feat: add network monitors (ICMP/TCP/HTTP/DNS) (#2266)

Co-authored-by: xiaomiku01 <xiaomiku01@outlook.com>
Co-authored-by: henrygd <hank@henrygd.me>

.dockerignore

@@ -1,6 +1,6 @@
 # Node.js dependencies
-node_modules
-internalsite/node_modules
+node_modules/
+**/node_modules/
 
 # Go build artifacts and binaries
 build

SECURITY.md

@@ -2,6 +2,8 @@
 
 ## Reporting a Vulnerability
 
-If you find a vulnerability in the latest version, please [submit a private advisory](https://github.com/henrygd/beszel/security/advisories/new).
+**PLEASE ONLY USE SECURITY ADVISORIES FOR REAL HIGH SEVERITY VULNERABILITIES.**
 
-If it's low severity (use best judgement) you may open an issue instead of an advisory.
+If you find a vulnerability in the latest version, and it is not high severity, open an issue instead of an advisory.
+
+I am overwhelmed with advisories, often erroneous, which are clearly found and written by AI. I don't have the capacity to review all of them.

agent/agent.go

@@ -48,7 +48,8 @@ type Agent struct {
 	keys                      []gossh.PublicKey                                     // SSH public keys
 	smartManager              *SmartManager                                         // Manages SMART data
 	systemdManager            *systemdManager                                       // Manages systemd services
-	zfsManager                *ZfsManager                                           // Manages ZFS pool and dataset data
+	monitorManager            *MonitorManager                                       // Manages network monitors
+	storagePoolManager        *StoragePoolManager                                   // Manages storage pool and dataset data
 }
 
 // NewAgent creates a new agent with the given data directory for persisting data.
@@ -122,12 +123,15 @@ func NewAgent(dataDir ...string) (agent *Agent, err error) {
 	// initialize handler registry
 	agent.handlerRegistry = NewHandlerRegistry()
 
-	agent.zfsManager = newZfsManager()
+	// initialize monitor manager
+	agent.monitorManager = newMonitorManager()
 
-	// ZFS_INTERVAL env var to update ZFS detail data at this interval
+	agent.storagePoolManager = newStoragePoolManager()
+
+	// Retain ZFS_INTERVAL for the shared storage pool detail refresh interval.
 	if zfsIntervalEnv, exists := utils.GetEnv("ZFS_INTERVAL"); exists {
 		if duration, err := time.ParseDuration(zfsIntervalEnv); err == nil && duration > 0 {
-			agent.zfsManager.detailInterval = duration
+			agent.storagePoolManager.detailInterval = duration
 			agent.systemDetails.ZfsInterval = duration
 			slog.Info("ZFS_INTERVAL", "duration", duration)
 		} else {
@@ -192,6 +196,11 @@ func (a *Agent) gatherStats(options common.DataRequestOptions) *system.CombinedD
 		}
 	}
 
+	if a.monitorManager != nil {
+		data.Monitors = a.monitorManager.GetResults(cacheTimeMs)
+		slog.Debug("Monitors", "data", data.Monitors)
+	}
+
 	// skip updating systemd services if cache time is not the default 60sec interval
 	if a.systemdManager != nil && cacheTimeMs == defaultDataCacheTimeMs {
 		totalCount := uint16(a.systemdManager.getServiceStatsCount())

agent/btrfs/btrfs.go

@@ -0,0 +1,26 @@
+// Package btrfs reads btrfs filesystem state from sysfs.
+package btrfs
+
+// Filesystem is a mounted btrfs filesystem read from /sys/fs/btrfs/<uuid>.
+type Filesystem struct {
+	UUID     string // stable filesystem UUID from sysfs
+	MountID  string // kernel filesystem identity for matching monitored mounts
+	IODevice string // sole member block-device name, empty for multi-device/unknown pools
+	Name     string // label, else first mountpoint, else UUID
+	Size     uint64 // effective usable capacity, or raw member capacity when Raw
+	Raw      bool   // capacity and usage are physical bytes, unsuitable for disk alerts
+	Alloc    uint64 // raw bytes allocated to data, metadata and system chunks
+	Health   string // ONLINE, or DEGRADED when a device is missing
+	NRead    uint64 // cumulative bytes read across member devices
+	NWrite   uint64 // cumulative bytes written across member devices
+	Devices  []Device
+}
+
+// Device is one member device (devinfo/<devid>) with its error counters.
+type Device struct {
+	Name           string // "devid N"; sysfs does not expose the block device path
+	State          string // ONLINE or MISSING
+	ReadErrs       uint64
+	WriteErrs      uint64
+	CorruptionErrs uint64
+}

agent/btrfs/btrfs_linux.go

@@ -0,0 +1,285 @@
+//go:build linux
+
+package btrfs
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"unsafe"
+
+	"github.com/henrygd/beszel/agent/utils"
+	"golang.org/x/sys/unix"
+)
+
+var (
+	sysfsPath       = "/sys/fs/btrfs"
+	mountsPath      = "/proc/self/mounts"
+	mountinfoPath   = "/proc/self/mountinfo"
+	mountUUID       = MountID
+	deviceSize      = ioctlDeviceSize
+	filesystemUsage = statfsUsage
+)
+
+// Filesystems returns all mounted btrfs filesystems, or nil when there are none.
+func Filesystems() ([]Filesystem, error) {
+	entries, err := os.ReadDir(sysfsPath)
+	if errors.Is(err, os.ErrNotExist) {
+		return nil, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	mounts := mountpointsByDevice()
+	var filesystems []Filesystem
+	for _, entry := range entries {
+		if !entry.IsDir() || entry.Name() == "features" {
+			continue
+		}
+		fs, err := readFilesystem(filepath.Join(sysfsPath, entry.Name()), mounts)
+		if err != nil {
+			return nil, fmt.Errorf("btrfs %s: %w", entry.Name(), err)
+		}
+		filesystems = append(filesystems, fs)
+	}
+	return filesystems, nil
+}
+
+func readFilesystem(dir string, mounts map[string]string) (Filesystem, error) {
+	fs := Filesystem{UUID: filepath.Base(dir), Name: utils.ReadStringFile(filepath.Join(dir, "label")), Health: "UNKNOWN"}
+	for _, kind := range []string{"data", "metadata", "system"} {
+		if value, ok := utils.ReadUintFile(filepath.Join(dir, "allocation", kind, "disk_used")); ok {
+			fs.Alloc += value
+		}
+	}
+	// devices/<name> links to the block device's sysfs directory.
+	devices, err := os.ReadDir(filepath.Join(dir, "devices"))
+	if err != nil && !errors.Is(err, os.ErrNotExist) {
+		return fs, err
+	}
+	mountpoint := mounts["uuid:"+fs.UUID]
+	if fs.Name == "" {
+		fs.Name = mountpoint
+	}
+	var backingSize uint64
+	for _, dev := range devices {
+		if mountpoint == "" {
+			mountpoint = mounts[dev.Name()]
+		}
+		if fs.Name == "" {
+			fs.Name = mountpoint
+		}
+		devDir := filepath.Join(dir, "devices", dev.Name())
+		if size, ok := utils.ReadUintFile(filepath.Join(devDir, "size")); ok {
+			backingSize += size * 512
+		}
+		if stat := strings.Fields(utils.ReadStringFile(filepath.Join(devDir, "stat"))); len(stat) >= 7 {
+			fs.NRead += parseUint(stat[2]) * 512
+			fs.NWrite += parseUint(stat[6]) * 512
+		}
+	}
+	devids, err := os.ReadDir(filepath.Join(dir, "devinfo"))
+	if err != nil && !errors.Is(err, os.ErrNotExist) {
+		return fs, err
+	}
+	capacityAvailable := len(devids) > 0
+	healthKnown := len(devids) > 0
+	for _, devid := range devids {
+		devDir := filepath.Join(dir, "devinfo", devid.Name())
+		// Replacement targets do not add filesystem capacity.
+		replaceTarget, _ := utils.ReadUintFile(filepath.Join(devDir, "replace_target"))
+		if replaceTarget != 1 {
+			devid, err := strconv.ParseUint(devid.Name(), 10, 64)
+			if err != nil {
+				return fs, err
+			}
+			size, err := deviceSize(mountpoint, devid)
+			if err != nil {
+				capacityAvailable = false
+			}
+			fs.Size += size
+		}
+		dev := Device{Name: "devid " + devid.Name(), State: "ONLINE"}
+		missing := utils.ReadStringFile(filepath.Join(devDir, "missing"))
+		if missing != "0" && missing != "1" {
+			healthKnown = false
+			dev.State = "UNKNOWN"
+		}
+		if missing == "1" {
+			dev.State = "MISSING"
+			fs.Health = "DEGRADED"
+		}
+		for line := range strings.Lines(utils.ReadStringFile(filepath.Join(devDir, "error_stats"))) {
+			if fields := strings.Fields(line); len(fields) == 2 {
+				switch fields[0] {
+				case "read_errs":
+					dev.ReadErrs = parseUint(fields[1])
+				case "write_errs":
+					dev.WriteErrs = parseUint(fields[1])
+				case "corruption_errs":
+					dev.CorruptionErrs = parseUint(fields[1])
+				}
+			}
+		}
+		fs.Devices = append(fs.Devices, dev)
+	}
+	// Use one capacity source for the whole filesystem: device IDs cannot be
+	// reliably matched to block-device names in sysfs. A partial ioctl result
+	// must not be added to the complete backing-device total.
+	if !capacityAvailable {
+		fs.Size = backingSize
+	}
+	if fs.Health != "DEGRADED" && healthKnown {
+		fs.Health = "ONLINE"
+	}
+	fs.MountID = mountUUID(mountpoint)
+	if len(devices) == 1 && len(devids) == 1 && fs.Health == "ONLINE" {
+		fs.IODevice = devices[0].Name()
+	}
+	fs.Raw = true
+	if used, available, err := filesystemUsage(mountpoint); err == nil {
+		// Effective capacity excludes reserved/unavailable space, so Size-Alloc
+		// is available to applications and the usage ratio matches df.
+		fs.Size, fs.Alloc, fs.Raw = used+available, used, false
+	}
+	if fs.Name == "" {
+		fs.Name = filepath.Base(dir)
+	}
+	return fs, nil
+}
+
+// mountpointsByDevice prefers UUID matches from mountinfo and retains source
+// device names as a fallback for environments where FS_INFO is unavailable.
+func mountpointsByDevice() map[string]string {
+	mounts := mountpointsByUUID(utils.ReadStringFile(mountinfoPath), mountUUID)
+	for line := range strings.Lines(utils.ReadStringFile(mountsPath)) {
+		fields := strings.Fields(line)
+		if len(fields) < 3 || fields[2] != "btrfs" {
+			continue
+		}
+		device := fields[0]
+		if resolved, err := filepath.EvalSymlinks(device); err == nil {
+			device = resolved
+		}
+		if _, seen := mounts[filepath.Base(device)]; !seen {
+			mounts[filepath.Base(device)] = unescapeMountPath(fields[1])
+		}
+	}
+	return mounts
+}
+
+func parseUint(s string) uint64 {
+	n, _ := strconv.ParseUint(s, 10, 64)
+	return n
+}
+
+// ioctlDeviceSize reads Btrfs's recorded device size, which can be smaller
+// than the block device after a filesystem resize. BTRFS_IOC_DEV_INFO is
+// _IOWR(0x94, 30, struct btrfs_ioctl_dev_info_args), a 4096-byte ABI structure.
+func ioctlDeviceSize(mountpoint string, devid uint64) (uint64, error) {
+	if mountpoint == "" {
+		return 0, errors.New("no accessible mountpoint")
+	}
+	f, err := os.Open(mountpoint)
+	if err != nil {
+		return 0, err
+	}
+	defer f.Close()
+	args := struct {
+		Devid      uint64
+		UUID       [16]byte
+		BytesUsed  uint64
+		TotalBytes uint64
+		Reserved   [4096 - 40]byte
+	}{Devid: devid}
+	_, _, errno := unix.Syscall(unix.SYS_IOCTL, f.Fd(), 0xd000941e, uintptr(unsafe.Pointer(&args)))
+	if errno != 0 {
+		return 0, errno
+	}
+	return args.TotalBytes, nil
+}
+
+// The filesystem magic is unsigned even when Statfs_t.Type is int32.
+func isBtrfs(stat *unix.Statfs_t) bool {
+	return uint32(stat.Type) == unix.BTRFS_SUPER_MAGIC
+}
+
+func statfsUsage(path string) (used, available uint64, err error) {
+	if path == "" {
+		return 0, 0, errors.New("no accessible mountpoint")
+	}
+	var stat unix.Statfs_t
+	if err = unix.Statfs(path, &stat); err != nil {
+		return
+	}
+	if !isBtrfs(&stat) {
+		return 0, 0, errors.New("mountpoint is not Btrfs")
+	}
+	blockSize := uint64(stat.Bsize)
+	return (stat.Blocks - min(stat.Blocks, stat.Bfree)) * blockSize, min(stat.Blocks, stat.Bavail) * blockSize, nil
+}
+
+// MountID returns the filesystem UUID via BTRFS_IOC_FS_INFO. Unlike statfs
+// f_fsid, this identity is shared by all subvolumes and bind mounts.
+func MountID(path string) string {
+	if path == "" {
+		return ""
+	}
+	var stat unix.Statfs_t
+	if unix.Statfs(path, &stat) != nil || !isBtrfs(&stat) {
+		return ""
+	}
+	f, err := os.Open(path)
+	if err != nil {
+		return ""
+	}
+	defer f.Close()
+	args := struct {
+		MaxID      uint64
+		NumDevices uint64
+		FSID       [16]byte
+		Reserved   [992]byte
+	}{}
+	// _IOR(0x94, 31, 1024). Reuse the platform's read-direction bits;
+	// MIPS/PowerPC use a different encoding than asm-generic.
+	request := uintptr(unix.FS_IOC_GETFLAGS&0xe0000000) | 0x0400941f
+	_, _, errno := unix.Syscall(unix.SYS_IOCTL, f.Fd(), request, uintptr(unsafe.Pointer(&args)))
+	if errno != 0 {
+		return ""
+	}
+	id := args.FSID
+	return fmt.Sprintf("%x-%x-%x-%x-%x", id[:4], id[4:6], id[6:8], id[8:10], id[10:])
+}
+
+// Btrfs mountinfo device numbers can be virtual (0:N), so query the UUID
+// through the mount instead of comparing those numbers with sysfs block devs.
+// Retry another path when a bind mount is inaccessible. Once resolved, reuse
+// the result for that mount device to avoid opening every Docker bind mount.
+func mountpointsByUUID(mountinfo string, identify func(string) string) map[string]string {
+	mounts := make(map[string]string)
+	resolved := make(map[string]bool)
+	for line := range strings.Lines(mountinfo) {
+		before, after, ok := strings.Cut(line, " - ")
+		fields, fs := strings.Fields(before), strings.Fields(after)
+		if !ok || len(fields) < 6 || len(fs) < 3 || fs[0] != "btrfs" || resolved[fields[2]] {
+			continue
+		}
+		path := unescapeMountPath(fields[4])
+		uuid := identify(path)
+		if uuid == "" {
+			continue
+		}
+		resolved[fields[2]] = true
+		if mounts["uuid:"+uuid] == "" {
+			mounts["uuid:"+uuid] = path
+		}
+	}
+	return mounts
+}
+
+func unescapeMountPath(path string) string {
+	return strings.NewReplacer(`\040`, " ", `\011`, "\t", `\012`, "\n", `\134`, `\`).Replace(path)
+}

agent/btrfs/btrfs_linux_test.go

@@ -0,0 +1,274 @@
+//go:build testing && linux
+
+package btrfs
+
+import (
+	"os"
+	"path/filepath"
+	"strconv"
+	"testing"
+
+	"github.com/henrygd/beszel/agent/utils"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/unix"
+)
+
+func TestFilesystems(t *testing.T) {
+	root := t.TempDir()
+	oldSysfs, oldMounts := sysfsPath, mountsPath
+	sysfsPath, mountsPath = root, filepath.Join(root, "mounts")
+	t.Cleanup(func() { sysfsPath, mountsPath = oldSysfs, oldMounts })
+
+	fsDir := filepath.Join(root, "1b2c3d4e-0000-0000-0000-000000000000")
+	write := func(rel, content string) {
+		path := filepath.Join(fsDir, rel)
+		require.NoError(t, os.MkdirAll(filepath.Dir(path), 0o755))
+		require.NoError(t, os.WriteFile(path, []byte(content), 0o644))
+	}
+	require.NoError(t, os.MkdirAll(filepath.Join(root, "features"), 0o755))
+	oldUsage := filesystemUsage
+	filesystemUsage = func(string) (uint64, uint64, error) { return 0, 0, os.ErrNotExist }
+	t.Cleanup(func() { filesystemUsage = oldUsage })
+	oldDeviceSize := deviceSize
+	t.Cleanup(func() { deviceSize = oldDeviceSize })
+	deviceSize = func(_ string, devid uint64) (uint64, error) {
+		value, _ := utils.ReadUintFile(filepath.Join(fsDir, "recorded-size", strconv.FormatUint(devid, 10)))
+		return value, nil
+	}
+	// Recorded member capacities differ from the unchanged backing devices.
+	write("recorded-size/1", "256000\n")
+	write("recorded-size/2", "128000\n")
+	write("label", "tank\n")
+	write("allocation/data/disk_used", "4096\n")
+	write("allocation/metadata/disk_used", "2048\n")
+	write("allocation/system/disk_used", "1024\n")
+	write("devices/sda/size", "1000\n")
+	write("devices/sda/stat", "10 0 200 0 20 0 400 0 0 0 0\n")
+	write("devices/sdb/size", "1000\n")
+	write("devices/sdb/stat", "10 0 100 0 20 0 100 0 0 0 0\n")
+	write("devinfo/1/missing", "0\n")
+	write("devinfo/1/error_stats", "write_errs 1\nread_errs 2\nflush_errs 0\ncorruption_errs 3\ngeneration_errs 0\n")
+	write("devinfo/2/missing", "1\n")
+
+	filesystems, err := Filesystems()
+	require.NoError(t, err)
+	require.Len(t, filesystems, 1)
+	assert.Equal(t, Filesystem{
+		UUID: "1b2c3d4e-0000-0000-0000-000000000000", Raw: true, Name: "tank", Size: 384000, Alloc: 7168, Health: "DEGRADED", NRead: 153600, NWrite: 256000,
+		Devices: []Device{
+			{Name: "devid 1", State: "ONLINE", ReadErrs: 2, WriteErrs: 1, CorruptionErrs: 3},
+			{Name: "devid 2", State: "MISSING"},
+		},
+	}, filesystems[0])
+
+	// Unlabeled filesystems fall back to the first mountpoint, then the UUID.
+	write("label", "\n")
+	require.NoError(t, os.WriteFile(mountsPath, []byte(
+		"/dev/sdz1 /other btrfs rw 0 0\n/dev/sdb /mnt/storage btrfs rw 0 0\n/dev/sdb /mnt/storage/sub btrfs rw,subvol=/sub 0 0\n",
+	), 0o644))
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	assert.Equal(t, "/mnt/storage", filesystems[0].Name)
+
+	require.NoError(t, os.Remove(mountsPath))
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	assert.Equal(t, "1b2c3d4e-0000-0000-0000-000000000000", filesystems[0].Name)
+	write("devinfo/3/replace_target", "1\n")
+	write("recorded-size/3", "512000\n")
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	assert.Equal(t, uint64(384000), filesystems[0].Size, "replacement target must not inflate capacity")
+
+	deviceSize = func(string, uint64) (uint64, error) { return 0, os.ErrPermission }
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	require.Len(t, filesystems, 1)
+	assert.Equal(t, uint64(1024000), filesystems[0].Size)
+	assert.Equal(t, "DEGRADED", filesystems[0].Health)
+	assert.Equal(t, uint64(153600), filesystems[0].NRead)
+
+	// A partial ioctl result must not be mixed with the backing-device total.
+	deviceSize = func(_ string, devid uint64) (uint64, error) {
+		if devid == 2 {
+			return 0, os.ErrPermission
+		}
+		return 256000, nil
+	}
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	assert.Equal(t, uint64(1024000), filesystems[0].Size)
+
+	// With no mount visible (e.g. Docker), the real lookup falls back too.
+	deviceSize = ioctlDeviceSize
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	require.Len(t, filesystems, 1)
+	assert.Equal(t, uint64(1024000), filesystems[0].Size)
+
+	filesystemUsage = func(string) (uint64, uint64, error) { return 100, 900, nil }
+	filesystems, err = Filesystems()
+	require.NoError(t, err)
+	assert.Equal(t, uint64(1000), filesystems[0].Size)
+	assert.Equal(t, uint64(100), filesystems[0].Alloc)
+	assert.False(t, filesystems[0].Raw)
+}
+
+func TestFilesystemsNoBtrfs(t *testing.T) {
+	oldPath := sysfsPath
+	sysfsPath = filepath.Join(t.TempDir(), "missing")
+	t.Cleanup(func() { sysfsPath = oldPath })
+
+	filesystems, err := Filesystems()
+	require.NoError(t, err)
+	assert.Nil(t, filesystems)
+}
+
+func TestIoctlDeviceSizeFailure(t *testing.T) {
+	_, err := ioctlDeviceSize("", 1)
+	require.Error(t, err)
+	_, err = ioctlDeviceSize(t.TempDir(), 1)
+	require.Error(t, err)
+	assert.ErrorIs(t, err, unix.ENOTTY)
+}
+
+func TestMountpointsDecodeEscapes(t *testing.T) {
+	oldMounts := mountsPath
+	mountsPath = filepath.Join(t.TempDir(), "mounts")
+	t.Cleanup(func() { mountsPath = oldMounts })
+	require.NoError(t, os.WriteFile(mountsPath, []byte("/dev/test-btrfs /mnt/my\\040data btrfs rw 0 0\n"), 0o644))
+	assert.Equal(t, "/mnt/my data", mountpointsByDevice()["test-btrfs"])
+}
+
+func TestFilesystemWithoutDevinfo(t *testing.T) {
+	root := t.TempDir()
+	require.NoError(t, os.MkdirAll(filepath.Join(root, "devices", "sda"), 0755))
+	require.NoError(t, os.WriteFile(filepath.Join(root, "devices", "sda", "size"), []byte("1000"), 0644))
+	fs, err := readFilesystem(root, nil)
+	require.NoError(t, err)
+	assert.Equal(t, uint64(512000), fs.Size)
+	assert.True(t, fs.Raw)
+	assert.Equal(t, "UNKNOWN", fs.Health)
+	assert.Empty(t, fs.Devices)
+
+	require.NoError(t, os.MkdirAll(filepath.Join(root, "devinfo", "1"), 0755))
+	fs, err = readFilesystem(root, nil)
+	require.NoError(t, err)
+	assert.Equal(t, "UNKNOWN", fs.Health)
+	require.Len(t, fs.Devices, 1)
+	assert.Equal(t, "UNKNOWN", fs.Devices[0].State)
+
+	// Some older interfaces lack the devices directory too.
+	fs, err = readFilesystem(t.TempDir(), nil)
+	require.NoError(t, err)
+	assert.Equal(t, "UNKNOWN", fs.Health)
+}
+
+func TestLocalBtrfsUsage(t *testing.T) {
+	path := os.Getenv("BESZEL_TEST_BTRFS_MOUNT")
+	if path == "" {
+		t.Skip("set BESZEL_TEST_BTRFS_MOUNT for read-only live validation")
+	}
+	used, available, err := statfsUsage(path)
+	require.NoError(t, err)
+	filesystems, err := Filesystems()
+	require.NoError(t, err)
+	for _, fs := range filesystems {
+		if !fs.Raw && fs.Alloc == used && fs.Size == used+available {
+			t.Logf("pool=%s used=%d available=%d effective_capacity=%d", fs.Name, used, available, fs.Size)
+			return
+		}
+	}
+	t.Fatal("collector did not report the mounted filesystem's usable capacity")
+}
+
+func TestMountID(t *testing.T) {
+	assert.Empty(t, MountID(""))
+	assert.Empty(t, MountID(filepath.Join(t.TempDir(), "missing")))
+	path := os.Getenv("BESZEL_TEST_BTRFS_MOUNT")
+	if path == "" {
+		t.Skip("set BESZEL_TEST_BTRFS_MOUNT for live identity validation")
+	}
+	id := MountID(path)
+	require.NotEmpty(t, id)
+	assert.Equal(t, id, MountID(filepath.Join(path, ".")))
+}
+
+func TestMountinfoUUIDLookup(t *testing.T) {
+	info := `1 0 0:40 /@ /inaccessible ro shared:1 - btrfs /dev/mapper/unavailable rw
+2 0 0:40 /@/docker/hosts /etc/hosts ro - btrfs /dev/mapper/unavailable rw
+3 0 0:40 /@/docker/hostname /etc/hostname ro - btrfs /dev/mapper/unavailable rw
+4 0 0:41 /subvol /extra-filesystems/my\040disk ro master:2 - btrfs /dev/missing rw
+5 0 0:42 / /ext4 ro - ext4 /dev/mapper/unavailable rw
+malformed
+6 0 0:43 / /bad ro - btrfs
+`
+	var calls []string
+	mounts := mountpointsByUUID(info, func(path string) string {
+		calls = append(calls, path)
+		switch path {
+		case "/etc/hosts":
+			return "root-uuid"
+		case "/extra-filesystems/my disk":
+			return "extra-uuid"
+		}
+		return ""
+	})
+	assert.Equal(t, map[string]string{"uuid:root-uuid": "/etc/hosts", "uuid:extra-uuid": "/extra-filesystems/my disk"}, mounts)
+	assert.Equal(t, []string{"/inaccessible", "/etc/hosts", "/extra-filesystems/my disk"}, calls)
+}
+
+func TestDockerFilesystemWithoutDeviceNodes(t *testing.T) {
+	root := t.TempDir()
+	oldSysfs, oldMounts, oldInfo, oldUUID, oldUsage := sysfsPath, mountsPath, mountinfoPath, mountUUID, filesystemUsage
+	t.Cleanup(func() {
+		sysfsPath, mountsPath, mountinfoPath, mountUUID, filesystemUsage = oldSysfs, oldMounts, oldInfo, oldUUID, oldUsage
+	})
+	sysfsPath = filepath.Join(root, "sysfs")
+	mountsPath = filepath.Join(root, "missing-mounts")
+	mountinfoPath = filepath.Join(root, "mountinfo")
+	uuid := "11111111-1111-4111-8111-111111111111"
+	dir := filepath.Join(sysfsPath, uuid)
+	for path, content := range map[string]string{"devices/dm-0/size": "1000", "devinfo/1/missing": "0"} {
+		target := filepath.Join(dir, path)
+		require.NoError(t, os.MkdirAll(filepath.Dir(target), 0755))
+		require.NoError(t, os.WriteFile(target, []byte(content), 0644))
+	}
+	require.NoError(t, os.WriteFile(mountinfoPath, []byte("2 1 0:40 /@/docker/hosts /etc/hosts ro - btrfs /dev/mapper/not-in-container rw\n"), 0644))
+	mountUUID = func(path string) string {
+		if path == "/etc/hosts" {
+			return uuid
+		}
+		return ""
+	}
+	filesystemUsage = func(path string) (uint64, uint64, error) { require.Equal(t, "/etc/hosts", path); return 100, 900, nil }
+	fs, err := Filesystems()
+	require.NoError(t, err)
+	require.Len(t, fs, 1)
+	assert.Equal(t, uuid, fs[0].MountID)
+	assert.Equal(t, "dm-0", fs[0].IODevice)
+	assert.False(t, fs[0].Raw)
+	assert.Equal(t, uint64(1000), fs[0].Size)
+}
+
+func TestLivePoolMountIdentity(t *testing.T) {
+	path := os.Getenv("BESZEL_TEST_BTRFS_MOUNT")
+	if path == "" {
+		t.Skip("set BESZEL_TEST_BTRFS_MOUNT for live validation")
+	}
+	id := MountID(path)
+	require.NotEmpty(t, id)
+	pools, err := Filesystems()
+	require.NoError(t, err)
+	for _, pool := range pools {
+		if pool.UUID != id {
+			continue
+		}
+		assert.Equal(t, id, pool.MountID)
+		assert.False(t, pool.Raw)
+		t.Logf("uuid=%s mount_identity=%s io_device=%s raw=%v", pool.UUID, pool.MountID, pool.IODevice, pool.Raw)
+		return
+	}
+	t.Fatal("mounted Btrfs filesystem was not discovered")
+}

agent/btrfs/btrfs_unsupported.go

@@ -0,0 +1,11 @@
+//go:build !linux
+
+package btrfs
+
+import "errors"
+
+func Filesystems() ([]Filesystem, error) {
+	return nil, errors.ErrUnsupported
+}
+
+func MountID(string) string { return "" }

agent/client.go

@@ -25,9 +25,16 @@ import (
 )
 
 const (
-	wsDeadline = 70 * time.Second
+	// Keep the connection alive long enough for a slow collection cycle to
+	// finish before the hub considers the agent disconnected.
+	wsDeadline = 120 * time.Second
 )
 
+// errNoHubURL is returned when HUB_URL is unset. This is not a failure
+// condition: an agent configured with only a public key runs in SSH-only mode,
+// where the hub dials the agent and no outbound WebSocket client is expected.
+var errNoHubURL = errors.New("HUB_URL environment variable not set")
+
 type caCertFileError struct {
 	err error
 }
@@ -61,7 +68,7 @@ type WebSocketClient struct {
 func newWebSocketClient(agent *Agent) (client *WebSocketClient, err error) {
 	hubURLStr, exists := utils.GetEnv("HUB_URL")
 	if !exists {
-		return nil, errors.New("HUB_URL environment variable not set")
+		return nil, errNoHubURL
 	}
 
 	client = &WebSocketClient{}

agent/client_test.go

@@ -32,6 +32,28 @@ import (
 	"golang.org/x/crypto/ssh"
 )
 
+// TestNewWebSocketClientNoHubURL verifies that an unset HUB_URL returns the
+// errNoHubURL sentinel rather than an opaque error. Callers rely on this to
+// distinguish SSH-only mode -- a supported configuration in which the hub dials
+// the agent -- from an actual misconfiguration.
+func TestNewWebSocketClientNoHubURL(t *testing.T) {
+	agent := createTestAgent(t)
+
+	// t.Setenv registers restoration of the original value; unset afterwards so
+	// GetEnv's LookupEnv reports the variable as absent rather than empty.
+	t.Setenv("BESZEL_AGENT_HUB_URL", "")
+	os.Unsetenv("BESZEL_AGENT_HUB_URL")
+	t.Setenv("HUB_URL", "")
+	os.Unsetenv("HUB_URL")
+	t.Setenv("BESZEL_AGENT_TOKEN", "test-token")
+
+	client, err := newWebSocketClient(agent)
+
+	require.Error(t, err)
+	assert.Nil(t, client)
+	assert.ErrorIs(t, err, errNoHubURL)
+}
+
 // TestNewWebSocketClient tests WebSocket client creation
 func TestNewWebSocketClient(t *testing.T) {
 	agent := createTestAgent(t)
@@ -700,3 +722,11 @@ func TestGetToken(t *testing.T) {
 		assert.Equal(t, expectedToken, token, "Whitespace should be stripped from token file content")
 	})
 }
+
+func TestWebSocketDeadlineCoversSlowCollection(t *testing.T) {
+	const minimumDeadline = 120 * time.Second
+
+	if wsDeadline < minimumDeadline {
+		t.Fatalf("WebSocket deadline %s is shorter than the slow-collection window of %s", wsDeadline, minimumDeadline)
+	}
+}

agent/connection_manager.go

@@ -91,7 +91,15 @@ func (c *ConnectionManager) Start(serverOptions ServerOptions) error {
 		if errors.As(err, &caCertErr) {
 			return err
 		}
-		slog.Warn("Error creating WebSocket client", "err", err)
+		disableSSH, _ := utils.GetEnv("DISABLE_SSH")
+		if errors.Is(err, errNoHubURL) && disableSSH != "true" {
+			// SSH-only mode: the hub dials the agent, so there is nothing to warn
+			// about. With SSH also disabled there is no connection method at all,
+			// so that case still warns.
+			slog.Debug("WebSocket client not configured", "err", err)
+		} else {
+			slog.Warn("Error creating WebSocket client", "err", err)
+		}
 	}
 	c.wsClient = wsClient
 
@@ -145,6 +153,7 @@ func (c *ConnectionManager) Start(serverOptions ServerOptions) error {
 //	}
 func (c *ConnectionManager) stop() error {
 	_ = c.agent.StopServer()
+	c.agent.monitorManager.Stop()
 	c.closeWebSocket()
 	return health.CleanUp()
 }

agent/disk.go

@@ -18,11 +18,11 @@ import (
 // fsRegistrationContext holds the shared lookup state needed to resolve a
 // filesystem into the tracked fsStats key and metadata.
 type fsRegistrationContext struct {
-	filesystem         string // device part of optional FILESYSTEM env var
-	filesystemName     string // optional custom name from FILESYSTEM=device__name
-	isWindows          bool
-	efPath             string // path to extra filesystems (default "/extra-filesystems")
-	diskIoCounters     map[string]disk.IOCountersStat
+	filesystem     string // device part of optional FILESYSTEM env var
+	filesystemName string // optional custom name from FILESYSTEM=device__name
+	isWindows      bool
+	efPath         string // path to extra filesystems (default "/extra-filesystems")
+	diskIoCounters map[string]disk.IOCountersStat
 }
 
 // diskDiscovery groups the transient state for a single initializeDiskInfo run so
@@ -325,11 +325,11 @@ func (a *Agent) initializeDiskInfo() {
 	}
 	slog.Debug("Disk I/O", "diskstats", diskIoCounters)
 	ctx := fsRegistrationContext{
-		filesystem:         filesystem,
-		filesystemName:     filesystemName,
-		isWindows:          isWindows,
-		diskIoCounters:     diskIoCounters,
-		efPath:             "/extra-filesystems",
+		filesystem:     filesystem,
+		filesystemName: filesystemName,
+		isWindows:      isWindows,
+		diskIoCounters: diskIoCounters,
+		efPath:         "/extra-filesystems",
 	}
 
 	// Get the appropriate root mount point for this system
@@ -540,8 +540,8 @@ func (a *Agent) initializeDiskIoStats(diskIoCounters map[string]disk.IOCountersS
 	// ZFS datasets have no /proc/diskstats entry, so they are excluded from
 	// I/O tracking instead of warning about a missing device (#1541).
 	var zfsMountpoints map[string]bool
-	if a.zfsManager != nil {
-		zfsMountpoints = a.zfsManager.ZfsMountpoints()
+	if a.storagePoolManager != nil {
+		zfsMountpoints = a.storagePoolManager.ZfsMountpoints()
 	}
 	for device, stats := range a.fsStats {
 		if zfsMountpoints[stats.Mountpoint] {
@@ -574,8 +574,8 @@ func (a *Agent) updateDiskUsage(systemStats *system.Stats) {
 	// ZFS dataset mountpoints use `zfs list` values because statfs(2) reports
 	// dataset-level usage that excludes child datasets (#1541).
 	var zfsUsage map[string]zfsDatasetUsage
-	if a.zfsManager != nil {
-		zfsUsage = a.zfsManager.DatasetUsage()
+	if a.storagePoolManager != nil {
+		zfsUsage = a.storagePoolManager.DatasetUsage()
 	}
 
 	// disk usage

agent/disk_zfs_test.go

@@ -4,6 +4,7 @@ package agent
 
 import (
 	"testing"
+	"time"
 
 	"github.com/henrygd/beszel/agent/zfs"
 	"github.com/henrygd/beszel/internal/entities/system"
@@ -16,8 +17,8 @@ import (
 // is a ZFS dataset reports `zfs list` usage (which includes child datasets)
 // instead of the dataset-scoped statfs values (#1541).
 func TestUpdateDiskUsageZfsMountpoint(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
 		return []zfs.Dataset{
 			{Name: "tank", Used: 12000000000000, Avail: 11999000000000, Mountpoint: "/tank"},
 		}, nil
@@ -26,7 +27,7 @@ func TestUpdateDiskUsageZfsMountpoint(t *testing.T) {
 		fsStats: map[string]*system.FsStats{
 			"tank": {Root: false, Mountpoint: "/tank"},
 		},
-		zfsManager: zm,
+		storagePoolManager: zm,
 	}
 
 	var stats system.Stats
@@ -43,8 +44,8 @@ func TestUpdateDiskUsageZfsMountpoint(t *testing.T) {
 // TestUpdateDiskUsageZfsRootPopulatesSystemStats verifies the root disk values
 // are derived from ZFS usage when the root mountpoint is a ZFS dataset.
 func TestUpdateDiskUsageZfsRootPopulatesSystemStats(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
 		return []zfs.Dataset{
 			{Name: "rpool/ROOT/pve-1", Used: 900000000000, Avail: 300000000000, Mountpoint: "/"},
 		}, nil
@@ -53,7 +54,7 @@ func TestUpdateDiskUsageZfsRootPopulatesSystemStats(t *testing.T) {
 		fsStats: map[string]*system.FsStats{
 			"rpool/ROOT/pve-1": {Root: true, Mountpoint: "/"},
 		},
-		zfsManager: zm,
+		storagePoolManager: zm,
 	}
 
 	var stats system.Stats
@@ -85,8 +86,8 @@ func TestUpdateDiskUsageWithoutZfsManager(t *testing.T) {
 // TestInitializeDiskIoStatsSkipsZfsMountpoints verifies ZFS filesystems are
 // excluded from diskstats I/O tracking instead of warning about a missing device.
 func TestInitializeDiskIoStatsSkipsZfsMountpoints(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
 		return []zfs.Dataset{{Name: "tank", Mountpoint: "/tank"}}, nil
 	}
 	agent := &Agent{
@@ -94,8 +95,8 @@ func TestInitializeDiskIoStatsSkipsZfsMountpoints(t *testing.T) {
 			"tank": {Root: false, Mountpoint: "/tank"},
 			"sda1": {Root: false, Mountpoint: "/mnt/data"},
 		},
-		zfsManager: zm,
-		diskPrev:   make(map[uint16]map[string]prevDisk),
+		storagePoolManager: zm,
+		diskPrev:           make(map[uint16]map[string]prevDisk),
 	}
 
 	agent.initializeDiskIoStats(map[string]disk.IOCountersStat{

agent/docker.go

@@ -65,10 +65,14 @@ type dockerManager struct {
 	dockerVersionChecked bool                        // Whether a version probe has completed successfully
 	isWindows            bool                        // Whether the Docker Engine API is running on Windows
 	buf                  *bytes.Buffer               // Buffer to store and read response bodies
-	apiStats             *container.ApiStats         // Reusable API stats object
 	excludeContainers    []string                    // Patterns to exclude containers by name
 	usingPodman          bool                        // Whether the Docker Engine API is running on Podman
 
+	registryClient      *http.Client                  // Client for registry requests; nil uses a client with a 10-second timeout
+	imageUpdatesMutex   sync.RWMutex                  // Protects imageUpdates, its entries, and imageUpdatesRunning
+	imageUpdates        map[string]*imageUpdateStatus // Shared update status keyed by normalized image reference
+	imageUpdatesRunning bool                          // Whether a background image-update batch is in progress
+
 	// Cache-time-aware tracking for CPU stats (similar to cpu.go)
 	// Maps cache time intervals to container-specific CPU usage tracking
 	lastCpuContainer map[uint16]map[string]uint64    // cacheTimeMs -> containerId -> last cpu container usage
@@ -161,6 +165,9 @@ func (dm *dockerManager) getDockerStats(cacheTimeMs uint16) ([]*container.Stats,
 		clear(dm.validIds)
 	}
 
+	// Only schedule auxiliary work here; metrics never wait for image discovery.
+	dm.refreshImageUpdates(dm.apiContainerList, time.Now())
+
 	var failedContainers []*container.ApiInfo
 
 	for _, ctr := range dm.apiContainerList {
@@ -506,6 +513,17 @@ func (dm *dockerManager) updateContainerStats(ctr *container.ApiInfo, cacheTimeM
 		}
 	}
 
+	// Read and decode the response before locking shared stats to avoid blocking
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("container stats request failed: %s", resp.Status)
+	}
+	res := &container.ApiStats{}
+	if err := json.NewDecoder(resp.Body).Decode(res); err != nil {
+		return err
+	}
+	updateAvailable := dm.cachedImageUpdate(ctr.Image)
+
 	dm.containerStatsMutex.Lock()
 	defer dm.containerStatsMutex.Unlock()
 
@@ -520,6 +538,9 @@ func (dm *dockerManager) updateContainerStats(ctr *container.ApiInfo, cacheTimeM
 	stats.Status = statusText
 	stats.Health = health
 
+	stats.Image = ctr.Image
+	stats.UpdateAvailable = updateAvailable
+
 	if len(ctr.Ports) > 0 {
 		stats.Ports = convertContainerPortsToString(ctr)
 	}
@@ -532,12 +553,6 @@ func (dm *dockerManager) updateContainerStats(ctr *container.ApiInfo, cacheTimeM
 	stats.NetworkSent = 0
 	stats.NetworkRecv = 0
 
-	res := dm.apiStats
-	res.Networks = nil
-	if err := dm.decode(resp, res); err != nil {
-		return err
-	}
-
 	// Initialize CPU tracking for this cache time interval
 	dm.initializeCpuTracking(cacheTimeMs)
 
@@ -695,7 +710,6 @@ func newDockerManager(agent *Agent) *dockerManager {
 		containerStatsMap: make(map[string]*container.Stats),
 		sem:               make(chan struct{}, 5),
 		apiContainerList:  []*container.ApiInfo{},
-		apiStats:          &container.ApiStats{},
 		excludeContainers: excludeContainers,
 
 		// Initialize cache-time-aware tracking structures

agent/docker_image_updates.go

@@ -0,0 +1,105 @@
+package agent
+
+import (
+	"log/slog"
+	"sync"
+	"time"
+
+	"github.com/distribution/reference"
+	"github.com/henrygd/beszel/internal/entities/container"
+)
+
+const imageUpdateInterval = time.Hour
+
+type imageUpdateStatus struct {
+	available bool
+	checkedAt time.Time
+}
+
+func normalizedImageReference(image string) string {
+	named, err := reference.ParseNormalizedNamed(image)
+	if err != nil {
+		return ""
+	}
+	// Digest-pinned references cannot move to a new version.
+	if _, pinned := named.(reference.Digested); pinned {
+		return ""
+	}
+	return reference.TagNameOnly(named).String()
+}
+
+// refreshImageUpdates starts at most one background batch. Neither its network
+// work nor its completion is part of the container metrics wait group.
+func (dm *dockerManager) refreshImageUpdates(containers []*container.ApiInfo, now time.Time) {
+	dm.imageUpdatesMutex.Lock()
+	defer dm.imageUpdatesMutex.Unlock()
+	if dm.imageUpdatesRunning {
+		return
+	}
+	if dm.imageUpdates == nil {
+		dm.imageUpdates = make(map[string]*imageUpdateStatus)
+	}
+	active := make(map[string]struct{}, len(containers))
+	pending := make(map[string]*imageUpdateStatus)
+	for _, ctr := range containers {
+		if len(ctr.Names) > 0 && dm.shouldExcludeContainer(ctr.Names[0][1:]) {
+			continue
+		}
+		key := normalizedImageReference(ctr.Image)
+		if key == "" {
+			continue
+		}
+		active[key] = struct{}{}
+		entry := dm.imageUpdates[key]
+		if entry == nil {
+			entry = &imageUpdateStatus{}
+			dm.imageUpdates[key] = entry
+		}
+		if entry.checkedAt.IsZero() || now.Sub(entry.checkedAt) >= imageUpdateInterval {
+			pending[key] = entry
+		}
+	}
+	for key := range dm.imageUpdates {
+		if _, ok := active[key]; !ok {
+			delete(dm.imageUpdates, key)
+		}
+	}
+	if len(pending) == 0 {
+		return
+	}
+	dm.imageUpdatesRunning = true
+	go func() {
+		// Limit auxiliary requests even on hosts running many different images.
+		sem := make(chan struct{}, 2)
+		var wg sync.WaitGroup
+		for key, entry := range pending {
+			sem <- struct{}{}
+			wg.Add(1)
+			go func() {
+				defer wg.Done()
+				defer func() { <-sem }()
+				available, err := dm.checkImageUpdate(key)
+				if err != nil {
+					available = false
+					slog.Debug("Image update check failed", "image", key, "err", err)
+				}
+				dm.imageUpdatesMutex.Lock()
+				entry.available = available
+				entry.checkedAt = time.Now()
+				dm.imageUpdatesMutex.Unlock()
+			}()
+		}
+		wg.Wait()
+		dm.imageUpdatesMutex.Lock()
+		dm.imageUpdatesRunning = false
+		dm.imageUpdatesMutex.Unlock()
+	}()
+}
+
+func (dm *dockerManager) cachedImageUpdate(image string) bool {
+	key := normalizedImageReference(image)
+	dm.imageUpdatesMutex.RLock()
+	defer dm.imageUpdatesMutex.RUnlock()
+	entry := dm.imageUpdates[key]
+	return entry != nil && entry.available
+}

agent/docker_image_updates_test.go

@@ -0,0 +1,225 @@
+//go:build testing
+
+package agent
+
+import (
+	"encoding/json"
+	"fmt"
+	"github.com/fxamacker/cbor/v2"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/container"
+	"github.com/stretchr/testify/require"
+)
+
+func waitForImageUpdates(t *testing.T, dm *dockerManager) {
+	t.Helper()
+	require.Eventually(t, func() bool {
+		dm.imageUpdatesMutex.RLock()
+		defer dm.imageUpdatesMutex.RUnlock()
+		return !dm.imageUpdatesRunning
+	}, time.Second*3, time.Millisecond)
+}
+
+func TestImageUpdateCacheAndStats(t *testing.T) {
+	local := "sha256:" + strings.Repeat("a", 64)
+	remote := "sha256:" + strings.Repeat("b", 64)
+	var inspections, lookups atomic.Int32
+	var fail atomic.Bool
+	var upToDate atomic.Bool
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case strings.HasPrefix(r.URL.Path, "/images/"):
+			inspections.Add(1)
+			fmt.Fprintf(w, `{"RepoDigests":["docker.io/library/nginx@%s"]}`, local)
+		case r.URL.Path == "/containers/json":
+			fmt.Fprint(w, `[{"Id":"aaaaaaaaaaaa","Names":["/one"],"Image":"nginx","Status":"Up 2 hours"},{"Id":"bbbbbbbbbbbb","Names":["/two"],"Image":"docker.io/library/nginx:latest","Status":"Up 2 hours"}]`)
+		case strings.Contains(r.URL.Path, "/stats"):
+			fmt.Fprint(w, `{"memory_stats":{"usage":1048576},"cpu_stats":{},"networks":{}}`)
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer server.Close()
+	dm := newDockerManagerForVersionTest(server)
+	dm.dockerVersionChecked = true
+	dm.registryClient = &http.Client{Timeout: time.Second, Transport: roundTripFunc(func(r *http.Request) (*http.Response, error) {
+		if fail.Load() {
+			return nil, fmt.Errorf("registry unavailable")
+		}
+		response := &http.Response{StatusCode: 200, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(`{"token":"test"}`))}
+		if r.Method == http.MethodHead {
+			lookups.Add(1)
+			digest := remote
+			if upToDate.Load() {
+				digest = local
+			}
+			response.Header.Set("Docker-Content-Digest", digest)
+		}
+		return response, nil
+	})}
+	stats, err := dm.getDockerStats(defaultCacheTimeMs)
+	require.NoError(t, err)
+	require.Len(t, stats, 2)
+	waitForImageUpdates(t, dm)
+	require.EqualValues(t, 1, lookups.Load())
+	require.EqualValues(t, 1, inspections.Load())
+	stats, err = dm.getDockerStats(defaultCacheTimeMs)
+	require.NoError(t, err)
+	for _, stat := range stats {
+		require.True(t, stat.UpdateAvailable)
+		if stat.Id == "aaaaaaaaaaaa" {
+			require.Equal(t, "nginx", stat.Image)
+		} else {
+			require.Equal(t, "docker.io/library/nginx:latest", stat.Image)
+		}
+	}
+	require.EqualValues(t, 1, lookups.Load())
+
+	expire := func() {
+		dm.imageUpdatesMutex.Lock()
+		dm.imageUpdates["docker.io/library/nginx:latest"].checkedAt = time.Now().Add(-imageUpdateInterval)
+		dm.imageUpdatesMutex.Unlock()
+	}
+	upToDate.Store(true)
+	expire()
+	_, err = dm.getDockerStats(defaultCacheTimeMs)
+	require.NoError(t, err)
+	waitForImageUpdates(t, dm)
+	require.EqualValues(t, 2, lookups.Load())
+	require.False(t, dm.cachedImageUpdate("nginx:latest"))
+
+	// An expired positive result is cleared on failure, and the failure itself
+	// is cached so realtime stats do not retry a broken registry every second.
+	dm.imageUpdatesMutex.Lock()
+	dm.imageUpdates["docker.io/library/nginx:latest"].available = true
+	dm.imageUpdatesMutex.Unlock()
+	fail.Store(true)
+	expire()
+	_, err = dm.getDockerStats(defaultCacheTimeMs)
+	require.NoError(t, err)
+	waitForImageUpdates(t, dm)
+	failedInspections := inspections.Load()
+	stats, err = dm.getDockerStats(defaultCacheTimeMs)
+	require.NoError(t, err)
+	require.Len(t, stats, 2)
+	require.Equal(t, failedInspections, inspections.Load())
+	for _, stat := range stats {
+		require.False(t, stat.UpdateAvailable)
+		require.Equal(t, 1.0, stat.Mem)
+	}
+}
+
+func TestImageDiscoveryDoesNotBlockStats(t *testing.T) {
+	started := make(chan struct{}, 1)
+	release := make(chan struct{})
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, "/images/") {
+			fmt.Fprintf(w, `{"RepoDigests":["example.com/app@sha256:%s"]}`, strings.Repeat("a", 64))
+		} else {
+			fmt.Fprint(w, `{"memory_stats":{"usage":1048576}}`)
+		}
+	}))
+	defer server.Close()
+	dm := newDockerManagerForVersionTest(server)
+	defer func() { close(release); waitForImageUpdates(t, dm) }()
+	dm.registryClient = &http.Client{Transport: roundTripFunc(func(r *http.Request) (*http.Response, error) {
+		started <- struct{}{}
+		<-release
+		return nil, fmt.Errorf("timeout")
+	})}
+	ctr := &container.ApiInfo{IdShort: "aaaaaaaaaaaa", Image: "example.com/app", Names: []string{"/one"}}
+	dm.refreshImageUpdates([]*container.ApiInfo{ctr}, time.Now())
+	select {
+	case <-started:
+	case <-time.After(3 * time.Second):
+		t.Fatal("check did not start")
+	}
+	done := make(chan error, 1)
+	go func() { done <- dm.updateContainerStats(ctr, defaultCacheTimeMs) }()
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(time.Second):
+		t.Fatal("registry blocked stats")
+	}
+	dm.imageUpdatesMutex.RLock()
+	require.True(t, dm.imageUpdatesRunning)
+	dm.imageUpdatesMutex.RUnlock()
+}
+
+func TestNormalizeImageUpdateReferences(t *testing.T) {
+	require.Equal(t, normalizedImageReference("nginx"), normalizedImageReference("docker.io/library/nginx:latest"))
+	require.Empty(t, normalizedImageReference("bad reference"))
+	require.Empty(t, normalizedImageReference("nginx@sha256:"+strings.Repeat("a", 64)))
+}
+
+// A stats request can return headers promptly and then stall while reading its
+// body. The stats-map mutex must remain available during that read.
+func TestStatsResponseBodyDoesNotHoldStatsLock(t *testing.T) {
+	started := make(chan struct{})
+	release := make(chan struct{})
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusOK)
+		w.(http.Flusher).Flush()
+		close(started)
+		<-release
+		fmt.Fprint(w, `{"memory_stats":{"usage":1048576}}`)
+	}))
+	defer server.Close()
+	dm := newDockerManagerForVersionTest(server)
+	done := make(chan error, 1)
+	go func() {
+		done <- dm.updateContainerStats(&container.ApiInfo{IdShort: "aaaaaaaaaaaa", Names: []string{"/one"}, Image: "nginx"}, defaultCacheTimeMs)
+	}()
+	<-started
+	locked := make(chan struct{})
+	go func() { dm.containerStatsMutex.Lock(); dm.containerStatsMutex.Unlock(); close(locked) }()
+	select {
+	case <-locked:
+	case <-time.After(time.Second):
+		close(release)
+		<-done
+		t.Fatal("Docker response body held the stats mutex")
+	}
+	close(release)
+	require.NoError(t, <-done)
+}
+
+func TestImageUpdateStatsEncoding(t *testing.T) {
+	original := container.Stats{Image: "nginx:latest", UpdateAvailable: true}
+	encoded, err := cbor.Marshal(original)
+	require.NoError(t, err)
+	var fields map[int]any
+	require.NoError(t, cbor.Unmarshal(encoded, &fields))
+	require.Equal(t, true, fields[11])
+	require.Equal(t, "nginx:latest", fields[8])
+	var decoded container.Stats
+	require.NoError(t, cbor.Unmarshal(encoded, &decoded))
+	require.True(t, decoded.UpdateAvailable)
+	require.Equal(t, original.Image, decoded.Image)
+	encoded, err = json.Marshal(original)
+	require.NoError(t, err)
+	require.Contains(t, string(encoded), `"u":true`)
+}
+
+func TestImageUpdateCacheExpiryBoundaryAndPruning(t *testing.T) {
+	now := time.Now()
+	key := normalizedImageReference("nginx")
+	dm := &dockerManager{imageUpdates: map[string]*imageUpdateStatus{
+		key:                           {available: true, checkedAt: now},
+		"unused.example/image:latest": {checkedAt: now},
+	}}
+	dm.refreshImageUpdates([]*container.ApiInfo{{Image: "nginx"}}, now.Add(imageUpdateInterval-time.Nanosecond))
+	require.False(t, dm.imageUpdatesRunning)
+	require.Len(t, dm.imageUpdates, 1)
+	require.True(t, dm.cachedImageUpdate("nginx:latest"))
+	dm.refreshImageUpdates(nil, now)
+	require.Empty(t, dm.imageUpdates)
+}

agent/docker_registry.go

@@ -0,0 +1,222 @@
+package agent
+
+import (
+	_ "crypto/sha256"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+
+	"github.com/distribution/reference"
+	"github.com/opencontainers/go-digest"
+)
+
+const imageRegistryTimeout = 10 * time.Second
+
+const imageManifestAccept = "application/vnd.docker.distribution.manifest.list.v2+json, " +
+	"application/vnd.docker.distribution.manifest.v2+json, " +
+	"application/vnd.oci.image.manifest.v1+json, " +
+	"application/vnd.oci.image.index.v1+json"
+
+// checkImageUpdate compares the digest recorded by Docker for image with the
+// digest currently advertised by its registry. A digest-pinned reference is
+// immutable and therefore never has an update available.
+func (dm *dockerManager) checkImageUpdate(image string) (bool, error) {
+	named, err := reference.ParseNormalizedNamed(image)
+	if err != nil {
+		return false, fmt.Errorf("parse image reference %q: %w", image, err)
+	}
+	if _, pinned := named.(reference.Digested); pinned {
+		return false, nil
+	}
+	named = reference.TagNameOnly(named)
+
+	registry := reference.Domain(named)
+	repository := reference.Path(named)
+	tag := named.(reference.Tagged).Tag()
+
+	localDigest, err := dm.inspectImageDigest(image, registry, repository)
+	if err != nil {
+		return false, err
+	}
+
+	remoteDigest, err := dm.registryImageDigest(registry, repository, tag)
+	if err != nil {
+		return false, err
+	}
+
+	return remoteDigest != localDigest, nil
+}
+
+// inspectImageDigest reads Docker's image metadata without using dm.decode.
+// The checker runs in the image-discovery goroutine, so it must not hold any
+// of the container statistics locks while waiting on the Docker API.
+func (dm *dockerManager) inspectImageDigest(image, registry, repository string) (string, error) {
+	if dm.client == nil {
+		return "", fmt.Errorf("inspect image %q: Docker client is unavailable", image)
+	}
+
+	endpoint := "http://localhost/images/" + url.PathEscape(image) + "/json"
+	resp, err := dm.client.Get(endpoint)
+	if err != nil {
+		return "", fmt.Errorf("inspect image %q: %w", image, err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("inspect image %q failed: %s", image, responseStatus(resp))
+	}
+
+	var inspect struct {
+		RepoDigests []string `json:"RepoDigests"`
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&inspect); err != nil {
+		return "", fmt.Errorf("decode image inspect %q: %w", image, err)
+	}
+	if len(inspect.RepoDigests) == 0 {
+		return "", fmt.Errorf("inspect image %q returned no repository digests", image)
+	}
+
+	localDigest, ok := matchingRepositoryDigest(inspect.RepoDigests, registry, repository)
+	if !ok {
+		return "", fmt.Errorf("inspect image %q returned no valid digest for %s/%s", image, registry, repository)
+	}
+	return localDigest, nil
+}
+
+// matchingRepositoryDigest returns a valid digest belonging to the requested
+// repository. Docker can return multiple RepoDigests for one local image; an
+// unrelated first entry must never be used for the comparison.
+func matchingRepositoryDigest(repoDigests []string, registry, repository string) (string, bool) {
+	for _, repoDigest := range repoDigests {
+		repoDigest = strings.TrimSpace(repoDigest)
+		at := strings.LastIndexByte(repoDigest, '@')
+		if at <= 0 || at == len(repoDigest)-1 || strings.Contains(repoDigest[:at], "@") {
+			continue
+		}
+
+		repoRef, err := reference.ParseNormalizedNamed(repoDigest[:at])
+		if err != nil || reference.Path(repoRef) != repository || !sameRegistry(reference.Domain(repoRef), registry) {
+			continue
+		}
+		if _, hasTag := repoRef.(reference.Tagged); hasTag {
+			continue
+		}
+
+		d, err := digest.Parse(repoDigest[at+1:])
+		if err != nil {
+			continue
+		}
+		return d.String(), true
+	}
+	return "", false
+}
+
+func sameRegistry(left, right string) bool {
+	left = canonicalRegistry(left)
+	right = canonicalRegistry(right)
+	return left == right ||
+		(left == "ghcr.io" && right == "lscr.io") ||
+		(left == "lscr.io" && right == "ghcr.io")
+}
+
+func canonicalRegistry(registry string) string {
+	if registry == "index.docker.io" {
+		return "docker.io"
+	}
+	return registry
+}
+
+func (dm *dockerManager) registryImageDigest(registry, repository, tag string) (string, error) {
+	client := dm.registryClient
+	if client == nil {
+		client = &http.Client{Timeout: imageRegistryTimeout}
+	}
+
+	token, err := dm.registryToken(client, registry, repository)
+	if err != nil {
+		return "", err
+	}
+
+	host := registry
+	if registry == "docker.io" {
+		host = "registry-1.docker.io"
+	}
+	manifestURL := "https://" + host + "/v2/" + repository + "/manifests/" + url.PathEscape(tag)
+	req, err := http.NewRequest(http.MethodHead, manifestURL, nil)
+	if err != nil {
+		return "", fmt.Errorf("create manifest request: %w", err)
+	}
+	req.Header.Set("Accept", imageManifestAccept)
+	if token != "" {
+		req.Header.Set("Authorization", "Bearer "+token)
+	}
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("fetch manifest %s:%s: %w", registry, repository, err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("manifest request for %s:%s failed: %s", repository, tag, responseStatus(resp))
+	}
+
+	remote := strings.TrimSpace(resp.Header.Get("Docker-Content-Digest"))
+	d, err := digest.Parse(remote)
+	if err != nil {
+		return "", fmt.Errorf("manifest request for %s:%s returned invalid digest: %w", repository, tag, err)
+	}
+	return d.String(), nil
+}
+
+func (dm *dockerManager) registryToken(client *http.Client, registry, repository string) (string, error) {
+	var authURL string
+	switch registry {
+	case "docker.io":
+		authURL = "https://auth.docker.io/token?service=registry.docker.io&scope=" + url.QueryEscape("repository:"+repository+":pull")
+	case "ghcr.io", "lscr.io":
+		// lscr.io is the LinuxServer alias for its GHCR-backed images.
+		authURL = "https://ghcr.io/token?service=ghcr.io&scope=" + url.QueryEscape("repository:"+repository+":pull")
+	default:
+		// Anonymous registries remain supported, as they were before the
+		// authenticated Docker Hub and GHCR paths were added.
+		return "", nil
+	}
+
+	req, err := http.NewRequest(http.MethodGet, authURL, nil)
+	if err != nil {
+		return "", fmt.Errorf("create registry auth request: %w", err)
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("fetch registry auth token for %s: %w", repository, err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("registry auth request for %s failed: %s", repository, responseStatus(resp))
+	}
+
+	var tokenResponse struct {
+		Token       string `json:"token"`
+		AccessToken string `json:"access_token"`
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&tokenResponse); err != nil {
+		return "", fmt.Errorf("decode registry auth response for %s: %w", repository, err)
+	}
+	token := strings.TrimSpace(tokenResponse.Token)
+	if token == "" {
+		token = strings.TrimSpace(tokenResponse.AccessToken)
+	}
+	if token == "" {
+		return "", fmt.Errorf("registry auth response for %s contained no token", repository)
+	}
+	return token, nil
+}
+
+func responseStatus(resp *http.Response) string {
+	if resp.Status != "" {
+		return resp.Status
+	}
+	return http.StatusText(resp.StatusCode)
+}

agent/docker_registry_test.go

@@ -0,0 +1,204 @@
+//go:build testing
+
+package agent
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+type registryTransportFunc func(*http.Request) (*http.Response, error)
+
+func (fn registryTransportFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return fn(req)
+}
+
+func registryResponse(status int, body string) *http.Response {
+	return &http.Response{
+		StatusCode: status,
+		Status:     fmt.Sprintf("%d %s", status, http.StatusText(status)),
+		Header:     make(http.Header),
+		Body:       io.NopCloser(strings.NewReader(body)),
+	}
+}
+
+func registryDigest(fill byte) string {
+	return "sha256:" + strings.Repeat(string(fill), 64)
+}
+
+func newRegistryChecker(t *testing.T, inspectBody string, transport http.RoundTripper) *dockerManager {
+	t.Helper()
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, "/images/") {
+			w.Header().Set("Content-Type", "application/json")
+			_, _ = io.WriteString(w, inspectBody)
+			return
+		}
+		http.NotFound(w, r)
+	}))
+	t.Cleanup(server.Close)
+
+	return &dockerManager{
+		client:         newDockerManagerForVersionTest(server).client,
+		registryClient: &http.Client{Transport: transport},
+	}
+}
+
+func TestCheckImageUpdateUsesInspectAndManifestDigests(t *testing.T) {
+	local := registryDigest('a')
+	remote := registryDigest('b')
+	var authCalls, manifestCalls atomic.Int32
+	dm := newRegistryChecker(t, fmt.Sprintf(`{"RepoDigests":["docker.io/library/alpine@%s"]}`, local), registryTransportFunc(func(req *http.Request) (*http.Response, error) {
+		switch {
+		case req.Method == http.MethodGet && req.URL.Host == "auth.docker.io":
+			authCalls.Add(1)
+			require.Equal(t, "/token", req.URL.Path)
+			return registryResponse(http.StatusOK, `{"token":"test-token"}`), nil
+		case req.Method == http.MethodHead && req.URL.Host == "registry-1.docker.io":
+			manifestCalls.Add(1)
+			require.Equal(t, "/v2/library/alpine/manifests/latest", req.URL.Path)
+			require.Equal(t, "Bearer test-token", req.Header.Get("Authorization"))
+			resp := registryResponse(http.StatusOK, "")
+			resp.Header.Set("Docker-Content-Digest", remote)
+			return resp, nil
+		default:
+			return registryResponse(http.StatusNotFound, ""), nil
+		}
+	}))
+
+	available, err := dm.checkImageUpdate("alpine")
+	require.NoError(t, err)
+	require.True(t, available)
+	require.EqualValues(t, 1, authCalls.Load())
+	require.EqualValues(t, 1, manifestCalls.Load())
+}
+
+func TestCheckImageUpdateReportsUnknownInspectState(t *testing.T) {
+	for _, test := range []struct {
+		name string
+		body string
+	}{
+		{name: "missing field", body: `{}`},
+		{name: "empty field", body: `{"RepoDigests":[]}`},
+		{name: "malformed reference", body: `{"RepoDigests":["not-a-repo-digest"]}`},
+		{name: "wrong repository", body: `{"RepoDigests":["docker.io/library/busybox@` + registryDigest('a') + `"]}`},
+		{name: "malformed digest", body: `{"RepoDigests":["docker.io/library/alpine@sha256:not-a-digest"]}`},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			var registryCalls atomic.Int32
+			dm := newRegistryChecker(t, test.body, registryTransportFunc(func(req *http.Request) (*http.Response, error) {
+				registryCalls.Add(1)
+				return registryResponse(http.StatusOK, `{"token":"unexpected"}`), nil
+			}))
+
+			available, err := dm.checkImageUpdate("alpine")
+			require.Error(t, err)
+			require.False(t, available)
+			require.EqualValues(t, 0, registryCalls.Load(), "invalid local state must not query a registry")
+		})
+	}
+}
+
+func TestCheckImageUpdateChecksInspectAuthAndManifestStatuses(t *testing.T) {
+	local := registryDigest('a')
+	validInspect := fmt.Sprintf(`{"RepoDigests":["docker.io/library/alpine@%s"]}`, local)
+
+	tests := []struct {
+		name         string
+		inspectCode  int
+		authCode     int
+		manifestCode int
+		remote       string
+		want         string
+	}{
+		{name: "inspect status", inspectCode: http.StatusNotFound, want: "inspect image"},
+		{name: "auth status", inspectCode: http.StatusOK, authCode: http.StatusUnauthorized, want: "registry auth"},
+		{name: "manifest status", inspectCode: http.StatusOK, authCode: http.StatusOK, manifestCode: http.StatusNotFound, remote: local, want: "manifest request"},
+		{name: "missing digest", inspectCode: http.StatusOK, authCode: http.StatusOK, manifestCode: http.StatusOK, want: "invalid digest"},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if test.inspectCode != http.StatusOK && strings.HasPrefix(r.URL.Path, "/images/") {
+					w.WriteHeader(test.inspectCode)
+					return
+				}
+				_, _ = io.WriteString(w, validInspect)
+			}))
+			t.Cleanup(server.Close)
+
+			calls := 0
+			dm := &dockerManager{client: newDockerManagerForVersionTest(server).client, registryClient: &http.Client{Transport: registryTransportFunc(func(req *http.Request) (*http.Response, error) {
+				calls++
+				if req.Method == http.MethodGet {
+					return registryResponse(test.authCode, `{"token":"test"}`), nil
+				}
+				response := registryResponse(test.manifestCode, "")
+				response.Header.Set("Docker-Content-Digest", test.remote)
+				return response, nil
+			})}}
+
+			_, err := dm.checkImageUpdate("alpine")
+			require.Error(t, err)
+			require.Contains(t, err.Error(), test.want)
+			if test.inspectCode != http.StatusOK {
+				require.Zero(t, calls)
+			}
+		})
+	}
+}
+
+func TestCheckImageUpdateSupportsAnonymousAndLSCRRegistries(t *testing.T) {
+	t.Run("anonymous registry", func(t *testing.T) {
+		local := registryDigest('a')
+		var calls atomic.Int32
+		dm := newRegistryChecker(t, fmt.Sprintf(`{"RepoDigests":["example.com/app@%s"]}`, local), registryTransportFunc(func(req *http.Request) (*http.Response, error) {
+			calls.Add(1)
+			require.Equal(t, http.MethodHead, req.Method)
+			require.Equal(t, "example.com", req.URL.Host)
+			resp := registryResponse(http.StatusOK, "")
+			resp.Header.Set("Docker-Content-Digest", local)
+			return resp, nil
+		}))
+		available, err := dm.checkImageUpdate("example.com/app")
+		require.NoError(t, err)
+		require.False(t, available)
+		require.EqualValues(t, 1, calls.Load())
+	})
+
+	t.Run("lscr ghcr alias", func(t *testing.T) {
+		local := registryDigest('a')
+		var authCalls, manifestCalls atomic.Int32
+		dm := newRegistryChecker(t, fmt.Sprintf(`{"RepoDigests":["ghcr.io/linuxserver/app@%s"]}`, local), registryTransportFunc(func(req *http.Request) (*http.Response, error) {
+			if req.Method == http.MethodGet {
+				authCalls.Add(1)
+				return registryResponse(http.StatusOK, `{"token":"test"}`), nil
+			}
+			manifestCalls.Add(1)
+			require.Equal(t, "lscr.io", req.URL.Host)
+			resp := registryResponse(http.StatusOK, "")
+			resp.Header.Set("Docker-Content-Digest", local)
+			return resp, nil
+		}))
+		available, err := dm.checkImageUpdate("lscr.io/linuxserver/app")
+		require.NoError(t, err)
+		require.False(t, available)
+		require.EqualValues(t, 1, authCalls.Load())
+		require.EqualValues(t, 1, manifestCalls.Load())
+	})
+}
+
+func TestCheckImageUpdateSkipsPinnedDigest(t *testing.T) {
+	image := "docker.io/library/alpine@" + registryDigest('a')
+	dm := &dockerManager{}
+	available, err := dm.checkImageUpdate(image)
+	require.NoError(t, err)
+	require.False(t, available)
+}

agent/docker_test.go

@@ -1184,7 +1184,6 @@ func TestUpdateContainerStatsPodmanCpuCalculation(t *testing.T) {
 			}
 		})},
 		containerStatsMap: make(map[string]*container.Stats),
-		apiStats:          &container.ApiStats{},
 		usingPodman:       true,
 		lastCpuContainer: map[uint16]map[string]uint64{
 			defaultCacheTimeMs: {"0123456789ab": prevCpuUsage},
@@ -1676,7 +1675,6 @@ func TestUpdateContainerStatsUsesPodmanInspectHealthFallback(t *testing.T) {
 			}
 		})},
 		containerStatsMap:   make(map[string]*container.Stats),
-		apiStats:            &container.ApiStats{},
 		usingPodman:         true,
 		lastCpuContainer:    make(map[uint16]map[string]uint64),
 		lastCpuSystem:       make(map[uint16]map[string]uint64),

agent/gpu_test.go

@@ -1119,7 +1119,6 @@ func TestCalculateGPUAverage(t *testing.T) {
 }
 
 func TestGPUCapabilitiesAndLegacyPriority(t *testing.T) {
-	// Save original PATH
 	hasAmdSysfs := (&GPUManager{}).hasAmdSysfs()
 
 	tests := []struct {
@@ -1213,7 +1212,7 @@ echo "[]"`
 		{
 			name: "no gpu tools available",
 			setupCommands: func(_ string) error {
-				t.Setenv("PATH", "")
+				// The subtest already restricts PATH to its empty temporary directory.
 				return nil
 			},
 			wantErr: true,

agent/handlers.go

@@ -7,6 +7,7 @@ import (
 
 	"github.com/fxamacker/cbor/v2"
 	"github.com/henrygd/beszel/internal/common"
+	"github.com/henrygd/beszel/internal/entities/monitor"
 	"github.com/henrygd/beszel/internal/entities/smart"
 
 	"log/slog"
@@ -51,6 +52,7 @@ func NewHandlerRegistry() *HandlerRegistry {
 	registry.Register(common.GetContainerInfo, &GetContainerInfoHandler{})
 	registry.Register(common.GetSmartData, &GetSmartDataHandler{})
 	registry.Register(common.GetSystemdInfo, &GetSystemdInfoHandler{})
+	registry.Register(common.SyncNetworkMonitors, &SyncNetworkMonitorsHandler{})
 	registry.Register(common.GetZfsData, &GetZfsDataHandler{})
 
 	return registry
@@ -186,14 +188,14 @@ func (h *GetSmartDataHandler) Handle(hctx *HandlerContext) error {
 type GetZfsDataHandler struct{}
 
 func (h *GetZfsDataHandler) Handle(hctx *HandlerContext) error {
-	if hctx.Agent.zfsManager == nil {
+	if hctx.Agent.storagePoolManager == nil {
 		return hctx.SendResponse(nil, hctx.RequestID)
 	}
 	var req common.ZfsDataRequest
 	if err := cbor.Unmarshal(hctx.Request.Data, &req); err != nil {
 		return err
 	}
-	return hctx.SendResponse(hctx.Agent.zfsManager.GetDetail(req.Force), hctx.RequestID)
+	return hctx.SendResponse(hctx.Agent.storagePoolManager.GetDetail(req.Force), hctx.RequestID)
 }
 
 ////////////////////////////////////////////////////////////////////////////
@@ -223,3 +225,21 @@ func (h *GetSystemdInfoHandler) Handle(hctx *HandlerContext) error {
 
 	return hctx.SendResponse(details, hctx.RequestID)
 }
+
+////////////////////////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////////////////////////
+
+// SyncNetworkMonitorsHandler handles monitor configuration sync from hub
+type SyncNetworkMonitorsHandler struct{}
+
+func (h *SyncNetworkMonitorsHandler) Handle(hctx *HandlerContext) error {
+	var req monitor.SyncRequest
+	if err := cbor.Unmarshal(hctx.Request.Data, &req); err != nil {
+		return err
+	}
+	resp, err := hctx.Agent.monitorManager.HandleSyncRequest(req)
+	if err != nil {
+		return err
+	}
+	return hctx.SendResponse(resp, hctx.RequestID)
+}

agent/handlers_test.go

@@ -34,19 +34,19 @@ func TestNewAgentResponseSmartData(t *testing.T) {
 
 func TestGetZfsDataHandlerForceRefresh(t *testing.T) {
 	poolCalls := 0
-	zm := &ZfsManager{detailInterval: time.Hour}
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
 		poolCalls++
 		return []zfs.PoolStat{{Name: "tank", Alloc: uint64(poolCalls)}}, nil
 	}
-	zm.poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, nil }
-	zm.datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
+	zm.backends[0].poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, nil }
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
 	zm.GetDetail(false)
 
 	requestData, err := cbor.Marshal(common.ZfsDataRequest{Force: true})
 	assert.NoError(t, err)
 	ctx := &HandlerContext{
-		Agent: &Agent{zfsManager: zm},
+		Agent: &Agent{storagePoolManager: zm},
 		Request: &common.HubRequest[cbor.RawMessage]{
 			Action: common.GetZfsData,
 			Data:   requestData,

agent/network_monitor.go

@@ -0,0 +1,176 @@
+package agent
+
+import (
+	"errors"
+	"fmt"
+	"net/http"
+	"sync"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+)
+
+// MonitorManager manages network monitor configurations and task lifetimes.
+type MonitorManager struct {
+	mu          sync.RWMutex
+	monitors    map[string]*monitorTask // keyed by monitor ID
+	probe       monitorProbe
+	resumeGuard monitorResumeGuard
+}
+
+func newMonitorManager() *MonitorManager {
+	return newMonitorManagerWithProbe(networkMonitorProbe(&http.Client{Timeout: monitor.MaxProbeTimeout}))
+}
+
+func newMonitorManagerWithProbe(probe monitorProbe) *MonitorManager {
+	return &MonitorManager{monitors: make(map[string]*monitorTask), probe: probe}
+}
+
+// SyncMonitors replaces all monitor tasks with the given configs.
+func (pm *MonitorManager) SyncMonitors(configs []monitor.Config) {
+	pm.mu.Lock()
+	defer pm.mu.Unlock()
+
+	// Build set of new keys
+	newKeys := make(map[string]monitor.Config, len(configs))
+	for _, cfg := range configs {
+		if cfg.ID == "" {
+			continue
+		}
+		newKeys[cfg.ID] = cfg
+	}
+
+	// Stop removed monitors
+	for key, task := range pm.monitors {
+		if _, exists := newKeys[key]; !exists {
+			task.cancel()
+			delete(pm.monitors, key)
+		}
+	}
+
+	// Start new monitors and restart tasks whose config changed.
+	for key, cfg := range newKeys {
+		task, exists := pm.monitors[key]
+		if exists && task.config == cfg {
+			continue
+		}
+		if exists {
+			task.cancel()
+		}
+		task = newMonitorTaskFromExisting(cfg, task)
+		task.resumeGuard = &pm.resumeGuard
+		pm.resumeGuard.start()
+		pm.monitors[key] = task
+		pm.startMonitor(task)
+	}
+	if len(pm.monitors) == 0 {
+		pm.resumeGuard.shutdown()
+	}
+}
+
+// HandleSyncRequest applies a full or incremental monitor sync request.
+func (pm *MonitorManager) HandleSyncRequest(req monitor.SyncRequest) (monitor.SyncResponse, error) {
+	switch req.Action {
+	case monitor.SyncActionReplace:
+		pm.SyncMonitors(req.Configs)
+		return monitor.SyncResponse{}, nil
+	case monitor.SyncActionUpsert:
+		result, err := pm.UpsertMonitor(req.Config, req.RunNow)
+		if err != nil {
+			return monitor.SyncResponse{}, err
+		}
+		if result == nil {
+			return monitor.SyncResponse{}, nil
+		}
+		return monitor.SyncResponse{Result: *result}, nil
+	case monitor.SyncActionDelete:
+		if req.Config.ID == "" {
+			return monitor.SyncResponse{}, errors.New("missing monitor ID for delete")
+		}
+		pm.DeleteMonitor(req.Config.ID)
+		return monitor.SyncResponse{}, nil
+	default:
+		return monitor.SyncResponse{}, fmt.Errorf("unknown monitor sync action: %d", req.Action)
+	}
+}
+
+// UpsertMonitor creates or replaces a single monitor task.
+func (pm *MonitorManager) UpsertMonitor(config monitor.Config, runNow bool) (*monitor.Result, error) {
+	if config.ID == "" {
+		return nil, errors.New("missing monitor ID")
+	}
+
+	pm.mu.Lock()
+	task, exists := pm.monitors[config.ID]
+	if exists && task.config == config {
+		pm.mu.Unlock()
+		if !runNow {
+			return nil, nil
+		}
+		return task.runProbe(pm.probe), nil
+	}
+	if exists {
+		task.cancel()
+	}
+	task = newMonitorTaskFromExisting(config, task)
+	task.resumeGuard = &pm.resumeGuard
+	pm.resumeGuard.start()
+	pm.monitors[config.ID] = task
+	pm.mu.Unlock()
+
+	if runNow {
+		result := task.runProbe(pm.probe)
+		pm.startMonitor(task)
+		return result, nil
+	}
+	pm.startMonitor(task)
+	return nil, nil
+}
+
+// DeleteMonitor stops and removes a single monitor task.
+func (pm *MonitorManager) DeleteMonitor(id string) {
+	if id == "" {
+		return
+	}
+	pm.mu.Lock()
+	defer pm.mu.Unlock()
+	if task, exists := pm.monitors[id]; exists {
+		task.cancel()
+		delete(pm.monitors, id)
+	}
+	if len(pm.monitors) == 0 {
+		pm.resumeGuard.shutdown()
+	}
+}
+
+// GetResults returns aggregated results for all monitors over the last supplied duration in ms.
+func (pm *MonitorManager) GetResults(durationMs uint16) map[string]monitor.Result {
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
+
+	results := make(map[string]monitor.Result, len(pm.monitors))
+	now := time.Now()
+	duration := time.Duration(durationMs) * time.Millisecond
+
+	for _, task := range pm.monitors {
+		result, ok := task.history.result(duration, now)
+
+		if !ok {
+			continue
+		}
+		results[task.config.ID] = result
+	}
+
+	return results
+}
+
+// Stop stops all monitor tasks.
+func (pm *MonitorManager) Stop() {
+	pm.mu.Lock()
+	defer pm.mu.Unlock()
+	for key, task := range pm.monitors {
+		task.cancel()
+		delete(pm.monitors, key)
+	}
+	pm.resumeGuard.shutdown()
+}

agent/network_monitor_history.go

@@ -0,0 +1,274 @@
+package agent
+
+import (
+	"math"
+	"sync"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+)
+
+// Monitors run at user-defined intervals (e.g., every 10s).
+// To keep memory usage low and constant, data is stored in two layers:
+// 1. Raw samples: The most recent individual results (kept for monitorRawRetention).
+// 2. Minute buckets: A ring buffer of 61 buckets, each representing one
+//    wall-clock minute. Samples collected within the same minute are aggregated
+//    (sum, min, max, count) into a single bucket.
+//
+// Short-term requests (<= 61s) use raw samples.
+// Long-term requests (up to 1h) use the minute buckets to avoid storing thousands
+// of individual data points.
+
+const (
+	// monitorRawRetention is the duration to keep individual samples
+	monitorRawRetention = 61 * time.Second
+	// monitorMinuteBucketLen is the number of 1-minute buckets to keep (1 hour + 1 for partials)
+	monitorMinuteBucketLen int32 = 61
+)
+
+// monitorHistory owns retention and aggregation, independently of probe execution.
+type monitorHistory struct {
+	mu          sync.Mutex
+	sampleCount int64
+	samples     []monitorSample
+	buckets     [monitorMinuteBucketLen]monitorBucket
+}
+
+func newMonitorHistory() *monitorHistory {
+	// Start small for typical intervals; append grows the buffer for faster probes.
+	return &monitorHistory{samples: make([]monitorSample, 0, 4)}
+}
+
+func (h *monitorHistory) clone() *monitorHistory {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	cloned := newMonitorHistory()
+	cloned.samples = append(cloned.samples, h.samples...)
+	cloned.buckets = h.buckets
+	cloned.sampleCount = h.sampleCount
+	return cloned
+}
+
+func (h *monitorHistory) result(duration time.Duration, now time.Time) (monitor.Result, bool) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	return h.resultLocked(duration, now)
+}
+
+func (h *monitorHistory) record(sample monitorSample) monitor.Result {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.addSampleLocked(sample)
+	result, _ := h.resultLocked(time.Minute, sample.timestamp)
+	return result
+}
+
+// monitorSample stores one monitor attempt and its collection time.
+type monitorSample struct {
+	responseUs int64 // -1 means loss
+	timestamp  time.Time
+}
+
+// monitorBucket stores one minute of aggregated monitor data.
+type monitorBucket struct {
+	minute int32
+	filled bool
+	stats  monitorAggregate
+}
+
+// monitorAggregate accumulates successful response stats and total sample counts.
+type monitorAggregate struct {
+	sumUs        int64
+	minUs        int64
+	maxUs        int64
+	totalCount   int64
+	successCount int64
+}
+
+// newMonitorAggregate initializes an aggregate with an unset minimum value.
+func newMonitorAggregate() monitorAggregate {
+	return monitorAggregate{minUs: math.MaxInt64}
+}
+
+// addResponse folds a single monitor sample into the aggregate.
+func (agg *monitorAggregate) addResponse(responseUs int64) {
+	agg.totalCount++
+	if responseUs < 0 {
+		return
+	}
+	agg.successCount++
+	agg.sumUs += responseUs
+	if responseUs < agg.minUs {
+		agg.minUs = responseUs
+	}
+	if responseUs > agg.maxUs {
+		agg.maxUs = responseUs
+	}
+}
+
+// addAggregate merges another aggregate into this one.
+func (agg *monitorAggregate) addAggregate(other monitorAggregate) {
+	if other.totalCount == 0 {
+		return
+	}
+	agg.totalCount += other.totalCount
+	agg.successCount += other.successCount
+	agg.sumUs += other.sumUs
+	if other.successCount == 0 {
+		return
+	}
+	if agg.minUs == math.MaxInt64 || other.minUs < agg.minUs {
+		agg.minUs = other.minUs
+	}
+	if other.maxUs > agg.maxUs {
+		agg.maxUs = other.maxUs
+	}
+}
+
+// hasData reports whether the aggregate contains any samples.
+func (agg monitorAggregate) hasData() bool {
+	return agg.totalCount > 0
+}
+
+// result converts the aggregate into the monitor result format.
+func (agg monitorAggregate) result() monitor.Result {
+	avg := agg.avgResponse()
+	result := monitor.Result{
+		AvgResponse:  avg,
+		MinResponse:  agg.minUs,
+		MaxResponse:  agg.maxUs,
+		PacketLoss:   agg.lossPercentage(),
+		TotalCount:   agg.totalCount,
+		SuccessCount: agg.successCount,
+		ResponseSum:  agg.sumUs,
+	}
+	if agg.successCount == 0 {
+		result.MinResponse, result.MaxResponse = 0, 0
+	}
+	return result
+}
+
+// avgResponse returns the rounded average of successful samples.
+func (agg monitorAggregate) avgResponse() int64 {
+	if agg.successCount == 0 {
+		return 0
+	}
+	return agg.sumUs / agg.successCount
+
+}
+
+// lossPercentage returns the rounded failure rate for the aggregate.
+func (agg monitorAggregate) lossPercentage() float64 {
+	if agg.totalCount == 0 {
+		return 0
+	}
+	return math.Round(float64(agg.totalCount-agg.successCount)/float64(agg.totalCount)*10000) / 100
+}
+
+// resultLocked returns the aggregated monitor result for the requested duration along with a bool indicating whether any data was available.
+func (h *monitorHistory) resultLocked(duration time.Duration, now time.Time) (monitor.Result, bool) {
+	agg := h.aggregateLocked(duration, now)
+	if !agg.hasData() {
+		// short realtime windows (e.g. the 1s window used for 1m/realtime charts) often fall
+		// between monitor samples since monitors run at longer, user-defined intervals; fall back to
+		// the most recent sample so realtime requests still report current status.
+		agg = h.latestSampleAggregateLocked()
+	}
+	hourAgg := h.aggregateLocked(time.Hour, now)
+	if !agg.hasData() {
+		return monitor.Result{}, false
+	}
+
+	result := agg.result()
+	if len(h.samples) > 0 {
+		result.LastProbeAt = h.samples[len(h.samples)-1].timestamp.UnixMilli()
+	}
+
+	result.AvgResponse1h = hourAgg.avgResponse()
+	result.MinResponse1h = hourAgg.minUs
+	result.MaxResponse1h = hourAgg.maxUs
+	result.PacketLoss1h = hourAgg.lossPercentage()
+	result.SampleCount = h.sampleCount
+
+	if hourAgg.successCount == 0 {
+		result.MinResponse1h, result.MaxResponse1h = 0, 0
+	}
+	return result, true
+}
+
+// latestSampleAggregateLocked returns an aggregate containing only the most recent sample, if any.
+func (h *monitorHistory) latestSampleAggregateLocked() monitorAggregate {
+	agg := newMonitorAggregate()
+	if len(h.samples) == 0 {
+		return agg
+	}
+	agg.addResponse(h.samples[len(h.samples)-1].responseUs)
+	return agg
+}
+
+// aggregateLocked collects monitor data for the requested time window.
+func (h *monitorHistory) aggregateLocked(duration time.Duration, now time.Time) monitorAggregate {
+	cutoff := now.Add(-duration)
+	// Keep short windows exact; longer windows read from minute buckets to avoid raw-sample retention.
+	if duration <= monitorRawRetention {
+		return aggregateSamplesSince(h.samples, cutoff)
+	}
+	return aggregateBucketsSince(h.buckets[:], cutoff, now)
+}
+
+// aggregateSamplesSince aggregates raw samples newer than the cutoff.
+func aggregateSamplesSince(samples []monitorSample, cutoff time.Time) monitorAggregate {
+	agg := newMonitorAggregate()
+	for _, sample := range samples {
+		if sample.timestamp.Before(cutoff) {
+			continue
+		}
+		agg.addResponse(sample.responseUs)
+	}
+	return agg
+}
+
+// aggregateBucketsSince aggregates minute buckets overlapping the requested window.
+func aggregateBucketsSince(buckets []monitorBucket, cutoff, now time.Time) monitorAggregate {
+	agg := newMonitorAggregate()
+	startMinute := int32(cutoff.Unix() / 60)
+	endMinute := int32(now.Unix() / 60)
+	for _, bucket := range buckets {
+		if !bucket.filled || bucket.minute < startMinute || bucket.minute > endMinute {
+			continue
+		}
+		agg.addAggregate(bucket.stats)
+	}
+	return agg
+}
+
+// addSampleLocked stores a fresh sample in both raw and per-minute retention buffers.
+func (h *monitorHistory) addSampleLocked(sample monitorSample) {
+	h.sampleCount++
+	cutoff := sample.timestamp.Add(-monitorRawRetention)
+	start := 0
+	for i := range h.samples {
+		if !h.samples[i].timestamp.Before(cutoff) {
+			start = i
+			break
+		}
+		if i == len(h.samples)-1 {
+			start = len(h.samples)
+		}
+	}
+	if start > 0 {
+		size := copy(h.samples, h.samples[start:])
+		h.samples = h.samples[:size]
+	}
+	h.samples = append(h.samples, sample)
+
+	minute := int32(sample.timestamp.Unix() / 60)
+	// Each slot stores one wall-clock minute, so the ring stays fixed-size at ~1h per monitor.
+	bucket := &h.buckets[minute%monitorMinuteBucketLen]
+	if !bucket.filled || bucket.minute != minute {
+		bucket.minute = minute
+		bucket.filled = true
+		bucket.stats = newMonitorAggregate()
+	}
+	bucket.stats.addResponse(sample.responseUs)
+}

agent/network_monitor_history_test.go

@@ -0,0 +1,154 @@
+package agent
+
+import (
+	"testing"
+	"time"
+
+	"github.com/fxamacker/cbor/v2"
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestMonitorHistoryWindowCounts(t *testing.T) {
+	history := newMonitorHistory()
+	now := time.Now()
+	// This older success counts toward lifetime warm-up, but not this window.
+	history.record(monitorSample{responseUs: 1000, timestamp: now.Add(-2 * time.Minute)})
+	history.record(monitorSample{responseUs: 10, timestamp: now.Add(-30 * time.Second)})
+	history.record(monitorSample{responseUs: 21, timestamp: now.Add(-20 * time.Second)})
+	history.record(monitorSample{responseUs: -1, timestamp: now.Add(-10 * time.Second)})
+	result, ok := history.result(time.Minute, now)
+	require.True(t, ok)
+	assert.EqualValues(t, 4, result.SampleCount)
+	assert.EqualValues(t, 3, result.TotalCount)
+	assert.EqualValues(t, 2, result.SuccessCount)
+	assert.EqualValues(t, 31, result.ResponseSum, "preserve the sum before average rounding")
+	assert.EqualValues(t, 15, result.AvgResponse)
+	assert.Equal(t, 33.33, result.PacketLoss)
+
+	encoded, err := cbor.Marshal(result)
+	require.NoError(t, err)
+	var decoded monitor.Result
+	require.NoError(t, cbor.Unmarshal(encoded, &decoded))
+	assert.Equal(t, result, decoded)
+	stats := monitor.Stats{}.FromResult(decoded)
+	assert.Equal(t, result.TotalCount, stats.TotalCount)
+	assert.Equal(t, result.SuccessCount, stats.SuccessCount)
+	assert.Equal(t, result.ResponseSum, stats.ResponseSum)
+
+	// Reads do not consume samples. A short window's latest-sample fallback
+	// carries the count for that single failure, not the minute or lifetime count.
+	repeated, _ := history.result(time.Minute, now)
+	assert.Equal(t, result, repeated)
+	fallback, ok := history.result(time.Second, now)
+	require.True(t, ok)
+	assert.EqualValues(t, 1, fallback.TotalCount)
+	assert.Zero(t, fallback.SuccessCount)
+	assert.Zero(t, fallback.ResponseSum)
+	assert.Equal(t, 100.0, fallback.PacketLoss)
+	assert.EqualValues(t, 4, fallback.SampleCount)
+}
+
+func TestMonitorHistoryAggregateLockedUsesRawSamplesForShortWindows(t *testing.T) {
+	now := time.Date(2026, time.April, 21, 12, 0, 0, 0, time.UTC)
+	history := newMonitorHistory()
+
+	history.addSampleLocked(monitorSample{responseUs: 10, timestamp: now.Add(-90 * time.Second)})
+	history.addSampleLocked(monitorSample{responseUs: 20, timestamp: now.Add(-30 * time.Second)})
+	history.addSampleLocked(monitorSample{responseUs: -1, timestamp: now.Add(-10 * time.Second)})
+
+	agg := history.aggregateLocked(time.Minute, now)
+	require.True(t, agg.hasData())
+	assert.Equal(t, int64(2), agg.totalCount)
+	assert.Equal(t, int64(1), agg.successCount)
+	result := agg.result()
+	assert.Equal(t, int64(20), result.AvgResponse)
+	assert.Equal(t, int64(20), result.MinResponse)
+	assert.Equal(t, int64(20), result.MaxResponse)
+	assert.Equal(t, 50.0, result.PacketLoss)
+}
+
+func TestMonitorHistoryAggregateLockedUsesMinuteBucketsForLongWindows(t *testing.T) {
+	now := time.Date(2026, time.April, 21, 12, 0, 30, 0, time.UTC)
+	history := newMonitorHistory()
+
+	history.addSampleLocked(monitorSample{responseUs: 10, timestamp: now.Add(-11 * time.Minute)})
+	history.addSampleLocked(monitorSample{responseUs: 20, timestamp: now.Add(-9 * time.Minute)})
+	history.addSampleLocked(monitorSample{responseUs: 40, timestamp: now.Add(-5 * time.Minute)})
+	history.addSampleLocked(monitorSample{responseUs: -1, timestamp: now.Add(-90 * time.Second)})
+	history.addSampleLocked(monitorSample{responseUs: 30, timestamp: now.Add(-30 * time.Second)})
+
+	agg := history.aggregateLocked(10*time.Minute, now)
+	require.True(t, agg.hasData())
+	assert.Equal(t, int64(4), agg.totalCount)
+	assert.Equal(t, int64(3), agg.successCount)
+	result := agg.result()
+	assert.Equal(t, int64(30), result.AvgResponse)
+	assert.Equal(t, int64(20), result.MinResponse)
+	assert.Equal(t, int64(40), result.MaxResponse)
+	assert.Equal(t, 25.0, result.PacketLoss)
+}
+
+func TestMonitorHistoryAddSampleLockedTrimsRawSamplesButKeepsBucketHistory(t *testing.T) {
+	now := time.Date(2026, time.April, 21, 12, 0, 0, 0, time.UTC)
+	history := newMonitorHistory()
+
+	history.addSampleLocked(monitorSample{responseUs: 10, timestamp: now.Add(-10 * time.Minute)})
+	history.addSampleLocked(monitorSample{responseUs: 20, timestamp: now})
+
+	require.Len(t, history.samples, 1)
+	assert.Equal(t, int64(20), history.samples[0].responseUs)
+
+	agg := history.aggregateLocked(10*time.Minute, now)
+	require.True(t, agg.hasData())
+	assert.Equal(t, int64(2), agg.totalCount)
+	assert.Equal(t, int64(2), agg.successCount)
+	result := agg.result()
+	assert.Equal(t, int64(15), result.AvgResponse)
+	assert.Equal(t, int64(10), result.MinResponse)
+	assert.Equal(t, int64(20), result.MaxResponse)
+	assert.Equal(t, 0.0, result.PacketLoss)
+}
+
+func TestMonitorHistoryProbeTimestamp(t *testing.T) {
+	history := newMonitorHistory()
+	start := time.Date(2026, time.September, 14, 12, 0, 0, 0, time.UTC)
+	_, ok := history.result(time.Minute, start)
+	require.False(t, ok)
+	first := history.record(monitorSample{responseUs: 20, timestamp: start})
+	assert.Equal(t, start.UnixMilli(), first.LastProbeAt)
+	for minute := 0; minute < 5; minute++ {
+		now := start.Add(time.Duration(minute)*time.Minute + time.Second)
+		// Realtime reads must not consume freshness for the persistence request.
+		for _, window := range []time.Duration{time.Second, time.Minute} {
+			result, ok := history.result(window, now)
+			require.True(t, ok)
+			assert.Equal(t, first.LastProbeAt, result.LastProbeAt)
+			assert.Equal(t, int64(20), result.AvgResponse)
+		}
+	}
+	next := start.Add(5 * time.Minute)
+	failed := history.record(monitorSample{responseUs: -1, timestamp: next})
+	assert.Equal(t, next.UnixMilli(), failed.LastProbeAt)
+	assert.Equal(t, float64(100), failed.PacketLoss)
+	repeated, ok := history.result(time.Minute, next.Add(2*time.Minute))
+	require.True(t, ok)
+	assert.Equal(t, failed.LastProbeAt, repeated.LastProbeAt)
+	assert.Equal(t, float64(100), repeated.PacketLoss)
+}
+
+func TestMonitorHistorySampleCount(t *testing.T) {
+	history := newMonitorHistory()
+	now := time.Now()
+	// Both failed and successful probes count, including older samples so
+	// monitors with hourly intervals can finish warming up.
+	history.record(monitorSample{responseUs: -1, timestamp: now.Add(-2 * time.Hour)})
+	for i, response := range []int64{10, -1, 20} {
+		result := history.record(monitorSample{responseUs: response, timestamp: now.Add(time.Duration(i) * time.Second)})
+		assert.EqualValues(t, i+2, result.SampleCount)
+	}
+	result, ok := history.clone().result(time.Minute, now.Add(3*time.Second))
+	require.True(t, ok)
+	assert.EqualValues(t, 4, result.SampleCount)
+}

agent/network_monitor_ping.go

@@ -0,0 +1,312 @@
+package agent
+
+import (
+	"bytes"
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"math"
+	"net"
+	"os"
+	"os/exec"
+	"regexp"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"golang.org/x/net/icmp"
+	"golang.org/x/net/ipv4"
+	"golang.org/x/net/ipv6"
+
+	"log/slog"
+)
+
+// Match the numeric RTT independently of the localized label used by Windows.
+var pingTimeRegex = regexp.MustCompile(`(?i)[=<]\s*([0-9]+(?:[.,][0-9]+)?)\s*ms\b`)
+
+var icmpSequence atomic.Uint32
+
+type icmpPacketConn interface {
+	Close() error
+}
+
+// icmpMethod tracks which ICMP approach to use. Once a method succeeds or
+// all native methods fail, the choice is cached so subsequent monitors skip
+// the trial-and-error overhead.
+type icmpMethod uint8
+
+const (
+	icmpUntried      icmpMethod = iota // haven't tried yet
+	icmpRaw                            // privileged raw socket
+	icmpDatagram                       // unprivileged datagram socket
+	icmpExecFallback                   // shell out to system ping command
+)
+
+// icmpFamily holds the network parameters and cached detection result for one address family.
+type icmpFamily struct {
+	rawNetwork   string    // e.g. "ip4:icmp" or "ip6:ipv6-icmp"
+	dgramNetwork string    // e.g. "udp4" or "udp6"
+	listenAddr   string    // "0.0.0.0" or "::"
+	echoType     icmp.Type // outgoing echo request type
+	replyType    icmp.Type // expected echo reply type
+	proto        int       // IANA protocol number for parsing replies
+	isIPv6       bool
+	mode         icmpMethod // cached detection result (guarded by icmpModeMu)
+}
+
+var (
+	icmpV4 = icmpFamily{
+		rawNetwork:   "ip4:icmp",
+		dgramNetwork: "udp4",
+		listenAddr:   "0.0.0.0",
+		echoType:     ipv4.ICMPTypeEcho,
+		replyType:    ipv4.ICMPTypeEchoReply,
+		proto:        1,
+	}
+	icmpV6 = icmpFamily{
+		rawNetwork:   "ip6:ipv6-icmp",
+		dgramNetwork: "udp6",
+		listenAddr:   "::",
+		echoType:     ipv6.ICMPTypeEchoRequest,
+		replyType:    ipv6.ICMPTypeEchoReply,
+		proto:        58,
+		isIPv6:       true,
+	}
+	icmpModeMu sync.Mutex
+	icmpListen = func(network, listenAddr string) (icmpPacketConn, error) {
+		return icmp.ListenPacket(network, listenAddr)
+	}
+)
+
+// monitorICMP sends an ICMP echo request and measures round-trip response.
+// Supports both IPv4 and IPv6 targets. The ICMP method (raw socket,
+// unprivileged datagram, or exec fallback) is detected once per address
+// family and cached for subsequent monitors.
+// Returns response in microseconds, or -1 and an error on failure.
+func monitorICMP(ctx context.Context, target string) (int64, error) {
+	ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+	defer cancel()
+
+	family, ip, err := resolveICMPTarget(ctx, target)
+	if err != nil {
+		return -1, err
+	}
+
+	icmpModeMu.Lock()
+	if family.mode == icmpUntried {
+		family.mode = detectICMPMode(family, icmpListen)
+	}
+	mode := family.mode
+	icmpModeMu.Unlock()
+
+	switch mode {
+	case icmpRaw:
+		return monitorICMPNative(ctx, family.rawNetwork, family, &net.IPAddr{IP: ip})
+	case icmpDatagram:
+		return monitorICMPNative(ctx, family.dgramNetwork, family, &net.UDPAddr{IP: ip})
+	case icmpExecFallback:
+		return monitorICMPExec(ctx, ip.String(), family.isIPv6)
+	default:
+		return -1, errors.New("unsupported ICMP mode")
+	}
+}
+
+// resolveICMPTarget resolves a target hostname or IP to determine the address
+// family and concrete IP address. Prefers IPv4 for dual-stack hostnames.
+func resolveICMPTarget(ctx context.Context, target string) (*icmpFamily, net.IP, error) {
+	if ip := net.ParseIP(target); ip != nil {
+		if ip.To4() != nil {
+			return &icmpV4, ip.To4(), nil
+		}
+		return &icmpV6, ip, nil
+	}
+
+	ips, err := net.DefaultResolver.LookupIP(ctx, "ip", target)
+	if err != nil || len(ips) == 0 {
+		return nil, nil, err
+	}
+	for _, ip := range ips {
+		if v4 := ip.To4(); v4 != nil {
+			return &icmpV4, v4, nil
+		}
+	}
+	return &icmpV6, ips[0], nil
+}
+
+func detectICMPMode(family *icmpFamily, listen func(network, listenAddr string) (icmpPacketConn, error)) icmpMethod {
+	label := "IPv4"
+	if family.isIPv6 {
+		label = "IPv6"
+	}
+
+	conn, err := listen(family.rawNetwork, family.listenAddr)
+	slog.Debug("ICMP raw socket test", "family", label, "err", err)
+	if err == nil {
+		conn.Close()
+		return icmpRaw
+	}
+
+	conn, err = listen(family.dgramNetwork, family.listenAddr)
+	slog.Debug("ICMP datagram socket test", "family", label, "err", err)
+	if err == nil {
+		conn.Close()
+		return icmpDatagram
+	}
+
+	return icmpExecFallback
+}
+
+// monitorICMPNative sends an ICMP echo request using Go's x/net/icmp package.
+func monitorICMPNative(ctx context.Context, network string, family *icmpFamily, dst net.Addr) (int64, error) {
+	conn, err := icmp.ListenPacket(network, family.listenAddr)
+	if err != nil {
+		return -1, err
+	}
+	defer conn.Close()
+
+	return monitorICMPPacket(ctx, conn, family, dst)
+}
+
+func monitorICMPPacket(ctx context.Context, conn net.PacketConn, family *icmpFamily, dst net.Addr) (int64, error) {
+	if err := ctx.Err(); err != nil {
+		return -1, err
+	}
+	// Closing the socket interrupts both reads and writes on cancellation.
+	stop := context.AfterFunc(ctx, func() { _ = conn.Close() })
+	defer stop()
+
+	// Prepare correlation data before starting the round-trip timer. The token
+	// also distinguishes delayed replies after the 16-bit sequence wraps.
+	token := make([]byte, 16)
+	if _, err := rand.Read(token); err != nil {
+		return -1, err
+	}
+	echo := &icmp.Echo{
+		ID:   os.Getpid() & 0xffff,
+		Seq:  int(icmpSequence.Add(1) & 0xffff),
+		Data: token,
+	}
+	// Linux ping sockets replace the Echo ID with their bound port. Darwin
+	// datagram sockets and raw sockets preserve the supplied ID.
+	if local, ok := conn.LocalAddr().(*net.UDPAddr); ok && runtime.GOOS == "linux" {
+		echo.ID = local.Port
+	}
+	targetIP := icmpAddrIP(dst)
+	msg := &icmp.Message{
+		Type: family.echoType,
+		Code: 0,
+		Body: echo,
+	}
+	msgBytes, err := msg.Marshal(nil)
+	if err != nil {
+		return -1, err
+	}
+
+	// Set deadline before sending
+	if err := conn.SetDeadline(time.Now().Add(3 * time.Second)); err != nil {
+		return -1, err
+	}
+
+	buf := make([]byte, 1500)
+	start := time.Now()
+	if _, err := conn.WriteTo(msgBytes, dst); err != nil {
+		return -1, err
+	}
+
+	// Read reply
+	for {
+		n, peer, err := conn.ReadFrom(buf)
+		received := time.Now()
+		if err != nil {
+			return -1, err
+		}
+		if !targetIP.Equal(icmpAddrIP(peer)) {
+			continue
+		}
+
+		reply, err := icmp.ParseMessage(family.proto, buf[:n])
+		if err != nil || reply.Type != family.replyType || reply.Code != 0 {
+			continue
+		}
+
+		body, ok := reply.Body.(*icmp.Echo)
+		if ok && body.ID == echo.ID && body.Seq == echo.Seq && bytes.Equal(body.Data, echo.Data) {
+			return received.Sub(start).Microseconds(), nil
+		}
+		// Keep waiting for our reply without extending the original deadline.
+	}
+}
+
+func icmpAddrIP(addr net.Addr) net.IP {
+	switch addr := addr.(type) {
+	case *net.IPAddr:
+		return addr.IP
+	case *net.UDPAddr:
+		return addr.IP
+	default:
+		return nil
+	}
+}
+
+// pingCommand selects the executable and arguments for the supported agent platforms.
+// The context deadline enforces the timeout: -W has incompatible meanings across
+// Linux, BSD IPv4 ping, and macOS ping6.
+func pingCommand(goos, target string, isIPv6 bool) (string, []string, error) {
+	family := "-4"
+	if isIPv6 {
+		family = "-6"
+	}
+	switch goos {
+	case "windows":
+		return "ping", []string{family, "-n", "1", "-w", "3000", target}, nil
+	case "linux":
+		return "ping", []string{family, "-n", "-c", "1", target}, nil
+	case "darwin", "freebsd", "openbsd":
+		command := "ping"
+		if isIPv6 {
+			command = "ping6"
+		}
+		return command, []string{"-n", "-c", "1", target}, nil
+	default:
+		return "", nil, fmt.Errorf("ping fallback is unsupported on %s", goos)
+	}
+}
+
+// monitorICMPExec falls back to the system ping command. Returns -1 and an error on failure.
+func monitorICMPExec(ctx context.Context, target string, isIPv6 bool) (int64, error) {
+	ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+	defer cancel()
+	name, args, err := pingCommand(runtime.GOOS, target, isIPv6)
+	if err != nil {
+		return -1, err
+	}
+	cmd := exec.CommandContext(ctx, name, args...)
+	// Keep Unix output and decimal formatting stable. Windows ignores LC_ALL.
+	cmd.Env = append(os.Environ(), "LC_ALL=C")
+	output, err := cmd.Output()
+	if ctx.Err() != nil {
+		return -1, ctx.Err()
+	}
+	if err != nil {
+		return -1, fmt.Errorf("%s failed: %w", name, err)
+	}
+	return parsePingResponse(output)
+}
+
+// parsePingResponse returns the reported RTT, never subprocess execution time.
+// For a bounded value such as Windows' time<1ms, retain the reported upper bound.
+func parsePingResponse(output []byte) (int64, error) {
+	matches := pingTimeRegex.FindSubmatch(output)
+	if len(matches) < 2 {
+		return -1, errors.New("ping output contains no round-trip time")
+	}
+	ms, err := strconv.ParseFloat(strings.ReplaceAll(string(matches[1]), ",", "."), 64)
+	if err != nil || math.IsInf(ms, 0) || ms >= float64(math.MaxInt64)/1000 {
+		return -1, errors.New("invalid round-trip time in ping output")
+	}
+	return int64(math.Round(ms * 1000)), nil
+}

agent/network_monitor_ping_test.go

@@ -0,0 +1,433 @@
+//go:build testing
+
+package agent
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/net/icmp"
+)
+
+type testICMPPacketConn struct{}
+
+func (testICMPPacketConn) Close() error { return nil }
+
+type blockingICMPConn struct {
+	net.PacketConn
+	reading chan struct{}
+}
+
+func (c *blockingICMPConn) WriteTo(p []byte, addr net.Addr) (int, error) {
+	return len(p), nil
+}
+
+func (c *blockingICMPConn) ReadFrom(p []byte) (int, net.Addr, error) {
+	close(c.reading)
+	return c.PacketConn.ReadFrom(p)
+}
+
+func TestMonitorICMPPacketCancellation(t *testing.T) {
+	conn, err := net.ListenPacket("udp4", "127.0.0.1:0")
+	require.NoError(t, err)
+	defer conn.Close()
+	blocking := &blockingICMPConn{PacketConn: conn, reading: make(chan struct{})}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	done := make(chan error, 1)
+	go func() {
+		_, err := monitorICMPPacket(ctx, blocking, &icmpV4, conn.LocalAddr())
+		done <- err
+	}()
+	select {
+	case <-blocking.reading:
+	case <-time.After(time.Second):
+		t.Fatal("probe did not begin reading")
+	}
+	cancel()
+	select {
+	case err := <-done:
+		require.Error(t, err)
+	case <-time.After(time.Second):
+		t.Fatal("cancellation did not interrupt the socket read")
+	}
+}
+
+func TestMonitorICMPExecCancellation(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("test uses a POSIX shell stub for ping")
+	}
+	dir := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "ping"), []byte("#!/bin/sh\nexec sleep 30\n"), 0o755))
+	t.Setenv("PATH", dir+string(os.PathListSeparator)+os.Getenv("PATH"))
+	ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer cancel()
+	done := make(chan error, 1)
+	go func() {
+		_, err := monitorICMPExec(ctx, "127.0.0.1", false)
+		done <- err
+	}()
+	select {
+	case err := <-done:
+		require.ErrorIs(t, err, context.DeadlineExceeded)
+	case <-time.After(time.Second):
+		t.Fatal("cancellation did not terminate ping")
+	}
+}
+
+func TestPingCommand(t *testing.T) {
+	for _, goos := range []string{"linux", "windows", "darwin", "freebsd", "openbsd"} {
+		for _, ipv6 := range []bool{false, true} {
+			t.Run(fmt.Sprintf("%s/ipv6=%t", goos, ipv6), func(t *testing.T) {
+				target, family := "192.0.2.1", "-4"
+				if ipv6 {
+					target, family = "2001:db8::1", "-6"
+				}
+				name, args, err := pingCommand(goos, target, ipv6)
+				require.NoError(t, err)
+				wantName := "ping"
+				wantArgs := []string{"-n", "-c", "1", target}
+				switch goos {
+				case "windows":
+					wantArgs = []string{family, "-n", "1", "-w", "3000", target}
+				case "linux":
+					wantArgs = append([]string{family}, wantArgs...)
+				default:
+					if ipv6 {
+						wantName = "ping6"
+					}
+				}
+				assert.Equal(t, wantName, name)
+				assert.Equal(t, wantArgs, args)
+			})
+		}
+	}
+	_, _, err := pingCommand("unsupported", "192.0.2.1", false)
+	require.Error(t, err)
+}
+
+func TestParsePingResponse(t *testing.T) {
+	for _, tc := range []struct {
+		name   string
+		output string
+		wantUs int64
+	}{
+		{"linux", "64 bytes from 192.0.2.1: icmp_seq=1 ttl=64 time=12.345 ms", 12345},
+		{"bsd", "64 bytes from 192.0.2.1: icmp_seq=0 ttl=64 time=0.023 ms", 23},
+		{"ipv6", "64 bytes from 2001:db8::1: icmp_seq=0 hlim=64 time=1.234 ms", 1234},
+		{"windows", "Reply from 192.0.2.1: bytes=32 time=12ms TTL=128", 12000},
+		{"windows submillisecond", "Reply from ::1: time<1ms", 1000},
+		{"localized windows", "Antwort von 192.0.2.1: Bytes=32 Zeit=12ms TTL=128", 12000},
+		{"decimal comma", "64 bytes from 192.0.2.1: time=1,234 ms", 1234},
+		{"rounding", "time=0.1236 ms", 124},
+		{"empty", "", -1},
+		{"timeout", "Request timed out.", -1},
+		{"unreachable", "Reply from 192.0.2.1: Destination host unreachable.", -1},
+		{"malformed", "time=oops ms", -1},
+		{"negative", "time=-1 ms", -1},
+		{"overflow", "time=999999999999999999999 ms", -1},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			responseUs, err := parsePingResponse([]byte(tc.output))
+			if tc.wantUs < 0 {
+				require.Error(t, err)
+			} else {
+				require.NoError(t, err)
+			}
+			assert.Equal(t, tc.wantUs, responseUs)
+		})
+	}
+}
+
+func TestMonitorICMPExecOutput(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("test uses a POSIX shell stub for ping")
+	}
+	for _, tc := range []struct {
+		name   string
+		output string
+		exit   int
+		wantUs int64
+	}{
+		{"success", "time=1.234 ms", 0, 1234},
+		{"missing RTT", "unrecognized output", 0, -1},
+		{"failed command with RTT", "time=1.234 ms", 1, -1},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			dir := t.TempDir()
+			// Also verify an inherited locale cannot override the C locale.
+			script := fmt.Sprintf("#!/bin/sh\n[ \"$LC_ALL\" = C ] || exit 2\nprintf '%%s\\n' '%s'\nexit %d\n", tc.output, tc.exit)
+			require.NoError(t, os.WriteFile(filepath.Join(dir, "ping"), []byte(script), 0o755))
+			t.Setenv("PATH", dir+string(os.PathListSeparator)+os.Getenv("PATH"))
+			t.Setenv("LC_ALL", "de_DE.UTF-8")
+			responseUs, err := monitorICMPExec(t.Context(), "127.0.0.1", false)
+			if tc.wantUs < 0 {
+				require.Error(t, err)
+			} else {
+				require.NoError(t, err)
+			}
+			assert.Equal(t, tc.wantUs, responseUs)
+		})
+	}
+}
+
+type icmpTestReply struct {
+	data []byte
+	peer net.Addr
+}
+
+type scriptedICMPConn struct {
+	net.PacketConn
+	local        net.Addr
+	onWrite      func([]byte, net.Addr)
+	replies      []icmpTestReply
+	reads        int
+	deadlineSets int
+}
+
+func (c *scriptedICMPConn) LocalAddr() net.Addr { return c.local }
+
+func (c *scriptedICMPConn) SetDeadline(deadline time.Time) error {
+	c.deadlineSets++
+	return nil
+}
+
+func (c *scriptedICMPConn) WriteTo(data []byte, dst net.Addr) (int, error) {
+	c.onWrite(data, dst)
+	return len(data), nil
+}
+
+func (c *scriptedICMPConn) ReadFrom(buf []byte) (int, net.Addr, error) {
+	c.reads++
+	if len(c.replies) == 0 {
+		return 0, nil, os.ErrDeadlineExceeded
+	}
+	reply := c.replies[0]
+	c.replies = c.replies[1:]
+	return copy(buf, reply.data), reply.peer, nil
+}
+
+func TestMonitorICMPReplyCorrelation(t *testing.T) {
+	for _, family := range []*icmpFamily{&icmpV4, &icmpV6} {
+		for _, datagram := range []bool{false, true} {
+			network := family.rawNetwork
+			ip, other := net.ParseIP("192.0.2.1"), net.ParseIP("192.0.2.2")
+			if family.isIPv6 {
+				ip, other = net.ParseIP("2001:db8::1"), net.ParseIP("2001:db8::2")
+			}
+			var dst net.Addr = &net.IPAddr{IP: ip}
+			var wrongPeer net.Addr = &net.IPAddr{IP: other}
+			if datagram {
+				network = family.dgramNetwork
+				dst = &net.UDPAddr{IP: ip}
+				wrongPeer = &net.UDPAddr{IP: other}
+			}
+			for _, mismatch := range []string{"source", "id", "sequence", "payload", "type", "code", "malformed"} {
+				for _, eventuallyMatches := range []bool{false, true} {
+					ending := "timeout"
+					if eventuallyMatches {
+						ending = "success"
+					}
+					t.Run(network+"/"+mismatch+"/"+ending, func(t *testing.T) {
+						conn := &scriptedICMPConn{local: &net.IPAddr{IP: net.IPv4zero}}
+						if datagram {
+							conn.local = &net.UDPAddr{Port: 12345}
+							if runtime.GOOS == "linux" {
+								// Deliberately differ from the process ID.
+								conn.local = &net.UDPAddr{Port: (os.Getpid() % 65534) + 1}
+							}
+						}
+						conn.onWrite = func(data []byte, target net.Addr) {
+							require.Equal(t, dst, target)
+							request, err := icmp.ParseMessage(family.proto, data)
+							require.NoError(t, err)
+							echo := request.Body.(*icmp.Echo)
+							expectedID := os.Getpid() & 0xffff
+							if datagram && runtime.GOOS == "linux" {
+								expectedID = conn.local.(*net.UDPAddr).Port
+							}
+							require.Equal(t, expectedID, echo.ID)
+							reply := &icmp.Message{Type: family.replyType, Body: echo}
+							valid, err := reply.Marshal(nil)
+							require.NoError(t, err)
+							peer := dst
+							switch mismatch {
+							case "source":
+								peer = wrongPeer
+							case "id":
+								echo.ID ^= 1
+							case "sequence":
+								echo.Seq ^= 1
+							case "payload":
+								echo.Data[0] ^= 1
+							case "type":
+								reply.Type = family.echoType
+							case "code":
+								reply.Code = 1
+							}
+							invalid, err := reply.Marshal(nil)
+							require.NoError(t, err)
+							if mismatch == "malformed" {
+								invalid = invalid[:2]
+							}
+							conn.replies = []icmpTestReply{{invalid, peer}}
+							if eventuallyMatches {
+								conn.replies = append(conn.replies, icmpTestReply{valid, dst})
+							}
+						}
+						elapsed, err := monitorICMPPacket(context.Background(), conn, family, dst)
+						if eventuallyMatches {
+							require.NoError(t, err)
+							assert.GreaterOrEqual(t, elapsed, int64(0))
+						} else {
+							require.ErrorIs(t, err, os.ErrDeadlineExceeded)
+							assert.Equal(t, int64(-1), elapsed)
+						}
+						assert.Equal(t, 2, conn.reads)
+						assert.Equal(t, 1, conn.deadlineSets)
+					})
+				}
+			}
+		}
+	}
+}
+
+func TestMonitorICMPLoopback(t *testing.T) {
+	for _, family := range []*icmpFamily{&icmpV4, &icmpV6} {
+		for _, network := range []string{family.rawNetwork, family.dgramNetwork} {
+			t.Run(network, func(t *testing.T) {
+				conn, err := icmp.ListenPacket(network, family.listenAddr)
+				if err != nil {
+					t.Skipf("ICMP socket unavailable: %v", err)
+				}
+				defer conn.Close()
+				ip := net.ParseIP("127.0.0.1")
+				if family.isIPv6 {
+					ip = net.ParseIP("::1")
+				}
+				var dst net.Addr = &net.IPAddr{IP: ip}
+				if network == family.dgramNetwork {
+					dst = &net.UDPAddr{IP: ip}
+				}
+				elapsed, err := monitorICMPPacket(context.Background(), conn, family, dst)
+				require.NoError(t, err)
+				assert.GreaterOrEqual(t, elapsed, int64(0))
+			})
+		}
+	}
+}
+
+func TestDetectICMPMode(t *testing.T) {
+	tests := []struct {
+		name         string
+		family       *icmpFamily
+		rawErr       error
+		udpErr       error
+		want         icmpMethod
+		wantNetworks []string
+	}{
+		{
+			name:         "IPv4 prefers raw socket when available",
+			family:       &icmpV4,
+			want:         icmpRaw,
+			wantNetworks: []string{"ip4:icmp"},
+		},
+		{
+			name:         "IPv4 uses datagram when raw unavailable",
+			family:       &icmpV4,
+			rawErr:       errors.New("operation not permitted"),
+			want:         icmpDatagram,
+			wantNetworks: []string{"ip4:icmp", "udp4"},
+		},
+		{
+			name:         "IPv4 falls back to exec when both unavailable",
+			family:       &icmpV4,
+			rawErr:       errors.New("operation not permitted"),
+			udpErr:       errors.New("protocol not supported"),
+			want:         icmpExecFallback,
+			wantNetworks: []string{"ip4:icmp", "udp4"},
+		},
+		{
+			name:         "IPv6 prefers raw socket when available",
+			family:       &icmpV6,
+			want:         icmpRaw,
+			wantNetworks: []string{"ip6:ipv6-icmp"},
+		},
+		{
+			name:         "IPv6 uses datagram when raw unavailable",
+			family:       &icmpV6,
+			rawErr:       errors.New("operation not permitted"),
+			want:         icmpDatagram,
+			wantNetworks: []string{"ip6:ipv6-icmp", "udp6"},
+		},
+		{
+			name:         "IPv6 falls back to exec when both unavailable",
+			family:       &icmpV6,
+			rawErr:       errors.New("operation not permitted"),
+			udpErr:       errors.New("protocol not supported"),
+			want:         icmpExecFallback,
+			wantNetworks: []string{"ip6:ipv6-icmp", "udp6"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			calls := make([]string, 0, 2)
+			listen := func(network, listenAddr string) (icmpPacketConn, error) {
+				require.Equal(t, tt.family.listenAddr, listenAddr)
+				calls = append(calls, network)
+				switch network {
+				case tt.family.rawNetwork:
+					if tt.rawErr != nil {
+						return nil, tt.rawErr
+					}
+				case tt.family.dgramNetwork:
+					if tt.udpErr != nil {
+						return nil, tt.udpErr
+					}
+				default:
+					t.Fatalf("unexpected network %q", network)
+				}
+				return testICMPPacketConn{}, nil
+			}
+
+			assert.Equal(t, tt.want, detectICMPMode(tt.family, listen))
+			assert.Equal(t, tt.wantNetworks, calls)
+		})
+	}
+}
+
+func TestResolveICMPTarget(t *testing.T) {
+	t.Run("IPv4 literal", func(t *testing.T) {
+		family, ip, err := resolveICMPTarget(context.Background(), "127.0.0.1")
+		require.NoError(t, err)
+		require.NotNil(t, family)
+		assert.False(t, family.isIPv6)
+		assert.Equal(t, "127.0.0.1", ip.String())
+	})
+
+	t.Run("IPv6 literal", func(t *testing.T) {
+		family, ip, err := resolveICMPTarget(context.Background(), "::1")
+		require.NoError(t, err)
+		require.NotNil(t, family)
+		assert.True(t, family.isIPv6)
+		assert.Equal(t, "::1", ip.String())
+	})
+
+	t.Run("IPv4-mapped IPv6 resolves as IPv4", func(t *testing.T) {
+		family, ip, err := resolveICMPTarget(context.Background(), "::ffff:127.0.0.1")
+		require.NoError(t, err)
+		require.NotNil(t, family)
+		assert.False(t, family.isIPv6)
+		assert.Equal(t, "127.0.0.1", ip.String())
+	})
+}

agent/network_monitor_probe.go

@@ -0,0 +1,105 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"net/http"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+)
+
+// monitorProbe performs one check. Errors are recorded as loss by the task runner.
+// Implementations must honor cancellation and bound their execution time.
+type monitorProbe func(context.Context, monitor.Config) (int64, error)
+
+func networkMonitorProbe(client *http.Client) monitorProbe {
+	return func(ctx context.Context, config monitor.Config) (int64, error) {
+		switch config.Protocol {
+		case "icmp":
+			return monitorICMP(ctx, config.Target)
+		case "tcp":
+			return monitorTCP(ctx, config.Target, config.Port)
+		case "http":
+			return monitorHTTP(ctx, client, config.Target)
+		case "dns":
+			return monitorDNS(ctx, config.Target)
+		default:
+			return -1, fmt.Errorf("unknown monitor protocol: %s", config.Protocol)
+		}
+	}
+}
+
+// monitorTCP measures connection establishment time, including address fallback
+// but excluding DNS resolution.
+// Returns -1 and an error on failure.
+func monitorTCP(ctx context.Context, target string, port uint16) (int64, error) {
+	ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+	defer cancel()
+
+	// Resolve DNS first, outside the timing window but within the probe deadline.
+	ips, err := net.DefaultResolver.LookupHost(ctx, target)
+	if err != nil {
+		return -1, err
+	}
+	if len(ips) == 0 {
+		return -1, errors.New("no addresses resolved for TCP monitor")
+	}
+	portString := fmt.Sprintf("%d", port)
+	deadline, _ := ctx.Deadline()
+
+	// Share the remaining probe budget across addresses so an unresponsive
+	// first address cannot consume all the time available for alternatives.
+	start := time.Now()
+	for i, ip := range ips {
+		if err := ctx.Err(); err != nil {
+			return -1, err
+		}
+		dialer := net.Dialer{Timeout: time.Until(deadline) / time.Duration(len(ips)-i)}
+		var conn net.Conn
+		conn, err = dialer.DialContext(ctx, "tcp", net.JoinHostPort(ip, portString))
+		if err != nil {
+			continue
+		}
+		responseUs := time.Since(start).Microseconds()
+		conn.Close()
+		return responseUs, nil
+	}
+	return -1, err
+}
+
+// monitorDNS measures DNS resolution response time in microseconds. Returns -1 and an error on failure.
+func monitorDNS(ctx context.Context, target string) (int64, error) {
+	ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	ips, err := net.DefaultResolver.LookupHost(ctx, target)
+	if err != nil || len(ips) == 0 {
+		return -1, err
+	}
+	return time.Since(start).Microseconds(), nil
+}
+
+// monitorHTTP measures HTTP GET request response in microseconds. Returns -1 and an error on failure.
+func monitorHTTP(ctx context.Context, client *http.Client, url string) (int64, error) {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	start := time.Now()
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return -1, err
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		return -1, err
+	}
+	resp.Body.Close()
+	if resp.StatusCode >= 400 {
+		return -1, fmt.Errorf("HTTP error: %s", resp.Status)
+	}
+	return time.Since(start).Microseconds(), nil
+}

agent/network_monitor_resume.go

@@ -0,0 +1,88 @@
+package agent
+
+import (
+	"sync"
+	"time"
+)
+
+const (
+	monitorResumeHeartbeat = 10 * time.Second
+	// Allow scheduling jitter without mistaking an ordinary tick for resume.
+	monitorResumeGap   = 2 * monitorResumeHeartbeat
+	monitorResumePause = 10 * time.Second
+)
+
+// monitorResumeGuard detects likely suspend/resume using wall time. A long
+// process stall or forward clock adjustment can also trigger the bounded pause.
+// One heartbeat is shared by all configured monitors.
+type monitorResumeGuard struct {
+	mu         sync.Mutex
+	stop       chan struct{}
+	lastTick   time.Time
+	pauseUntil time.Time
+	generation uint32
+}
+
+func (g *monitorResumeGuard) start() {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if g.stop != nil {
+		return
+	}
+	stop := make(chan struct{})
+	g.stop = stop
+	g.lastTick = time.Now().Round(0)
+	g.pauseUntil = time.Time{}
+	go func() {
+		ticker := time.NewTicker(monitorResumeHeartbeat)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-stop:
+				return
+			case <-ticker.C:
+				g.mu.Lock()
+				if g.stop == stop {
+					g.observe(time.Now())
+				}
+				g.mu.Unlock()
+			}
+		}
+	}()
+}
+
+func (g *monitorResumeGuard) shutdown() {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if g.stop != nil {
+		close(g.stop)
+		g.stop = nil
+		g.generation++
+	}
+}
+
+// observe requires mu. Strip the monotonic component because it can stop during
+// suspend. Read the current time rather than the ticker's queued timestamp.
+func (g *monitorResumeGuard) observe(now time.Time) {
+	now = now.Round(0)
+	if now.Sub(g.lastTick) > monitorResumeGap {
+		g.pauseUntil = now.Add(monitorResumePause)
+		g.generation++
+	}
+	g.lastTick = now
+}
+
+// snapshot also observes time so a probe waking before the heartbeat detects
+// resume itself. A changed generation invalidates probes spanning suspend.
+func (g *monitorResumeGuard) snapshot() (generation uint32, allowed bool) {
+	if g == nil {
+		return 0, true
+	}
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if g.stop == nil {
+		return g.generation, true
+	}
+	g.observe(time.Now())
+	return g.generation, !g.lastTick.Before(g.pauseUntil)
+}

agent/network_monitor_resume_test.go

@@ -0,0 +1,121 @@
+//go:build testing
+
+package agent
+
+import (
+	"context"
+	"errors"
+	"sync/atomic"
+	"testing"
+	"testing/synctest"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func simulateMonitorSleep(g *monitorResumeGuard) {
+	g.mu.Lock()
+	g.lastTick = time.Now().Add(-time.Hour).Round(0)
+	g.mu.Unlock()
+}
+
+func TestMonitorResumePause(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		var g monitorResumeGuard
+		g.start()
+		defer g.shutdown()
+		generation, allowed := g.snapshot()
+		require.True(t, allowed)
+		// Heartbeats alone must keep the guard current between infrequent probes.
+		time.Sleep(time.Minute)
+		synctest.Wait()
+		steadyGeneration, allowed := g.snapshot()
+		require.True(t, allowed)
+		require.Equal(t, generation, steadyGeneration)
+		// The probe, rather than the heartbeat, must detect this gap.
+		simulateMonitorSleep(&g)
+		next, allowed := g.snapshot()
+		assert.False(t, allowed)
+		assert.NotEqual(t, generation, next)
+		time.Sleep(9 * time.Second)
+		_, allowed = g.snapshot()
+		assert.False(t, allowed)
+		time.Sleep(time.Second)
+		end, allowed := g.snapshot()
+		assert.True(t, allowed)
+		assert.Equal(t, next, end)
+	})
+}
+
+func TestMonitorResumeGuardLifecycle(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		pm := newMonitorManagerWithProbe(func(context.Context, monitor.Config) (int64, error) { return 1, nil })
+		defer pm.Stop()
+		assert.Nil(t, pm.resumeGuard.stop)
+		pm.SyncMonitors([]monitor.Config{{ID: "a", Interval: 3600}, {ID: "b", Interval: 3600}})
+		stop := pm.resumeGuard.stop
+		require.NotNil(t, stop)
+		pm.DeleteMonitor("a")
+		assert.Equal(t, stop, pm.resumeGuard.stop)
+		pm.DeleteMonitor("b")
+		assert.Nil(t, pm.resumeGuard.stop)
+		select {
+		case <-stop:
+		default:
+			t.Fatal("heartbeat was not stopped")
+		}
+		time.Sleep(time.Hour)
+		_, err := pm.UpsertMonitor(monitor.Config{ID: "c", Interval: 3600}, false)
+		require.NoError(t, err)
+		_, allowed := pm.resumeGuard.snapshot()
+		assert.True(t, allowed, "idle time must not trigger a resume pause")
+		pm.SyncMonitors(nil)
+		assert.Nil(t, pm.resumeGuard.stop)
+	})
+}
+
+func TestMonitorResumeDiscardsInflightProbe(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		var g monitorResumeGuard
+		g.start()
+		defer g.shutdown()
+		task := newMonitorTask(monitor.Config{ID: "test"})
+		defer task.cancel()
+		task.resumeGuard = &g
+		result := task.runProbe(func(context.Context, monitor.Config) (int64, error) {
+			simulateMonitorSleep(&g)
+			return 0, errors.New("network not ready")
+		})
+		assert.Nil(t, result)
+		assert.Empty(t, task.history.samples)
+		// Explicit requests may still run during the pause and record real failures.
+		result = task.runProbe(func(context.Context, monitor.Config) (int64, error) {
+			return 0, errors.New("unreachable")
+		})
+		require.NotNil(t, result)
+		assert.Equal(t, 100.0, result.PacketLoss)
+	})
+}
+
+func TestMonitorResumeSkipsScheduledProbes(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		var calls atomic.Int32
+		pm := newMonitorManagerWithProbe(func(context.Context, monitor.Config) (int64, error) {
+			calls.Add(1)
+			return 1, nil
+		})
+		defer pm.Stop()
+		pm.SyncMonitors([]monitor.Config{{ID: "test", Interval: 1}})
+		simulateMonitorSleep(&pm.resumeGuard)
+		pm.resumeGuard.snapshot()
+		time.Sleep(9 * time.Second)
+		synctest.Wait()
+		assert.Zero(t, calls.Load())
+		assert.Empty(t, pm.GetResults(1000))
+		time.Sleep(2 * time.Second)
+		synctest.Wait()
+		assert.Positive(t, calls.Load())
+	})
+}

agent/network_monitor_schedule.go

@@ -0,0 +1,60 @@
+package agent
+
+import (
+	"context"
+	"log/slog"
+	"math/rand"
+	"time"
+)
+
+func (pm *MonitorManager) startMonitor(task *monitorTask) {
+	interval := time.Duration(task.config.Interval) * time.Second
+	if interval < time.Second {
+		interval = 30 * time.Second
+	}
+	delay := getStagger(interval.Milliseconds())
+	slog.Debug("starting monitor task", "target", task.config.Target, "delay", delay, "interval", interval)
+	go runMonitorSchedule(task.ctx, interval, delay, func() {
+		if _, allowed := task.resumeGuard.snapshot(); allowed {
+			task.runProbe(pm.probe)
+		}
+	})
+}
+
+// runMonitorSchedule owns only timing. Checks run serially, and slow checks
+// naturally drop missed ticks rather than building an execution backlog.
+func runMonitorSchedule(ctx context.Context, interval, delay time.Duration, run func()) {
+	timer := time.NewTimer(delay)
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return
+	case <-timer.C:
+	}
+	if ctx.Err() != nil {
+		return
+	}
+	run()
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			if ctx.Err() != nil {
+				return
+			}
+			run()
+		}
+	}
+}
+
+// getStagger returns an initial delay between half an interval and one interval.
+func getStagger(intervalMilli int64) time.Duration {
+	delay := rand.Intn(int(intervalMilli))
+	if delay < int(intervalMilli)/2 {
+		delay += int(intervalMilli) / 2
+	}
+	return time.Duration(delay) * time.Millisecond
+}

agent/network_monitor_schedule_test.go

@@ -0,0 +1,167 @@
+//go:build testing
+
+package agent
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+	"testing/synctest"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestMonitorScheduleTiming(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		ctx, cancel := context.WithCancel(t.Context())
+		defer cancel()
+		var calls atomic.Int32
+		go runMonitorSchedule(ctx, 10*time.Second, 5*time.Second, func() { calls.Add(1) })
+		synctest.Wait()
+		time.Sleep(4 * time.Second)
+		synctest.Wait()
+		assert.Equal(t, 0, int(calls.Load()))
+		time.Sleep(time.Second)
+		synctest.Wait()
+		assert.Equal(t, 1, int(calls.Load()))
+		time.Sleep(10 * time.Second)
+		synctest.Wait()
+		assert.Equal(t, 2, int(calls.Load()))
+		cancel()
+		synctest.Wait()
+		time.Sleep(time.Minute)
+		synctest.Wait()
+		assert.Equal(t, 2, int(calls.Load()))
+	})
+}
+
+func TestMonitorScheduleSlowProbe(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		ctx, cancel := context.WithCancel(t.Context())
+		defer cancel()
+		var calls atomic.Int32
+		release := make(chan struct{})
+		go runMonitorSchedule(ctx, time.Second, 0, func() {
+			calls.Add(1)
+			select {
+			case <-release:
+			case <-ctx.Done():
+			}
+		})
+		synctest.Wait()
+		assert.Equal(t, 1, int(calls.Load()))
+		time.Sleep(time.Minute)
+		synctest.Wait()
+		assert.Equal(t, 1, int(calls.Load()), "a slow probe must not spawn overlapping checks")
+		close(release)
+		synctest.Wait()
+		assert.Equal(t, 1, int(calls.Load()), "missed intervals must not accumulate a backlog")
+		time.Sleep(time.Second)
+		synctest.Wait()
+		assert.Equal(t, 2, int(calls.Load()))
+		cancel()
+		synctest.Wait()
+	})
+}
+
+func TestMonitorScheduledAndImmediateRequestsShareProbe(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		var calls atomic.Int32
+		release := make(chan struct{})
+		cfg := monitor.Config{ID: "test", Interval: 10}
+		pm := newMonitorManagerWithProbe(func(ctx context.Context, config monitor.Config) (int64, error) {
+			assert.Equal(t, cfg, config)
+			calls.Add(1)
+			<-release
+			return 42, nil
+		})
+		defer pm.Stop()
+		task := newMonitorTask(cfg)
+		pm.monitors[cfg.ID] = task
+		go runMonitorSchedule(task.ctx, 10*time.Second, 0, func() { task.runProbe(pm.probe) })
+		synctest.Wait()
+		results := make(chan *monitor.Result, 2)
+		for range 2 {
+			go func() {
+				result, _ := pm.UpsertMonitor(cfg, true)
+				results <- result
+			}()
+		}
+		synctest.Wait()
+		assert.Equal(t, 1, int(calls.Load()))
+		assert.Empty(t, pm.GetResults(1000), "reading history must not wait for network I/O")
+		close(release)
+		synctest.Wait()
+		first, second := <-results, <-results
+		require.NotNil(t, first)
+		require.NotNil(t, second)
+		assert.Equal(t, int64(42), first.AvgResponse)
+		assert.Equal(t, first, second)
+		assert.NotSame(t, first, second, "callers must not share mutable result pointers")
+		assert.Len(t, task.history.samples, 1)
+		// A later explicit request must still perform a fresh probe.
+		_, err := pm.UpsertMonitor(cfg, true)
+		require.NoError(t, err)
+		assert.Equal(t, 2, int(calls.Load()))
+		assert.Len(t, task.history.samples, 2)
+	})
+}
+
+func TestMonitorReplacementCancelsSharedProbe(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		cfg := monitor.Config{ID: "test", Interval: 10}
+		pm := newMonitorManagerWithProbe(func(ctx context.Context, config monitor.Config) (int64, error) {
+			if config.Interval == 10 {
+				<-ctx.Done()
+				return 0, ctx.Err()
+			}
+			return 30, nil
+		})
+		defer pm.Stop()
+		task := newMonitorTask(cfg)
+		task.history.record(monitorSample{responseUs: 10, timestamp: time.Now()})
+		pm.monitors[cfg.ID] = task
+		results := make(chan *monitor.Result, 2)
+		for range 2 {
+			go func() {
+				result, _ := pm.UpsertMonitor(cfg, true)
+				results <- result
+			}()
+		}
+		synctest.Wait()
+		updated := cfg
+		updated.Interval = 20
+		result, err := pm.UpsertMonitor(updated, true)
+		require.NoError(t, err)
+		require.NotNil(t, result)
+		assert.Equal(t, int64(20), result.AvgResponse)
+		assert.Zero(t, result.PacketLoss)
+		synctest.Wait()
+		assert.Nil(t, <-results)
+		assert.Nil(t, <-results)
+		assert.Len(t, task.history.samples, 1)
+		assert.Len(t, pm.monitors[cfg.ID].history.samples, 2)
+	})
+}
+
+func TestMonitorInjectedProbeTimeoutRecordsLoss(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		pm := newMonitorManagerWithProbe(func(ctx context.Context, _ monitor.Config) (int64, error) {
+			ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
+			defer cancel()
+			<-ctx.Done()
+			return 0, ctx.Err()
+		})
+		defer pm.Stop()
+		start := time.Now()
+		result, err := pm.UpsertMonitor(monitor.Config{ID: "test", Interval: 3600}, true)
+		require.NoError(t, err)
+		require.NotNil(t, result)
+		assert.Equal(t, 3*time.Second, time.Since(start))
+		assert.Equal(t, 100.0, result.PacketLoss)
+		assert.NoError(t, pm.monitors["test"].ctx.Err())
+	})
+}

agent/network_monitor_task.go

@@ -0,0 +1,116 @@
+package agent
+
+import (
+	"context"
+	"log/slog"
+	"sync"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+)
+
+const monitorFailureLogInterval = 5 * time.Minute
+
+// monitorTask coordinates a probe and its history for one immutable configuration.
+type monitorTask struct {
+	config         monitor.Config
+	ctx            context.Context
+	cancel         context.CancelFunc
+	history        *monitorHistory
+	resumeGuard    *monitorResumeGuard
+	runMu          sync.Mutex
+	inflight       *monitorRun
+	lastFailureLog int64 // Unix nanoseconds
+}
+
+type monitorRun struct {
+	done   chan struct{}
+	result *monitor.Result // published by closing done; never mutated afterwards
+}
+
+func newMonitorTask(config monitor.Config) *monitorTask {
+	ctx, cancel := context.WithCancel(context.Background())
+	task := &monitorTask{config: config, ctx: ctx, history: newMonitorHistory()}
+	// Serialize cancellation with publication, so canceled probes cannot enter
+	// history copied into a replacement task.
+	task.cancel = func() {
+		task.runMu.Lock()
+		cancel()
+		task.runMu.Unlock()
+	}
+	return task
+}
+
+func newMonitorTaskFromExisting(config monitor.Config, existing *monitorTask) *monitorTask {
+	task := newMonitorTask(config)
+	if existing != nil {
+		task.history = existing.history.clone()
+	}
+	return task
+}
+
+// runProbe shares an in-flight check between scheduled and immediate requests.
+// Every completed check contributes exactly one sample, regardless of how many
+// callers were waiting for it. No task or history lock is held during network I/O.
+func (task *monitorTask) runProbe(probe monitorProbe) *monitor.Result {
+	task.runMu.Lock()
+	if task.ctx.Err() != nil {
+		task.runMu.Unlock()
+		return nil
+	}
+	if run := task.inflight; run != nil {
+		task.runMu.Unlock()
+		select {
+		case <-task.ctx.Done():
+			return nil
+		case <-run.done:
+			if task.ctx.Err() != nil {
+				return nil
+			}
+			return copyMonitorResult(run.result)
+		}
+	}
+	run := &monitorRun{done: make(chan struct{})}
+	task.inflight = run
+	task.runMu.Unlock()
+
+	generation, _ := task.resumeGuard.snapshot()
+	responseUs, err := probe(task.ctx, task.config)
+	var logFailure bool
+	task.runMu.Lock()
+	currentGeneration, _ := task.resumeGuard.snapshot()
+	if task.ctx.Err() == nil && generation == currentGeneration {
+		now := time.Now()
+		if err != nil {
+			responseUs = -1
+			logAt := now.UnixNano()
+			if task.lastFailureLog == 0 || logAt < task.lastFailureLog || logAt-task.lastFailureLog >= int64(monitorFailureLogInterval) {
+				logFailure = true
+				task.lastFailureLog = logAt
+			}
+		} else {
+			task.lastFailureLog = 0
+		}
+		result := task.history.record(monitorSample{responseUs: responseUs, timestamp: now})
+		run.result = &result
+	}
+
+	task.inflight = nil
+	close(run.done)
+	task.runMu.Unlock()
+	if logFailure {
+		slog.Warn("monitor failed", "err", err, "target", task.config.Target, "protocol", task.config.Protocol)
+	}
+	if task.ctx.Err() != nil {
+		return nil
+	}
+	return copyMonitorResult(run.result)
+}
+
+func copyMonitorResult(result *monitor.Result) *monitor.Result {
+	if result == nil {
+		return nil
+	}
+	copy := *result
+	return &copy
+}

agent/network_monitor_task_test.go

@@ -0,0 +1,79 @@
+//go:build testing
+
+package agent
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"log/slog"
+	"testing"
+	"testing/synctest"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestMonitorFailureLogCooldown(t *testing.T) {
+	var logs bytes.Buffer
+	previous := slog.Default()
+	slog.SetDefault(slog.New(slog.NewTextHandler(&logs, nil)))
+	t.Cleanup(func() { slog.SetDefault(previous) })
+
+	synctest.Test(t, func(t *testing.T) {
+		task := newMonitorTask(monitor.Config{ID: "test", Target: "example.test", Protocol: "tcp"})
+		defer task.cancel()
+		failure := errors.New("connection refused")
+		probe := func(context.Context, monitor.Config) (int64, error) { return 42, failure }
+		var samples int64
+		check := func(wantLog bool) {
+			t.Helper()
+			logs.Reset()
+			result := task.runProbe(probe)
+			require.NotNil(t, result)
+			samples++
+			assert.Equal(t, samples, result.SampleCount, "suppressed warnings must still record samples")
+			if !wantLog {
+				assert.Empty(t, logs.String())
+			} else {
+				assert.Contains(t, logs.String(), `msg="monitor failed"`)
+				assert.Equal(t, 1, bytes.Count(logs.Bytes(), []byte("\n")))
+			}
+		}
+
+		check(true)
+		check(false)
+		time.Sleep(5*time.Minute - time.Nanosecond)
+		check(false)
+		time.Sleep(time.Nanosecond)
+		check(true)
+		check(false)
+		time.Sleep(5 * time.Minute)
+		check(true)
+		check(false)
+
+		// Recovery clears the cooldown.
+		failure = nil
+		check(false)
+		failure = errors.New("connection refused again")
+		check(true)
+
+		// Another monitor has its own cooldown.
+		other := newMonitorTask(task.config)
+		defer other.cancel()
+		logs.Reset()
+		require.NotNil(t, other.runProbe(probe))
+		assert.Contains(t, logs.String(), `msg="monitor failed"`)
+
+		// A canceled probe must not publish a failure or emit a warning.
+		logs.Reset()
+		result := other.runProbe(func(context.Context, monitor.Config) (int64, error) {
+			other.cancel()
+			return -1, context.Canceled
+		})
+		assert.Nil(t, result)
+		assert.Empty(t, logs.String())
+	})
+}

agent/network_monitor_test.go

@@ -0,0 +1,524 @@
+package agent
+
+import (
+	"context"
+	"encoding/binary"
+	"io"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/net/dns/dnsmessage"
+)
+
+func TestMonitorManagerGetResultsIncludesHourResponseRange(t *testing.T) {
+	now := time.Now().UTC()
+	task := newMonitorTask(monitor.Config{ID: "monitor-1"})
+	task.history.addSampleLocked(monitorSample{responseUs: 10, timestamp: now.Add(-30 * time.Minute)})
+	task.history.addSampleLocked(monitorSample{responseUs: 20, timestamp: now.Add(-9 * time.Minute)})
+	task.history.addSampleLocked(monitorSample{responseUs: 40, timestamp: now.Add(-5 * time.Minute)})
+	task.history.addSampleLocked(monitorSample{responseUs: 30, timestamp: now.Add(-50 * time.Second)})
+	task.history.addSampleLocked(monitorSample{responseUs: -1, timestamp: now.Add(-30 * time.Second)})
+
+	pm := newMonitorManager()
+	pm.monitors = map[string]*monitorTask{"icmp:example.com": task}
+
+	results := pm.GetResults(uint16(time.Minute / time.Millisecond))
+	result, ok := results["monitor-1"]
+	require.True(t, ok)
+	assert.Equal(t, int64(30), result.AvgResponse)
+	assert.Equal(t, int64(25), result.AvgResponse1h)
+	assert.Equal(t, int64(30), result.MinResponse)
+	assert.Equal(t, int64(10), result.MinResponse1h)
+	assert.Equal(t, int64(30), result.MaxResponse)
+	assert.Equal(t, int64(40), result.MaxResponse1h)
+	assert.Equal(t, 50.0, result.PacketLoss)
+	assert.Equal(t, 20.0, result.PacketLoss1h)
+}
+
+func TestMonitorManagerGetResultsIncludesLossOnlyHourData(t *testing.T) {
+	now := time.Now().UTC()
+	task := newMonitorTask(monitor.Config{ID: "monitor-1"})
+	task.history.addSampleLocked(monitorSample{responseUs: -1, timestamp: now.Add(-30 * time.Second)})
+	task.history.addSampleLocked(monitorSample{responseUs: -1, timestamp: now.Add(-10 * time.Second)})
+
+	pm := newMonitorManager()
+	pm.monitors = map[string]*monitorTask{"icmp:example.com": task}
+
+	results := pm.GetResults(uint16(time.Minute / time.Millisecond))
+	result, ok := results["monitor-1"]
+	require.True(t, ok)
+	assert.Equal(t, int64(0), result.AvgResponse)
+	assert.Equal(t, int64(0), result.AvgResponse1h)
+	assert.Equal(t, int64(0), result.MinResponse)
+	assert.Equal(t, int64(0), result.MinResponse1h)
+	assert.Equal(t, int64(0), result.MaxResponse)
+	assert.Equal(t, int64(0), result.MaxResponse1h)
+	assert.Equal(t, 100.0, result.PacketLoss)
+	assert.Equal(t, 100.0, result.PacketLoss1h)
+}
+
+func TestMonitorConfigResultKeyUsesSyncedID(t *testing.T) {
+	cfg := monitor.Config{ID: "monitor-1", Target: "1.1.1.1", Protocol: "icmp", Interval: 10}
+	assert.Equal(t, "monitor-1", cfg.ID)
+}
+
+func TestMonitorManagerSyncMonitorsSkipsConfigsWithoutStableID(t *testing.T) {
+	validCfg := monitor.Config{ID: "monitor-1", Target: "ignored", Protocol: "noop", Interval: 10}
+	invalidCfg := monitor.Config{Target: "ignored", Protocol: "noop", Interval: 10}
+
+	pm := newMonitorManager()
+	pm.SyncMonitors([]monitor.Config{validCfg, invalidCfg})
+	defer pm.Stop()
+
+	_, validExists := pm.monitors[validCfg.ID]
+	_, invalidExists := pm.monitors[invalidCfg.ID]
+	assert.True(t, validExists)
+	assert.False(t, invalidExists)
+}
+
+func TestMonitorManagerSyncMonitorsStopsRemovedTasksButKeepsExisting(t *testing.T) {
+	keepCfg := monitor.Config{ID: "monitor-1", Target: "ignored", Protocol: "noop", Interval: 10}
+	removeCfg := monitor.Config{ID: "monitor-2", Target: "ignored", Protocol: "noop", Interval: 10}
+
+	keptTask := newMonitorTask(keepCfg)
+	removedTask := newMonitorTask(removeCfg)
+	pm := newMonitorManager()
+	pm.monitors = map[string]*monitorTask{
+		keepCfg.ID:   keptTask,
+		removeCfg.ID: removedTask,
+	}
+
+	pm.SyncMonitors([]monitor.Config{keepCfg})
+
+	assert.Same(t, keptTask, pm.monitors[keepCfg.ID])
+	_, exists := pm.monitors[removeCfg.ID]
+	assert.False(t, exists)
+
+	select {
+	case <-removedTask.ctx.Done():
+	default:
+		t.Fatal("expected removed monitor task to be cancelled")
+	}
+
+	select {
+	case <-keptTask.ctx.Done():
+		t.Fatal("expected existing monitor task to remain active")
+	default:
+	}
+}
+
+func TestMonitorManagerSyncMonitorsRestartsChangedConfig(t *testing.T) {
+	originalCfg := monitor.Config{ID: "monitor-1", Target: "ignored-a", Protocol: "noop", Interval: 10}
+	updatedCfg := monitor.Config{ID: "monitor-1", Target: "ignored-b", Protocol: "noop", Interval: 10}
+	originalTask := newMonitorTask(originalCfg)
+	pm := newMonitorManager()
+	pm.monitors = map[string]*monitorTask{
+		originalCfg.ID: originalTask,
+	}
+
+	pm.SyncMonitors([]monitor.Config{updatedCfg})
+	defer pm.Stop()
+
+	restartedTask := pm.monitors[updatedCfg.ID]
+	assert.NotSame(t, originalTask, restartedTask)
+	assert.Equal(t, updatedCfg, restartedTask.config)
+
+	select {
+	case <-originalTask.ctx.Done():
+	default:
+		t.Fatal("expected changed monitor task to be cancelled")
+	}
+}
+
+func TestMonitorManagerApplySyncUpsertRunsImmediatelyAndReturnsResult(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	pm := &MonitorManager{
+		monitors: make(map[string]*monitorTask),
+		probe:    networkMonitorProbe(server.Client()),
+	}
+
+	resp, err := pm.HandleSyncRequest(monitor.SyncRequest{
+		Action: monitor.SyncActionUpsert,
+		Config: monitor.Config{ID: "monitor-1", Target: server.URL, Protocol: "http", Interval: 10},
+		RunNow: true,
+	})
+	defer pm.Stop()
+
+	require.NoError(t, err)
+	assert.GreaterOrEqual(t, resp.Result.AvgResponse, int64(0))
+	assert.Equal(t, 0.0, resp.Result.PacketLoss)
+	assert.Equal(t, 0.0, resp.Result.PacketLoss1h)
+
+	task := pm.monitors["monitor-1"]
+	require.NotNil(t, task)
+	task.history.mu.Lock()
+	defer task.history.mu.Unlock()
+	require.Len(t, task.history.samples, 1)
+}
+
+func TestMonitorManagerUpsertMonitorKeepsHistoryWhenOnlyIntervalChanges(t *testing.T) {
+	originalCfg := monitor.Config{ID: "monitor-1", Target: "1.1.1.1", Protocol: "icmp", Interval: 10}
+	updatedCfg := monitor.Config{ID: "monitor-1", Target: "1.1.1.1", Protocol: "icmp", Interval: 30}
+	now := time.Now().UTC()
+
+	existingTask := newMonitorTask(originalCfg)
+	existingTask.history.addSampleLocked(monitorSample{responseUs: 12, timestamp: now.Add(-50 * time.Minute)})
+	existingTask.history.addSampleLocked(monitorSample{responseUs: 24, timestamp: now.Add(-30 * time.Second)})
+
+	pm := newMonitorManager()
+	pm.monitors = map[string]*monitorTask{originalCfg.ID: existingTask}
+
+	result, err := pm.UpsertMonitor(updatedCfg, false)
+	defer pm.Stop()
+
+	require.NoError(t, err)
+	assert.Nil(t, result)
+
+	updatedTask := pm.monitors[updatedCfg.ID]
+	require.NotNil(t, updatedTask)
+	assert.NotSame(t, existingTask, updatedTask)
+	assert.Equal(t, updatedCfg, updatedTask.config)
+
+	updatedTask.history.mu.Lock()
+	defer updatedTask.history.mu.Unlock()
+	require.Len(t, updatedTask.history.samples, 1)
+	assert.Equal(t, int64(24), updatedTask.history.samples[0].responseUs)
+
+	agg := updatedTask.history.aggregateLocked(time.Hour, now)
+	require.True(t, agg.hasData())
+	assert.Equal(t, int64(2), agg.totalCount)
+	assert.Equal(t, int64(2), agg.successCount)
+	assert.Equal(t, int64(18), agg.avgResponse())
+
+	select {
+	case <-existingTask.ctx.Done():
+	default:
+		t.Fatal("expected original monitor task to be cancelled")
+	}
+}
+
+func TestMonitorManagerApplySyncDeleteRemovesTask(t *testing.T) {
+	config := monitor.Config{ID: "monitor-1", Target: "1.1.1.1", Protocol: "icmp", Interval: 10}
+	task := newMonitorTask(config)
+	pm := newMonitorManager()
+	pm.monitors = map[string]*monitorTask{config.ID: task}
+
+	_, err := pm.HandleSyncRequest(monitor.SyncRequest{
+		Action: monitor.SyncActionDelete,
+		Config: monitor.Config{ID: config.ID},
+	})
+
+	require.NoError(t, err)
+	_, exists := pm.monitors[config.ID]
+	assert.False(t, exists)
+
+	select {
+	case <-task.ctx.Done():
+	default:
+		t.Fatal("expected deleted monitor task to be cancelled")
+	}
+}
+
+func TestMonitorManagerGetRandomDelay(t *testing.T) {
+	for i := 1000; i < 360_000; i += 1000 {
+		delay := getStagger(int64(i))
+		assert.GreaterOrEqual(t, delay, time.Duration(i/2)*time.Millisecond)
+		assert.LessOrEqual(t, delay, time.Duration(i)*time.Millisecond)
+	}
+}
+
+func TestMonitorHTTP(t *testing.T) {
+	t.Run("success", func(t *testing.T) {
+		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			w.WriteHeader(http.StatusNoContent)
+		}))
+		defer server.Close()
+
+		responseUs, err := monitorHTTP(context.Background(), server.Client(), server.URL)
+		require.NoError(t, err)
+		assert.GreaterOrEqual(t, responseUs, int64(0))
+	})
+
+	t.Run("server error", func(t *testing.T) {
+		server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			http.Error(w, "boom", http.StatusInternalServerError)
+		}))
+		defer server.Close()
+
+		responseUs, err := monitorHTTP(context.Background(), server.Client(), server.URL)
+		assert.Equal(t, int64(-1), responseUs)
+		require.Error(t, err)
+	})
+}
+
+func TestMonitorTCP(t *testing.T) {
+	t.Run("success", func(t *testing.T) {
+		listener, err := net.Listen("tcp", "127.0.0.1:0")
+		require.NoError(t, err)
+		defer listener.Close()
+
+		accepted := make(chan struct{})
+		go func() {
+			defer close(accepted)
+			conn, err := listener.Accept()
+			if err == nil {
+				_ = conn.Close()
+			}
+		}()
+
+		port := uint16(listener.Addr().(*net.TCPAddr).Port)
+		responseUs, err := monitorTCP(context.Background(), "127.0.0.1", port)
+		require.NoError(t, err)
+		assert.GreaterOrEqual(t, responseUs, int64(0))
+		<-accepted
+	})
+
+	t.Run("connection failure", func(t *testing.T) {
+		listener, err := net.Listen("tcp", "127.0.0.1:0")
+		require.NoError(t, err)
+
+		port := uint16(listener.Addr().(*net.TCPAddr).Port)
+		require.NoError(t, listener.Close())
+
+		responseUs, err := monitorTCP(context.Background(), "127.0.0.1", port)
+		assert.Equal(t, int64(-1), responseUs)
+		require.Error(t, err)
+	})
+}
+
+func TestMonitorTCPAddressFallback(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		ips  []string
+		loss bool
+	}{
+		{"first address fails", []string{"127.0.0.2", "127.0.0.1"}, false},
+		{"first address succeeds", []string{"127.0.0.1", "127.0.0.2"}, false},
+		{"all addresses fail", []string{"127.0.0.2", "127.0.0.3"}, true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			listener, err := net.Listen("tcp4", "127.0.0.1:0")
+			require.NoError(t, err)
+			defer listener.Close()
+			original := net.DefaultResolver
+			net.DefaultResolver = tcpMonitorTestResolver(tc.ips)
+			defer func() { net.DefaultResolver = original }()
+
+			// Verify the resolver preserves the intended order, so success cannot
+			// accidentally bypass the failed first address in the regression case.
+			ips, err := net.DefaultResolver.LookupHost(t.Context(), "tcp-monitor.invalid.")
+			require.NoError(t, err)
+			require.Equal(t, tc.ips, ips)
+			responseUs, err := monitorTCP(t.Context(), "tcp-monitor.invalid.", uint16(listener.Addr().(*net.TCPAddr).Port))
+			if tc.loss {
+				require.Error(t, err)
+				assert.Equal(t, int64(-1), responseUs)
+			} else {
+				require.NoError(t, err)
+				assert.GreaterOrEqual(t, responseUs, int64(0))
+			}
+		})
+	}
+}
+
+// tcpMonitorTestResolver supplies multiple A records without external DNS.
+func tcpMonitorTestResolver(ips []string) *net.Resolver {
+	return &net.Resolver{PreferGo: true, Dial: func(ctx context.Context, network, address string) (net.Conn, error) {
+		client, server := net.Pipe()
+		go func() {
+			defer server.Close()
+			// net.Resolver uses TCP framing when its connection is not a PacketConn.
+			var size uint16
+			if err := binary.Read(server, binary.BigEndian, &size); err != nil {
+				return
+			}
+			packet := make([]byte, size)
+			if _, err := io.ReadFull(server, packet); err != nil {
+				return
+			}
+			var msg dnsmessage.Message
+			if err := msg.Unpack(packet); err != nil {
+				return
+			}
+			msg.Header.Response = true
+			msg.Header.RecursionAvailable = true
+			for _, question := range msg.Questions {
+				if question.Type != dnsmessage.TypeA {
+					continue
+				}
+				for _, ip := range ips {
+					msg.Answers = append(msg.Answers, dnsmessage.Resource{
+						Header: dnsmessage.ResourceHeader{Name: question.Name, Type: dnsmessage.TypeA, Class: dnsmessage.ClassINET},
+						Body:   &dnsmessage.AResource{A: [4]byte(net.ParseIP(ip).To4())},
+					})
+				}
+			}
+			packet, err := msg.Pack()
+			if err != nil {
+				return
+			}
+			response := binary.BigEndian.AppendUint16(nil, uint16(len(packet)))
+			_, _ = server.Write(append(response, packet...))
+		}()
+		return client, nil
+	}}
+}
+
+func TestMonitorDNS(t *testing.T) {
+	t.Run("success", func(t *testing.T) {
+		responseUs, err := monitorDNS(context.Background(), "localhost")
+		require.NoError(t, err)
+		assert.GreaterOrEqual(t, responseUs, int64(0))
+	})
+
+	t.Run("lookup failure", func(t *testing.T) {
+		responseUs, err := monitorDNS(context.Background(), "")
+		assert.Equal(t, int64(-1), responseUs)
+		require.Error(t, err)
+	})
+}
+
+func TestMonitorManagerCancelsActiveProbe(t *testing.T) {
+	for _, action := range []string{"stop", "delete", "upsert", "sync replace", "sync remove"} {
+		t.Run(action, func(t *testing.T) {
+			started := make(chan struct{})
+			canceled := make(chan struct{})
+			release := make(chan struct{})
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				close(started)
+				select {
+				case <-r.Context().Done():
+					close(canceled)
+				case <-release:
+				}
+			}))
+			defer server.Close()
+			defer close(release)
+			pm := newMonitorManager()
+			defer pm.Stop()
+			cfg := monitor.Config{ID: "test", Protocol: "http", Target: server.URL, Interval: 3600}
+			task := newMonitorTask(cfg)
+			// Seed history to ensure a canceled RunNow does not return an old result.
+			task.history.addSampleLocked(monitorSample{responseUs: 123, timestamp: time.Now()})
+			pm.monitors[cfg.ID] = task
+			done := make(chan *monitor.Result, 1)
+			go func() {
+				result, _ := pm.UpsertMonitor(cfg, true)
+				done <- result
+			}()
+			select {
+			case <-started:
+			case <-time.After(time.Second):
+				t.Fatal("probe did not start")
+			}
+			updated := cfg
+			updated.Interval--
+			switch action {
+			case "stop":
+				pm.Stop()
+			case "delete":
+				pm.DeleteMonitor(cfg.ID)
+			case "upsert":
+				_, err := pm.UpsertMonitor(updated, false)
+				require.NoError(t, err)
+			case "sync replace":
+				pm.SyncMonitors([]monitor.Config{updated})
+			case "sync remove":
+				pm.SyncMonitors(nil)
+			}
+			select {
+			case <-canceled:
+			case <-time.After(time.Second):
+				t.Fatal("active HTTP request was not canceled")
+			}
+			select {
+			case result := <-done:
+				assert.Nil(t, result)
+			case <-time.After(time.Second):
+				t.Fatal("RunNow did not return after cancellation")
+			}
+			task.history.mu.Lock()
+			assert.Len(t, task.history.samples, 1, "cancellation must not record packet loss")
+			task.history.mu.Unlock()
+		})
+	}
+}
+
+func TestMonitorResolutionCancellation(t *testing.T) {
+	for _, protocol := range []string{"tcp", "dns", "icmp"} {
+		t.Run(protocol, func(t *testing.T) {
+			started := make(chan struct{}, 1)
+			original := net.DefaultResolver
+			net.DefaultResolver = &net.Resolver{PreferGo: true, Dial: func(ctx context.Context, network, address string) (net.Conn, error) {
+				select {
+				case started <- struct{}{}:
+				default:
+				}
+				<-ctx.Done()
+				return nil, ctx.Err()
+			}}
+			defer func() { net.DefaultResolver = original }()
+			ctx, cancel := context.WithCancel(context.Background())
+			defer cancel()
+			done := make(chan error, 1)
+			go func() {
+				var err error
+				switch protocol {
+				case "tcp":
+					_, err = monitorTCP(ctx, "monitor-cancellation.invalid.", 80)
+				case "dns":
+					_, err = monitorDNS(ctx, "monitor-cancellation.invalid.")
+				case "icmp":
+					_, err = monitorICMP(ctx, "monitor-cancellation.invalid.")
+				}
+				done <- err
+			}()
+			select {
+			case <-started:
+			case <-time.After(time.Second):
+				t.Fatal("lookup did not start")
+			}
+			cancel()
+			select {
+			case err := <-done:
+				require.Error(t, err)
+			case <-time.After(time.Second):
+				t.Fatal("lookup did not cancel")
+			}
+		})
+	}
+}
+
+func TestMonitorProbeTimeoutRecordsLoss(t *testing.T) {
+	release := make(chan struct{})
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		select {
+		case <-r.Context().Done():
+		case <-release:
+		}
+	}))
+	defer server.Close()
+	defer close(release)
+	pm := newMonitorManager()
+	pm.probe = networkMonitorProbe(&http.Client{Timeout: 20 * time.Millisecond})
+	task := newMonitorTask(monitor.Config{ID: "timeout", Protocol: "http", Target: server.URL})
+	defer task.cancel()
+
+	result := task.runProbe(pm.probe)
+	require.NotNil(t, result)
+	assert.Equal(t, 100.0, result.PacketLoss)
+	assert.Equal(t, 100.0, result.PacketLoss1h)
+	require.Len(t, task.history.samples, 1)
+	assert.Equal(t, int64(-1), task.history.samples[0].responseUs)
+	assert.NoError(t, task.ctx.Err(), "a probe timeout must not cancel the task")
+}

agent/smart.go

@@ -931,9 +931,6 @@ func (sm *SmartManager) parseSmartForSata(output []byte, deviceType string) (boo
 		if parsed, ok := smart.ParseSmartRawValueString(attr.Raw.String); ok {
 			rawValue = parsed
 		}
-		if smartData.SmartStatus == "PASSED" && rawValue > 0 && (attr.ID == 5 || attr.ID == 197 || attr.ID == 198) {
-			smartData.SmartStatus = "WARNING"
-		}
 		smartAttr := &smart.SmartAttribute{
 			ID:         attr.ID,
 			Name:       attr.Name,

agent/smart_test.go

@@ -7,7 +7,6 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
-	"strconv"
 	"testing"
 
 	"github.com/henrygd/beszel/internal/entities/smart"
@@ -90,27 +89,6 @@ func TestParseSmartForSata(t *testing.T) {
 	}
 }
 
-func TestParseSmartForSataWarnsForCriticalAttributes(t *testing.T) {
-	for _, attrID := range []int{5, 197, 198} {
-		t.Run("attribute "+strconv.Itoa(attrID), func(t *testing.T) {
-			jsonPayload := []byte(fmt.Sprintf(`{
-				"smartctl": {"exit_status": 0},
-				"device": {"name": "/dev/sda", "type": "sat"},
-				"model_name": "Example",
-				"serial_number": "WARNING%d",
-				"smart_status": {"passed": true},
-				"temperature": {"current": 30},
-				"ata_smart_attributes": {"table": [{"id": %d, "raw": {"value": 1, "string": "1"}}]}
-			}`, attrID, attrID))
-
-			sm := &SmartManager{SmartDataMap: make(map[string]*smart.SmartData)}
-			hasData, _ := sm.parseSmartForSata(jsonPayload, "")
-			require.True(t, hasData)
-			assert.Equal(t, "WARNING", sm.SmartDataMap[fmt.Sprintf("WARNING%d", attrID)].SmartStatus)
-		})
-	}
-}
-
 func TestParseSmartForSataPreservesFailedAndUnknownStatus(t *testing.T) {
 	for _, test := range []struct {
 		name        string

agent/storage_pool.go

@@ -0,0 +1,462 @@
+package agent
+
+import (
+	"errors"
+	"log/slog"
+	"os/exec"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/henrygd/beszel/agent/btrfs"
+	"github.com/henrygd/beszel/agent/zfs"
+	"github.com/henrygd/beszel/internal/entities/system"
+	zfsentity "github.com/henrygd/beszel/internal/entities/zfs"
+)
+
+// zfsDatasetUsage holds usage values for a ZFS dataset mountpoint.
+type zfsDatasetUsage struct {
+	used  uint64
+	avail uint64
+}
+
+// datasetUsageRefreshInterval controls how often `zfs list` is re-run for the
+// mountpoint usage map. Dataset inventory changes rarely.
+const datasetUsageRefreshInterval = 5 * time.Minute
+
+// poolStatsRefreshInterval controls how often `zpool list` is re-run for pool
+// capacity. Health and I/O are read from procfs on Linux, so the utility only
+// needs to refresh slow-moving space accounting.
+const poolStatsRefreshInterval = time.Minute
+
+// btrfsFilesystems is the btrfs source; overridable in tests.
+var btrfsFilesystems = btrfs.Filesystems
+
+type poolKernelSample struct {
+	nread  uint64
+	nwrite uint64
+	at     time.Time
+}
+
+// StoragePoolManager combines independent backend inventories. Metrics and
+// dataset usage require the agent lock; GetDetail is safe for concurrent calls.
+type StoragePoolManager struct {
+	backends       []*poolBackend
+	detailInterval time.Duration
+}
+
+// poolBackend owns one backend's collectors and caches. Collector functions
+// are immutable after construction and may run concurrently for metrics/details.
+type poolBackend struct {
+	name           string
+	poolStatsFn    func() ([]zfs.PoolStat, error)       // capacity/health source
+	datasetsFn     func() ([]zfs.Dataset, error)        // dataset inventory source
+	kernelStatsFn  func() ([]zfs.PoolKernelStat, error) // procfs pool state/I/O source
+	poolStatusesFn func() ([]zfs.PoolStatus, error)     // scrub/vdev detail source
+
+	poolData      []zfs.PoolStat // cached pool inventory (TTL below)
+	lastPoolStats time.Time
+	kernelSamples map[string]poolKernelSample
+
+	datasetUsage     map[string]zfsDatasetUsage // mountpoint -> usage
+	lastUsageRefresh time.Time
+
+	// Detail data (pools, vdevs, scrub, datasets) is cached and refreshed on
+	// an interval. Accessed from handler goroutines, so it is mutex-protected.
+	detailMu          sync.Mutex
+	detail            *zfsentity.ZfsData
+	lastDetailRefresh time.Time
+	detailFailed      bool
+}
+
+func newStoragePoolManager() *StoragePoolManager {
+	return &StoragePoolManager{
+		backends:       []*poolBackend{newZfsBackend(), newBtrfsBackend()},
+		detailInterval: time.Hour,
+	}
+}
+
+func newZfsBackend() *poolBackend {
+	return &poolBackend{
+		name:           "zfs",
+		poolStatsFn:    optionalPoolSource(zfs.PoolStats),
+		datasetsFn:     optionalPoolSource(zfs.Datasets),
+		kernelStatsFn:  optionalPoolSource(zfs.PoolKernelStats),
+		poolStatusesFn: optionalPoolSource(zfs.PoolStatuses),
+	}
+}
+
+func newBtrfsBackend() *poolBackend {
+	return &poolBackend{
+		name:           "btrfs",
+		poolStatsFn:    btrfsSource(btrfsPoolStats),
+		kernelStatsFn:  btrfsSource(btrfsKernelStats),
+		poolStatusesFn: btrfsSource(btrfsPoolStatuses),
+	}
+}
+
+// datasets is optional: only backends that expose datasets provide a collector.
+func (b *poolBackend) datasets() ([]zfs.Dataset, error) {
+	if b.datasetsFn == nil {
+		return nil, nil
+	}
+	return b.datasetsFn()
+}
+
+// A missing utility/interface is a successfully observed absent backend.
+func optionalPoolSource[T any](source func() ([]T, error)) func() ([]T, error) {
+	return func() ([]T, error) {
+		items, err := source()
+		if errors.Is(err, zfs.ErrNoZfs) || errors.Is(err, exec.ErrNotFound) || errors.Is(err, errors.ErrUnsupported) {
+			return nil, nil
+		}
+		return items, err
+	}
+}
+
+func btrfsSource[T any](convert func(btrfs.Filesystem) T) func() ([]T, error) {
+	return func() ([]T, error) {
+		filesystems, err := optionalPoolSource(btrfsFilesystems)()
+		if err != nil {
+			return nil, err
+		}
+		items := make([]T, 0, len(filesystems))
+		for _, fs := range filesystems {
+			items = append(items, convert(fs))
+		}
+		return items, nil
+	}
+}
+
+// Update refreshes systemStats.ZfsPools with the latest pool data. I/O
+// throughput and health come from inexpensive kernel kstats on Linux. Pool
+// capacity and dataset usage come from separately cached utility calls. The
+// pool map is empty when both backends are absent.
+func (m *StoragePoolManager) Update(systemStats *system.Stats) {
+	// Rebuild the combined map so successful pool removals clear old samples.
+	systemStats.ZfsPools = nil
+	for _, backend := range m.backends {
+		backend.updateBackendStats(systemStats)
+	}
+}
+
+func (b *poolBackend) updateBackendStats(systemStats *system.Stats) {
+	pools := b.poolStats()
+	if len(pools) == 0 {
+		b.kernelSamples = nil
+		return
+	}
+
+	kernelStats, ioRates := b.kernelStats()
+
+	if systemStats.ZfsPools == nil {
+		systemStats.ZfsPools = make(map[string]*system.ZfsPool, len(pools))
+	}
+	for i := range pools {
+		pool := &pools[i]
+		// Full precision, matching the dataset values below; the frontend
+		// formats any magnitude.
+		stats := &system.ZfsPool{
+			DisplayName: pool.DisplayName,
+			Raw:         pool.Raw,
+			Total:       float64(pool.Size) / (1024 * 1024 * 1024),
+			Used:        float64(pool.Alloc) / (1024 * 1024 * 1024),
+			Health:      pool.Health,
+		}
+		if kernel, exists := kernelStats[pool.Name]; exists && kernel.Health != "" {
+			stats.Health = kernel.Health
+		}
+		if io, exists := ioRates[pool.Name]; exists {
+			stats.ReadBytes = io.NRead
+			stats.WriteBytes = io.NWrite
+		}
+		slog.Debug("Storage pool sample", "backend", b.name, "pool", pool.Name, "health", stats.Health, "used_gb", stats.Used, "read_bps", stats.ReadBytes, "write_bps", stats.WriteBytes)
+		systemStats.ZfsPools[pool.Name] = stats
+	}
+
+}
+
+// poolStats returns the cached pool inventory, calling its collector at most
+// every poolStatsRefreshInterval. On failure the previous inventory is
+// retained and the refresh is retried on the next cadence.
+func (b *poolBackend) poolStats() []zfs.PoolStat {
+	if b.lastPoolStats.IsZero() || time.Since(b.lastPoolStats) >= poolStatsRefreshInterval {
+		pools, err := b.poolStatsFn()
+		if err != nil {
+			slog.Debug("Storage pool stats unavailable", "backend", b.name, "err", err)
+		} else {
+			b.poolData = pools
+		}
+		b.lastPoolStats = time.Now()
+	}
+	return b.poolData
+}
+
+// kernelStats reads cumulative pool counters and converts them to per-second
+// rates. Counter decreases indicate a pool export/import and reset the
+// baseline instead of producing an underflow spike.
+func (b *poolBackend) kernelStats() (map[string]zfs.PoolKernelStat, map[string]zfs.PoolIoStats) {
+	if b.kernelStatsFn == nil {
+		return nil, nil
+	}
+	stats, err := b.kernelStatsFn()
+	if err != nil {
+		slog.Debug("Storage pool kernel stats unavailable", "backend", b.name, "err", err)
+		return nil, nil
+	}
+	now := time.Now()
+	byName := make(map[string]zfs.PoolKernelStat, len(stats))
+	rates := make(map[string]zfs.PoolIoStats, len(stats))
+	nextSamples := make(map[string]poolKernelSample, len(stats))
+	for _, stat := range stats {
+		byName[stat.Name] = stat
+		if previous, ok := b.kernelSamples[stat.Name]; ok && now.After(previous.at) &&
+			stat.NRead >= previous.nread && stat.NWrite >= previous.nwrite {
+			seconds := now.Sub(previous.at).Seconds()
+			rates[stat.Name] = zfs.PoolIoStats{
+				NRead:  uint64(float64(stat.NRead-previous.nread) / seconds),
+				NWrite: uint64(float64(stat.NWrite-previous.nwrite) / seconds),
+			}
+		}
+		nextSamples[stat.Name] = poolKernelSample{nread: stat.NRead, nwrite: stat.NWrite, at: now}
+	}
+	b.kernelSamples = nextSamples
+	return byName, rates
+}
+
+// refreshDatasetUsage re-runs `zfs list` when the refresh window has elapsed
+// and rebuilds the mountpoint-keyed usage map.
+func (b *poolBackend) refreshDatasetUsage() {
+	if !b.lastUsageRefresh.IsZero() && time.Since(b.lastUsageRefresh) < datasetUsageRefreshInterval {
+		return
+	}
+	datasets, err := b.datasets()
+	if err != nil {
+		slog.Debug("Storage pool dataset usage unavailable", "backend", b.name, "err", err)
+	} else {
+		usage := make(map[string]zfsDatasetUsage, len(datasets))
+		for _, ds := range datasets {
+			if ds.Mountpoint != "" && ds.Mountpoint != "-" {
+				usage[ds.Mountpoint] = zfsDatasetUsage{used: ds.Used, avail: ds.Avail}
+			}
+		}
+		b.datasetUsage = usage
+	}
+	b.lastUsageRefresh = time.Now()
+}
+
+// DatasetUsage returns ZFS dataset usage keyed by mountpoint, refreshed at
+// most every datasetUsageRefreshInterval. On failure the previous map is
+// retained and a debug log is emitted.
+func (m *StoragePoolManager) DatasetUsage() map[string]zfsDatasetUsage {
+	for _, backend := range m.backends {
+		if backend.name == "zfs" {
+			backend.refreshDatasetUsage()
+			return backend.datasetUsage
+		}
+	}
+	return nil
+}
+
+// GetDetail combines backend snapshots, identifying successful inventories so
+// the hub can accept partial updates without deleting failed backend records.
+func (m *StoragePoolManager) GetDetail(force bool) *zfsentity.ZfsData {
+	data := &zfsentity.ZfsData{Complete: true}
+	for _, backend := range m.backends {
+		snapshot := backend.getBackendDetail(force, m.detailInterval)
+		data.Pools = append(data.Pools, snapshot.Pools...)
+		if snapshot.Complete {
+			data.CompleteBackends = append(data.CompleteBackends, backend.name)
+		} else {
+			data.Complete = false
+		}
+	}
+	return data
+}
+
+func (b *poolBackend) getBackendDetail(force bool, interval time.Duration) *zfsentity.ZfsData {
+	b.detailMu.Lock()
+	defer b.detailMu.Unlock()
+
+	if force || b.detailFailed || b.detail == nil || time.Since(b.lastDetailRefresh) >= interval {
+		if data, err := b.collectDetail(b.detail); err != nil {
+			b.detailFailed = true
+			slog.Debug("Storage pool detail collection failed", "backend", b.name, "err", err)
+			if b.detail == nil {
+				return &zfsentity.ZfsData{}
+			}
+			return &zfsentity.ZfsData{Pools: b.detail.Pools}
+		} else {
+			b.detailFailed = false
+			b.detail = data
+			b.lastDetailRefresh = time.Now()
+		}
+	}
+	if b.detail == nil {
+		return &zfsentity.ZfsData{}
+	}
+	return b.detail
+}
+
+// collectDetail builds a ZfsData payload from the current system state.
+func (b *poolBackend) collectDetail(previous *zfsentity.ZfsData) (*zfsentity.ZfsData, error) {
+	pools, err := b.poolStatsFn()
+	if err != nil {
+		return nil, err
+	}
+	if len(pools) == 0 {
+		return &zfsentity.ZfsData{Pools: []*zfsentity.PoolDetail{}, Complete: true}, nil
+	}
+
+	statuses, statusErr := b.poolStatusesFn()
+	if statusErr != nil {
+		slog.Debug("Storage pool status unavailable", "backend", b.name, "err", statusErr)
+	}
+	datasets, datasetsErr := b.datasets()
+	if datasetsErr != nil {
+		slog.Debug("Storage pool datasets unavailable", "backend", b.name, "err", datasetsErr)
+	}
+
+	statusByPool := make(map[string]zfs.PoolStatus, len(statuses))
+	for _, st := range statuses {
+		statusByPool[st.Name] = st
+	}
+
+	previousByPool := make(map[string]*zfsentity.PoolDetail)
+	if previous != nil {
+		for _, pool := range previous.Pools {
+			if pool != nil {
+				previousByPool[pool.Name] = pool
+			}
+		}
+	}
+
+	data := &zfsentity.ZfsData{Pools: make([]*zfsentity.PoolDetail, 0, len(pools)), Complete: true}
+	for i := range pools {
+		p := &pools[i]
+		detail := &zfsentity.PoolDetail{
+			DisplayName: p.DisplayName,
+			Raw:         p.Raw,
+			Name:        p.Name,
+			Health:      p.Health,
+			Size:        p.Size,
+			Alloc:       p.Alloc,
+			Free:        p.Free,
+		}
+		if st, ok := statusByPool[p.Name]; statusErr == nil && ok {
+			if st.Scrub.State != "" && st.Scrub.State != "NONE" {
+				detail.Scrub = &zfsentity.Scrub{
+					State:    st.Scrub.State,
+					Progress: st.Scrub.Progress,
+					Errors:   st.Scrub.Errors,
+				}
+			}
+			for _, v := range st.Vdevs {
+				detail.Vdevs = append(detail.Vdevs, &zfsentity.Vdev{
+					Name:         v.Name,
+					State:        v.State,
+					ReadErrs:     v.ReadErrs,
+					WriteErrs:    v.WriteErrs,
+					ChecksumErrs: v.ChecksumErrs,
+				})
+			}
+		} else {
+			if cached := previousByPool[p.Name]; cached != nil {
+				detail.Scrub = cached.Scrub
+				detail.Vdevs = cached.Vdevs
+			}
+		}
+		if datasetsErr == nil {
+			foundDataset := false
+			for _, ds := range datasets {
+				if poolOfDataset(ds.Name) == p.Name {
+					foundDataset = true
+					detail.Datasets = append(detail.Datasets, &zfsentity.Dataset{
+						Name:       ds.Name,
+						Used:       ds.Used,
+						Avail:      ds.Avail,
+						Mountpoint: ds.Mountpoint,
+					})
+				}
+			}
+			if !foundDataset {
+				if cached := previousByPool[p.Name]; cached != nil {
+					detail.Datasets = cached.Datasets
+				}
+			}
+		} else if cached := previousByPool[p.Name]; cached != nil {
+			detail.Datasets = cached.Datasets
+		}
+		data.Pools = append(data.Pools, detail)
+	}
+	return data, nil
+}
+
+// poolOfDataset returns the pool name for a dataset name (everything before
+// the first '/'). Datasets without a separator belong to a pool of the same
+// name.
+func poolOfDataset(name string) string {
+	if idx := strings.IndexByte(name, '/'); idx >= 0 {
+		return name[:idx]
+	}
+	return name
+}
+
+// ZfsMountpoints returns the set of mountpoints backed by ZFS datasets.
+func (m *StoragePoolManager) ZfsMountpoints() map[string]bool {
+	usage := m.DatasetUsage()
+	mountpoints := make(map[string]bool, len(usage))
+	for mountpoint := range usage {
+		mountpoints[mountpoint] = true
+	}
+	return mountpoints
+}
+
+func btrfsPoolStats(fs btrfs.Filesystem) zfs.PoolStat {
+	return zfs.PoolStat{MountID: fs.MountID, IODevice: fs.IODevice, Raw: fs.Raw, DisplayName: fs.Name, Name: "b:" + fs.UUID, Size: fs.Size, Alloc: fs.Alloc, Free: fs.Size - min(fs.Alloc, fs.Size), Health: fs.Health}
+}
+
+func btrfsKernelStats(fs btrfs.Filesystem) zfs.PoolKernelStat {
+	return zfs.PoolKernelStat{Name: "b:" + fs.UUID, Health: fs.Health, NRead: fs.NRead, NWrite: fs.NWrite}
+}
+
+func btrfsPoolStatuses(fs btrfs.Filesystem) zfs.PoolStatus {
+	status := zfs.PoolStatus{Name: "b:" + fs.UUID, State: fs.Health, Scrub: zfs.ScrubStatus{State: "NONE"}}
+	for _, dev := range fs.Devices {
+		status.Vdevs = append(status.Vdevs, zfs.VdevStatus{
+			Name: dev.Name, State: dev.State,
+			ReadErrs: dev.ReadErrs, WriteErrs: dev.WriteErrs, ChecksumErrs: dev.CorruptionErrs,
+		})
+	}
+	return status
+}
+
+// markDuplicateCharts leaves pool telemetry and detail intact, but tells the
+// hub which charts already have a filesystem equivalent. Only exact kernel
+// filesystem and I/O-device matches qualify; labels are never used.
+func (m *StoragePoolManager) markDuplicateCharts(stats *system.Stats, filesystems map[string]*system.FsStats, mountID func(string) string) {
+	identities := make(map[string]string, len(filesystems))
+	for device, fs := range filesystems {
+		if fs.DiskTotal > 0 {
+			identities[device] = mountID(fs.Mountpoint)
+		}
+	}
+	for _, backend := range m.backends {
+		for _, pool := range backend.poolData {
+			sample := stats.ZfsPools[pool.Name]
+			if sample == nil || pool.MountID == "" {
+				continue
+			}
+			for device, identity := range identities {
+				if identity != pool.MountID {
+					continue
+				}
+				// Raw physical usage is not equivalent to a filesystem usage chart.
+				sample.HideUsage = !pool.Raw
+				if pool.IODevice != "" && pool.IODevice == device {
+					sample.HideIO = true
+				}
+			}
+		}
+	}
+}

agent/storage_pool_test.go

@@ -0,0 +1,520 @@
+//go:build testing
+
+package agent
+
+import (
+	"errors"
+	"fmt"
+	"os/exec"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/agent/btrfs"
+	"github.com/henrygd/beszel/agent/zfs"
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestOptionalPoolSource(t *testing.T) {
+	failure := errors.New("timeout")
+	for _, err := range []error{nil, zfs.ErrNoZfs, fmt.Errorf("zpool: %w", exec.ErrNotFound), errors.ErrUnsupported, failure} {
+		_, got := optionalPoolSource(func() ([]zfs.PoolStat, error) { return nil, err })()
+		if err == failure {
+			assert.ErrorIs(t, got, failure)
+		} else {
+			assert.NoError(t, got)
+		}
+	}
+}
+
+type poolTestBackend struct {
+	name  string
+	err   error
+	alloc uint64
+	read  uint64
+	empty bool
+}
+
+func (state *poolTestBackend) backend() *poolBackend {
+	name := "zfs"
+	if strings.HasPrefix(state.name, "b:") {
+		name = "btrfs"
+	}
+	return &poolBackend{
+		name: name,
+		poolStatsFn: func() ([]zfs.PoolStat, error) {
+			if state.empty {
+				return nil, state.err
+			}
+			return []zfs.PoolStat{{Name: state.name, Size: 100, Alloc: state.alloc}}, state.err
+		},
+		kernelStatsFn: func() ([]zfs.PoolKernelStat, error) {
+			return []zfs.PoolKernelStat{{Name: state.name, NRead: state.read}}, state.err
+		},
+		poolStatusesFn: func() ([]zfs.PoolStatus, error) { return nil, nil },
+		datasetsFn:     func() ([]zfs.Dataset, error) { return nil, nil },
+	}
+}
+
+func TestIndependentPoolBackendCaches(t *testing.T) {
+	for _, failed := range []int{0, 1} {
+		t.Run([]string{"zfs", "btrfs"}[failed], func(t *testing.T) {
+			states := []*poolTestBackend{{name: "tank", alloc: 10}, {name: "b:uuid", alloc: 10}}
+			managers := []*poolBackend{states[0].backend(), states[1].backend()}
+			zm := &StoragePoolManager{backends: managers, detailInterval: time.Hour}
+			var stats system.Stats
+			zm.Update(&stats)
+			require.Len(t, stats.ZfsPools, 2)
+			require.True(t, zm.GetDetail(true).Complete)
+			baseline := poolKernelSample{at: time.Now().Add(-time.Second)}
+			for i, m := range managers {
+				m.lastPoolStats = time.Time{}
+				m.kernelSamples[states[i].name] = baseline
+				states[i].alloc = 20
+				states[i].read = 100
+			}
+			states[failed].err = errors.New("collection failed")
+			zm.Update(&stats)
+			healthy := 1 - failed
+			assert.Equal(t, uint64(10), managers[failed].poolData[0].Alloc)
+			assert.Equal(t, uint64(20), managers[healthy].poolData[0].Alloc)
+			assert.Equal(t, baseline, managers[failed].kernelSamples[states[failed].name])
+			assert.Zero(t, stats.ZfsPools[states[failed].name].ReadBytes)
+			assert.Positive(t, stats.ZfsPools[states[healthy].name].ReadBytes)
+			partial := zm.GetDetail(true)
+			assert.False(t, partial.Complete)
+			assert.False(t, partial.CanRefreshPool(states[failed].name))
+			assert.True(t, partial.CanRefreshPool(states[healthy].name))
+			assert.Equal(t, uint64(10), partial.Pools[failed].Alloc)
+			assert.Equal(t, uint64(20), partial.Pools[healthy].Alloc)
+			assert.False(t, zm.GetDetail(false).Complete, "a failed forced refresh must not become complete from cache")
+
+			// Successful empty inventory removes only the healthy backend's pool.
+			states[healthy].empty = true
+			managers[healthy].lastPoolStats = time.Time{}
+			zm.Update(&stats)
+			require.Len(t, stats.ZfsPools, 1)
+			assert.Contains(t, stats.ZfsPools, states[failed].name)
+			partial = zm.GetDetail(true)
+			require.Len(t, partial.Pools, 1)
+			assert.True(t, partial.CanRefreshPool(states[healthy].name))
+
+			// Recovery uses the retained I/O baseline, then normal removal works.
+			states[failed].err = nil
+			managers[failed].lastPoolStats = time.Time{}
+			zm.Update(&stats)
+			assert.Positive(t, stats.ZfsPools[states[failed].name].ReadBytes)
+			assert.True(t, zm.GetDetail(true).Complete)
+			states[failed].empty = true
+			managers[failed].lastPoolStats = time.Time{}
+			zm.Update(&stats)
+			assert.Empty(t, stats.ZfsPools)
+			assert.Empty(t, zm.GetDetail(true).Pools)
+		})
+	}
+}
+
+func TestIndependentBackendsWithoutCache(t *testing.T) {
+	z := &poolTestBackend{name: "tank", err: errors.New("ZFS failure")}
+	b := &poolTestBackend{name: "b:uuid", alloc: 20}
+	zm := &StoragePoolManager{backends: []*poolBackend{z.backend(), b.backend()}, detailInterval: time.Hour}
+	var stats system.Stats
+	zm.Update(&stats)
+	require.Len(t, stats.ZfsPools, 1)
+	assert.Contains(t, stats.ZfsPools, "b:uuid")
+	detail := zm.GetDetail(true)
+	require.Len(t, detail.Pools, 1)
+	assert.False(t, detail.Complete)
+	assert.Equal(t, []string{"btrfs"}, detail.CompleteBackends)
+}
+
+func TestConcurrentBackendDetailsAndMetrics(t *testing.T) {
+	zm := &StoragePoolManager{backends: []*poolBackend{(&poolTestBackend{name: "tank"}).backend(), (&poolTestBackend{name: "b:uuid"}).backend()}, detailInterval: time.Hour}
+	var wg sync.WaitGroup
+	for i := 0; i < 3; i++ {
+		wg.Add(1)
+		go func(metrics bool) {
+			defer wg.Done()
+			for j := 0; j < 10; j++ {
+				if metrics {
+					zm.Update(&system.Stats{})
+				} else {
+					zm.GetDetail(true)
+				}
+			}
+		}(i == 0)
+	}
+	wg.Wait()
+}
+
+func TestStoragePoolBackendOrder(t *testing.T) {
+	z := (&poolTestBackend{name: "tank"}).backend()
+	b := (&poolTestBackend{name: "b:uuid"}).backend()
+	b.datasetsFn = nil // Btrfs does not expose datasets.
+	z.datasetsFn = func() ([]zfs.Dataset, error) {
+		return []zfs.Dataset{{Name: "tank/data", Mountpoint: "/tank", Used: 10}}, nil
+	}
+	m := &StoragePoolManager{backends: []*poolBackend{b, z}, detailInterval: time.Hour}
+	var stats system.Stats
+	m.Update(&stats)
+	require.Len(t, stats.ZfsPools, 2)
+	detail := m.GetDetail(true)
+	require.True(t, detail.Complete)
+	assert.Equal(t, []string{"btrfs", "zfs"}, detail.CompleteBackends)
+	assert.Empty(t, detail.Pools[0].Datasets)
+	assert.Len(t, detail.Pools[1].Datasets, 1)
+	assert.Equal(t, uint64(10), m.DatasetUsage()["/tank"].used)
+
+	b.poolData[0].MountID = "uuid"
+	b.poolData[0].IODevice = "sda"
+	calls := 0
+	m.markDuplicateCharts(&stats, map[string]*system.FsStats{
+		"sda": {Mountpoint: "/", DiskTotal: 100},
+	}, func(string) string { calls++; return "uuid" })
+	assert.Equal(t, 1, calls, "resolve each filesystem once across all backends")
+	assert.True(t, stats.ZfsPools["b:uuid"].HideUsage)
+	assert.True(t, stats.ZfsPools["b:uuid"].HideIO)
+}
+
+func TestUpdatePopulatesZfsPools(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		return []zfs.PoolStat{{Name: "tank", Size: 23999000000000, Alloc: 12000000000000, Free: 11999000000000, Health: "DEGRADED"}}, nil
+	}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
+		return []zfs.Dataset{
+			{Name: "tank/apps", Used: 5000000000000, Avail: 11999000000000, Mountpoint: "/tank/apps"},
+			{Name: "tank/backup", Used: 6000000000000, Avail: 11999000000000, Mountpoint: "/tank/backup"},
+			// Small zvol (Proxmox VM EFI disk): must not round to zero.
+			{Name: "rpool/vm-100-disk-2", Used: 4194304, Avail: 0, Mountpoint: "-"},
+		}, nil
+	}
+	var kernelCalls int
+	zm.backends[0].kernelStatsFn = func() ([]zfs.PoolKernelStat, error) {
+		kernelCalls++
+		return []zfs.PoolKernelStat{{
+			Name: "tank", Health: "ONLINE",
+			NRead: uint64(kernelCalls-1) * 1250, NWrite: uint64(kernelCalls-1) * 5120,
+		}}, nil
+	}
+
+	var stats system.Stats
+	// The first kernel sample establishes the cumulative-counter baseline.
+	zm.Update(&stats)
+	zm.backends[0].kernelSamples["tank"] = poolKernelSample{at: time.Now().Add(-time.Second)}
+	zm.Update(&stats)
+	require.NotNil(t, stats.ZfsPools)
+	require.Contains(t, stats.ZfsPools, "tank")
+	assert.InDelta(t, 22350.8105, stats.ZfsPools["tank"].Total, 0.0001) // Size in GiB
+	assert.InDelta(t, 11175.8709, stats.ZfsPools["tank"].Used, 0.0001)  // Alloc in GiB
+	assert.Equal(t, "ONLINE", stats.ZfsPools["tank"].Health)
+	assert.InDelta(t, 1250, stats.ZfsPools["tank"].ReadBytes, 5)
+	assert.InDelta(t, 5120, stats.ZfsPools["tank"].WriteBytes, 5)
+
+}
+
+// TestUpdateKernelStatsMissing verifies pools without a kernel sample report zero
+// I/O instead of erroring.
+func TestUpdateKernelStatsMissing(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		return []zfs.PoolStat{{Name: "tank", Size: 1, Alloc: 1, Health: "ONLINE"}}, nil
+	}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
+	zm.backends[0].kernelStatsFn = func() ([]zfs.PoolKernelStat, error) {
+		return nil, zfs.ErrNoZfs
+	}
+
+	var stats system.Stats
+	zm.Update(&stats)
+	require.NotNil(t, stats.ZfsPools)
+	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].ReadBytes)
+	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].WriteBytes)
+}
+
+func TestUpdateKernelCounterReset(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		return []zfs.PoolStat{{Name: "tank", Health: "ONLINE"}}, nil
+	}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
+	zm.backends[0].kernelSamples = map[string]poolKernelSample{
+		"tank": {nread: 100, nwrite: 200, at: time.Now().Add(-time.Second)},
+	}
+	zm.backends[0].kernelStatsFn = func() ([]zfs.PoolKernelStat, error) {
+		return []zfs.PoolKernelStat{{Name: "tank", Health: "ONLINE", NRead: 10, NWrite: 20}}, nil
+	}
+
+	var stats system.Stats
+	zm.Update(&stats)
+	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].ReadBytes)
+	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].WriteBytes)
+}
+
+func TestUpdateNoZfs(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	calls := 0
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		calls++
+		return nil, zfs.ErrNoZfs
+	}
+
+	var stats system.Stats
+	zm.Update(&stats)
+	zm.Update(&stats)
+	assert.Nil(t, stats.ZfsPools)
+	assert.Equal(t, 1, calls, "failed pool discovery should be cached until the next refresh interval")
+}
+
+func TestUpdateEmptyPools(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	calls := 0
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		calls++
+		return nil, nil
+	}
+
+	var stats system.Stats
+	zm.Update(&stats)
+	zm.Update(&stats)
+	assert.Nil(t, stats.ZfsPools)
+	assert.Equal(t, 1, calls, "an empty pool inventory should be cached until the next refresh interval")
+}
+
+func TestDatasetUsage(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	calls := 0
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
+		calls++
+		return []zfs.Dataset{
+			{Name: "tank", Used: 12000000000000, Avail: 11999000000000, Mountpoint: "/tank"},
+			{Name: "tank/apps", Used: 1000000000000, Avail: 11999000000000, Mountpoint: "/tank/apps"},
+			{Name: "rpool", Used: 900000000000, Avail: 300000000000, Mountpoint: "-"}, // zvol/unmounted: excluded
+		}, nil
+	}
+
+	usage := zm.DatasetUsage()
+	require.Len(t, usage, 2)
+	assert.Equal(t, zfsDatasetUsage{used: 12000000000000, avail: 11999000000000}, usage["/tank"])
+	assert.Equal(t, zfsDatasetUsage{used: 1000000000000, avail: 11999000000000}, usage["/tank/apps"])
+	assert.Equal(t, 1, calls)
+
+	// Second call within the refresh window must not re-run the collector.
+	zm.DatasetUsage()
+	assert.Equal(t, 1, calls)
+}
+
+func TestDatasetUsageRefreshOnErrorKeepsPrevious(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
+		return []zfs.Dataset{{Name: "tank", Used: 1, Avail: 1, Mountpoint: "/tank"}}, nil
+	}
+	assert.Len(t, zm.DatasetUsage(), 1)
+
+	// Force refresh window expiry, then a failing collector.
+	zm.backends[0].lastUsageRefresh = time.Now().Add(-10 * time.Minute)
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
+		return nil, zfs.ErrNoZfs
+	}
+	usage := zm.DatasetUsage()
+	assert.Len(t, usage, 1, "previous usage should be retained on error")
+}
+
+func TestDatasetUsageClearsAbsentBackend(t *testing.T) {
+	b := newZfsBackend()
+	b.datasetUsage = map[string]zfsDatasetUsage{"/tank": {used: 1, avail: 1}}
+	b.datasetsFn = optionalPoolSource(func() ([]zfs.Dataset, error) {
+		return nil, zfs.ErrNoZfs
+	})
+
+	datasets, err := b.datasets()
+	require.NoError(t, err, "an absent backend must not produce an error to log")
+	assert.Empty(t, datasets)
+	b.refreshDatasetUsage()
+	assert.Empty(t, b.datasetUsage)
+	assert.False(t, b.lastUsageRefresh.IsZero())
+}
+
+func TestGetDetailForceRefresh(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	poolCalls := 0
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		poolCalls++
+		return []zfs.PoolStat{{Name: "tank", Alloc: uint64(poolCalls)}}, nil
+	}
+	zm.backends[0].poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, nil }
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
+
+	first := zm.GetDetail(false)
+	assert.True(t, first.Complete)
+	require.Len(t, first.Pools, 1)
+	assert.Equal(t, uint64(1), first.Pools[0].Alloc)
+
+	cached := zm.GetDetail(false)
+	require.Len(t, cached.Pools, 1)
+	assert.Equal(t, uint64(1), cached.Pools[0].Alloc)
+	assert.Equal(t, 1, poolCalls)
+
+	refreshed := zm.GetDetail(true)
+	assert.True(t, refreshed.Complete)
+	require.Len(t, refreshed.Pools, 1)
+	assert.Equal(t, uint64(2), refreshed.Pools[0].Alloc)
+	assert.Equal(t, 2, poolCalls)
+}
+
+func TestGetDetailSuccessfulEmptyInventoryClearsCache(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		return []zfs.PoolStat{{Name: "tank"}}, nil
+	}
+	zm.backends[0].poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, nil }
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
+
+	require.Len(t, zm.GetDetail(false).Pools, 1)
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) { return nil, nil }
+	empty := zm.GetDetail(true)
+	assert.True(t, empty.Complete)
+	assert.Empty(t, empty.Pools)
+}
+
+func TestGetDetailFailureReturnsIncompleteCachedInventory(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) {
+		return []zfs.PoolStat{{Name: "tank"}}, nil
+	}
+	zm.backends[0].poolStatusesFn = func() ([]zfs.PoolStatus, error) {
+		return []zfs.PoolStatus{{Name: "tank", Vdevs: []zfs.VdevStatus{{Name: "mirror-0"}}}}, nil
+	}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
+		return []zfs.Dataset{{Name: "tank/data"}}, nil
+	}
+	first := zm.GetDetail(false)
+	require.True(t, first.Complete)
+	require.Len(t, first.Pools[0].Vdevs, 1)
+	require.Len(t, first.Pools[0].Datasets, 1)
+
+	zm.backends[0].poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, zfs.ErrNoZfs }
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) { return nil, zfs.ErrNoZfs }
+	partial := zm.GetDetail(true)
+	require.True(t, partial.Complete)
+	require.Len(t, partial.Pools[0].Vdevs, 1)
+	require.Len(t, partial.Pools[0].Datasets, 1)
+
+	zm.backends[0].poolStatsFn = func() ([]zfs.PoolStat, error) { return nil, zfs.ErrNoZfs }
+	lastSuccessfulRefresh := zm.backends[0].lastDetailRefresh
+	failed := zm.GetDetail(true)
+	assert.False(t, failed.Complete)
+	require.Len(t, failed.Pools, 1)
+	assert.Equal(t, "tank", failed.Pools[0].Name)
+	assert.Equal(t, lastSuccessfulRefresh, zm.backends[0].lastDetailRefresh)
+}
+
+func TestZfsMountpoints(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs"}}}
+	zm.backends[0].datasetsFn = func() ([]zfs.Dataset, error) {
+		return []zfs.Dataset{
+			{Name: "tank", Mountpoint: "/tank"},
+			{Name: "rpool/ROOT/pve-1", Mountpoint: "/"},
+		}, nil
+	}
+	mountpoints := zm.ZfsMountpoints()
+	assert.Len(t, mountpoints, 2)
+	assert.True(t, mountpoints["/tank"])
+	assert.True(t, mountpoints["/"])
+}
+
+func TestBtrfsRawCapacityPropagates(t *testing.T) {
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs", poolStatsFn: func() ([]zfs.PoolStat, error) {
+		return []zfs.PoolStat{btrfsPoolStats(btrfs.Filesystem{UUID: "raw", Name: "raw", Size: 200, Alloc: 100, Raw: true})}, nil
+	},
+		poolStatusesFn: func() ([]zfs.PoolStatus, error) { return nil, nil },
+		datasetsFn:     func() ([]zfs.Dataset, error) { return nil, nil }}}}
+	var stats system.Stats
+	zm.Update(&stats)
+	require.True(t, stats.ZfsPools["b:raw"].Raw)
+	detail := zm.GetDetail(true)
+	require.True(t, detail.Complete)
+	require.Len(t, detail.Pools, 1)
+	assert.True(t, detail.Pools[0].Raw)
+}
+
+func TestMarkDuplicatePoolCharts(t *testing.T) {
+	for _, tc := range []struct {
+		name, poolID, device string
+		raw                  bool
+		diskTotal            float64
+		wantUsage, wantIO    bool
+	}{
+		{"single device root", "fs1", "dm-0", false, 100, true, true},
+		{"multi device", "fs1", "", false, 100, true, false},
+		{"different IO device", "fs1", "nvme0n1", false, 100, true, false},
+		{"different filesystem", "fs2", "dm-0", false, 100, false, false},
+		{"unknown identity", "", "dm-0", false, 100, false, false},
+		{"raw usage", "fs1", "dm-0", true, 100, false, true},
+		{"failed disk collection", "fs1", "dm-0", false, 0, false, false},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs", poolData: []zfs.PoolStat{{Name: "arbitrary label", MountID: tc.poolID, IODevice: tc.device, Raw: tc.raw}}}}}
+			stats := &system.Stats{ZfsPools: map[string]*system.ZfsPool{"arbitrary label": {}}}
+			fs := map[string]*system.FsStats{"dm-0": {Root: true, Mountpoint: "/", DiskTotal: tc.diskTotal}}
+			zm.markDuplicateCharts(stats, fs, func(string) string { return "fs1" })
+			assert.Equal(t, tc.wantUsage, stats.ZfsPools["arbitrary label"].HideUsage)
+			assert.Equal(t, tc.wantIO, stats.ZfsPools["arbitrary label"].HideIO)
+			// Bind mounts and custom extra-filesystem names have the same identity.
+			fs["dm-0"].Root = false
+			fs["dm-0"].Mountpoint = "/extra-filesystems/storage"
+			fs["dm-0"].Name = "custom name"
+			stats.ZfsPools["arbitrary label"] = &system.ZfsPool{}
+			zm.markDuplicateCharts(stats, fs, func(string) string { return "fs1" })
+			assert.Equal(t, tc.wantUsage, stats.ZfsPools["arbitrary label"].HideUsage)
+			assert.Equal(t, tc.wantIO, stats.ZfsPools["arbitrary label"].HideIO)
+		})
+	}
+}
+
+func TestBtrfsPoolIdentities(t *testing.T) {
+	old := btrfsFilesystems
+	t.Cleanup(func() { btrfsFilesystems = old })
+	label := "tank"
+	btrfsFilesystems = func() ([]btrfs.Filesystem, error) {
+		return []btrfs.Filesystem{
+			{UUID: "11111111-1111-4111-8111-111111111111", Name: label, Size: 100, Health: "ONLINE", NRead: 100, Devices: []btrfs.Device{{Name: "first"}}},
+			{UUID: "22222222-2222-4222-8222-222222222222", Name: "tank", Size: 200, Health: "DEGRADED", NRead: 200, Devices: []btrfs.Device{{Name: "second"}}},
+		}, nil
+	}
+	zm := &StoragePoolManager{detailInterval: time.Hour, backends: []*poolBackend{{name: "zfs", poolStatsFn: func() ([]zfs.PoolStat, error) { return []zfs.PoolStat{{Name: "tank", Size: 300}}, nil },
+		kernelStatsFn: func() ([]zfs.PoolKernelStat, error) { return []zfs.PoolKernelStat{{Name: "tank", NRead: 300}}, nil },
+		poolStatusesFn: func() ([]zfs.PoolStatus, error) {
+			return []zfs.PoolStatus{{Name: "tank", Vdevs: []zfs.VdevStatus{{Name: "zfs-device"}}}}, nil
+		},
+
+		datasetsFn: func() ([]zfs.Dataset, error) { return []zfs.Dataset{{Name: "tank/data"}}, nil }}, newBtrfsBackend()}}
+	first := "b:11111111-1111-4111-8111-111111111111"
+	second := "b:22222222-2222-4222-8222-222222222222"
+	var stats system.Stats
+	zm.Update(&stats)
+	require.Len(t, stats.ZfsPools, 3)
+	assert.Contains(t, stats.ZfsPools, "tank")
+	assert.Equal(t, "ONLINE", stats.ZfsPools[first].Health)
+	assert.Equal(t, "DEGRADED", stats.ZfsPools[second].Health)
+	assert.Equal(t, uint64(100), zm.backends[1].kernelSamples[first].nread)
+	assert.Equal(t, uint64(200), zm.backends[1].kernelSamples[second].nread)
+	detail := zm.GetDetail(true)
+	require.Len(t, detail.Pools, 3)
+	assert.Equal(t, "zfs-device", detail.Pools[0].Vdevs[0].Name)
+	assert.Len(t, detail.Pools[0].Datasets, 1)
+	assert.Equal(t, "first", detail.Pools[1].Vdevs[0].Name)
+	assert.Empty(t, detail.Pools[1].Datasets)
+	assert.Equal(t, "second", detail.Pools[2].Vdevs[0].Name)
+	label = "renamed"
+	zm.backends[0].lastPoolStats = time.Time{}
+	zm.backends[1].lastPoolStats = time.Time{}
+	zm.Update(&stats)
+	require.Len(t, stats.ZfsPools, 3)
+	assert.Equal(t, "renamed", stats.ZfsPools[first].DisplayName)
+	assert.Equal(t, first, zm.GetDetail(true).Pools[1].Name)
+	assert.Equal(t, "renamed", zm.GetDetail(true).Pools[1].DisplayName)
+}

agent/system.go

@@ -12,6 +12,7 @@ import (
 
 	"github.com/henrygd/beszel"
 	"github.com/henrygd/beszel/agent/battery"
+	"github.com/henrygd/beszel/agent/btrfs"
 	"github.com/henrygd/beszel/agent/utils"
 	"github.com/henrygd/beszel/agent/zfs"
 	"github.com/henrygd/beszel/internal/entities/container"
@@ -219,8 +220,9 @@ func (a *Agent) getSystemStats(cacheTimeMs uint16) system.Stats {
 	// disk i/o (cache-aware per interval)
 	a.updateDiskIo(cacheTimeMs, &systemStats)
 
-	// zfs pool stats
-	a.zfsManager.Update(&systemStats)
+	// storage pool stats
+	a.storagePoolManager.Update(&systemStats)
+	a.storagePoolManager.markDuplicateCharts(&systemStats, a.fsStats, btrfs.MountID)
 
 	// network stats (per cache interval)
 	a.updateNetworkStats(cacheTimeMs, &systemStats)

agent/zfs/zfs.go

@@ -33,11 +33,15 @@ var ErrNoZfs = errors.New("zfs utilities unavailable")
 
 // PoolStat is a snapshot of a ZFS pool's capacity and health.
 type PoolStat struct {
-	Name   string
-	Size   uint64 // total capacity in bytes
-	Alloc  uint64 // allocated bytes
-	Free   uint64 // free bytes
-	Health string // ONLINE, DEGRADED, FAULTED, ...
+	DisplayName string // optional friendly name; Name remains the stable key
+	MountID     string // Btrfs filesystem identity, empty for other backends
+	IODevice    string // sole Btrfs member device, if known
+	Raw         bool   // physical accounting rather than usable filesystem space
+	Name        string
+	Size        uint64 // total capacity in bytes
+	Alloc       uint64 // allocated bytes
+	Free        uint64 // free bytes
+	Health      string // ONLINE, DEGRADED, FAULTED, ...
 }
 
 // PoolKernelStat is the inexpensive pool telemetry exposed by the ZFS kernel.
@@ -66,6 +70,9 @@ type Dataset struct {
 // PoolStats returns capacity and health for all pools on the system using
 // `zpool list`. Frequent health and I/O sampling uses PoolKernelStats instead.
 func PoolStats() ([]PoolStat, error) {
+	if err := checkZfsDevice(); err != nil {
+		return nil, err
+	}
 	out, err := commandOutput("zpool", "list", "-Hp", "-o", "name,size,alloc,free,health")
 	if err != nil {
 		var exitErr *exec.ExitError
@@ -80,6 +87,9 @@ func PoolStats() ([]PoolStat, error) {
 // Datasets returns all datasets on the system with usage and mountpoint
 // information using `zfs list` (recursive by default).
 func Datasets() ([]Dataset, error) {
+	if err := checkZfsDevice(); err != nil {
+		return nil, err
+	}
 	out, err := commandOutput("zfs", "list", "-Hp", "-o", "name,used,avail,mountpoint")
 	if err != nil {
 		return nil, fmt.Errorf("zfs list: %w", err)

agent/zfs/zfs_linux.go

@@ -13,7 +13,10 @@ import (
 	"strings"
 )
 
-var procZfsPath = "/proc/spl/kstat/zfs"
+var (
+	procZfsPath = "/proc/spl/kstat/zfs"
+	devZfsPath  = "/dev/zfs"
+)
 
 func ARCSize() (uint64, error) {
 	file, err := os.Open(filepath.Join(procZfsPath, "arcstats"))
@@ -40,6 +43,19 @@ func ARCSize() (uint64, error) {
 	return 0, fmt.Errorf("size field not found in arcstats")
 }
 
+// checkZfsDevice lets containers without /dev/zfs fail fast instead of
+// waiting for ZFS utility commands to time out.
+func checkZfsDevice() error {
+	_, err := os.Stat(devZfsPath)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return ErrNoZfs
+		}
+		return err
+	}
+	return nil
+}
+
 // PoolKernelStats reads pool state and cumulative I/O counters directly from
 // procfs. These kstats are the same interfaces used by node_exporter's Linux
 // ZFS collector and avoid keeping a `zpool iostat` subprocess alive.

agent/zfs/zfs_linux_test.go

@@ -88,3 +88,66 @@ func TestReadObjsetIORequiresAllCounters(t *testing.T) {
 	_, _, err := readObjsetIO(path)
 	require.Error(t, err)
 }
+
+func TestCollectorsSkipCommandsWhenDevZfsMissing(t *testing.T) {
+	root := t.TempDir()
+	oldDevZfsPath := devZfsPath
+	devZfsPath = filepath.Join(root, "missing")
+	t.Cleanup(func() { devZfsPath = oldDevZfsPath })
+
+	oldCommandOutput := commandOutput
+	commandOutput = func(name string, args ...string) ([]byte, error) {
+		t.Fatalf("unexpected %s call with %v", name, args)
+		return nil, nil
+	}
+	t.Cleanup(func() { commandOutput = oldCommandOutput })
+
+	_, err := PoolStats()
+	assert.ErrorIs(t, err, ErrNoZfs)
+	_, err = Datasets()
+	assert.ErrorIs(t, err, ErrNoZfs)
+}
+
+func TestDatasetsDelegatesWhenDevZfsPresent(t *testing.T) {
+	oldDevZfsPath := devZfsPath
+	devZfsPath = filepath.Join(t.TempDir(), "zfs")
+	require.NoError(t, os.WriteFile(devZfsPath, nil, 0o644))
+	t.Cleanup(func() { devZfsPath = oldDevZfsPath })
+
+	oldCommandOutput := commandOutput
+	commandOutput = func(name string, args ...string) ([]byte, error) {
+		assert.Equal(t, "zfs", name)
+		assert.Equal(t, []string{"list", "-Hp", "-o", "name,used,avail,mountpoint"}, args)
+		return []byte("tank\t50\t50\t/tank\n"), nil
+	}
+	t.Cleanup(func() { commandOutput = oldCommandOutput })
+
+	datasets, err := Datasets()
+	require.NoError(t, err)
+	assert.Equal(t, []Dataset{{Name: "tank", Used: 50, Avail: 50, Mountpoint: "/tank"}}, datasets)
+}
+
+func TestPoolStatsDelegatesToZpoolWhenDevZfsPresent(t *testing.T) {
+	root := t.TempDir()
+	devFile := filepath.Join(root, "zfs")
+	require.NoError(t, os.WriteFile(devFile, []byte(""), 0o644))
+
+	oldDevZfsPath := devZfsPath
+	devZfsPath = devFile
+	t.Cleanup(func() { devZfsPath = oldDevZfsPath })
+
+	oldCommandOutput := commandOutput
+	called := false
+	commandOutput = func(name string, args ...string) ([]byte, error) {
+		called = true
+		assert.Equal(t, "zpool", name)
+		assert.Equal(t, []string{"list", "-Hp", "-o", "name,size,alloc,free,health"}, args)
+		return []byte("tank\t100\t50\t50\tONLINE\n"), nil
+	}
+	t.Cleanup(func() { commandOutput = oldCommandOutput })
+
+	pools, err := PoolStats()
+	require.NoError(t, err)
+	assert.True(t, called)
+	assert.Equal(t, []PoolStat{{Name: "tank", Size: 100, Alloc: 50, Free: 50, Health: "ONLINE"}}, pools)
+}

agent/zfs/zfs_nonlinux.go

@@ -0,0 +1,9 @@
+//go:build !linux
+
+package zfs
+
+// The /dev/zfs probe is Linux-specific. Other platforms detect availability
+// through the ZFS utilities themselves.
+func checkZfsDevice() error {
+	return nil
+}

agent/zfs/zfs_nonlinux_test.go

@@ -0,0 +1,33 @@
+//go:build testing && !linux
+
+package zfs
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCollectorsUseUtilitiesOnNonLinux(t *testing.T) {
+	oldCommandOutput := commandOutput
+	commandOutput = func(name string, args ...string) ([]byte, error) {
+		switch name {
+		case "zpool":
+			return []byte("tank\t100\t50\t50\tONLINE\n"), nil
+		case "zfs":
+			return []byte("tank\t50\t50\t/tank\n"), nil
+		default:
+			t.Fatalf("unexpected command %s", name)
+			return nil, nil
+		}
+	}
+	t.Cleanup(func() { commandOutput = oldCommandOutput })
+
+	pools, err := PoolStats()
+	require.NoError(t, err)
+	assert.Equal(t, []PoolStat{{Name: "tank", Size: 100, Alloc: 50, Free: 50, Health: "ONLINE"}}, pools)
+	datasets, err := Datasets()
+	require.NoError(t, err)
+	assert.Equal(t, []Dataset{{Name: "tank", Used: 50, Avail: 50, Mountpoint: "/tank"}}, datasets)
+}

agent/zfs_pool.go

@@ -1,320 +0,0 @@
-package agent
-
-import (
-	"log/slog"
-	"strings"
-	"sync"
-	"time"
-
-	"github.com/henrygd/beszel/agent/zfs"
-	"github.com/henrygd/beszel/internal/entities/system"
-	zfsentity "github.com/henrygd/beszel/internal/entities/zfs"
-)
-
-// zfsDatasetUsage holds usage values for a ZFS dataset mountpoint.
-type zfsDatasetUsage struct {
-	used  uint64
-	avail uint64
-}
-
-// datasetUsageRefreshInterval controls how often `zfs list` is re-run for the
-// mountpoint usage map. Dataset inventory changes rarely.
-const datasetUsageRefreshInterval = 5 * time.Minute
-
-// poolStatsRefreshInterval controls how often `zpool list` is re-run for pool
-// capacity. Health and I/O are read from procfs on Linux, so the utility only
-// needs to refresh slow-moving space accounting.
-const poolStatsRefreshInterval = time.Minute
-
-type poolKernelSample struct {
-	nread  uint64
-	nwrite uint64
-	at     time.Time
-}
-
-// ZfsManager collects ZFS pool and dataset statistics. Collection functions
-// are fields so unit tests can substitute them (same pattern as
-// diskDiscovery.usageFn). It is safe for concurrent use by a single goroutine
-// only; callers must hold the agent lock like updateDiskUsage does.
-type ZfsManager struct {
-	poolStatsFn    func() ([]zfs.PoolStat, error)       // capacity/health source
-	datasetsFn     func() ([]zfs.Dataset, error)        // dataset inventory source
-	kernelStatsFn  func() ([]zfs.PoolKernelStat, error) // procfs pool state/I/O source
-	poolStatusesFn func() ([]zfs.PoolStatus, error)     // scrub/vdev detail source
-
-	poolData      []zfs.PoolStat // cached pool inventory (TTL below)
-	lastPoolStats time.Time
-	kernelSamples map[string]poolKernelSample
-
-	datasetUsage     map[string]zfsDatasetUsage // mountpoint -> usage
-	lastUsageRefresh time.Time
-
-	// Detail data (pools, vdevs, scrub, datasets) is cached and refreshed on
-	// an interval. Accessed from handler goroutines, so it is mutex-protected.
-	detailMu          sync.Mutex
-	detail            *zfsentity.ZfsData
-	lastDetailRefresh time.Time
-	detailInterval    time.Duration
-}
-
-// newZfsManager creates a ZfsManager wired to the system's ZFS utilities.
-func newZfsManager() *ZfsManager {
-	return &ZfsManager{
-		poolStatsFn:    zfs.PoolStats,
-		datasetsFn:     zfs.Datasets,
-		kernelStatsFn:  zfs.PoolKernelStats,
-		poolStatusesFn: zfs.PoolStatuses,
-		detailInterval: time.Hour,
-	}
-}
-
-// Update refreshes systemStats.ZfsPools with the latest pool data. I/O
-// throughput and health come from inexpensive kernel kstats on Linux. Pool
-// capacity and dataset usage come from separately cached utility calls. It is
-// a no-op when ZFS is absent.
-func (zm *ZfsManager) Update(systemStats *system.Stats) {
-	pools := zm.poolStats()
-	if len(pools) == 0 {
-		return
-	}
-
-	kernelStats, ioRates := zm.kernelStats()
-
-	if systemStats.ZfsPools == nil {
-		systemStats.ZfsPools = make(map[string]*system.ZfsPool, len(pools))
-	}
-	for i := range pools {
-		pool := &pools[i]
-		// Full precision, matching the dataset values below; the frontend
-		// formats any magnitude.
-		stats := &system.ZfsPool{
-			Total:  float64(pool.Size) / (1024 * 1024 * 1024),
-			Used:   float64(pool.Alloc) / (1024 * 1024 * 1024),
-			Health: pool.Health,
-		}
-		if kernel, exists := kernelStats[pool.Name]; exists && kernel.Health != "" {
-			stats.Health = kernel.Health
-		}
-		if io, exists := ioRates[pool.Name]; exists {
-			stats.ReadBytes = io.NRead
-			stats.WriteBytes = io.NWrite
-		}
-		slog.Debug("ZFS pool sample", "pool", pool.Name, "health", stats.Health, "used_gb", stats.Used, "read_bps", stats.ReadBytes, "write_bps", stats.WriteBytes)
-		systemStats.ZfsPools[pool.Name] = stats
-	}
-
-}
-
-// poolStats returns the cached pool inventory, re-running `zpool list` at most
-// every poolStatsRefreshInterval. On failure the previous inventory is
-// retained and the refresh is retried on the next cadence.
-func (zm *ZfsManager) poolStats() []zfs.PoolStat {
-	if zm.lastPoolStats.IsZero() || time.Since(zm.lastPoolStats) >= poolStatsRefreshInterval {
-		pools, err := zm.poolStatsFn()
-		if err != nil {
-			slog.Debug("ZFS pool stats unavailable", "err", err)
-		} else {
-			zm.poolData = pools
-		}
-		zm.lastPoolStats = time.Now()
-	}
-	return zm.poolData
-}
-
-// kernelStats reads cumulative pool counters and converts them to per-second
-// rates. Counter decreases indicate a pool export/import and reset the
-// baseline instead of producing an underflow spike.
-func (zm *ZfsManager) kernelStats() (map[string]zfs.PoolKernelStat, map[string]zfs.PoolIoStats) {
-	if zm.kernelStatsFn == nil {
-		return nil, nil
-	}
-	stats, err := zm.kernelStatsFn()
-	if err != nil {
-		slog.Debug("ZFS kernel stats unavailable", "err", err)
-		return nil, nil
-	}
-	now := time.Now()
-	byName := make(map[string]zfs.PoolKernelStat, len(stats))
-	rates := make(map[string]zfs.PoolIoStats, len(stats))
-	nextSamples := make(map[string]poolKernelSample, len(stats))
-	for _, stat := range stats {
-		byName[stat.Name] = stat
-		if previous, ok := zm.kernelSamples[stat.Name]; ok && now.After(previous.at) &&
-			stat.NRead >= previous.nread && stat.NWrite >= previous.nwrite {
-			seconds := now.Sub(previous.at).Seconds()
-			rates[stat.Name] = zfs.PoolIoStats{
-				NRead:  uint64(float64(stat.NRead-previous.nread) / seconds),
-				NWrite: uint64(float64(stat.NWrite-previous.nwrite) / seconds),
-			}
-		}
-		nextSamples[stat.Name] = poolKernelSample{nread: stat.NRead, nwrite: stat.NWrite, at: now}
-	}
-	zm.kernelSamples = nextSamples
-	return byName, rates
-}
-
-// refreshDatasetUsage re-runs `zfs list` when the refresh window has elapsed
-// and rebuilds the mountpoint-keyed usage map.
-func (zm *ZfsManager) refreshDatasetUsage() {
-	if !zm.lastUsageRefresh.IsZero() && time.Since(zm.lastUsageRefresh) < datasetUsageRefreshInterval {
-		return
-	}
-	datasets, err := zm.datasetsFn()
-	if err != nil {
-		slog.Debug("ZFS dataset usage unavailable", "err", err)
-	} else {
-		usage := make(map[string]zfsDatasetUsage, len(datasets))
-		for _, ds := range datasets {
-			if ds.Mountpoint != "" && ds.Mountpoint != "-" {
-				usage[ds.Mountpoint] = zfsDatasetUsage{used: ds.Used, avail: ds.Avail}
-			}
-		}
-		zm.datasetUsage = usage
-	}
-	zm.lastUsageRefresh = time.Now()
-}
-
-// DatasetUsage returns ZFS dataset usage keyed by mountpoint, refreshed at
-// most every datasetUsageRefreshInterval. On failure the previous map is
-// retained and a debug log is emitted.
-func (zm *ZfsManager) DatasetUsage() map[string]zfsDatasetUsage {
-	zm.refreshDatasetUsage()
-	return zm.datasetUsage
-}
-
-// GetDetail returns ZFS detail data (pool health, scrub, vdevs, datasets).
-// Scheduled requests use the cached snapshot until stale; manual requests can
-// force collection. On failure the previous snapshot is retained.
-func (zm *ZfsManager) GetDetail(force bool) *zfsentity.ZfsData {
-	zm.detailMu.Lock()
-	defer zm.detailMu.Unlock()
-
-	if force || zm.detail == nil || time.Since(zm.lastDetailRefresh) >= zm.detailInterval {
-		if data, err := zm.collectDetail(zm.detail); err != nil {
-			slog.Debug("ZFS detail collection failed", "err", err)
-			if zm.detail == nil {
-				return &zfsentity.ZfsData{}
-			}
-			return &zfsentity.ZfsData{Pools: zm.detail.Pools}
-		} else {
-			zm.detail = data
-			zm.lastDetailRefresh = time.Now()
-		}
-	}
-	if zm.detail == nil {
-		return &zfsentity.ZfsData{}
-	}
-	return zm.detail
-}
-
-// collectDetail builds a ZfsData payload from the current system state.
-func (zm *ZfsManager) collectDetail(previous *zfsentity.ZfsData) (*zfsentity.ZfsData, error) {
-	pools, err := zm.poolStatsFn()
-	if err != nil {
-		return nil, err
-	}
-	if len(pools) == 0 {
-		return &zfsentity.ZfsData{Pools: []*zfsentity.PoolDetail{}, Complete: true}, nil
-	}
-
-	statuses, statusErr := zm.poolStatusesFn()
-	if statusErr != nil {
-		slog.Debug("ZFS pool status unavailable", "err", statusErr)
-	}
-	datasets, datasetsErr := zm.datasetsFn()
-	if datasetsErr != nil {
-		slog.Debug("ZFS datasets unavailable", "err", datasetsErr)
-	}
-
-	statusByPool := make(map[string]zfs.PoolStatus, len(statuses))
-	for _, st := range statuses {
-		statusByPool[st.Name] = st
-	}
-
-	previousByPool := make(map[string]*zfsentity.PoolDetail)
-	if previous != nil {
-		for _, pool := range previous.Pools {
-			if pool != nil {
-				previousByPool[pool.Name] = pool
-			}
-		}
-	}
-
-	data := &zfsentity.ZfsData{Pools: make([]*zfsentity.PoolDetail, 0, len(pools)), Complete: true}
-	for i := range pools {
-		p := &pools[i]
-		detail := &zfsentity.PoolDetail{
-			Name:   p.Name,
-			Health: p.Health,
-			Size:   p.Size,
-			Alloc:  p.Alloc,
-			Free:   p.Free,
-		}
-		if st, ok := statusByPool[p.Name]; statusErr == nil && ok {
-			if st.Scrub.State != "" && st.Scrub.State != "NONE" {
-				detail.Scrub = &zfsentity.Scrub{
-					State:    st.Scrub.State,
-					Progress: st.Scrub.Progress,
-					Errors:   st.Scrub.Errors,
-				}
-			}
-			for _, v := range st.Vdevs {
-				detail.Vdevs = append(detail.Vdevs, &zfsentity.Vdev{
-					Name:         v.Name,
-					State:        v.State,
-					ReadErrs:     v.ReadErrs,
-					WriteErrs:    v.WriteErrs,
-					ChecksumErrs: v.ChecksumErrs,
-				})
-			}
-		} else {
-			if cached := previousByPool[p.Name]; cached != nil {
-				detail.Scrub = cached.Scrub
-				detail.Vdevs = cached.Vdevs
-			}
-		}
-		if datasetsErr == nil {
-			foundDataset := false
-			for _, ds := range datasets {
-				if poolOfDataset(ds.Name) == p.Name {
-					foundDataset = true
-					detail.Datasets = append(detail.Datasets, &zfsentity.Dataset{
-						Name:       ds.Name,
-						Used:       ds.Used,
-						Avail:      ds.Avail,
-						Mountpoint: ds.Mountpoint,
-					})
-				}
-			}
-			if !foundDataset {
-				if cached := previousByPool[p.Name]; cached != nil {
-					detail.Datasets = cached.Datasets
-				}
-			}
-		} else if cached := previousByPool[p.Name]; cached != nil {
-			detail.Datasets = cached.Datasets
-		}
-		data.Pools = append(data.Pools, detail)
-	}
-	return data, nil
-}
-
-// poolOfDataset returns the pool name for a dataset name (everything before
-// the first '/'). Datasets without a separator belong to a pool of the same
-// name.
-func poolOfDataset(name string) string {
-	if idx := strings.IndexByte(name, '/'); idx >= 0 {
-		return name[:idx]
-	}
-	return name
-}
-
-// ZfsMountpoints returns the set of mountpoints backed by ZFS datasets.
-func (zm *ZfsManager) ZfsMountpoints() map[string]bool {
-	usage := zm.DatasetUsage()
-	mountpoints := make(map[string]bool, len(usage))
-	for mountpoint := range usage {
-		mountpoints[mountpoint] = true
-	}
-	return mountpoints
-}

agent/zfs_pool_test.go

@@ -1,245 +0,0 @@
-//go:build testing
-
-package agent
-
-import (
-	"testing"
-	"time"
-
-	"github.com/henrygd/beszel/agent/zfs"
-	"github.com/henrygd/beszel/internal/entities/system"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestUpdatePopulatesZfsPools(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		return []zfs.PoolStat{{Name: "tank", Size: 23999000000000, Alloc: 12000000000000, Free: 11999000000000, Health: "DEGRADED"}}, nil
-	}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
-		return []zfs.Dataset{
-			{Name: "tank/apps", Used: 5000000000000, Avail: 11999000000000, Mountpoint: "/tank/apps"},
-			{Name: "tank/backup", Used: 6000000000000, Avail: 11999000000000, Mountpoint: "/tank/backup"},
-			// Small zvol (Proxmox VM EFI disk): must not round to zero.
-			{Name: "rpool/vm-100-disk-2", Used: 4194304, Avail: 0, Mountpoint: "-"},
-		}, nil
-	}
-	var kernelCalls int
-	zm.kernelStatsFn = func() ([]zfs.PoolKernelStat, error) {
-		kernelCalls++
-		return []zfs.PoolKernelStat{{
-			Name: "tank", Health: "ONLINE",
-			NRead: uint64(kernelCalls-1) * 1250, NWrite: uint64(kernelCalls-1) * 5120,
-		}}, nil
-	}
-
-	var stats system.Stats
-	// The first kernel sample establishes the cumulative-counter baseline.
-	zm.Update(&stats)
-	zm.kernelSamples["tank"] = poolKernelSample{at: time.Now().Add(-time.Second)}
-	zm.Update(&stats)
-	require.NotNil(t, stats.ZfsPools)
-	require.Contains(t, stats.ZfsPools, "tank")
-	assert.InDelta(t, 22350.8105, stats.ZfsPools["tank"].Total, 0.0001) // Size in GiB
-	assert.InDelta(t, 11175.8709, stats.ZfsPools["tank"].Used, 0.0001)  // Alloc in GiB
-	assert.Equal(t, "ONLINE", stats.ZfsPools["tank"].Health)
-	assert.InDelta(t, 1250, stats.ZfsPools["tank"].ReadBytes, 5)
-	assert.InDelta(t, 5120, stats.ZfsPools["tank"].WriteBytes, 5)
-
-}
-
-// TestUpdateKernelStatsMissing verifies pools without a kernel sample report zero
-// I/O instead of erroring.
-func TestUpdateKernelStatsMissing(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		return []zfs.PoolStat{{Name: "tank", Size: 1, Alloc: 1, Health: "ONLINE"}}, nil
-	}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
-	zm.kernelStatsFn = func() ([]zfs.PoolKernelStat, error) {
-		return nil, zfs.ErrNoZfs
-	}
-
-	var stats system.Stats
-	zm.Update(&stats)
-	require.NotNil(t, stats.ZfsPools)
-	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].ReadBytes)
-	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].WriteBytes)
-}
-
-func TestUpdateKernelCounterReset(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		return []zfs.PoolStat{{Name: "tank", Health: "ONLINE"}}, nil
-	}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
-	zm.kernelSamples = map[string]poolKernelSample{
-		"tank": {nread: 100, nwrite: 200, at: time.Now().Add(-time.Second)},
-	}
-	zm.kernelStatsFn = func() ([]zfs.PoolKernelStat, error) {
-		return []zfs.PoolKernelStat{{Name: "tank", Health: "ONLINE", NRead: 10, NWrite: 20}}, nil
-	}
-
-	var stats system.Stats
-	zm.Update(&stats)
-	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].ReadBytes)
-	assert.Equal(t, uint64(0), stats.ZfsPools["tank"].WriteBytes)
-}
-
-func TestUpdateNoZfs(t *testing.T) {
-	zm := &ZfsManager{}
-	calls := 0
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		calls++
-		return nil, zfs.ErrNoZfs
-	}
-
-	var stats system.Stats
-	zm.Update(&stats)
-	zm.Update(&stats)
-	assert.Nil(t, stats.ZfsPools)
-	assert.Equal(t, 1, calls, "failed pool discovery should be cached until the next refresh interval")
-}
-
-func TestUpdateEmptyPools(t *testing.T) {
-	zm := &ZfsManager{}
-	calls := 0
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		calls++
-		return nil, nil
-	}
-
-	var stats system.Stats
-	zm.Update(&stats)
-	zm.Update(&stats)
-	assert.Nil(t, stats.ZfsPools)
-	assert.Equal(t, 1, calls, "an empty pool inventory should be cached until the next refresh interval")
-}
-
-func TestDatasetUsage(t *testing.T) {
-	zm := &ZfsManager{}
-	calls := 0
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
-		calls++
-		return []zfs.Dataset{
-			{Name: "tank", Used: 12000000000000, Avail: 11999000000000, Mountpoint: "/tank"},
-			{Name: "tank/apps", Used: 1000000000000, Avail: 11999000000000, Mountpoint: "/tank/apps"},
-			{Name: "rpool", Used: 900000000000, Avail: 300000000000, Mountpoint: "-"}, // zvol/unmounted: excluded
-		}, nil
-	}
-
-	usage := zm.DatasetUsage()
-	require.Len(t, usage, 2)
-	assert.Equal(t, zfsDatasetUsage{used: 12000000000000, avail: 11999000000000}, usage["/tank"])
-	assert.Equal(t, zfsDatasetUsage{used: 1000000000000, avail: 11999000000000}, usage["/tank/apps"])
-	assert.Equal(t, 1, calls)
-
-	// Second call within the refresh window must not re-run the collector.
-	zm.DatasetUsage()
-	assert.Equal(t, 1, calls)
-}
-
-func TestDatasetUsageRefreshOnErrorKeepsPrevious(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
-		return []zfs.Dataset{{Name: "tank", Used: 1, Avail: 1, Mountpoint: "/tank"}}, nil
-	}
-	assert.Len(t, zm.DatasetUsage(), 1)
-
-	// Force refresh window expiry, then a failing collector.
-	zm.lastUsageRefresh = time.Now().Add(-10 * time.Minute)
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
-		return nil, zfs.ErrNoZfs
-	}
-	usage := zm.DatasetUsage()
-	assert.Len(t, usage, 1, "previous usage should be retained on error")
-}
-
-func TestGetDetailForceRefresh(t *testing.T) {
-	zm := &ZfsManager{detailInterval: time.Hour}
-	poolCalls := 0
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		poolCalls++
-		return []zfs.PoolStat{{Name: "tank", Alloc: uint64(poolCalls)}}, nil
-	}
-	zm.poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, nil }
-	zm.datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
-
-	first := zm.GetDetail(false)
-	assert.True(t, first.Complete)
-	require.Len(t, first.Pools, 1)
-	assert.Equal(t, uint64(1), first.Pools[0].Alloc)
-
-	cached := zm.GetDetail(false)
-	require.Len(t, cached.Pools, 1)
-	assert.Equal(t, uint64(1), cached.Pools[0].Alloc)
-	assert.Equal(t, 1, poolCalls)
-
-	refreshed := zm.GetDetail(true)
-	assert.True(t, refreshed.Complete)
-	require.Len(t, refreshed.Pools, 1)
-	assert.Equal(t, uint64(2), refreshed.Pools[0].Alloc)
-	assert.Equal(t, 2, poolCalls)
-}
-
-func TestGetDetailSuccessfulEmptyInventoryClearsCache(t *testing.T) {
-	zm := &ZfsManager{detailInterval: time.Hour}
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		return []zfs.PoolStat{{Name: "tank"}}, nil
-	}
-	zm.poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, nil }
-	zm.datasetsFn = func() ([]zfs.Dataset, error) { return nil, nil }
-
-	require.Len(t, zm.GetDetail(false).Pools, 1)
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) { return nil, nil }
-	empty := zm.GetDetail(true)
-	assert.True(t, empty.Complete)
-	assert.Empty(t, empty.Pools)
-}
-
-func TestGetDetailFailureReturnsIncompleteCachedInventory(t *testing.T) {
-	zm := &ZfsManager{detailInterval: time.Hour}
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) {
-		return []zfs.PoolStat{{Name: "tank"}}, nil
-	}
-	zm.poolStatusesFn = func() ([]zfs.PoolStatus, error) {
-		return []zfs.PoolStatus{{Name: "tank", Vdevs: []zfs.VdevStatus{{Name: "mirror-0"}}}}, nil
-	}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
-		return []zfs.Dataset{{Name: "tank/data"}}, nil
-	}
-	first := zm.GetDetail(false)
-	require.True(t, first.Complete)
-	require.Len(t, first.Pools[0].Vdevs, 1)
-	require.Len(t, first.Pools[0].Datasets, 1)
-
-	zm.poolStatusesFn = func() ([]zfs.PoolStatus, error) { return nil, zfs.ErrNoZfs }
-	zm.datasetsFn = func() ([]zfs.Dataset, error) { return nil, zfs.ErrNoZfs }
-	partial := zm.GetDetail(true)
-	require.True(t, partial.Complete)
-	require.Len(t, partial.Pools[0].Vdevs, 1)
-	require.Len(t, partial.Pools[0].Datasets, 1)
-
-	zm.poolStatsFn = func() ([]zfs.PoolStat, error) { return nil, zfs.ErrNoZfs }
-	lastSuccessfulRefresh := zm.lastDetailRefresh
-	failed := zm.GetDetail(true)
-	assert.False(t, failed.Complete)
-	require.Len(t, failed.Pools, 1)
-	assert.Equal(t, "tank", failed.Pools[0].Name)
-	assert.Equal(t, lastSuccessfulRefresh, zm.lastDetailRefresh)
-}
-
-func TestZfsMountpoints(t *testing.T) {
-	zm := &ZfsManager{}
-	zm.datasetsFn = func() ([]zfs.Dataset, error) {
-		return []zfs.Dataset{
-			{Name: "tank", Mountpoint: "/tank"},
-			{Name: "rpool/ROOT/pve-1", Mountpoint: "/"},
-		}, nil
-	}
-	mountpoints := zm.ZfsMountpoints()
-	assert.Len(t, mountpoints, 2)
-	assert.True(t, mountpoints["/tank"])
-	assert.True(t, mountpoints["/"])
-}

beszel.go

@@ -6,7 +6,7 @@ import "github.com/blang/semver"
 
 const (
 	// Version is the current version of the application.
-	Version = "0.19.0"
+	Version = "0.20.0"
 	// AppName is the name of the application.
 	AppName = "beszel"
 )
@@ -19,3 +19,6 @@ var MinVersionAgentResponse = semver.MustParse("0.13.0")
 
 // MinVersionZfsData is the minimum agent version that supports ZFS detail requests.
 var MinVersionZfsData = semver.MustParse("0.18.9")
+
+// MinVersionNetworkMonitors is the minimum agent version that supports network monitor sync.
+var MinVersionNetworkMonitors = semver.MustParse("0.20.0")

go.mod

@@ -5,12 +5,13 @@ go 1.27.1
 require (
 	github.com/blang/semver v3.5.1+incompatible
 	github.com/coreos/go-systemd/v22 v22.7.0
+	github.com/distribution/reference v0.6.0
 	github.com/ebitengine/purego v0.11.0
 	github.com/fxamacker/cbor/v2 v2.9.3
 	github.com/gliderlabs/ssh v0.3.8
-	github.com/google/uuid v1.6.0
 	github.com/lxzan/gws v1.10.1
-	github.com/nicholas-fedor/shoutrrr v0.19.0
+	github.com/nicholas-fedor/shoutrrr v0.20.0
+	github.com/opencontainers/go-digest v1.0.0
 	github.com/pocketbase/dbx v1.12.0
 	github.com/pocketbase/pocketbase v0.40.2
 	github.com/shirou/gopsutil/v4 v4.26.8
@@ -41,6 +42,7 @@ require (
 	github.com/go-sql-driver/mysql v1.9.1 // indirect
 	github.com/godbus/dbus/v5 v5.2.2 // indirect
 	github.com/golang-jwt/jwt/v5 v5.3.1 // indirect
+	github.com/google/uuid v1.6.0 // indirect
 	github.com/gorilla/websocket v1.5.3 // indirect
 	github.com/inconshreveable/mousetrap v1.1.0 // indirect
 	github.com/klauspost/compress v1.20.0 // indirect

go.sum

@@ -15,6 +15,8 @@ github.com/cpuguy83/go-md2man/v2 v2.0.6/go.mod h1:oOW0eioCTA6cOiMLiUPZOpcVxMig6N
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/disintegration/imaging v1.6.2 h1:w1LecBlG2Lnp8B3jk5zSuNqd7b4DXhcjwek1ei82L+c=
 github.com/disintegration/imaging v1.6.2/go.mod h1:44/5580QXChDfwIclfc/PCwrr44amcmDAg8hxG0Ewe4=
+github.com/distribution/reference v0.6.0 h1:0IXCQ5g4/QMHHkarYzh5l+u8T3t73zM5QvfrDyIgxBk=
+github.com/distribution/reference v0.6.0/go.mod h1:BbU0aIcezP1/5jX/8MP0YiH4SdvB5Y4f/wlDRiLyi3E=
 github.com/domodwyer/mailyak/v3 v3.6.2 h1:x3tGMsyFhTCaxp6ycgR0FE/bu5QiNp+hetUuCOBXMn8=
 github.com/domodwyer/mailyak/v3 v3.6.2/go.mod h1:lOm/u9CyCVWHeaAmHIdF4RiKVxKUT/H5XX10lIKAL6c=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
@@ -54,8 +56,8 @@ github.com/golang-jwt/jwt/v5 v5.3.1/go.mod h1:fxCRLWMO43lRc8nhHWY6LGqRcf+1gQWArs
 github.com/golang/protobuf v1.3.1/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U=
 github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
 github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
-github.com/google/pprof v0.0.0-20260902005441-ca85771921e4 h1:/6mPXfWmhv8eKck12I0YNIcIjwHtxP3YRIMKiEgTjWg=
-github.com/google/pprof v0.0.0-20260902005441-ca85771921e4/go.mod h1:jl5iWTm0/hd5PjEYEOuwAJ57L/CibdZfrqZ5XA5GrCk=
+github.com/google/pprof v0.0.0-20260906184651-6331bc6350fe h1:QAinXoAFJdGQYztXn3VpFey7KCwpedbZ/EkzbplQ0cY=
+github.com/google/pprof v0.0.0-20260906184651-6331bc6350fe/go.mod h1:jl5iWTm0/hd5PjEYEOuwAJ57L/CibdZfrqZ5XA5GrCk=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/gorilla/websocket v1.5.3 h1:saDtZ6Pbx/0u+bgYQ3q96pZgCzfhKXGPqt7kZ72aNNg=
@@ -83,12 +85,14 @@ github.com/mattn/go-isatty v0.0.24 h1:tGZZoVgT/KiqK1c8ocVLeDS8BSWMRd47J3Lbz7vsRe
 github.com/mattn/go-isatty v0.0.24/go.mod h1:nMCL3Zebbrt45jsMDgnfIwz6ydEQApk5oEI3HqDio6A=
 github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=
 github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=
-github.com/nicholas-fedor/shoutrrr v0.19.0 h1:Rl6bpK3DXuR2Trtx2JV8t+wjUwkHdRHrc8nBKoEpHr0=
-github.com/nicholas-fedor/shoutrrr v0.19.0/go.mod h1:Glfdi8AGTbnEn2k2+hW62n8oL0i9vqRVFtXaUIthNks=
+github.com/nicholas-fedor/shoutrrr v0.20.0 h1:hMAxIYlfAeZ1FcTDgU0kUOvVXUsOirWo8IWlnzGLkac=
+github.com/nicholas-fedor/shoutrrr v0.20.0/go.mod h1:hgde37yNWCXh8+N6WemyDRMNYLOFTf326GsBx8Z7CFA=
 github.com/onsi/ginkgo/v2 v2.32.1 h1:6tlvcDm/3sE8lGJbZ4+d4mO3RLy24/tQWOFzVSQNIfw=
 github.com/onsi/ginkgo/v2 v2.32.1/go.mod h1:+aXOY+vzZ5mu2iI2HpTZUPmM//oQfsNFX6gU9kNcA44=
 github.com/onsi/gomega v1.43.0 h1:VlG/1FxqNxhSO+lq/OHBNaaqwiBK/mO8JbVkX9Y+FeU=
 github.com/onsi/gomega v1.43.0/go.mod h1:REff/hsDsodHoKlWsP2mAPhu1+5/6hVYNf9rIEBpeSg=
+github.com/opencontainers/go-digest v1.0.0 h1:apOUWs51W5PlhuyGyz9FCeeBIOUDA/6nW8Oi/yOhh5U=
+github.com/opencontainers/go-digest v1.0.0/go.mod h1:0JzlMkj0TRzQZfJkVvzbP0HBR3IKzErnv2BNG4W4MAM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/pocketbase/dbx v1.12.0 h1:/oLErM+A0b4xI0PWTGPqSDVjzix48PqI/bng2l0PzoA=
 github.com/pocketbase/dbx v1.12.0/go.mod h1:xXRCIAKTHMgUCyCKZm55pUOdvFziJjQfXaWKhu2vhMs=

internal/alerts/alerts.go

@@ -20,10 +20,11 @@ type hubLike interface {
 }
 
 type AlertManager struct {
-	hub                    hubLike
-	stopOnce               sync.Once
-	pendingAlerts          sync.Map
-	alertsCache            *AlertsCache
+	hub             hubLike
+	stopOnce        sync.Once
+	pendingAlerts   sync.Map
+	alertsCache     *AlertsCache
+	networkMonitors *networkMonitorCache
 }
 
 type AlertMessageData struct {
@@ -66,6 +67,7 @@ type SystemAlertGPUData struct {
 }
 
 type SystemAlertZfsPool struct {
+	Raw   bool    `json:"raw,omitempty"`
 	Total float64 `json:"d"`
 	Used  float64 `json:"du"`
 }
@@ -106,8 +108,9 @@ var supportsTitle = map[string]struct{}{
 // NewAlertManager creates a new AlertManager instance.
 func NewAlertManager(app hubLike) *AlertManager {
 	am := &AlertManager{
-		hub:         app,
-		alertsCache: NewAlertsCache(app),
+		hub:             app,
+		alertsCache:     NewAlertsCache(app),
+		networkMonitors: newNetworkMonitorCache(app),
 	}
 	am.bindEvents()
 	return am
@@ -115,6 +118,7 @@ func NewAlertManager(app hubLike) *AlertManager {
 
 // Bind events to the alerts collection lifecycle
 func (am *AlertManager) bindEvents() {
+	am.bindNetworkMonitorAlertEvents()
 	am.hub.OnRecordAfterUpdateSuccess("alerts").BindFunc(updateHistoryOnAlertUpdate)
 	am.hub.OnRecordAfterDeleteSuccess("alerts").BindFunc(resolveHistoryOnAlertDelete)
 	am.hub.OnRecordAfterUpdateSuccess("smart_devices").BindFunc(am.handleSmartDeviceAlert)
@@ -231,8 +235,20 @@ func (am *AlertManager) SendAlert(data AlertMessageData) error {
 		am.hub.Logger().Error("Failed to unmarshal user settings", "err", err)
 	}
 	// send alerts via webhooks
+	send := sendPublicNotification
+	if len(userAlertSettings.Webhooks) > 0 {
+		// Read the owner's current role at delivery time, including for URLs
+		// saved before an admin was demoted. Never fall back on lookup failure.
+		owner, err := am.hub.FindRecordById("users", data.UserID)
+		if err != nil {
+			return fmt.Errorf("load notification owner: %w", err)
+		}
+		if owner.GetString("role") == "admin" {
+			send = shoutrrr.Send
+		}
+	}
 	for _, webhook := range userAlertSettings.Webhooks {
-		if err := am.SendShoutrrrAlert(webhook, data.Title, data.Message, data.Link, data.LinkText); err != nil {
+		if err := am.sendShoutrrrAlert(webhook, data.Title, data.Message, data.Link, data.LinkText, send); err != nil {
 			am.hub.Logger().Error("Failed to send shoutrrr alert", "err", err)
 		}
 	}
@@ -263,6 +279,10 @@ func (am *AlertManager) SendAlert(data AlertMessageData) error {
 
 // SendShoutrrrAlert sends an alert via a Shoutrrr URL
 func (am *AlertManager) SendShoutrrrAlert(notificationUrl, title, message, link, linkText string) error {
+	return am.sendShoutrrrAlert(notificationUrl, title, message, link, linkText, shoutrrr.Send)
+}
+
+func (am *AlertManager) sendShoutrrrAlert(notificationUrl, title, message, link, linkText string, send func(string, string) error) error {
 	// Parse the URL
 	parsedURL, err := url.Parse(notificationUrl)
 	if err != nil {
@@ -305,7 +325,7 @@ func (am *AlertManager) SendShoutrrrAlert(notificationUrl, title, message, link,
 	parsedURL.RawQuery = queryParams.Encode()
 	// log.Println("URL after modification:", parsedURL.String())
 
-	err = shoutrrr.Send(parsedURL.String(), message)
+	err = send(parsedURL.String(), message)
 
 	if err == nil {
 		am.hub.Logger().Info("Sent shoutrrr alert", "title", title)

internal/alerts/alerts_api.go

@@ -3,13 +3,11 @@ package alerts
 import (
 	"database/sql"
 	"errors"
-	"net"
 	"net/http"
-	"net/url"
 	"slices"
-	"strings"
 
 	"github.com/henrygd/beszel/internal/hub/utils"
+	"github.com/nicholas-fedor/shoutrrr"
 	"github.com/pocketbase/dbx"
 	"github.com/pocketbase/pocketbase/core"
 )
@@ -31,6 +29,13 @@ func UpsertUserAlerts(e *core.RequestEvent) error {
 		return e.BadRequestError("Bad data", err)
 	}
 
+	if reqData.Name == alertNameNetworkMonitorLoss {
+		if reqData.Value < 0 || reqData.Value >= 100 {
+			return e.BadRequestError("Monitor loss threshold must be at least 0 and below 100", nil)
+		}
+		reqData.Min = 0
+	}
+
 	alertsCollection, err := e.App.FindCachedCollectionByNameOrId("alerts")
 	if err != nil {
 		return err
@@ -147,72 +152,16 @@ func (am *AlertManager) SendTestNotification(e *core.RequestEvent) error {
 	if err != nil || data.URL == "" {
 		return e.BadRequestError("URL is required", err)
 	}
-	// Only allow admins to send test notifications to internal URLs
+	send := shoutrrr.Send
 	if !e.Auth.IsSuperuser() && e.Auth.GetString("role") != "admin" {
-		internalURL, err := isInternalURL(data.URL)
-		if err != nil {
-			return e.BadRequestError(err.Error(), nil)
-		}
-		if internalURL {
-			return e.ForbiddenError("Only admins can send to internal destinations", nil)
-		}
+		send = sendPublicNotification
+	}
+	err = am.sendShoutrrrAlert(data.URL, "Test Alert", "This is a notification from Beszel.", am.hub.Settings().Meta.AppURL, "View Beszel", send)
+	if errors.Is(err, errInternalDestination) || errors.Is(err, errUnrestrictedService) {
+		return e.ForbiddenError(err.Error(), nil)
 	}
-	err = am.SendShoutrrrAlert(data.URL, "Test Alert", "This is a notification from Beszel.", am.hub.Settings().Meta.AppURL, "View Beszel")
 	if err != nil {
 		return e.JSON(200, map[string]string{"err": err.Error()})
 	}
 	return e.JSON(200, map[string]bool{"err": false})
 }
-
-// isInternalURL checks if the given shoutrrr URL points to an internal destination (localhost or private IP)
-func isInternalURL(rawURL string) (bool, error) {
-	parsedURL, err := url.Parse(rawURL)
-	if err != nil {
-		return false, err
-	}
-
-	host := parsedURL.Hostname()
-	if host == "" {
-		return false, nil
-	}
-
-	if strings.EqualFold(host, "localhost") {
-		return true, nil
-	}
-
-	if ip := net.ParseIP(host); ip != nil {
-		return isInternalIP(ip), nil
-	}
-
-	// Some Shoutrrr URLs use the host position for service identifiers rather than a
-	// network hostname (for example, discord://token@webhookid). Restrict DNS lookups
-	// to names that look like actual hostnames so valid service URLs keep working.
-	if !strings.Contains(host, ".") {
-		return false, nil
-	}
-
-	ips, err := net.LookupIP(host)
-	if err != nil {
-		return false, nil
-	}
-
-	if slices.ContainsFunc(ips, isInternalIP) {
-		return true, nil
-	}
-
-	return false, nil
-}
-
-var cgnatNetwork = &net.IPNet{
-	IP:   net.IPv4(100, 64, 0, 0),
-	Mask: net.CIDRMask(10, 32),
-}
-
-func isInternalIP(ip net.IP) bool {
-	return ip.IsPrivate() ||
-		ip.IsLoopback() ||
-		ip.IsUnspecified() ||
-		ip.IsLinkLocalUnicast() ||
-		ip.IsMulticast() ||
-		cgnatNetwork.Contains(ip)
-}

internal/alerts/alerts_api_test.go

@@ -7,10 +7,11 @@ import (
 	"encoding/json"
 	"io"
 	"net/http"
+	"net/http/httptest"
 	"strings"
+	"sync/atomic"
 	"testing"
 
-	"github.com/henrygd/beszel/internal/alerts"
 	beszelTests "github.com/henrygd/beszel/internal/tests"
 	pbTests "github.com/pocketbase/pocketbase/tests"
 
@@ -29,43 +30,6 @@ func jsonReader(v any) io.Reader {
 	return bytes.NewReader(data)
 }
 
-func TestIsInternalURL(t *testing.T) {
-	testCases := []struct {
-		name     string
-		url      string
-		internal bool
-	}{
-		{name: "loopback ipv4", url: "generic://127.0.0.1", internal: true},
-		{name: "private ipv4", url: "generic://10.0.0.1", internal: true},
-		{name: "localhost hostname", url: "generic://localhost", internal: true},
-		{name: "localhost with path", url: "generic+http://localhost/api/v1/postStuff", internal: true},
-		{name: "loopback with port and path", url: "generic+http://127.0.0.1:8080/api/v1/postStuff", internal: true},
-		{name: "public hostname", url: "generic+https://beszel.dev/api/v1/postStuff", internal: false},
-		{name: "cloud metadata ipv4", url: "generic://169.254.169.254", internal: true},
-		{name: "link-local ipv4", url: "generic://169.254.1.1", internal: true},
-		{name: "link-local ipv6", url: "generic://[fe80::1]", internal: true},
-		{name: "mapped link-local ipv4", url: "generic://[::ffff:169.254.169.254]", internal: true},
-		{name: "cgnat lower boundary", url: "generic://100.64.0.0", internal: true},
-		{name: "cgnat upper boundary", url: "generic://100.127.255.255", internal: true},
-		{name: "below cgnat", url: "generic://100.63.255.255", internal: false},
-		{name: "above cgnat", url: "generic://100.128.0.0", internal: false},
-		{name: "multicast ipv4", url: "generic://224.0.0.1", internal: true},
-		{name: "multicast ipv6", url: "generic://[ff02::1]", internal: true},
-		{name: "public ipv4", url: "generic://8.8.8.8", internal: false},
-		{name: "public ipv6", url: "generic://[2001:4860:4860::8888]", internal: false},
-		{name: "token style service url", url: "discord://abc123@123456789", internal: false},
-		{name: "single label service url", url: "slack://token@team/channel", internal: false},
-	}
-
-	for _, testCase := range testCases {
-		t.Run(testCase.name, func(t *testing.T) {
-			internal, err := alerts.IsInternalURL(testCase.url)
-			assert.NoError(t, err)
-			assert.Equal(t, testCase.internal, internal)
-		})
-	}
-}
-
 func TestUserAlertsApi(t *testing.T) {
 	hub, _ := beszelTests.NewTestHub(t.TempDir())
 	defer hub.Cleanup()
@@ -457,6 +421,17 @@ func TestSendTestNotification(t *testing.T) {
 	hub, user := beszelTests.GetHubWithUser(t)
 	defer hub.Cleanup()
 
+	var delivered atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		delivered.Add(1)
+	}))
+	defer server.Close()
+	localURL := "generic+" + server.URL
+
+	readonlyUser, err := beszelTests.CreateUserWithRole(hub, "readonly@example.com", "password123", "readonly")
+	assert.NoError(t, err)
+	readonlyToken, err := readonlyUser.NewAuthToken()
+	assert.NoError(t, err)
 	userToken, err := user.NewAuthToken()
 
 	adminUser, err := beszelTests.CreateUserWithRole(hub, "admin@example.com", "password123", "admin")
@@ -481,11 +456,11 @@ func TestSendTestNotification(t *testing.T) {
 			ExpectedContent: []string{"requires valid"},
 			TestAppFactory:  testAppFactory,
 			Body: jsonReader(map[string]any{
-				"url": "generic://127.0.0.1",
+				"url": localURL,
 			}),
 		},
 		{
-			Name:           "POST /test-notification - with external auth should succeed",
+			Name:           "POST /test-notification - invalid service reports error",
 			Method:         http.MethodPost,
 			URL:            "/api/beszel/test-notification",
 			TestAppFactory: testAppFactory,
@@ -493,7 +468,7 @@ func TestSendTestNotification(t *testing.T) {
 				"Authorization": userToken,
 			},
 			Body: jsonReader(map[string]any{
-				"url": "generic://8.8.8.8",
+				"url": "unknown://example.com",
 			}),
 			ExpectedStatus:  200,
 			ExpectedContent: []string{"\"err\":"},
@@ -535,10 +510,10 @@ func TestSendTestNotification(t *testing.T) {
 				"Authorization": adminUserToken,
 			},
 			Body: jsonReader(map[string]any{
-				"url": "generic://127.0.0.1",
+				"url": localURL,
 			}),
 			ExpectedStatus:  200,
-			ExpectedContent: []string{"\"err\":"},
+			ExpectedContent: []string{"\"err\":false"},
 		},
 		{
 			Name:           "POST /test-notification - internal url with superuser auth should succeed",
@@ -549,14 +524,28 @@ func TestSendTestNotification(t *testing.T) {
 				"Authorization": superuserToken,
 			},
 			Body: jsonReader(map[string]any{
-				"url": "generic://127.0.0.1",
+				"url": localURL,
 			}),
 			ExpectedStatus:  200,
 			ExpectedContent: []string{"\"err\":"},
 		},
 	}
 
+	for _, url := range []string{localURL, "smtp://user:pass@127.0.0.1/?fromAddress=sender@example.com&toAddresses=recipient@example.com", "mqtt://127.0.0.1/topic"} {
+		scenarios = append(scenarios, beszelTests.ApiScenario{
+			Name:            "readonly cannot send to " + url,
+			Method:          http.MethodPost,
+			URL:             "/api/beszel/test-notification",
+			TestAppFactory:  testAppFactory,
+			Headers:         map[string]string{"Authorization": readonlyToken},
+			Body:            jsonReader(map[string]any{"url": url}),
+			ExpectedStatus:  403,
+			ExpectedContent: []string{"Only admins"},
+		})
+	}
+
 	for _, scenario := range scenarios {
 		scenario.Test(t)
 	}
+	assert.EqualValues(t, 2, delivered.Load(), "only admin and superuser requests should reach the server")
 }

internal/alerts/alerts_cache.go

@@ -1,6 +1,7 @@
 package alerts
 
 import (
+	"sync"
 	"time"
 
 	"github.com/pocketbase/dbx"
@@ -18,6 +19,9 @@ type CachedAlertData struct {
 	Triggered    bool
 	Min          uint8
 	PendingSince time.Time
+	// Immutable after publication; decoded only when the alert record changes.
+	MonitorStates      map[string]string
+	MonitorStatesValid bool
 	// Created   types.DateTime
 }
 
@@ -30,11 +34,18 @@ func (a *CachedAlertData) PopulateFromRecord(record *core.Record) {
 	a.Triggered = record.GetBool("triggered")
 	a.Min = uint8(record.GetInt("min"))
 	a.PendingSince = record.GetDateTime("pending_since").Time()
+	if a.Name == alertNameNetworkMonitorLoss {
+		var state networkMonitorAlertState
+		a.MonitorStatesValid = record.UnmarshalJSONField("state", &state) == nil
+		a.MonitorStates = state.Monitors
+	}
 	// a.Created = record.GetDateTime("created")
 }
 
 // AlertsCache provides an in-memory cache for system alerts.
 type AlertsCache struct {
+	// Serialize lazy loads with updates so a late load cannot replace newer state.
+	loadMu    sync.Mutex
 	app       core.App
 	store     *store.Store[string, *store.Store[string, CachedAlertData]]
 	populated bool
@@ -69,6 +80,8 @@ func (c *AlertsCache) bindEvents() *AlertsCache {
 
 // PopulateFromDB clears current entries and loads all alerts from the database into the cache.
 func (c *AlertsCache) PopulateFromDB(force bool) error {
+	c.loadMu.Lock()
+	defer c.loadMu.Unlock()
 	if !force && c.populated {
 		return nil
 	}
@@ -78,7 +91,7 @@ func (c *AlertsCache) PopulateFromDB(force bool) error {
 	}
 	c.store.RemoveAll()
 	for _, record := range records {
-		c.Update(record)
+		c.update(record)
 	}
 	c.populated = true
 	return nil
@@ -86,6 +99,12 @@ func (c *AlertsCache) PopulateFromDB(force bool) error {
 
 // Update adds or updates an alert record in the cache.
 func (c *AlertsCache) Update(record *core.Record) {
+	c.loadMu.Lock()
+	defer c.loadMu.Unlock()
+	c.update(record)
+}
+
+func (c *AlertsCache) update(record *core.Record) {
 	systemID := record.GetString("system")
 	if systemID == "" {
 		return
@@ -102,6 +121,8 @@ func (c *AlertsCache) Update(record *core.Record) {
 
 // Delete removes an alert record from the cache.
 func (c *AlertsCache) Delete(record *core.Record) {
+	c.loadMu.Lock()
+	defer c.loadMu.Unlock()
 	systemID := record.GetString("system")
 	if systemID == "" {
 		return
@@ -115,18 +136,23 @@ func (c *AlertsCache) Delete(record *core.Record) {
 func (c *AlertsCache) GetSystemAlerts(systemID string) []CachedAlertData {
 	systemStore, ok := c.store.GetOk(systemID)
 	if !ok {
-		// Populate cache for this system
-		records, err := c.app.FindAllRecords("alerts", dbx.NewExp("system={:system}", dbx.Params{"system": systemID}))
-		if err != nil {
-			return nil
+		c.loadMu.Lock()
+		defer c.loadMu.Unlock()
+		systemStore, ok = c.store.GetOk(systemID)
+		if !ok {
+			// Populate cache for this system
+			records, err := c.app.FindAllRecords("alerts", dbx.NewExp("system={:system}", dbx.Params{"system": systemID}))
+			if err != nil {
+				return nil
+			}
+			systemStore = store.New(map[string]CachedAlertData{})
+			for _, record := range records {
+				var ca CachedAlertData
+				ca.PopulateFromRecord(record)
+				systemStore.Set(record.Id, ca)
+			}
+			c.store.Set(systemID, systemStore)
 		}
-		systemStore = store.New(map[string]CachedAlertData{})
-		for _, record := range records {
-			var ca CachedAlertData
-			ca.PopulateFromRecord(record)
-			systemStore.Set(record.Id, ca)
-		}
-		c.store.Set(systemID, systemStore)
 	}
 	all := systemStore.GetAll()
 	alerts := make([]CachedAlertData, 0, len(all))

internal/alerts/alerts_history.go

@@ -9,6 +9,12 @@ import (
 
 // On triggered alert record delete, set matching alert history record to resolved
 func resolveHistoryOnAlertDelete(e *core.RecordEvent) error {
+	if e.Record.GetString("name") == alertNameNetworkMonitorLoss {
+		if err := resolveNetworkMonitorHistory(e.App, e.Record.Id); err != nil {
+			return err
+		}
+		return e.Next()
+	}
 	if !e.Record.GetBool("triggered") {
 		return e.Next()
 	}
@@ -18,6 +24,10 @@ func resolveHistoryOnAlertDelete(e *core.RecordEvent) error {
 
 // On alert record update, update alert history record
 func updateHistoryOnAlertUpdate(e *core.RecordEvent) error {
+	// Network monitor incidents have separate history entries per monitor.
+	if e.Record.GetString("name") == alertNameNetworkMonitorLoss {
+		return e.Next()
+	}
 	original := e.Record.Original()
 	new := e.Record
 

internal/alerts/alerts_network_monitors.go

@@ -0,0 +1,269 @@
+package alerts
+
+import (
+	"database/sql"
+	"errors"
+	"fmt"
+	"math"
+	"net"
+	"strconv"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/core"
+)
+
+const alertNameNetworkMonitorLoss = "NetworkMonitorLoss"
+
+// networkMonitorAlertState is this alert type's persisted runtime state.
+// Monitor IDs map to their open history entries independently of history retention.
+type networkMonitorAlertState struct {
+	Monitors map[string]string `json:"monitors"`
+}
+
+func (am *AlertManager) bindNetworkMonitorAlertEvents() {
+	// Hidden fields are still writable through the record API unless protected.
+	protectState := func(e *core.RecordRequestEvent) error {
+		e.Record.Set("state", e.Record.Original().Get("state"))
+		oldName, newName := e.Record.Original().GetString("name"), e.Record.GetString("name")
+		if oldName != "" && (oldName == alertNameNetworkMonitorLoss || newName == alertNameNetworkMonitorLoss) &&
+			(oldName != newName || e.Record.GetString("system") != e.Record.Original().GetString("system")) {
+			return e.BadRequestError("Delete and recreate the alert to change its type or system", nil)
+		}
+		if e.Record.GetString("name") == alertNameNetworkMonitorLoss {
+			if !e.HasSuperuserAuth() && (e.Auth == nil || !userHasSystem(e.App, e.Auth.Id, e.Record.GetString("system"))) {
+				return e.ForbiddenError("You do not have access to this system", nil)
+			}
+			e.Record.Set("triggered", e.Record.Original().GetBool("triggered"))
+			value := e.Record.GetFloat("value")
+			if math.IsNaN(value) || math.IsInf(value, 0) || value < 0 || value >= 100 {
+				return e.BadRequestError("Monitor loss threshold must be at least 0 and below 100", nil)
+			}
+			e.Record.Set("min", 0)
+		}
+		return e.Next()
+	}
+	am.hub.OnRecordCreateRequest("alerts").BindFunc(protectState)
+	am.hub.OnRecordUpdateRequest("alerts").BindFunc(protectState)
+	cleanup := func(e *core.RecordEvent) error {
+		if err := e.Next(); err != nil {
+			return err
+		}
+		return am.evaluateNetworkMonitorAlerts(e.App, e.Record.GetString("system"), nil)
+	}
+	am.hub.OnRecordAfterDeleteSuccess("network_monitors").BindFunc(cleanup)
+	am.hub.OnRecordAfterUpdateSuccess("network_monitors").BindFunc(func(e *core.RecordEvent) error {
+		if e.Record.GetBool("enabled") || !e.Record.Original().GetBool("enabled") {
+			return e.Next()
+		}
+		return cleanup(e)
+	})
+}
+
+// HandleNetworkMonitorAlerts runs after the full monitoring transaction commits,
+// using its exact payload (dashboard requests can replace the cached payload).
+// Omitted results and disconnected systems never imply recovery.
+func (am *AlertManager) HandleNetworkMonitorAlerts(systemRecord *core.Record, results map[string]monitor.Result) error {
+	if systemRecord.GetString("status") != "up" {
+		return nil
+	}
+	alerts := am.alertsCache.GetAlertsByName(systemRecord.Id, alertNameNetworkMonitorLoss)
+	if len(alerts) == 0 {
+		return nil
+	}
+	monitors, err := am.networkMonitors.get(systemRecord.Id)
+	if err != nil {
+		return err
+	}
+	if !networkMonitorTransitionPending(alerts, monitors, results, time.Now()) {
+		return nil
+	}
+	// The cache only predicts a transition. Reload and recheck under the DB
+	// transaction before persisting, including current system/monitor status.
+	return am.evaluateNetworkMonitorAlerts(am.hub, systemRecord.Id, results)
+}
+
+// networkMonitorTransitionPending does no IO and never mutates cached maps.
+func networkMonitorTransitionPending(alerts []CachedAlertData, monitors map[string]int, results map[string]monitor.Result, now time.Time) bool {
+	for _, alert := range alerts {
+		if !alert.MonitorStatesValid || alert.Triggered != (len(alert.MonitorStates) > 0) {
+			return true
+		}
+		for id := range alert.MonitorStates {
+			if _, enabled := monitors[id]; !enabled {
+				return true
+			}
+		}
+		for id, result := range results {
+			interval, enabled := monitors[id]
+			if !enabled || !monitorResultReady(result, interval, now) {
+				continue
+			}
+			_, active := alert.MonitorStates[id]
+			if (result.PacketLoss1h > alert.Value) != active {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func (am *AlertManager) evaluateNetworkMonitorAlerts(app core.App, systemID string, results map[string]monitor.Result) error {
+	var messages []AlertMessageData
+	err := app.RunInTransaction(func(tx core.App) error {
+		// Read configuration inside the transaction so concurrent threshold changes,
+		// disabling, and evaluations cannot overwrite each other's incident state.
+		alerts, err := tx.FindAllRecords("alerts", dbx.HashExp{"system": systemID, "name": alertNameNetworkMonitorLoss})
+		if err != nil || len(alerts) == 0 {
+			return err
+		}
+		system, err := tx.FindRecordById("systems", systemID)
+		if errors.Is(err, sql.ErrNoRows) {
+			// System deletion cascades to its alerts.
+			return nil
+		}
+		if err != nil {
+			return err
+		}
+		monitors, err := tx.FindAllRecords("network_monitors", dbx.HashExp{"system": systemID, "enabled": true})
+		if err != nil {
+			return err
+		}
+		enabled := make(map[string]*core.Record, len(monitors))
+		for _, m := range monitors {
+			enabled[m.Id] = m
+		}
+		now := time.Now()
+		for _, alert := range alerts {
+			var state networkMonitorAlertState
+			if err := alert.UnmarshalJSONField("state", &state); err != nil {
+				return err
+			}
+			states := state.Monitors
+			if states == nil {
+				states = map[string]string{}
+			}
+			changed := false
+			// Removing or disabling a monitor closes its incident silently.
+			for id, historyID := range states {
+				if _, ok := enabled[id]; !ok {
+					if err := resolveMonitorIncident(tx, historyID, now); err != nil {
+						return err
+					}
+					delete(states, id)
+					changed = true
+				}
+			}
+			if system.GetString("status") == "up" {
+				for _, m := range monitors {
+					result, ok := results[m.Id]
+					if !ok || !monitorResultReady(result, m.GetInt("interval"), now) {
+						continue
+					}
+					historyID, active := states[m.Id]
+					triggered := result.PacketLoss1h > alert.GetFloat("value")
+					if triggered == active {
+						continue
+					}
+					label := m.GetString("target")
+					if m.GetString("protocol") == "tcp" {
+						label = net.JoinHostPort(label, strconv.Itoa(m.GetInt("port")))
+					}
+					if triggered {
+						collection, err := tx.FindCachedCollectionByNameOrId("alerts_history")
+						if err != nil {
+							return err
+						}
+						history := core.NewRecord(collection)
+						history.Load(map[string]any{
+							"alert_id": alert.Id, "user": alert.GetString("user"), "system": systemID,
+							"name": alertNameNetworkMonitorLoss, "monitor_name": label, "value": result.PacketLoss1h,
+						})
+						if err := tx.Save(history); err != nil {
+							return err
+						}
+						states[m.Id] = history.Id
+					} else {
+						if err := resolveMonitorIncident(tx, historyID, now); err != nil {
+							return err
+						}
+						delete(states, m.Id)
+					}
+					changed = true
+					state, comparison := "loss", "exceeds"
+					if !triggered {
+						state, comparison = "recovered", "is at or below"
+					}
+					messages = append(messages, AlertMessageData{
+						UserID: alert.GetString("user"), SystemID: systemID,
+						Title:   fmt.Sprintf("Network monitor %s on %s: %s", state, system.GetString("name"), label),
+						Message: fmt.Sprintf("%s on %s: loss over the past hour is %.2f%%, which %s the %.2f%% threshold.", label, system.GetString("name"), result.PacketLoss1h, comparison, alert.GetFloat("value")),
+						Link:    am.hub.MakeLink("system", systemID), LinkText: "View " + system.GetString("name"),
+					})
+				}
+			}
+			if changed || alert.GetBool("triggered") != (len(states) > 0) {
+				alert.Set("state", networkMonitorAlertState{Monitors: states})
+				alert.Set("triggered", len(states) > 0)
+				if err := tx.Save(alert); err != nil {
+					return err
+				}
+			}
+		}
+		return nil
+	})
+	if err != nil {
+		return err
+	}
+	// Match other alert types: persist transitions before delivery, and respect
+	// the user's existing notification destinations and quiet hours.
+	for _, message := range messages {
+		if err := am.SendAlert(message); err != nil {
+			app.Logger().Error("Failed to send network monitor alert", "err", err)
+		}
+	}
+	return nil
+}
+
+func monitorResultReady(result monitor.Result, interval int, now time.Time) bool {
+	// Three completed attempts provide a short warm-up, including after an agent
+	// restart.
+	if result.SampleCount < 3 || result.LastProbeAt <= 0 || math.IsNaN(result.PacketLoss1h) || math.IsInf(result.PacketLoss1h, 0) || result.PacketLoss1h < 0 || result.PacketLoss1h > 100 {
+		return false
+	}
+	// Never interpret an empty one-hour window as zero loss.
+	maxAge := min(time.Hour, max(3*time.Duration(interval)*time.Second, 3*time.Minute))
+	age := now.Sub(time.UnixMilli(result.LastProbeAt))
+	return age >= -time.Minute && age <= maxAge
+}
+
+func resolveMonitorIncident(app core.App, id string, now time.Time) error {
+	record, err := app.FindRecordById("alerts_history", id)
+	if errors.Is(err, sql.ErrNoRows) {
+		// History can be purged independently.
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	if !record.GetDateTime("resolved").IsZero() {
+		return nil
+	}
+	record.Set("resolved", now.UTC())
+	return app.Save(record)
+}
+
+func resolveNetworkMonitorHistory(app core.App, alertID string) error {
+	records, err := app.FindAllRecords("alerts_history", dbx.HashExp{"alert_id": alertID, "resolved": ""})
+	if err != nil {
+		return err
+	}
+	for _, record := range records {
+		record.Set("resolved", time.Now().UTC())
+		if err := app.Save(record); err != nil {
+			return err
+		}
+	}
+	return nil
+}

internal/alerts/alerts_network_monitors_test.go

@@ -0,0 +1,513 @@
+//go:build testing
+
+package alerts_test
+
+import (
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/internal/alerts"
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	beszelTests "github.com/henrygd/beszel/internal/tests"
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/core"
+	pbTests "github.com/pocketbase/pocketbase/tests"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func networkAlertSetup(t *testing.T) (*beszelTests.TestHub, *core.Record, *core.Record, []*core.Record) {
+	t.Helper()
+	hub, system, alert := systemdTestSetup(t, false)
+	t.Cleanup(hub.Cleanup)
+	alert.Set("name", "NetworkMonitorLoss")
+	alert.Set("value", 5)
+	require.NoError(t, hub.Save(alert))
+	var monitors []*core.Record
+	for _, name := range []string{"gateway", "website"} {
+		record, err := beszelTests.CreateRecord(hub, "network_monitors", map[string]any{
+			"system": system.Id, "target": name + ".example.com", "protocol": "icmp", "interval": 60, "enabled": true,
+		})
+		require.NoError(t, err)
+		monitors = append(monitors, record)
+	}
+	// Avoid starting a system update worker in tests.
+	_, err := hub.DB().Update("systems", dbx.Params{"status": "up"}, dbx.HashExp{"id": system.Id}).Execute()
+	require.NoError(t, err)
+	system.Set("status", "up")
+	return hub, system, alert, monitors
+}
+
+func monitorResult(loss float64) monitor.Result {
+	return monitor.Result{LastProbeAt: time.Now().UnixMilli(), SampleCount: 60, PacketLoss1h: loss}
+}
+
+func TestNetworkMonitorAlertIndependentIncidents(t *testing.T) {
+	hub, system, alert, monitors := networkAlertSetup(t)
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+	count := hub.TestMailer.TotalSend()
+	results := map[string]monitor.Result{monitors[0].Id: monitorResult(10), monitors[1].Id: monitorResult(0)}
+	check := func(active bool, open, sent int) {
+		t.Helper()
+		record, err := hub.FindRecordById("alerts", alert.Id)
+		require.NoError(t, err)
+		assert.Equal(t, active, record.GetBool("triggered"))
+		total, err := hub.CountRecords("alerts_history", dbx.HashExp{"alert_id": alert.Id, "resolved": ""})
+		require.NoError(t, err)
+		assert.EqualValues(t, open, total)
+		assert.Equal(t, count+sent, hub.TestMailer.TotalSend())
+	}
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	check(true, 1, 1)
+	message := hub.TestMailer.Messages()[count]
+	assert.Contains(t, message.Text, "gateway.example.com")
+	assert.Contains(t, message.Text, "10.00%")
+	assert.Contains(t, message.Text, "5.00%")
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	check(true, 1, 1)
+	// Persisted monitor state prevents duplicate notifications after a hub restart.
+	am = alerts.NewTestAlertManagerWithoutWorker(hub)
+	results[monitors[1].Id] = monitorResult(20)
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	check(true, 2, 2)
+	results[monitors[0].Id] = monitorResult(5) // Equality is a recovery.
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	check(true, 1, 3)
+	results[monitors[1].Id] = monitorResult(0)
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	check(false, 0, 4)
+	histories, err := hub.FindAllRecords("alerts_history", dbx.HashExp{"alert_id": alert.Id})
+	require.NoError(t, err)
+	require.Len(t, histories, 2)
+	for _, history := range histories {
+		assert.NotEmpty(t, history.GetString("monitor_name"))
+	}
+}
+
+func TestNetworkMonitorAlertTargetLabel(t *testing.T) {
+	for _, tc := range []struct {
+		protocol, target, label string
+		port                    int
+	}{
+		{"icmp", "gateway.example.com", "gateway.example.com", 0},
+		{"http", "https://example.com/health", "https://example.com/health", 0},
+		{"tcp", "example.com", "example.com:8443", 8443},
+		{"tcp", "2001:db8::1", "[2001:db8::1]:443", 443},
+	} {
+		t.Run(tc.label, func(t *testing.T) {
+			hub, system, alert, monitors := networkAlertSetup(t)
+			m := monitors[0]
+			m.Set("protocol", tc.protocol)
+			m.Set("target", tc.target)
+			m.Set("port", tc.port)
+			require.NoError(t, hub.Save(m))
+			am := alerts.NewTestAlertManagerWithoutWorker(hub)
+			require.NoError(t, am.HandleNetworkMonitorAlerts(system, map[string]monitor.Result{m.Id: monitorResult(10)}))
+			histories, err := hub.FindAllRecords("alerts_history", dbx.HashExp{"alert_id": alert.Id})
+			require.NoError(t, err)
+			require.Len(t, histories, 1)
+			assert.Equal(t, tc.label, histories[0].GetString("monitor_name"))
+			assert.Contains(t, hub.TestMailer.Messages()[hub.TestMailer.TotalSend()-1].Text, tc.label)
+		})
+	}
+}
+
+func TestNetworkMonitorAlertIgnoresUnknownResults(t *testing.T) {
+	for _, scenario := range []string{"missing", "stale", "warmup", "no probes", "down", "paused", "future", "expired hourly window"} {
+		t.Run(scenario, func(t *testing.T) {
+			hub, system, alert, monitors := networkAlertSetup(t)
+			am := alerts.NewTestAlertManagerWithoutWorker(hub)
+			apply := func(loss float64) {
+				result := monitorResult(loss)
+				results := map[string]monitor.Result{monitors[0].Id: result}
+				switch scenario {
+				case "missing":
+					results = nil
+				case "stale":
+					result.LastProbeAt = time.Now().Add(-10 * time.Minute).UnixMilli()
+					results[monitors[0].Id] = result
+				case "expired hourly window":
+					monitors[0].Set("interval", 3600)
+					require.NoError(t, hub.Save(monitors[0]))
+					result.LastProbeAt = time.Now().Add(-2 * time.Hour).UnixMilli()
+					results[monitors[0].Id] = result
+				case "future":
+					result.LastProbeAt = time.Now().Add(time.Hour).UnixMilli()
+					results[monitors[0].Id] = result
+				case "warmup":
+					result.SampleCount = 2
+					results[monitors[0].Id] = result
+				case "no probes":
+					result.SampleCount = 0
+					results[monitors[0].Id] = result
+				case "down":
+					_, err := hub.DB().Update("systems", dbx.Params{"status": scenario}, dbx.HashExp{"id": system.Id}).Execute()
+					require.NoError(t, err)
+				case "paused":
+					record, err := hub.FindRecordById("systems", system.Id)
+					require.NoError(t, err)
+					record.Set("status", "paused")
+					require.NoError(t, hub.Save(record))
+				}
+				require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+			}
+			count := hub.TestMailer.TotalSend()
+			apply(100)
+			assert.Equal(t, count, hub.TestMailer.TotalSend())
+			_, err := hub.DB().Update("systems", dbx.Params{"status": "up"}, dbx.HashExp{"id": system.Id}).Execute()
+			require.NoError(t, err)
+			require.NoError(t, am.HandleNetworkMonitorAlerts(system, map[string]monitor.Result{monitors[0].Id: monitorResult(10)}))
+			apply(0)
+			assert.Equal(t, count+1, hub.TestMailer.TotalSend(), "unknown data must not recover an incident")
+			record, err := hub.FindRecordById("alerts", alert.Id)
+			require.NoError(t, err)
+			assert.True(t, record.GetBool("triggered"))
+		})
+	}
+}
+
+func TestNetworkMonitorAlertCleanup(t *testing.T) {
+	for _, scenario := range []string{"disable monitor", "delete monitor", "disable alert", "purge history", "delete system"} {
+		t.Run(scenario, func(t *testing.T) {
+			hub, system, alert, monitors := networkAlertSetup(t)
+			am := alerts.NewTestAlertManagerWithoutWorker(hub)
+			results := map[string]monitor.Result{monitors[0].Id: monitorResult(10), monitors[1].Id: monitorResult(20)}
+			require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+			count := hub.TestMailer.TotalSend()
+			switch scenario {
+			case "disable monitor":
+				monitors[0].Set("enabled", false)
+				require.NoError(t, hub.Save(monitors[0]))
+			case "delete monitor":
+				require.NoError(t, hub.Delete(monitors[0]))
+			case "disable alert":
+				require.NoError(t, hub.Delete(alert))
+			case "delete system":
+				require.NoError(t, hub.Delete(system))
+			case "purge history":
+				history, err := hub.FindAllRecords("alerts_history")
+				require.NoError(t, err)
+				for _, record := range history {
+					require.NoError(t, hub.Delete(record))
+				}
+			}
+			require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+			assert.Equal(t, count, hub.TestMailer.TotalSend())
+			open, err := hub.CountRecords("alerts_history", dbx.HashExp{"alert_id": alert.Id, "resolved": ""})
+			require.NoError(t, err)
+			if scenario == "disable monitor" || scenario == "delete monitor" {
+				assert.EqualValues(t, 1, open)
+				record, err := hub.FindRecordById("alerts", alert.Id)
+				require.NoError(t, err)
+				assert.True(t, record.GetBool("triggered"))
+				require.NoError(t, hub.Delete(monitors[1]))
+				record, err = hub.FindRecordById("alerts", alert.Id)
+				require.NoError(t, err)
+				assert.False(t, record.GetBool("triggered"))
+			} else {
+				assert.Zero(t, open)
+			}
+		})
+	}
+}
+
+func TestNetworkMonitorAlertPerUserThresholds(t *testing.T) {
+	hub, system, alert, monitors := networkAlertSetup(t)
+	user, err := beszelTests.CreateUser(hub, "monitor2@example.com", "password")
+	require.NoError(t, err)
+	other, err := beszelTests.CreateRecord(hub, "alerts", map[string]any{"name": "NetworkMonitorLoss", "system": system.Id, "user": user.Id, "value": 20})
+	require.NoError(t, err)
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+	results := map[string]monitor.Result{monitors[0].Id: monitorResult(10)}
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	other, err = hub.FindRecordById("alerts", other.Id)
+	require.NoError(t, err)
+	assert.False(t, other.GetBool("triggered"))
+	// Editing the threshold re-evaluates on the next batch, without losing state.
+	alert, err = hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	alert.Set("value", 15)
+	require.NoError(t, hub.Save(alert))
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	alert, err = hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alert.GetBool("triggered"))
+}
+
+func TestNetworkMonitorAlertAPI(t *testing.T) {
+	for _, tc := range []struct {
+		name                  string
+		value                 float64
+		direct, denied, patch bool
+		status                int
+	}{
+		{name: "zero threshold", value: 0, status: 200},
+		{name: "fractional threshold", value: 5.5, status: 200},
+		{name: "negative threshold", value: -1, status: 400},
+		{name: "unreachable threshold", value: 100, status: 400},
+		{name: "bulk inaccessible system", value: 5, denied: true, status: 200},
+		{name: "direct inaccessible system", value: 5, direct: true, denied: true, status: 403},
+		{name: "direct invalid threshold", value: -1, direct: true, status: 400},
+		{name: "direct private state", value: 5, direct: true, status: 200},
+		{name: "patch preserves state", value: 10, direct: true, patch: true, status: 200},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			hub, user := beszelTests.GetHubWithUser(t)
+			defer hub.Cleanup()
+			owner := user.Id
+			if tc.denied {
+				other, err := beszelTests.CreateUser(hub, "other@example.com", "password")
+				require.NoError(t, err)
+				owner = other.Id
+			}
+			systems, err := beszelTests.CreateSystems(hub, 1, owner, "paused")
+			require.NoError(t, err)
+			token, err := user.NewAuthToken()
+			require.NoError(t, err)
+			body := map[string]any{"name": "NetworkMonitorLoss", "value": tc.value, "min": 60, "systems": []string{systems[0].Id}, "overwrite": true}
+			url, method := "/api/beszel/user-alerts", "POST"
+			if tc.direct {
+				url = "/api/collections/alerts/records"
+				body["system"], body["user"] = systems[0].Id, user.Id
+				body["state"], body["triggered"] = map[string]any{"monitors": map[string]string{"fake": "fake"}}, true
+			}
+			if tc.patch {
+				alert, err := beszelTests.CreateRecord(hub, "alerts", map[string]any{"name": "NetworkMonitorLoss", "system": systems[0].Id, "user": user.Id, "value": 5, "triggered": true, "state": map[string]any{"monitors": map[string]string{"real": "history"}}})
+				require.NoError(t, err)
+				url += "/" + alert.Id
+				method = "PATCH"
+				body["triggered"] = false
+			}
+			content := `"success":true`
+			if tc.direct {
+				content = `"name":"NetworkMonitorLoss"`
+			}
+			if tc.status == 400 {
+				content = `"status":400`
+			}
+			if tc.status == 403 {
+				content = `"status":403`
+			}
+			scenario := beszelTests.ApiScenario{
+				Name: tc.name, Method: method, URL: url, Body: jsonReader(body),
+				Headers: map[string]string{"Authorization": token}, ExpectedStatus: tc.status, ExpectedContent: []string{content},
+				TestAppFactory: func(testing.TB) *pbTests.TestApp { return hub.TestApp },
+			}
+			scenario.Test(t)
+			records, err := hub.FindAllRecords("alerts")
+			require.NoError(t, err)
+			if tc.status != 200 || tc.denied {
+				assert.Empty(t, records)
+				return
+			}
+			require.Len(t, records, 1)
+			assert.Equal(t, tc.value, records[0].GetFloat("value"))
+			assert.Zero(t, records[0].GetInt("min"))
+			state := struct {
+				Monitors map[string]string `json:"monitors"`
+			}{}
+			require.NoError(t, records[0].UnmarshalJSONField("state", &state))
+			states := state.Monitors
+			if tc.patch {
+				assert.Equal(t, map[string]string{"real": "history"}, states)
+				assert.True(t, records[0].GetBool("triggered"))
+			} else {
+				assert.Empty(t, states)
+				assert.False(t, records[0].GetBool("triggered"))
+			}
+		})
+	}
+}
+
+type monitorCountingHub struct {
+	*beszelTests.TestHub
+	transactions      atomic.Int64
+	beforeTransaction func()
+}
+
+func (h *monitorCountingHub) RunInTransaction(fn func(core.App) error) error {
+	h.transactions.Add(1)
+	if h.beforeTransaction != nil {
+		h.beforeTransaction()
+	}
+	return h.App.RunInTransaction(fn)
+}
+
+// Count actual SQL on both DB connections, including queries through record APIs.
+func monitorSQLCounter(t *testing.T, app core.App) *atomic.Int64 {
+	t.Helper()
+	count := &atomic.Int64{}
+	for _, builder := range []dbx.Builder{app.ConcurrentDB(), app.NonconcurrentDB()} {
+		db := builder.(*dbx.DB)
+		old := db.LogFunc
+		db.LogFunc = func(string, ...any) { count.Add(1) }
+		t.Cleanup(func() { db.LogFunc = old })
+	}
+	return count
+}
+
+func TestNetworkMonitorAlertSteadyStateNoDatabaseWork(t *testing.T) {
+	hub, system, alert, monitors := networkAlertSetup(t)
+	counted := &monitorCountingHub{TestHub: hub}
+	am := alerts.NewTestAlertManagerWithoutWorker(counted)
+	sql := monitorSQLCounter(t, hub)
+	results := map[string]monitor.Result{monitors[0].Id: monitorResult(0)}
+	evaluate := func() { t.Helper(); require.NoError(t, am.HandleNetworkMonitorAlerts(system, results)) }
+	noWork := func() {
+		t.Helper()
+		sql.Store(0)
+		counted.transactions.Store(0)
+		for range 100 {
+			evaluate()
+		}
+		assert.Zero(t, sql.Load(), "steady state must not issue SQL")
+		assert.Zero(t, counted.transactions.Load(), "steady state must not open transactions")
+	}
+	// One-time lazy loads are permitted, including on hub restart.
+	evaluate()
+	assert.Positive(t, sql.Load())
+	noWork()
+	// Realtime metric saves invoke record hooks but must not invalidate config.
+	fresh, err := hub.FindRecordById("network_monitors", monitors[0].Id)
+	require.NoError(t, err)
+	monitors[0] = fresh
+	monitors[0].Set("loss1h", 0)
+	monitors[0].Set("res", 100)
+	require.NoError(t, hub.Save(monitors[0]))
+	noWork()
+	results[monitors[0].Id] = monitorResult(10)
+	evaluate()
+	assert.Positive(t, sql.Load(), "transitions must still be persisted")
+	assert.EqualValues(t, 1, counted.transactions.Load())
+	noWork()
+	// Missing and stale observations must not enter the transaction either.
+	results = nil
+	noWork()
+	results = map[string]monitor.Result{monitors[0].Id: {SampleCount: 60, LastProbeAt: time.Now().Add(-10 * time.Minute).UnixMilli()}}
+	noWork()
+	results[monitors[0].Id] = monitorResult(0)
+	evaluate()
+	noWork()
+	require.NoError(t, hub.Delete(alert))
+	noWork()
+}
+
+func TestNetworkMonitorAlertConfigCacheInvalidation(t *testing.T) {
+	hub, system, alert, monitors := networkAlertSetup(t)
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+	results := map[string]monitor.Result{monitors[0].Id: monitorResult(0)}
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	count := hub.TestMailer.TotalSend()
+	// Widening the interval makes this observation fresh. A stale interval cache
+	// would miss the failure indefinitely, even though results keep arriving.
+	result := monitorResult(10)
+	result.LastProbeAt = time.Now().Add(-4 * time.Minute).UnixMilli()
+	results[monitors[0].Id] = result
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count, hub.TestMailer.TotalSend())
+	monitors[0].Set("interval", 120)
+	require.NoError(t, hub.Save(monitors[0]))
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count+1, hub.TestMailer.TotalSend())
+	// Disable, then re-enable the same ID: its new failure must be detected.
+	monitors[0].Set("enabled", false)
+	require.NoError(t, hub.Save(monitors[0]))
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	monitors[0].Set("enabled", true)
+	require.NoError(t, hub.Save(monitors[0]))
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count+2, hub.TestMailer.TotalSend())
+	// A new monitor must also become eligible without restarting the hub.
+	created, err := beszelTests.CreateRecord(hub, "network_monitors", map[string]any{
+		"system": system.Id, "name": "new", "target": "new.example.com", "protocol": "icmp", "interval": 60, "enabled": true,
+	})
+	require.NoError(t, err)
+	results[created.Id] = monitorResult(10)
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count+3, hub.TestMailer.TotalSend())
+	// Threshold changes refresh cached config and preserve the active incidents.
+	alert, err = hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	alert.Set("value", 15)
+	require.NoError(t, hub.Save(alert))
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count+5, hub.TestMailer.TotalSend())
+}
+
+func TestNetworkMonitorAlertRevalidatesCandidate(t *testing.T) {
+	for _, change := range []string{"threshold", "disable alert", "disable monitor", "down"} {
+		t.Run(change, func(t *testing.T) {
+			hub, system, alert, monitors := networkAlertSetup(t)
+			counted := &monitorCountingHub{TestHub: hub}
+			am := alerts.NewTestAlertManagerWithoutWorker(counted)
+			results := map[string]monitor.Result{monitors[0].Id: monitorResult(0)}
+			require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+			count := hub.TestMailer.TotalSend()
+			// Change the DB after the cache predicts a transition, before its transaction.
+			counted.beforeTransaction = func() {
+				counted.beforeTransaction = nil
+				switch change {
+				case "threshold":
+					alert.Set("value", 20)
+					require.NoError(t, hub.Save(alert))
+				case "disable alert":
+					require.NoError(t, hub.Delete(alert))
+				case "disable monitor":
+					monitors[0].Set("enabled", false)
+					require.NoError(t, hub.Save(monitors[0]))
+				case "down":
+					_, err := hub.DB().Update("systems", dbx.Params{"status": "down"}, dbx.HashExp{"id": system.Id}).Execute()
+					require.NoError(t, err)
+				}
+			}
+			results[monitors[0].Id] = monitorResult(10)
+			require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+			assert.EqualValues(t, 1, counted.transactions.Load())
+			assert.Equal(t, count, hub.TestMailer.TotalSend())
+			histories, err := hub.CountRecords("alerts_history")
+			require.NoError(t, err)
+			assert.Zero(t, histories)
+		})
+	}
+}
+
+func TestNetworkMonitorAlertConcurrentEvaluations(t *testing.T) {
+	hub, system, _, monitors := networkAlertSetup(t)
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+	results := map[string]monitor.Result{monitors[0].Id: monitorResult(10)}
+	count := hub.TestMailer.TotalSend()
+	var wg sync.WaitGroup
+	errs := make(chan error, 8)
+	for range 8 {
+		wg.Go(func() { errs <- am.HandleNetworkMonitorAlerts(system, results) })
+	}
+	wg.Wait()
+	close(errs)
+	for err := range errs {
+		require.NoError(t, err)
+	}
+	assert.Equal(t, count+1, hub.TestMailer.TotalSend())
+}
+
+func TestNetworkMonitorAlertCacheAfterRollback(t *testing.T) {
+	hub, system, _, monitors := networkAlertSetup(t)
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+	results := map[string]monitor.Result{monitors[0].Id: monitorResult(0)}
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	_, err := hub.DB().NewQuery(`CREATE TRIGGER fail_alert BEFORE UPDATE ON alerts BEGIN SELECT RAISE(ABORT, 'test rollback'); END`).Execute()
+	require.NoError(t, err)
+	count := hub.TestMailer.TotalSend()
+	results[monitors[0].Id] = monitorResult(10)
+	require.Error(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count, hub.TestMailer.TotalSend())
+	histories, err := hub.CountRecords("alerts_history")
+	require.NoError(t, err)
+	assert.Zero(t, histories)
+	_, err = hub.DB().NewQuery("DROP TRIGGER fail_alert").Execute()
+	require.NoError(t, err)
+	// A failed transition must not be published to the cache and mask the retry.
+	require.NoError(t, am.HandleNetworkMonitorAlerts(system, results))
+	assert.Equal(t, count+1, hub.TestMailer.TotalSend())
+}

internal/alerts/alerts_system.go

@@ -47,7 +47,7 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 		return nil
 	}
 
-	alerts := am.alertsCache.GetAlertsExcludingNames(systemRecord.Id, "Status", alertNameSystemdFailed, containerAlertName)
+	alerts := am.alertsCache.GetAlertsExcludingNames(systemRecord.Id, "Status", alertNameSystemdFailed, containerAlertName, alertNameNetworkMonitorLoss)
 	if len(alerts) == 0 {
 		return nil
 	}
@@ -78,7 +78,7 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 				}
 			}
 			for _, pool := range data.Stats.ZfsPools {
-				if pool != nil && pool.Total > 0 {
+				if pool != nil && !pool.Raw && pool.Total > 0 {
 					usedPct := pool.Used / pool.Total * 100
 					if usedPct > maxUsedPct {
 						maxUsedPct = usedPct
@@ -256,7 +256,7 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 				}
 				// add zfs pool usage from historical record
 				for key, pool := range stats.ZfsPools {
-					if pool.Total > 0 {
+					if !pool.Raw && pool.Total > 0 {
 						zfsKey := zfsDiskAlertKey(key)
 						if _, ok := alert.mapSums[zfsKey]; !ok {
 							alert.mapSums[zfsKey] = 0.0
@@ -319,6 +319,11 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 				if sumPct > maxPct {
 					maxPct = sumPct
 					alert.descriptor = diskAlertDescriptor(key)
+					if poolKey, ok := strings.CutPrefix(key, "zfs:"); ok {
+						if pool := data.Stats.ZfsPools[poolKey]; pool != nil && pool.DisplayName != "" {
+							alert.descriptor = diskAlertDescriptor(zfsDiskAlertKey(pool.DisplayName))
+						}
+					}
 				}
 			}
 			alert.val = float64(maxPct / float32(alert.count))
@@ -370,7 +375,7 @@ func zfsDiskAlertKey(poolName string) string {
 
 func diskAlertDescriptor(key string) string {
 	if poolName, ok := strings.CutPrefix(key, "zfs:"); ok {
-		return fmt.Sprintf("Usage of ZFS pool %s", poolName)
+		return fmt.Sprintf("Usage of storage pool %s", poolName)
 	}
 	return fmt.Sprintf("Usage of %s", key)
 }

internal/alerts/alerts_test_helpers.go

@@ -11,8 +11,9 @@ import (
 
 func NewTestAlertManagerWithoutWorker(app hubLike) *AlertManager {
 	return &AlertManager{
-		hub:         app,
-		alertsCache: NewAlertsCache(app),
+		hub:             app,
+		alertsCache:     NewAlertsCache(app),
+		networkMonitors: newNetworkMonitorCache(app),
 	}
 }
 
@@ -100,10 +101,6 @@ func (am *AlertManager) SetAlertTriggered(alert CachedAlertData, triggered bool)
 	return am.setAlertTriggered(alert, triggered)
 }
 
-func IsInternalURL(rawURL string) (bool, error) {
-	return isInternalURL(rawURL)
-}
-
 // BuildContainerLogExcerpt exposes buildContainerLogExcerpt for testing.
 func BuildContainerLogExcerpt(raw string) string {
 	return buildContainerLogExcerpt(raw)

internal/alerts/alerts_webhooks_test.go

@@ -0,0 +1,66 @@
+//go:build testing
+
+package alerts_test
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+
+	"github.com/henrygd/beszel/internal/alerts"
+	beszelTests "github.com/henrygd/beszel/internal/tests"
+	"github.com/pocketbase/dbx"
+	"github.com/stretchr/testify/require"
+)
+
+func TestPersistedWebhooksUseCurrentOwnerRole(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	var delivered atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		delivered.Add(1)
+	}))
+	defer server.Close()
+
+	settings, err := hub.FindFirstRecordByFilter("user_settings", "user={:user}", dbx.Params{"user": user.Id})
+	require.NoError(t, err)
+	settings.Set("settings", alerts.UserNotificationSettings{Webhooks: []string{"generic+" + server.URL}})
+	require.NoError(t, hub.Save(settings))
+	message := alerts.AlertMessageData{UserID: user.Id, Title: "Test", Message: "Persisted webhook"}
+
+	// Keep the same URL and manager while changing roles, so cached privileges
+	// or treating previously saved URLs as trusted would fail this test.
+	for _, tc := range []struct {
+		name string
+		role string
+		want int32
+	}{
+		{"regular user", "user", 0},
+		{"readonly user", "readonly", 0},
+		{"promoted admin", "admin", 1},
+		{"demoted admin", "user", 1},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			user.Set("role", tc.role)
+			require.NoError(t, hub.Save(user))
+			// Webhook errors are logged; SendAlert continues to email delivery.
+			require.NoError(t, am.SendAlert(message))
+			require.Equal(t, tc.want, delivered.Load())
+		})
+	}
+
+	t.Run("missing owner fails closed", func(t *testing.T) {
+		// Model an orphaned settings record without deleting it through the
+		// normal user deletion cascade.
+		const missingOwner = "missingowner123"
+		settings.Set("user", missingOwner)
+		require.NoError(t, hub.SaveNoValidate(settings))
+		message.UserID = missingOwner
+		err := am.SendAlert(message)
+		require.ErrorContains(t, err, "load notification owner")
+		require.EqualValues(t, 1, delivered.Load())
+	})
+}

internal/alerts/alerts_zfs.go

@@ -46,12 +46,15 @@ func (am *AlertManager) handleZfsPoolHealthAlert(e *core.RecordEvent, oldHealth
 	}
 
 	systemName := systemRecord.GetString("name")
-	poolName := e.Record.GetString("name")
+	poolName := e.Record.GetString("display_name")
+	if poolName == "" {
+		poolName = e.Record.GetString("name")
+	}
 
-	title := fmt.Sprintf("ZFS pool %s on %s: %s", newHealth, systemName, poolName)
-	message := fmt.Sprintf("ZFS pool %s (%s) was first observed as %s", poolName, systemName, newHealth)
+	title := fmt.Sprintf("Storage pool %s on %s: %s", newHealth, systemName, poolName)
+	message := fmt.Sprintf("Storage pool %s (%s) was first observed as %s", poolName, systemName, newHealth)
 	if oldSeverity > 0 {
-		message = fmt.Sprintf("ZFS pool %s (%s) health changed from %s to %s", poolName, systemName, oldHealth, newHealth)
+		message = fmt.Sprintf("Storage pool %s (%s) health changed from %s to %s", poolName, systemName, oldHealth, newHealth)
 	}
 
 	userIDs := systemRecord.GetStringSlice("users")
@@ -116,7 +119,7 @@ func createZfsPoolHistoryRecord(app core.App, userID, systemID, alertID, poolNam
 	record.Set("user", userID)
 	record.Set("system", systemID)
 	record.Set("alert_id", alertID)
-	record.Set("name", "ZFS Pool: "+poolName)
+	record.Set("name", "Storage Pool: "+poolName)
 	return app.Save(record)
 }
 

internal/alerts/alerts_zfs_disk_test.go

@@ -143,3 +143,37 @@ func TestDiskAlertZfsPoolMultiMinute(t *testing.T) {
 	assert.False(t, diskAlert.GetBool("triggered"),
 		"Alert should be resolved when ZFS pool average (50%%) drops below threshold (80%%)")
 }
+
+func TestDiskAlertIgnoresRawPool(t *testing.T) {
+	for _, minutes := range []int{0, 2} {
+		hub, user := beszelTests.GetHubWithUser(t)
+		systems, err := beszelTests.CreateSystems(hub, 1, user.Id, "up")
+		require.NoError(t, err)
+		alert, err := beszelTests.CreateRecord(hub, "alerts", map[string]any{"name": "Disk", "system": systems[0].Id, "user": user.Id, "value": 80, "min": minutes})
+		require.NoError(t, err)
+		pools := map[string]*system.ZfsPool{"btrfs": {Total: 100, Used: 99, Raw: true}}
+		for _, offset := range []time.Duration{-180, -90, -60, -30} {
+			data, err := json.Marshal(system.Stats{ZfsPools: pools})
+			require.NoError(t, err)
+			record, err := beszelTests.CreateRecord(hub, "system_stats", map[string]any{"system": systems[0].Id, "type": "1m", "stats": string(data)})
+			require.NoError(t, err)
+			record.SetRaw("created", time.Now().UTC().Add(offset*time.Second).Format(types.DefaultDateLayout))
+			require.NoError(t, hub.SaveNoValidate(record))
+		}
+		require.NoError(t, hub.GetAlertManager().HandleSystemAlerts(systems[0], &system.CombinedData{Stats: system.Stats{ZfsPools: pools}}))
+		time.Sleep(20 * time.Millisecond)
+		record, err := hub.FindRecordById("alerts", alert.Id)
+		require.NoError(t, err)
+		assert.False(t, record.GetBool("triggered"))
+		if minutes > 0 {
+			// A current usable sample must not make raw historical values eligible.
+			pools["btrfs"].Raw = false
+			require.NoError(t, hub.GetAlertManager().HandleSystemAlerts(systems[0], &system.CombinedData{Stats: system.Stats{ZfsPools: pools}}))
+			time.Sleep(20 * time.Millisecond)
+			record, err = hub.FindRecordById("alerts", alert.Id)
+			require.NoError(t, err)
+			assert.False(t, record.GetBool("triggered"))
+		}
+		hub.Cleanup()
+	}
+}

internal/alerts/alerts_zfs_key_test.go

@@ -10,6 +10,6 @@ import (
 
 func TestZfsDiskAlertKeyIsNamespaced(t *testing.T) {
 	assert.Equal(t, "zfs:tank", zfsDiskAlertKey("tank"))
-	assert.Equal(t, "Usage of ZFS pool tank", diskAlertDescriptor(zfsDiskAlertKey("tank")))
+	assert.Equal(t, "Usage of storage pool tank", diskAlertDescriptor(zfsDiskAlertKey("tank")))
 	assert.Equal(t, "Usage of tank", diskAlertDescriptor("tank"))
 }

internal/alerts/alerts_zfs_test.go

@@ -42,7 +42,7 @@ func TestZfsPoolAlertOnlineToDegraded(t *testing.T) {
 
 	assert.EqualValues(t, 1, hub.TestMailer.TotalSend(), "should have 1 email sent after pool became DEGRADED")
 	lastMessage := hub.TestMailer.LastMessage()
-	assert.Contains(t, lastMessage.Subject, "ZFS pool DEGRADED on test-system")
+	assert.Contains(t, lastMessage.Subject, "Storage pool DEGRADED on test-system")
 	assert.Contains(t, lastMessage.Subject, "tank")
 	assert.Contains(t, lastMessage.Text, "ONLINE to DEGRADED")
 }
@@ -76,7 +76,7 @@ func TestZfsPoolAlertDegradedToFaulted(t *testing.T) {
 
 	assert.EqualValues(t, 2, hub.TestMailer.TotalSend(), "should alert on initial DEGRADED state and later FAULTED transition")
 	lastMessage := hub.TestMailer.LastMessage()
-	assert.Contains(t, lastMessage.Subject, "ZFS pool FAULTED on test-system")
+	assert.Contains(t, lastMessage.Subject, "Storage pool FAULTED on test-system")
 }
 
 func TestZfsPoolAlertNoAlertOnRecovery(t *testing.T) {
@@ -239,7 +239,7 @@ func TestZfsPoolAlertWritesHistory(t *testing.T) {
 	history, err := hub.FindRecordsByFilter("alerts_history", "alert_id={:alert_id}", "", 0, 0, map[string]any{"alert_id": pool.Id})
 	assert.NoError(t, err)
 	require.Len(t, history, 1, "expected one history entry per user")
-	assert.Equal(t, "ZFS Pool: tank", history[0].GetString("name"))
+	assert.Equal(t, "Storage Pool: tank", history[0].GetString("name"))
 	assert.Equal(t, system.Id, history[0].GetString("system"))
 }
 

internal/alerts/network_monitor_cache.go

@@ -0,0 +1,79 @@
+package alerts
+
+import (
+	"sync"
+
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/core"
+)
+
+// networkMonitorCache keeps just the enabled monitor IDs and probe intervals
+// needed for the alert fast path. Names and targets are read only on transitions.
+// Returned maps are immutable; configuration changes invalidate the whole entry.
+type networkMonitorCache struct {
+	app     core.App
+	mu      sync.RWMutex
+	systems map[string]map[string]int
+}
+
+func newNetworkMonitorCache(app core.App) *networkMonitorCache {
+	c := &networkMonitorCache{app: app, systems: make(map[string]map[string]int)}
+	invalidate := func(e *core.RecordEvent) error {
+		c.invalidate(e.Record.GetString("system"))
+		return e.Next()
+	}
+	app.OnRecordAfterCreateSuccess("network_monitors").BindFunc(invalidate)
+	app.OnRecordAfterDeleteSuccess("network_monitors").BindFunc(invalidate)
+	app.OnRecordAfterUpdateSuccess("network_monitors").BindFunc(func(e *core.RecordEvent) error {
+		old := e.Record.Original()
+		// Realtime metric saves also invoke this hook. They must not evict config.
+		if old.GetString("system") != e.Record.GetString("system") ||
+			old.GetBool("enabled") != e.Record.GetBool("enabled") ||
+			old.GetInt("interval") != e.Record.GetInt("interval") {
+			c.invalidate(old.GetString("system"))
+			c.invalidate(e.Record.GetString("system"))
+		}
+		return e.Next()
+	})
+	app.OnRecordAfterDeleteSuccess("systems").BindFunc(func(e *core.RecordEvent) error {
+		c.invalidate(e.Record.Id)
+		return e.Next()
+	})
+	return c
+}
+
+func (c *networkMonitorCache) invalidate(systemID string) {
+	c.mu.Lock()
+	delete(c.systems, systemID)
+	c.mu.Unlock()
+}
+
+func (c *networkMonitorCache) get(systemID string) (map[string]int, error) {
+	c.mu.RLock()
+	monitors, ok := c.systems[systemID]
+	c.mu.RUnlock()
+	if ok {
+		return monitors, nil
+	}
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if monitors, ok := c.systems[systemID]; ok {
+		return monitors, nil
+	}
+	// Keep the lock through the load so a concurrent config change cannot be
+	// invalidated first and then overwritten by the older query result.
+	var rows []struct {
+		ID       string `db:"id"`
+		Interval int    `db:"interval"`
+	}
+	if err := c.app.DB().Select("id", "interval").From("network_monitors").
+		Where(dbx.HashExp{"system": systemID, "enabled": true}).All(&rows); err != nil {
+		return nil, err
+	}
+	monitors = make(map[string]int, len(rows))
+	for _, row := range rows {
+		monitors[row.ID] = row.Interval
+	}
+	c.systems[systemID] = monitors
+	return monitors, nil
+}

internal/alerts/notification_client.go

@@ -0,0 +1,150 @@
+package alerts
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/netip"
+	"sync/atomic"
+	"syscall"
+	"time"
+
+	"github.com/nicholas-fedor/shoutrrr/pkg/router"
+	"github.com/nicholas-fedor/shoutrrr/pkg/types"
+)
+
+var (
+	errInternalDestination   = errors.New("Only admins can send to internal destinations")
+	errUnrestrictedService   = errors.New("Only admins can use this notification service") // Restrict services w/o custom connection support
+	publicNotificationDialer = &net.Dialer{
+		Timeout: 10 * time.Second,
+		// Control checks each resolved address immediately before connecting.
+		Control: func(_, address string, _ syscall.RawConn) error { return checkNotificationAddress(address) },
+	}
+	publicNotificationClient = newPublicNotificationClient()
+)
+
+func newPublicNotificationClient() *http.Client {
+	return &http.Client{
+		Timeout: 15 * time.Second,
+		Transport: &http.Transport{
+			// Do not use proxies: they can resolve the target themselves and
+			// bypass the destination check on our socket.
+			DialContext:         publicNotificationDialer.DialContext,
+			TLSHandshakeTimeout: 10 * time.Second,
+			IdleConnTimeout:     90 * time.Second,
+		},
+	}
+}
+
+func checkNotificationAddress(address string) error {
+	addr, err := netip.ParseAddrPort(address)
+	if err != nil || addr.Addr().Zone() != "" {
+		return errInternalDestination
+	}
+	ip := net.IP(addr.Addr().AsSlice())
+	if !ip.IsGlobalUnicast() || isInternalIP(ip) {
+		return errInternalDestination
+	}
+	return nil
+}
+
+func sendPublicNotification(rawURL, message string) error {
+	client := &notificationClient{Client: publicNotificationClient}
+	service, err := newPublicNotificationService(rawURL, types.SenderOptions{HTTPClient: client, DialContext: client.dialContext})
+	if err == nil {
+		if closer, ok := service.(io.Closer); ok {
+			defer closer.Close()
+		}
+		err = service.Send(message, &types.Params{})
+	}
+	// Some services format errors without preserving their error chain.
+	if client.blocked.Load() {
+		return errInternalDestination
+	}
+	return err
+}
+
+type notificationClient struct {
+	*http.Client
+	blocked atomic.Bool
+}
+
+func (c *notificationClient) Do(req *http.Request) (*http.Response, error) {
+	response, err := c.Client.Do(req)
+	if errors.Is(err, errInternalDestination) {
+		c.blocked.Store(true)
+	}
+	return response, err
+}
+
+func (c *notificationClient) dialContext(ctx context.Context, network, address string) (net.Conn, error) {
+	conn, err := publicNotificationDialer.DialContext(ctx, network, address)
+	if errors.Is(err, errInternalDestination) {
+		c.blocked.Store(true)
+	}
+	return conn, err
+}
+
+func newPublicNotificationService(rawURL string, opts types.SenderOptions) (types.Service, error) {
+	r := &router.ServiceRouter{}
+	scheme, serviceURL, err := r.ExtractServiceName(rawURL)
+	if err != nil {
+		return nil, err
+	}
+	service, err := r.NewService(scheme)
+	if err != nil {
+		return nil, err
+	}
+	httpSetter, httpOK := service.(types.HTTPClientSetter)
+	dialSetter, dialOK := service.(types.DialContextSetter)
+	if (!httpOK || opts.HTTPClient == nil) && (!dialOK || opts.DialContext == nil) {
+		return nil, errUnrestrictedService
+	}
+	if serviceURL.Scheme != scheme {
+		custom, ok := service.(types.CustomURLService)
+		if !ok {
+			return nil, fmt.Errorf("%w: %s", router.ErrCustomURLsNotSupported, scheme)
+		}
+		serviceURL, err = custom.GetServiceURLFromCustom(serviceURL)
+		if err != nil {
+			return nil, err
+		}
+	}
+	// Shoutrrr v0.20.0 CreateSenderWithOptions injects only AFTER Initialize.
+	// Matrix can log in during Initialize, so inject before it as well.
+	if httpOK {
+		httpSetter.SetHTTPClient(opts.HTTPClient)
+	}
+	if dialOK {
+		dialSetter.SetDialContext(opts.DialContext)
+	}
+	if err := service.Initialize(serviceURL, nil); err != nil {
+		return nil, err
+	}
+	// Some initializers replace their HTTP client with a default client.
+	if httpOK {
+		httpSetter.SetHTTPClient(opts.HTTPClient)
+	}
+	if dialOK {
+		dialSetter.SetDialContext(opts.DialContext)
+	}
+	return service, nil
+}
+
+var cgnatNetwork = &net.IPNet{
+	IP:   net.IPv4(100, 64, 0, 0),
+	Mask: net.CIDRMask(10, 32),
+}
+
+func isInternalIP(ip net.IP) bool {
+	return ip.IsPrivate() ||
+		ip.IsLoopback() ||
+		ip.IsUnspecified() ||
+		ip.IsLinkLocalUnicast() ||
+		ip.IsMulticast() ||
+		cgnatNetwork.Contains(ip)
+}

internal/alerts/notification_client_test.go

@@ -0,0 +1,217 @@
+package alerts
+
+import (
+	"context"
+	"encoding/binary"
+	"errors"
+	"io"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"github.com/nicholas-fedor/shoutrrr/pkg/types"
+	"golang.org/x/net/dns/dnsmessage"
+)
+
+func TestCheckNotificationAddress(t *testing.T) {
+	for _, host := range []string{"127.0.0.1", "10.0.0.1", "172.16.0.1", "192.168.0.1", "169.254.169.254", "100.64.0.0", "100.127.255.255", "0.0.0.0", "224.0.0.1", "255.255.255.255", "::1", "::", "fc00::1", "fe80::1", "ff02::1", "::ffff:127.0.0.1", "::ffff:169.254.169.254", "fe80::1%lo", "localhost", "consul"} {
+		t.Run(host, func(t *testing.T) {
+			if err := checkNotificationAddress(net.JoinHostPort(host, "80")); !errors.Is(err, errInternalDestination) {
+				t.Fatalf("expected blocked address, got %v", err)
+			}
+		})
+	}
+	for _, host := range []string{"8.8.8.8", "100.63.255.255", "100.128.0.0", "2001:4860:4860::8888"} {
+		if err := checkNotificationAddress(net.JoinHostPort(host, "443")); err != nil {
+			t.Errorf("public address %s: %v", host, err)
+		}
+	}
+}
+
+func TestPublicNotificationBlocksInternalRequests(t *testing.T) {
+	var hits atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		hits.Add(1)
+	}))
+	defer server.Close()
+	host := strings.TrimPrefix(server.URL, "http://")
+	for _, rawURL := range []string{
+		"generic+http://" + host,
+		"generic+https://" + host,
+		"generic+http://localhost:" + strings.Split(host, ":")[1],
+		"matrix://user:password@" + host + "/room?disabletls=yes",
+		"mattermost://" + host + "/token?disabletls=yes",
+	} {
+		t.Run(rawURL, func(t *testing.T) {
+			if err := sendPublicNotification(rawURL, "test"); !errors.Is(err, errInternalDestination) {
+				t.Fatalf("expected internal destination error, got %v", err)
+			}
+		})
+	}
+	if hits.Load() != 0 {
+		t.Fatal("internal server received a request")
+	}
+
+}
+
+type notificationRoundTripper func(*http.Request) (*http.Response, error)
+
+func (f notificationRoundTripper) RoundTrip(r *http.Request) (*http.Response, error) { return f(r) }
+
+func TestPublicNotificationRedirect(t *testing.T) {
+	client := newPublicNotificationClient()
+	defer client.CloseIdleConnections()
+	transport := client.Transport
+	client.Transport = notificationRoundTripper(func(r *http.Request) (*http.Response, error) {
+		if r.URL.Host == "public.example" {
+			return &http.Response{StatusCode: 307, Header: http.Header{"Location": {"http://127.0.0.1/"}}, Body: io.NopCloser(strings.NewReader("")), Request: r}, nil
+		}
+		return transport.RoundTrip(r)
+	})
+	_, err := client.Get("http://public.example/")
+	if !errors.Is(err, errInternalDestination) {
+		t.Fatalf("expected redirect to be blocked, got %v", err)
+	}
+}
+
+func TestPublicNotificationServiceClient(t *testing.T) {
+	for _, rawURL := range []string{"generic+http://public.example/path", "discord://token@123456789", "slack://hook:AAAAAAAAA-BBBBBBBBB-123456789123456789123456@webhook"} {
+		t.Run(rawURL, func(t *testing.T) {
+			var hits int
+			client := &http.Client{Transport: notificationRoundTripper(func(r *http.Request) (*http.Response, error) {
+				hits++
+				body := `{"ok":true}`
+				if strings.HasPrefix(rawURL, "slack:") {
+					body = "ok"
+				}
+				return &http.Response{StatusCode: 200, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(body)), Request: r}, nil
+			})}
+			service, err := newPublicNotificationService(rawURL, types.SenderOptions{HTTPClient: client})
+			if err != nil {
+				t.Fatal(err)
+			}
+			if err := service.Send("test", &types.Params{}); err != nil {
+				t.Fatal(err)
+			}
+			if hits == 0 {
+				t.Fatal("injected client was not used")
+			}
+		})
+	}
+}
+
+func TestPublicNotificationDNS(t *testing.T) {
+	// Supply deterministic DNS responses over an in-memory TCP connection.
+	// The first lookup sees a public IP; subsequent lookups see loopback.
+	var rebound atomic.Bool
+	resolver := &net.Resolver{PreferGo: true, Dial: func(ctx context.Context, network, address string) (net.Conn, error) {
+		client, server := net.Pipe()
+		go func() {
+			defer server.Close()
+			var size [2]byte
+			if _, err := io.ReadFull(server, size[:]); err != nil {
+				return
+			}
+			buf := make([]byte, binary.BigEndian.Uint16(size[:]))
+			if _, err := io.ReadFull(server, buf); err != nil {
+				return
+			}
+			var msg dnsmessage.Message
+			if err := msg.Unpack(buf); err != nil {
+				return
+			}
+			msg.Header.Response = true
+			msg.Header.RecursionAvailable = true
+			q := msg.Questions[0]
+			if q.Type == dnsmessage.TypeA {
+				ip := [4]byte{8, 8, 8, 8}
+				if rebound.Load() {
+					ip = [4]byte{127, 0, 0, 1}
+				}
+				msg.Answers = []dnsmessage.Resource{{Header: dnsmessage.ResourceHeader{Name: q.Name, Type: q.Type, Class: dnsmessage.ClassINET}, Body: &dnsmessage.AResource{A: ip}}}
+			}
+			buf, err := msg.Pack()
+			if err != nil {
+				return
+			}
+			binary.BigEndian.PutUint16(size[:], uint16(len(buf)))
+			server.Write(append(size[:], buf...))
+		}()
+		return client, nil
+	}}
+	// These tests do not run in parallel; restore the process resolver afterward.
+	previous := net.DefaultResolver
+	net.DefaultResolver = resolver
+	t.Cleanup(func() { net.DefaultResolver = previous })
+	ips, err := resolver.LookupIP(context.Background(), "ip4", "rebind.example")
+	if err != nil || len(ips) != 1 || !ips[0].Equal(net.IPv4(8, 8, 8, 8)) {
+		t.Fatalf("initial DNS lookup: %v, %v", ips, err)
+	}
+	rebound.Store(true)
+	client := newPublicNotificationClient()
+	defer client.CloseIdleConnections()
+	for _, host := range []string{"rebind.example", "consul"} {
+		guarded := &notificationClient{Client: client}
+		conn, dialErr := guarded.dialContext(context.Background(), "tcp", net.JoinHostPort(host, "25"))
+		if conn != nil {
+			conn.Close()
+		}
+		if !errors.Is(dialErr, errInternalDestination) || !guarded.blocked.Load() {
+			t.Errorf("expected TCP dial-time rejection for %s, got %v", host, dialErr)
+		}
+		_, err := client.Get("http://" + host + "/")
+		if !errors.Is(err, errInternalDestination) {
+			t.Errorf("expected dial-time rejection for %s, got %v", host, err)
+		}
+	}
+}
+
+func TestPublicNotificationTCP(t *testing.T) {
+	for _, rawURL := range []string{
+		"smtp://user:pass@HOST:25/?fromAddress=sender@example.com&toAddresses=recipient@example.com",
+		"smtp://user:pass@HOST:465/?fromAddress=sender@example.com&toAddresses=recipient@example.com",
+		"mqtt://HOST:1883/topic",
+		"mqtts://HOST:8883/topic",
+	} {
+		t.Run(rawURL, func(t *testing.T) {
+			t.Parallel()
+			t.Run("internal destination", func(t *testing.T) {
+				err := sendPublicNotification(strings.ReplaceAll(rawURL, "HOST", "127.0.0.1"), "test")
+				if !errors.Is(err, errInternalDestination) {
+					t.Fatalf("expected blocked destination, got %v", err)
+				}
+			})
+			t.Run("public destination uses injected dialer", func(t *testing.T) {
+				var calls atomic.Int32
+				stopped := errors.New("test dial stopped")
+				service, err := newPublicNotificationService(strings.ReplaceAll(rawURL, "HOST", "8.8.8.8"), types.SenderOptions{
+					DialContext: func(ctx context.Context, network, address string) (net.Conn, error) {
+						calls.Add(1)
+						if network != "tcp" || !strings.HasPrefix(address, "8.8.8.8:") {
+							t.Errorf("unexpected dial: %s %s", network, address)
+						}
+						if err := checkNotificationAddress(address); err != nil {
+							t.Error(err)
+						}
+						return nil, stopped
+					},
+				})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if closer, ok := service.(io.Closer); ok {
+					defer closer.Close()
+				}
+				if err := service.Send("test", &types.Params{}); err == nil {
+					t.Fatal("expected dial failure")
+				}
+				if calls.Load() == 0 {
+					t.Fatal("custom dialer was not used")
+				}
+			})
+		})
+	}
+}

internal/common/common-ws.go

@@ -24,6 +24,8 @@ const (
 	GetSystemdInfo
 	// Request ZFS detail data from agent
 	GetZfsData
+	// Sync network monitor configuration to agent
+	SyncNetworkMonitors
 	// Add new actions here...
 )
 

internal/dockerfile_agent_nvidia

@@ -43,6 +43,11 @@ COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
 # Copy smartmontools binaries and config files
 COPY --from=smartmontools-builder /usr/sbin/smartctl /usr/sbin/smartctl
 
+# Install ZFS userspace utilities (zpool, zfs) for pool/dataset monitoring
+RUN apt-get update && apt-get install -y --no-install-recommends \
+  zfsutils-linux \
+  && apt-get clean && rm -rf /var/lib/apt/lists/*
+
 # Ensure data persistence across container recreations
 VOLUME ["/var/lib/beszel-agent"]
 

internal/dockerfile_agent_nvidia_slim

@@ -65,6 +65,32 @@ RUN set -eux; \
     cp -v "$interp" "/out/rootfs$interp"; \
   fi
 
+# --------------------------
+# ZFS utilities builder stage
+# --------------------------
+FROM --platform=$TARGETPLATFORM debian:bookworm-slim AS zfsutils-builder
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+  zfsutils-linux \
+  && rm -rf /var/lib/apt/lists/*
+
+# Copy the zpool/zfs binaries and their required runtime libraries
+RUN set -eux; \
+  mkdir -p /out/rootfs/lib /out/rootfs/lib64 /out/rootfs/usr/lib; \
+  for bin in /usr/sbin/zpool /usr/sbin/zfs; do \
+    mkdir -p "/out/rootfs$(dirname "$bin")"; \
+    cp -v "$bin" "/out/rootfs$bin"; \
+    ldd "$bin" \
+      | awk '{print $3}' \
+      | grep '^/' \
+      | xargs -r -I '{}' sh -c 'mkdir -p "/out/rootfs$(dirname "{}")"; cp -v "{}" "/out/rootfs{}"'; \
+    interp="$(ldd "$bin" | awk "/ld-linux/ {print \$1}")"; \
+    if [ -n "$interp" ] && [ -e "$interp" ]; then \
+      mkdir -p "/out/rootfs$(dirname "$interp")"; \
+      cp -v "$interp" "/out/rootfs$interp"; \
+    fi; \
+  done
+
 # --------------------------
 # Final image: lightweight multi-arch NVIDIA agent (slim)
 # --------------------------
@@ -78,6 +104,9 @@ COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
 COPY --from=smartmontools-builder /usr/sbin/smartctl /usr/sbin/smartctl
 COPY --from=smartmontools-builder /out/rootfs/ /
 
+# Copy ZFS utilities (zpool, zfs) binaries and required runtime libraries
+COPY --from=zfsutils-builder /out/rootfs/ /
+
 # nvidia-smi is intentionally not bundled.
 # Mount the host binary instead, for example:
 # - /usr/bin/nvidia-smi:/usr/bin/nvidia-smi:ro

internal/entities/container/container.go

@@ -186,11 +186,12 @@ type Stats struct {
 	NetworkRecv float64   `json:"nr,omitzero" cbor:"4,keyasint,omitzero"` // deprecated 0.18.3 (MB) - keep field for old agents/records
 	Bandwidth   [2]uint64 `json:"b,omitzero" cbor:"9,keyasint,omitzero"`  // [sent bytes, recv bytes]
 
-	Health DockerHealth `json:"-" cbor:"5,keyasint"`
-	Status string       `json:"-" cbor:"6,keyasint"`
-	Id     string       `json:"-" cbor:"7,keyasint"`
-	Image  string       `json:"-" cbor:"8,keyasint"`
-	Ports  string       `json:"-" cbor:"10,keyasint"`
+	Health          DockerHealth `json:"-" cbor:"5,keyasint"`
+	Status          string       `json:"-" cbor:"6,keyasint"`
+	Id              string       `json:"-" cbor:"7,keyasint"`
+	Image           string       `json:"-" cbor:"8,keyasint"`
+	Ports           string       `json:"-" cbor:"10,keyasint"`
+	UpdateAvailable bool         `json:"u,omitzero" cbor:"11,keyasint,omitzero"`
 	// PrevCpu     [2]uint64    `json:"-"`
 	CpuSystem    uint64       `json:"-"`
 	CpuContainer uint64       `json:"-"`

internal/entities/monitor/monitor.go

@@ -0,0 +1,102 @@
+package monitor
+
+import "time"
+
+// MaxProbeTimeout is the longest agent probe timeout (currently HTTP).
+// Hub requests that run a probe must allow this time in addition to transport overhead.
+const MaxProbeTimeout = 10 * time.Second
+
+type SyncAction uint8
+
+const (
+	// SyncActionReplace indicates a full sync where the provided configs should replace all existing monitors for the system.
+	SyncActionReplace SyncAction = iota
+	// SyncActionUpsert indicates an incremental sync where the provided config should be added or updated.
+	SyncActionUpsert
+	// SyncActionDelete indicates an incremental sync where the provided config should be removed.
+	SyncActionDelete
+)
+
+// Config defines a network monitor task sent from hub to agent.
+type Config struct {
+	// ID is the stable network_monitors record ID generated by the hub.
+	ID       string `cbor:"0,keyasint"`
+	Target   string `cbor:"1,keyasint"`
+	Protocol string `cbor:"2,keyasint"` // "icmp", "tcp", "http", or "dns"
+	Port     uint16 `cbor:"3,keyasint,omitempty"`
+	Interval uint16 `cbor:"4,keyasint"` // seconds
+}
+
+// SyncRequest defines an incremental or full monitor sync request sent to the agent.
+type SyncRequest struct {
+	Action  SyncAction `cbor:"0,keyasint"`
+	Config  Config     `cbor:"1,keyasint,omitempty"`
+	Configs []Config   `cbor:"2,keyasint,omitempty"`
+	RunNow  bool       `cbor:"3,keyasint,omitempty"`
+}
+
+// SyncResponse returns the immediate result for an upsert when requested.
+type SyncResponse struct {
+	Result Result `cbor:"0,keyasint,omitempty"`
+}
+
+// Result holds aggregated monitor results for a single target.
+//
+// 0: avg response in microseconds
+//
+// 1: 1h average response in microseconds
+//
+// 2: min response in microseconds
+//
+// 3: 1h min response in microseconds
+//
+// 4: max response in microseconds
+//
+// 5: 1h max response in microseconds
+//
+// 6: packet loss percentage (0-100)
+//
+// 7: 1h packet loss percentage (0-100)
+type Result struct {
+	AvgResponse   int64   `cbor:"0,keyasint,omitempty"`
+	AvgResponse1h int64   `cbor:"1,keyasint,omitempty"`
+	MinResponse   int64   `cbor:"2,keyasint,omitempty"`
+	MinResponse1h int64   `cbor:"3,keyasint,omitempty"`
+	MaxResponse   int64   `cbor:"4,keyasint,omitempty"`
+	MaxResponse1h int64   `cbor:"5,keyasint,omitempty"`
+	PacketLoss    float64 `cbor:"6,keyasint,omitempty"`
+	PacketLoss1h  float64 `cbor:"7,keyasint,omitempty"`
+	// LastProbeAt is the latest completed probe's Unix timestamp in milliseconds.
+	LastProbeAt int64 `cbor:"8,keyasint"`
+	// SampleCount includes all completed probes since this monitor started.
+	// Used for alert warm-up even when the interval is longer than 20 minutes.
+	SampleCount int64 `cbor:"9,keyasint,omitempty"`
+	// Counts and sum cover the current response window (or latest-sample
+	// fallback), not the hourly window or lifetime SampleCount.
+	TotalCount   int64 `cbor:"10,keyasint"`
+	SuccessCount int64 `cbor:"11,keyasint"`
+	ResponseSum  int64 `cbor:"12,keyasint"`
+}
+
+// Stats holds response times in microseconds and packet loss percentage (0-100).
+type Stats struct {
+	ResAvg       float64 `json:"res_avg" db:"-"` // Derived for display; not stored.
+	ResMin       float64 `json:"res_min" db:"res_min"`
+	ResMax       float64 `json:"res_max" db:"res_max"`
+	Loss         float64 `json:"loss" db:"-"` // Derived for display; not stored.
+	TotalCount   int64   `json:"-" db:"total_count"`
+	SuccessCount int64   `json:"-" db:"success_count"`
+	ResponseSum  int64   `json:"-" db:"res_sum"`
+}
+
+func (s Stats) FromResult(result Result) Stats {
+	return Stats{
+		ResAvg:       float64(result.AvgResponse),
+		ResMin:       float64(result.MinResponse),
+		ResMax:       float64(result.MaxResponse),
+		Loss:         result.PacketLoss,
+		TotalCount:   result.TotalCount,
+		SuccessCount: result.SuccessCount,
+		ResponseSum:  result.ResponseSum,
+	}
+}

internal/entities/system/system.go

@@ -7,6 +7,7 @@ import (
 	"time"
 
 	"github.com/henrygd/beszel/internal/entities/container"
+	"github.com/henrygd/beszel/internal/entities/monitor"
 	"github.com/henrygd/beszel/internal/entities/systemd"
 )
 
@@ -59,11 +60,15 @@ type Stats struct {
 
 // ZfsPool holds per-pool ZFS metrics for a single collection interval.
 type ZfsPool struct {
-	Total      float64 `json:"d" cbor:"0,keyasint"`                     // total capacity in GiB
-	Used       float64 `json:"du" cbor:"1,keyasint"`                    // allocated in GiB
-	ReadBytes  uint64  `json:"rb,omitzero" cbor:"2,keyasint,omitzero"`  // read throughput in bytes/s
-	WriteBytes uint64  `json:"wb,omitzero" cbor:"3,keyasint,omitzero"`  // write throughput in bytes/s
-	Health     string  `json:"h,omitempty" cbor:"4,keyasint,omitempty"` // ONLINE, DEGRADED, FAULTED, ...
+	DisplayName string  `json:"n,omitempty" cbor:"8,keyasint,omitempty"`
+	HideUsage   bool    `json:"hu,omitempty" cbor:"6,keyasint,omitempty"` // equivalent filesystem usage chart exists
+	HideIO      bool    `json:"hi,omitempty" cbor:"7,keyasint,omitempty"` // equivalent filesystem I/O chart exists
+	Raw         bool    `json:"raw,omitempty" cbor:"5,keyasint,omitempty"`
+	Total       float64 `json:"d" cbor:"0,keyasint"`                     // total capacity in GiB
+	Used        float64 `json:"du" cbor:"1,keyasint"`                    // allocated in GiB
+	ReadBytes   uint64  `json:"rb,omitzero" cbor:"2,keyasint,omitzero"`  // read throughput in bytes/s
+	WriteBytes  uint64  `json:"wb,omitzero" cbor:"3,keyasint,omitzero"`  // write throughput in bytes/s
+	Health      string  `json:"h,omitempty" cbor:"4,keyasint,omitempty"` // ONLINE, DEGRADED, FAULTED, ...
 }
 
 // Uint8Slice wraps []uint8 to customize JSON encoding while keeping CBOR efficient.
@@ -206,5 +211,6 @@ type CombinedData struct {
 	Details         *Details           `cbor:"4,keyasint,omitempty"`
 	// SystemdServicesUpdated distinguishes a fresh empty snapshot from a response
 	// that omitted systemd data (for example, a short-cache dashboard request).
-	SystemdServicesUpdated bool `json:"systemdUpdated,omitempty" cbor:"5,keyasint,omitempty"`
+	SystemdServicesUpdated bool                      `json:"systemdUpdated,omitempty" cbor:"5,keyasint,omitempty"`
+	Monitors               map[string]monitor.Result `cbor:"6,keyasint"`
 }

internal/entities/zfs/zfs.go

@@ -1,23 +1,47 @@
 // Package zfs defines the ZFS detail data exchanged between agent and hub.
 package zfs
 
+import "strings"
+
 // ZfsData is the detail payload returned by the agent for the GetZfsData action.
 type ZfsData struct {
 	Pools    []*PoolDetail `json:"pools,omitempty"`
 	Complete bool          `json:"complete,omitempty"`
+	// Backends whose inventories are complete, even when another backend failed.
+	CompleteBackends []string `json:"completeBackends,omitempty"`
+}
+
+// CanRefreshPool also governs deletion: missing pools may only be removed
+// after a successful inventory of their backend. Complete supports old agents.
+func (data *ZfsData) CanRefreshPool(name string) bool {
+	if data.Complete {
+		return true
+	}
+	backend := "zfs"
+	if strings.HasPrefix(name, "b:") {
+		backend = "btrfs"
+	}
+	for _, complete := range data.CompleteBackends {
+		if complete == backend {
+			return true
+		}
+	}
+	return false
 }
 
 // PoolDetail holds the verbose state of a single pool: capacity, health,
 // scrub, vdev, and dataset information.
 type PoolDetail struct {
-	Name     string     `json:"name"`
-	Health   string     `json:"health,omitempty"`
-	Size     uint64     `json:"size,omitempty"`  // bytes
-	Alloc    uint64     `json:"alloc,omitempty"` // bytes
-	Free     uint64     `json:"free,omitempty"`  // bytes
-	Scrub    *Scrub     `json:"scrub,omitempty"`
-	Vdevs    []*Vdev    `json:"vdevs,omitempty"`
-	Datasets []*Dataset `json:"datasets,omitempty"`
+	DisplayName string     `json:"displayName,omitempty"`
+	Raw         bool       `json:"raw,omitempty"`
+	Name        string     `json:"name"`
+	Health      string     `json:"health,omitempty"`
+	Size        uint64     `json:"size,omitempty"`  // bytes
+	Alloc       uint64     `json:"alloc,omitempty"` // bytes
+	Free        uint64     `json:"free,omitempty"`  // bytes
+	Scrub       *Scrub     `json:"scrub,omitempty"`
+	Vdevs       []*Vdev    `json:"vdevs,omitempty"`
+	Datasets    []*Dataset `json:"datasets,omitempty"`
 }
 
 // Scrub holds the scrub (or resilver) status of a pool.

internal/hub/api.go

@@ -2,13 +2,17 @@ package hub
 
 import (
 	"context"
+	"fmt"
+	"log/slog"
+	"net"
 	"net/http"
+	"net/netip"
 	"regexp"
 	"strings"
 	"time"
+	"uuid"
 
 	"github.com/blang/semver"
-	"github.com/google/uuid"
 	"github.com/henrygd/beszel"
 	"github.com/henrygd/beszel/internal/alerts"
 	"github.com/henrygd/beszel/internal/ghupdate"
@@ -78,12 +82,81 @@ func (h *Hub) registerMiddlewares(se *core.ServeEvent) {
 	}
 	// authenticate with trusted header
 	if trustedHeader, _ := utils.GetEnv("TRUSTED_AUTH_HEADER"); trustedHeader != "" {
+		// only honor the header from these peers, if set
+		trustedProxies, restricted := parseTrustedProxies()
 		se.Router.BindFunc(func(e *core.RequestEvent) error {
+			if restricted && !isTrustedProxy(trustedProxies, e.Request.RemoteAddr) {
+				return e.Next()
+			}
 			return authorizeRequestWithEmail(e, e.Request.Header.Get(trustedHeader))
 		})
 	}
 }
 
+// parseTrustedProxies reads TRUSTED_PROXY_IPS (comma-separated IPs or CIDRs).
+// restricted is false when the variable is unset or empty, meaning the trusted
+// header is accepted from any peer. Invalid entries are skipped with a warning,
+// so a typo narrows the allowlist rather than widening it.
+func parseTrustedProxies() (prefixes []netip.Prefix, restricted bool) {
+	value, _ := utils.GetEnv("TRUSTED_PROXY_IPS")
+	if value == "" {
+		return nil, false
+	}
+	for entry := range strings.SplitSeq(value, ",") {
+		entry = strings.TrimSpace(entry)
+		if entry == "" {
+			continue
+		}
+		if prefix, err := parseProxyPrefix(entry); err == nil {
+			prefixes = append(prefixes, prefix)
+		} else {
+			slog.Warn("Ignoring invalid TRUSTED_PROXY_IPS entry", "entry", entry)
+		}
+	}
+	return prefixes, true
+}
+
+// parseProxyPrefix parses an IP or CIDR into a masked prefix. IPv4-mapped IPv6
+// entries are converted to IPv4 so they match IPv4 peers.
+func parseProxyPrefix(entry string) (netip.Prefix, error) {
+	prefix, err := netip.ParsePrefix(entry)
+	if err != nil {
+		addr, err := netip.ParseAddr(entry)
+		if err != nil {
+			return netip.Prefix{}, err
+		}
+		addr = addr.Unmap()
+		return netip.PrefixFrom(addr, addr.BitLen()), nil
+	}
+	if prefix.Addr().Is4In6() {
+		if prefix.Bits() < 96 {
+			return netip.Prefix{}, fmt.Errorf("%s covers more than the IPv4-mapped range", entry)
+		}
+		prefix = netip.PrefixFrom(prefix.Addr().Unmap(), prefix.Bits()-96)
+	}
+	return prefix.Masked(), nil
+}
+
+// isTrustedProxy reports whether the peer address of a request (host:port) is
+// within one of the prefixes.
+func isTrustedProxy(prefixes []netip.Prefix, remoteAddr string) bool {
+	host, _, err := net.SplitHostPort(remoteAddr)
+	if err != nil {
+		host = remoteAddr
+	}
+	addr, err := netip.ParseAddr(host)
+	if err != nil {
+		return false
+	}
+	addr = addr.Unmap().WithZone("")
+	for _, prefix := range prefixes {
+		if prefix.Contains(addr) {
+			return true
+		}
+	}
+	return false
+}
+
 // registerApiRoutes registers custom API routes
 func (h *Hub) registerApiRoutes(se *core.ServeEvent) error {
 	// auth protected routes

internal/hub/api_test.go

@@ -6,12 +6,16 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/http/httptest"
+	"sort"
 	"testing"
+	"time"
 
 	beszelTests "github.com/henrygd/beszel/internal/tests"
 
 	"github.com/henrygd/beszel/internal/migrations"
 	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/apis"
 	"github.com/pocketbase/pocketbase/core"
 	pbTests "github.com/pocketbase/pocketbase/tests"
 	"github.com/stretchr/testify/require"
@@ -26,6 +30,59 @@ func jsonReader(v any) io.Reader {
 	return bytes.NewReader(data)
 }
 
+type gatedReader struct {
+	data    []byte
+	started chan struct{}
+	release chan struct{}
+	offset  int
+}
+
+func (r *gatedReader) Read(p []byte) (int, error) {
+	if r.offset == 0 {
+		close(r.started)
+		<-r.release
+	}
+	if r.offset >= len(r.data) {
+		return 0, io.EOF
+	}
+	n := copy(p, r.data[r.offset:])
+	r.offset += n
+	return n, nil
+}
+
+func firstUserTestMux(t *testing.T) (*beszelTests.TestHub, http.Handler) {
+	t.Helper()
+	hub, err := beszelTests.NewTestHub(t.TempDir())
+	require.NoError(t, err)
+	_ = hub.StartHub()
+
+	router, err := apis.NewRouter(hub.TestApp)
+	require.NoError(t, err)
+	serveEvent := &core.ServeEvent{App: hub.TestApp, Router: router}
+
+	var handler http.Handler
+	err = hub.TestApp.OnServe().Trigger(serveEvent, func(e *core.ServeEvent) error {
+		var buildErr error
+		handler, buildErr = e.Router.BuildMux()
+		return buildErr
+	})
+	require.NoError(t, err)
+	require.NotNil(t, handler)
+	return hub, handler
+}
+
+func postFirstUser(handler http.Handler, email string) *httptest.ResponseRecorder {
+	body, _ := json.Marshal(map[string]string{
+		"email":    email,
+		"password": "password123",
+	})
+	req := httptest.NewRequest(http.MethodPost, "/api/beszel/create-user", bytes.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	recorder := httptest.NewRecorder()
+	handler.ServeHTTP(recorder, req)
+	return recorder
+}
+
 func TestApiRoutesAuthentication(t *testing.T) {
 	hub, user := beszelTests.GetHubWithUser(t)
 	defer hub.Cleanup()
@@ -789,6 +846,87 @@ func TestFirstUserCreation(t *testing.T) {
 	})
 }
 
+func TestFirstUserBootstrapAtomicity(t *testing.T) {
+	t.Run("concurrent complete requests produce exactly one winner", func(t *testing.T) {
+		hub, handler := firstUserTestMux(t)
+		defer hub.Cleanup()
+
+		start := make(chan struct{})
+		statuses := make(chan int, 2)
+		for _, email := range []string{"first@example.com", "second@example.com"} {
+			go func(email string) {
+				<-start
+				statuses <- postFirstUser(handler, email).Code
+			}(email)
+		}
+		close(start)
+
+		got := []int{<-statuses, <-statuses}
+		sort.Ints(got)
+		require.Equal(t, []int{http.StatusOK, http.StatusForbidden}, got)
+
+		users, err := hub.FindAllRecords("users")
+		require.NoError(t, err)
+		require.Len(t, users, 1)
+		superusers, err := hub.FindAllRecords(core.CollectionNameSuperusers)
+		require.NoError(t, err)
+		require.Len(t, superusers, 1)
+		require.NotEqual(t, migrations.TempAdminEmail, superusers[0].Email())
+	})
+
+	t.Run("partial body cannot retain stale bootstrap authorization", func(t *testing.T) {
+		hub, handler := firstUserTestMux(t)
+		defer hub.Cleanup()
+
+		body, err := json.Marshal(map[string]string{
+			"email":    "parked@example.com",
+			"password": "password123",
+		})
+		require.NoError(t, err)
+		gated := &gatedReader{
+			data:    body,
+			started: make(chan struct{}),
+			release: make(chan struct{}),
+		}
+		parkedRequest := httptest.NewRequest(http.MethodPost, "/api/beszel/create-user", gated)
+		parkedRequest.Header.Set("Content-Type", "application/json")
+		parkedRecorder := httptest.NewRecorder()
+		parkedDone := make(chan struct{})
+		go func() {
+			handler.ServeHTTP(parkedRecorder, parkedRequest)
+			close(parkedDone)
+		}()
+
+		select {
+		case <-gated.started:
+		case <-time.After(2 * time.Second):
+			t.Fatal("parked request did not begin reading its body")
+		}
+
+		operatorRecorder := postFirstUser(handler, "operator@example.com")
+		require.Equal(t, http.StatusOK, operatorRecorder.Code)
+		lateRecorder := postFirstUser(handler, "late@example.com")
+		require.Equal(t, http.StatusForbidden, lateRecorder.Code)
+
+		close(gated.release)
+		select {
+		case <-parkedDone:
+		case <-time.After(2 * time.Second):
+			t.Fatal("parked request did not finish")
+		}
+		require.Equal(t, http.StatusForbidden, parkedRecorder.Code)
+
+		users, err := hub.FindAllRecords("users")
+		require.NoError(t, err)
+		require.Len(t, users, 1)
+		require.Equal(t, "operator@example.com", users[0].Email())
+		superusers, err := hub.FindAllRecords(core.CollectionNameSuperusers)
+		require.NoError(t, err)
+		require.Len(t, superusers, 1)
+		require.Equal(t, "operator@example.com", superusers[0].Email())
+	})
+}
+
 func TestCreateUserEndpointAvailability(t *testing.T) {
 	t.Run("CreateUserEndpoint available when no users exist", func(t *testing.T) {
 		hub, _ := beszelTests.NewTestHub(t.TempDir())
@@ -969,6 +1107,79 @@ func TestTrustedHeaderMiddleware(t *testing.T) {
 	}
 }
 
+func TestTrustedHeaderProxyAllowlist(t *testing.T) {
+	var hubs []*beszelTests.TestHub
+
+	defer func() {
+		for _, hub := range hubs {
+			hub.Cleanup()
+		}
+	}()
+
+	testAppFactory := func(t testing.TB) *pbTests.TestApp {
+		hub, _ := beszelTests.NewTestHub(t.TempDir())
+		hubs = append(hubs, hub)
+		hub.StartHub()
+		return hub.TestApp
+	}
+
+	// httptest requests arrive from 192.0.2.1:1234
+	testCases := []struct {
+		name            string
+		proxies         string
+		expectedStatus  int
+		expectedContent []string
+	}{
+		{
+			name:            "peer inside an allowed range",
+			proxies:         "10.0.0.0/8, 192.0.2.0/24",
+			expectedStatus:  200,
+			expectedContent: []string{"\"key\":", "\"v\":"},
+		},
+		{
+			name:            "peer is the listed address",
+			proxies:         "192.0.2.1",
+			expectedStatus:  200,
+			expectedContent: []string{"\"key\":", "\"v\":"},
+		},
+		{
+			name:            "peer outside the allowlist",
+			proxies:         "10.0.0.0/8",
+			expectedStatus:  401,
+			expectedContent: []string{"requires valid"},
+		},
+		{
+			name:            "allowlist with no valid entry",
+			proxies:         "proxy.internal",
+			expectedStatus:  401,
+			expectedContent: []string{"requires valid"},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv("TRUSTED_AUTH_HEADER", "X-Beszel-Trusted")
+			t.Setenv("TRUSTED_PROXY_IPS", tc.proxies)
+
+			scenario := beszelTests.ApiScenario{
+				Name:   "GET /getkey - with trusted header",
+				Method: http.MethodGet,
+				URL:    "/api/beszel/getkey",
+				Headers: map[string]string{
+					"X-Beszel-Trusted": "user@test.com",
+				},
+				ExpectedStatus:  tc.expectedStatus,
+				ExpectedContent: tc.expectedContent,
+				TestAppFactory:  testAppFactory,
+				BeforeTestFunc: func(t testing.TB, app *pbTests.TestApp, e *core.ServeEvent) {
+					beszelTests.CreateUser(app, "user@test.com", "password123")
+				},
+			}
+			scenario.Test(t)
+		})
+	}
+}
+
 func TestUpdateEndpoint(t *testing.T) {
 	t.Setenv("CHECK_UPDATES", "true")
 

internal/hub/collections.go

@@ -78,7 +78,7 @@ func setCollectionAuthSettings(app core.App) error {
 		return err
 	}
 
-	if err := applyCollectionRules(app, []string{"containers", "container_stats", "system_stats", "systemd_services"}, collectionRules{
+	if err := applyCollectionRules(app, []string{"containers", "container_stats", "system_stats", "systemd_services", "network_monitor_stats"}, collectionRules{
 		list: &systemScopedReadRule,
 	}); err != nil {
 		return err
@@ -99,6 +99,16 @@ func setCollectionAuthSettings(app core.App) error {
 	}
 
 	if err := applyCollectionRules(app, []string{"fingerprints"}, collectionRules{
+		list:   &systemScopedWriteRule,
+		view:   &systemScopedWriteRule,
+		create: &systemScopedWriteRule,
+		update: &systemScopedWriteRule,
+		delete: &systemScopedWriteRule,
+	}); err != nil {
+		return err
+	}
+
+	if err := applyCollectionRules(app, []string{"network_monitors"}, collectionRules{
 		list:   &systemScopedReadRule,
 		view:   &systemScopedReadRule,
 		create: &systemScopedWriteRule,

internal/hub/collections_test.go

@@ -88,8 +88,8 @@ func TestCollectionRulesDefault(t *testing.T) {
 	// fingerprints collection
 	fingerprintsCollection, err := hub.FindCollectionByNameOrId("fingerprints")
 	require.NoError(t, err, "Failed to find fingerprints collection")
-	assert.Equal(t, isUserInSystemUsers, *fingerprintsCollection.ListRule)
-	assert.Equal(t, isUserInSystemUsers, *fingerprintsCollection.ViewRule)
+	assert.Equal(t, isUserInSystemUsersNotReadonly, *fingerprintsCollection.ListRule)
+	assert.Equal(t, isUserInSystemUsersNotReadonly, *fingerprintsCollection.ViewRule)
 	assert.Equal(t, isUserInSystemUsersNotReadonly, *fingerprintsCollection.CreateRule)
 	assert.Equal(t, isUserInSystemUsersNotReadonly, *fingerprintsCollection.UpdateRule)
 	assert.Equal(t, isUserInSystemUsersNotReadonly, *fingerprintsCollection.DeleteRule)
@@ -216,8 +216,8 @@ func TestCollectionRulesShareAllSystems(t *testing.T) {
 	// fingerprints collection
 	fingerprintsCollection, err := hub.FindCollectionByNameOrId("fingerprints")
 	require.NoError(t, err, "Failed to find fingerprints collection")
-	assert.Equal(t, isUser, *fingerprintsCollection.ListRule)
-	assert.Equal(t, isUser, *fingerprintsCollection.ViewRule)
+	assert.Equal(t, isUserNotReadonly, *fingerprintsCollection.ListRule)
+	assert.Equal(t, isUserNotReadonly, *fingerprintsCollection.ViewRule)
 	assert.Equal(t, isUserNotReadonly, *fingerprintsCollection.CreateRule)
 	assert.Equal(t, isUserNotReadonly, *fingerprintsCollection.UpdateRule)
 	assert.Equal(t, isUserNotReadonly, *fingerprintsCollection.DeleteRule)

internal/hub/config/config.go

@@ -6,8 +6,8 @@ import (
 	"log"
 	"os"
 	"path/filepath"
+	"uuid"
 
-	"github.com/google/uuid"
 	"github.com/henrygd/beszel/internal/entities/system"
 
 	"github.com/pocketbase/dbx"

internal/hub/hub.go

@@ -109,6 +109,8 @@ func (h *Hub) StartHub() error {
 	h.App.OnRecordCreate("users").BindFunc(h.um.InitializeUserRole)
 	h.App.OnRecordCreate("user_settings").BindFunc(h.um.InitializeUserSettings)
 
+	bindNetworkMonitorsEvents(h)
+
 	pb, ok := h.App.(*pocketbase.PocketBase)
 	if !ok {
 		return errors.New("not a pocketbase app")
@@ -122,6 +124,8 @@ func (h *Hub) initialize(app core.App) error {
 	settings := app.Settings()
 	// batch requests (for alerts)
 	settings.Batch.Enabled = true
+	settings.Batch.MaxRequests = 100
+	settings.Batch.MaxBodySize = 1 << 20 // 1 MiB
 	// set URL if APP_URL env is set
 	if appURL, isSet := utils.GetEnv("APP_URL"); isSet {
 		h.appURL = appURL

internal/hub/network_monitors.go

@@ -0,0 +1,158 @@
+package hub
+
+import (
+	"strconv"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/henrygd/beszel/internal/hub/systems"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/pocketbase/pocketbase/tools/types"
+)
+
+// generateMonitorID creates a stable hash ID for a monitor based on its configuration and the system it belongs to.
+func generateMonitorID(systemId string, config monitor.Config) string {
+	args := []string{systemId, config.Target, config.Protocol}
+	// only use port for TCP monitors, since for other protocols it's not relevant as standalone value
+	if config.Protocol == "tcp" {
+		args = append(args, strconv.FormatUint(uint64(config.Port), 10))
+	}
+	return systems.MakeStableHashId(args...)
+}
+
+// bindNetworkMonitorsEvents keeps monitor records and agent monitor state in sync.
+func bindNetworkMonitorsEvents(hub *Hub) {
+	// on create, make sure the id is set to a stable hash
+	hub.OnRecordCreate("network_monitors").BindFunc(func(e *core.RecordEvent) error {
+		systemID := e.Record.GetString("system")
+		config := monitorConfigFromRecord(e.Record)
+		id := generateMonitorID(systemID, *config)
+		e.Record.Set("id", id)
+		return e.Next()
+	})
+
+	// sync monitor to agent on creation and persist the first result immediately when available
+	hub.OnRecordAfterCreateSuccess("network_monitors").BindFunc(func(e *core.RecordEvent) error {
+		err := e.Next()
+		if err != nil {
+			return err
+		}
+		if !e.Record.GetBool("enabled") {
+			return nil
+		}
+		// If connected, run the monitor immediately. Paused systems may be absent
+		// from the manager; their monitors will sync when they reconnect.
+		system, err := hub.sm.GetSystem(e.Record.GetString("system"))
+		if err == nil && system.Status == "up" {
+			go hub.upsertNetworkMonitor(e.Record, true)
+		}
+		return nil
+	})
+
+	// On API update requests, if the monitor config changed in a way that requires a new ID, create a new
+	// record with the new ID and delete the old one. Otherwise, just update the existing monitor on the agent.
+	hub.OnRecordUpdateRequest("network_monitors").BindFunc(func(e *core.RecordRequestEvent) error {
+		systemID := e.Record.GetString("system")
+		// only tcp uses port - set other protocols port to zero
+		if e.Record.GetString("protocol") != "tcp" {
+			e.Record.Set("port", 0)
+		}
+		ID := generateMonitorID(systemID, *monitorConfigFromRecord(e.Record))
+		if ID != e.Record.Id {
+			newRecord := copyMonitorToNewRecord(e.Record, ID)
+			if err := e.App.Save(newRecord); err != nil {
+				return err
+			}
+			if err := e.App.Delete(e.Record); err != nil {
+				return err
+			}
+			return nil
+		}
+		err := e.Next()
+		if err != nil {
+			return err
+		}
+		if e.Record.GetBool("enabled") {
+			// if the monitor is enabled, sync the updated config to the agent now
+			runNow := !e.Record.Original().GetBool("enabled")
+			err = hub.upsertNetworkMonitor(e.Record, runNow)
+		} else {
+			// if the monitor is paused, remove it from the agent
+			err = hub.deleteNetworkMonitor(e.Record)
+		}
+		if err != nil {
+			hub.Logger().Warn("failed to sync updated monitor", "system", systemID, "monitor", e.Record.Id, "err", err)
+		}
+		return nil
+	})
+
+	// sync monitor to agent on delete
+	hub.OnRecordAfterDeleteSuccess("network_monitors").BindFunc(func(e *core.RecordEvent) error {
+		if err := hub.deleteNetworkMonitor(e.Record); err != nil {
+			hub.Logger().Warn("failed to delete monitor on agent", "system", e.Record.GetString("system"), "monitor", e.Record.Id, "err", err)
+		}
+		return e.Next()
+	})
+}
+
+// monitorConfigFromRecord builds a monitor config from a network_monitors record.
+func monitorConfigFromRecord(record *core.Record) *monitor.Config {
+	return &monitor.Config{
+		ID:       record.Id,
+		Target:   record.GetString("target"),
+		Protocol: record.GetString("protocol"),
+		Port:     uint16(record.GetInt("port")),
+		Interval: uint16(record.GetInt("interval")),
+	}
+}
+
+// setMonitorResultFields stores the latest monitor result values on the record.
+func setMonitorResultFields(record *core.Record, result monitor.Result) {
+	nowString := time.Now().UTC().Format(types.DefaultDateLayout)
+	record.Set("res", result.AvgResponse)
+	record.Set("resAvg1h", result.AvgResponse1h)
+	record.Set("resMin1h", result.MinResponse1h)
+	record.Set("resMax1h", result.MaxResponse1h)
+	record.Set("loss1h", result.PacketLoss1h)
+	record.Set("updated", nowString)
+}
+
+// copyMonitorToNewRecord creates a new record with the same field values as the old one.
+// This is used when the monitor config changes in a way that requires a new ID, so we need
+// to create a new record with the new ID and delete the old one.
+func copyMonitorToNewRecord(oldRecord *core.Record, newID string) *core.Record {
+	collection := oldRecord.Collection()
+	newRecord := core.NewRecord(collection)
+	newRecord.Id = newID
+	fields := []string{"system", "target", "protocol", "port", "interval", "enabled"}
+	for _, field := range fields {
+		newRecord.Set(field, oldRecord.Get(field))
+	}
+	return newRecord
+}
+
+// upsertNetworkMonitor creates or updates the record's monitor on the target system. If runNow
+// is true, it will also trigger an immediate monitor run and update the record with the result.
+func (h *Hub) upsertNetworkMonitor(record *core.Record, runNow bool) error {
+	systemID := record.GetString("system")
+	system, err := h.sm.GetSystem(systemID)
+	if err != nil {
+		return err
+	}
+	result, err := system.UpsertNetworkMonitor(*monitorConfigFromRecord(record), runNow)
+	if err != nil || result == nil {
+		return err
+	}
+	setMonitorResultFields(record, *result)
+	return h.App.SaveNoValidate(record)
+}
+
+// deleteNetworkMonitor removes the record's monitor from the target system.
+func (h *Hub) deleteNetworkMonitor(record *core.Record) error {
+	systemID := record.GetString("system")
+	system, err := h.sm.GetSystem(systemID)
+	if err != nil {
+		return err
+	}
+	return system.DeleteNetworkMonitor(record.Id)
+}

internal/hub/network_monitors_test.go

@@ -0,0 +1,225 @@
+package hub
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/pocketbase/pocketbase/apis"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCreateNetworkMonitorsOnPausedSystem(t *testing.T) {
+	for _, batch := range []bool{false, true} {
+		name := "single"
+		if batch {
+			name = "batch"
+		}
+		t.Run(name, func(t *testing.T) {
+			hub, testApp, err := createTestHub(t)
+			require.NoError(t, err)
+			defer cleanupTestHub(hub, testApp)
+			bindNetworkMonitorsEvents(hub)
+
+			user, err := createTestUser(hub)
+			require.NoError(t, err)
+			system, err := createTestRecord(hub, "systems", map[string]any{
+				"name": "Paused", "host": "localhost", "port": "45876",
+				"status": "paused", "users": []string{user.Id},
+			})
+			require.NoError(t, err)
+			// Paused systems are not loaded into the manager at startup.
+			_, err = hub.sm.GetSystem(system.Id)
+			require.Error(t, err)
+
+			payload := func(target string) map[string]any {
+				return map[string]any{
+					"system": system.Id, "target": target, "protocol": "icmp",
+					"interval": 60, "enabled": true,
+				}
+			}
+			url := "/api/collections/network_monitors/records"
+			var body any = payload("1.1.1.1")
+			count := 1
+			if batch {
+				body = map[string]any{"requests": []map[string]any{
+					{"method": "POST", "url": url, "body": payload("1.1.1.1")},
+					{"method": "POST", "url": url, "body": payload("8.8.8.8")},
+				}}
+				url = "/api/batch"
+				count = 2
+			}
+			data, err := json.Marshal(body)
+			require.NoError(t, err)
+			token, err := user.NewAuthToken()
+			require.NoError(t, err)
+			router, err := apis.NewRouter(hub)
+			require.NoError(t, err)
+			handler, err := router.BuildMux()
+			require.NoError(t, err)
+			request := httptest.NewRequest(http.MethodPost, url, bytes.NewReader(data))
+			request.Header.Set("Content-Type", "application/json")
+			request.Header.Set("Authorization", token)
+			response := httptest.NewRecorder()
+			handler.ServeHTTP(response, request)
+			assert.Equal(t, http.StatusOK, response.Code, response.Body.String())
+
+			records, err := hub.FindAllRecords("network_monitors")
+			require.NoError(t, err)
+			require.Len(t, records, count)
+			for _, record := range records {
+				assert.Equal(t, system.Id, record.GetString("system"))
+				assert.True(t, record.GetBool("enabled"))
+			}
+		})
+	}
+}
+
+func TestGenerateMonitorID(t *testing.T) {
+	tests := []struct {
+		name     string
+		systemID string
+		config   monitor.Config
+		expected string
+	}{
+		{
+			name:     "HTTP monitor on example.com",
+			systemID: "sys123",
+			config: monitor.Config{
+				Protocol: "http",
+				Target:   "example.com",
+				Port:     0,
+				Interval: 60,
+			},
+			expected: "a20a5827",
+		},
+		{
+			name:     "HTTP monitor on example.com with different port",
+			systemID: "sys123",
+			config: monitor.Config{
+				Protocol: "http",
+				Target:   "example.com",
+				Port:     8080,
+				Interval: 60,
+			},
+			expected: "a20a5827",
+		},
+		{
+			name:     "HTTP monitor on example.com with different system ID",
+			systemID: "sys1234",
+			config: monitor.Config{
+				Protocol: "http",
+				Target:   "example.com",
+				Port:     80,
+				Interval: 60,
+			},
+			expected: "ab602ae7",
+		},
+		{
+			name:     "Same monitor, different interval",
+			systemID: "sys1234",
+			config: monitor.Config{
+				Protocol: "http",
+				Target:   "example.com",
+				Port:     80,
+				Interval: 120,
+			},
+			expected: "ab602ae7",
+		},
+		{
+			name:     "ICMP monitor on 1.1.1.1",
+			systemID: "sys456",
+			config: monitor.Config{
+				Protocol: "icmp",
+				Target:   "1.1.1.1",
+				Port:     0,
+				Interval: 10,
+			},
+			expected: "6d13a4a4",
+		}, {
+			name:     "ICMP monitor on 1.1.1.1 with different system ID",
+			systemID: "sys4567",
+			config: monitor.Config{
+				Protocol: "icmp",
+				Target:   "1.1.1.1",
+				Port:     0,
+				Interval: 10,
+			},
+			expected: "ddd6c81",
+		},
+		{
+			name:     "TCP monitor on example.com with port 443",
+			systemID: "sys789",
+			config: monitor.Config{
+				Protocol: "tcp",
+				Target:   "example.com",
+				Port:     443,
+				Interval: 30,
+			},
+			expected: "677b991",
+		},
+		{
+			name:     "TCP monitor on example.com with port 8443",
+			systemID: "sys789",
+			config: monitor.Config{
+				Protocol: "tcp",
+				Target:   "example.com",
+				Port:     8443,
+				Interval: 30,
+			},
+			expected: "84167969",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := generateMonitorID(tt.systemID, tt.config)
+			assert.Equal(t, tt.expected, got, "generateMonitorID() = %v, want %v", got, tt.expected)
+		})
+	}
+}
+
+func TestCopyMonitorToNewRecordDropsResultFields(t *testing.T) {
+	hub, testApp, err := createTestHub(t)
+	require.NoError(t, err)
+	defer cleanupTestHub(hub, testApp)
+
+	collection, err := hub.FindCachedCollectionByNameOrId("network_monitors")
+	require.NoError(t, err)
+	assert.Nil(t, collection.Fields.GetByName("name"))
+
+	oldRecord := core.NewRecord(collection)
+	oldRecord.Load(map[string]any{
+		"system":   "sys123",
+		"target":   "https://example.com",
+		"protocol": "http",
+		"port":     443,
+		"interval": 60,
+		"enabled":  true,
+		"res":      1200,
+		"resAvg1h": 1300,
+		"resMin1h": 900,
+		"resMax1h": 1600,
+		"loss1h":   5,
+		"updated":  "2026-04-29 12:00:00.000Z",
+	})
+
+	newRecord := copyMonitorToNewRecord(oldRecord, "next12345")
+
+	assert.Equal(t, "next12345", newRecord.Id)
+	assert.Equal(t, "https://example.com", newRecord.GetString("target"))
+	assert.Equal(t, "http", newRecord.GetString("protocol"))
+	assert.Equal(t, 443, newRecord.GetInt("port"))
+	assert.True(t, newRecord.GetBool("enabled"))
+	assert.Zero(t, newRecord.GetFloat("res"))
+	assert.Zero(t, newRecord.GetFloat("resAvg1h"))
+	assert.Zero(t, newRecord.GetFloat("resMin1h"))
+	assert.Zero(t, newRecord.GetFloat("resMax1h"))
+	assert.Zero(t, newRecord.GetFloat("loss1h"))
+	assert.Equal(t, "", newRecord.GetString("updated"))
+}

internal/hub/systems/container_records_test.go

@@ -0,0 +1,46 @@
+//go:build testing
+
+package systems
+
+import (
+	"testing"
+
+	"github.com/henrygd/beszel/internal/entities/container"
+	"github.com/pocketbase/dbx"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCreateContainerRecordsPersistsImageUpdateAvailability(t *testing.T) {
+	_, app := newTestSystemWithHub(t)
+
+	const (
+		systemID    = "system123"
+		containerID = "abcdef123456"
+		image       = "nginx:latest"
+	)
+
+	data := &container.Stats{
+		Id:              containerID,
+		Name:            "web",
+		Image:           image,
+		UpdateAvailable: true,
+	}
+	require.NoError(t, createContainerRecords(app, []*container.Stats{data}, systemID))
+
+	var record struct {
+		Image           string `db:"image"`
+		UpdateAvailable bool   `db:"updatable"`
+	}
+	require.NoError(t, app.DB().Select("image", "updatable").From("containers").
+		Where(dbx.HashExp{"id": containerID}).One(&record))
+	assert.Equal(t, image, record.Image)
+	assert.True(t, record.UpdateAvailable)
+
+	data.UpdateAvailable = false
+	require.NoError(t, createContainerRecords(app, []*container.Stats{data}, systemID))
+	require.NoError(t, app.DB().Select("image", "updatable").From("containers").
+		Where(dbx.HashExp{"id": containerID}).One(&record))
+	assert.Equal(t, image, record.Image)
+	assert.False(t, record.UpdateAvailable)
+}

internal/hub/systems/network_monitor_stats_test.go

@@ -0,0 +1,224 @@
+//go:build testing
+
+package systems
+
+import (
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/fxamacker/cbor/v2"
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/pocketbase/pocketbase/tools/subscriptions"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestNetworkMonitorProbePruning(t *testing.T) {
+	for _, tc := range []struct {
+		name     string
+		monitors map[string]monitor.Result
+		fail     bool
+		want     map[string]int64
+	}{
+		{"nil report", nil, false, map[string]int64{"monitor1": 1000, "monitor2": 1000}},
+		{"empty report", map[string]monitor.Result{}, false, map[string]int64{}},
+		{"removed monitor", map[string]monitor.Result{"monitor1": {LastProbeAt: 1000}}, false, map[string]int64{"monitor1": 1000}},
+		{"rolled back report", map[string]monitor.Result{}, true, map[string]int64{"monitor1": 1000, "monitor2": 1000}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			sys, app := newTestSystemWithHub(t)
+			sys.lastSavedMonitorProbe = map[string]int64{"monitor1": 1000, "monitor2": 1000}
+			// Preserve the distinction between nil and empty across the agent transport.
+			encoded, err := cbor.Marshal(system.CombinedData{Monitors: tc.monitors})
+			require.NoError(t, err)
+			var data system.CombinedData
+			require.NoError(t, cbor.Unmarshal(encoded, &data))
+			if tc.fail {
+				_, err = app.DB().NewQuery(`CREATE TRIGGER fail_system_update BEFORE UPDATE ON systems BEGIN SELECT RAISE(ABORT, 'test rollback'); END`).Execute()
+				require.NoError(t, err)
+			}
+			_, err = sys.createRecords(&data)
+			if tc.fail {
+				require.Error(t, err)
+			} else {
+				require.NoError(t, err)
+			}
+			assert.Equal(t, tc.want, sys.lastSavedMonitorProbe)
+		})
+	}
+}
+
+func TestNetworkMonitorStatsFreshness(t *testing.T) {
+	for _, realtime := range []bool{false, true} {
+		name := "sql"
+		if realtime {
+			name = "realtime"
+		}
+		t.Run(name, func(t *testing.T) {
+			sys, app := newTestSystemWithHub(t)
+			if realtime {
+				client := subscriptions.NewDefaultClient()
+				client.Subscribe("network_monitors/*")
+				app.SubscriptionsBroker().Register(client)
+				t.Cleanup(func() { app.SubscriptionsBroker().Unregister(client.Id()) })
+			}
+			col, err := app.FindCachedCollectionByNameOrId("network_monitors")
+			require.NoError(t, err)
+			for _, id := range []string{"monitor1", "monitor2"} {
+				record := core.NewRecord(col)
+				record.Id = id
+				record.Set("system", sys.Id)
+				require.NoError(t, app.SaveNoValidate(record))
+			}
+			data := &system.CombinedData{Monitors: map[string]monitor.Result{
+				"monitor1": {LastProbeAt: 1000, AvgResponse: 20, TotalCount: 6, SuccessCount: 6, ResponseSum: 123},
+				"monitor2": {LastProbeAt: 1000, PacketLoss: 100, TotalCount: 1},
+			}}
+			count := func(want int64) {
+				t.Helper()
+				got, err := app.CountRecords("network_monitor_stats")
+				require.NoError(t, err)
+				assert.Equal(t, want, got)
+			}
+			save := func() {
+				t.Helper()
+				_, err := sys.createRecords(data)
+				require.NoError(t, err)
+			}
+			save()
+			count(2)
+			stored, err := app.FindAllRecords("network_monitor_stats")
+			require.NoError(t, err)
+			for _, record := range stored {
+				result := data.Monitors[record.GetString("monitor")]
+				assert.EqualValues(t, result.TotalCount, record.GetInt("total_count"))
+				assert.EqualValues(t, result.SuccessCount, record.GetInt("success_count"))
+				assert.EqualValues(t, result.ResponseSum, record.GetInt("res_sum"))
+			}
+			// A resume can overlap the scheduled update with the same probe.
+			errs := make(chan error, 4)
+			for range 4 {
+				go func() {
+					_, err := sys.createRecords(data)
+					errs <- err
+				}()
+			}
+			for range 4 {
+				require.NoError(t, <-errs)
+			}
+			count(2)
+
+			// A rolling hourly value can change without a new probe.
+			result := data.Monitors["monitor1"]
+			result.AvgResponse1h = 42
+			data.Monitors["monitor1"] = result
+			save()
+			count(2)
+			record, err := app.FindRecordById("network_monitors", "monitor1")
+			require.NoError(t, err)
+			assert.Equal(t, 42, record.GetInt("resAvg1h"))
+
+			// Identical response values and failed probes still count as new measurements.
+			for id, result := range data.Monitors {
+				result.LastProbeAt = 301000
+				data.Monitors[id] = result
+			}
+			save()
+			count(4)
+
+			// A failed individual insert must remain retryable, even if others commit.
+			_, err = app.DB().NewQuery(`CREATE TRIGGER fail_monitor_insert BEFORE INSERT ON network_monitor_stats WHEN NEW.monitor = 'monitor1' BEGIN SELECT RAISE(ABORT, 'test insert failure'); END`).Execute()
+			require.NoError(t, err)
+			for id, result := range data.Monitors {
+				result.LastProbeAt = 601000
+				data.Monitors[id] = result
+			}
+			save()
+			count(5)
+			assert.Equal(t, int64(301000), sys.lastSavedMonitorProbe["monitor1"])
+			assert.Equal(t, int64(601000), sys.lastSavedMonitorProbe["monitor2"])
+			_, err = app.DB().NewQuery("DROP TRIGGER fail_monitor_insert").Execute()
+			require.NoError(t, err)
+			save()
+			count(6)
+
+			// Failure after inserting stats rolls back the whole transaction and its markers.
+			_, err = app.DB().NewQuery(`CREATE TRIGGER fail_system_update BEFORE UPDATE ON systems BEGIN SELECT RAISE(ABORT, 'test rollback'); END`).Execute()
+			require.NoError(t, err)
+			result = data.Monitors["monitor1"]
+			result.LastProbeAt = 901000
+			data.Monitors["monitor1"] = result
+			_, err = sys.createRecords(data)
+			require.Error(t, err)
+			count(6)
+			assert.Equal(t, int64(601000), sys.lastSavedMonitorProbe["monitor1"])
+			_, err = app.DB().NewQuery("DROP TRIGGER fail_system_update").Execute()
+			require.NoError(t, err)
+			save()
+			count(7)
+
+			// Clock rollback is a new probe identity, not a reason to stall writes.
+			result.LastProbeAt = 500
+			data.Monitors["monitor1"] = result
+			save()
+			count(8)
+
+			// Recreated systems intentionally accept the first result without restoring state.
+			sys = &System{Id: sys.Id, manager: sys.manager}
+			save()
+			count(10)
+		})
+	}
+}
+
+// Observes the committed DB through the hub, not the transaction's app.
+type monitorAlertHub struct {
+	stubHub
+	handle func(*core.Record, map[string]monitor.Result) error
+}
+
+func (h monitorAlertHub) HandleNetworkMonitorAlerts(record *core.Record, results map[string]monitor.Result) error {
+	return h.handle(record, results)
+}
+
+func TestNetworkMonitorAlertsAfterCommit(t *testing.T) {
+	for _, realtime := range []bool{false, true} {
+		t.Run(fmt.Sprint(realtime), func(t *testing.T) {
+			sys, app := newTestSystemWithHub(t)
+			if realtime {
+				client := subscriptions.NewDefaultClient()
+				client.Subscribe("network_monitors/*")
+				app.SubscriptionsBroker().Register(client)
+			}
+			collection, err := app.FindCachedCollectionByNameOrId("network_monitors")
+			require.NoError(t, err)
+			record := core.NewRecord(collection)
+			record.Set("system", sys.Id)
+			require.NoError(t, app.SaveNoValidate(record))
+			called := 0
+			result := monitor.Result{LastProbeAt: time.Now().UnixMilli(), SampleCount: 3, PacketLoss1h: 10}
+			sys.manager.hub = monitorAlertHub{stubHub: stubHub{app}, handle: func(systemRecord *core.Record, results map[string]monitor.Result) error {
+				called++
+				assert.Equal(t, sys.Id, systemRecord.Id)
+				assert.Equal(t, result, results[record.Id])
+				saved, err := app.FindRecordById("network_monitors", record.Id)
+				require.NoError(t, err)
+				assert.Equal(t, 10.0, saved.GetFloat("loss1h"))
+				return nil
+			}}
+			data := &system.CombinedData{Monitors: map[string]monitor.Result{record.Id: result}}
+			_, err = sys.createRecords(data)
+			require.NoError(t, err)
+			assert.Equal(t, 1, called)
+			// A transaction that fails after writing monitor stats must not notify.
+			_, err = app.DB().NewQuery(`CREATE TRIGGER fail_system BEFORE UPDATE ON systems BEGIN SELECT RAISE(ABORT, 'test rollback'); END`).Execute()
+			require.NoError(t, err)
+			_, err = sys.createRecords(data)
+			require.Error(t, err)
+			assert.Equal(t, 1, called)
+		})
+	}
+}

internal/hub/systems/network_monitor_sync_test.go

@@ -0,0 +1,144 @@
+//go:build testing
+
+package systems
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/blang/semver"
+	"github.com/fxamacker/cbor/v2"
+	"github.com/henrygd/beszel/internal/common"
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/henrygd/beszel/internal/hub/ws"
+	"github.com/lxzan/gws"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/stretchr/testify/require"
+)
+
+type monitorSyncClient struct {
+	gws.BuiltinEventHandler
+	requests chan common.HubRequest[monitor.SyncRequest]
+}
+
+func (c *monitorSyncClient) OnMessage(conn *gws.Conn, message *gws.Message) {
+	defer message.Close()
+	var req common.HubRequest[monitor.SyncRequest]
+	if err := cbor.Unmarshal(message.Bytes(), &req); err != nil {
+		return
+	}
+	c.requests <- req
+	data, _ := cbor.Marshal(monitor.SyncResponse{})
+	response, _ := cbor.Marshal(common.AgentResponse{Id: req.Id, Data: data})
+	_ = conn.WriteMessage(gws.OpcodeBinary, response)
+}
+
+// Avoid the production delayed disconnect notification; these tests explicitly
+// remove each connection from the manager before reconnecting.
+type monitorSyncServer struct{ ws.Handler }
+
+func (*monitorSyncServer) OnClose(*gws.Conn, error) {}
+
+func TestNetworkMonitorSyncSkipsOlderAgents(t *testing.T) {
+	for _, version := range []string{"0.0.0", "0.18.0", "0.19.0"} {
+		t.Run(version, func(t *testing.T) {
+			// No transport: attempting to send any request would fail.
+			sys := &System{agentVersion: semver.MustParse(version)}
+			require.NoError(t, sys.SyncNetworkMonitors(nil))
+			result, err := sys.UpsertNetworkMonitor(monitor.Config{ID: "test"}, true)
+			require.NoError(t, err)
+			require.Nil(t, result)
+			require.NoError(t, sys.DeleteNetworkMonitor("test"))
+		})
+	}
+}
+
+func TestNetworkMonitorReconnectSync(t *testing.T) {
+	for _, change := range []string{"delete", "disable"} {
+		t.Run(change, func(t *testing.T) {
+			sys, app := newTestSystemWithHub(t)
+			record, err := app.FindRecordById("systems", sys.Id)
+			require.NoError(t, err)
+			// Suppress unrelated system-stat requests while exercising reconnects.
+			record.Set("status", paused)
+			require.NoError(t, app.SaveNoValidate(record))
+			collection, err := app.FindCachedCollectionByNameOrId("network_monitors")
+			require.NoError(t, err)
+			probe := core.NewRecord(collection)
+			probe.Load(map[string]any{
+				"system": sys.Id, "target": "localhost", "protocol": "tcp",
+				"port": 80, "interval": 60, "enabled": true,
+			})
+			require.NoError(t, app.SaveNoValidate(probe))
+
+			sm := NewSystemManager(stubHub{app})
+			t.Cleanup(func() {
+				sm.cancel()
+				_ = sm.RemoveSystem(sys.Id)
+				sm.smartFetchMap.StopCleaner()
+				sm.zfsFetchMap.StopCleaner()
+			})
+			version := semver.MustParse("0.20.0")
+			connections := make(chan *ws.WsConn, 1)
+			upgrader := gws.NewUpgrader(&monitorSyncServer{}, nil)
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				conn, err := upgrader.Upgrade(w, r)
+				if err != nil {
+					t.Error(err)
+					return
+				}
+				wsConn := ws.NewWsConnection(conn, version)
+				conn.Session().Store("wsConn", wsConn)
+				connections <- wsConn
+				conn.ReadLoop()
+			}))
+			t.Cleanup(server.Close)
+			client := &monitorSyncClient{requests: make(chan common.HubRequest[monitor.SyncRequest], 2)}
+			connect := func() monitor.SyncRequest {
+				t.Helper()
+				conn, _, err := gws.NewClient(client, &gws.ClientOption{Addr: "ws" + strings.TrimPrefix(server.URL, "http")})
+				require.NoError(t, err)
+				t.Cleanup(func() { _ = conn.NetConn().Close() })
+				go conn.ReadLoop()
+				select {
+				case wsConn := <-connections:
+					require.NoError(t, sm.AddWebSocketSystem(sys.Id, version, wsConn))
+				case <-time.After(3 * time.Second):
+					t.Fatal("websocket connection was not established")
+				}
+				select {
+				case req := <-client.requests:
+					require.Equal(t, common.SyncNetworkMonitors, req.Action)
+					require.Equal(t, monitor.SyncActionReplace, req.Data.Action)
+					return req.Data
+				case <-time.After(3 * time.Second):
+					t.Fatal("reconnected agent did not receive a monitor replacement")
+					return monitor.SyncRequest{}
+				}
+			}
+
+			initial := connect()
+			require.Len(t, initial.Configs, 1)
+			require.Equal(t, probe.Id, initial.Configs[0].ID)
+			require.NoError(t, sm.RemoveSystem(sys.Id))
+			if change == "delete" {
+				require.NoError(t, app.Delete(probe))
+			} else {
+				probe.Set("enabled", false)
+				require.NoError(t, app.SaveNoValidate(probe))
+			}
+			require.Empty(t, connect().Configs, "reconnect must clear the agent's previous probe")
+		})
+	}
+}
+
+func TestGetMonitorConfigsForSystemQueryError(t *testing.T) {
+	sys, app := newTestSystemWithHub(t)
+	_, err := app.DB().NewQuery("DROP TABLE network_monitors").Execute()
+	require.NoError(t, err)
+	_, err = sys.manager.GetMonitorConfigsForSystem(sys.Id)
+	require.Error(t, err, "a failed query must not be treated as an empty monitor set")
+}

internal/hub/systems/network_monitors.go

@@ -0,0 +1,58 @@
+package systems
+
+import (
+	"context"
+	"time"
+
+	"github.com/henrygd/beszel"
+	"github.com/henrygd/beszel/internal/common"
+	"github.com/henrygd/beszel/internal/entities/monitor"
+)
+
+// SyncNetworkMonitors sends monitor configurations to the agent.
+func (sys *System) SyncNetworkMonitors(configs []monitor.Config) error {
+	_, err := sys.syncNetworkMonitors(monitor.SyncRequest{Action: monitor.SyncActionReplace, Configs: configs})
+	return err
+}
+
+// UpsertNetworkMonitor sends a single monitor configuration change to the agent.
+func (sys *System) UpsertNetworkMonitor(config monitor.Config, runNow bool) (*monitor.Result, error) {
+	resp, err := sys.syncNetworkMonitors(monitor.SyncRequest{
+		Action: monitor.SyncActionUpsert,
+		Config: config,
+		RunNow: runNow,
+	})
+	if err != nil {
+		return nil, err
+	}
+	if resp.Result == (monitor.Result{}) {
+		return nil, nil
+	}
+	result := resp.Result
+	return &result, nil
+}
+
+// DeleteNetworkMonitor removes a single monitor task from the agent.
+func (sys *System) DeleteNetworkMonitor(id string) error {
+	_, err := sys.syncNetworkMonitors(monitor.SyncRequest{
+		Action: monitor.SyncActionDelete,
+		Config: monitor.Config{ID: id},
+	})
+	return err
+}
+
+func (sys *System) syncNetworkMonitors(req monitor.SyncRequest) (monitor.SyncResponse, error) {
+	if sys.agentVersion.LT(beszel.MinVersionNetworkMonitors) {
+		return monitor.SyncResponse{}, nil
+	}
+	timeout := 5 * time.Second
+	if req.Action == monitor.SyncActionUpsert && req.RunNow {
+		// Allow the probe to finish, including a timeout result, while preserving
+		// the normal request budget for transport and response handling.
+		timeout += monitor.MaxProbeTimeout
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	var result monitor.SyncResponse
+	return result, sys.request(ctx, common.SyncNetworkMonitors, req, &result)
+}

internal/hub/systems/system.go

@@ -9,6 +9,7 @@ import (
 	"math/rand"
 	"net"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"time"
 
@@ -18,6 +19,7 @@ import (
 	"github.com/henrygd/beszel/internal/hub/ws"
 
 	"github.com/henrygd/beszel/internal/entities/container"
+	"github.com/henrygd/beszel/internal/entities/monitor"
 	"github.com/henrygd/beszel/internal/entities/smart"
 	"github.com/henrygd/beszel/internal/entities/system"
 	"github.com/henrygd/beszel/internal/entities/systemd"
@@ -30,6 +32,8 @@ import (
 	"github.com/lxzan/gws"
 	"github.com/pocketbase/dbx"
 	"github.com/pocketbase/pocketbase/core"
+	"github.com/pocketbase/pocketbase/tools/security"
+	"github.com/pocketbase/pocketbase/tools/types"
 	"golang.org/x/crypto/ssh"
 )
 
@@ -52,6 +56,10 @@ type System struct {
 	smartInterval  time.Duration              // Interval for periodic SMART data updates
 	zfsFetching    atomic.Bool                // True if ZFS pools are currently being fetched
 	zfsInterval    time.Duration              // Interval for periodic ZFS detail data updates
+	// Serialize persistence from scheduled updates and resumes through commit.
+	recordsMu sync.Mutex
+	// Protected by recordsMu; realtime reads don't consume probes.
+	lastSavedMonitorProbe map[string]int64
 }
 
 func (sm *SystemManager) NewSystem(systemId string) *System {
@@ -211,11 +219,15 @@ func (sys *System) handlePaused() {
 
 // createRecords updates the system record and adds system_stats and container_stats records
 func (sys *System) createRecords(data *system.CombinedData) (*core.Record, error) {
+	sys.recordsMu.Lock()
+	defer sys.recordsMu.Unlock()
+
 	systemRecord, err := sys.getRecord(sys.manager.hub)
 	if err != nil {
 		return nil, err
 	}
 	hub := sys.manager.hub
+	savedMonitorProbes := make(map[string]int64)
 	err = hub.RunInTransaction(func(txApp core.App) error {
 		// add system_stats record
 		systemStatsCollection, err := txApp.FindCachedCollectionByNameOrId("system_stats")
@@ -266,19 +278,56 @@ func (sys *System) createRecords(data *system.CombinedData) (*core.Record, error
 			}
 		}
 
+		if data.Monitors != nil {
+			if err := sys.updateNetworkMonitorsRecords(txApp, data.Monitors, savedMonitorProbes); err != nil {
+				return err
+			}
+		}
+
 		if err := sys.syncZfsPoolHealth(txApp, data.Stats.ZfsPools); err != nil {
 			return err
 		}
 
 		// update system record (do this last because it triggers alerts and we need above records to be inserted first)
 		systemRecord.Set("status", up)
-		systemRecord.Set("info", data.Info)
+		// Distinguish an idle GPU from a system without GPU data (#2312)
+		info := struct {
+			system.Info
+			GpuPct *float64 `json:"g,omitempty"`
+		}{Info: data.Info}
+		if len(data.Stats.GPUData) > 0 {
+			info.GpuPct = &data.Info.GpuPct
+		}
+		systemRecord.Set("info", info)
 		if err := txApp.SaveNoValidate(systemRecord); err != nil {
 			return err
 		}
 		return nil
 	})
 
+	// Publish only successful inserts after the entire transaction commits.
+	if err == nil && len(savedMonitorProbes) > 0 {
+		if sys.lastSavedMonitorProbe == nil {
+			sys.lastSavedMonitorProbe = savedMonitorProbes
+		} else {
+			for id, timestamp := range savedMonitorProbes {
+				sys.lastSavedMonitorProbe[id] = timestamp
+			}
+		}
+	}
+	// A non-nil report includes cached results for all remaining monitors.
+	if err == nil && data.Monitors != nil {
+		for id := range sys.lastSavedMonitorProbe {
+			if _, exists := data.Monitors[id]; !exists {
+				delete(sys.lastSavedMonitorProbe, id)
+			}
+		}
+	}
+	if err == nil {
+		if alertErr := hub.HandleNetworkMonitorAlerts(systemRecord, data.Monitors); alertErr != nil {
+			hub.Logger().Error("Error handling network monitor alerts", "err", alertErr)
+		}
+	}
 	return systemRecord, err
 }
 
@@ -322,9 +371,14 @@ func createSystemdStatsRecords(app core.App, data []*systemd.Service, systemId s
 
 	valueStrings := make([]string, 0, len(data))
 	for i, service := range data {
+		// Agent payloads can contain null entries. Reject the snapshot before
+		// executing any queries so existing service records remain intact.
+		if service == nil {
+			return fmt.Errorf("null systemd service at index %d", i)
+		}
 		suffix := fmt.Sprintf("%d", i)
 		valueStrings = append(valueStrings, fmt.Sprintf("({:id%[1]s}, {:system}, {:name%[1]s}, {:state%[1]s}, {:sub%[1]s}, {:cpu%[1]s}, {:cpuPeak%[1]s}, {:memory%[1]s}, {:memPeak%[1]s}, {:updated})", suffix))
-		params["id"+suffix] = makeStableHashId(systemId, service.Name)
+		params["id"+suffix] = MakeStableHashId(systemId, service.Name)
 		params["name"+suffix] = service.Name
 		params["state"+suffix] = service.State
 		params["sub"+suffix] = service.Sub
@@ -350,6 +404,106 @@ func createSystemdStatsRecords(app core.App, data []*systemd.Service, systemId s
 	return err
 }
 
+func (sys *System) updateNetworkMonitorsRecords(app core.App, monitorResults map[string]monitor.Result, savedProbes map[string]int64) error {
+	if len(monitorResults) == 0 {
+		return nil
+	}
+	var err error
+	systemId := sys.Id
+	const monitorCollectionName = "network_monitors"
+
+	// If realtime updates are active, we save via PocketBase records to trigger realtime events.
+	// Otherwise we can do a more efficient direct update via SQL
+	realtimeActive := utils.RealtimeActiveForCollection(app, monitorCollectionName, func(filterQuery string) bool {
+		return !strings.Contains(filterQuery, "system") || strings.Contains(filterQuery, systemId)
+	})
+
+	now := time.Now().UTC()
+	nowMilli := now.UnixMilli()
+	nowString := now.Format(types.DefaultDateLayout)
+	var db dbx.Builder
+	var updateQuery *dbx.Query
+	if !realtimeActive {
+		db = app.DB()
+		monitorFields := []string{"res", "resMin1h", "resMax1h", "resAvg1h", "loss1h", "updated"}
+		setClauses := make([]string, len(monitorFields))
+		for i, f := range monitorFields {
+			setClauses[i] = fmt.Sprintf("%s={:%s}", f, f)
+		}
+		queryString := fmt.Sprintf("UPDATE %s SET %s WHERE id={:id}", monitorCollectionName, strings.Join(setClauses, ", "))
+		updateQuery = db.NewQuery(queryString)
+	}
+
+	// update network_monitors records
+	for id, result := range monitorResults {
+		monitorData := map[string]any{
+			"id":       id,
+			"res":      result.AvgResponse,
+			"resAvg1h": result.AvgResponse1h,
+			"resMin1h": result.MinResponse1h,
+			"resMax1h": result.MaxResponse1h,
+			"loss1h":   result.PacketLoss1h,
+			"updated":  nowString,
+		}
+		switch realtimeActive {
+		case true:
+			var record *core.Record
+			record, err = app.FindRecordById(monitorCollectionName, id)
+			if err == nil {
+				record.Load(monitorData)
+				err = app.SaveNoValidate(record)
+			}
+		default:
+			_, err = updateQuery.Bind(dbx.Params(monitorData)).Execute()
+		}
+		if err != nil {
+			app.Logger().Warn("Failed to update monitor", "system", systemId, "monitor", id, "err", err)
+		}
+	}
+
+	// handle stats collection — one record per monitor
+	const statsCollectionName = "network_monitor_stats"
+
+	var statsCollection *core.Collection
+	if realtimeActive {
+		statsCollection, _ = app.FindCachedCollectionByNameOrId(statsCollectionName)
+	}
+
+	for monitorId, result := range monitorResults {
+		// Compare identity, not ordering, so agent clock changes don't stall writes.
+		if result.LastProbeAt == sys.lastSavedMonitorProbe[monitorId] {
+			continue
+		}
+		statsRecordData := map[string]any{
+			"system":        systemId,
+			"monitor":       monitorId,
+			"type":          "1m",
+			"created":       nowMilli,
+			"res_min":       result.MinResponse,
+			"res_max":       result.MaxResponse,
+			"total_count":   result.TotalCount,
+			"success_count": result.SuccessCount,
+			"res_sum":       result.ResponseSum,
+		}
+		switch realtimeActive {
+		case true:
+			record := core.NewRecord(statsCollection)
+			record.Load(statsRecordData)
+			err = app.SaveNoValidate(record)
+		default:
+			statsRecordData["id"] = security.PseudorandomStringWithAlphabet(10, core.DefaultIdAlphabet)
+			_, err = db.Insert(statsCollectionName, dbx.Params(statsRecordData)).Execute()
+		}
+		if err != nil {
+			app.Logger().Error("Failed to update monitor stats", "system", systemId, "monitor", monitorId, "err", err)
+		} else {
+			savedProbes[monitorId] = result.LastProbeAt
+		}
+	}
+
+	return nil
+}
+
 // createContainerRecords creates container records
 func createContainerRecords(app core.App, data []*container.Stats, systemId string) error {
 	if len(data) == 0 {
@@ -363,7 +517,7 @@ func createContainerRecords(app core.App, data []*container.Stats, systemId stri
 	valueStrings := make([]string, 0, len(data))
 	for i, container := range data {
 		suffix := fmt.Sprintf("%d", i)
-		valueStrings = append(valueStrings, fmt.Sprintf("({:id%[1]s}, {:system}, {:name%[1]s}, {:image%[1]s}, {:ports%[1]s}, {:status%[1]s}, {:health%[1]s}, {:cpu%[1]s}, {:memory%[1]s}, {:net%[1]s}, {:updated})", suffix))
+		valueStrings = append(valueStrings, fmt.Sprintf("({:id%[1]s}, {:system}, {:name%[1]s}, {:image%[1]s}, {:ports%[1]s}, {:status%[1]s}, {:health%[1]s}, {:cpu%[1]s}, {:memory%[1]s}, {:net%[1]s}, {:updateAvailable%[1]s}, {:updated})", suffix))
 		params["id"+suffix] = container.Id
 		params["name"+suffix] = container.Name
 		params["image"+suffix] = container.Image
@@ -377,9 +531,10 @@ func createContainerRecords(app core.App, data []*container.Stats, systemId stri
 			netBytes = uint64((container.NetworkSent + container.NetworkRecv) * 1024 * 1024)
 		}
 		params["net"+suffix] = netBytes
+		params["updateAvailable"+suffix] = container.UpdateAvailable
 	}
 	queryString := fmt.Sprintf(
-		"INSERT INTO containers (id, system, name, image, ports, status, health, cpu, memory, net, updated) VALUES %s ON CONFLICT(id) DO UPDATE SET system = excluded.system, name = excluded.name, image = excluded.image, ports = excluded.ports, status = excluded.status, health = excluded.health, cpu = excluded.cpu, memory = excluded.memory, net = excluded.net, updated = excluded.updated",
+		"INSERT INTO containers (id, system, name, image, ports, status, health, cpu, memory, net, updatable, updated) VALUES %s ON CONFLICT(id) DO UPDATE SET system = excluded.system, name = excluded.name, image = excluded.image, ports = excluded.ports, status = excluded.status, health = excluded.health, cpu = excluded.cpu, memory = excluded.memory, net = excluded.net, updatable = excluded.updatable, updated = excluded.updated",
 		strings.Join(valueStrings, ","),
 	)
 	_, err := app.DB().NewQuery(queryString).Bind(params).Execute()
@@ -608,7 +763,7 @@ func (sys *System) FetchZfsDataFromAgent(force bool) (*zfs.ZfsData, error) {
 	return &result, err
 }
 
-func makeStableHashId(strings ...string) string {
+func MakeStableHashId(strings ...string) string {
 	hash := fnv.New32a()
 	for _, str := range strings {
 		hash.Write([]byte(str))

internal/hub/systems/system_gpu_test.go

@@ -0,0 +1,44 @@
+//go:build testing
+
+package systems
+
+import (
+	"testing"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCreateRecordsGPUUtilization(t *testing.T) {
+	sys, app := newTestSystemWithHub(t)
+	for _, tc := range []struct {
+		name  string
+		gpu   bool
+		usage float64
+	}{
+		{"no GPU", false, 0},
+		{"active GPU", true, 42.5},
+		{"idle GPU", true, 0},
+		{"GPU removed", false, 0},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			data := &system.CombinedData{Info: system.Info{GpuPct: tc.usage, Cpu: 12.5}}
+			if tc.gpu {
+				data.Stats.GPUData = map[string]system.GPUData{"0": {Name: "GPU", Usage: tc.usage}}
+			}
+			_, err := sys.createRecords(data)
+			require.NoError(t, err)
+			record, err := app.FindRecordById("systems", sys.Id)
+			require.NoError(t, err)
+			var info map[string]any
+			require.NoError(t, record.UnmarshalJSONField("info", &info))
+			assert.Equal(t, 12.5, info["cpu"])
+			if tc.gpu {
+				assert.Equal(t, tc.usage, info["g"])
+			} else {
+				assert.NotContains(t, info, "g")
+			}
+		})
+	}
+}

internal/hub/systems/system_manager.go

@@ -4,10 +4,12 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"sync"
 	"time"
 
 	"github.com/henrygd/beszel/internal/hub/ws"
 
+	"github.com/henrygd/beszel/internal/entities/monitor"
 	"github.com/henrygd/beszel/internal/entities/system"
 	"github.com/henrygd/beszel/internal/hub/expirymap"
 
@@ -16,6 +18,7 @@ import (
 	"github.com/henrygd/beszel"
 
 	"github.com/blang/semver"
+	"github.com/pocketbase/dbx"
 	"github.com/pocketbase/pocketbase/core"
 	"github.com/pocketbase/pocketbase/tools/store"
 	"golang.org/x/crypto/ssh"
@@ -42,13 +45,17 @@ var errSystemExists = errors.New("system exists")
 // SystemManager manages a collection of monitored systems and their connections.
 // It handles system lifecycle, status updates, and maintains both SSH and WebSocket connections.
 type SystemManager struct {
-	hub           hubLike                               // Hub interface for database and alert operations
-	systems       *store.Store[string, *System]         // Thread-safe store of active systems
-	sshConfig     *ssh.ClientConfig                     // SSH client configuration for system connections
-	smartFetchMap *expirymap.ExpiryMap[smartFetchState] // Stores last SMART fetch time/result; TTL is only for cleanup
-	zfsFetchMap   *expirymap.ExpiryMap[zfsFetchState]   // Stores last ZFS fetch time/result; TTL is only for cleanup
-	ctx           context.Context                       // Cancelled when the app terminates
-	cancel        context.CancelFunc                    // Cancels ctx and all child system contexts
+	hub                 hubLike                               // Hub interface for database and alert operations
+	systems             *store.Store[string, *System]         // Thread-safe store of active systems
+	sshConfig           *ssh.ClientConfig                     // SSH client configuration for system connections
+	smartFetchMap       *expirymap.ExpiryMap[smartFetchState] // Stores last SMART fetch time/result; TTL is only for cleanup
+	zfsFetchMap         *expirymap.ExpiryMap[zfsFetchState]   // Stores last ZFS fetch time/result; TTL is only for cleanup
+	realtimeMutex       sync.Mutex                            // Protects all realtime worker and subscription state
+	activeSubscriptions map[string]*subscriptionInfo          // Realtime subscriptions keyed by system ID
+	realtimeWorkerStop  chan struct{}                         // Stops the current realtime worker generation
+	realtimeWorkerRun   bool                                  // Whether a realtime worker has been started
+	ctx                 context.Context                       // Cancelled when the app terminates
+	cancel              context.CancelFunc                    // Cancels ctx and all child system contexts
 }
 
 // hubLike defines the interface requirements for the hub dependency.
@@ -57,6 +64,7 @@ type hubLike interface {
 	core.App
 	GetSSHKey(dataDir string) (ssh.Signer, error)
 	HandleSystemAlerts(systemRecord *core.Record, data *system.CombinedData) error
+	HandleNetworkMonitorAlerts(systemRecord *core.Record, results map[string]monitor.Result) error
 	HandleStatusAlerts(status string, systemRecord *core.Record) error
 	HandleContainerAlerts(systemRecord *core.Record, data *system.CombinedData, fetchLogs func(containerID string) (string, error)) error
 	CancelPendingStatusAlerts(systemID string)
@@ -67,10 +75,11 @@ type hubLike interface {
 // The hub must implement the hubLike interface to provide database and alert functionality.
 func NewSystemManager(hub hubLike) *SystemManager {
 	sm := &SystemManager{
-		systems:       store.New(map[string]*System{}),
-		hub:           hub,
-		smartFetchMap: expirymap.New[smartFetchState](time.Hour),
-		zfsFetchMap:   expirymap.New[zfsFetchState](time.Hour),
+		systems:             store.New(map[string]*System{}),
+		hub:                 hub,
+		smartFetchMap:       expirymap.New[smartFetchState](time.Hour),
+		zfsFetchMap:         expirymap.New[zfsFetchState](time.Hour),
+		activeSubscriptions: make(map[string]*subscriptionInfo),
 	}
 	sm.ctx, sm.cancel = context.WithCancel(context.Background())
 	return sm
@@ -138,6 +147,7 @@ func (sm *SystemManager) bindEventHooks() {
 // onTerminate cancels SystemManager context on app shutdown
 func (sm *SystemManager) onTerminate(e *core.TerminateEvent) error {
 	sm.cancel()
+	sm.stopRealtimeWorker()
 	return e.Next()
 }
 
@@ -343,6 +353,20 @@ func (sm *SystemManager) AddWebSocketSystem(systemId string, agentVersion semver
 	if err := sm.AddRecord(systemRecord, system); err != nil {
 		return err
 	}
+
+	// Sync network monitors to the newly connected agent
+	go func() {
+		configs, err := sm.GetMonitorConfigsForSystem(systemId)
+		if err != nil {
+			sm.hub.Logger().Warn("failed to load monitors for agent", "system", systemId, "err", err)
+			return
+		}
+		// An empty set must also replace any probes retained across a disconnect.
+		if err := system.SyncNetworkMonitors(configs); err != nil {
+			sm.hub.Logger().Warn("failed to sync monitors to agent", "system", systemId, "err", err)
+		}
+	}()
+
 	return nil
 }
 
@@ -355,6 +379,16 @@ func (sm *SystemManager) resetFailedSmartFetchState(systemID string) {
 	}
 }
 
+// GetMonitorConfigsForSystem returns all enabled monitor configs for a system.
+func (sm *SystemManager) GetMonitorConfigsForSystem(systemID string) ([]monitor.Config, error) {
+	var configs []monitor.Config
+	err := sm.hub.DB().
+		NewQuery("SELECT id, target, protocol, port, interval FROM network_monitors WHERE system = {:system} AND enabled = true").
+		Bind(dbx.Params{"system": systemID}).
+		All(&configs)
+	return configs, err
+}
+
 // resetFailedZfsFetchState clears only failed ZFS cooldown entries so a fresh
 // agent reconnect retries ZFS discovery immediately after configuration changes.
 func (sm *SystemManager) resetFailedZfsFetchState(systemID string) {
@@ -390,11 +424,12 @@ func (sm *SystemManager) createSSHClientConfig() error {
 
 // deactivateAlerts finds all triggered alerts for a system and sets them to inactive.
 // This is called when a system is paused or goes offline to prevent continued alerts.
+// Monitor incidents remain open: a missing observation does not establish recovery.
 func deactivateAlerts(app core.App, systemID string) error {
 	// Note: Direct SQL updates don't trigger SSE, so we use the PocketBase API
 	// _, err := app.DB().NewQuery(fmt.Sprintf("UPDATE alerts SET triggered = false WHERE system = '%s'", systemID)).Execute()
 
-	alerts, err := app.FindRecordsByFilter("alerts", fmt.Sprintf("system = '%s' && triggered = 1", systemID), "", -1, 0)
+	alerts, err := app.FindRecordsByFilter("alerts", fmt.Sprintf("system = '%s' && triggered = 1 && name != 'NetworkMonitorLoss'", systemID), "", -1, 0)
 	if err != nil {
 		return err
 	}

internal/hub/systems/system_realtime.go

@@ -3,25 +3,29 @@ package systems
 import (
 	"encoding/json"
 	"strings"
-	"sync"
 	"time"
 
 	"github.com/henrygd/beszel/internal/common"
+	"github.com/henrygd/beszel/internal/entities/monitor"
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/henrygd/beszel/internal/hub/utils"
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/apis"
 	"github.com/pocketbase/pocketbase/core"
 	"github.com/pocketbase/pocketbase/tools/subscriptions"
 )
 
 type subscriptionInfo struct {
 	subscription     string
-	connectedClients uint8
+	connectedClients int
+	fetching         bool
 }
 
-var (
-	activeSubscriptions = make(map[string]*subscriptionInfo)
-	workerRunning       bool
-	tickerStopChan      chan struct{}
-	realtimeMutex       sync.Mutex
-)
+type realtimeFetch struct {
+	systemID     string
+	subscription string
+	info         *subscriptionInfo
+}
 
 // onRealtimeConnectRequest handles client connection events for realtime subscriptions.
 // It cleans up existing subscriptions when a client connects.
@@ -38,6 +42,19 @@ func (sm *SystemManager) onRealtimeConnectRequest(e *core.RealtimeConnectRequest
 // onRealtimeSubscribeRequest handles client subscription events for realtime metrics.
 // It tracks new subscriptions and unsubscriptions to manage the realtime worker lifecycle.
 func (sm *SystemManager) onRealtimeSubscribeRequest(e *core.RealtimeSubscribeRequestEvent) error {
+	// Parse with PocketBase's own subscription parser before changing the real
+	// client. Reject the entire request if any metrics target is inaccessible.
+	requested := subscriptions.NewDefaultClient()
+	requested.Subscribe(e.Subscriptions...)
+	for topic, options := range requested.Subscriptions() {
+		if !strings.HasPrefix(topic, "rt_metrics") {
+			continue
+		}
+		system, err := sm.GetSystem(options.Query["system"])
+		if err != nil || !system.HasUser(e.App, e.Auth) {
+			return e.NotFoundError("", nil)
+		}
+	}
 	oldSubs := e.Client.Subscriptions()
 	// after e.Next() is the result of the subscribe request
 	err := e.Next()
@@ -47,14 +64,7 @@ func (sm *SystemManager) onRealtimeSubscribeRequest(e *core.RealtimeSubscribeReq
 	for k, options := range newSubs {
 		if _, ok := oldSubs[k]; !ok {
 			if strings.HasPrefix(k, "rt_metrics") {
-				systemId := options.Query["system"]
-				if _, ok := activeSubscriptions[systemId]; !ok {
-					activeSubscriptions[systemId] = &subscriptionInfo{
-						subscription: k,
-					}
-				}
-				activeSubscriptions[systemId].connectedClients += 1
-				sm.onRealtimeSubscriptionAdded()
+				sm.addRealtimeSubscription(options.Query["system"], k)
 			}
 		}
 	}
@@ -68,72 +78,76 @@ func (sm *SystemManager) onRealtimeSubscribeRequest(e *core.RealtimeSubscribeReq
 	return err
 }
 
-// onRealtimeSubscriptionAdded initializes or starts the realtime worker when the first subscription is added.
-// It ensures only one worker runs at a time.
-func (sm *SystemManager) onRealtimeSubscriptionAdded() {
-	realtimeMutex.Lock()
-	defer realtimeMutex.Unlock()
+// addRealtimeSubscription tracks a subscriber and starts a worker if necessary.
+func (sm *SystemManager) addRealtimeSubscription(systemID, subscription string) {
+	sm.realtimeMutex.Lock()
+	defer sm.realtimeMutex.Unlock()
 
-	// Start the worker if it's not already running
-	if !workerRunning {
-		workerRunning = true
-		// Create a new stop channel for this worker instance
-		tickerStopChan = make(chan struct{})
-		go sm.startRealtimeWorker()
+	if sm.activeSubscriptions == nil {
+		sm.activeSubscriptions = make(map[string]*subscriptionInfo)
+	}
+	info, ok := sm.activeSubscriptions[systemID]
+	if !ok {
+		info = &subscriptionInfo{subscription: subscription}
+		sm.activeSubscriptions[systemID] = info
+	}
+	info.connectedClients++
+
+	if !sm.realtimeWorkerRun {
+		sm.realtimeWorkerRun = true
+		stop := make(chan struct{})
+		sm.realtimeWorkerStop = stop
+		go sm.startRealtimeWorker(stop)
 	}
 }
 
-// checkSubscriptions stops the realtime worker when there are no active subscriptions.
-// This prevents unnecessary resource usage when no clients are listening for realtime data.
-func (sm *SystemManager) checkSubscriptions() {
-	if !workerRunning || len(activeSubscriptions) > 0 {
+// stopRealtimeWorker stops the current worker generation, if any.
+func (sm *SystemManager) stopRealtimeWorker() {
+	sm.realtimeMutex.Lock()
+	defer sm.realtimeMutex.Unlock()
+	sm.stopRealtimeWorkerLocked()
+}
+
+func (sm *SystemManager) stopRealtimeWorkerLocked() {
+	if !sm.realtimeWorkerRun {
 		return
 	}
-
-	realtimeMutex.Lock()
-	defer realtimeMutex.Unlock()
-
-	// Signal the worker to stop
-	if tickerStopChan != nil {
-		select {
-		case tickerStopChan <- struct{}{}:
-		default:
-		}
-	}
-
-	// Mark worker as stopped (will be reset when next subscription comes in)
-	workerRunning = false
+	close(sm.realtimeWorkerStop)
+	sm.realtimeWorkerStop = nil
+	sm.realtimeWorkerRun = false
 }
 
 // removeRealtimeSubscription removes a realtime subscription and checks if the worker should be stopped.
 // It only processes subscriptions with the "rt_metrics" prefix and triggers cleanup when subscriptions are removed.
 func (sm *SystemManager) removeRealtimeSubscription(subscription string, options subscriptions.SubscriptionOptions) {
 	if strings.HasPrefix(subscription, "rt_metrics") {
-		systemId := options.Query["system"]
-		if info, ok := activeSubscriptions[systemId]; ok {
-			info.connectedClients -= 1
+		systemID := options.Query["system"]
+		sm.realtimeMutex.Lock()
+		if info, ok := sm.activeSubscriptions[systemID]; ok {
+			info.connectedClients--
 			if info.connectedClients <= 0 {
-				delete(activeSubscriptions, systemId)
+				delete(sm.activeSubscriptions, systemID)
 			}
 		}
-		sm.checkSubscriptions()
+		if len(sm.activeSubscriptions) == 0 {
+			sm.stopRealtimeWorkerLocked()
+		}
+		sm.realtimeMutex.Unlock()
 	}
 }
 
 // startRealtimeWorker runs the main loop for fetching realtime data from agents.
 // It continuously fetches system data and broadcasts it to subscribed clients via WebSocket.
-func (sm *SystemManager) startRealtimeWorker() {
+func (sm *SystemManager) startRealtimeWorker(stop <-chan struct{}) {
 	sm.fetchRealtimeDataAndNotify()
-	tick := time.Tick(1 * time.Second)
+	ticker := time.NewTicker(time.Second)
+	defer ticker.Stop()
 
 	for {
 		select {
-		case <-tickerStopChan:
+		case <-stop:
 			return
-		case <-tick:
-			if len(activeSubscriptions) == 0 {
-				return
-			}
+		case <-ticker.C:
 			sm.fetchRealtimeDataAndNotify()
 		}
 	}
@@ -141,27 +155,95 @@ func (sm *SystemManager) startRealtimeWorker() {
 
 // fetchRealtimeDataAndNotify fetches realtime data for all active subscriptions and notifies the clients.
 func (sm *SystemManager) fetchRealtimeDataAndNotify() {
-	for systemId, info := range activeSubscriptions {
-		system, err := sm.GetSystem(systemId)
+	for _, fetch := range sm.claimRealtimeFetches() {
+		system, err := sm.GetSystem(fetch.systemID)
 		if err != nil {
+			sm.finishRealtimeFetch(fetch)
 			continue
 		}
-		go func() {
+		go func(fetch realtimeFetch) {
+			defer sm.finishRealtimeFetch(fetch)
 			data, err := system.fetchDataFromAgent(common.DataRequestOptions{CacheTimeMs: 1000})
 			if err != nil {
 				return
 			}
-			bytes, err := json.Marshal(data)
+			bytes, err := marshalRealtimeData(data)
 			if err == nil {
-				notify(sm.hub, info.subscription, bytes)
+				notify(sm.hub, system, fetch.subscription, bytes)
 			}
-		}()
+		}(fetch)
 	}
 }
 
+// claimRealtimeFetches takes a stable snapshot and marks each selected system as
+// in flight. Slow agents are skipped on later ticks until their fetch completes.
+func (sm *SystemManager) claimRealtimeFetches() []realtimeFetch {
+	sm.realtimeMutex.Lock()
+	defer sm.realtimeMutex.Unlock()
+
+	fetches := make([]realtimeFetch, 0, len(sm.activeSubscriptions))
+	for systemID, info := range sm.activeSubscriptions {
+		if info.fetching {
+			continue
+		}
+		info.fetching = true
+		fetches = append(fetches, realtimeFetch{
+			systemID:     systemID,
+			subscription: info.subscription,
+			info:         info,
+		})
+	}
+	return fetches
+}
+
+func (sm *SystemManager) finishRealtimeFetch(fetch realtimeFetch) {
+	sm.realtimeMutex.Lock()
+	defer sm.realtimeMutex.Unlock()
+	// A subscription may have been removed and recreated while the old request
+	// was running. Only release the exact entry claimed by this request.
+	if info := sm.activeSubscriptions[fetch.systemID]; info == fetch.info {
+		info.fetching = false
+	}
+}
+
+// marshalRealtimeData marshals combined agent data for a realtime broadcast, converting
+// the per-monitor results into the derived metric fields the frontend charts expect.
+func marshalRealtimeData(data *system.CombinedData) ([]byte, error) {
+	if len(data.Monitors) == 0 {
+		return json.Marshal(data)
+	}
+	monitorStats := make(map[string]monitor.Stats, len(data.Monitors))
+	for id, result := range data.Monitors {
+		monitorStats[id] = monitor.Stats{}.FromResult(result)
+	}
+	return json.Marshal(struct {
+		*system.CombinedData
+		Monitors map[string]monitor.Stats `json:"Monitors"`
+	}{data, monitorStats})
+}
+
 // notify broadcasts realtime data to all clients subscribed to a specific subscription.
-// It iterates through all connected clients and sends the data only to those with matching subscriptions.
-func notify(app core.App, subscription string, data []byte) error {
+// Custom topics bypass collection rules, so check current access for every
+// recipient, including clients whose authentication or membership was revoked.
+func notify(app core.App, system *System, subscription string, data []byte) error {
+	shareAll, _ := utils.GetEnv("SHARE_ALL_SYSTEMS")
+	members := make(map[string]struct{})
+	if shareAll != "true" {
+		// Refresh once per broadcast so membership changes take effect on the
+		// next update without querying the database for every recipient.
+		var recordData struct{ Users string }
+		if err := app.DB().NewQuery("SELECT users FROM systems WHERE id={:id}").
+			Bind(dbx.Params{"id": system.Id}).One(&recordData); err != nil {
+			return err
+		}
+		var userIDs []string
+		if err := json.Unmarshal([]byte(recordData.Users), &userIDs); err != nil {
+			return err
+		}
+		for _, id := range userIDs {
+			members[id] = struct{}{}
+		}
+	}
 	message := subscriptions.Message{
 		Name: subscription,
 		Data: data,
@@ -170,6 +252,13 @@ func notify(app core.App, subscription string, data []byte) error {
 		if !client.HasSubscription(subscription) {
 			continue
 		}
+		auth, _ := client.Get(apis.RealtimeClientAuthKey).(*core.Record)
+		if auth == nil {
+			continue
+		}
+		if _, member := members[auth.Id]; shareAll != "true" && !member {
+			continue
+		}
 		client.Send(message)
 	}
 	return nil

internal/hub/systems/system_realtime_test.go

@@ -0,0 +1,229 @@
+package systems
+
+import (
+	"testing"
+	"time"
+
+	"github.com/pocketbase/pocketbase/apis"
+	"github.com/pocketbase/pocketbase/core"
+	pbtests "github.com/pocketbase/pocketbase/tests"
+	"github.com/pocketbase/pocketbase/tools/hook"
+	"github.com/pocketbase/pocketbase/tools/store"
+	"github.com/pocketbase/pocketbase/tools/subscriptions"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestRealtimeAuthorization(t *testing.T) {
+	t.Setenv("SHARE_ALL_SYSTEMS", "false")
+	t.Setenv("BESZEL_HUB_SHARE_ALL_SYSTEMS", "")
+	app, err := pbtests.NewTestApp(t.TempDir())
+	require.NoError(t, err)
+	t.Cleanup(app.Cleanup)
+	_, err = app.DB().NewQuery(`CREATE TABLE IF NOT EXISTS systems (id TEXT PRIMARY KEY, users TEXT)`).Execute()
+	require.NoError(t, err)
+	_, err = app.DB().NewQuery(`INSERT INTO systems (id, users) VALUES ('target', '["member"]')`).Execute()
+	require.NoError(t, err)
+	member := core.NewRecord(core.NewAuthCollection("users"))
+	member.Id = "member"
+	outsider := core.NewRecord(member.Collection())
+	outsider.Id = "outsider"
+	system := &System{Id: "target"}
+	sm := newRealtimeTestManager()
+	sm.systems.Set(system.Id, system)
+	// Keep the lifecycle bookkeeping active without starting an agent worker.
+	sm.realtimeWorkerRun = true
+	sm.realtimeWorkerStop = make(chan struct{})
+	t.Cleanup(sm.stopRealtimeWorker)
+	topic := `rt_metrics?options={"query":{"system":"target"}}`
+
+	for _, tc := range []struct {
+		name    string
+		auth    *core.Record
+		topic   string
+		share   bool
+		allowed bool
+	}{
+		{"guest", nil, topic, false, false},
+		{"outsider", outsider, topic, false, false},
+		{"member", member, topic, false, true},
+		{"missing system", member, `rt_metrics`, false, false},
+		{"unknown system", member, `rt_metrics?options={"query":{"system":"missing"}}`, false, false},
+		{"malformed options", member, `rt_metrics?options=invalid`, false, false},
+		{"prefix variant", outsider, `rt_metrics_extra?options={"query":{"system":"target"}}`, false, false},
+		{"shared outsider", outsider, topic, true, true},
+		{"shared guest", nil, topic, true, false},
+		{"other topic", nil, "systems/*", false, true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			if tc.share {
+				t.Setenv("BESZEL_HUB_SHARE_ALL_SYSTEMS", "true")
+			}
+			client := subscriptions.NewDefaultClient()
+			client.Subscribe("existing")
+			e := &core.RealtimeSubscribeRequestEvent{
+				RequestEvent: &core.RequestEvent{App: app, Auth: tc.auth},
+				Client:       client, Subscriptions: []string{tc.topic},
+			}
+			called := false
+			h := &hook.Hook[*core.RealtimeSubscribeRequestEvent]{}
+			h.BindFunc(sm.onRealtimeSubscribeRequest)
+			err := h.Trigger(e, func(e *core.RealtimeSubscribeRequestEvent) error {
+				called = true
+				client.Unsubscribe()
+				client.Subscribe(e.Subscriptions...)
+				return nil
+			})
+			if tc.allowed {
+				require.NoError(t, err)
+				assert.True(t, called)
+			} else {
+				require.Error(t, err)
+				assert.False(t, called)
+				assert.True(t, client.HasSubscription("existing"))
+				assert.False(t, client.HasSubscription(tc.topic))
+			}
+		})
+	}
+
+	t.Run("broadcast checks current access", func(t *testing.T) {
+		client := subscriptions.NewDefaultClient()
+		client.Subscribe(topic)
+		app.SubscriptionsBroker().Register(client)
+		defer app.SubscriptionsBroker().Unregister(client.Id())
+		secondClient := subscriptions.NewDefaultClient()
+		secondClient.Subscribe(topic)
+		app.SubscriptionsBroker().Register(secondClient)
+		defer app.SubscriptionsBroker().Unregister(secondClient.Id())
+		check := func(auth *core.Record, allowed bool) {
+			t.Helper()
+			client.Set(apis.RealtimeClientAuthKey, auth)
+			secondClient.Set(apis.RealtimeClientAuthKey, auth)
+			done := make(chan struct{})
+			go func() {
+				notify(app, system, topic, []byte(`{"cpu":42}`))
+				close(done)
+			}()
+			// Even on failure, drain pending sends and join the broadcaster before
+			// unregistering clients, which closes their channels.
+			defer func() {
+				for {
+					select {
+					case <-client.Channel():
+					case <-secondClient.Channel():
+					case <-done:
+						return
+					}
+				}
+			}()
+			var received [2]int
+			timer := time.NewTimer(time.Second)
+			defer timer.Stop()
+			for {
+				select {
+				case msg := <-client.Channel():
+					received[0]++
+					assert.Equal(t, topic, msg.Name)
+				case msg := <-secondClient.Channel():
+					received[1]++
+					assert.Equal(t, topic, msg.Name)
+				case <-done:
+					want := [2]int{}
+					if allowed {
+						want = [2]int{1, 1}
+					}
+					assert.Equal(t, want, received)
+					return
+				case <-timer.C:
+					t.Fatal("broadcast did not finish")
+				}
+			}
+		}
+		check(nil, false)
+		check(outsider, false)
+		check(member, true)
+		_, err := app.DB().NewQuery(`UPDATE systems SET users = '[]'`).Execute()
+		require.NoError(t, err)
+		check(member, false)
+		t.Setenv("BESZEL_HUB_SHARE_ALL_SYSTEMS", "true")
+		check(outsider, true)
+		check(nil, false)
+	})
+}
+
+func newRealtimeTestManager() *SystemManager {
+	return &SystemManager{
+		systems:             store.New(map[string]*System{}),
+		activeSubscriptions: make(map[string]*subscriptionInfo),
+	}
+}
+
+func TestRealtimeFetchesDoNotOverlapPerSystem(t *testing.T) {
+	sm := newRealtimeTestManager()
+	sm.activeSubscriptions["one"] = &subscriptionInfo{subscription: "rt_metrics_one"}
+	sm.activeSubscriptions["two"] = &subscriptionInfo{subscription: "rt_metrics_two"}
+
+	first := sm.claimRealtimeFetches()
+	require.Len(t, first, 2)
+	assert.Empty(t, sm.claimRealtimeFetches())
+
+	sm.finishRealtimeFetch(first[0])
+	next := sm.claimRealtimeFetches()
+	require.Len(t, next, 1)
+	assert.Equal(t, first[0].systemID, next[0].systemID)
+
+	sm.finishRealtimeFetch(first[1])
+	sm.finishRealtimeFetch(next[0])
+}
+
+func TestFinishingOldRealtimeFetchDoesNotReleaseReplacement(t *testing.T) {
+	sm := newRealtimeTestManager()
+	oldInfo := &subscriptionInfo{subscription: "old"}
+	sm.activeSubscriptions["system"] = oldInfo
+
+	fetch := sm.claimRealtimeFetches()[0]
+	newInfo := &subscriptionInfo{subscription: "new", fetching: true}
+	sm.activeSubscriptions["system"] = newInfo
+
+	sm.finishRealtimeFetch(fetch)
+	assert.True(t, newInfo.fetching)
+}
+
+func TestRealtimeSubscriptionLifecycle(t *testing.T) {
+	sm := newRealtimeTestManager()
+	options := subscriptions.SubscriptionOptions{Query: map[string]string{"system": "system"}}
+
+	sm.addRealtimeSubscription("system", "rt_metrics")
+	sm.addRealtimeSubscription("system", "rt_metrics")
+
+	sm.realtimeMutex.Lock()
+	firstStop := sm.realtimeWorkerStop
+	assert.True(t, sm.realtimeWorkerRun)
+	assert.Equal(t, 2, sm.activeSubscriptions["system"].connectedClients)
+	sm.realtimeMutex.Unlock()
+
+	sm.removeRealtimeSubscription("rt_metrics", options)
+	sm.realtimeMutex.Lock()
+	assert.True(t, sm.realtimeWorkerRun)
+	assert.Equal(t, 1, sm.activeSubscriptions["system"].connectedClients)
+	sm.realtimeMutex.Unlock()
+
+	sm.removeRealtimeSubscription("rt_metrics", options)
+	sm.realtimeMutex.Lock()
+	assert.False(t, sm.realtimeWorkerRun)
+	assert.Empty(t, sm.activeSubscriptions)
+	sm.realtimeMutex.Unlock()
+	select {
+	case <-firstStop:
+	default:
+		t.Fatal("worker stop channel was not closed")
+	}
+
+	// A later subscription must get a new stop channel owned by its worker.
+	sm.addRealtimeSubscription("system", "rt_metrics")
+	sm.realtimeMutex.Lock()
+	secondStop := sm.realtimeWorkerStop
+	assert.NotEqual(t, firstStop, secondStop)
+	sm.realtimeMutex.Unlock()
+	sm.stopRealtimeWorker()
+}

internal/hub/systems/system_smart.go

@@ -77,7 +77,7 @@ func (sys *System) saveSmartDevices(smartData map[string]smart.SmartData, comple
 
 	currentIDs := make(map[string]struct{}, len(smartData))
 	for deviceKey := range smartData {
-		currentIDs[makeStableHashId(sys.Id, deviceKey)] = struct{}{}
+		currentIDs[MakeStableHashId(sys.Id, deviceKey)] = struct{}{}
 	}
 
 	err = hub.RunInTransaction(func(txApp core.App) error {
@@ -115,7 +115,7 @@ func (sys *System) saveSmartDevices(smartData map[string]smart.SmartData, comple
 }
 
 func (sys *System) upsertSmartDeviceRecord(app core.App, collection *core.Collection, deviceKey string, device smart.SmartData) error {
-	recordID := makeStableHashId(sys.Id, deviceKey)
+	recordID := MakeStableHashId(sys.Id, deviceKey)
 
 	record, err := app.FindRecordById(collection, recordID)
 	if err != nil {

internal/hub/systems/system_smart_test.go

@@ -7,6 +7,7 @@ import (
 	"testing"
 	"time"
 
+	"github.com/henrygd/beszel/internal/entities/monitor"
 	"github.com/henrygd/beszel/internal/entities/smart"
 	esystem "github.com/henrygd/beszel/internal/entities/system"
 	"github.com/henrygd/beszel/internal/hub/expirymap"
@@ -28,7 +29,8 @@ func (stubHub) GetSSHKey(dataDir string) (ssh.Signer, error) { return nil, nil }
 func (stubHub) HandleSystemAlerts(systemRecord *core.Record, data *esystem.CombinedData) error {
 	return nil
 }
-func (stubHub) HandleStatusAlerts(status string, systemRecord *core.Record) error { return nil }
+func (stubHub) HandleNetworkMonitorAlerts(*core.Record, map[string]monitor.Result) error { return nil }
+func (stubHub) HandleStatusAlerts(status string, systemRecord *core.Record) error        { return nil }
 func (stubHub) HandleContainerAlerts(systemRecord *core.Record, data *esystem.CombinedData, fetchLogs func(containerID string) (string, error)) error {
 	return nil
 }
@@ -212,7 +214,7 @@ func TestSaveSmartDevices_IncompleteDataDoesNotRemoveDevices(t *testing.T) {
 	}, false))
 
 	assert.Len(t, countSmartDeviceRecords(t, testApp, sys.Id), 2)
-	recordA, err := testApp.FindRecordById("smart_devices", makeStableHashId(sys.Id, "AAA"))
+	recordA, err := testApp.FindRecordById("smart_devices", MakeStableHashId(sys.Id, "AAA"))
 	require.NoError(t, err)
 	assert.EqualValues(t, 42, recordA.GetInt("temp"))
 }

internal/hub/systems/system_systemd_test.go

@@ -14,9 +14,9 @@ func TestGetSystemdServiceId(t *testing.T) {
 		serviceName := "nginx.service"
 
 		// Call multiple times and ensure same result
-		id1 := makeStableHashId(systemId, serviceName)
-		id2 := makeStableHashId(systemId, serviceName)
-		id3 := makeStableHashId(systemId, serviceName)
+		id1 := MakeStableHashId(systemId, serviceName)
+		id2 := MakeStableHashId(systemId, serviceName)
+		id3 := MakeStableHashId(systemId, serviceName)
 
 		assert.Equal(t, id1, id2)
 		assert.Equal(t, id2, id3)
@@ -29,10 +29,10 @@ func TestGetSystemdServiceId(t *testing.T) {
 		serviceName1 := "nginx.service"
 		serviceName2 := "apache.service"
 
-		id1 := makeStableHashId(systemId1, serviceName1)
-		id2 := makeStableHashId(systemId2, serviceName1)
-		id3 := makeStableHashId(systemId1, serviceName2)
-		id4 := makeStableHashId(systemId2, serviceName2)
+		id1 := MakeStableHashId(systemId1, serviceName1)
+		id2 := MakeStableHashId(systemId2, serviceName1)
+		id3 := MakeStableHashId(systemId1, serviceName2)
+		id4 := MakeStableHashId(systemId2, serviceName2)
 
 		// All IDs should be different
 		assert.NotEqual(t, id1, id2)
@@ -56,14 +56,14 @@ func TestGetSystemdServiceId(t *testing.T) {
 		}
 
 		for _, tc := range testCases {
-			id := makeStableHashId(tc.systemId, tc.serviceName)
+			id := MakeStableHashId(tc.systemId, tc.serviceName)
 			// FNV-32 produces 8 hex characters
 			assert.Len(t, id, 8, "ID should be 8 characters for systemId='%s', serviceName='%s'", tc.systemId, tc.serviceName)
 		}
 	})
 
 	t.Run("hexadecimal output", func(t *testing.T) {
-		id := makeStableHashId("test-system", "test-service")
+		id := MakeStableHashId("test-system", "test-service")
 		assert.NotEmpty(t, id)
 
 		// Should only contain hexadecimal characters

internal/hub/systems/system_zfs.go

@@ -32,13 +32,12 @@ func (sys *System) FetchAndSaveZfsPools(force bool) error {
 		sys.recordZfsFetchResult(err, 0)
 		return err
 	}
-	if zfsData == nil || !zfsData.Complete {
-		err = errIncompleteZfsData
-		sys.recordZfsFetchResult(err, 0)
-		return err
-	}
 	err = sys.saveZfsPools(zfsData)
-	sys.recordZfsFetchResult(err, len(zfsData.Pools))
+	poolCount := 0
+	if zfsData != nil {
+		poolCount = len(zfsData.Pools)
+	}
+	sys.recordZfsFetchResult(err, poolCount)
 	return err
 }
 
@@ -79,7 +78,7 @@ func (sys *System) zfsFetchInterval() time.Duration {
 // saveZfsPools saves ZFS pool detail data to the zfs_pools collection and
 // removes records for pools no longer reported by a complete agent inventory.
 func (sys *System) saveZfsPools(zfsData *zfs.ZfsData) error {
-	if zfsData == nil || !zfsData.Complete {
+	if zfsData == nil || (!zfsData.CanRefreshPool("zfs") && !zfsData.CanRefreshPool("b:")) {
 		return errIncompleteZfsData
 	}
 
@@ -89,10 +88,10 @@ func (sys *System) saveZfsPools(zfsData *zfs.ZfsData) error {
 		return err
 	}
 
-	return hub.RunInTransaction(func(txApp core.App) error {
+	err = hub.RunInTransaction(func(txApp core.App) error {
 		alive := make(map[string]bool, len(zfsData.Pools))
 		for _, pool := range zfsData.Pools {
-			if pool == nil {
+			if pool == nil || !zfsData.CanRefreshPool(pool.Name) {
 				continue
 			}
 			alive[pool.Name] = true
@@ -111,7 +110,7 @@ func (sys *System) saveZfsPools(zfsData *zfs.ZfsData) error {
 			return err
 		}
 		for _, record := range existing {
-			if !alive[record.GetString("name")] {
+			if name := record.GetString("name"); zfsData.CanRefreshPool(name) && !alive[name] {
 				if err := txApp.Delete(record); err != nil {
 					return err
 				}
@@ -119,10 +118,18 @@ func (sys *System) saveZfsPools(zfsData *zfs.ZfsData) error {
 		}
 		return nil
 	})
+	if err != nil {
+		return err
+	}
+	// Report partial failure only after committing healthy backend updates.
+	if !zfsData.Complete {
+		return errIncompleteZfsData
+	}
+	return nil
 }
 
 func (sys *System) upsertZfsPoolRecord(app core.App, collection *core.Collection, pool *zfs.PoolDetail) error {
-	recordID := makeStableHashId(sys.Id, pool.Name)
+	recordID := MakeStableHashId(sys.Id, pool.Name)
 
 	record, err := app.FindRecordById(collection, recordID)
 	if err != nil {
@@ -135,10 +142,12 @@ func (sys *System) upsertZfsPoolRecord(app core.App, collection *core.Collection
 
 	record.Set("system", sys.Id)
 	record.Set("name", pool.Name)
+	record.Set("display_name", pool.DisplayName)
 	record.Set("health", pool.Health)
 	record.Set("size", pool.Size)
 	record.Set("alloc", pool.Alloc)
 	record.Set("free", pool.Free)
+	record.Set("raw", pool.Raw)
 	record.Set("scrub", pool.Scrub)
 	record.Set("vdevs", pool.Vdevs)
 	record.Set("datasets", pool.Datasets)
@@ -162,7 +171,7 @@ func (sys *System) syncZfsPoolHealth(app core.App, pools map[string]*system.ZfsP
 		if pool == nil {
 			continue
 		}
-		recordID := makeStableHashId(sys.Id, name)
+		recordID := MakeStableHashId(sys.Id, name)
 		record, err := app.FindRecordById(collection, recordID)
 		if err != nil {
 			if !errors.Is(err, sql.ErrNoRows) {
@@ -172,7 +181,9 @@ func (sys *System) syncZfsPoolHealth(app core.App, pools map[string]*system.ZfsP
 			record.Set("id", recordID)
 			record.Set("system", sys.Id)
 			record.Set("name", name)
+			record.Set("display_name", pool.DisplayName)
 			record.Set("health", pool.Health)
+			record.Set("raw", pool.Raw)
 			record.Set("size", uint64(pool.Total*gib))
 			record.Set("alloc", uint64(pool.Used*gib))
 			record.Set("free", uint64(max(pool.Total-pool.Used, 0)*gib))
@@ -181,10 +192,15 @@ func (sys *System) syncZfsPoolHealth(app core.App, pools map[string]*system.ZfsP
 			}
 			continue
 		}
-		if record.GetString("health") == pool.Health {
+		if record.GetString("health") == pool.Health && record.GetBool("raw") == pool.Raw && record.GetString("display_name") == pool.DisplayName {
 			continue
 		}
+		record.Set("display_name", pool.DisplayName)
 		record.Set("health", pool.Health)
+		record.Set("raw", pool.Raw)
+		record.Set("size", uint64(pool.Total*gib))
+		record.Set("alloc", uint64(pool.Used*gib))
+		record.Set("free", uint64(max(pool.Total-pool.Used, 0)*gib))
 		if err := app.SaveNoValidate(record); err != nil {
 			return fmt.Errorf("updating ZFS pool health %q: %w", name, err)
 		}

internal/hub/systems/system_zfs_test.go

@@ -123,6 +123,44 @@ func TestSaveZfsPoolsIncompletePreservesRecords(t *testing.T) {
 	assert.Len(t, records, 1)
 }
 
+func TestSavePartialBackendInventory(t *testing.T) {
+	for _, healthy := range []string{"zfs", "btrfs"} {
+		t.Run(healthy, func(t *testing.T) {
+			sys, app := newTestSystemWithHub(t)
+			healthyKey, failedKey := "tank", "b:uuid"
+			if healthy == "btrfs" {
+				healthyKey, failedKey = failedKey, healthyKey
+			}
+			initial := &zfs.ZfsData{Complete: true, Pools: []*zfs.PoolDetail{
+				{Name: healthyKey, Alloc: 10}, {Name: failedKey, Alloc: 10},
+			}}
+			require.NoError(t, sys.saveZfsPools(initial))
+			failedID := MakeStableHashId(sys.Id, failedKey)
+			before, err := app.FindRecordById("zfs_pools", failedID)
+			require.NoError(t, err)
+			partial := &zfs.ZfsData{CompleteBackends: []string{healthy}, Pools: []*zfs.PoolDetail{
+				{Name: healthyKey, Alloc: 20}, {Name: failedKey, Alloc: 99},
+			}}
+			assert.ErrorIs(t, sys.saveZfsPools(partial), errIncompleteZfsData)
+			fresh, err := app.FindRecordById("zfs_pools", MakeStableHashId(sys.Id, healthyKey))
+			require.NoError(t, err)
+			assert.EqualValues(t, 20, fresh.GetInt("alloc"))
+			cached, err := app.FindRecordById("zfs_pools", failedID)
+			require.NoError(t, err)
+			assert.EqualValues(t, 10, cached.GetInt("alloc"))
+			assert.Equal(t, before.GetDateTime("details_updated"), cached.GetDateTime("details_updated"))
+			// An empty successful backend can prune, even while the other fails.
+			partial.Pools = nil
+			assert.ErrorIs(t, sys.saveZfsPools(partial), errIncompleteZfsData)
+			records, err := app.FindRecordsByFilter("zfs_pools", "system={:system}", "", 0, 0, map[string]any{"system": sys.Id})
+			require.NoError(t, err)
+			require.Len(t, records, 1)
+			assert.Equal(t, failedKey, records[0].GetString("name"))
+			require.NoError(t, sys.saveZfsPools(&zfs.ZfsData{Complete: true}))
+		})
+	}
+}
+
 func TestSyncZfsPoolHealthWritesOnlyTransitions(t *testing.T) {
 	sys, app := newTestSystemWithHub(t)
 	collection, err := app.FindCachedCollectionByNameOrId("zfs_pools")
@@ -131,7 +169,7 @@ func TestSyncZfsPoolHealthWritesOnlyTransitions(t *testing.T) {
 	require.NoError(t, sys.syncZfsPoolHealth(app, map[string]*system.ZfsPool{
 		"tank": {Total: 100, Used: 25, Health: "ONLINE"},
 	}))
-	record, err := app.FindRecordById(collection, makeStableHashId(sys.Id, "tank"))
+	record, err := app.FindRecordById(collection, MakeStableHashId(sys.Id, "tank"))
 	require.NoError(t, err)
 	firstUpdated := record.GetDateTime("updated")
 	assert.Equal(t, "ONLINE", record.GetString("health"))
@@ -151,3 +189,42 @@ func TestSyncZfsPoolHealthWritesOnlyTransitions(t *testing.T) {
 	require.NoError(t, err)
 	assert.Equal(t, "DEGRADED", record.GetString("health"))
 }
+
+func TestZfsRawCapacityPersistence(t *testing.T) {
+	sys, app := newTestSystemWithHub(t)
+	require.NoError(t, sys.saveZfsPools(&zfs.ZfsData{Complete: true, Pools: []*zfs.PoolDetail{{Name: "btrfs", Size: 200, Alloc: 10, Raw: true}}}))
+	record, err := app.FindRecordById("zfs_pools", MakeStableHashId(sys.Id, "btrfs"))
+	require.NoError(t, err)
+	require.True(t, record.GetBool("raw"))
+	require.NoError(t, sys.syncZfsPoolHealth(app, map[string]*system.ZfsPool{"btrfs": {Total: 1, Used: 0.25}}))
+	record, err = app.FindRecordById("zfs_pools", record.Id)
+	require.NoError(t, err)
+	assert.False(t, record.GetBool("raw"))
+	assert.EqualValues(t, 1024*1024*1024, record.GetInt("size"))
+}
+
+func TestBtrfsDisplayNameKeepsRecordIdentity(t *testing.T) {
+	sys, app := newTestSystemWithHub(t)
+	key := "b:11111111-1111-4111-8111-111111111111"
+	require.NoError(t, sys.syncZfsPoolHealth(app, map[string]*system.ZfsPool{
+		key:    {DisplayName: "tank", Health: "ONLINE"},
+		"tank": {Health: "ONLINE"},
+	}))
+	id := MakeStableHashId(sys.Id, key)
+	record, err := app.FindRecordById("zfs_pools", id)
+	require.NoError(t, err)
+	assert.Equal(t, "tank", record.GetString("display_name"))
+	require.NoError(t, sys.syncZfsPoolHealth(app, map[string]*system.ZfsPool{key: {DisplayName: "renamed", Health: "ONLINE"}}))
+	record, err = app.FindRecordById("zfs_pools", id)
+	require.NoError(t, err)
+	assert.Equal(t, key, record.GetString("name"))
+	assert.Equal(t, "renamed", record.GetString("display_name"))
+	require.NoError(t, sys.saveZfsPools(&zfs.ZfsData{Complete: true, Pools: []*zfs.PoolDetail{
+		{Name: key, DisplayName: "detail name", Health: "ONLINE"}, {Name: "tank", Health: "ONLINE"},
+	}}))
+	record, err = app.FindRecordById("zfs_pools", id)
+	require.NoError(t, err)
+	assert.Equal(t, "detail name", record.GetString("display_name"))
+	_, err = app.FindRecordById("zfs_pools", MakeStableHashId(sys.Id, "tank"))
+	require.NoError(t, err)
+}

internal/hub/systems/systemd_records_test.go

@@ -3,9 +3,11 @@
 package systems_test
 
 import (
+	"encoding/json"
 	"testing"
 	"time"
 
+	"github.com/fxamacker/cbor/v2"
 	"github.com/henrygd/beszel/internal/entities/system"
 	"github.com/henrygd/beszel/internal/entities/systemd"
 	"github.com/henrygd/beszel/internal/hub/systems"
@@ -15,6 +17,42 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func TestCreateRecordsRejectsNullSystemdService(t *testing.T) {
+	hub, user := tests.GetHubWithUser(t)
+	defer hub.Cleanup()
+	records, err := tests.CreateSystems(hub, 1, user.Id, "paused")
+	require.NoError(t, err)
+	sys, err := hub.GetSystemManager().GetSystem(records[0].Id)
+	require.NoError(t, err)
+	require.NoError(t, systems.CreateSystemdStatsRecords(hub, []*systemd.Service{
+		{Name: "existing.service", State: systemd.StatusFailed},
+	}, records[0].Id))
+
+	for _, services := range []string{`[null]`, `[{"name":"new.service"},null]`, `[null,{"name":"new.service"}]`} {
+		for _, encoding := range []string{"json", "cbor"} {
+			t.Run(encoding+"/"+services, func(t *testing.T) {
+				var data system.CombinedData
+				require.NoError(t, json.Unmarshal([]byte(`{"systemd":`+services+`}`), &data))
+				if encoding == "cbor" {
+					encoded, err := cbor.Marshal(data)
+					require.NoError(t, err)
+					data = system.CombinedData{}
+					require.NoError(t, cbor.Unmarshal(encoded, &data))
+				}
+				_, err := sys.CreateRecords(&data)
+				require.ErrorContains(t, err, "null systemd service")
+				var names []string
+				require.NoError(t, hub.DB().Select("name").From("systemd_services").
+					Where(dbx.HashExp{"system": records[0].Id}).Column(&names))
+				assert.Equal(t, []string{"existing.service"}, names)
+				count, err := hub.CountRecords("system_stats", dbx.HashExp{"system": records[0].Id})
+				require.NoError(t, err)
+				assert.Zero(t, count, "invalid snapshot must roll back system stats")
+			})
+		}
+	}
+}
+
 func TestCreateRecordsHandlesSystemdAlertLifecycle(t *testing.T) {
 	hub, user := tests.GetHubWithUser(t)
 	defer hub.Cleanup()

internal/hub/trusted_proxy_test.go

@@ -0,0 +1,127 @@
+//go:build testing
+
+package hub
+
+import (
+	"net/netip"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestParseTrustedProxies(t *testing.T) {
+	testCases := []struct {
+		name       string
+		value      string
+		prefixes   []string
+		restricted bool
+	}{
+		{
+			name:       "empty",
+			value:      "",
+			restricted: false,
+		},
+		{
+			name:       "blank",
+			value:      "  , ",
+			prefixes:   nil,
+			restricted: true,
+		},
+		{
+			name:       "single addresses become host prefixes",
+			value:      "10.0.0.5, 2001:db8::1",
+			prefixes:   []string{"10.0.0.5/32", "2001:db8::1/128"},
+			restricted: true,
+		},
+		{
+			name:       "cidrs are masked",
+			value:      "172.16.5.9/12,fd00::1/64",
+			prefixes:   []string{"172.16.0.0/12", "fd00::/64"},
+			restricted: true,
+		},
+		{
+			name:       "ipv4-mapped entries become ipv4",
+			value:      "::ffff:10.0.0.5, ::ffff:10.0.0.0/104",
+			prefixes:   []string{"10.0.0.5/32", "10.0.0.0/8"},
+			restricted: true,
+		},
+		{
+			name:       "invalid entries are skipped, valid ones kept",
+			value:      "proxy.internal, 10.0.0.0/8, 300.1.1.1, ::ffff:0.0.0.0/64",
+			prefixes:   []string{"10.0.0.0/8"},
+			restricted: true,
+		},
+		{
+			name:       "only invalid entries trust nobody",
+			value:      "proxy.internal",
+			prefixes:   nil,
+			restricted: true,
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv("TRUSTED_PROXY_IPS", tc.value)
+			prefixes, restricted := parseTrustedProxies()
+			assert.Equal(t, tc.restricted, restricted)
+			var got []string
+			for _, p := range prefixes {
+				got = append(got, p.String())
+			}
+			assert.Equal(t, tc.prefixes, got)
+		})
+	}
+
+	t.Run("unset", func(t *testing.T) {
+		t.Setenv("TRUSTED_PROXY_IPS", "")
+		os.Unsetenv("TRUSTED_PROXY_IPS")
+		prefixes, restricted := parseTrustedProxies()
+		assert.False(t, restricted)
+		assert.Nil(t, prefixes)
+	})
+
+	t.Run("prefixed env var takes precedence", func(t *testing.T) {
+		t.Setenv("TRUSTED_PROXY_IPS", "10.0.0.0/8")
+		t.Setenv("BESZEL_HUB_TRUSTED_PROXY_IPS", "192.168.0.0/16")
+		prefixes, restricted := parseTrustedProxies()
+		assert.True(t, restricted)
+		require.Len(t, prefixes, 1)
+		assert.Equal(t, "192.168.0.0/16", prefixes[0].String())
+	})
+}
+
+func TestIsTrustedProxy(t *testing.T) {
+	prefixes := []netip.Prefix{
+		netip.MustParsePrefix("10.0.0.0/8"),
+		netip.MustParsePrefix("2001:db8::/32"),
+		netip.MustParsePrefix("fe80::/10"),
+	}
+
+	testCases := []struct {
+		name       string
+		remoteAddr string
+		trusted    bool
+	}{
+		{"ipv4 in prefix", "10.20.30.40:51234", true},
+		{"ipv4 outside prefix", "11.0.0.1:51234", false},
+		{"ipv6 in prefix", "[2001:db8:1::2]:443", true},
+		{"ipv6 outside prefix", "[2001:db9::1]:443", false},
+		{"ipv4-mapped ipv6 matches ipv4 prefix", "[::ffff:10.1.2.3]:80", true},
+		{"zone is ignored", "[fe80::1%eth0]:80", true},
+		{"no port", "10.1.2.3", true},
+		{"empty", "", false},
+		{"garbage", "not-an-address:80", false},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.trusted, isTrustedProxy(prefixes, tc.remoteAddr))
+		})
+	}
+
+	t.Run("empty allowlist trusts nobody", func(t *testing.T) {
+		assert.False(t, isTrustedProxy(nil, "10.0.0.1:1"))
+	})
+}

internal/hub/utils/utils.go

@@ -1,7 +1,11 @@
 // Package utils provides utility functions for the hub.
 package utils
 
-import "os"
+import (
+	"os"
+
+	"github.com/pocketbase/pocketbase/core"
+)
 
 // GetEnv retrieves an environment variable with a "BESZEL_HUB_" prefix, or falls back to the unprefixed key.
 func GetEnv(key string) (value string, exists bool) {
@@ -10,3 +14,26 @@ func GetEnv(key string) (value string, exists bool) {
 	}
 	return os.LookupEnv(key)
 }
+
+// realtimeActiveForCollection checks if there are active WebSocket subscriptions for the given collection.
+func RealtimeActiveForCollection(app core.App, collectionName string, validateFn func(filterQuery string) bool) bool {
+	broker := app.SubscriptionsBroker()
+	if broker.TotalClients() == 0 {
+		return false
+	}
+	for _, client := range broker.Clients() {
+		subs := client.Subscriptions(collectionName)
+		if len(subs) > 0 {
+			if validateFn == nil {
+				return true
+			}
+			for k := range subs {
+				filter := subs[k].Query["filter"]
+				if validateFn(filter) {
+					return true
+				}
+			}
+		}
+	}
+	return false
+}

internal/migrations/0_collections_snapshot_0_20_0_dev_5.go

@@ -83,7 +83,8 @@ func init() {
 					"ContainerHealth",
 					"SystemdFailed",
 					"CPUIOWait",
-					"CPUSteal"
+					"CPUSteal",
+					"NetworkMonitorLoss"
 				]
 			},
 			{
@@ -119,6 +120,16 @@ func init() {
 				"system": false,
 				"type": "bool"
 			},
+			{
+				"hidden": true,
+				"id": "json4000656575",
+				"maxSize": 0,
+				"name": "state",
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "json"
+			},
 			{
 				"hidden": true,
 				"id": "date1302749137",
@@ -152,6 +163,7 @@ func init() {
 			}
 		],
 		"indexes": [
+			"CREATE INDEX idx_alerts_system_name ON alerts (system, name)",
 			"CREATE UNIQUE INDEX ` + "`" + `idx_MnhEt21L5r` + "`" + ` ON ` + "`" + `alerts` + "`" + ` (\n  ` + "`" + `user` + "`" + `,\n  ` + "`" + `system` + "`" + `,\n  ` + "`" + `name` + "`" + `\n)"
 		],
 		"system": false
@@ -234,6 +246,20 @@ func init() {
 					"system": false,
 					"type": "text"
 				},
+				{
+					"autogeneratePattern": "",
+					"hidden": false,
+					"id": "text3888135399",
+					"max": 0,
+					"min": 0,
+					"name": "monitor_name",
+					"pattern": "",
+					"presentable": false,
+					"primaryKey": false,
+					"required": false,
+					"system": false,
+					"type": "text"
+				},
 				{
 					"hidden": false,
 					"id": "number494360628",
@@ -1721,6 +1747,7 @@ func init() {
 		"fields": [
 			{
 				"autogeneratePattern": "[a-z0-9]{15}",
+				"help": "",
 				"hidden": false,
 				"id": "text3208210256",
 				"max": 15,
@@ -1736,6 +1763,7 @@ func init() {
 			{
 				"cascadeDelete": true,
 				"collectionId": "2hz5ncl8tizk5nx",
+				"help": "",
 				"hidden": false,
 				"id": "relation1204987316",
 				"maxSelect": 1,
@@ -1748,6 +1776,7 @@ func init() {
 			},
 			{
 				"autogeneratePattern": "",
+				"help": "",
 				"hidden": false,
 				"id": "text7739291048",
 				"max": 0,
@@ -1762,6 +1791,7 @@ func init() {
 			},
 			{
 				"autogeneratePattern": "",
+				"help": "",
 				"hidden": false,
 				"id": "text5528164482",
 				"max": 0,
@@ -1775,6 +1805,7 @@ func init() {
 				"type": "text"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "number8862034195",
 				"max": null,
@@ -1787,6 +1818,7 @@ func init() {
 				"type": "number"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "number4418907321",
 				"max": null,
@@ -1799,6 +1831,7 @@ func init() {
 				"type": "number"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "number2904183765",
 				"max": null,
@@ -1811,6 +1844,7 @@ func init() {
 				"type": "number"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "json4466109723",
 				"maxSize": 0,
@@ -1821,6 +1855,7 @@ func init() {
 				"type": "json"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "json9012873456",
 				"maxSize": 0,
@@ -1831,6 +1866,7 @@ func init() {
 				"type": "json"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "json7182045639",
 				"maxSize": 0,
@@ -1841,6 +1877,7 @@ func init() {
 				"type": "json"
 			},
 			{
+				"help": "",
 				"hidden": false,
 				"id": "date9274163058",
 				"max": "",
@@ -1860,19 +1897,401 @@ func init() {
 				"presentable": false,
 				"system": false,
 				"type": "autodate"
+			},
+			{
+				"autogeneratePattern": "",
+				"help": "",
+				"hidden": false,
+				"id": "text3578368839",
+				"max": 0,
+				"min": 0,
+				"name": "display_name",
+				"pattern": "",
+				"presentable": false,
+				"primaryKey": false,
+				"required": false,
+				"system": false,
+				"type": "text"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "bool447994709",
+				"name": "raw",
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "bool"
 			}
 		],
 		"id": "pbc_8441057391",
 		"indexes": [
 			"CREATE INDEX ` + "`" + `idx_zfsPoolsSystem` + "`" + ` ON ` + "`" + `zfs_pools` + "`" + ` (` + "`" + `system` + "`" + `)"
 		],
-		"listRule": null,
+		"listRule": "@request.auth.id != \"\" && system.users.id ?= @request.auth.id",
 		"name": "zfs_pools",
 		"system": false,
 		"type": "base",
 		"updateRule": null,
+		"viewRule": "@request.auth.id != \"\" && system.users.id ?= @request.auth.id"
+	},
+	{
+		"createRule": null,
+		"deleteRule": null,
+		"fields": [
+			{
+				"autogeneratePattern": "[a-z0-9]{10}",
+				"help": "",
+				"hidden": false,
+				"id": "text3208210256",
+				"max": 10,
+				"min": 6,
+				"name": "id",
+				"pattern": "^[a-z0-9]+$",
+				"presentable": false,
+				"primaryKey": true,
+				"required": true,
+				"system": true,
+				"type": "text"
+			},
+			{
+				"cascadeDelete": true,
+				"collectionId": "2hz5ncl8tizk5nx",
+				"help": "",
+				"hidden": false,
+				"id": "nm_system",
+				"maxSelect": 1,
+				"minSelect": 0,
+				"name": "system",
+				"presentable": false,
+				"required": true,
+				"system": false,
+				"type": "relation"
+			},
+			{
+				"autogeneratePattern": "",
+				"help": "",
+				"hidden": false,
+				"id": "nm_target",
+				"max": 500,
+				"min": 1,
+				"name": "target",
+				"pattern": "",
+				"presentable": false,
+				"primaryKey": false,
+				"required": true,
+				"system": false,
+				"type": "text"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "nm_protocol",
+				"maxSelect": 1,
+				"name": "protocol",
+				"presentable": false,
+				"required": true,
+				"system": false,
+				"type": "select",
+				"values": [
+					"icmp",
+					"tcp",
+					"http",
+					"dns"
+				]
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "nm_port",
+				"max": 65535,
+				"min": 0,
+				"name": "port",
+				"onlyInt": true,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "nm_interval",
+				"max": 3600,
+				"min": 1,
+				"name": "interval",
+				"onlyInt": true,
+				"presentable": false,
+				"required": true,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "number926446584",
+				"max": null,
+				"min": null,
+				"name": "res",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "number1006954605",
+				"max": null,
+				"min": null,
+				"name": "resAvg1h",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "number4267669802",
+				"max": null,
+				"min": null,
+				"name": "resMin1h",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "number591433223",
+				"max": null,
+				"min": null,
+				"name": "resMax1h",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "number3726709001",
+				"max": null,
+				"min": null,
+				"name": "loss1h",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "nm_enabled",
+				"name": "enabled",
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "bool"
+			},
+			{
+				"hidden": false,
+				"id": "autodate2990389176",
+				"name": "created",
+				"onCreate": true,
+				"onUpdate": false,
+				"presentable": false,
+				"system": false,
+				"type": "autodate"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "date3332085495",
+				"max": "",
+				"min": "",
+				"name": "updated",
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "date"
+			}
+		],
+		"id": "nm_monitors_001",
+		"indexes": [
+			"CREATE INDEX ` + "`" + `idx_nm_system_enabled` + "`" + ` ON ` + "`" + `network_monitors` + "`" + ` (` + "`" + `system` + "`" + `, ` + "`" + `enabled` + "`" + `)"
+		],
+		"listRule": null,
+		"name": "network_monitors",
+		"system": false,
+		"type": "base",
+		"updateRule": null,
 		"viewRule": null
-		}
+	},
+	{
+		"createRule": null,
+		"deleteRule": null,
+		"fields": [
+			{
+				"autogeneratePattern": "[a-z0-9]{10}",
+				"help": "",
+				"hidden": false,
+				"id": "text3208210256",
+				"max": 10,
+				"min": 10,
+				"name": "id",
+				"pattern": "^[a-z0-9]+$",
+				"presentable": false,
+				"primaryKey": true,
+				"required": true,
+				"system": true,
+				"type": "text"
+			},
+			{
+				"cascadeDelete": true,
+				"collectionId": "2hz5ncl8tizk5nx",
+				"help": "",
+				"hidden": false,
+				"id": "nms_system",
+				"maxSelect": 1,
+				"minSelect": 0,
+				"name": "system",
+				"presentable": false,
+				"required": true,
+				"system": false,
+				"type": "relation"
+			},
+			{
+				"cascadeDelete": true,
+				"collectionId": "nm_monitors_001",
+				"help": "",
+				"hidden": false,
+				"id": "nms_monitor",
+				"maxSelect": 1,
+				"minSelect": 0,
+				"name": "monitor",
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "relation"
+			},
+			{
+				"help": "Number of probe attempts",
+				"hidden": false,
+				"id": "nms_total_count",
+				"max": null,
+				"min": 0,
+				"name": "total_count",
+				"onlyInt": true,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "Number of successful probe attempts",
+				"hidden": false,
+				"id": "nms_success_count",
+				"max": null,
+				"min": 0,
+				"name": "success_count",
+				"onlyInt": true,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "Sum of successful response times in microseconds",
+				"hidden": false,
+				"id": "nms_res_sum",
+				"max": null,
+				"min": 0,
+				"name": "res_sum",
+				"onlyInt": true,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "Response time in microseconds",
+				"hidden": false,
+				"id": "nms_res_min",
+				"max": null,
+				"min": 0,
+				"name": "res_min",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "Response time in microseconds",
+				"hidden": false,
+				"id": "nms_res_max",
+				"max": null,
+				"min": 0,
+				"name": "res_max",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "nms_type",
+				"maxSelect": 1,
+				"name": "type",
+				"presentable": false,
+				"required": true,
+				"system": false,
+				"type": "select",
+				"values": [
+					"1m",
+					"10m",
+					"20m",
+					"120m",
+					"480m"
+				]
+			},
+			{
+				"help": "",
+				"hidden": false,
+				"id": "number2990389176",
+				"max": null,
+				"min": null,
+				"name": "created",
+				"onlyInt": false,
+				"presentable": false,
+				"required": false,
+				"system": false,
+				"type": "number"
+			}
+		],
+		"id": "nm_stats_001",
+		"indexes": [
+			"CREATE INDEX IF NOT EXISTS ` + "`" + `idx_nms_system_type_created` + "`" + ` ON ` + "`" + `network_monitor_stats` + "`" + ` (` + "`" + `system` + "`" + `, ` + "`" + `type` + "`" + `, ` + "`" + `created` + "`" + `)",
+			"CREATE INDEX IF NOT EXISTS ` + "`" + `idx_nms_monitor_type_created` + "`" + ` ON ` + "`" + `network_monitor_stats` + "`" + ` (` + "`" + `monitor` + "`" + `, ` + "`" + `type` + "`" + `, ` + "`" + `created` + "`" + `)",
+			"CREATE INDEX IF NOT EXISTS ` + "`" + `idx_nms_type_created` + "`" + ` ON ` + "`" + `network_monitor_stats` + "`" + ` (` + "`" + `type` + "`" + `, ` + "`" + `created` + "`" + `)"
+		],
+		"listRule": null,
+		"name": "network_monitor_stats",
+		"system": false,
+		"type": "base",
+		"updateRule": null,
+		"viewRule": null
+	}
 ]`
 
 		err := app.ImportCollectionsByMarshaledJSON([]byte(jsonData), false)

internal/migrations/1788955200_pool_updates.go

@@ -0,0 +1,27 @@
+package migrations
+
+import (
+	"github.com/pocketbase/pocketbase/core"
+	m "github.com/pocketbase/pocketbase/migrations"
+)
+
+func init() {
+	m.Register(func(app core.App) error {
+		c, err := app.FindCollectionByNameOrId("zfs_pools")
+		if err != nil {
+			return err
+		}
+		c.Fields.Add(&core.TextField{Name: "display_name"})
+		c.Fields.Add(&core.BoolField{Name: "raw"})
+		return app.Save(c)
+	}, func(app core.App) error {
+		c, err := app.FindCollectionByNameOrId("zfs_pools")
+		if err != nil {
+			return err
+		}
+		c.Fields.RemoveByName("display_name")
+		c.Fields.RemoveByName("raw")
+
+		return app.Save(c)
+	})
+}