diff --git a/scripts/state-db-stabilize.py b/scripts/state-db-stabilize.py
index 476eecd1..3fc288bc 100755
--- a/scripts/state-db-stabilize.py
+++ b/scripts/state-db-stabilize.py
@@ -71,8 +71,9 @@ def stabilize(dry_run: bool = False) -> int:
     log("\n[1/5] 完整性体检")
     report = {"started_at": now_iso(), "actions": [], "warnings": []}
 
-    # 用只读连接快速体检（不持有写锁，避免和 gateway 抢）
-    health_conn = sqlite3.connect(f"file:{DB_PATH}?mode=ro", uri=True, timeout=10)
+    # 用 immutable 只读连接快速体检（完全跳过 wal/lock，gateway 运行时零风险）
+    # 🔴 2026-09-04 教训：gateway 运行中普通只读连接（mode=ro 非 immutable）也会触发 wal 访问
+    health_conn = sqlite3.connect(f"file:{DB_PATH}?mode=ro&immutable=1", uri=True, timeout=10)
     health_conn.execute("PRAGMA busy_timeout=5000")
 
     ok, integrity = safe_run(health_conn, "PRAGMA integrity_check")
@@ -124,6 +125,58 @@ def stabilize(dry_run: bool = False) -> int:
         log(f"  ⚠️ 发现 0 字节 WAL 残留，将清理（不破坏数据，SQLite 会自动重建）")
         report["warnings"].append("zero_wal_residue")
 
+    # === 阶段 1.5：损坏自动恢复（2026-09-03 新增 — 防止 gateway 启动失败导致 agent 失联）===
+    # ExecStartPre 场景（gateway 尚未启动）若检测到真损坏 → 从最近健康快照恢复
+    # 前提：无 gateway 进程持有 DB（手动运行时保护）
+    if (not ok or integrity != "ok") and not dry_run:
+        import subprocess as _sp
+        _gw = _sp.run(["pgrep", "-f", "hermes_cli.main gateway run"],
+                      capture_output=True, text=True)
+        if _gw.returncode == 0:
+            log("  ⚠️ gateway 正在运行，跳过自动恢复（避免覆盖活动 DB）")
+        else:
+            SNAP_DIR = Path.home() / ".hermes" / "backups" / "state-db-snap"
+            if SNAP_DIR.exists():
+                snaps = sorted(SNAP_DIR.glob("state-*.db"))
+                if snaps:
+                    latest = snaps[-1]
+                    try:
+                        _sc = sqlite3.connect(f"file:{latest}?mode=ro", uri=True, timeout=10)
+                        _sok = _sc.execute("PRAGMA integrity_check").fetchone()[0]
+                        _sc.close()
+                    except sqlite3.Error as _e:
+                        _sok = None
+                        log(f"  ⚠️ 快照 {latest.name} 读取失败: {_e}")
+                    if _sok == "ok":
+                        import shutil as _sh
+                        _ts = datetime.datetime.now().strftime('%Y%m%d_%H%M%S')
+                        _bad = DB_PATH.with_name(f"state.db.bad-auto-{_ts}")
+                        _sh.copy2(DB_PATH, _bad)
+                        _sh.copy2(latest, DB_PATH)
+                        for _p in (WAL_PATH, SHM_PATH):
+                            if _p.exists():
+                                try:
+                                    _p.unlink()
+                                except OSError:
+                                    pass
+                        log(f"  🔄 自动恢复：DB 损坏 → 从 {latest.name} 恢复（坏库备份 {_bad.name}）")
+                        report["actions"].append(f"auto_restore_from={latest.name}")
+                        # 重新体检
+                        try:
+                            _hc = sqlite3.connect(f"file:{DB_PATH}?mode=ro", uri=True, timeout=10)
+                            _ok2, _int2 = safe_run(_hc, "PRAGMA integrity_check")
+                            report["integrity_after_restore"] = _int2
+                            log(f"  恢复后 integrity_check = {_int2}")
+                            _hc.close()
+                        except sqlite3.Error as _e:
+                            log(f"  ⚠️ 恢复后体检失败: {_e}")
+                    else:
+                        log(f"  ⚠️ 最近快照 {latest.name} 也不健康（{_sok}），跳过自动恢复")
+                else:
+                    log("  ⚠️ 无可用快照（backups/state-db-snap 为空），跳过自动恢复")
+            else:
+                log("  ⚠️ 无快照目录 backups/state-db-snap，跳过自动恢复")
+
     # === 阶段 3：核心修复（写连接，序列化）===
     log("\n[3/5] 应用稳定修复")
 
@@ -221,10 +274,14 @@ def stabilize(dry_run: bool = False) -> int:
 
 
 def main() -> int:
-    p = argparse.ArgumentParser(description="state.db 止血脚本")
-    p.add_argument("--dry-run", action="store_true", help="只体检，不修改")
+    p = argparse.ArgumentParser(description="state.db 止血脚本（安全默认：只体检 dry-run，--apply 才执行修改）")
+    p.add_argument("--apply", action="store_true", help="执行实际修改（默认只体检不修改）")
+    p.add_argument("--dry-run", action="store_true", help="强制只体检（兼容旧用法）")
     args = p.parse_args()
-    return stabilize(dry_run=args.dry_run)
+    # 🔴 安全默认反转（2026-09-04）：无参数跑 = 只体检。显式 --apply（且未 --dry-run）才执行修改/自动恢复
+    dry_run = (not args.apply) or args.dry_run
+    log(f"模式: {'DRY-RUN 只体检（--apply 才执行修改）' if dry_run else 'APPLY 执行修改'}")
+    return stabilize(dry_run=dry_run)
 
 
 if __name__ == "__main__":