diff --git a/prisma/migrations/0032_host_metric/migration.sql b/prisma/migrations/0032_host_metric/migration.sql
new file mode 100644
index 000000000..e76902ec3
--- /dev/null
+++ b/prisma/migrations/0032_host_metric/migration.sql
@@ -0,0 +1,31 @@
+-- v1.4.16 phase B3: per-minute host-load snapshot table.
+--
+-- The admin /system-status section now renders a 2-hour host-load chart
+-- above the existing facts grid (Marc: "über dem Systemstatus vielleicht
+-- nochmal so ein Graf der Auslastung des Hosts der letzten zwei Stunden").
+--
+-- A pg-boss `host-metric-sample` job inserts one row per minute carrying
+-- `os.loadavg()` (1/5/15-minute) + `os.totalmem() - os.freemem()` and,
+-- on Linux, a rolling cumulative byte counter from /proc/diskstats.
+-- Disk fields are nullable so non-Linux dev hosts keep working.
+--
+-- Retention: 7 days, enforced inside the sampler at the end of each
+-- tick. Indexed by `captured_at` so the API's `WHERE captured_at >=
+-- now() - 2h ORDER BY captured_at` query is index-only.
+
+CREATE TABLE "host_metrics" (
+  "id"               TEXT    NOT NULL,
+  "captured_at"      TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  "load_avg_1"       DOUBLE PRECISION NOT NULL,
+  "load_avg_5"       DOUBLE PRECISION NOT NULL,
+  "load_avg_15"      DOUBLE PRECISION NOT NULL,
+  "mem_used_bytes"   BIGINT NOT NULL,
+  "mem_total_bytes"  BIGINT NOT NULL,
+  "disk_read_bytes"  BIGINT,
+  "disk_write_bytes" BIGINT,
+
+  CONSTRAINT "host_metrics_pkey" PRIMARY KEY ("id")
+);
+
+CREATE INDEX "host_metrics_captured_at_idx"
+  ON "host_metrics" ("captured_at");
diff --git a/prisma/schema.prisma b/prisma/schema.prisma
index 0b9aea443..7aa78a626 100644
--- a/prisma/schema.prisma
+++ b/prisma/schema.prisma
@@ -757,3 +757,29 @@ model UserAchievement {
   @@index([userId])
   @@map("user_achievements")
 }
+
+// ─── Host Metrics (admin system-status host-load chart) ──────
+
+/// Per-minute snapshot of host-process load + memory captured by the
+/// pg-boss `host-metric-sample` job. Read by `/api/admin/host-metrics`
+/// to render the host-load trend above the system-status section.
+/// 7-day retention is enforced inside the sampler — we keep the table
+/// small enough that a single `findMany` over the last 2 hours stays
+/// well under 200 rows.
+model HostMetric {
+  id             String   @id @default(cuid())
+  capturedAt     DateTime @default(now()) @map("captured_at")
+  loadAvg1       Float    @map("load_avg_1")
+  loadAvg5       Float    @map("load_avg_5")
+  loadAvg15      Float    @map("load_avg_15")
+  memUsedBytes   BigInt   @map("mem_used_bytes")
+  memTotalBytes  BigInt   @map("mem_total_bytes")
+  // Optional rolling-window byte counters from /proc/diskstats. Null
+  // outside Linux (macOS dev machines, Windows). The API endpoint
+  // converts deltas into bytes-per-second client-side.
+  diskReadBytes  BigInt?  @map("disk_read_bytes")
+  diskWriteBytes BigInt?  @map("disk_write_bytes")
+
+  @@index([capturedAt])
+  @@map("host_metrics")
+}
diff --git a/tests/integration/setup.ts b/tests/integration/setup.ts
index 846c6200b..5a5c0b1ba 100644
--- a/tests/integration/setup.ts
+++ b/tests/integration/setup.ts
@@ -43,6 +43,7 @@ export async function truncateAllTables(client: PrismaClient): Promise<void> {
     "data_backups",
     "devices",
     "feedback",
+    "host_metrics",
     "idempotency_keys",
     "integration_statuses",
     "measurements",