multi-agent-mux-status`**: 활성화된 모든 세션의 실시간 작동 상태를 쿼리하여 PID 정합성, 실행 명령 포맷, herdr 실제 상태와 데이터베이스 간의 동기화 드리프트를 감지합니다.
-*   **`multi-agent-mux-monitor`**: 백그라운드에서 Kanban Reconcile 프로세스로 실행되어, 실시간 herdr 세션 변화를 모니터링하고 `.mam/agent-sessions.yaml` 메타데이터 파일에 상태를 동기화합니다.
+*   **`multi-agent-mux-monitor`**: 백그라운드 Reconcile 프로세스로 실행되어, 실시간 herdr 세션 변화를 모니터링하고 `.mam/agent-sessions.yaml` 메타데이터 파일에 상태를 동기화합니다.
 *   **`multi-agent-mux-delegate-job`**: 태스크를 비동기식 독립 잡으로 위임 및 관리하는 핵심 모듈입니다:
    *   `registry.py`: 파일 락(`fcntl`)을 활용해 경쟁 조건 없이 잡을 원자적으로 등록 및 점유(claim)합니다.
    *   `job_subscriber.py`: MQTT 백플레인 채널을 구독하여 실시간 상태 이벤트를 수집하고 이를 감사 로그(audit trail)에 기록합니다.
@@ -42,7 +42,7 @@ All orchestration functionalities are structured under the `.agents/skills/` dir
 *   **`multi-agent-mux-stop`**: Gracefully terminates agent CLI sessions (using key macros like `/exit` or `Exit`) and handles disk purge operations (removing conversation JSON files and SQLite logs for deleted workspaces).
 *   **`multi-agent-mux-resume`**: Restores stopped sessions by resolving workspace UUIDs from disk or cache, and invokes the underlying agent using session-resume parameters (e.g., `claude -r <uuid>` or `hermes --resume <uuid>`).
 *   **`multi-agent-mux-status`**: Queries the running states of all active sessions, detecting PID mismatches, command signatures, and drifts between actual herdr instances and the registry database.
-*   **`multi-agent-mux-monitor`**: A long-running Kanban reconcile worker that dynamically monitors herdr sessions and synchronizes states to `.mam/agent-sessions.yaml`.
+*   **`multi-agent-mux-monitor`**: A long-running reconcile worker that dynamically monitors herdr sessions and synchronizes states to `.mam/agent-sessions.yaml`.
 *   **`multi-agent-mux-delegate-job`**: The core asynchronous task distribution module containing:
    *   `registry.py`: Atomically registers and claims jobs using file advisory locks (`fcntl`).
    *   `job_subscriber.py`: Connects to the MQTT backplane, captures live events, and appends them to audit trails.
@@ -0,0 +1,176 @@
 #!/usr/bin/env python3
 """
 tests/test_o1_rebuttal.py — O-1 Rebuttal & Re-adjudication Protocol regression suite (V-1..V-10).
 """
 import os
 import subprocess
 import pytest
 REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 RUN_LOOP_SH = os.path.join(REPO_ROOT, ".agents", "skills", "multi-agent-mux-loop", "scripts", "run_loop.sh")
 RULES_MD = os.path.join(REPO_ROOT, ".agents", "MULTI_AGENT_RULES.md")
 RULES_KO_MD = os.path.join(REPO_ROOT, ".agents", "MULTI_AGENT_RULES.ko.md")
 SKILL_MD = os.path.join(REPO_ROOT, ".agents", "skills", "multi-agent-mux-loop", "SKILL.md")
 IMPROVEMENTS_MD = os.path.join(REPO_ROOT, "IMPROVEMENTS.md")
 def _get_run_loop_helpers():
    """Extract helper functions from run_loop.sh to run standalone in bash tests."""
    with open(RUN_LOOP_SH, "r", encoding="utf-8") as f:
        content = f.read()
    # Extract has_tag_line and extract_tag_values definitions
    has_tag_line_code = content.split("has_tag_line() {")[1].split("extract_tag_values() {")[0]
    extract_tag_values_code = content.split("extract_tag_values() {")[1].split("find_report() {")[0]
    return f"""
 has_tag_line() {{
 {has_tag_line_code}
 extract_tag_values() {{
 {extract_tag_values_code}
 """
 def test_v1_has_tag_line_single_and_first_line(tmp_path):
    """V-1: has_tag_line matches tag on line 1 of report file."""
    report_file = tmp_path / "report-final.md"
    report_file.write_text("[ESCALATE: PLANNER]\nDetailed feedback follows...")
    helpers = _get_run_loop_helpers()
    cmd = f"""
 {helpers}
 has_tag_line "{report_file}" "ESCALATE: PLANNER"
 """
    res = subprocess.run(["bash", "-c", cmd], capture_output=True, text=True)
    assert res.returncode == 0
 def test_v2_extract_tag_values_multiple_rebuts(tmp_path):
    """V-2: extract_tag_values extracts all [REBUT: target] tags."""
    report_file = tmp_path / "report-final.md"
    report_file.write_text("[REBUT: reviewer_a]\n[REBUT: reviewer_b]\nContent...")
    helpers = _get_run_loop_helpers()
    cmd = f"""
 {helpers}
 extract_tag_values "{report_file}" "REBUT"
 """
    res = subprocess.run(["bash", "-c", cmd], capture_output=True, text=True)
    assert res.returncode == 0
    lines = [line.strip() for line in res.stdout.splitlines() if line.strip()]
    assert lines == ["reviewer_a", "reviewer_b"]
 def test_v3_rebut_option_parsing_and_budget():
    """V-3: --max-rebut N option parsing and error handling."""
    # Check help
    res_help = subprocess.run(["bash", RUN_LOOP_SH, "--help"], capture_output=True, text=True)
    assert "--max-rebut N" in res_help.stdout or "--max-rebut N" in res_help.stderr
    # Check invalid value
    res_invalid = subprocess.run(["bash", RUN_LOOP_SH, "--max-rebut", "abc"], capture_output=True, text=True)
    assert res_invalid.returncode != 0
    assert "ERROR: --max-rebut requires a non-negative integer." in res_invalid.stdout or "ERROR: --max-rebut requires a non-negative integer." in res_invalid.stderr
 def test_v4_rebuttal_deduplication():
    """V-4: REBUT_TARGETS deduplication logic."""
    cmd = """
    FAIL_REVS=("rev1" "rev2")
    REBUT_TARGETS=()
    INPUT_TAGS=("rev1" "rev1" "rev2" "rev3")
    for _t in "${INPUT_TAGS[@]}"; do
        for _fv in "${FAIL_REVS[@]}"; do
            [ "$_t" = "$_fv" ] || continue
            _dup=0
            for _e in ${REBUT_TARGETS[@]+"${REBUT_TARGETS[@]}"}; do
                if [ "$_e" = "$_t" ]; then _dup=1; break; fi
            done
            [ "$_dup" -eq 0 ] && REBUT_TARGETS+=("$_t")
            break
        done
    done
    echo "${REBUT_TARGETS[*]}"
    """
    res = subprocess.run(["bash", "-c", cmd], capture_output=True, text=True)
    assert res.returncode == 0
    assert res.stdout.strip() == "rev1 rev2"
 def test_v5_fail_closed_on_missing_sustained(tmp_path):
    """V-5: Non-SUSTAINED adjudication report fails closed (objection stands)."""
    report_overruled = tmp_path / "report_overruled.md"
    report_overruled.write_text("[ADJUDICATION: OVERRULED]\nObjection maintained.")
    report_empty = tmp_path / "report_empty.md"
    report_empty.write_text("No explicit adjudication tag present.")
    helpers = _get_run_loop_helpers()
    cmd = f"""
 {helpers}
 if has_tag_line "{report_overruled}" "ADJUDICATION: SUSTAINED"; then echo "SUSTAINED"; else echo "OVERRULED"; fi
 if has_tag_line "{report_empty}" "ADJUDICATION: SUSTAINED"; then echo "SUSTAINED"; else echo "OVERRULED"; fi
 """
    res = subprocess.run(["bash", "-c", cmd], capture_output=True, text=True)
    assert res.returncode == 0
    lines = res.stdout.splitlines()
    assert lines[0] == "OVERRULED"
    assert lines[1] == "OVERRULED"
 def test_v6_bash_3_2_empty_array_safety():
    """V-6: Empty array expansion with set -u on bash 3.2 syntax."""
    cmd = """
    set -euo pipefail
    declare -a EMPTY_ARR=()
    for item in ${EMPTY_ARR[@]+"${EMPTY_ARR[@]}"}; do
        echo "$item"
    done
    echo "SUCCESS"
    """
    res = subprocess.run(["bash", "-c", cmd], capture_output=True, text=True)
    assert res.returncode == 0
    assert "SUCCESS" in res.stdout
 def test_v7_rules_docs_contain_section_3_1():
    """V-7: MULTI_AGENT_RULES.md and MULTI_AGENT_RULES.ko.md contain section 3.1."""
    with open(RULES_MD, "r", encoding="utf-8") as f:
        content_en = f.read()
    with open(RULES_KO_MD, "r", encoding="utf-8") as f:
        content_ko = f.read()
    assert "### 3.1 Rebuttal & Adjudication Protocol" in content_en
    assert "### 3.1 Rebuttal & Adjudication Protocol" in content_ko
    assert "[REBUT:" in content_en and "[ADJUDICATION: SUSTAINED]" in content_en
    assert "[REBUT:" in content_ko and "[ADJUDICATION: SUSTAINED]" in content_ko
 def test_v8_skill_md_contains_max_rebut():
    """V-8: multi-agent-mux-loop/SKILL.md contains --max-rebut documentation."""
    with open(SKILL_MD, "r", encoding="utf-8") as f:
        content = f.read()
    assert "`--max-rebut`" in content
    assert "Rebuttal & Adjudication Protocol" in content
 def test_v9_per_iteration_budget_reset():
    """V-9: run_loop.sh resets REBUT_BUDGET per iteration while tracking REBUT_TOTAL_BUDGET."""
    with open(RUN_LOOP_SH, "r", encoding="utf-8") as f:
        content = f.read()
    assert 'REBUT_BUDGET="$MAX_REBUT"' in content
    assert "REBUT_TOTAL_BUDGET=$((MAX_REBUT * MAX_LOOP))" in content
    assert "REBUT_TOTAL_BUDGET=$((REBUT_TOTAL_BUDGET - 1))" in content
 def test_v10_improvements_md_o1_completed():
    """V-10: IMPROVEMENTS.md has O-1 moved to completed tasks section."""
    with open(IMPROVEMENTS_MD, "r", encoding="utf-8") as f:
        content = f.read()
    assert "O-1" in content
    assert "[O-1] Rebuttal & Re-adjudication Protocol" in content or "O-1: 타당하지 않은 리뷰 피드백 거부/반론 프로토콜 미지원" in content
Author	SHA1	Message	Date
Godopu	cdeb9e2c8f	feat(o1): implement reviewer feedback rebuttal and re-adjudication protocol (100% PASS)	2026-08-06 09:21:28 +09:00
Godopu	2d5e20a6aa	docs(c1): purge all 29 legacy Kanban references from SKILL.md and documentation (100% PASS)	2026-08-06 08:03:15 +09:00