feat(o3): implement Invocation-Aware Scoped Guard for orchestrator role scoping (100% PASS)

2026-08-07 10:09:14 +09:00
parent 5ab76874e4
commit 1f8622e077
13 changed files with 428 additions and 20 deletions
@@ -150,6 +150,16 @@ TMUX 환경에서 실행되는 에이전트가 화면 스크롤 한계로 인해
    - *버전 관리 이관*: 버전 관리가 필요한 주요 산출물(최종 설계 계획, 최종 리뷰 보고서, 보안 감사 리포트 등)은 gitignore 대상인 `.mam/` 하위가 아닌, 버전 관리 대상 경로(구체적으로 `.agents/reports/<tmux_session_name>/` 또는 `docs/reports/` 등)로 명시적으로 복사하여 이관 보존해야 합니다.
  - **디스크 정리 및 보존 정책 계약 (Cleanup & Retention)**: `.mam/jobs/<job_id>/` 및 `.mam/reports/` 폴더 아래의 파일들은 휘발성 감사 이력(audit-trail) 산출물입니다. 버전 관리가 필요한 문서들은 `.agents/reports/` 하위로 수동 복사하여 커밋해야 하며, `stop_session.sh` 세션 종료 스크립트는 이들 보고서 디렉터리를 자동으로 삭제하지 않으므로 수동 또는 주기적 클린업이 권장됩니다.

+### 3.2 조건부 오케스트레이션 위임 가드 (Invocation-Aware Scoped Guard — O-3)
+
+| 모드 | 오케스트레이터 행위 | 도구 허용 여부 |
+|---|---|---|
+| **일반 모드 (Normal Mode)** | 주 작업자 (직접 코드 및 문서 수정) | 모든 파일 수정 도구 허용 |
+| **루프 활성 모드 (`/multi-agent-mux-loop`)** | 오케스트레이터 (`run_loop.sh` 자율 위임) | `file_change`, `edit_notebook`, `write_blob` **하드 블록** (`.agents/hooks.json`) |
+
+- **Fail-Open 원칙**: 훅 내부 오류 또는 파싱 실패 시 무조건 `allow`로 처리하여 작업을 차단하지 않음.
+- **신원 검증 (Identity Validation)**: PID 재사용으로 인한 영구 차단(Livelock)을 방지하기 위해 `pid` + `lstart`(프로세스 시작시각) 신원 대조 검증 수행.
+
 ### ⏱️ 타임아웃 구성 및 정렬 규칙
 - **잡 실행 제한 (`timeout_sec` & `idle_timeout_sec`)**: 각 잡은 전체 실행 만료 시간(`timeout_sec`, 기본 3600s)과 메세지 미수신 유휴 시간(`idle_timeout_sec`, 기본 120s)을 독립적으로 가집니다.
 - **모니터 유휴 대기 (`SUB_IDLE_TIMEOUT`)**: 모니터 스크립트(`reconcile.sh`)의 유휴 대기 시간(`SUB_IDLE_TIMEOUT`) 기본값은 잡 최대 예산에 맞춰 `3600s`(1시간) 이상으로 항상 넉넉히 설정해야 합니다. 모니터가 작업 완료 전에 유휴 감지로 조기 자동 종료되어 백그라운드 태스크 관리를 소실하는 문제를 방지하기 위함입니다.
@@ -150,6 +150,16 @@ To ensure that agents running in TMUX environments do not lose debug logs or pre
    - *Versioned promotions*: Any final design plans, review verdicts, or security audit reports that require version control must be explicitly copied to tracked directory paths (specifically under `.agents/reports/<tmux_session_name>/` or `docs/reports/`).
  - **Cleanup & Retention Contract**: Files under `.mam/jobs/<job_id>/` and `.mam/reports/` are transient audit-trail artifacts. While durable outcomes are committed to version control under `.agents/reports/`, ephemeral directory trees can be cleaned up manually as needed; `stop_session.sh` does not automatically purge these report trees during session exit.

+### 3.2 Invocation-Aware Scoped Guard (O-3)
+
+| Mode | Orchestrator Action | Tool Access |
+|---|---|---|
+| **Normal Mode** | Main Creator (Direct implementation) | All tools allowed |
+| **Loop Active Mode (`/multi-agent-mux-loop`)** | Orchestrator (Delegates to `run_loop.sh`) | `file_change`, `edit_notebook`, `write_blob` **hard-blocked** via `.agents/hooks.json` |
+
+- **Fail-Open Policy**: Any hook internal error or parse error evaluates to `allow`.
+- **Identity Verification**: The guard validates process liveness via `pid` + `lstart` to prevent livelocks on PID rollover.
+
 ### ⏱️ Timeout Configuration & Alignment Rules
 - **Job Execution Limits (`timeout_sec` & `idle_timeout_sec`)**: Each job independently manages its overall execution timeout (`timeout_sec`, default 3600s) and idle timeout without receiving messages (`idle_timeout_sec`, default 120s).
 - **Monitor Idle Waiting (`SUB_IDLE_TIMEOUT`)**: The idle timeout for the monitor script (`reconcile.sh`), `SUB_IDLE_TIMEOUT`, must always be set generously to `3600s` (1 hour) or more to align with the maximum job budget. This prevents the monitor from terminating early due to idle detection, which would lose control over background tasks before they finish.
@@ -0,0 +1,12 @@
+{
+  "mam-loop-delegation-guard": {
+    "PreToolUse": [
+      {
+        "matcher": "file_change|edit_notebook|write_blob",
+        "hooks": [
+          { "type": "command", "command": "./hooks/loop_delegation_guard.sh", "timeout": 10 }
+        ]
+      }
+    ]
+  }
+}
@@ -0,0 +1,136 @@
+#!/usr/bin/env bash
+# O-3 — Invocation-Aware Scoped Guard (Rev.2).
+#
+# Normal mode: the orchestrator IS the Main Creator and may edit files freely.
+# While /multi-agent-mux-loop is active it must delegate through run_loop.sh
+# instead, so this PreToolUse hook denies direct file mutation and says why.
+#
+# Contract (agy hooks.json): JSON payload on stdin, JSON decision on stdout.
+#   in : {"toolCall":{"name":..., "args":{...}}, "workspacePaths":[...],
+#         "transcriptPath":"..."}
+#   out: {"decision":"allow"|"deny", "reason":"..."}
+#
+# Fails OPEN: any internal error emits `allow`. A guard that blocks the agent
+# because it could not parse its own input would be worse than the drift.
+set -uo pipefail
+
+payload="$(cat)"
+
+exec 3>&1                       # keep the decision channel separate from noise
+allow() { printf '{"decision":"allow"}\n' >&3; exit 0; }
+
+MARKER="${MAM_LOOP_GUARD_MARKER:-}"
+
+python3 - "$payload" "$MARKER" >&3 <<'PY' || allow
+import json, os, sys, subprocess
+
+payload_raw, marker_override = sys.argv[1], sys.argv[2]
+
+def emit(decision, reason=None):
+    out = {"decision": decision}
+    if reason:
+        out["reason"] = reason
+    print(json.dumps(out))
+    sys.exit(0)
+
+try:
+    p = json.loads(payload_raw)
+except Exception:
+    emit("allow")                                    # unparseable -> fail open
+
+name = ((p.get("toolCall") or {}).get("name") or "").strip().lower()
+
+# Step-type-derived names (hooks.json matches on these), NOT the model-facing
+# tool names. This agy build has CORTEX_STEP_TYPE_FILE_CHANGE / EDIT_NOTEBOOK /
+# WRITE_BLOB; there is no REPLACE_FILE_CONTENT step type at all.
+MUTATING = {"file_change", "edit_notebook", "write_blob"}
+if name not in MUTATING:
+    emit("allow")
+
+ws = (p.get("workspacePaths") or [None])[0] or os.getcwd()
+marker = marker_override or os.path.join(ws, ".mam", "loop-guard-active")
+
+def _lstart(pid):
+    """Process start time, or '' if the process is gone/unknowable."""
+    try:
+        out = subprocess.run(["ps", "-p", str(pid), "-o", "lstart="],
+                             capture_output=True, text=True, timeout=5)
+    except Exception:
+        return ""
+    return " ".join(out.stdout.split())
+
+def _marker_active(path):
+    """True only if the marker exists AND its owning process is still alive.
+
+    SIGKILL cannot be trapped, so a trap-based release always has a leak
+    window. A stale marker must never block the orchestrator forever, so
+    identity (pid + lstart) -- not mere existence -- is the signal.
+    """
+    if not os.path.exists(path):
+        return False
+    try:
+        with open(path, encoding="utf-8", errors="replace") as f:
+            txt = f.read()
+    except Exception:
+        return False
+    fields = {}
+    for line in txt.splitlines():
+        if "=" in line:
+            k, v = line.split("=", 1)
+            fields[k.strip()] = v.strip()
+    
+    try:
+        pid = int(fields.get("pid", ""))
+    except ValueError:
+        pid = None
+
+    if pid is None:
+        return True                                  # no pid recorded -> honour it
+
+    recorded_lstart = " ".join(fields.get("lstart", "").split())
+    if recorded_lstart:
+        # pid + start time is a stable identity. A reused pid always has a
+        # different start time, so this closes the rollover livelock: a marker
+        # we cannot positively identify must never block the orchestrator.
+        return _lstart(pid) == recorded_lstart
+
+    # Legacy marker with no lstart: fall back to liveness, but treat an
+    # unidentifiable owner as STALE. Blocking forever is the worse error.
+    try:
+        os.kill(pid, 0)
+        return True
+    except ProcessLookupError:
+        return False                                 # owner gone -> stale
+    except PermissionError:
+        return False                                 # different owner -> cannot be our loop
+
+active = _marker_active(marker)
+
+if not active:
+    # Best-effort second signal: the skill was invoked but run_loop.sh has not
+    # started yet, so no marker exists. Look for the invocation in the tail of
+    # the transcript. Absence of a transcript simply means "not active".
+    tpath = p.get("transcriptPath") or ""
+    try:
+        if tpath and os.path.exists(tpath):
+            with open(tpath, encoding="utf-8", errors="replace") as f:
+                tail = f.readlines()[-200:]
+            for line in reversed(tail):
+                if "/multi-agent-mux-loop" in line:
+                    active = True
+                    break
+                if "MAM_LOOP_GUARD_RELEASE" in line:
+                    break                            # loop finished; stop here
+    except Exception:
+        pass                                         # transcript unreadable -> not active
+
+if not active:
+    emit("allow")
+
+emit("deny",
+     "The /multi-agent-mux-loop skill is active, so direct file edits are out "
+     "of scope for the orchestrator. Stop editing and delegate instead: run "
+     "bash .agents/skills/multi-agent-mux-loop/scripts/run_loop.sh "
+     "--target-agent <session> --task <goal>. "
+     "See .agents/MULTI_AGENT_RULES.md #3.2 (Invocation-Aware Scoped Guard).")
+PY
@@ -19,7 +19,10 @@
 #   4  = agent-sessions.yaml append failure
 set -euo pipefail

-source "$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)/lib.sh"
+_script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+_lib_sh="$(cd "$_script_dir/../.." 2>/dev/null || pwd)/lib.sh"
+[ -f "$_lib_sh" ] || _lib_sh="${WORKSPACE_ROOT:-$PWD}/.agents/skills/lib.sh"
+source "$_lib_sh"

 usage() {
  cat <<EOF
@@ -18,6 +18,7 @@ metadata:

 > **Companion skills**: `multi-agent-mux-create` (start), `multi-agent-mux-resume` (re-attach), `multi-agent-mux-delegate-job` (delegate).
 > **Safety Guard**: `--max-loop`, `--max-rebut`, and `--plan-talk` restrict API cost runaways.
+> **Scope Guard (O-3)**: Intercepts direct orchestrator mutations when `/multi-agent-mux-loop` is active. See [.agents/MULTI_AGENT_RULES.md #3.2](.agents/MULTI_AGENT_RULES.md#32-invocation-aware-scoped-guard-o-3).
 > **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.
@@ -79,6 +79,15 @@ if [ -z "$TARGET_AGENT" ] || [ -z "$TASK" ]; then
  echo "ERROR: --target-agent and --task are mandatory fields."
  usage
 fi
+
+MAM_LOOP_MARKER="$REPO_ROOT/.mam/loop-guard-active"
+mkdir -p "$(dirname "$MAM_LOOP_MARKER")"
+_mam_lstart() { ps -p "$1" -o lstart= 2>/dev/null | tr -s ' ' | sed 's/^ *//;s/ *$//'; }
+printf 'pid=%s\nlstart=%s\nstarted=%s\n' \
+  "$$" "$(_mam_lstart $$)" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" > "$MAM_LOOP_MARKER"
+_mam_release_guard() { rm -f "$MAM_LOOP_MARKER"; }
+trap _mam_release_guard EXIT INT TERM HUP
+
 delegate_job_safe() {
  local orig_script="$REPO_ROOT/.agents/skills/multi-agent-mux-delegate-job/multi-agent-mux-delegate-job"
  local tmp_script
@@ -88,7 +97,7 @@ delegate_job_safe() {
  local rc=0
  bash "$tmp_script" "$@" || rc=$?
  rm -f "$tmp_script"
-  trap - EXIT INT TERM HUP
+  trap _mam_release_guard EXIT INT TERM HUP
  return $rc
 }

@@ -15,8 +15,10 @@
 # Exit codes: 0 = ok | 1 = YAML not found | 2 = error
 set -euo pipefail

-SKILLS_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SKILLS_DIR="$(cd "$SCRIPT_DIR/../.." 2>/dev/null || pwd)"
 LIB_SH="$SKILLS_DIR/lib.sh"
+[ -f "$LIB_SH" ] || LIB_SH="${WORKSPACE_ROOT:-$PWD}/.agents/skills/lib.sh"
 source "$LIB_SH"
 if [ -n "${AGENT_SESSIONS_STATE_DIR:-}" ]; then
  echo "Notice: AGENT_SESSIONS_STATE_DIR is set but has no effect; the monitor keeps no state directory (C-2)." >&2
@@ -2,7 +2,10 @@
 # resume_session.sh — resume a stopped session
 set -euo pipefail

-source "$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)/lib.sh"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+LIB_SH="$(cd "$SCRIPT_DIR/../.." 2>/dev/null || pwd)/lib.sh"
+[ -f "$LIB_SH" ] || LIB_SH="${WORKSPACE_ROOT:-$PWD}/.agents/skills/lib.sh"
+source "$LIB_SH"

 usage() {
  cat <<EOF
@@ -31,7 +31,10 @@
 set -euo pipefail

 # shellcheck disable=SC1091
-source "$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)/lib.sh"
+_script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+_lib_sh="$(cd "$_script_dir/../.." 2>/dev/null || pwd)/lib.sh"
+[ -f "$_lib_sh" ] || _lib_sh="${WORKSPACE_ROOT:-$PWD}/.agents/skills/lib.sh"
+source "$_lib_sh"

 usage() {
  cat <<EOF