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

.github/workflows/helm-charts.yml

@@ -63,7 +63,7 @@ jobs:
         uses: actions/checkout@v7
 
       - name: Set up Helm
-        uses: azure/setup-helm@v4
+        uses: azure/setup-helm@v5
 
       - name: Lint chart
         run: helm lint "${{ matrix.chart.path }}" --set env.KEY=ci-placeholder

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

@@ -48,6 +48,8 @@ type Agent struct {
 	keys                      []gossh.PublicKey                                     // SSH public keys
 	smartManager              *SmartManager                                         // Manages SMART data
 	systemdManager            *systemdManager                                       // Manages systemd services
+	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.
@@ -121,6 +123,22 @@ func NewAgent(dataDir ...string) (agent *Agent, err error) {
 	// initialize handler registry
 	agent.handlerRegistry = NewHandlerRegistry()
 
+	// initialize monitor manager
+	agent.monitorManager = newMonitorManager()
+
+	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.storagePoolManager.detailInterval = duration
+			agent.systemDetails.ZfsInterval = duration
+			slog.Info("ZFS_INTERVAL", "duration", duration)
+		} else {
+			slog.Warn("Invalid ZFS_INTERVAL", "err", err)
+		}
+	}
+
 	// initialize disk info
 	agent.initializeDiskInfo()
 
@@ -178,6 +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())
@@ -187,13 +210,25 @@ func (a *Agent) gatherStats(options common.DataRequestOptions) *system.CombinedD
 		}
 		if a.systemdManager.hasFreshStats {
 			data.SystemdServices = a.systemdManager.getServiceStats(nil, false)
+			data.SystemdServicesUpdated = true
+			// Preserve an explicit zero count so the hub can distinguish a fresh
+			// empty snapshot from a response that omitted systemd data.
+			if totalCount == 0 {
+				data.Info.Services = []uint16{0, 0}
+			}
 		}
 	}
 
 	data.Stats.ExtraFs = make(map[string]*system.FsStats)
 	data.Info.ExtraFsPct = make(map[string]float64)
 	for name, stats := range a.fsStats {
-		if !stats.Root && stats.DiskTotal > 0 {
+		if stats.Root {
+			if stats.Name != "" {
+				data.Info.RootDiskName = stats.Name
+			}
+			continue
+		}
+		if stats.DiskTotal > 0 {
 			// Use custom name if available, otherwise use device name
 			key := name
 			if stats.Name != "" {

agent/battery/battery.go

@@ -33,7 +33,10 @@ var errNoBatteries = errors.New("no readable batteries")
 func normalizeBatteries(batteries []Battery) []Battery {
 	nameCounts := make(map[string]int, len(batteries))
 	for i := range batteries {
-		name := strings.TrimSpace(batteries[i].Name)
+		// Names come from firmware (e.g. sysfs model_name) and are not guaranteed to
+		// be valid UTF-8. Invalid bytes are rejected when the hub decodes the CBOR
+		// payload, which drops every metric for the system, so strip them here.
+		name := strings.TrimSpace(strings.ToValidUTF8(batteries[i].Name, ""))
 		if name == "" {
 			name = "Battery " + strconv.Itoa(i+1)
 		}

agent/battery/battery_test.go

@@ -2,6 +2,7 @@ package battery
 
 import (
 	"testing"
+	"unicode/utf8"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -33,3 +34,15 @@ func TestNormalizeBatteriesFallbackNames(t *testing.T) {
 	bats := normalizeBatteries([]Battery{{}, {}, {Name: "Mouse"}, {Name: "Mouse"}})
 	assert.Equal(t, []string{"Battery 1", "Battery 2", "Mouse", "Mouse (2)"}, []string{bats[0].Name, bats[1].Name, bats[2].Name, bats[3].Name})
 }
+
+func TestNormalizeBatteriesStripsInvalidUTF8(t *testing.T) {
+	// Firmware occasionally reports names that are not valid UTF-8 (a ThinkPad
+	// reporting "LNV-5B11K63024@\xd0" in model_name is a real example).
+	bats := normalizeBatteries([]Battery{{Name: "LNV-5B11K63024@\xd0"}, {Name: "\xff\xfe"}})
+	assert.Equal(t, "LNV-5B11K63024@", bats[0].Name)
+	// A name made up entirely of invalid bytes falls back to the generic name.
+	assert.Equal(t, "Battery 2", bats[1].Name)
+	for _, b := range bats {
+		assert.True(t, utf8.ValidString(b.Name))
+	}
+}

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

@@ -2,6 +2,7 @@ package agent
 
 import (
 	"crypto/tls"
+	"crypto/x509"
 	"errors"
 	"fmt"
 	"log/slog"
@@ -24,9 +25,28 @@ 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
+}
+
+func (e *caCertFileError) Error() string {
+	return e.err.Error()
+}
+
+func (e *caCertFileError) Unwrap() error {
+	return e.err
+}
+
 // WebSocketClient manages the WebSocket connection between the agent and hub.
 // It handles authentication, message routing, and connection lifecycle management.
 type WebSocketClient struct {
@@ -40,6 +60,7 @@ type WebSocketClient struct {
 	hubRequest         *common.HubRequest[cbor.RawMessage] // Reusable request structure for message parsing
 	lastConnectAttempt time.Time                           // Timestamp of last connection attempt
 	hubVerified        bool                                // Whether the hub has been cryptographically verified
+	tlsConfig          *tls.Config                         // Optional TLS configuration with custom CA certificates
 }
 
 // newWebSocketClient creates a new WebSocket client for the given agent.
@@ -47,7 +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{}
@@ -61,6 +82,10 @@ func newWebSocketClient(agent *Agent) (client *WebSocketClient, err error) {
 	if err != nil {
 		return nil, err
 	}
+	client.tlsConfig, err = getTLSConfig()
+	if err != nil {
+		return nil, err
+	}
 
 	client.agent = agent
 	client.hubRequest = &common.HubRequest[cbor.RawMessage]{}
@@ -87,7 +112,52 @@ func getToken() (string, error) {
 	if err != nil {
 		return "", err
 	}
-	return strings.TrimSpace(string(tokenBytes)), nil
+	return parseTokenFile(string(tokenBytes), tokenFile)
+}
+
+// parseTokenFile reads a single token from TOKEN_FILE.
+// Blank lines and comments are ignored. Multiple tokens are rejected because
+// the agent supports only one outbound hub connection.
+func parseTokenFile(contents, path string) (string, error) {
+	var token string
+	for line := range strings.Lines(contents) {
+		line = strings.TrimSpace(line)
+		if len(line) == 0 || strings.HasPrefix(line, "#") {
+			continue
+		}
+		if token != "" {
+			return "", fmt.Errorf("%s must contain a single token", path)
+		}
+		token = line
+	}
+	// An empty file keeps returning an empty token, as before: the caller decides
+	// what to do about it.
+	return token, nil
+}
+
+// getTLSConfig returns a TLS configuration containing the system certificate
+// pool plus any certificates configured through CA_CERT_FILE. A nil config lets
+// gws use Go's default TLS configuration and system roots.
+func getTLSConfig() (*tls.Config, error) {
+	caCertFile, _ := utils.GetEnv("CA_CERT_FILE")
+	if caCertFile == "" {
+		return nil, nil
+	}
+
+	caCertPEM, err := os.ReadFile(caCertFile)
+	if err != nil {
+		return nil, &caCertFileError{fmt.Errorf("read CA_CERT_FILE %q: %w", caCertFile, err)}
+	}
+
+	rootCAs, err := x509.SystemCertPool()
+	if err != nil {
+		return nil, &caCertFileError{fmt.Errorf("load system CA certificate pool: %w", err)}
+	}
+	if !rootCAs.AppendCertsFromPEM(caCertPEM) {
+		return nil, &caCertFileError{fmt.Errorf("CA_CERT_FILE %q does not contain any valid PEM certificates", caCertFile)}
+	}
+
+	return &tls.Config{RootCAs: rootCAs}, nil
 }
 
 // getOptions returns the WebSocket client options, creating them if necessary.
@@ -112,7 +182,7 @@ func (client *WebSocketClient) getOptions() *gws.ClientOption {
 
 	client.options = &gws.ClientOption{
 		Addr:      client.hubURL.String(),
-		TlsConfig: &tls.Config{InsecureSkipVerify: true},
+		TlsConfig: client.tlsConfig,
 		RequestHeader: http.Header{
 			"User-Agent": []string{getUserAgent()},
 			"X-Token":    []string{client.token},

agent/client_test.go

@@ -4,8 +4,19 @@ package agent
 
 import (
 	"crypto/ed25519"
+	"crypto/rand"
+	"crypto/rsa"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/pem"
+	"math/big"
+	"net"
+	"net/http"
+	"net/http/httptest"
 	"net/url"
 	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 	"time"
@@ -15,11 +26,34 @@ import (
 	"github.com/henrygd/beszel/internal/common"
 
 	"github.com/fxamacker/cbor/v2"
+	"github.com/lxzan/gws"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 	"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)
@@ -164,6 +198,155 @@ func TestWebSocketClient_GetOptions(t *testing.T) {
 	}
 }
 
+func TestWebSocketClient_TLSVerification(t *testing.T) {
+	agent := createTestAgent(t)
+	serverCert, serverCertPEM := newSelfSignedServerCertificate(t)
+	upgrader := gws.NewUpgrader(&gws.BuiltinEventHandler{}, nil)
+	server := httptest.NewUnstartedServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		conn, err := upgrader.Upgrade(w, r)
+		if err == nil {
+			go conn.ReadLoop()
+		}
+	}))
+	server.TLS = &tls.Config{Certificates: []tls.Certificate{serverCert}}
+	server.StartTLS()
+	t.Cleanup(server.Close)
+
+	caCertFile := filepath.Join(t.TempDir(), "hub-ca.crt")
+	require.NoError(t, os.WriteFile(caCertFile, serverCertPEM, 0600))
+
+	newClient := func(t *testing.T, caCertFile string) *WebSocketClient {
+		t.Helper()
+		t.Setenv("BESZEL_AGENT_HUB_URL", server.URL)
+		t.Setenv("BESZEL_AGENT_TOKEN", "test-token")
+		t.Setenv("BESZEL_AGENT_CA_CERT_FILE", caCertFile)
+		client, err := newWebSocketClient(agent)
+		require.NoError(t, err)
+		return client
+	}
+
+	t.Run("system roots are used by default", func(t *testing.T) {
+		client := newClient(t, "")
+		assert.Nil(t, client.getOptions().TlsConfig)
+		_, _, err := gws.NewClient(&gws.BuiltinEventHandler{}, client.getOptions())
+		require.Error(t, err)
+	})
+
+	t.Run("custom CA trusts self-signed certificate", func(t *testing.T) {
+		systemRoots, err := x509.SystemCertPool()
+		require.NoError(t, err)
+		client := newClient(t, caCertFile)
+		assert.Greater(t, len(client.getOptions().TlsConfig.RootCAs.Subjects()), len(systemRoots.Subjects()))
+		conn, _, err := gws.NewClient(&gws.BuiltinEventHandler{}, client.getOptions())
+		require.NoError(t, err)
+		require.NoError(t, conn.NetConn().Close())
+	})
+
+	t.Run("custom CA does not bypass hostname verification", func(t *testing.T) {
+		client := newClient(t, caCertFile)
+		client.getOptions().TlsConfig.ServerName = "wrong.example.com"
+		_, _, err := gws.NewClient(&gws.BuiltinEventHandler{}, client.getOptions())
+		require.Error(t, err)
+	})
+}
+
+func TestWebSocketClient_NonTLSConnection(t *testing.T) {
+	agent := createTestAgent(t)
+	upgrader := gws.NewUpgrader(&gws.BuiltinEventHandler{}, nil)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		conn, err := upgrader.Upgrade(w, r)
+		if err == nil {
+			go conn.ReadLoop()
+		}
+	}))
+	t.Cleanup(server.Close)
+
+	t.Setenv("BESZEL_AGENT_HUB_URL", server.URL)
+	t.Setenv("BESZEL_AGENT_TOKEN", "test-token")
+	t.Setenv("BESZEL_AGENT_CA_CERT_FILE", "")
+	client, err := newWebSocketClient(agent)
+	require.NoError(t, err)
+	assert.Nil(t, client.getOptions().TlsConfig)
+
+	conn, _, err := gws.NewClient(&gws.BuiltinEventHandler{}, client.getOptions())
+	require.NoError(t, err)
+	require.NoError(t, conn.NetConn().Close())
+}
+
+func TestGetTLSConfigErrors(t *testing.T) {
+	tempDir := t.TempDir()
+	testCases := []struct {
+		name       string
+		path       string
+		contents   []byte
+		errorMatch string
+	}{
+		{
+			name:       "missing file",
+			path:       filepath.Join(tempDir, "missing.pem"),
+			errorMatch: "read CA_CERT_FILE",
+		},
+		{
+			name:       "unreadable path",
+			path:       tempDir,
+			errorMatch: "read CA_CERT_FILE",
+		},
+		{
+			name:       "empty file",
+			path:       filepath.Join(tempDir, "empty.pem"),
+			contents:   []byte{},
+			errorMatch: "does not contain any valid PEM certificates",
+		},
+		{
+			name:       "malformed file",
+			path:       filepath.Join(tempDir, "malformed.pem"),
+			contents:   []byte("not a PEM certificate"),
+			errorMatch: "does not contain any valid PEM certificates",
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			if tc.contents != nil {
+				require.NoError(t, os.WriteFile(tc.path, tc.contents, 0600))
+			}
+			t.Setenv("BESZEL_AGENT_CA_CERT_FILE", tc.path)
+
+			tlsConfig, err := getTLSConfig()
+			require.Error(t, err)
+			assert.Nil(t, tlsConfig)
+			assert.Contains(t, err.Error(), tc.errorMatch)
+			assert.Contains(t, err.Error(), tc.path)
+		})
+	}
+}
+
+func newSelfSignedServerCertificate(t *testing.T) (tls.Certificate, []byte) {
+	t.Helper()
+	privateKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	template := &x509.Certificate{
+		SerialNumber:          big.NewInt(1),
+		Subject:               pkix.Name{CommonName: "127.0.0.1"},
+		NotBefore:             time.Now().Add(-time.Hour),
+		NotAfter:              time.Now().Add(time.Hour),
+		IPAddresses:           []net.IP{net.ParseIP("127.0.0.1")},
+		KeyUsage:              x509.KeyUsageDigitalSignature | x509.KeyUsageKeyEncipherment | x509.KeyUsageCertSign,
+		ExtKeyUsage:           []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+		BasicConstraintsValid: true,
+		IsCA:                  true,
+	}
+	certDER, err := x509.CreateCertificate(rand.Reader, template, template, &privateKey.PublicKey, privateKey)
+	require.NoError(t, err)
+
+	certPEM := pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: certDER})
+	keyPEM := pem.EncodeToMemory(&pem.Block{Type: "RSA PRIVATE KEY", Bytes: x509.MarshalPKCS1PrivateKey(privateKey)})
+	certificate, err := tls.X509KeyPair(certPEM, keyPEM)
+	require.NoError(t, err)
+	return certificate, certPEM
+}
+
 // TestWebSocketClient_VerifySignature tests signature verification
 func TestWebSocketClient_VerifySignature(t *testing.T) {
 	agent := createTestAgent(t)
@@ -409,6 +592,41 @@ func TestGetToken(t *testing.T) {
 		assert.Equal(t, expectedToken, token)
 	})
 
+	t.Run("TOKEN_FILE with surrounding blank lines and comments", func(t *testing.T) {
+		expectedToken := "test-token-with-noise"
+		tokenFile := filepath.Join(t.TempDir(), "token")
+		require.NoError(t, os.WriteFile(tokenFile, []byte("# hub token\n\n"+expectedToken+"\n\n"), 0o600))
+
+		t.Setenv("TOKEN_FILE", tokenFile)
+
+		token, err := getToken()
+		assert.NoError(t, err)
+		assert.Equal(t, expectedToken, token)
+	})
+
+	t.Run("TOKEN_FILE with multiple tokens is rejected", func(t *testing.T) {
+		tokenFile := filepath.Join(t.TempDir(), "token")
+		require.NoError(t, os.WriteFile(tokenFile, []byte("11111111-1111-1111-1111-111111111111\n22222222-2222-2222-2222-222222222222\n"), 0o600))
+
+		t.Setenv("TOKEN_FILE", tokenFile)
+
+		token, err := getToken()
+		require.Error(t, err)
+		assert.Empty(t, token)
+		assert.Contains(t, err.Error(), "must contain a single token")
+	})
+
+	t.Run("TOKEN_FILE holding only comments behaves like an empty file", func(t *testing.T) {
+		tokenFile := filepath.Join(t.TempDir(), "token")
+		require.NoError(t, os.WriteFile(tokenFile, []byte("\n# only a comment\n"), 0o600))
+
+		t.Setenv("TOKEN_FILE", tokenFile)
+
+		token, err := getToken()
+		assert.NoError(t, err)
+		assert.Equal(t, "", token)
+	})
+
 	t.Run("token from BESZEL_AGENT_TOKEN_FILE", func(t *testing.T) {
 		// Create a temporary token file
 		expectedToken := "test-token-from-beszel-file"
@@ -504,3 +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

@@ -87,7 +87,19 @@ func (c *ConnectionManager) Start(serverOptions ServerOptions) error {
 
 	wsClient, err := newWebSocketClient(c.agent)
 	if err != nil {
-		slog.Warn("Error creating WebSocket client", "err", err)
+		var caCertErr *caCertFileError
+		if errors.As(err, &caCertErr) {
+			return 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
 
@@ -141,6 +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()
 }
@@ -151,7 +164,9 @@ func (c *ConnectionManager) handleEvent(event ConnectionEvent) {
 	case WebSocketConnect:
 		c.handleStateChange(WebSocketConnected)
 	case SSHConnect:
-		c.handleStateChange(SSHConnected)
+		if c.State == Disconnected {
+			c.handleStateChange(SSHConnected)
+		}
 	case WebSocketDisconnect:
 		if c.State == WebSocketConnected {
 			c.handleStateChange(Disconnected)

agent/connection_manager_test.go

@@ -114,6 +114,12 @@ func TestConnectionManager_EventHandling(t *testing.T) {
 			event:         SSHConnect,
 			expectedState: SSHConnected,
 		},
+		{
+			name:          "SSH connect from WebSocket connected (no change)",
+			initialState:  WebSocketConnected,
+			event:         SSHConnect,
+			expectedState: WebSocketConnected,
+		},
 		{
 			name:          "WebSocket disconnect from connected",
 			initialState:  WebSocketConnected,
@@ -265,6 +271,19 @@ func TestConnectionManager_StartWithInvalidConfig(t *testing.T) {
 	assert.Error(t, err, "Should error when starting already started connection manager")
 }
 
+func TestConnectionManager_StartRejectsInvalidCACertFile(t *testing.T) {
+	agent := createTestAgent(t)
+	cm := agent.connectionManager
+	t.Setenv("BESZEL_AGENT_HUB_URL", "https://hub.example.com")
+	t.Setenv("BESZEL_AGENT_TOKEN", "test-token")
+	t.Setenv("BESZEL_AGENT_CA_CERT_FILE", t.TempDir())
+
+	err := cm.Start(ServerOptions{})
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "read CA_CERT_FILE")
+	assert.Nil(t, cm.eventChan)
+}
+
 // TestConnectionManager_CloseWebSocket tests WebSocket closing
 func TestConnectionManager_CloseWebSocket(t *testing.T) {
 	agent := createTestAgent(t)

agent/data_dir_test.go

@@ -12,6 +12,14 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func invalidDataDir(t *testing.T) string {
+	t.Helper()
+
+	filePath := filepath.Join(t.TempDir(), "file")
+	require.NoError(t, os.WriteFile(filePath, nil, 0644))
+	return filepath.Join(filePath, "data")
+}
+
 func TestGetDataDir(t *testing.T) {
 	// Test with explicit dataDir parameter
 	t.Run("explicit data dir", func(t *testing.T) {
@@ -48,7 +56,7 @@ func TestGetDataDir(t *testing.T) {
 
 	// Test with invalid explicit dataDir
 	t.Run("invalid explicit data dir", func(t *testing.T) {
-		invalidPath := "/invalid/path/that/cannot/be/created"
+		invalidPath := invalidDataDir(t)
 		_, err := GetDataDir(invalidPath)
 		assert.Error(t, err)
 	})
@@ -78,7 +86,7 @@ func TestTestDataDirs(t *testing.T) {
 	// Test with multiple directories, first one valid
 	t.Run("multiple dirs - first valid", func(t *testing.T) {
 		tempDir := t.TempDir()
-		invalidDir := "/invalid/path"
+		invalidDir := invalidDataDir(t)
 		result, err := testDataDirs([]string{tempDir, invalidDir})
 		require.NoError(t, err)
 		assert.Equal(t, tempDir, result)
@@ -87,7 +95,7 @@ func TestTestDataDirs(t *testing.T) {
 	// Test with multiple directories, second one valid
 	t.Run("multiple dirs - second valid", func(t *testing.T) {
 		tempDir := t.TempDir()
-		invalidDir := "/invalid/path"
+		invalidDir := invalidDataDir(t)
 		result, err := testDataDirs([]string{invalidDir, tempDir})
 		require.NoError(t, err)
 		assert.Equal(t, tempDir, result)
@@ -109,7 +117,7 @@ func TestTestDataDirs(t *testing.T) {
 
 	// Test with no valid directories
 	t.Run("no valid directories", func(t *testing.T) {
-		invalidPaths := []string{"/invalid/path1", "/invalid/path2"}
+		invalidPaths := []string{invalidDataDir(t), invalidDataDir(t)}
 		_, err := testDataDirs(invalidPaths)
 		assert.Error(t, err)
 		assert.Contains(t, err.Error(), "data directory not found")

agent/disk.go

@@ -18,7 +18,8 @@ import (
 // fsRegistrationContext holds the shared lookup state needed to resolve a
 // filesystem into the tracked fsStats key and metadata.
 type fsRegistrationContext struct {
-	filesystem     string // value of optional FILESYSTEM env var
+	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
@@ -177,7 +178,7 @@ func (d *diskDiscovery) addConfiguredRootFs() bool {
 
 	for _, p := range d.partitions {
 		if filesystemMatchesPartitionSetting(d.ctx.filesystem, p) {
-			d.addFsStat(p.Device, p.Mountpoint, true, "")
+			d.addFsStat(p.Device, p.Mountpoint, true, d.ctx.filesystemName)
 			return true
 		}
 	}
@@ -185,7 +186,7 @@ func (d *diskDiscovery) addConfiguredRootFs() bool {
 	// FILESYSTEM may name a physical disk absent from partitions (e.g. ZFS lists
 	// dataset paths like zroot/ROOT/default, not block devices).
 	if ioKey, match := findIoDevice(d.ctx.filesystem, d.ctx.diskIoCounters); match {
-		d.agent.fsStats[ioKey] = &system.FsStats{Root: true, Mountpoint: d.rootMountPoint}
+		d.agent.fsStats[ioKey] = &system.FsStats{Root: true, Mountpoint: d.rootMountPoint, Name: d.ctx.filesystemName}
 		return true
 	}
 
@@ -300,7 +301,8 @@ func (d *diskDiscovery) addExtraFilesystemFolders(folderNames []string) {
 
 // Sets up the filesystems to monitor for disk usage and I/O.
 func (a *Agent) initializeDiskInfo() {
-	filesystem, _ := utils.GetEnv("FILESYSTEM")
+	filesystemRaw, _ := utils.GetEnv("FILESYSTEM")
+	filesystem, filesystemName := parseFilesystemEntry(filesystemRaw)
 	hasRoot := false
 	isWindows := runtime.GOOS == "windows"
 
@@ -324,6 +326,7 @@ func (a *Agent) initializeDiskInfo() {
 	slog.Debug("Disk I/O", "diskstats", diskIoCounters)
 	ctx := fsRegistrationContext{
 		filesystem:     filesystem,
+		filesystemName: filesystemName,
 		isWindows:      isWindows,
 		diskIoCounters: diskIoCounters,
 		efPath:         "/extra-filesystems",
@@ -534,7 +537,16 @@ func normalizeDeviceName(value string) string {
 func (a *Agent) initializeDiskIoStats(diskIoCounters map[string]disk.IOCountersStat) {
 	a.fsNames = a.fsNames[:0]
 	now := time.Now()
+	// 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.storagePoolManager != nil {
+		zfsMountpoints = a.storagePoolManager.ZfsMountpoints()
+	}
 	for device, stats := range a.fsStats {
+		if zfsMountpoints[stats.Mountpoint] {
+			continue
+		}
 		// skip if not in diskIoCounters
 		d, exists := diskIoCounters[device]
 		if !exists {
@@ -559,20 +571,31 @@ func (a *Agent) updateDiskUsage(systemStats *system.Stats) {
 		!a.lastDiskUsageUpdate.IsZero() &&
 		time.Since(a.lastDiskUsageUpdate) < a.diskUsageCacheDuration
 
+	// 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.storagePoolManager != nil {
+		zfsUsage = a.storagePoolManager.DatasetUsage()
+	}
+
 	// disk usage
 	for _, stats := range a.fsStats {
 		// Skip non-root filesystems if caching is active
 		if cacheExtraFs && !stats.Root {
 			continue
 		}
-		if d, err := disk.Usage(stats.Mountpoint); err == nil {
-			stats.DiskTotal = utils.BytesToGigabytes(d.Total)
-			stats.DiskUsed = utils.BytesToGigabytes(d.Used)
-			if stats.Root {
-				systemStats.DiskTotal = utils.BytesToGigabytes(d.Total)
-				systemStats.DiskUsed = utils.BytesToGigabytes(d.Used)
-				systemStats.DiskPct = utils.TwoDecimals(d.UsedPercent)
+		var total, used uint64
+		var usedPct float64
+		if u, ok := zfsUsage[stats.Mountpoint]; ok {
+			total = u.used + u.avail
+			used = u.used
+			if total > 0 {
+				usedPct = float64(used) / float64(total) * 100
 			}
+		} else if d, err := disk.Usage(stats.Mountpoint); err == nil {
+			total = d.Total
+			used = d.Used
+			usedPct = d.UsedPercent
 		} else {
 			// reset stats if error (likely unmounted)
 			slog.Error("Error getting disk stats", "name", stats.Mountpoint, "err", err)
@@ -580,6 +603,14 @@ func (a *Agent) updateDiskUsage(systemStats *system.Stats) {
 			stats.DiskUsed = 0
 			stats.TotalRead = 0
 			stats.TotalWrite = 0
+			continue
+		}
+		stats.DiskTotal = utils.BytesToGigabytes(total)
+		stats.DiskUsed = utils.BytesToGigabytes(used)
+		if stats.Root {
+			systemStats.DiskTotal = stats.DiskTotal
+			systemStats.DiskUsed = stats.DiskUsed
+			systemStats.DiskPct = utils.TwoDecimals(usedPct)
 		}
 	}
 
@@ -696,6 +727,8 @@ func (a *Agent) updateDiskIo(cacheTimeMs uint16, systemStats *system.Stats) {
 				systemStats.DiskWritePs = stats.DiskWritePs
 				systemStats.DiskIO[0] = diskIORead
 				systemStats.DiskIO[1] = diskIOWrite
+				systemStats.DiskIOTotal[0] = d.ReadBytes
+				systemStats.DiskIOTotal[1] = d.WriteBytes
 				systemStats.DiskIoStats[0] = diskReadTime
 				systemStats.DiskIoStats[1] = diskWriteTime
 				systemStats.DiskIoStats[2] = diskIoUtilPct

agent/disk_test.go

@@ -78,14 +78,7 @@ func TestParseFilesystemEntry(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			fsEntry := strings.TrimSpace(tt.input)
-			var fs, customName string
-			if parts := strings.SplitN(fsEntry, "__", 2); len(parts) == 2 {
-				fs = strings.TrimSpace(parts[0])
-				customName = strings.TrimSpace(parts[1])
-			} else {
-				fs = fsEntry
-			}
+			fs, customName := parseFilesystemEntry(tt.input)
 
 			assert.Equal(t, tt.expectedFs, fs)
 			assert.Equal(t, tt.expectedName, customName)
@@ -287,8 +280,9 @@ func TestAddConfiguredRootFs(t *testing.T) {
 			rootMountPoint: "/",
 			partitions:     []disk.PartitionStat{{Device: "/dev/ada0p2", Mountpoint: "/"}},
 			ctx: fsRegistrationContext{
-				filesystem: "/dev/ada0p2",
-				isWindows:  false,
+				filesystem:     "/dev/ada0p2",
+				filesystemName: "root disk",
+				isWindows:      false,
 				diskIoCounters: map[string]disk.IOCountersStat{
 					"ada0": {Name: "ada0", ReadBytes: 1000, WriteBytes: 1000},
 				},
@@ -302,6 +296,7 @@ func TestAddConfiguredRootFs(t *testing.T) {
 		assert.True(t, exists)
 		assert.True(t, stats.Root)
 		assert.Equal(t, "/", stats.Mountpoint)
+		assert.Equal(t, "root disk", stats.Name)
 	})
 
 	t.Run("adds root from io device when partition is missing", func(t *testing.T) {
@@ -310,8 +305,9 @@ func TestAddConfiguredRootFs(t *testing.T) {
 			agent:          agent,
 			rootMountPoint: "/sysroot",
 			ctx: fsRegistrationContext{
-				filesystem: "zroot",
-				isWindows:  false,
+				filesystem:     "zroot",
+				filesystemName: "root pool",
+				isWindows:      false,
 				diskIoCounters: map[string]disk.IOCountersStat{
 					"nda0": {Name: "nda0", Label: "zroot", ReadBytes: 1000, WriteBytes: 1000},
 				},
@@ -325,6 +321,7 @@ func TestAddConfiguredRootFs(t *testing.T) {
 		assert.True(t, exists)
 		assert.True(t, stats.Root)
 		assert.Equal(t, "/sysroot", stats.Mountpoint)
+		assert.Equal(t, "root pool", stats.Name)
 	})
 
 	t.Run("returns false when filesystem cannot be resolved", func(t *testing.T) {

agent/disk_zfs_test.go

@@ -0,0 +1,110 @@
+//go:build testing
+
+package agent
+
+import (
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/agent/zfs"
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/shirou/gopsutil/v4/disk"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestUpdateDiskUsageZfsMountpoint verifies that a filesystem whose mountpoint
+// 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 := &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
+	}
+	agent := &Agent{
+		fsStats: map[string]*system.FsStats{
+			"tank": {Root: false, Mountpoint: "/tank"},
+		},
+		storagePoolManager: zm,
+	}
+
+	var stats system.Stats
+	agent.updateDiskUsage(&stats)
+
+	fs := agent.fsStats["tank"]
+	require.NotNil(t, fs)
+	assert.Equal(t, 22350.81, fs.DiskTotal) // (used + avail) in GiB
+	assert.Equal(t, 11175.87, fs.DiskUsed)
+	// Non-root filesystems do not populate system-level stats.
+	assert.Equal(t, float64(0), stats.DiskTotal)
+}
+
+// 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 := &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
+	}
+	agent := &Agent{
+		fsStats: map[string]*system.FsStats{
+			"rpool/ROOT/pve-1": {Root: true, Mountpoint: "/"},
+		},
+		storagePoolManager: zm,
+	}
+
+	var stats system.Stats
+	agent.updateDiskUsage(&stats)
+
+	assert.Equal(t, 1117.59, agent.fsStats["rpool/ROOT/pve-1"].DiskTotal)
+	assert.Equal(t, 838.19, agent.fsStats["rpool/ROOT/pve-1"].DiskUsed)
+	assert.Equal(t, 75.0, stats.DiskPct)
+	assert.Equal(t, 1117.59, stats.DiskTotal)
+	assert.Equal(t, 838.19, stats.DiskUsed)
+}
+
+// TestUpdateDiskUsageWithoutZfsManager falls back to statfs when no manager is
+// present (e.g. tests constructing bare Agent values).
+func TestUpdateDiskUsageWithoutZfsManager(t *testing.T) {
+	agent := &Agent{
+		fsStats: map[string]*system.FsStats{
+			"root": {Root: true, Mountpoint: "/"},
+		},
+	}
+
+	var stats system.Stats
+	agent.updateDiskUsage(&stats)
+
+	assert.True(t, agent.fsStats["root"].DiskTotal > 0, "root usage should come from statfs")
+	assert.True(t, stats.DiskTotal > 0)
+}
+
+// TestInitializeDiskIoStatsSkipsZfsMountpoints verifies ZFS filesystems are
+// excluded from diskstats I/O tracking instead of warning about a missing device.
+func TestInitializeDiskIoStatsSkipsZfsMountpoints(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"}}, nil
+	}
+	agent := &Agent{
+		fsStats: map[string]*system.FsStats{
+			"tank": {Root: false, Mountpoint: "/tank"},
+			"sda1": {Root: false, Mountpoint: "/mnt/data"},
+		},
+		storagePoolManager: zm,
+		diskPrev:           make(map[uint16]map[string]prevDisk),
+	}
+
+	agent.initializeDiskIoStats(map[string]disk.IOCountersStat{
+		"sda1": {Name: "sda1", ReadBytes: 100, WriteBytes: 100},
+	})
+
+	assert.Equal(t, []string{"sda1"}, agent.fsNames)
+	assert.Equal(t, uint64(100), agent.fsStats["sda1"].TotalRead)
+	// ZFS entry is present but untouched by diskstats initialization.
+	assert.Equal(t, uint64(0), agent.fsStats["tank"].TotalRead)
+}

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

@@ -729,6 +729,7 @@ func TestGetDockerStatsChecksDockerVersionAfterContainerList(t *testing.T) {
 
 			stats, err := dm.getDockerStats(defaultCacheTimeMs)
 			require.NoError(t, err)
+			require.NotNil(t, stats, "A successful empty snapshot must remain distinguishable from a collection failure")
 			assert.Empty(t, stats)
 			assert.True(t, dm.dockerVersionChecked)
 			assert.Equal(t, tt.expectedGood, dm.goodDockerVersion)
@@ -742,6 +743,7 @@ func TestGetDockerStatsChecksDockerVersionAfterContainerList(t *testing.T) {
 
 			stats, err = dm.getDockerStats(defaultCacheTimeMs)
 			require.NoError(t, err)
+			require.NotNil(t, stats, "A successful empty snapshot must remain distinguishable from a collection failure")
 			assert.Empty(t, stats)
 			assert.Equal(t, tt.expectedGood, dm.goodDockerVersion)
 			assert.Equal(t, tt.expectedPodman, dm.usingPodman)
@@ -1182,7 +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},
@@ -1674,7 +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/fans.go

@@ -72,14 +72,30 @@ func discoverHwmonFans(root string) ([]fanSensor, error) {
 	var sensors []fanSensor
 	for _, entry := range entries {
 		chipDir := filepath.Join(root, entry.Name())
-		chipName := utils.ReadStringFile(filepath.Join(chipDir, "name"))
+		sensorDir := chipDir
+		inputs, _ := filepath.Glob(filepath.Join(sensorDir, "fan*_input"))
+
+		// Some legacy hwmon drivers (notably applesmc) register a hwmon class
+		// device but create fan attributes on the parent platform device. In
+		// sysfs that parent is exposed through hwmonN/device.
+		if len(inputs) == 0 {
+			deviceDir := filepath.Join(chipDir, "device")
+			if deviceInputs, _ := filepath.Glob(filepath.Join(deviceDir, "fan*_input")); len(deviceInputs) > 0 {
+				sensorDir = deviceDir
+				inputs = deviceInputs
+			}
+		}
+
+		chipName := utils.ReadStringFile(filepath.Join(sensorDir, "name"))
+		if chipName == "" {
+			chipName = utils.ReadStringFile(filepath.Join(chipDir, "name"))
+		}
 		if chipName == "" {
 			chipName = entry.Name()
 		}
-		inputs, _ := filepath.Glob(filepath.Join(chipDir, "fan*_input"))
 		for _, inputPath := range inputs {
 			base := strings.TrimSuffix(filepath.Base(inputPath), "_input")
-			label := utils.ReadStringFile(filepath.Join(chipDir, base+"_label"))
+			label := utils.ReadStringFile(filepath.Join(sensorDir, base+"_label"))
 			key := chipName + "_" + base
 			if label != "" {
 				key = chipName + "_" + label

agent/fans_test.go

@@ -50,6 +50,24 @@ func TestReadHwmonFans(t *testing.T) {
 	}, fans)
 }
 
+// TestReadHwmonFansLegacyParent verifies legacy hwmon layouts such as applesmc,
+// where the hwmon class node exists but fan attributes live on hwmonN/device.
+func TestReadHwmonFansLegacyParent(t *testing.T) {
+	root := t.TempDir()
+	deviceDir := filepath.Join(root, "devices", "applesmc.768")
+	writeFile(t, filepath.Join(deviceDir, "name"), "applesmc\n")
+	writeFile(t, filepath.Join(deviceDir, "fan1_input"), "1202\n")
+	writeFile(t, filepath.Join(deviceDir, "fan1_label"), "Exhaust\n")
+
+	chipDir := filepath.Join(root, "hwmon1")
+	require.NoError(t, os.MkdirAll(chipDir, 0o755))
+	require.NoError(t, os.Symlink(deviceDir, filepath.Join(chipDir, "device")))
+
+	fans, err := readHwmonFans(root)
+	require.NoError(t, err)
+	assert.Equal(t, map[string]uint16{"applesmc_Exhaust": 1202}, fans)
+}
+
 // TestReadHwmonFansMissingRoot returns an error rather than panicking when the
 // hwmon root doesn't exist (e.g. running on a kernel without hwmon support).
 func TestReadHwmonFansMissingRoot(t *testing.T) {

agent/fingerprint.go

@@ -50,6 +50,9 @@ func generateFingerprint(hostname, cpuModel string) string {
 			if info, err := cpu.Info(); err == nil && len(info) > 0 {
 				cpuModel = info[0].ModelName
 			}
+			if cpuModel == "" {
+				cpuModel = getCpuModelFromCpuinfo()
+			}
 		}
 		fingerprint = hostname + cpuModel
 	}

agent/gpu.go

@@ -361,12 +361,16 @@ func (gm *GPUManager) calculateGPUAverage(id string, gpu *system.GPUData, cacheK
 
 	// If no new data arrived
 	if deltaCount == 0 {
-		// If GPU appears suspended (instantaneous values are 0), return zero values
-		// Otherwise return last known average for temporary collection gaps
-		if gpu.Temperature == 0 && gpu.MemoryUsed == 0 {
+		// Only discrete GPUs report temp/memory, so treat all-zero as suspended (return zeros).
+		// Engine-based (Intel) GPUs don't, so carry the last average forward across sample gaps.
+		if gpu.Engines == nil && gpu.Temperature == 0 && gpu.MemoryUsed == 0 {
 			return system.GPUData{Name: gpu.Name}
 		}
-		return gm.lastAvgData[id] // zero value if not found
+		lastAvg := gm.lastAvgData[id] // zero value if not found
+		if lastAvg.Name == "" {
+			lastAvg.Name = gpu.Name
+		}
+		return lastAvg
 	}
 
 	// Calculate new average

agent/gpu_test.go

@@ -566,6 +566,42 @@ func TestGetCurrentData(t *testing.T) {
 		assert.EqualValues(t, 2, gm.GpuDataMap["0"].Count, "Count should still be 2")
 	})
 
+	t.Run("carries Intel GPU average forward between samples", func(t *testing.T) {
+		// Intel GPUs report no temp/memory, so between-sample gaps (delta 0) must
+		// reuse the last average instead of returning zeros and blanking the chart.
+		gm := &GPUManager{
+			GpuDataMap: map[string]*system.GPUData{
+				"0": {
+					Name:     "GPU",
+					Usage:    0,   // derived from engines for Intel
+					Power:    200, // averages to 100 over 2 counts
+					PowerPkg: 60,  // averages to 30 over 2 counts
+					Count:    2,
+					Engines: map[string]float64{
+						"Render/3D": 80, // averages to 40
+						"Video":     20, // averages to 10
+					},
+				},
+			},
+		}
+
+		cacheKey := uint16(1000) // realtime cache key
+
+		// First collection - computes and stores averages
+		result1 := gm.GetCurrentData(cacheKey)
+		assert.InDelta(t, 100.0, result1["0"].Power, 0.01)
+		assert.InDelta(t, 30.0, result1["0"].PowerPkg, 0.01)
+		assert.InDelta(t, 40.0, result1["0"].Engines["Render/3D"], 0.01)
+
+		// Second collection with no new sample (count unchanged, temp/mem still 0).
+		// Must carry the last average forward rather than blanking to zero.
+		result2 := gm.GetCurrentData(cacheKey)
+		assert.Equal(t, "GPU", result2["0"].Name, "Name should be preserved")
+		assert.InDelta(t, 100.0, result2["0"].Power, 0.01, "Should reuse last average power, not 0")
+		assert.InDelta(t, 30.0, result2["0"].PowerPkg, 0.01, "Should reuse last average package power, not 0")
+		assert.InDelta(t, 40.0, result2["0"].Engines["Render/3D"], 0.01, "Should reuse last average engine usage")
+	})
+
 	t.Run("tracks separate averages per cache key", func(t *testing.T) {
 		gm := &GPUManager{
 			GpuDataMap: map[string]*system.GPUData{
@@ -1083,7 +1119,6 @@ func TestCalculateGPUAverage(t *testing.T) {
 }
 
 func TestGPUCapabilitiesAndLegacyPriority(t *testing.T) {
-	// Save original PATH
 	hasAmdSysfs := (&GPUManager{}).hasAmdSysfs()
 
 	tests := []struct {
@@ -1177,7 +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 +52,8 @@ 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
 }
@@ -178,6 +181,23 @@ func (h *GetSmartDataHandler) Handle(hctx *HandlerContext) error {
 	}, hctx.RequestID)
 }
 
+////////////////////////////////////////////////////////////////////////////
+////////////////////////////////////////////////////////////////////////////
+
+// GetZfsDataHandler handles ZFS detail data requests
+type GetZfsDataHandler struct{}
+
+func (h *GetZfsDataHandler) Handle(hctx *HandlerContext) error {
+	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.storagePoolManager.GetDetail(req.Force), hctx.RequestID)
+}
+
 ////////////////////////////////////////////////////////////////////////////
 ////////////////////////////////////////////////////////////////////////////
 ////////////////////////////////////////////////////////////////////////////
@@ -205,3 +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

@@ -4,8 +4,10 @@ package agent
 
 import (
 	"testing"
+	"time"
 
 	"github.com/fxamacker/cbor/v2"
+	"github.com/henrygd/beszel/agent/zfs"
 	"github.com/henrygd/beszel/internal/common"
 	"github.com/henrygd/beszel/internal/entities/smart"
 	"github.com/stretchr/testify/assert"
@@ -30,6 +32,32 @@ func TestNewAgentResponseSmartData(t *testing.T) {
 	assert.True(t, response.SmartComplete)
 }
 
+func TestGetZfsDataHandlerForceRefresh(t *testing.T) {
+	poolCalls := 0
+	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.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{storagePoolManager: zm},
+		Request: &common.HubRequest[cbor.RawMessage]{
+			Action: common.GetZfsData,
+			Data:   requestData,
+		},
+		SendResponse: func(any, *uint32) error { return nil },
+	}
+
+	assert.NoError(t, (&GetZfsDataHandler{}).Handle(ctx))
+	assert.Equal(t, 2, poolCalls)
+}
+
 func (m *MockHandler) Handle(ctx *HandlerContext) error {
 	if m.handleFunc != nil {
 		return m.handleFunc(ctx)

agent/mdraid_linux.go

@@ -17,15 +17,17 @@ import (
 var mdraidSysfsRoot = "/sys"
 
 type mdraidHealth struct {
-	level         string
-	arrayState    string
-	degraded      uint64
-	raidDisks     uint64
-	syncAction    string
-	syncCompleted string
-	syncSpeed     string
-	mismatchCnt   uint64
-	capacity      uint64
+	level          string
+	arrayState     string
+	degraded       uint64
+	faultyDisks    uint64
+	populatedDisks uint64
+	raidDisks      uint64
+	syncAction     string
+	syncCompleted  string
+	syncSpeed      string
+	mismatchCnt    uint64
+	capacity       uint64
 }
 
 // scanMdraidDevices discovers Linux md arrays exposed in sysfs.
@@ -92,6 +94,9 @@ func (sm *SmartManager) collectMdraidHealth(deviceInfo *DeviceInfo) (bool, error
 	if health.degraded > 0 {
 		attrs = append(attrs, &smart.SmartAttribute{Name: "Degraded", RawValue: health.degraded})
 	}
+	if health.faultyDisks > 0 {
+		attrs = append(attrs, &smart.SmartAttribute{Name: "FaultyDisks", RawValue: health.faultyDisks})
+	}
 	if health.syncAction != "" {
 		attrs = append(attrs, &smart.SmartAttribute{Name: "SyncAction", RawString: health.syncAction})
 	}
@@ -152,6 +157,7 @@ func readMdraidHealth(blockName string) (mdraidHealth, bool) {
 	if val, ok := utils.ReadUintFile(filepath.Join(mdDir, "degraded")); ok {
 		out.degraded = val
 	}
+	out.faultyDisks, out.populatedDisks = countMdraidMemberStates(blockName, mdraidSysfsRoot)
 	if val, ok := utils.ReadUintFile(filepath.Join(mdDir, "mismatch_cnt")); ok {
 		out.mismatchCnt = val
 	}
@@ -177,7 +183,19 @@ func mdraidSmartStatus(health mdraidHealth) string {
 	case "resync", "recover", "reshape":
 		return "WARNING"
 	}
+	// Use actual faulty member count rather than the degraded counter, which
+	// equals raid_disks minus active_disks. On QNAP systems raid_disks may be
+	// set to a large value (e.g. 32) while only a few slots are ever used,
+	// making degraded misleadingly large despite zero failed disks.
+	if health.faultyDisks > 0 {
+		return "FAILED"
+	}
 	if health.degraded > 0 {
+		if isSparseSlotDegraded(health) {
+			// A sysfs snapshot cannot distinguish reserved slots from a removed
+			// member on sparse arrays, so report the ambiguity as a warning.
+			return "WARNING"
+		}
 		return "FAILED"
 	}
 	if health.mismatchCnt > 0 {
@@ -196,6 +214,43 @@ func mdraidSmartStatus(health mdraidHealth) string {
 	return "UNKNOWN"
 }
 
+// countMdraidMemberStates reads member device directories under
+// block/<name>/md and returns how many are explicitly marked "faulty", plus
+// how many are populated at all (regardless of state). populatedDisks lets
+// callers distinguish RAID slots that were never used (QNAP reserves far
+// more raid_disks than it ever populates) from members that went missing.
+func countMdraidMemberStates(blockName, root string) (faultyDisks, populatedDisks uint64) {
+	devDir := filepath.Join(root, "block", blockName, "md")
+	entries, err := os.ReadDir(devDir)
+	if err != nil {
+		return 0, 0
+	}
+	for _, ent := range entries {
+		if !strings.HasPrefix(ent.Name(), "dev-") {
+			continue
+		}
+		populatedDisks++
+		statePath := filepath.Join(devDir, ent.Name(), "state")
+		state := utils.ReadStringFile(statePath)
+		if strings.Contains(state, "faulty") {
+			faultyDisks++
+		}
+	}
+	return faultyDisks, populatedDisks
+}
+
+// isSparseSlotDegraded reports whether a non-zero "degraded" count may be
+// explained by RAID slots that were never populated. QNAP configures system
+// arrays with raid_disks set to a large fixed maximum (e.g. 32) far beyond the
+// handful of slots it ever populates, so sparse slots outnumber populated ones.
+func isSparseSlotDegraded(health mdraidHealth) bool {
+	if health.populatedDisks == 0 || health.raidDisks <= health.populatedDisks {
+		return false
+	}
+	sparseSlots := health.raidDisks - health.populatedDisks
+	return sparseSlots > health.populatedDisks
+}
+
 // isMdraidBlockName matches /dev/mdN-style block device names.
 func isMdraidBlockName(name string) bool {
 	if !strings.HasPrefix(name, "md") {

agent/mdraid_linux_test.go

@@ -40,6 +40,15 @@ func TestMdraidMockSysfsScanAndCollect(t *testing.T) {
 	write(filepath.Join(mdDir, "sync_completed"), "10%\n")
 	write(filepath.Join(mdDir, "sync_speed"), "100M\n")
 	write(filepath.Join(mdDir, "mismatch_cnt"), "0\n")
+
+	// Simulate two healthy member devices (no faulty state).
+	for _, dev := range []string{"dev-sda", "dev-sdb"} {
+		devPath := filepath.Join(mdDir, dev)
+		if err := os.MkdirAll(devPath, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		write(filepath.Join(devPath, "state"), "in_sync\n")
+	}
 	write(filepath.Join(queueDir, "logical_block_size"), "512\n")
 	write(filepath.Join(tmp, "block", "md0", "size"), "2048\n")
 
@@ -81,15 +90,77 @@ func TestMdraidMockSysfsScanAndCollect(t *testing.T) {
 	}
 }
 
+func TestCountMdraidMemberStates(t *testing.T) {
+	tmp := t.TempDir()
+
+	write := func(path, content string) {
+		t.Helper()
+		if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	mdDir := filepath.Join(tmp, "block", "md0", "md")
+
+	// No dev-* entries: zero faulty, zero populated.
+	if faulty, populated := countMdraidMemberStates("md0", tmp); faulty != 0 || populated != 0 {
+		t.Fatalf("no members: got (faulty=%d populated=%d), want (0,0)", faulty, populated)
+	}
+
+	// Two healthy members.
+	write(filepath.Join(mdDir, "dev-sda", "state"), "in_sync\n")
+	write(filepath.Join(mdDir, "dev-sdb", "state"), "in_sync\n")
+	if faulty, populated := countMdraidMemberStates("md0", tmp); faulty != 0 || populated != 2 {
+		t.Fatalf("all in_sync: got (faulty=%d populated=%d), want (0,2)", faulty, populated)
+	}
+
+	// One faulty member.
+	write(filepath.Join(mdDir, "dev-sdb", "state"), "faulty\n")
+	if faulty, populated := countMdraidMemberStates("md0", tmp); faulty != 1 || populated != 2 {
+		t.Fatalf("one faulty: got (faulty=%d populated=%d), want (1,2)", faulty, populated)
+	}
+
+	// QNAP-style: 28 degraded slots but no dev-* entries for them, 4 in_sync.
+	write(filepath.Join(mdDir, "dev-sdb", "state"), "in_sync\n")
+	write(filepath.Join(mdDir, "dev-sdc", "state"), "in_sync\n")
+	write(filepath.Join(mdDir, "dev-sdd", "state"), "in_sync\n")
+	if faulty, populated := countMdraidMemberStates("md0", tmp); faulty != 0 || populated != 4 {
+		t.Fatalf("qnap sparse: got (faulty=%d populated=%d), want (0,4)", faulty, populated)
+	}
+}
+
 func TestMdraidSmartStatus(t *testing.T) {
 	if got := mdraidSmartStatus(mdraidHealth{arrayState: "inactive"}); got != "FAILED" {
 		t.Fatalf("mdraidSmartStatus(inactive) = %q, want FAILED", got)
 	}
-	if got := mdraidSmartStatus(mdraidHealth{arrayState: "active", degraded: 1, syncAction: "recover"}); got != "WARNING" {
+	if got := mdraidSmartStatus(mdraidHealth{arrayState: "active", degraded: 1, faultyDisks: 1, syncAction: "recover"}); got != "WARNING" {
 		t.Fatalf("mdraidSmartStatus(degraded+recover) = %q, want WARNING", got)
 	}
-	if got := mdraidSmartStatus(mdraidHealth{arrayState: "active", degraded: 1}); got != "FAILED" {
-		t.Fatalf("mdraidSmartStatus(degraded) = %q, want FAILED", got)
+	if got := mdraidSmartStatus(mdraidHealth{arrayState: "active", degraded: 1, faultyDisks: 1}); got != "FAILED" {
+		t.Fatalf("mdraidSmartStatus(degraded+faulty) = %q, want FAILED", got)
+	}
+	// QNAP-style: raid_disks=32 but only 4 populated; degraded=28 but no faulty devices.
+	if got := mdraidSmartStatus(mdraidHealth{arrayState: "clean", degraded: 28, faultyDisks: 0, raidDisks: 32, populatedDisks: 4}); got != "WARNING" {
+		t.Fatalf("mdraidSmartStatus(qnap sparse) = %q, want WARNING", got)
+	}
+	// A member disappearing from the same sparse array is indistinguishable
+	// from another reserved slot, so it must not be reported as healthy.
+	if got := mdraidSmartStatus(mdraidHealth{arrayState: "clean", degraded: 29, faultyDisks: 0, raidDisks: 32, populatedDisks: 3}); got != "WARNING" {
+		t.Fatalf("mdraidSmartStatus(qnap sparse missing member) = %q, want WARNING", got)
+	}
+	// A genuinely missing member (removed dev-* entry, not just an unpopulated
+	// QNAP reserve slot) must still fail: raid_disks=4, only 3 populated, all
+	// of them in_sync, so faultyDisks==0 but degraded==1.
+	if got := mdraidSmartStatus(mdraidHealth{arrayState: "clean", degraded: 1, faultyDisks: 0, raidDisks: 4, populatedDisks: 3}); got != "FAILED" {
+		t.Fatalf("mdraidSmartStatus(missing member) = %q, want FAILED", got)
+	}
+	// Degraded with no member-state info at all (e.g. sysfs read failed) must
+	// still fail rather than being silently treated as a sparse QNAP array.
+	if got := mdraidSmartStatus(mdraidHealth{arrayState: "clean", degraded: 1, faultyDisks: 0, raidDisks: 4, populatedDisks: 0}); got != "FAILED" {
+		t.Fatalf("mdraidSmartStatus(degraded, no member info) = %q, want FAILED", got)
 	}
 	if got := mdraidSmartStatus(mdraidHealth{arrayState: "active", syncAction: "recover"}); got != "WARNING" {
 		t.Fatalf("mdraidSmartStatus(recover) = %q, want WARNING", got)

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/sensors_test.go

@@ -602,8 +602,9 @@ func TestUpdateTemperaturesSkipsOnTimeout(t *testing.T) {
 		},
 	}
 
+	originalGetSensorTemps := getSensorTemps
 	t.Cleanup(func() {
-		getSensorTemps = sensors.TemperaturesWithContext
+		getSensorTemps = originalGetSensorTemps
 	})
 	getSensorTemps = func(ctx context.Context) ([]sensors.TemperatureStat, error) {
 		time.Sleep(50 * time.Millisecond)

agent/sensors_windows.go

@@ -214,9 +214,12 @@ func (lhm *lhmProcess) getTemps(ctx context.Context) (temps []sensors.Temperatur
 	return temps, nil
 }
 
-// getSensorTemps attempts to pull sensor temperatures from the embedded LHM process.
+// getSensorTemps is a variable so tests can replace the platform sensor collector.
+var getSensorTemps = getWindowsSensorTemps
+
+// getWindowsSensorTemps attempts to pull sensor temperatures from the embedded LHM process.
 // NB: LibreHardwareMonitorLib requires admin privileges to access all available sensors.
-func getSensorTemps(ctx context.Context) (temps []sensors.TemperatureStat, err error) {
+func getWindowsSensorTemps(ctx context.Context) (temps []sensors.TemperatureStat, err error) {
 	defer func() {
 		if err != nil {
 			slog.Debug("Error reading sensors", "err", err)

agent/server.go

@@ -265,6 +265,5 @@ func (a *Agent) StopServer() error {
 	slog.Info("Stopping SSH server")
 	_ = a.server.Close()
 	a.server = nil
-	a.connectionManager.eventChan <- SSHDisconnect
 	return nil
 }

agent/server_test.go

@@ -198,6 +198,28 @@ func TestStartServerDisableSSH(t *testing.T) {
 	assert.Contains(t, err.Error(), "SSH disabled")
 }
 
+func TestStopServerDoesNotBlockWhenEventQueueFull(t *testing.T) {
+	agent := createTestAgent(t)
+	agent.server = &ssh.Server{}
+	agent.connectionManager.eventChan = make(chan ConnectionEvent, 1)
+	agent.connectionManager.eventChan <- WebSocketConnect
+
+	done := make(chan error, 1)
+	go func() {
+		done <- agent.StopServer()
+	}()
+
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(time.Second):
+		t.Fatal("StopServer blocked on the connection event queue")
+	}
+
+	assert.Nil(t, agent.server)
+	assert.Equal(t, WebSocketConnect, <-agent.connectionManager.eventChan)
+}
+
 /////////////////////////////////////////////////////////////////
 //////////////////// ParseKeys Tests ////////////////////////////
 /////////////////////////////////////////////////////////////////

agent/smart_test.go

@@ -4,6 +4,7 @@ package agent
 
 import (
 	"errors"
+	"fmt"
 	"os"
 	"path/filepath"
 	"testing"
@@ -88,6 +89,31 @@ func TestParseSmartForSata(t *testing.T) {
 	}
 }
 
+func TestParseSmartForSataPreservesFailedAndUnknownStatus(t *testing.T) {
+	for _, test := range []struct {
+		name        string
+		temperature int
+		want        string
+	}{
+		{name: "failed", temperature: 30, want: "FAILED"},
+		{name: "unknown", want: "UNKNOWN"},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			jsonPayload := []byte(fmt.Sprintf(`{
+				"device": {"name": "/dev/sda", "type": "sat"},
+				"serial_number": "PRESERVE%s",
+				"temperature": {"current": %d},
+				"ata_smart_attributes": {"table": [{"id": 197, "raw": {"value": 1, "string": "1"}}]}
+			}`, test.name, test.temperature))
+
+			sm := &SmartManager{SmartDataMap: make(map[string]*smart.SmartData)}
+			hasData, _ := sm.parseSmartForSata(jsonPayload, "")
+			require.True(t, hasData)
+			assert.Equal(t, test.want, sm.SmartDataMap["PRESERVE"+test.name].SmartStatus)
+		})
+	}
+}
+
 func TestParseSmartForSataDeviceStatisticsTemperature(t *testing.T) {
 	jsonPayload := []byte(`{
 		"smartctl": {"exit_status": 0},

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

@@ -4,6 +4,7 @@ import (
 	"bufio"
 	"errors"
 	"fmt"
+	"io"
 	"log/slog"
 	"os"
 	"runtime"
@@ -11,6 +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"
@@ -31,7 +33,11 @@ func (a *Agent) refreshSystemDetails() {
 
 	if a.dockerManager != nil {
 		a.systemDetails.Podman = a.dockerManager.IsPodman()
-		hostInfo, _ = a.dockerManager.GetHostInfo()
+		// Docker's host info describes the machine its daemon runs on. On macOS and
+		// Windows that is a Linux VM, so its CPU and memory totals are not this host's.
+		if runtime.GOOS != "darwin" && runtime.GOOS != "windows" {
+			hostInfo, _ = a.dockerManager.GetHostInfo()
+		}
 	}
 
 	a.systemDetails.Hostname, _ = os.Hostname()
@@ -78,6 +84,12 @@ func (a *Agent) refreshSystemDetails() {
 	if info, err := cpu.Info(); err == nil && len(info) > 0 {
 		a.systemDetails.CpuModel = info[0].ModelName
 	}
+	// gopsutil doesn't parse the "cpu model" field from /proc/cpuinfo, which
+	// is the only source of the CPU model name on MIPS. Fall back to reading
+	// it directly when ModelName is empty.
+	if a.systemDetails.CpuModel == "" {
+		a.systemDetails.CpuModel = getCpuModelFromCpuinfo()
+	}
 	// cores / threads
 	cores, _ := cpu.Counts(false)
 	threads := hostInfo.NCPU
@@ -164,9 +176,9 @@ func (a *Agent) getSystemStats(cacheTimeMs uint16) system.Stats {
 
 	// load average
 	if avgstat, err := load.Avg(); err == nil {
-		systemStats.LoadAvg[0] = avgstat.Load1
-		systemStats.LoadAvg[1] = avgstat.Load5
-		systemStats.LoadAvg[2] = avgstat.Load15
+		systemStats.LoadAvg[0] = utils.TwoDecimals(avgstat.Load1)
+		systemStats.LoadAvg[1] = utils.TwoDecimals(avgstat.Load5)
+		systemStats.LoadAvg[2] = utils.TwoDecimals(avgstat.Load15)
 		slog.Debug("Load average", "5m", avgstat.Load5, "15m", avgstat.Load15)
 	} else {
 		slog.Error("Error getting load average", "err", err)
@@ -208,6 +220,10 @@ func (a *Agent) getSystemStats(cacheTimeMs uint16) system.Stats {
 	// disk i/o (cache-aware per interval)
 	a.updateDiskIo(cacheTimeMs, &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)
 
@@ -258,13 +274,66 @@ func (a *Agent) getSystemStats(cacheTimeMs uint16) system.Stats {
 	a.systemInfo.MemPct = systemStats.MemPct
 	a.systemInfo.DiskPct = systemStats.DiskPct
 	a.systemInfo.Battery = systemStats.Battery
-	a.systemInfo.Uptime, _ = host.Uptime()
+	a.systemInfo.Uptime, _ = getUptime()
 	a.systemInfo.BandwidthBytes = systemStats.Bandwidth[0] + systemStats.Bandwidth[1]
 	a.systemInfo.Threads = a.systemDetails.Threads
 
 	return systemStats
 }
 
+// cpuModelFallbackKeys are the field names to look for in /proc/cpuinfo when
+// gopsutil fails to return a ModelName. The "cpu model" key is used on MIPS
+// (e.g. "MIPS 1004Kc V2.15"), while "system type" provides SoC information
+// on various embedded architectures.
+var cpuModelFallbackKeys = []string{"cpu model", "system type"}
+
+// getCpuModelFromCpuinfo reads /proc/cpuinfo and returns a CPU model string.
+// This is a fallback for architectures where gopsutil's cpu.Info() does not
+// populate ModelName, most notably MIPS.
+func getCpuModelFromCpuinfo() string {
+	file, err := os.Open("/proc/cpuinfo")
+	if err != nil {
+		return ""
+	}
+	defer file.Close()
+	return parseCpuModel(file)
+}
+
+// parseCpuModel scans r (expected to be /proc/cpuinfo content) and returns
+// a combined CPU model string. It collects values from all matching keys
+// and joins them with " / " when multiple are found.
+func parseCpuModel(r io.Reader) string {
+	lines := readLines(r)
+	var parts []string
+	for _, key := range cpuModelFallbackKeys {
+		for _, line := range lines {
+			after, found := strings.CutPrefix(line, key)
+			if !found {
+				continue
+			}
+			after = strings.TrimSpace(after)
+			if len(after) < 2 || after[0] != ':' {
+				continue
+			}
+			if value := strings.TrimSpace(after[1:]); value != "" {
+				parts = append(parts, value)
+				break
+			}
+		}
+	}
+	return strings.Join(parts, " / ")
+}
+
+// readLines reads all lines from r into a slice.
+func readLines(r io.Reader) []string {
+	scanner := bufio.NewScanner(r)
+	var lines []string
+	for scanner.Scan() {
+		lines = append(lines, scanner.Text())
+	}
+	return lines
+}
+
 // calculateHostMemoryUsage derives counters defensively because /proc/meminfo may
 // change while gopsutil reads it. Invalid unsigned subtractions saturate at zero.
 func calculateHostMemoryUsage(v *mem.VirtualMemoryStat, htop bool) (used, cacheBuff, swapUsed uint64) {
@@ -283,7 +352,8 @@ func calculateHostMemoryUsage(v *mem.VirtualMemoryStat, htop bool) (used, cacheB
 	if htop {
 		used = saturatingSub(v.Total, v.Free, cacheBuff)
 	}
-	return used, cacheBuff, saturatingSub(v.SwapTotal, v.SwapFree, v.SwapCached)
+	// Cached swap pages still occupy swap slots and are included in `free`'s used value.
+	return used, cacheBuff, saturatingSub(v.SwapTotal, v.SwapFree)
 }
 
 // saturatingSub subtracts each value, returning zero on underflow.

agent/system_test.go

@@ -1,6 +1,7 @@
 package agent
 
 import (
+	"strings"
 	"testing"
 
 	"github.com/henrygd/beszel/internal/common"
@@ -46,14 +47,14 @@ func TestCalculateHostMemoryUsage(t *testing.T) {
 			memory:    mem.VirtualMemoryStat{Total: 100, Available: 40, Used: 60, Free: 20, Cached: 25, Buffers: 10, Shared: 5, SwapTotal: 20, SwapFree: 8, SwapCached: 2},
 			used:      60,
 			cacheBuff: 30,
-			swapUsed:  10,
+			swapUsed:  12,
 		},
 		{
 			name:      "inconsistent counters saturate",
 			memory:    mem.VirtualMemoryStat{Total: 100, Available: 110, Used: ^uint64(0) - 9, Free: 90, Cached: 5, Buffers: 10, Shared: 20, SwapTotal: 10, SwapFree: 9, SwapCached: 2},
 			used:      0,
 			cacheBuff: 0,
-			swapUsed:  0,
+			swapUsed:  1,
 		},
 		{
 			name:      "htop subtraction saturates",
@@ -61,7 +62,7 @@ func TestCalculateHostMemoryUsage(t *testing.T) {
 			htop:      true,
 			used:      0,
 			cacheBuff: 25,
-			swapUsed:  15,
+			swapUsed:  20,
 		},
 		{
 			name:      "zero cache from shared cancellation does not fall back",
@@ -113,3 +114,81 @@ func TestUpdateSystemDetailsMarksDetailsDirty(t *testing.T) {
 	assert.False(t, agent.detailsDirty)
 	assert.Nil(t, original.Details)
 }
+
+func TestParseCpuModel(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected string
+	}{
+		{
+			name: "MIPS with both cpu model and system type",
+			input: `system type             : MediaTek MT7621 ver:1 eco:3
+machine                 : ASUS RT-AX53U
+processor               : 0
+cpu model               : MIPS 1004Kc V2.15
+BogoMIPS                : 586.13
+wait instruction        : yes`,
+			expected: "MIPS 1004Kc V2.15 / MediaTek MT7621 ver:1 eco:3",
+		},
+		{
+			name: "MIPS with different SoC",
+			input: `system type             : Atheros AR7161 rev 2
+machine                 : NETGEAR WNDR3700
+processor               : 0
+cpu model               : MIPS 24Kc V7.4
+BogoMIPS                : 452.19`,
+			expected: "MIPS 24Kc V7.4 / Atheros AR7161 rev 2",
+		},
+		{
+			name: "only system type when cpu model missing",
+			input: `system type             : Broadcom BCM47xx
+processor               : 0
+BogoMIPS                : 296.11`,
+			expected: "Broadcom BCM47xx",
+		},
+		{
+			name: "only cpu model when system type missing",
+			input: `processor               : 0
+cpu model               : MIPS 34Kc V2.15
+BogoMIPS                : 300.00`,
+			expected: "MIPS 34Kc V2.15",
+		},
+		{
+			name: "x86 cpuinfo returns empty",
+			input: `processor	: 0
+vendor_id	: GenuineIntel
+cpu family	: 6
+model		: 142
+model name	: Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz
+stepping	: 10`,
+			expected: "",
+		},
+		{
+			name:     "empty input",
+			input:    "",
+			expected: "",
+		},
+		{
+			name: "cpu model with extra whitespace",
+			input: `processor               : 0
+cpu model               :   MIPS 34Kc V2.15
+BogoMIPS                : 300.00`,
+			expected: "MIPS 34Kc V2.15",
+		},
+		{
+			name: "cpu model without value",
+			input: `processor               : 0
+cpu model               :
+BogoMIPS                : 300.00`,
+			expected: "",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := parseCpuModel(strings.NewReader(tt.input))
+			assert.Equal(t, tt.expected, result)
+		})
+	}
+}

agent/test-data/zfs/zfs_list.txt

@@ -0,0 +1,9 @@
+tank	12000000000000	11999000000000	/tank
+tank/apps	1000000000000	11999000000000	/tank/apps
+tank/backup	2000000000000	11999000000000	/tank/backup
+tank/media	1000000000000	11999000000000	/tank/my media
+rpool	900000000000	300000000000	-
+rpool/ROOT	1000000000	300000000000	-
+rpool/ROOT/pve-1	890000000000	300000000000	/
+rpool/data	9000000000	300000000000	-
+rpool/data/subvol-100-disk-0	400000000000	300000000000	/subvol-100-disk-0

agent/test-data/zfs/zpool_list.txt

@@ -0,0 +1,2 @@
+tank	23999000000000	12000000000000	11999000000000	ONLINE
+rpool	1200000000000	900000000000	300000000000	DEGRADED

agent/test-data/zfs/zpool_status.txt

@@ -0,0 +1,29 @@
+  pool: tank
+ state: ONLINE
+  scan: scrub repaired 0B in 00:05:12 with 0 errors on Sun Jun  1 02:00:12 2025
+config:
+
+	NAME        STATE     READ WRITE CKSUM
+	tank        ONLINE       0     0     0
+	  mirror-0  ONLINE       0     0     0
+	    sda     ONLINE       0     0     0
+	    sdb     ONLINE       0     0     0
+
+errors: No known data errors
+
+  pool: rpool
+ state: DEGRADED
+ status: One or more devices could not be used because the label is missing or
+	invalid.  Sufficient replicas exist for the pool to continue functioning in a
+	degraded state.
+  scan: scrub in progress since Sun Jun  8 01:00:00 2025
+	10.00% done, 01:30:00 to go, 0.00/s
+config:
+
+	NAME        STATE     READ WRITE CKSUM
+	rpool       DEGRADED     0     0     0
+	  mirror-0  DEGRADED     0     0     0
+	    sda     ONLINE       0     0     0
+	    sdb     FAULTED      1     2     3
+
+errors: 1 data errors, use '-v' for a list

agent/uptime_linux.go

@@ -0,0 +1,44 @@
+//go:build linux
+
+package agent
+
+import (
+	"math"
+	"os"
+	"strconv"
+	"strings"
+
+	"github.com/shirou/gopsutil/v4/host"
+)
+
+// uptimeFilePath is a variable so tests can point it at a fixture.
+var uptimeFilePath = "/proc/uptime"
+
+// getUptime returns the system uptime in seconds.
+//
+// This reads /proc/uptime instead of using host.Uptime(), which calls the
+// sysinfo(2) syscall. Inside an LXC container lxcfs virtualizes /proc/uptime
+// but cannot intercept a syscall, so sysinfo(2) reports the host's uptime
+// rather than the container's.
+//
+// Falls back to host.Uptime() if /proc/uptime is missing or unparseable, so
+// behavior is unchanged anywhere the file isn't available.
+func getUptime() (uint64, error) {
+	data, err := os.ReadFile(uptimeFilePath)
+	if err != nil {
+		return host.Uptime()
+	}
+	fields := strings.Fields(string(data))
+	if len(fields) == 0 {
+		return host.Uptime()
+	}
+	seconds, err := strconv.ParseFloat(fields[0], 64)
+	if err != nil ||
+		math.IsNaN(seconds) ||
+		math.IsInf(seconds, 0) ||
+		seconds < 0 ||
+		seconds >= 1<<64 {
+		return host.Uptime()
+	}
+	return uint64(seconds), nil
+}

agent/uptime_linux_test.go

@@ -0,0 +1,101 @@
+//go:build linux
+
+package agent
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestGetUptimeFromProc(t *testing.T) {
+	tests := []struct {
+		name     string
+		contents string
+		want     uint64
+	}{
+		{"typical", "12345.67 98765.43\n", 12345},
+		{"zero", "0.00 0.00\n", 0},
+		{"no trailing newline", "42.99 7.00", 42},
+		{"single field", "600.5", 600},
+		{"large value", "266030.12 1000000.00\n", 266030},
+	}
+
+	prev := uptimeFilePath
+	t.Cleanup(func() { uptimeFilePath = prev })
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			path := filepath.Join(t.TempDir(), "uptime")
+			if err := os.WriteFile(path, []byte(tt.contents), 0o644); err != nil {
+				t.Fatal(err)
+			}
+			uptimeFilePath = path
+
+			got, err := getUptime()
+			if err != nil {
+				t.Fatalf("getUptime() returned error: %v", err)
+			}
+			if got != tt.want {
+				t.Errorf("getUptime() = %d, want %d", got, tt.want)
+			}
+		})
+	}
+}
+
+func writeUptime(contents string) func(t *testing.T) string {
+	return func(t *testing.T) string {
+		path := filepath.Join(t.TempDir(), "uptime")
+		if err := os.WriteFile(path, []byte(contents), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		return path
+	}
+}
+
+// Malformed, missing, or out-of-range input must fall back to host.Uptime()
+// rather than returning a bogus value, so the agent still reports something sane.
+func TestGetUptimeFallsBack(t *testing.T) {
+	prev := uptimeFilePath
+	t.Cleanup(func() { uptimeFilePath = prev })
+
+	for _, tt := range []struct {
+		name    string
+		prepare func(t *testing.T) string
+	}{
+		{"missing file", func(t *testing.T) string {
+			return filepath.Join(t.TempDir(), "does-not-exist")
+		}},
+		{"empty file", func(t *testing.T) string {
+			path := filepath.Join(t.TempDir(), "uptime")
+			if err := os.WriteFile(path, nil, 0o644); err != nil {
+				t.Fatal(err)
+			}
+			return path
+		}},
+		{"unparseable", func(t *testing.T) string {
+			path := filepath.Join(t.TempDir(), "uptime")
+			if err := os.WriteFile(path, []byte("not-a-number 1.0\n"), 0o644); err != nil {
+				t.Fatal(err)
+			}
+			return path
+		}},
+		{"NaN", writeUptime("NaN 1.0\n")},
+		{"positive infinity", writeUptime("+Inf 1.0\n")},
+		{"negative infinity", writeUptime("-Inf 1.0\n")},
+		{"negative", writeUptime("-42.5 1.0\n")},
+		{"exceeds uint64 range", writeUptime("1e20 1.0\n")},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			uptimeFilePath = tt.prepare(t)
+
+			got, err := getUptime()
+			if err != nil {
+				t.Fatalf("getUptime() returned error: %v", err)
+			}
+			if got == 0 {
+				t.Error("getUptime() = 0, expected fallback to host.Uptime()")
+			}
+		})
+	}
+}

agent/uptime_stub.go

@@ -0,0 +1,10 @@
+//go:build !linux
+
+package agent
+
+import "github.com/shirou/gopsutil/v4/host"
+
+// getUptime returns the system uptime in seconds.
+func getUptime() (uint64, error) {
+	return host.Uptime()
+}

agent/zfs/zfs.go

@@ -0,0 +1,170 @@
+// Package zfs provides functions to read ZFS statistics.
+package zfs
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"strconv"
+	"strings"
+	"time"
+)
+
+var commandTimeout = 10 * time.Second
+
+var commandOutput = func(name string, args ...string) ([]byte, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), commandTimeout)
+	defer cancel()
+	cmd := exec.CommandContext(ctx, name, args...)
+	cmd.Env = append(os.Environ(), "LC_ALL=C", "LANG=C")
+	out, err := cmd.Output()
+	if ctx.Err() != nil {
+		return nil, fmt.Errorf("%s timed out after %s: %w", name, commandTimeout, ctx.Err())
+	}
+	return out, err
+}
+
+// ErrNoZfs is returned when the ZFS utilities or kernel interfaces are unavailable.
+var ErrNoZfs = errors.New("zfs utilities unavailable")
+
+// PoolStat is a snapshot of a ZFS pool's capacity and health.
+type PoolStat struct {
+	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.
+// NRead and NWrite are cumulative byte counters since the pool was imported.
+type PoolKernelStat struct {
+	Name   string
+	Health string
+	NRead  uint64
+	NWrite uint64
+}
+
+// PoolIoStats holds calculated per-second I/O rates for a pool.
+type PoolIoStats struct {
+	NRead  uint64
+	NWrite uint64
+}
+
+// Dataset is a single ZFS dataset with usage information.
+type Dataset struct {
+	Name       string
+	Used       uint64
+	Avail      uint64
+	Mountpoint string
+}
+
+// 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
+		if errors.As(err, &exitErr) && strings.Contains(string(exitErr.Stderr), "no pools available") {
+			return nil, nil
+		}
+		return nil, fmt.Errorf("zpool list: %w", err)
+	}
+	return parseZpoolListOutput(out)
+}
+
+// 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)
+	}
+	return parseZfsListOutput(out)
+}
+
+// parseZpoolListOutput parses `zpool list -Hp -o name,size,alloc,free,health` output.
+// Columns are tab-separated; numeric columns are raw bytes.
+func parseZpoolListOutput(out []byte) ([]PoolStat, error) {
+	var pools []PoolStat
+	scanner := bufio.NewScanner(bytes.NewReader(out))
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" {
+			continue
+		}
+		if line == "no pools available" && len(pools) == 0 {
+			return nil, nil
+		}
+		fields := strings.Split(line, "\t")
+		if len(fields) < 5 {
+			return nil, fmt.Errorf("unexpected zpool list line: %q", line)
+		}
+		size, err := strconv.ParseUint(fields[1], 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("parsing size for pool %q: %w", fields[0], err)
+		}
+		alloc, err := strconv.ParseUint(fields[2], 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("parsing alloc for pool %q: %w", fields[0], err)
+		}
+		free, err := strconv.ParseUint(fields[3], 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("parsing free for pool %q: %w", fields[0], err)
+		}
+		pools = append(pools, PoolStat{
+			Name:   fields[0],
+			Size:   size,
+			Alloc:  alloc,
+			Free:   free,
+			Health: fields[4],
+		})
+	}
+	return pools, scanner.Err()
+}
+
+// parseZfsListOutput parses `zfs list -Hp -o name,used,avail,mountpoint` output.
+// The mountpoint column may contain spaces, so it is split on tabs only.
+func parseZfsListOutput(out []byte) ([]Dataset, error) {
+	var datasets []Dataset
+	scanner := bufio.NewScanner(bytes.NewReader(out))
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" {
+			continue
+		}
+		fields := strings.SplitN(line, "\t", 4)
+		if len(fields) < 4 {
+			return nil, fmt.Errorf("unexpected zfs list line: %q", line)
+		}
+		used, err := strconv.ParseUint(fields[1], 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("parsing used for dataset %q: %w", fields[0], err)
+		}
+		avail, err := strconv.ParseUint(fields[2], 10, 64)
+		if err != nil {
+			return nil, fmt.Errorf("parsing avail for dataset %q: %w", fields[0], err)
+		}
+		datasets = append(datasets, Dataset{
+			Name:       fields[0],
+			Used:       used,
+			Avail:      avail,
+			Mountpoint: fields[3],
+		})
+	}
+	return datasets, scanner.Err()
+}

agent/zfs/zfs_freebsd.go

@@ -3,9 +3,17 @@
 package zfs
 
 import (
+	"errors"
+
 	"golang.org/x/sys/unix"
 )
 
 func ARCSize() (uint64, error) {
 	return unix.SysctlUint64("kstat.zfs.misc.arcstats.size")
 }
+
+// FreeBSD does not expose Linux's per-pool procfs kstats. Capacity, health,
+// and detail collection still work through the cached utilities.
+func PoolKernelStats() ([]PoolKernelStat, error) {
+	return nil, errors.ErrUnsupported
+}

agent/zfs/zfs_linux.go

@@ -5,14 +5,21 @@ package zfs
 
 import (
 	"bufio"
+	"errors"
 	"fmt"
 	"os"
+	"path/filepath"
 	"strconv"
 	"strings"
 )
 
+var (
+	procZfsPath = "/proc/spl/kstat/zfs"
+	devZfsPath  = "/dev/zfs"
+)
+
 func ARCSize() (uint64, error) {
-	file, err := os.Open("/proc/spl/kstat/zfs/arcstats")
+	file, err := os.Open(filepath.Join(procZfsPath, "arcstats"))
 	if err != nil {
 		return 0, err
 	}
@@ -29,6 +36,180 @@ func ARCSize() (uint64, error) {
 			return strconv.ParseUint(fields[2], 10, 64)
 		}
 	}
+	if err := scanner.Err(); err != nil {
+		return 0, err
+	}
 
 	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.
+func PoolKernelStats() ([]PoolKernelStat, error) {
+	poolDirs := make(map[string]struct{})
+	for _, filename := range []string{"state", "io", "objset-*"} {
+		paths, err := filepath.Glob(filepath.Join(procZfsPath, "*", filename))
+		if err != nil {
+			return nil, err
+		}
+		for _, path := range paths {
+			poolDirs[filepath.Dir(path)] = struct{}{}
+		}
+	}
+	if len(poolDirs) == 0 {
+		return nil, ErrNoZfs
+	}
+	pools := make([]PoolKernelStat, 0, len(poolDirs))
+	for poolDir := range poolDirs {
+		nread, nwrite, err := readPoolCounters(poolDir)
+		if err != nil {
+			if errors.Is(err, os.ErrNotExist) {
+				continue // pool may have been exported after the glob
+			}
+			return nil, err
+		}
+		state, err := os.ReadFile(filepath.Join(poolDir, "state"))
+		if err != nil && !errors.Is(err, os.ErrNotExist) {
+			return nil, err
+		}
+		pools = append(pools, PoolKernelStat{
+			Name: filepath.Base(poolDir), Health: strings.ToUpper(strings.TrimSpace(string(state))),
+			NRead: nread, NWrite: nwrite,
+		})
+	}
+	if len(pools) == 0 {
+		return nil, ErrNoZfs
+	}
+	return pools, nil
+}
+
+// readPoolCounters supports both ZFS kernel interfaces. OpenZFS through 2.3
+// exposes aggregate vdev counters in "io". When that file is unavailable, sum
+// the logical I/O counters exposed for each dataset in the pool.
+func readPoolCounters(poolDir string) (uint64, uint64, error) {
+	nread, nwrite, err := readPoolIO(filepath.Join(poolDir, "io"))
+	if err == nil || !errors.Is(err, os.ErrNotExist) {
+		return nread, nwrite, err
+	}
+	return readPoolObjsets(poolDir)
+}
+
+func readPoolIO(path string) (uint64, uint64, error) {
+	file, err := os.Open(path)
+	if err != nil {
+		return 0, 0, err
+	}
+	defer file.Close()
+
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		fields := strings.Fields(scanner.Text())
+		if len(fields) < 2 || fields[0] != "nread" {
+			continue
+		}
+		if !scanner.Scan() {
+			break
+		}
+		values := strings.Fields(scanner.Text())
+		if len(values) < 2 {
+			break
+		}
+		nread, err := strconv.ParseUint(values[0], 10, 64)
+		if err != nil {
+			return 0, 0, fmt.Errorf("parsing nread in %s: %w", path, err)
+		}
+		nwrite, err := strconv.ParseUint(values[1], 10, 64)
+		if err != nil {
+			return 0, 0, fmt.Errorf("parsing nwritten in %s: %w", path, err)
+		}
+		return nread, nwrite, nil
+	}
+	if err := scanner.Err(); err != nil {
+		return 0, 0, err
+	}
+	return 0, 0, fmt.Errorf("I/O counters not found in %s", path)
+}
+
+func readPoolObjsets(poolDir string) (uint64, uint64, error) {
+	paths, err := filepath.Glob(filepath.Join(poolDir, "objset-*"))
+	if err != nil {
+		return 0, 0, err
+	}
+	if len(paths) == 0 {
+		return 0, 0, fmt.Errorf("dataset I/O counters not found in %s", poolDir)
+	}
+
+	var totalRead, totalWrite uint64
+	objsetsRead := 0
+	for _, path := range paths {
+		nread, nwrite, err := readObjsetIO(path)
+		if errors.Is(err, os.ErrNotExist) {
+			continue // dataset may have been destroyed after the glob
+		}
+		if err != nil {
+			return 0, 0, err
+		}
+		totalRead += nread
+		totalWrite += nwrite
+		objsetsRead++
+	}
+	if objsetsRead == 0 {
+		return 0, 0, fmt.Errorf("dataset I/O counters not found in %s", poolDir)
+	}
+	return totalRead, totalWrite, nil
+}
+
+func readObjsetIO(path string) (uint64, uint64, error) {
+	file, err := os.Open(path)
+	if err != nil {
+		return 0, 0, err
+	}
+	defer file.Close()
+
+	var nread, nwrite uint64
+	var foundRead, foundWrite bool
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		fields := strings.Fields(scanner.Text())
+		if len(fields) < 3 {
+			continue
+		}
+		var target *uint64
+		switch fields[0] {
+		case "nread":
+			target = &nread
+			foundRead = true
+		case "nwritten":
+			target = &nwrite
+			foundWrite = true
+		default:
+			continue
+		}
+		value, err := strconv.ParseUint(fields[2], 10, 64)
+		if err != nil {
+			return 0, 0, fmt.Errorf("parsing %s in %s: %w", fields[0], path, err)
+		}
+		*target = value
+	}
+	if err := scanner.Err(); err != nil {
+		return 0, 0, err
+	}
+	if !foundRead || !foundWrite {
+		return 0, 0, fmt.Errorf("incomplete I/O counters in %s", path)
+	}
+	return nread, nwrite, nil
+}

agent/zfs/zfs_linux_test.go

@@ -0,0 +1,153 @@
+//go:build testing && linux
+
+package zfs
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestPoolKernelStats(t *testing.T) {
+	root := t.TempDir()
+	oldPath := procZfsPath
+	procZfsPath = root
+	t.Cleanup(func() { procZfsPath = oldPath })
+
+	poolDir := filepath.Join(root, "tank")
+	require.NoError(t, os.MkdirAll(poolDir, 0o755))
+	require.NoError(t, os.WriteFile(filepath.Join(poolDir, "io"), []byte(
+		"11 3 0x00 1 80 0 0\n"+
+			"nread nwritten reads writes wtime wlentime wupdate rtime rlentime rupdate wcnt rcnt\n"+
+			"1884160 6450688 22 978 0 0 0 0 0 0 0 0\n",
+	), 0o644))
+	require.NoError(t, os.WriteFile(filepath.Join(poolDir, "state"), []byte("DEGRADED\n"), 0o644))
+
+	stats, err := PoolKernelStats()
+	require.NoError(t, err)
+	require.Len(t, stats, 1)
+	assert.Equal(t, PoolKernelStat{
+		Name: "tank", Health: "DEGRADED", NRead: 1884160, NWrite: 6450688,
+	}, stats[0])
+}
+
+func TestPoolKernelStatsOpenZfs24(t *testing.T) {
+	root := t.TempDir()
+	oldPath := procZfsPath
+	procZfsPath = root
+	t.Cleanup(func() { procZfsPath = oldPath })
+
+	poolDir := filepath.Join(root, "tank")
+	require.NoError(t, os.MkdirAll(poolDir, 0o755))
+	require.NoError(t, os.WriteFile(filepath.Join(poolDir, "state"), []byte("ONLINE\n"), 0o644))
+	require.NoError(t, os.WriteFile(filepath.Join(poolDir, "objset-0x1"), []byte(
+		"34 1 0x01 28 7872 0 0\n"+
+			"name type data\n"+
+			"dataset_name 7 tank\n"+
+			"nwritten 4 2000\n"+
+			"nread 4 1000\n",
+	), 0o644))
+	require.NoError(t, os.WriteFile(filepath.Join(poolDir, "objset-0x2"), []byte(
+		"34 1 0x01 28 7872 0 0\n"+
+			"name type data\n"+
+			"dataset_name 7 tank/videos\n"+
+			"nwritten 4 400\n"+
+			"nread 4 300\n",
+	), 0o644))
+
+	stats, err := PoolKernelStats()
+	require.NoError(t, err)
+	require.Len(t, stats, 1)
+	assert.Equal(t, PoolKernelStat{
+		Name: "tank", Health: "ONLINE", NRead: 1300, NWrite: 2400,
+	}, stats[0])
+}
+
+func TestPoolKernelStatsNoZfs(t *testing.T) {
+	oldPath := procZfsPath
+	procZfsPath = t.TempDir()
+	t.Cleanup(func() { procZfsPath = oldPath })
+
+	_, err := PoolKernelStats()
+	assert.ErrorIs(t, err, ErrNoZfs)
+}
+
+func TestReadPoolIORejectsMalformedCounters(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "io")
+	require.NoError(t, os.WriteFile(path, []byte("nread nwritten\nnope 10\n"), 0o644))
+	_, _, err := readPoolIO(path)
+	require.Error(t, err)
+}
+
+func TestReadObjsetIORequiresAllCounters(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "objset-0x1")
+	require.NoError(t, os.WriteFile(path, []byte("nread 4 10\n"), 0o644))
+	_, _, 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/zfs_status.go

@@ -0,0 +1,150 @@
+package zfs
+
+import (
+	"bufio"
+	"bytes"
+	"fmt"
+	"regexp"
+	"strconv"
+	"strings"
+)
+
+// PoolStatus holds parsed `zpool status` information for one pool.
+type PoolStatus struct {
+	Name  string
+	State string // ONLINE, DEGRADED, FAULTED, ...
+	Scrub ScrubStatus
+	Vdevs []VdevStatus
+}
+
+// ScrubStatus holds the scrub (or resilver) status parsed from the scan line.
+type ScrubStatus struct {
+	State    string // NONE, SCANNING, FINISHED, CANCELED
+	Progress string // e.g. "10.00%" while scanning
+	Errors   uint64
+}
+
+// VdevStatus is a single vdev row (mirror, raidz, or leaf disk).
+type VdevStatus struct {
+	Name         string
+	State        string
+	ReadErrs     uint64
+	WriteErrs    uint64
+	ChecksumErrs uint64
+}
+
+var (
+	progressRe = regexp.MustCompile(`(\d+\.\d+)%\s+done`)
+	errorsRe   = regexp.MustCompile(`with\s+(\d+)\s+errors`)
+)
+
+// PoolStatuses runs `zpool status` and parses per-pool state, scrub, and vdev
+// information. The human-readable format has been stable across OpenZFS
+// releases; rows are matched by their tabular shape rather than position.
+func PoolStatuses() ([]PoolStatus, error) {
+	out, err := commandOutput("zpool", "status")
+	if err != nil {
+		return nil, fmt.Errorf("zpool status: %w", err)
+	}
+	return parseZpoolStatusOutput(out)
+}
+
+// parseZpoolStatusOutput parses the output of `zpool status`.
+func parseZpoolStatusOutput(out []byte) ([]PoolStatus, error) {
+	var pools []PoolStatus
+	var current *PoolStatus
+	inConfig := false
+	scanContinuation := false // next non-blank line continues the scan line (progress)
+
+	scanner := bufio.NewScanner(bytes.NewReader(out))
+	for scanner.Scan() {
+		line := scanner.Text()
+		trimmed := strings.TrimSpace(line)
+
+		switch {
+		case strings.HasPrefix(trimmed, "pool:"):
+			pools = append(pools, PoolStatus{Name: strings.TrimSpace(strings.TrimPrefix(trimmed, "pool:"))})
+			current = &pools[len(pools)-1]
+			inConfig = false
+			scanContinuation = false
+		case current == nil:
+			continue
+		case strings.HasPrefix(trimmed, "state:"):
+			current.State = strings.TrimSpace(strings.TrimPrefix(trimmed, "state:"))
+		case strings.HasPrefix(trimmed, "scan:"):
+			current.Scrub = parseScanLine(trimmed)
+			// zpool status prints the progress percentage on the line after scan.
+			scanContinuation = true
+		case trimmed == "config:":
+			inConfig = true
+		case scanContinuation:
+			// The line after scan: may be an indented progress continuation.
+			if m := progressRe.FindStringSubmatch(trimmed); m != nil {
+				current.Scrub.Progress = m[1] + "%"
+			}
+			scanContinuation = false
+		case inConfig && (line == "" || strings.HasPrefix(line, " ") || strings.HasPrefix(line, "\t")):
+			// Table rows are indented; blank lines separate sections. The
+			// column header and the pool's own row are skipped.
+			if trimmed != "" && !strings.HasPrefix(trimmed, "NAME") {
+				if vdev, ok := parseVdevLine(trimmed, current.Name); ok {
+					current.Vdevs = append(current.Vdevs, vdev)
+				}
+			}
+		case inConfig:
+			// unindented line (errors:, status:, next pool:) ends the table
+			inConfig = false
+		}
+	}
+	return pools, scanner.Err()
+}
+
+// parseScanLine maps a `scan:` line to a ScrubStatus.
+func parseScanLine(line string) ScrubStatus {
+	var scrub ScrubStatus
+	switch {
+	case strings.Contains(line, "in progress"):
+		scrub.State = "SCANNING"
+	case strings.Contains(line, "canceled"):
+		scrub.State = "CANCELED"
+	case strings.Contains(line, "repaired"), strings.Contains(line, "resilvered"):
+		scrub.State = "FINISHED"
+	default:
+		scrub.State = "NONE"
+	}
+	if m := progressRe.FindStringSubmatch(line); m != nil {
+		scrub.Progress = m[1] + "%"
+	}
+	if m := errorsRe.FindStringSubmatch(line); m != nil {
+		if n, err := strconv.ParseUint(m[1], 10, 64); err == nil {
+			scrub.Errors = n
+		}
+	}
+	return scrub
+}
+
+// parseVdevLine parses one row of the config table. Rows have the shape
+// "NAME STATE READ WRITE CKSUM [extra...]". The first data row is the pool
+// itself and is skipped since it duplicates pool-level info.
+func parseVdevLine(line, poolName string) (VdevStatus, bool) {
+	fields := strings.Fields(line)
+	if len(fields) < 5 {
+		return VdevStatus{}, false
+	}
+	if fields[0] == poolName {
+		return VdevStatus{}, false
+	}
+	read, err1 := strconv.ParseUint(fields[2], 10, 64)
+	write, err2 := strconv.ParseUint(fields[3], 10, 64)
+	cksum, err3 := strconv.ParseUint(fields[4], 10, 64)
+	if err1 != nil || err2 != nil || err3 != nil {
+		return VdevStatus{}, false
+	}
+	return VdevStatus{
+		Name:         fields[0],
+		State:        fields[1],
+		ReadErrs:     read,
+		WriteErrs:    write,
+		ChecksumErrs: cksum,
+	}, true
+}

agent/zfs/zfs_test.go

@@ -0,0 +1,143 @@
+//go:build testing
+
+package zfs
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func fixturePath(name string) string {
+	return filepath.Join("..", "test-data", "zfs", name)
+}
+
+func TestParseZpoolListOutput(t *testing.T) {
+	data, err := os.ReadFile(fixturePath("zpool_list.txt"))
+	require.NoError(t, err)
+
+	pools, err := parseZpoolListOutput(data)
+	require.NoError(t, err)
+	require.Len(t, pools, 2)
+
+	assert.Equal(t, PoolStat{Name: "tank", Size: 23999000000000, Alloc: 12000000000000, Free: 11999000000000, Health: "ONLINE"}, pools[0])
+	assert.Equal(t, PoolStat{Name: "rpool", Size: 1200000000000, Alloc: 900000000000, Free: 300000000000, Health: "DEGRADED"}, pools[1])
+}
+
+func TestParseZpoolListOutputIgnoresEmptyLines(t *testing.T) {
+	pools, err := parseZpoolListOutput([]byte("tank\t100\t50\t50\tONLINE\n\n"))
+	require.NoError(t, err)
+	require.Len(t, pools, 1)
+	assert.Equal(t, "tank", pools[0].Name)
+}
+
+func TestParseZpoolListOutputNoPools(t *testing.T) {
+	pools, err := parseZpoolListOutput([]byte("no pools available\n"))
+	require.NoError(t, err)
+	assert.Empty(t, pools)
+}
+
+func TestParseZpoolListOutputRejectsMalformedLine(t *testing.T) {
+	_, err := parseZpoolListOutput([]byte("tank\t100\t50\n"))
+	require.Error(t, err)
+
+	_, err = parseZpoolListOutput([]byte("tank\tnotanumber\t50\t50\tONLINE\n"))
+	require.Error(t, err)
+}
+
+func TestParseZfsListOutput(t *testing.T) {
+	data, err := os.ReadFile(fixturePath("zfs_list.txt"))
+	require.NoError(t, err)
+
+	datasets, err := parseZfsListOutput(data)
+	require.NoError(t, err)
+	require.Len(t, datasets, 9)
+
+	// Mountpoint with a space must be kept intact (tab-split only).
+	assert.Equal(t, "/tank/my media", datasets[3].Mountpoint)
+	// Unmounted datasets/zvols report "-".
+	assert.Equal(t, "-", datasets[4].Mountpoint)
+	assert.Equal(t, uint64(12000000000000), datasets[0].Used)
+	assert.Equal(t, uint64(11999000000000), datasets[0].Avail)
+}
+
+func TestParseZpoolStatusOutput(t *testing.T) {
+	data, err := os.ReadFile(fixturePath("zpool_status.txt"))
+	require.NoError(t, err)
+
+	pools, err := parseZpoolStatusOutput(data)
+	require.NoError(t, err)
+	require.Len(t, pools, 2)
+
+	tank := pools[0]
+	assert.Equal(t, "tank", tank.Name)
+	assert.Equal(t, "ONLINE", tank.State)
+	assert.Equal(t, "FINISHED", tank.Scrub.State)
+	assert.Equal(t, "", tank.Scrub.Progress)
+	assert.Equal(t, uint64(0), tank.Scrub.Errors)
+	// Pool row itself is skipped; mirror + 2 disks remain.
+	require.Len(t, tank.Vdevs, 3)
+	assert.Equal(t, "mirror-0", tank.Vdevs[0].Name)
+	assert.Equal(t, "sda", tank.Vdevs[1].Name)
+	assert.Equal(t, "sdb", tank.Vdevs[2].Name)
+
+	rpool := pools[1]
+	assert.Equal(t, "rpool", rpool.Name)
+	assert.Equal(t, "DEGRADED", rpool.State)
+	assert.Equal(t, "SCANNING", rpool.Scrub.State)
+	assert.Equal(t, "10.00%", rpool.Scrub.Progress)
+	require.Len(t, rpool.Vdevs, 3)
+	assert.Equal(t, "FAULTED", rpool.Vdevs[2].State)
+	assert.Equal(t, uint64(1), rpool.Vdevs[2].ReadErrs)
+	assert.Equal(t, uint64(2), rpool.Vdevs[2].WriteErrs)
+	assert.Equal(t, uint64(3), rpool.Vdevs[2].ChecksumErrs)
+}
+
+func TestParseScanLine(t *testing.T) {
+	assert.Equal(t, "FINISHED", parseScanLine("scan: scrub repaired 0B in 00:05:12 with 0 errors on Sun Jun  1 02:00:12 2025").State)
+	assert.Equal(t, uint64(3), parseScanLine("scan: scrub repaired 10G in 01:00:00 with 3 errors on Sun Jun  1 02:00:12 2025").Errors)
+	assert.Equal(t, "SCANNING", parseScanLine("scan: scrub in progress since Sun Jun  8 01:00:00 2025").State)
+	assert.Equal(t, "CANCELED", parseScanLine("scan: scrub canceled on Sun Jun  1 02:00:12 2025").State)
+	assert.Equal(t, "FINISHED", parseScanLine("scan: resilvered 1.23G in 00:01:00 with 0 errors on Sun Jun  1 02:00:12 2025").State)
+	assert.Equal(t, "NONE", parseScanLine("scan: none requested").State)
+}
+
+func TestCommandOutputForcesLocaleAndTimesOut(t *testing.T) {
+	t.Setenv("BESZEL_ZFS_COMMAND_HELPER", "1")
+	out, err := commandOutput(os.Args[0], "-test.run=TestZfsCommandHelperProcess", "--", "locale")
+	require.NoError(t, err)
+	assert.Equal(t, "C/C", string(out))
+
+	oldTimeout := commandTimeout
+	commandTimeout = 20 * time.Millisecond
+	t.Cleanup(func() { commandTimeout = oldTimeout })
+	_, err = commandOutput(os.Args[0], "-test.run=TestZfsCommandHelperProcess", "--", "sleep")
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "timed out")
+}
+
+func TestZfsCommandHelperProcess(t *testing.T) {
+	if os.Getenv("BESZEL_ZFS_COMMAND_HELPER") != "1" {
+		return
+	}
+	mode := ""
+	for i, arg := range os.Args {
+		if arg == "--" && i+1 < len(os.Args) {
+			mode = os.Args[i+1]
+			break
+		}
+	}
+	switch strings.TrimSpace(mode) {
+	case "locale":
+		_, _ = fmt.Printf("%s/%s", os.Getenv("LC_ALL"), os.Getenv("LANG"))
+	case "sleep":
+		time.Sleep(time.Second)
+	}
+	os.Exit(0)
+}

agent/zfs/zfs_unsupported.go

@@ -7,3 +7,7 @@ import "errors"
 func ARCSize() (uint64, error) {
 	return 0, errors.ErrUnsupported
 }
+
+func PoolKernelStats() ([]PoolKernelStat, error) {
+	return nil, errors.ErrUnsupported
+}

beszel.go

@@ -6,7 +6,7 @@ import "github.com/blang/semver"
 
 const (
 	// Version is the current version of the application.
-	Version = "0.18.8"
+	Version = "0.20.0"
 	// AppName is the name of the application.
 	AppName = "beszel"
 )
@@ -16,3 +16,9 @@ var MinVersionCbor = semver.MustParse("0.12.0")
 
 // MinVersionAgentResponse is the minimum supported version for AgentResponse compatibility.
 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

@@ -1,25 +1,26 @@
 module github.com/henrygd/beszel
 
-go 1.26.6
+go 1.27.1
 
 require (
 	github.com/blang/semver v3.5.1+incompatible
 	github.com/coreos/go-systemd/v22 v22.7.0
-	github.com/ebitengine/purego v0.10.2
-	github.com/fxamacker/cbor/v2 v2.9.2
+	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 v0.17.0
+	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 v0.39.11
-	github.com/shirou/gopsutil/v4 v4.26.7
+	github.com/pocketbase/pocketbase v0.40.2
+	github.com/shirou/gopsutil/v4 v4.26.8
 	github.com/spf13/cast v1.10.0
 	github.com/spf13/cobra v1.10.2
 	github.com/spf13/pflag v1.0.10
-	github.com/stretchr/testify v1.12.0
-	golang.org/x/crypto v0.55.0
-	golang.org/x/exp v0.0.0-20260813180055-c1d0aacb2297
+	github.com/stretchr/testify v1.12.1
+	golang.org/x/crypto v0.56.0
+	golang.org/x/exp v0.0.0-20260824195058-e88cd73687aa
 	golang.org/x/net v0.58.0
 	golang.org/x/sys v0.47.0
 	gopkg.in/yaml.v3 v3.0.1
@@ -41,9 +42,10 @@ 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 v1.19.2 // indirect
+	github.com/klauspost/compress v1.20.0 // indirect
 	github.com/lufia/plan9stats v0.0.0-20260802145828-341c2f0c90b5 // indirect
 	github.com/mattn/go-colorable v0.1.15 // indirect
 	github.com/mattn/go-isatty v0.0.24 // indirect
@@ -55,13 +57,14 @@ require (
 	github.com/tklauser/numcpus v0.12.0 // indirect
 	github.com/x448/float16 v0.8.4 // indirect
 	github.com/yusufpapurcu/wmi v1.2.4 // indirect
+	go.yaml.in/yaml/v3 v3.0.5 // indirect
 	golang.org/x/image v0.45.0 // indirect
 	golang.org/x/oauth2 v0.36.0 // indirect
 	golang.org/x/sync v0.22.0 // indirect
 	golang.org/x/term v0.45.0 // indirect
 	golang.org/x/text v0.41.0 // indirect
-	modernc.org/libc v1.74.1 // indirect
+	modernc.org/libc v1.74.4 // indirect
 	modernc.org/mathutil v1.7.1 // indirect
-	modernc.org/memory v1.12.0 // indirect
-	modernc.org/sqlite v1.55.0 // indirect
+	modernc.org/memory v1.12.1 // indirect
+	modernc.org/sqlite v1.57.0 // indirect
 )

go.sum

@@ -15,12 +15,14 @@ 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=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
-github.com/ebitengine/purego v0.10.2 h1:W809HbnvzAxgdm+aOvlSekrM16wGCdT/e76+9tS7gzE=
-github.com/ebitengine/purego v0.10.2/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=
+github.com/ebitengine/purego v0.11.0 h1:jhp/D+Nyv7UUW8HAcmcjt2N2rYrYi9m3SL21k0Ua/NI=
+github.com/ebitengine/purego v0.11.0/go.mod h1:DCHPP08djqhNSoTfImcnHYQRZmd0qhakvrozqaEYhGQ=
 github.com/eclipse/paho.golang v0.23.0 h1:KHgl2wz6EJo7cMBmkuhpt7C576vP+kpPv7jjvSyR6Mk=
 github.com/eclipse/paho.golang v0.23.0/go.mod h1:nQRhTkoZv8EAiNs5UU0/WdQIx2NrnWUpL9nsGJTQN04=
 github.com/fatih/color v1.19.0 h1:Zp3PiM21/9Ld6FzSKyL5c/BULoe/ONr9KlbYVOfG8+w=
@@ -29,8 +31,8 @@ github.com/frankban/quicktest v1.14.6 h1:7Xjx+VpznH+oBnejlPUj8oUpdxnVs4f8XU8WnHk
 github.com/frankban/quicktest v1.14.6/go.mod h1:4ptaffx2x8+WTWXmUCuVU6aPUX1/Mz7zb5vbUoiM6w0=
 github.com/fsnotify/fsnotify v1.10.1 h1:b0/UzAf9yR5rhf3RPm9gf3ehBPpf0oZKIjtpKrx59Ho=
 github.com/fsnotify/fsnotify v1.10.1/go.mod h1:TLheqan6HD6GBK6PrDWyDPBaEV8LspOxvPSjC+bVfgo=
-github.com/fxamacker/cbor/v2 v2.9.2 h1:X4Ksno9+x3cz0TZv69ec1hxP/+tymuR8PXQJyDwfh78=
-github.com/fxamacker/cbor/v2 v2.9.2/go.mod h1:vM4b+DJCtHn+zz7h3FFp/hDAI9WNWCsZj23V5ytsSxQ=
+github.com/fxamacker/cbor/v2 v2.9.3 h1:oQBnFATpNdY8gJHTndDDv5Xl4QqNaz51G5LLEPhng3Q=
+github.com/fxamacker/cbor/v2 v2.9.3/go.mod h1:vM4b+DJCtHn+zz7h3FFp/hDAI9WNWCsZj23V5ytsSxQ=
 github.com/gabriel-vasile/mimetype v1.4.15 h1:05iP/CYtZ/w455R/KZM6rZ5ieAdh99UPtd+d3YzLmaI=
 github.com/gabriel-vasile/mimetype v1.4.15/go.mod h1:azpTcoLcDZRNgFou5j+APrqQx9HqVPWa6ijYQIIVswQ=
 github.com/ganigeorgiev/fexpr v0.6.0 h1:Fza3O/QMBKEudUvxV862qe6GjxM60GJjjKytdp+VQus=
@@ -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 v0.0.0-20260802141513-ef3492d7dac3 h1:LMLX+LgTNWpfvCBdFebv6EsYotImrt/Ppc5cXIriCSo=
-github.com/google/pprof v0.0.0-20260802141513-ef3492d7dac3/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=
@@ -67,8 +69,8 @@ github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLf
 github.com/jarcoal/httpmock v1.4.2 h1:dKwiP/9zITCPfBLsDn3kchbSOu16JrnxtVEmL0fPRcI=
 github.com/jarcoal/httpmock v1.4.2/go.mod h1:ftW1xULwo+j0R0JJkJIIi7UKigZUXCLLanykgjwBXL0=
 github.com/jessevdk/go-flags v1.4.0/go.mod h1:4FA24M0QyGHXBuZZK/XkWh8h0e1EYbRYJSGM75WSRxI=
-github.com/klauspost/compress v1.19.2 h1:hMRETovs/pu/dVWN7zIT1PGG8t509MwT6bO7XSi26R8=
-github.com/klauspost/compress v1.19.2/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
+github.com/klauspost/compress v1.20.0 h1:a3C1ke2ohxFymNlb2HWAHjDeKCI90scRskErZkR0ezA=
+github.com/klauspost/compress v1.20.0/go.mod h1:LUdAzn7YLVvxLpc7y3V1m40wESHTgc1422pwwBSKYuI=
 github.com/kr/pretty v0.3.1 h1:flRD4NNwYAUpkphVc1HcthR4KEIFJ65n8Mw5qdRn3LE=
 github.com/kr/pretty v0.3.1/go.mod h1:hoEshYVHaxMs3cyo3Yncou5ZscifuDolrwPKZanG3xk=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
@@ -83,19 +85,21 @@ 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 v0.17.0 h1:xfp3z5QbE8jXvUhUEwWDk47SJ/b912VoB8MJJDU+q4E=
-github.com/nicholas-fedor/shoutrrr v0.17.0/go.mod h1:s4ldyLs6uwBy9lIjYrY+8lyTqJtPvZSrILw0CyMLock=
-github.com/onsi/ginkgo/v2 v2.32.0 h1:Hw7s2pVrQo/8Yz5N77qdnpHaoc+c6cC9WIV1Jce+J6E=
-github.com/onsi/ginkgo/v2 v2.32.0/go.mod h1:+aXOY+vzZ5mu2iI2HpTZUPmM//oQfsNFX6gU9kNcA44=
-github.com/onsi/gomega v1.42.1 h1:iN1rCUX+44NZ1Dc97MPoeFYbFR0vh8zxoxMFwKdyZ6I=
-github.com/onsi/gomega v1.42.1/go.mod h1:REff/hsDsodHoKlWsP2mAPhu1+5/6hVYNf9rIEBpeSg=
+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=
 github.com/pocketbase/ozzo-validation/v4 v4.3.0 h1:uKBDVma7bZqgR2a6AwE+k9hkuDFfiZMpBHQdZ1z3iQs=
 github.com/pocketbase/ozzo-validation/v4 v4.3.0/go.mod h1:6XNjSTw/Jb2F8LOkKO3oyzIWExbrGiYoS4uVxVwz90g=
-github.com/pocketbase/pocketbase v0.39.11 h1:cl/Kh13ukof/4BAEku3OozYrLl85M5/bmH62Ny4szFc=
-github.com/pocketbase/pocketbase v0.39.11/go.mod h1:5CaCvp/52fZJ5/qyYsqpCGyVue/kjez889cQAATD2cY=
+github.com/pocketbase/pocketbase v0.40.2 h1:7gTqvt3bmilkphyZZ1QNhX19g3BXHqT7ynDyU81RVT4=
+github.com/pocketbase/pocketbase v0.40.2/go.mod h1:jc3YuyToy+ZXM4CeO7uSCN/htgR8yv+tjSE3eJZ8eh8=
 github.com/power-devops/perfstat v0.0.0-20260805114148-88456608a4f6 h1:jL3a8soXdzuTCcRnKhOmtcsVOObdDTFf4O2B403HPRU=
 github.com/power-devops/perfstat v0.0.0-20260805114148-88456608a4f6/go.mod h1:OmDBASR4679mdNQnz2pUhc2G8CO2JrUAVFDRBDP/hJE=
 github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec h1:W09IVJc94icq4NjY3clb7Lk8O1qJ8BdBEF8z0ibU0rE=
@@ -103,8 +107,8 @@ github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec/go.mod h1:qq
 github.com/rogpeppe/go-internal v1.9.0 h1:73kH8U+JUqXU8lRuOHeVHaa/SZPifC7BkcraZVejAe8=
 github.com/rogpeppe/go-internal v1.9.0/go.mod h1:WtVeX8xhTBvf0smdhujwtBcq4Qrzq/fJaraNFVN+nFs=
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
-github.com/shirou/gopsutil/v4 v4.26.7 h1:IXzpHz/dkMRYAhKkOXr1HB6SuzWU3eoyyeWe7g3bNZc=
-github.com/shirou/gopsutil/v4 v4.26.7/go.mod h1:5O9FjBiXoTDFatIWjZZosqj4pV0DRtLx598xGbBehzM=
+github.com/shirou/gopsutil/v4 v4.26.8 h1:YQMTF/1J50B5+Y0vlo1eDRf5DoR7Gk69hY+8wjYkQeo=
+github.com/shirou/gopsutil/v4 v4.26.8/go.mod h1:5O9FjBiXoTDFatIWjZZosqj4pV0DRtLx598xGbBehzM=
 github.com/spf13/cast v1.10.0 h1:h2x0u2shc1QuLHfxi+cTJvs30+ZAHOGRic8uyGTDWxY=
 github.com/spf13/cast v1.10.0/go.mod h1:jNfB8QC9IA6ZuY2ZjDp0KtFO2LZZlg4S/7bzP6qqeHo=
 github.com/spf13/cobra v1.10.2 h1:DMTTonx5m65Ic0GOoRY2c16WCbHxOOw6xxezuLaBpcU=
@@ -116,8 +120,8 @@ github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+
 github.com/stretchr/objx v0.5.3 h1:jmXUvGomnU1o3W/V5h2VEradbpJDwGrzugQQvL0POH4=
 github.com/stretchr/objx v0.5.3/go.mod h1:rDQraq+vQZU7Fde9LOZLr8Tax6zZvy4kuNKF+QYS+U0=
 github.com/stretchr/testify v1.4.0/go.mod h1:j7eGeouHqKxXV5pUuKE4zz7dFj8WfuZ+81PSLYec5m4=
-github.com/stretchr/testify v1.12.0 h1:K6Mr6jO9JICuend/5xzTM03ydSV3vdNRYAdPSukj8uI=
-github.com/stretchr/testify v1.12.0/go.mod h1:bOYBZb5qJ00vPzWfIqBUZPaxK8jWiXc6d3ErP4Ca9Gw=
+github.com/stretchr/testify v1.12.1 h1:EuwCh5fleGS7H32xRwO3wRGT7DxrDhLAT6FF8MpWDWE=
+github.com/stretchr/testify v1.12.1/go.mod h1:MDEgiDPPsNp5cuIrHPPCyornHKgEVbtFUmoNlxoYthg=
 github.com/tklauser/go-sysconf v0.4.0 h1:7H0uAN+7RkwWRaxhYXDLqa5V3LPrJeV8wmD9dRUgPQU=
 github.com/tklauser/go-sysconf v0.4.0/go.mod h1:8mTNWyog7H+MpKijp4VmKJAd2bbYQ2zuUwkYRbUArPI=
 github.com/tklauser/numcpus v0.12.0 h1:NR85qdvHA9pFse3x3weVZ0r0ST8R6l5RHbZrlRaqob4=
@@ -132,10 +136,10 @@ go.yaml.in/yaml/v3 v3.0.4/go.mod h1:DhzuOOF2ATzADvBadXxruRBLzYTpT36CKvDb3+aBEFg=
 go.yaml.in/yaml/v3 v3.0.5 h1:N6y/pJk8buWs9NY5ERU2HSMfm+IuD/OtfdAnq6kESPw=
 go.yaml.in/yaml/v3 v3.0.5/go.mod h1:HVTZu1O7/Vkt2N+BFy8Zza+lnLsABggaTM2ZpNIGuKg=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
-golang.org/x/crypto v0.55.0 h1:+KWHjbgOaAQ66dh/YlkZKHlz9ZUlq61AFirAR9ntP8M=
-golang.org/x/crypto v0.55.0/go.mod h1:uq0V9dE/fzQuJtbnL+2EhWOE63vo164FY8xqEnV9xis=
-golang.org/x/exp v0.0.0-20260813180055-c1d0aacb2297 h1:YXnL44eJ77R+ji4/ooy8UsXIhz+lbi2Qgdlc8iRN0gY=
-golang.org/x/exp v0.0.0-20260813180055-c1d0aacb2297/go.mod h1:Mkmymgv+uMpSQ/XxJ/7GpdrdYoqm3u72jEbpCLiJmNk=
+golang.org/x/crypto v0.56.0 h1:GUh5Ii4J5jtcseSMiRqr1jXCNHoxjeV9Fmekc2oLy6Y=
+golang.org/x/crypto v0.56.0/go.mod h1:OMW5y6CY9l38uPLmxU6l6pwcXp1obtLo3e6gT7gQR2I=
+golang.org/x/exp v0.0.0-20260824195058-e88cd73687aa h1:QSyA8ishJCyT21kER9KwNt0b7BM3iRK4x9QXhjN5Fdk=
+golang.org/x/exp v0.0.0-20260824195058-e88cd73687aa/go.mod h1:zeBbvyFKDaLwa7CH/zI8KXt7gTl14SF7sO08Pl5jBCM=
 golang.org/x/image v0.0.0-20191009234506-e7c1f5e7dbb8/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.45.0 h1:FMb1nTbH5H9vF55SriQHgFw5GnNL9Jg6L25BwXKzhB0=
 golang.org/x/image v0.45.0/go.mod h1:n62x/7RqlwXDvGsSU4u6IUTUf6KghUZ9Bt7cG/T9Fx4=
@@ -164,17 +168,16 @@ golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGm
 golang.org/x/tools v0.49.0 h1:3NI7VXzL9+1WZD52Dx2ttoPwD5DWrFGpl9mFZDlmisI=
 golang.org/x/tools v0.49.0/go.mod h1:SJNXV9DBKT0UbdttsQjbfJlAE/q+y36++zo3uL3N0Oo=
 google.golang.org/appengine v1.6.5/go.mod h1:8WjMMxjGQR8xUklV/ARdw2HLXBOI7O7uCIDZVag1xfc=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
-gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=
-gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=
 gopkg.in/yaml.v1 v1.0.0-20140924161607-9f9df34309c0/go.mod h1:WDnlLJ4WF5VGsH/HVa3CI79GS0ol3YnhVnKP89i0kNg=
 gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 howett.net/plist v1.0.1 h1:37GdZ8tP09Q35o9ych3ehygcsL+HqKSwzctveSlarvM=
 howett.net/plist v1.0.1/go.mod h1:lqaXoTrLY4hg8tnEzNru53gicrbv7rrk+2xJA/7hw9g=
-modernc.org/cc/v4 v4.29.0 h1:CXgwL8cvxmyzBQZzbSl/6xFtMCryb6u8IOqDci39cgc=
-modernc.org/cc/v4 v4.29.0/go.mod h1:OnovgIhbbMXMu1aISnJ0wvVD1KnW+cAUJkIrAWh+kVI=
+modernc.org/cc/v4 v4.29.1 h1:MKgdCV3WykTSPqpVrnxdEDS0HEd2FHpKZDzxzU5LyeI=
+modernc.org/cc/v4 v4.29.1/go.mod h1:OnovgIhbbMXMu1aISnJ0wvVD1KnW+cAUJkIrAWh+kVI=
 modernc.org/ccgo/v4 v4.34.6 h1:sBgfIwyN0TQ9C5hwIeuqyeAKyMWnbvj2fvpF4L11uzU=
 modernc.org/ccgo/v4 v4.34.6/go.mod h1:SZ8YcN9NG7XVsQYdm6jYBvi8PQP1qi+kqB6OhjqI3Fk=
 modernc.org/fileutil v1.4.0 h1:j6ZzNTftVS054gi281TyLjHPp6CPHr2KCxEXjEbD6SM=
@@ -185,18 +188,18 @@ modernc.org/gc/v3 v3.1.4 h1:2g65LGVSmFQrXeITAw97x7hCRvZFcyE1uDP+7Vng7JI=
 modernc.org/gc/v3 v3.1.4/go.mod h1:HFK/6AGESC7Ex+EZJhJ2Gni6cTaYpSMmU/cT9RmlfYY=
 modernc.org/goabi0 v0.2.0 h1:HvEowk7LxcPd0eq6mVOAEMai46V+i7Jrj13t4AzuNks=
 modernc.org/goabi0 v0.2.0/go.mod h1:CEFRnnJhKvWT1c1JTI3Avm+tgOWbkOu5oPA8eH8LnMI=
-modernc.org/libc v1.74.1 h1:bdR4VTKFMC4966QSNZ05XLGI/VwzVa2kTUX51Dm0riQ=
-modernc.org/libc v1.74.1/go.mod h1:uH4t5bOx3G3g9Xcmj10YKlTcVISlRDwv8VoQJG9n8Os=
+modernc.org/libc v1.74.4 h1:fX1Omw4o2/1C2iRkkIsrQTasJQldLhRmuPreXLoWs9k=
+modernc.org/libc v1.74.4/go.mod h1:eeQAS9W3sZeKYMFubydxJpII9ybHWshk+7or7bLG9co=
 modernc.org/mathutil v1.7.1 h1:GCZVGXdaN8gTqB1Mf/usp1Y/hSqgI2vAGGP4jZMCxOU=
 modernc.org/mathutil v1.7.1/go.mod h1:4p5IwJITfppl0G4sUEDtCr4DthTaT47/N3aT6MhfgJg=
-modernc.org/memory v1.12.0 h1:twkmYNkGXCvtYWzoux02jtK6eovjZbdI0uHFUYp6kuU=
-modernc.org/memory v1.12.0/go.mod h1:/JP4VbVC+K5sU2wZi9bHoq2MAkCnrt2r98UGeSK7Mjw=
+modernc.org/memory v1.12.1 h1:nFMiWrpStgZczNl6XI9GnIk/rWhYIyHGUaR04pGbp9g=
+modernc.org/memory v1.12.1/go.mod h1:/JP4VbVC+K5sU2wZi9bHoq2MAkCnrt2r98UGeSK7Mjw=
 modernc.org/opt v0.2.0 h1:tGyef5ApycA7FSEOMraay9SaTk5zmbx7Tu+cJs4QKZg=
 modernc.org/opt v0.2.0/go.mod h1:03fq9lsNfvkYSfxrfUhZCWPk1lm4cq4N+Bh//bEtgns=
 modernc.org/sortutil v1.2.1 h1:+xyoGf15mM3NMlPDnFqrteY07klSFxLElE2PVuWIJ7w=
 modernc.org/sortutil v1.2.1/go.mod h1:7ZI3a3REbai7gzCLcotuw9AC4VZVpYMjDzETGsSMqJE=
-modernc.org/sqlite v1.55.0 h1:hIFh0MCH0rGinQ/4KYb5/UbCkRkb+UP+OkLCVWa5MTM=
-modernc.org/sqlite v1.55.0/go.mod h1:4ntCLuNmnH8+GNqjka1wNg7KJd5/Hi5FYp8K+XQ7GZw=
+modernc.org/sqlite v1.57.0 h1:qNQP6xnx5M0ISNtlnxoOX0+cD5bJ0/gr9aMmndFczzg=
+modernc.org/sqlite v1.57.0/go.mod h1:yCJ2cmAaIkHQ25oXWrF8H4O1lIfPYPR26yCEDj2P3pQ=
 modernc.org/strutil v1.2.1 h1:UneZBkQA+DX2Rp35KcM69cSsNES9ly8mQWD71HKlOA0=
 modernc.org/strutil v1.2.1/go.mod h1:EHkiggD70koQxjVdSBM3JKM7k6L0FbGE5eymy9i3B9A=
 modernc.org/token v1.1.0 h1:Xl7Ap9dKaEs5kLoOQeQmPWevfnk/DM5qcLcYlA8ys6Y=

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 {
@@ -48,6 +49,7 @@ type SystemAlertFsStats struct {
 // Values pulled from system_stats.stats that are relevant to alerts.
 type SystemAlertStats struct {
 	Cpu          float64                       `json:"cpu"`
+	CpuBreakdown []float64                     `json:"cpub"`
 	Mem          float64                       `json:"mp"`
 	Disk         float64                       `json:"dp"`
 	Bandwidth    [2]uint64                     `json:"b"`
@@ -57,12 +59,19 @@ type SystemAlertStats struct {
 	Battery      [2]uint8                      `json:"bat"`
 	Batteries    map[string]uint8              `json:"bats"`
 	ExtraFs      map[string]SystemAlertFsStats `json:"efs"`
+	ZfsPools     map[string]SystemAlertZfsPool `json:"z"`
 }
 
 type SystemAlertGPUData struct {
 	Usage float64 `json:"u"`
 }
 
+type SystemAlertZfsPool struct {
+	Raw   bool    `json:"raw,omitempty"`
+	Total float64 `json:"d"`
+	Used  float64 `json:"du"`
+}
+
 type SystemAlertData struct {
 	systemRecord *core.Record
 	alertData    CachedAlertData
@@ -99,8 +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
@@ -108,9 +118,13 @@ 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)
+	am.hub.OnRecordAfterCreateSuccess("zfs_pools").BindFunc(am.handleZfsPoolCreateAlert)
+	am.hub.OnRecordAfterUpdateSuccess("zfs_pools").BindFunc(am.handleZfsPoolAlert)
+	am.hub.OnRecordAfterDeleteSuccess("zfs_pools").BindFunc(resolveZfsPoolHistoryOnDelete)
 
 	am.hub.OnServe().BindFunc(func(e *core.ServeEvent) error {
 		// Populate all alerts into cache on startup
@@ -119,6 +133,9 @@ func (am *AlertManager) bindEvents() {
 		if err := resolveStatusAlerts(e.App); err != nil {
 			e.App.Logger().Error("Failed to resolve stale status alerts", "err", err)
 		}
+		if err := resolveSystemdAlerts(e.App); err != nil {
+			e.App.Logger().Error("Failed to resolve stale systemd alerts", "err", err)
+		}
 		if err := am.restorePendingStatusAlerts(); err != nil {
 			e.App.Logger().Error("Failed to restore pending status alerts", "err", err)
 		}
@@ -218,8 +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)
 		}
 	}
@@ -250,6 +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 {
@@ -292,7 +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

@@ -3,12 +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"
 )
@@ -30,6 +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
@@ -37,6 +43,9 @@ func UpsertUserAlerts(e *core.RequestEvent) error {
 
 	err = e.App.RunInTransaction(func(txApp core.App) error {
 		for _, systemId := range reqData.Systems {
+			if !userHasSystem(txApp, userID, systemId) {
+				continue
+			}
 			// find existing matching alert
 			alertRecord, err := txApp.FindFirstRecordByFilter(alertsCollection,
 				"system={:system} && name={:name} && user={:user}",
@@ -94,6 +103,9 @@ func DeleteUserAlerts(e *core.RequestEvent) error {
 
 	err = e.App.RunInTransaction(func(txApp core.App) error {
 		for _, systemId := range reqData.Systems {
+			if !userHasSystem(txApp, userID, systemId) {
+				continue
+			}
 			// Find existing alert to delete
 			alertRecord, err := txApp.FindFirstRecordByFilter("alerts",
 				"system={:system} && name={:name} && user={:user}",
@@ -122,6 +134,15 @@ func DeleteUserAlerts(e *core.RequestEvent) error {
 	return e.JSON(http.StatusOK, map[string]any{"success": true, "count": numDeleted})
 }
 
+func userHasSystem(app core.App, userID, systemID string) bool {
+	system, err := app.FindRecordById("systems", systemID)
+	if err != nil {
+		return false
+	}
+	shareAll, _ := utils.GetEnv("SHARE_ALL_SYSTEMS")
+	return shareAll == "true" || slices.Contains(system.GetStringSlice("users"), userID)
+}
+
 // SendTestNotification handles API request to send a test notification to a specified Shoutrrr URL
 func (am *AlertManager) SendTestNotification(e *core.RequestEvent) error {
 	var data struct {
@@ -131,62 +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
-}
-
-func isInternalIP(ip net.IP) bool {
-	return ip.IsPrivate() || ip.IsLoopback() || ip.IsUnspecified()
-}

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"
 
@@ -29,31 +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: "localhost hostname", url: "generic://localhost", internal: true},
-		{name: "localhost hostname", url: "generic+http://localhost/api/v1/postStuff", internal: true},
-		{name: "localhost hostname", url: "generic+http://127.0.0.1:8080/api/v1/postStuff", internal: true},
-		{name: "localhost hostname", url: "generic+https://beszel.dev/api/v1/postStuff", internal: false},
-		{name: "public ipv4", url: "generic://8.8.8.8", 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()
@@ -190,6 +166,30 @@ func TestUserAlertsApi(t *testing.T) {
 				assert.EqualValues(t, 3, user1Alerts, "should have 3 alerts")
 			},
 		},
+		{
+			Name:   "POST ignores systems the user cannot access",
+			Method: http.MethodPost,
+			URL:    "/api/beszel/user-alerts",
+			Headers: map[string]string{
+				"Authorization": user2Token,
+			},
+			ExpectedStatus:  200,
+			ExpectedContent: []string{"\"success\":true"},
+			TestAppFactory:  testAppFactory,
+			Body: jsonReader(map[string]any{
+				"name":    "CPU",
+				"systems": []string{system1.Id},
+				"value":   90,
+				"min":     10,
+			}),
+			BeforeTestFunc: func(t testing.TB, app *pbTests.TestApp, e *core.ServeEvent) {
+				beszelTests.ClearCollection(t, app, "alerts")
+			},
+			AfterTestFunc: func(t testing.TB, app *pbTests.TestApp, res *http.Response) {
+				alerts, _ := app.CountRecords("alerts")
+				assert.Zero(t, alerts)
+			},
+		},
 		{
 			Name:   "Overwrite: false, should not overwrite existing alert",
 			Method: http.MethodPost,
@@ -347,6 +347,31 @@ func TestUserAlertsApi(t *testing.T) {
 				assert.Zero(t, alerts, "should have 0 alerts")
 			},
 		},
+		{
+			Name:   "DELETE ignores systems the user cannot access",
+			Method: http.MethodDelete,
+			URL:    "/api/beszel/user-alerts",
+			Headers: map[string]string{
+				"Authorization": user2Token,
+			},
+			ExpectedStatus:  200,
+			ExpectedContent: []string{"\"count\":0", "\"success\":true"},
+			TestAppFactory:  testAppFactory,
+			Body: jsonReader(map[string]any{
+				"name":    "CPU",
+				"systems": []string{system1.Id},
+			}),
+			BeforeTestFunc: func(t testing.TB, app *pbTests.TestApp, e *core.ServeEvent) {
+				beszelTests.ClearCollection(t, app, "alerts")
+				beszelTests.CreateRecord(app, "alerts", map[string]any{
+					"name": "CPU", "system": system1.Id, "user": user2.Id, "value": 80,
+				})
+			},
+			AfterTestFunc: func(t testing.TB, app *pbTests.TestApp, res *http.Response) {
+				alerts, _ := app.CountRecords("alerts")
+				assert.EqualValues(t, 1, alerts)
+			},
+		},
 		{
 			Name:   "User 2 should not be able to delete alert of user 1",
 			Method: http.MethodDelete,
@@ -396,6 +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")
@@ -420,11 +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,
@@ -432,7 +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\":"},
@@ -474,10 +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",
@@ -488,14 +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 +1,9 @@
 package alerts
 
 import (
+	"sync"
+	"time"
+
 	"github.com/pocketbase/dbx"
 	"github.com/pocketbase/pocketbase/core"
 	"github.com/pocketbase/pocketbase/tools/store"
@@ -8,13 +11,17 @@ import (
 
 // CachedAlertData represents the relevant fields of an alert record for status checking and updates.
 type CachedAlertData struct {
-	Id        string
-	SystemID  string
-	UserID    string
-	Name      string
-	Value     float64
-	Triggered bool
-	Min       uint8
+	Id           string
+	SystemID     string
+	UserID       string
+	Name         string
+	Value        float64
+	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
 }
 
@@ -26,11 +33,19 @@ func (a *CachedAlertData) PopulateFromRecord(record *core.Record) {
 	a.Value = record.GetFloat("value")
 	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
@@ -65,6 +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
 	}
@@ -74,7 +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
@@ -82,6 +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
@@ -98,6 +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
@@ -111,18 +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_container.go

@@ -0,0 +1,318 @@
+package alerts
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/container"
+	"github.com/henrygd/beszel/internal/entities/system"
+
+	"github.com/pocketbase/pocketbase/core"
+)
+
+const (
+	// containerAlertName is the value stored in the alerts.name field for this alert type.
+	containerAlertName = "ContainerHealth"
+
+	// containerLogMaxLines caps how many matched (error/fatal) log lines are kept.
+	containerLogMaxLines = 12
+	// containerLogFallbackLines is how many trailing raw log lines are used when no
+	// line matches "error" or "fatal", so the notification still carries some context.
+	containerLogFallbackLines = 6
+	// containerLogExcerptMaxChars bounds a single container's log excerpt so a
+	// handful of containers can't blow past Discord's message size limit.
+	containerLogExcerptMaxChars = 500
+	// containerAlertMaxLogged is the max number of unhealthy containers we fetch
+	// and embed logs for in a single alert message.
+	containerAlertMaxLogged = 2
+	// containerAlertMessageMaxChars is a final safety cap on the whole message body.
+	containerAlertMessageMaxChars = 1800
+)
+
+// FetchContainerLogsFunc retrieves recent logs for a container ID from its
+// connected agent. Implementations should apply their own timeout. This is a
+// type alias (not a defined type) so it satisfies the hubLike interface in
+// internal/hub/systems, which declares the same func signature without
+// importing this package.
+type FetchContainerLogsFunc = func(containerID string) (string, error)
+
+// containerAlertTarget is an immutable snapshot of the fields needed after the
+// alert fires. Keeping agent-owned container records out of notification work
+// avoids retaining and concurrently reading data that is refreshed in place.
+type containerAlertTarget struct {
+	id   string
+	name string
+}
+
+// HandleContainerAlerts checks configured "ContainerHealth" alerts for a system
+// against the Docker container health data included in the latest agent update.
+// It persists when containers first become unhealthy, fires from a fresh poll
+// once the configured delay has elapsed, and resolves once containers recover.
+// fetchLogs is used when an alert actually fires so the notification can include
+// a log excerpt (prioritizing lines containing "error"/"fatal") for context.
+func (am *AlertManager) HandleContainerAlerts(systemRecord *core.Record, data *system.CombinedData, fetchLogs FetchContainerLogsFunc) error {
+	alerts := am.alertsCache.GetAlertsByName(systemRecord.Id, containerAlertName)
+	if len(alerts) == 0 {
+		return nil
+	}
+	if data.Containers == nil {
+		// An unknown Docker state must not resolve a triggered alert or count
+		// toward the minimum unhealthy duration.
+		var result error
+		for _, alertData := range alerts {
+			if err := am.clearPendingContainerAlert(alertData); err != nil {
+				result = errors.Join(result, err)
+			}
+		}
+		return result
+	}
+
+	var unhealthy []*container.Stats
+	for _, c := range data.Containers {
+		if c.Health == container.DockerHealthUnhealthy {
+			unhealthy = append(unhealthy, c)
+		}
+	}
+
+	systemName := systemRecord.GetString("name")
+	now := time.Now().UTC()
+	var result error
+	for _, alertData := range alerts {
+		if len(unhealthy) > 0 {
+			if alertData.Triggered {
+				continue
+			}
+			min := max(1, int(alertData.Min))
+			if alertData.PendingSince.IsZero() {
+				pendingSince, err := am.setPendingContainerAlert(alertData, now)
+				if err != nil {
+					result = errors.Join(result, err)
+					continue
+				}
+				if pendingSince.IsZero() {
+					continue
+				}
+				alertData.PendingSince = pendingSince
+				if min > 1 {
+					continue
+				}
+			}
+			if min > 1 && now.Before(alertData.PendingSince.Add(time.Duration(min)*time.Minute)) {
+				continue
+			}
+			if err := am.sendContainerHealthAlert(true, systemName, alertData, snapshotContainerAlertTargets(unhealthy), fetchLogs); err != nil {
+				result = errors.Join(result, err)
+			}
+			continue
+		}
+
+		// no unhealthy containers right now
+		if err := am.clearPendingContainerAlert(alertData); err != nil {
+			result = errors.Join(result, err)
+		}
+		if !alertData.Triggered {
+			continue
+		}
+		if err := am.sendContainerHealthAlert(false, systemName, alertData, nil, fetchLogs); err != nil {
+			result = errors.Join(result, err)
+		}
+	}
+	return result
+}
+
+func snapshotContainerAlertTargets(containers []*container.Stats) []containerAlertTarget {
+	targets := make([]containerAlertTarget, len(containers))
+	for i, c := range containers {
+		targets[i] = containerAlertTarget{id: c.Id, name: c.Name}
+	}
+	return targets
+}
+
+// setPendingContainerAlert durably records the first unhealthy observation and
+// returns the persisted generation used to claim delivery.
+func (am *AlertManager) setPendingContainerAlert(alertData CachedAlertData, since time.Time) (time.Time, error) {
+	record, err := am.hub.FindRecordById("alerts", alertData.Id)
+	if err != nil {
+		return time.Time{}, err
+	}
+	if record.GetBool("triggered") {
+		return time.Time{}, nil
+	}
+	if pendingSince := record.GetDateTime("pending_since").Time(); !pendingSince.IsZero() {
+		return pendingSince, nil
+	}
+	// PocketBase date fields are persisted with millisecond precision. Normalize
+	// before saving so the update-hook cache and a subsequent database read agree.
+	since = since.Truncate(time.Millisecond)
+	record.Set("pending_since", since)
+	return since, am.hub.Save(record)
+}
+
+func (am *AlertManager) clearPendingContainerAlert(alertData CachedAlertData) error {
+	if alertData.PendingSince.IsZero() {
+		return nil
+	}
+	record, err := am.hub.FindRecordById("alerts", alertData.Id)
+	if err != nil {
+		return err
+	}
+	if record.GetDateTime("pending_since").Time().IsZero() {
+		return nil
+	}
+	record.Set("pending_since", nil)
+	return am.hub.Save(record)
+}
+
+// claimPendingContainerAlert marks an alert triggered only if the pending
+// generation is still current. A healthy/unknown update can clear the timestamp
+// while logs are being fetched, causing this claim to become a no-op.
+func (am *AlertManager) claimPendingContainerAlert(alertData CachedAlertData) (bool, error) {
+	record, err := am.hub.FindRecordById("alerts", alertData.Id)
+	if err != nil {
+		return false, err
+	}
+	pendingSince := record.GetDateTime("pending_since").Time()
+	if record.GetBool("triggered") || pendingSince.IsZero() || pendingSince.UnixMilli() != alertData.PendingSince.UnixMilli() {
+		return false, nil
+	}
+	record.Set("pending_since", nil)
+	record.Set("triggered", true)
+	return true, am.hub.Save(record)
+}
+
+// CancelPendingContainerAlerts clears pending container-health durations for a
+// system. Called when monitoring pauses or the system goes down.
+func (am *AlertManager) CancelPendingContainerAlerts(systemID string) {
+	for _, alertData := range am.alertsCache.GetAlertsByName(systemID, containerAlertName) {
+		if err := am.clearPendingContainerAlert(alertData); err != nil {
+			am.hub.Logger().Error("Failed to clear pending container alert", "err", err)
+		}
+	}
+}
+
+// sendContainerHealthAlert updates the alert's triggered state and sends the
+// notification. When unhealthy is true, it embeds a log excerpt (prioritizing
+// error/fatal lines) for up to containerAlertMaxLogged of the affected containers.
+func (am *AlertManager) sendContainerHealthAlert(unhealthy bool, systemName string, alertData CachedAlertData, containers []containerAlertTarget, fetchLogs FetchContainerLogsFunc) error {
+	link := am.hub.MakeLink("system", alertData.SystemID)
+	linkText := "View " + systemName
+
+	if !unhealthy {
+		if err := am.setAlertTriggered(alertData, false); err != nil {
+			return err
+		}
+		title := fmt.Sprintf("%s containers are healthy ✅", systemName)
+		return am.SendAlert(AlertMessageData{
+			UserID:   alertData.UserID,
+			SystemID: alertData.SystemID,
+			Title:    title,
+			Message:  strings.TrimSuffix(title, " ✅"),
+			Link:     link,
+			LinkText: linkText,
+		})
+	}
+
+	names := make([]string, len(containers))
+	for i, c := range containers {
+		names[i] = c.name
+	}
+
+	var title string
+	if len(names) == 1 {
+		title = fmt.Sprintf("Unhealthy container %s on %s \U0001F534", names[0], systemName)
+	} else {
+		title = fmt.Sprintf("%d unhealthy containers on %s \U0001F534", len(names), systemName)
+	}
+
+	var body strings.Builder
+	fmt.Fprintf(&body, "Unhealthy: %s", strings.Join(names, ", "))
+	body.WriteString(am.buildContainerLogsSection(containers, fetchLogs))
+
+	message := body.String()
+	if len(message) > containerAlertMessageMaxChars {
+		message = message[:containerAlertMessageMaxChars] + "\n…(truncated)"
+	}
+
+	claimed, err := am.claimPendingContainerAlert(alertData)
+	if err != nil || !claimed {
+		return err
+	}
+
+	return am.SendAlert(AlertMessageData{
+		UserID:   alertData.UserID,
+		SystemID: alertData.SystemID,
+		Title:    title,
+		Message:  message,
+		Link:     link,
+		LinkText: linkText,
+	})
+}
+
+// buildContainerLogsSection attempts to fetch and format log excerpts for up to
+// containerAlertMaxLogged unhealthy containers, to append to an alert message.
+func (am *AlertManager) buildContainerLogsSection(containers []containerAlertTarget, fetchLogs FetchContainerLogsFunc) string {
+	if fetchLogs == nil {
+		return ""
+	}
+
+	var section strings.Builder
+	attempts := min(len(containers), containerAlertMaxLogged)
+	for _, c := range containers[:attempts] {
+		rawLogs, err := fetchLogs(c.id)
+		if err != nil {
+			am.hub.Logger().Warn("Failed to fetch container logs for alert", "container", c.name, "err", err)
+			continue
+		}
+		excerpt := buildContainerLogExcerpt(rawLogs)
+		if excerpt == "" {
+			continue
+		}
+		fmt.Fprintf(&section, "\n\n%s logs:\n```\n%s\n```", c.name, excerpt)
+	}
+
+	if len(containers) > containerAlertMaxLogged {
+		fmt.Fprintf(&section, "\n\n(+%d more unhealthy container(s), logs omitted)", len(containers)-containerAlertMaxLogged)
+	}
+
+	return section.String()
+}
+
+// buildContainerLogExcerpt filters raw container log output down to the lines
+// most likely to explain why the container is unhealthy: lines containing
+// "error" or "fatal" (case-insensitive) are preferred. If none match, the tail
+// of the raw output is used instead so the notification still carries context.
+func buildContainerLogExcerpt(raw string) string {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return ""
+	}
+	lines := strings.Split(raw, "\n")
+
+	var matched []string
+	for _, line := range lines {
+		line = strings.TrimRight(line, "\r")
+		if line == "" {
+			continue
+		}
+		lower := strings.ToLower(line)
+		if strings.Contains(lower, "error") || strings.Contains(lower, "fatal") {
+			matched = append(matched, line)
+		}
+	}
+
+	selected := matched
+	if len(selected) == 0 {
+		start := max(0, len(lines)-containerLogFallbackLines)
+		selected = lines[start:]
+	} else if len(selected) > containerLogMaxLines {
+		selected = selected[len(selected)-containerLogMaxLines:]
+	}
+
+	excerpt := strings.TrimSpace(strings.Join(selected, "\n"))
+	if len(excerpt) > containerLogExcerptMaxChars {
+		excerpt = "…" + excerpt[len(excerpt)-containerLogExcerptMaxChars:]
+	}
+	return excerpt
+}

internal/alerts/alerts_container_test.go

@@ -0,0 +1,349 @@
+//go:build testing
+
+package alerts_test
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+	"testing/synctest"
+	"time"
+
+	"github.com/henrygd/beszel/internal/alerts"
+	"github.com/henrygd/beszel/internal/entities/container"
+	"github.com/henrygd/beszel/internal/entities/system"
+	beszelTests "github.com/henrygd/beszel/internal/tests"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+type containerAlertTestFixture struct {
+	hub          *beszelTests.TestHub
+	am           *alerts.AlertManager
+	alertID      string
+	systemRecord *core.Record
+}
+
+func newContainerAlertTestFixture(t *testing.T, min int) *containerAlertTestFixture {
+	t.Helper()
+
+	hub, user := beszelTests.GetHubWithUser(t)
+
+	systems, err := beszelTests.CreateSystems(hub, 1, user.Id, "up")
+	require.NoError(t, err)
+	systemRecord := systems[0]
+
+	userSettings, err := hub.FindFirstRecordByFilter("user_settings", "user={:user}", map[string]any{"user": user.Id})
+	require.NoError(t, err)
+	userSettings.Set("settings", `{"emails":["test@example.com"],"webhooks":[]}`)
+	require.NoError(t, hub.Save(userSettings))
+
+	alertRecord, err := beszelTests.CreateRecord(hub, "alerts", map[string]any{
+		"name":   "ContainerHealth",
+		"system": systemRecord.Id,
+		"user":   user.Id,
+		"min":    min,
+	})
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "Alert should not be triggered initially")
+
+	return &containerAlertTestFixture{
+		hub:          hub,
+		am:           alerts.NewTestAlertManagerWithoutWorker(hub),
+		alertID:      alertRecord.Id,
+		systemRecord: systemRecord,
+	}
+}
+
+func (f *containerAlertTestFixture) cleanup() {
+	f.hub.Cleanup()
+}
+
+func (f *containerAlertTestFixture) submit(t *testing.T, containers []*container.Stats, fetchLogs alerts.FetchContainerLogsFunc) {
+	t.Helper()
+	data := &system.CombinedData{Containers: containers}
+	require.NoError(t, f.am.HandleContainerAlerts(f.systemRecord, data, fetchLogs))
+}
+
+func (f *containerAlertTestFixture) submitInvalid(t *testing.T) {
+	t.Helper()
+	require.NoError(t, f.am.HandleContainerAlerts(f.systemRecord, &system.CombinedData{}, nil))
+}
+
+func (f *containerAlertTestFixture) assertTriggered(t *testing.T, triggered bool, message string) {
+	t.Helper()
+	alertRecord, err := f.hub.FindRecordById("alerts", f.alertID)
+	require.NoError(t, err)
+	assert.Equal(t, triggered, alertRecord.GetBool("triggered"), message)
+}
+
+func (f *containerAlertTestFixture) assertPending(t *testing.T, pending bool) {
+	t.Helper()
+	alertRecord, err := f.hub.FindRecordById("alerts", f.alertID)
+	require.NoError(t, err)
+	assert.Equal(t, pending, !alertRecord.GetDateTime("pending_since").Time().IsZero())
+}
+
+func waitForContainerAlert(d time.Duration) {
+	time.Sleep(d)
+	synctest.Wait()
+}
+
+func healthyContainer(name string) *container.Stats {
+	return &container.Stats{Name: name, Id: "abc123def456", Health: container.DockerHealthHealthy}
+}
+
+func unhealthyContainer(name string) *container.Stats {
+	return &container.Stats{Name: name, Id: "abc123def456", Health: container.DockerHealthUnhealthy}
+}
+
+func TestContainerHealthAlertTriggersAndResolves(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 1)
+	defer fixture.cleanup()
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, []*container.Stats{unhealthyContainer("web")}, nil)
+
+		fixture.assertTriggered(t, true, "A one-minute alert should trigger on the first unhealthy update")
+		require.Equal(t, 1, fixture.hub.TestMailer.TotalSend(), "An email should have been sent")
+
+		msg := fixture.hub.TestMailer.LastMessage()
+		assert.Contains(t, msg.Subject, "web", "Subject should name the unhealthy container")
+		assert.Contains(t, strings.ToLower(msg.Subject), "unhealthy")
+
+		fixture.submit(t, []*container.Stats{unhealthyContainer("web")}, nil)
+		fixture.assertPending(t, false)
+
+		fixture.submitInvalid(t)
+		fixture.assertTriggered(t, true, "An invalid container snapshot should not resolve the alert")
+		assert.Equal(t, 1, fixture.hub.TestMailer.TotalSend(), "An invalid snapshot should not send a recovery")
+
+		fixture.submit(t, []*container.Stats{}, nil)
+		waitForContainerAlert(time.Second)
+
+		fixture.assertTriggered(t, false, "Alert should resolve once the container is healthy again")
+		assert.Equal(t, 2, fixture.hub.TestMailer.TotalSend(), "A second email should have been sent for the recovery")
+		assert.Contains(t, fixture.hub.TestMailer.LastMessage().Subject, " healthy")
+	})
+}
+
+func TestContainerHealthAlertInvalidSnapshotCancelsPending(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 5)
+	defer fixture.cleanup()
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, []*container.Stats{unhealthyContainer("db")}, nil)
+		fixture.assertPending(t, true)
+		waitForContainerAlert(time.Minute)
+		fixture.submitInvalid(t)
+		fixture.assertPending(t, false)
+		waitForContainerAlert(10 * time.Minute)
+		fixture.submit(t, []*container.Stats{unhealthyContainer("db")}, nil)
+
+		fixture.assertTriggered(t, false, "Stale unhealthy data should not trigger an alert")
+		fixture.assertPending(t, true)
+		assert.Equal(t, 0, fixture.hub.TestMailer.TotalSend())
+	})
+}
+
+func TestContainerHealthAlertSystemDownCancelsPending(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 5)
+	defer fixture.cleanup()
+
+	// Use the hub's alert manager because the system-manager status hook invokes
+	// cancellation on that instance.
+	am := fixture.hub.GetAlertManager()
+	require.NoError(t, am.HandleContainerAlerts(
+		fixture.systemRecord,
+		&system.CombinedData{Containers: []*container.Stats{unhealthyContainer("db")}},
+		nil,
+	))
+	fixture.assertPending(t, true)
+
+	fixture.systemRecord.Set("status", "down")
+	require.NoError(t, fixture.hub.Save(fixture.systemRecord))
+
+	fixture.assertPending(t, false)
+}
+
+func TestContainerHealthAlertResolvesBeforeMinDelayCancelsPending(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 5)
+	defer fixture.cleanup()
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, []*container.Stats{unhealthyContainer("db")}, nil)
+		waitForContainerAlert(time.Minute)
+
+		fixture.assertTriggered(t, false, "Alert should not fire until the min delay elapses")
+		fixture.assertPending(t, true)
+		assert.Equal(t, 0, fixture.hub.TestMailer.TotalSend())
+
+		// container recovers before the 5 minute delay elapses
+		fixture.submit(t, []*container.Stats{healthyContainer("db")}, nil)
+		waitForContainerAlert(10 * time.Minute)
+		fixture.submit(t, []*container.Stats{healthyContainer("db")}, nil)
+
+		fixture.assertTriggered(t, false, "Alert should remain untriggered")
+		fixture.assertPending(t, false)
+		assert.Equal(t, 0, fixture.hub.TestMailer.TotalSend(), "No email should be sent for a container that recovered before the delay")
+	})
+}
+
+func TestContainerHealthAlertPreservesPendingDurationAcrossManagerRestart(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 2)
+	defer fixture.cleanup()
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, []*container.Stats{unhealthyContainer("db")}, nil)
+		waitForContainerAlert(30 * time.Second)
+
+		restarted := alerts.NewTestAlertManagerWithoutWorker(fixture.hub)
+		waitForContainerAlert(91 * time.Second)
+		require.NoError(t, restarted.HandleContainerAlerts(
+			fixture.systemRecord,
+			&system.CombinedData{Containers: []*container.Stats{unhealthyContainer("db")}},
+			nil,
+		))
+
+		fixture.assertTriggered(t, true, "Restart should preserve the original unhealthy start time")
+		fixture.assertPending(t, false)
+		assert.Equal(t, 1, fixture.hub.TestMailer.TotalSend())
+	})
+}
+
+func TestContainerHealthAlertClaimsPendingTimestampAtDatabasePrecision(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 1)
+	defer fixture.cleanup()
+
+	alertRecord, err := fixture.hub.FindRecordById("alerts", fixture.alertID)
+	require.NoError(t, err)
+	// PocketBase persists dates to milliseconds, while record update hooks can
+	// retain the original sub-millisecond value in the in-memory alert cache.
+	alertRecord.Set("pending_since", time.Now().UTC().Add(-2*time.Minute).Truncate(time.Millisecond).Add(123*time.Nanosecond))
+	require.NoError(t, fixture.hub.Save(alertRecord))
+
+	fixture.submit(t, []*container.Stats{unhealthyContainer("db")}, nil)
+
+	fixture.assertTriggered(t, true, "Equivalent persisted and cached timestamps should claim the alert")
+	fixture.assertPending(t, false)
+	assert.Equal(t, 1, fixture.hub.TestMailer.TotalSend())
+}
+
+func TestContainerHealthAlertRecoveryWhileFetchingLogsCancelsDelivery(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 1)
+	defer fixture.cleanup()
+
+	synctest.Test(t, func(t *testing.T) {
+		fetchLogs := func(containerID string) (string, error) {
+			fixture.submit(t, []*container.Stats{healthyContainer("api")}, nil)
+			return "FATAL stale failure", nil
+		}
+		fixture.submit(t, []*container.Stats{unhealthyContainer("api")}, fetchLogs)
+
+		fixture.assertTriggered(t, false, "Recovery should cancel delivery while logs are fetched")
+		fixture.assertPending(t, false)
+		assert.Equal(t, 0, fixture.hub.TestMailer.TotalSend())
+	})
+}
+
+func TestContainerHealthAlertIncludesLogExcerpt(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 1)
+	defer fixture.cleanup()
+
+	rawLogs := strings.Join([]string{
+		"2026-08-16T10:00:00Z booting",
+		"2026-08-16T10:00:01Z ERROR could not reach upstream",
+		"2026-08-16T10:00:02Z FATAL giving up after 3 retries",
+	}, "\n")
+	fetchLogs := func(containerID string) (string, error) {
+		assert.Equal(t, "abc123def456", containerID)
+		return rawLogs, nil
+	}
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, []*container.Stats{unhealthyContainer("api")}, fetchLogs)
+
+		fixture.assertTriggered(t, true, "Alert should be triggered")
+		require.Equal(t, 1, fixture.hub.TestMailer.TotalSend())
+
+		body := fixture.hub.TestMailer.LastMessage().Text
+		assert.Contains(t, body, "could not reach upstream")
+		assert.Contains(t, body, "giving up after 3 retries")
+		assert.NotContains(t, body, "booting", "non error/fatal lines should be dropped when matches exist")
+	})
+}
+
+func TestContainerHealthAlertSkipsLogsOnFetchError(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 1)
+	defer fixture.cleanup()
+
+	fetchLogs := func(containerID string) (string, error) {
+		return "", fmt.Errorf("agent unreachable")
+	}
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, []*container.Stats{unhealthyContainer("api")}, fetchLogs)
+
+		fixture.assertTriggered(t, true, "Alert should still be triggered even if logs can't be fetched")
+		require.Equal(t, 1, fixture.hub.TestMailer.TotalSend())
+	})
+}
+
+func TestContainerHealthAlertCapsLogFetchAttempts(t *testing.T) {
+	fixture := newContainerAlertTestFixture(t, 1)
+	defer fixture.cleanup()
+
+	containers := make([]*container.Stats, 100)
+	for i := range containers {
+		containers[i] = &container.Stats{
+			Name:   fmt.Sprintf("container-%d", i),
+			Id:     fmt.Sprintf("id-%d", i),
+			Health: container.DockerHealthUnhealthy,
+		}
+	}
+	attempts := 0
+	fetchLogs := func(containerID string) (string, error) {
+		attempts++
+		return "", fmt.Errorf("agent unreachable")
+	}
+
+	synctest.Test(t, func(t *testing.T) {
+		fixture.submit(t, containers, fetchLogs)
+
+		fixture.assertTriggered(t, true, "Alert should still fire when log retrieval fails")
+		assert.Equal(t, 2, attempts, "Log retrieval should attempt at most two containers")
+		require.Equal(t, 1, fixture.hub.TestMailer.TotalSend())
+	})
+}
+
+func TestBuildContainerLogExcerptPrefersErrorAndFatalLines(t *testing.T) {
+	raw := strings.Join([]string{
+		"2026-08-16T10:00:00Z starting up",
+		"2026-08-16T10:00:01Z listening on :8080",
+		"2026-08-16T10:00:02Z ERROR failed to connect to db",
+		"2026-08-16T10:00:03Z retrying connection",
+		"2026-08-16T10:00:04Z FATAL could not recover, exiting",
+	}, "\n")
+
+	excerpt := alerts.BuildContainerLogExcerpt(raw)
+	assert.Contains(t, excerpt, "failed to connect to db")
+	assert.Contains(t, excerpt, "could not recover, exiting")
+	assert.NotContains(t, excerpt, "starting up", "non-matching lines should be dropped when error/fatal lines exist")
+}
+
+func TestBuildContainerLogExcerptFallsBackToTailWhenNoMatches(t *testing.T) {
+	var lines []string
+	for i := range 20 {
+		lines = append(lines, fmt.Sprintf("line %d: all good here", i))
+	}
+	raw := strings.Join(lines, "\n")
+
+	excerpt := alerts.BuildContainerLogExcerpt(raw)
+	assert.Contains(t, excerpt, "line 19", "should keep the tail of the output")
+	assert.NotContains(t, excerpt, "line 0:", "should not keep the very start when falling back to a short tail")
+}
+
+func TestBuildContainerLogExcerptEmpty(t *testing.T) {
+	assert.Equal(t, "", alerts.BuildContainerLogExcerpt("   \n  \n"))
+}

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

@@ -13,8 +13,41 @@ import (
 	"github.com/pocketbase/pocketbase/tools/types"
 )
 
+var cpuStateAlerts = map[string]struct {
+	index int
+	label string
+}{
+	"CPUIOWait": {2, "CPU I/O Wait"},
+	"CPUSteal":  {3, "CPU Steal Time"},
+}
+
+func cpuStateAlertValue(name string, breakdown []float64) (float64, bool) {
+	state, ok := cpuStateAlerts[name]
+	if !ok || len(breakdown) < 5 {
+		return 0, false
+	}
+	var total float64
+	for _, value := range breakdown {
+		total += value
+	}
+	if total <= 0 {
+		return 0, false
+	}
+	return breakdown[state.index], true
+}
+
 func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *system.CombinedData) error {
-	alerts := am.alertsCache.GetAlertsExcludingNames(systemRecord.Id, "Status")
+	// Systemd alerts are binary state, not numeric thresholds, so they're handled
+	// separately. They read their own state from the database and don't use data.
+	if err := am.HandleSystemdAlerts(systemRecord); err != nil {
+		am.hub.Logger().Error("Error handling systemd alerts", "err", err)
+	}
+
+	if data == nil {
+		return nil
+	}
+
+	alerts := am.alertsCache.GetAlertsExcludingNames(systemRecord.Id, "Status", alertNameSystemdFailed, containerAlertName, alertNameNetworkMonitorLoss)
 	if len(alerts) == 0 {
 		return nil
 	}
@@ -44,6 +77,14 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 					maxUsedPct = usedPct
 				}
 			}
+			for _, pool := range data.Stats.ZfsPools {
+				if pool != nil && !pool.Raw && pool.Total > 0 {
+					usedPct := pool.Used / pool.Total * 100
+					if usedPct > maxUsedPct {
+						maxUsedPct = usedPct
+					}
+				}
+			}
 			val = maxUsedPct
 		case "Temperature":
 			if data.Info.DashboardTemp < 1 {
@@ -67,6 +108,11 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 				continue
 			}
 			val = float64(data.Stats.Battery[0])
+		default:
+			var ok bool
+			if val, ok = cpuStateAlertValue(name, data.Stats.CpuBreakdown); !ok {
+				continue
+			}
 		}
 
 		triggered := alertData.Triggered
@@ -208,6 +254,16 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 						alert.mapSums[key] += float32(fs.DiskUsed / fs.DiskTotal * 100)
 					}
 				}
+				// add zfs pool usage from historical record
+				for key, pool := range stats.ZfsPools {
+					if !pool.Raw && pool.Total > 0 {
+						zfsKey := zfsDiskAlertKey(key)
+						if _, ok := alert.mapSums[zfsKey]; !ok {
+							alert.mapSums[zfsKey] = 0.0
+						}
+						alert.mapSums[zfsKey] += float32(pool.Used / pool.Total * 100)
+					}
+				}
 			case "Temperature":
 				if alert.mapSums == nil {
 					alert.mapSums = make(map[string]float32, len(stats.Temperatures))
@@ -241,13 +297,20 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 				}
 				alert.val += float64(stats.Battery[0])
 			default:
-				continue
+				value, ok := cpuStateAlertValue(alert.name, stats.CpuBreakdown)
+				if !ok {
+					continue
+				}
+				alert.val += value
 			}
 			alert.count++
 		}
 	}
 	// sum up vals for each alert
 	for _, alert := range validAlerts {
+		if alert.count == 0 {
+			continue
+		}
 		switch alert.name {
 		case "Disk":
 			maxPct := float32(0)
@@ -255,7 +318,12 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 				sumPct := float32(value)
 				if sumPct > maxPct {
 					maxPct = sumPct
-					alert.descriptor = fmt.Sprintf("Usage of %s", key)
+					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))
@@ -301,6 +369,17 @@ func (am *AlertManager) HandleSystemAlerts(systemRecord *core.Record, data *syst
 	return nil
 }
 
+func zfsDiskAlertKey(poolName string) string {
+	return "zfs:" + poolName
+}
+
+func diskAlertDescriptor(key string) string {
+	if poolName, ok := strings.CutPrefix(key, "zfs:"); ok {
+		return fmt.Sprintf("Usage of storage pool %s", poolName)
+	}
+	return fmt.Sprintf("Usage of %s", key)
+}
+
 func hasRepresentativeBattery(legacy [2]uint8, batteries map[string]uint8) bool {
 	return legacy != [2]uint8{} || len(batteries) > 0
 }
@@ -309,6 +388,9 @@ func (am *AlertManager) sendSystemAlert(alert SystemAlertData) {
 	// log.Printf("Sending alert %s: val %f | count %d | threshold %f\n", alert.name, alert.val, alert.count, alert.threshold)
 	systemName := alert.systemRecord.GetString("name")
 
+	if state, ok := cpuStateAlerts[alert.name]; ok {
+		alert.name = state.label
+	}
 	// change Disk to Disk usage
 	if alert.name == "Disk" {
 		alert.name += " usage"
@@ -320,7 +402,7 @@ func (am *AlertManager) sendSystemAlert(alert SystemAlertData) {
 
 	// make title alert name lowercase if not CPU or GPU
 	titleAlertName := alert.name
-	if titleAlertName != "CPU" && titleAlertName != "GPU" {
+	if titleAlertName != "CPU" && titleAlertName != "GPU" && !strings.HasPrefix(titleAlertName, "CPU") {
 		titleAlertName = strings.ToLower(titleAlertName)
 	}
 

internal/alerts/alerts_system_test.go

@@ -146,6 +146,20 @@ func setCPUAlertValue(info *system.Info, stats *system.Stats, value float64) {
 	stats.Cpu = value
 }
 
+func setCPUStateAlertValue(_ *system.Info, stats *system.Stats, value []float64) {
+	stats.CpuBreakdown = value
+}
+
+var cpuStateAlertTests = []struct {
+	name     string
+	trigger  []float64
+	resolve  []float64
+	baseline []float64
+}{
+	{"CPUIOWait", []float64{0, 0, 51, 0, 49}, []float64{0, 0, 48, 0, 52}, []float64{0, 0, 10, 0, 90}},
+	{"CPUSteal", []float64{0, 0, 0, 51, 49}, []float64{0, 0, 0, 48, 52}, []float64{0, 0, 0, 10, 90}},
+}
+
 func setMemoryAlertValue(info *system.Info, stats *system.Stats, value float64) {
 	info.MemPct = value
 	stats.MemPct = value
@@ -191,6 +205,11 @@ func setBatteryAlertValue(info *system.Info, stats *system.Stats, value [2]uint8
 
 func TestSystemAlertsOneMin(t *testing.T) {
 	testOneMinuteSystemAlert(t, "CPU", 50, setCPUAlertValue, 51, 49)
+	for _, test := range cpuStateAlertTests {
+		t.Run(test.name, func(t *testing.T) {
+			testOneMinuteSystemAlert(t, test.name, 50, setCPUStateAlertValue, test.trigger, test.resolve)
+		})
+	}
 	testOneMinuteSystemAlert(t, "Memory", 50, setMemoryAlertValue, 51, 49)
 	testOneMinuteSystemAlert(t, "Disk", 50, setDiskAlertValue, 51, 49)
 	testOneMinuteSystemAlert(t, "Bandwidth", 50, setBandwidthAlertValue, [2]uint64{megabytesToBytes(26), megabytesToBytes(25)}, [2]uint64{megabytesToBytes(25), megabytesToBytes(24)})
@@ -204,6 +223,11 @@ func TestSystemAlertsOneMin(t *testing.T) {
 
 func TestSystemAlertsTwoMin(t *testing.T) {
 	testMultiMinuteSystemAlert(t, "CPU", 50, 2, setCPUAlertValue, 10, 51, 48)
+	for _, test := range cpuStateAlertTests {
+		t.Run(test.name, func(t *testing.T) {
+			testMultiMinuteSystemAlert(t, test.name, 50, 2, setCPUStateAlertValue, test.baseline, test.trigger, test.resolve)
+		})
+	}
 	testMultiMinuteSystemAlert(t, "Memory", 50, 2, setMemoryAlertValue, 10, 51, 48)
 	testMultiMinuteSystemAlert(t, "Disk", 50, 2, setDiskAlertValue, 10, 51, 48)
 	testMultiMinuteSystemAlert(t, "Bandwidth", 50, 2, setBandwidthAlertValue, [2]uint64{megabytesToBytes(10), megabytesToBytes(10)}, [2]uint64{megabytesToBytes(26), megabytesToBytes(25)}, [2]uint64{megabytesToBytes(10), megabytesToBytes(10)})
@@ -214,3 +238,17 @@ func TestSystemAlertsTwoMin(t *testing.T) {
 	testMultiMinuteSystemAlert(t, "LoadAvg15", 4, 2, setLoadAvgAlertValue, [3]float64{0, 0, 2}, [3]float64{0, 0, 4.1}, [3]float64{0, 0, 3.5})
 	testMultiMinuteSystemAlert(t, "Battery", 20, 2, setBatteryAlertValue, [2]uint8{21, 0}, [2]uint8{19, 0}, [2]uint8{25, 1})
 }
+
+func TestCPUStateAlertWithoutBreakdown(t *testing.T) {
+	fixture := newSystemAlertTestFixture(t, "CPUSteal", 1, 1)
+	defer fixture.cleanup()
+
+	synctest.Test(t, func(t *testing.T) {
+		submitValue(fixture, t, []float64(nil), setCPUStateAlertValue)
+		submitValue(fixture, t, []float64{0, 0, 0, 0, 0}, setCPUStateAlertValue)
+		waitForSystemAlert(time.Second)
+
+		fixture.assertTriggered(t, false, "Alert should ignore missing CPU breakdown data")
+		assert.Zero(t, fixture.hub.TestMailer.TotalSend(), "No email should be sent without CPU breakdown data")
+	})
+}

internal/alerts/alerts_systemd.go

@@ -0,0 +1,190 @@
+package alerts
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/henrygd/beszel/internal/entities/systemd"
+
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/core"
+)
+
+// alertNameSystemdFailed is the alerts.name value for the failed systemd services alert.
+const alertNameSystemdFailed = "SystemdFailed"
+
+// maxListedServices caps how many service names are listed in a notification body.
+const maxListedServices = 10
+
+// HandleSystemdAlerts manages alerts for systemd services in the failed state.
+//
+// This is a binary state alert and fires on the first observation of a failed
+// service rather than using a delay. The agent only refreshes systemd state every
+// 10 minutes, so a shorter delay could never observe new data before expiring, and
+// that poll interval already hides services that fail and restart quickly.
+func (am *AlertManager) HandleSystemdAlerts(systemRecord *core.Record) error {
+	alerts := am.alertsCache.GetAlertsByName(systemRecord.Id, alertNameSystemdFailed)
+	if len(alerts) == 0 {
+		return nil
+	}
+
+	// State is read from the systemd_services snapshot rather than the update payload.
+	// The payload is not a reliable source here: realtime dashboard subscriptions fetch
+	// from the agent with a shorter cache time, and the agent omits systemd services from
+	// those responses, overwriting the cached payload roughly once a second while a system
+	// is being viewed. The snapshot table is only written by the full update cycle.
+	total, failed, err := am.queryServiceStates(systemRecord.Id)
+	if err != nil {
+		return err
+	}
+	if total == 0 {
+		// No rows normally means no systemd data for this system (agent without
+		// systemd, or not yet reported), which must not be treated as a recovery.
+		// Read info only in this ambiguous case. The record being saved is used
+		// instead of data because dashboard polling can replace the system's
+		// in-memory payload concurrently.
+		var currentInfo system.Info
+		if err := systemRecord.UnmarshalJSONField("info", &currentInfo); err != nil ||
+			len(currentInfo.Services) == 0 || currentInfo.Services[0] != 0 {
+			return nil
+		}
+	}
+
+	systemName := systemRecord.GetString("name")
+
+	for _, alertData := range alerts {
+		triggered := len(failed) > 0
+		// Only notify on a change of state, so a service that stays failed across
+		// cycles doesn't re-notify every update.
+		if triggered == alertData.Triggered {
+			continue
+		}
+		if err := am.sendSystemdAlert(triggered, systemName, alertData, failed); err != nil {
+			am.hub.Logger().Error("Failed to send alert", "err", err)
+		}
+	}
+	return nil
+}
+
+// queryServiceStates returns the number of services reported in the most recent update
+// for a system, and the names of those in the failed state.
+//
+// Rows are restricted to the latest update because systemd_services is upserted, never
+// pruned on change: a service that no longer exists on the host stops being reported and
+// its row keeps its last known state until the retention sweep removes it. Every row
+// written in one cycle shares a single updated timestamp, so the newest timestamp
+// identifies exactly the services the agent last reported.
+func (am *AlertManager) queryServiceStates(systemID string) (total int, failed []string, err error) {
+	var rows []struct {
+		Name  string               `db:"name"`
+		State systemd.ServiceState `db:"state"`
+	}
+	err = am.hub.DB().
+		Select("name", "state").
+		From("systemd_services").
+		Where(dbx.NewExp(
+			"system={:system} AND updated=(SELECT MAX(updated) FROM systemd_services WHERE system={:system})",
+			dbx.Params{"system": systemID},
+		)).
+		OrderBy("name").
+		All(&rows)
+	if err != nil {
+		return 0, nil, err
+	}
+	for _, row := range rows {
+		if row.State == systemd.StatusFailed {
+			failed = append(failed, row.Name)
+		}
+	}
+	return len(rows), failed, nil
+}
+
+// sendSystemdAlert sends a failed or recovered systemd services alert to the alert's user.
+func (am *AlertManager) sendSystemdAlert(triggered bool, systemName string, alertData CachedAlertData, failed []string) error {
+	// Update trigger state for alert record before sending alert
+	if err := am.setAlertTriggered(alertData, triggered); err != nil {
+		return err
+	}
+
+	var title, message string
+	if triggered {
+		title = fmt.Sprintf("Failed services on %s %v", systemName, "\U0001F534") // Red alert emoji
+		message = fmt.Sprintf("%s on %s: %s", pluralizeServices(len(failed)), systemName, formatServiceList(failed))
+	} else {
+		title = fmt.Sprintf("Services recovered on %s %v", systemName, "✅") // Green checkmark emoji
+		message = fmt.Sprintf("No services are in the failed state on %s.", systemName)
+	}
+
+	systemID := alertData.SystemID
+
+	return am.SendAlert(AlertMessageData{
+		UserID:   alertData.UserID,
+		SystemID: systemID,
+		Title:    title,
+		Message:  message,
+		Link:     am.hub.MakeLink("system", systemID),
+		LinkText: "View " + systemName,
+	})
+}
+
+// pluralizeServices returns a count label like "1 failed service" or "3 failed services".
+func pluralizeServices(count int) string {
+	if count == 1 {
+		return "1 failed service"
+	}
+	return fmt.Sprintf("%d failed services", count)
+}
+
+// formatServiceList joins service names, truncating long lists.
+func formatServiceList(names []string) string {
+	if len(names) <= maxListedServices {
+		return strings.Join(names, ", ")
+	}
+	remaining := len(names) - maxListedServices
+	return fmt.Sprintf("%s and %d more", strings.Join(names[:maxListedServices], ", "), remaining)
+}
+
+// resolveSystemdAlerts resolves triggered systemd alerts for systems that no longer
+// have any failed services. This clears stale state left by a hub restart.
+func resolveSystemdAlerts(app core.App) error {
+	db := app.DB()
+	var alertIds []string
+	err := db.NewQuery(`
+		SELECT a.id
+		FROM alerts a
+		JOIN systems sys ON sys.id = a.system
+		WHERE a.name = {:name}
+		AND a.triggered = true
+		AND (
+			EXISTS (
+				SELECT 1 FROM systemd_services cur
+				WHERE cur.system = a.system
+				AND cur.updated = (SELECT MAX(updated) FROM systemd_services WHERE system = a.system)
+			)
+			OR json_extract(sys.info, '$.sv[0]') = 0
+		)
+		AND NOT EXISTS (
+			SELECT 1 FROM systemd_services s
+			WHERE s.system = a.system AND s.state = {:state}
+			AND s.updated = (SELECT MAX(updated) FROM systemd_services WHERE system = a.system)
+		)
+	`).Bind(dbx.Params{
+		"name":  alertNameSystemdFailed,
+		"state": systemd.StatusFailed,
+	}).Column(&alertIds)
+	if err != nil {
+		return err
+	}
+	for _, alertId := range alertIds {
+		alert, err := app.FindRecordById("alerts", alertId)
+		if err != nil {
+			return err
+		}
+		alert.Set("triggered", false)
+		if err := app.Save(alert); err != nil {
+			return err
+		}
+	}
+	return nil
+}

internal/alerts/alerts_systemd_test.go

@@ -0,0 +1,383 @@
+//go:build testing
+
+package alerts_test
+
+import (
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/internal/alerts"
+	systemEntity "github.com/henrygd/beszel/internal/entities/system"
+	"github.com/henrygd/beszel/internal/entities/systemd"
+	beszelTests "github.com/henrygd/beszel/internal/tests"
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// setSystemdServiceState upserts a systemd_services row mirroring the raw SQL write
+// path used by the hub (createSystemdStatsRecords), which bypasses record hooks.
+func setSystemdServiceState(t *testing.T, hub core.App, systemID, name string, state systemd.ServiceState, updated int64) {
+	t.Helper()
+
+	_, err := hub.DB().NewQuery(
+		"INSERT INTO systemd_services (id, system, name, state, sub, cpu, cpuPeak, memory, memPeak, updated) " +
+			"VALUES ({:id}, {:system}, {:name}, {:state}, 0, 0, 0, 0, 0, {:updated}) " +
+			"ON CONFLICT(id) DO UPDATE SET state = excluded.state, updated = excluded.updated",
+	).Bind(dbx.Params{
+		"id":      systemID + "-" + name,
+		"system":  systemID,
+		"name":    name,
+		"state":   state,
+		"updated": updated,
+	}).Execute()
+	require.NoError(t, err)
+}
+
+// seedServices writes a set of services into the systemd_services snapshot, which is the
+// source HandleSystemdAlerts reads from. All rows share one updated timestamp, matching
+// how the hub writes a batch in createSystemdStatsRecords.
+func seedServices(t *testing.T, hub core.App, systemID string, states ...systemd.ServiceState) {
+	t.Helper()
+	seedServicesAt(t, hub, systemID, time.Now().UTC().UnixMilli(), states...)
+}
+
+// seedServicesAt writes services with an explicit batch timestamp.
+func seedServicesAt(t *testing.T, hub core.App, systemID string, updated int64, states ...systemd.ServiceState) {
+	t.Helper()
+	for i, state := range states {
+		setSystemdServiceState(t, hub, systemID, serviceName(i), state, updated)
+	}
+}
+
+func serviceName(i int) string {
+	return string(rune('a'+i)) + ".service"
+}
+
+// systemdTestSetup creates a user with an email, a system, and a SystemdFailed alert.
+func systemdTestSetup(t *testing.T, triggered bool) (*beszelTests.TestHub, *core.Record, *core.Record) {
+	t.Helper()
+
+	hub, user := beszelTests.GetHubWithUser(t)
+
+	userSettings, err := hub.FindFirstRecordByFilter("user_settings", "user={:user}", map[string]any{"user": user.Id})
+	require.NoError(t, err)
+	userSettings.Set("settings", `{"emails":["test@example.com"],"webhooks":[]}`)
+	require.NoError(t, hub.Save(userSettings))
+
+	// "paused" avoids spawning a background updater goroutine that would outlive
+	// the test hub; these tests drive HandleSystemdAlerts directly.
+	systems, err := beszelTests.CreateSystems(hub, 1, user.Id, "paused")
+	require.NoError(t, err)
+	system := systems[0]
+
+	alert, err := beszelTests.CreateRecord(hub, "alerts", map[string]any{
+		"name":      "SystemdFailed",
+		"system":    system.Id,
+		"user":      user.Id,
+		"triggered": triggered,
+	})
+	require.NoError(t, err)
+
+	return hub, system, alert
+}
+
+func TestSystemdAlertFiresImmediately(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, false)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	seedServices(t, hub, system.Id, systemd.StatusFailed, systemd.StatusActive)
+	require.NoError(t, am.HandleSystemdAlerts(system))
+
+	assert.Equal(t, initialEmailCount+1, hub.TestMailer.TotalSend(), "failed service should notify on first observation")
+
+	messages := hub.TestMailer.Messages()
+	require.NotEmpty(t, messages)
+	last := messages[len(messages)-1]
+	assert.Contains(t, last.Subject, "Failed services")
+	assert.Contains(t, last.Text, "a.service", "notification should name the failed service")
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.True(t, alertRecord.GetBool("triggered"), "alert should be marked triggered")
+
+	// history record should be created via the alerts update hook
+	historyCount, err := hub.CountRecords("alerts_history", dbx.HashExp{"resolved": ""})
+	require.NoError(t, err)
+	assert.EqualValues(t, 1, historyCount, "should have one unresolved alert history record")
+}
+
+func TestSystemdAlertFullCycle(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, false)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	// Fail, then recover.
+	seedServices(t, hub, system.Id, systemd.StatusFailed)
+	require.NoError(t, am.HandleSystemdAlerts(system))
+	seedServices(t, hub, system.Id, systemd.StatusActive)
+	require.NoError(t, am.HandleSystemdAlerts(system))
+
+	assert.Equal(t, initialEmailCount+2, hub.TestMailer.TotalSend(), "should send a failure and a recovery notification")
+
+	messages := hub.TestMailer.Messages()
+	require.Len(t, messages, 2)
+	assert.Contains(t, messages[0].Subject, "Failed services")
+	assert.Contains(t, messages[1].Subject, "Services recovered")
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "alert should be cleared after recovery")
+
+	// history record should be resolved
+	historyCount, err := hub.CountRecords("alerts_history", dbx.HashExp{"resolved": ""})
+	require.NoError(t, err)
+	assert.Zero(t, historyCount, "alert history record should be resolved")
+}
+
+func TestSystemdAlertSendsRecoveryWhenTriggered(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	seedServices(t, hub, system.Id, systemd.StatusActive, systemd.StatusInactive)
+	require.NoError(t, am.HandleSystemdAlerts(system))
+
+	assert.Equal(t, initialEmailCount+1, hub.TestMailer.TotalSend(), "recovery notification should be sent")
+	messages := hub.TestMailer.Messages()
+	require.NotEmpty(t, messages)
+	assert.Contains(t, messages[len(messages)-1].Subject, "Services recovered")
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "alert should be cleared after recovery")
+}
+
+func TestSystemdAlertDoesNotResendWhileTriggered(t *testing.T) {
+	hub, system, _ := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	// Still failing across several cycles — should not re-notify.
+	for range 3 {
+		seedServices(t, hub, system.Id, systemd.StatusFailed)
+		require.NoError(t, am.HandleSystemdAlerts(system))
+	}
+
+	assert.Equal(t, initialEmailCount, hub.TestMailer.TotalSend(), "should not re-notify while still triggered")
+}
+
+func TestSystemdAlertRepeatedFailureNotifiesOnce(t *testing.T) {
+	hub, system, _ := systemdTestSetup(t, false)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	for range 3 {
+		seedServices(t, hub, system.Id, systemd.StatusFailed)
+		require.NoError(t, am.HandleSystemdAlerts(system))
+	}
+
+	assert.Equal(t, initialEmailCount+1, hub.TestMailer.TotalSend(), "repeated failures should only notify once")
+}
+
+// A service that no longer exists on the host stops being reported, but its row stays
+// in systemd_services with its last known state until the retention sweep. That stale
+// row must not keep the alert triggered.
+func TestSystemdAlertIgnoresServicesNoLongerReported(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	now := time.Now().UTC().UnixMilli()
+	// Older batch still holding a failed service that has since been removed.
+	setSystemdServiceState(t, hub, system.Id, "gone.service", systemd.StatusFailed, now-60_000)
+	// Current batch reports only healthy services.
+	seedServicesAt(t, hub, system.Id, now, systemd.StatusActive, systemd.StatusActive)
+
+	require.NoError(t, am.HandleSystemdAlerts(system))
+
+	assert.Equal(t, initialEmailCount+1, hub.TestMailer.TotalSend(), "stale failed row should not block recovery")
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "alert should resolve once the service stops being reported")
+}
+
+func TestResolveSystemdAlertsIgnoresStaleFailedRows(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	now := time.Now().UTC().UnixMilli()
+	setSystemdServiceState(t, hub, system.Id, "gone.service", systemd.StatusFailed, now-60_000)
+	seedServicesAt(t, hub, system.Id, now, systemd.StatusActive)
+
+	require.NoError(t, alerts.ResolveSystemdAlerts(hub))
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "stale failed row should not keep the alert triggered")
+}
+
+func TestSystemdAlertNoSystemdDataIsIgnored(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	// A system with no systemd_services rows (agent without systemd, or nothing
+	// reported yet) must not be treated as a recovery.
+	require.NoError(t, am.HandleSystemdAlerts(system))
+	require.NoError(t, am.HandleSystemdAlerts(system))
+
+	assert.Equal(t, initialEmailCount, hub.TestMailer.TotalSend(), "missing systemd data should not send a recovery")
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.True(t, alertRecord.GetBool("triggered"), "triggered state should be preserved when data is absent")
+}
+
+func TestSystemdAlertFreshEmptySnapshotResolves(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	// An explicit zero service count on the saved system record distinguishes a
+	// confirmed empty snapshot from an agent response that omitted systemd data.
+	system.Set("info", systemEntity.Info{Services: []uint16{0, 0}})
+	require.NoError(t, am.HandleSystemAlerts(system, nil))
+
+	assert.Equal(t, initialEmailCount+1, hub.TestMailer.TotalSend(), "fresh empty snapshot should send a recovery")
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "fresh empty snapshot should resolve the alert")
+}
+
+func TestSystemdAlertNoAlertRecord(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	systems, err := beszelTests.CreateSystems(hub, 1, user.Id, "paused")
+	require.NoError(t, err)
+	system := systems[0]
+
+	initialEmailCount := hub.TestMailer.TotalSend()
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+
+	seedServices(t, hub, system.Id, systemd.StatusFailed)
+	require.NoError(t, am.HandleSystemdAlerts(system))
+	assert.Equal(t, initialEmailCount, hub.TestMailer.TotalSend(), "no email when no alert record exists")
+}
+
+func TestResolveSystemdAlertsClearsStaleTriggered(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	// No failed services in the snapshot, but the alert is still marked triggered
+	// (e.g. the hub restarted while the alert was active).
+	setSystemdServiceState(t, hub, system.Id, "a.service", systemd.StatusActive, time.Now().UTC().UnixMilli())
+
+	require.NoError(t, alerts.ResolveSystemdAlerts(hub))
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "stale triggered flag should be cleared")
+}
+
+func TestResolveSystemdAlertsKeepsTriggeredWithoutSystemdData(t *testing.T) {
+	hub, _, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	// Missing rows do not prove recovery. This can happen when a system is offline
+	// and its last service snapshot has been removed by retention.
+	require.NoError(t, alerts.ResolveSystemdAlerts(hub))
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.True(t, alertRecord.GetBool("triggered"), "missing systemd data should preserve triggered state")
+}
+
+func TestResolveSystemdAlertsClearsConfirmedEmptySnapshot(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	// Update the persisted snapshot directly so record hooks don't alter alert state
+	// before the startup resolver is exercised.
+	_, err := hub.DB().NewQuery(
+		"UPDATE systems SET info = {:info} WHERE id = {:id}",
+	).Bind(dbx.Params{"info": `{"sv":[0,0]}`, "id": system.Id}).Execute()
+	require.NoError(t, err)
+	require.NoError(t, alerts.ResolveSystemdAlerts(hub))
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.False(t, alertRecord.GetBool("triggered"), "confirmed empty snapshot should clear triggered state")
+}
+
+func TestResolveSystemdAlertsKeepsStillFailing(t *testing.T) {
+	hub, system, alert := systemdTestSetup(t, true)
+	defer hub.Cleanup()
+
+	setSystemdServiceState(t, hub, system.Id, "a.service", systemd.StatusFailed, time.Now().UTC().UnixMilli())
+
+	require.NoError(t, alerts.ResolveSystemdAlerts(hub))
+
+	alertRecord, err := hub.FindRecordById("alerts", alert.Id)
+	require.NoError(t, err)
+	assert.True(t, alertRecord.GetBool("triggered"), "alert should stay triggered while a service is still failed")
+}
+
+func TestSystemdAlertMultipleUsersRespectOwnAlerts(t *testing.T) {
+	hub, user1 := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	setStatusAlertEmail(t, hub, user1.Id, "user1@example.com")
+
+	user2, err := beszelTests.CreateUser(hub, "user2@example.com", "password")
+	require.NoError(t, err)
+	_, err = beszelTests.CreateRecord(hub, "user_settings", map[string]any{
+		"user": user2.Id,
+		"settings": map[string]any{
+			"emails":   []string{"user2@example.com"},
+			"webhooks": []string{},
+		},
+	})
+	require.NoError(t, err)
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "shared-system",
+		"users": []string{user1.Id, user2.Id},
+		"host":  "127.0.0.1",
+	})
+	require.NoError(t, err)
+
+	for _, user := range []*core.Record{user1, user2} {
+		_, err = beszelTests.CreateRecord(hub, "alerts", map[string]any{
+			"name":   "SystemdFailed",
+			"system": system.Id,
+			"user":   user.Id,
+		})
+		require.NoError(t, err)
+	}
+
+	am := alerts.NewTestAlertManagerWithoutWorker(hub)
+	seedServices(t, hub, system.Id, systemd.StatusFailed)
+	require.NoError(t, am.HandleSystemdAlerts(system))
+
+	messages := hub.TestMailer.Messages()
+	require.Len(t, messages, 2, "each user should receive their own alert")
+}

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),
 	}
 }
 
@@ -88,6 +89,10 @@ func ResolveStatusAlerts(app core.App) error {
 	return resolveStatusAlerts(app)
 }
 
+func ResolveSystemdAlerts(app core.App) error {
+	return resolveSystemdAlerts(app)
+}
+
 func (am *AlertManager) RestorePendingStatusAlerts() error {
 	return am.restorePendingStatusAlerts()
 }
@@ -96,6 +101,7 @@ 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

@@ -0,0 +1,145 @@
+package alerts
+
+import (
+	"fmt"
+	"time"
+
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/core"
+)
+
+// handleZfsPoolAlert sends alerts when a ZFS pool health state worsens and
+// resolves the alert history entry when the pool recovers. Like the SMART
+// hook, this is automatic and does not require user opt-in.
+func (am *AlertManager) handleZfsPoolAlert(e *core.RecordEvent) error {
+	return am.handleZfsPoolHealthAlert(e, e.Record.Original().GetString("health"))
+}
+
+func (am *AlertManager) handleZfsPoolCreateAlert(e *core.RecordEvent) error {
+	return am.handleZfsPoolHealthAlert(e, "")
+}
+
+func (am *AlertManager) handleZfsPoolHealthAlert(e *core.RecordEvent, oldHealth string) error {
+	newHealth := e.Record.GetString("health")
+	oldSeverity := zfsPoolSeverity(oldHealth)
+	newSeverity := zfsPoolSeverity(newHealth)
+
+	systemID := e.Record.GetString("system")
+	if systemID == "" {
+		return e.Next()
+	}
+
+	systemRecord, err := e.App.FindRecordById("systems", systemID)
+	if err != nil {
+		e.App.Logger().Error("Failed to find system for ZFS alert", "err", err, "systemID", systemID)
+		return e.Next()
+	}
+
+	// Pool recovered to a healthy state: resolve any open history entries.
+	if newSeverity == 1 && oldSeverity > 1 {
+		resolveAllAlertHistoryRecords(e.App, e.Record.Id)
+		return e.Next()
+	}
+
+	if !shouldSendZfsPoolAlert(oldSeverity, newSeverity) {
+		return e.Next()
+	}
+
+	systemName := systemRecord.GetString("name")
+	poolName := e.Record.GetString("display_name")
+	if poolName == "" {
+		poolName = e.Record.GetString("name")
+	}
+
+	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("Storage pool %s (%s) health changed from %s to %s", poolName, systemName, oldHealth, newHealth)
+	}
+
+	userIDs := systemRecord.GetStringSlice("users")
+	if len(userIDs) == 0 {
+		return e.Next()
+	}
+
+	for _, userID := range userIDs {
+		if err := am.SendAlert(AlertMessageData{
+			UserID:   userID,
+			SystemID: systemID,
+			Title:    title,
+			Message:  message,
+			Link:     am.hub.MakeLink("system", systemID),
+			LinkText: "View " + systemName,
+		}); err != nil {
+			e.App.Logger().Error("Failed to send ZFS alert", "err", err, "userID", userID)
+		}
+		_ = createZfsPoolHistoryRecord(e.App, userID, systemID, e.Record.Id, poolName)
+	}
+
+	return e.Next()
+}
+
+// resolveZfsPoolHistoryOnDelete resolves open alert history entries when a
+// pool record is deleted (manually or because the pool disappeared), so the
+// UI does not keep showing an ongoing alert for a pool that no longer exists.
+func resolveZfsPoolHistoryOnDelete(e *core.RecordEvent) error {
+	resolveAllAlertHistoryRecords(e.App, e.Record.Id)
+	return e.Next()
+}
+
+// shouldSendZfsPoolAlert reports whether a health transition warrants an alert.
+// First observations of unhealthy pools and worsening transitions are reported.
+func shouldSendZfsPoolAlert(oldSeverity, newSeverity int) bool {
+	return newSeverity > 1 && (oldSeverity == 0 || newSeverity > oldSeverity)
+}
+
+// zfsPoolSeverity ranks pool health states: healthy (1), degraded (2),
+// failed/unavailable (3), unknown (0).
+func zfsPoolSeverity(health string) int {
+	switch health {
+	case "ONLINE":
+		return 1
+	case "DEGRADED":
+		return 2
+	case "FAULTED", "OFFLINE", "UNAVAIL", "REMOVED", "SUSPENDED":
+		return 3
+	default:
+		return 0
+	}
+}
+
+// createZfsPoolHistoryRecord logs a pool health alert in the alerts history so
+// it is visible in the UI without creating an editable alert configuration.
+func createZfsPoolHistoryRecord(app core.App, userID, systemID, alertID, poolName string) error {
+	collection, err := app.FindCachedCollectionByNameOrId("alerts_history")
+	if err != nil {
+		return err
+	}
+	record := core.NewRecord(collection)
+	record.Set("user", userID)
+	record.Set("system", systemID)
+	record.Set("alert_id", alertID)
+	record.Set("name", "Storage Pool: "+poolName)
+	return app.Save(record)
+}
+
+// resolveAllAlertHistoryRecords resolves every open history entry for an alert
+// record id (one per system user).
+func resolveAllAlertHistoryRecords(app core.App, alertID string) {
+	records, err := app.FindRecordsByFilter(
+		"alerts_history",
+		"alert_id={:alert_id} && resolved=null",
+		"", 0, 0,
+		dbx.Params{"alert_id": alertID},
+	)
+	if err != nil || len(records) == 0 {
+		return
+	}
+	now := time.Now().UTC()
+	for _, record := range records {
+		record.Set("resolved", now)
+		if err := app.Save(record); err != nil {
+			app.Logger().Error("Failed to resolve ZFS alert history", "err", err, "recordId", record.Id)
+		}
+	}
+}

internal/alerts/alerts_zfs_disk_test.go

@@ -0,0 +1,179 @@
+//go:build testing
+
+package alerts_test
+
+import (
+	"encoding/json"
+	"testing"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+	beszelTests "github.com/henrygd/beszel/internal/tests"
+
+	"github.com/pocketbase/dbx"
+	"github.com/pocketbase/pocketbase/tools/types"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestDiskAlertZfsPoolMultiMinute verifies that ZFS pool usage participates in
+// the Disk threshold alert using historical per-minute values, mirroring the
+// extra-filesystem behavior.
+func TestDiskAlertZfsPoolMultiMinute(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	systems, err := beszelTests.CreateSystems(hub, 1, user.Id, "up")
+	require.NoError(t, err)
+	systemRecord := systems[0]
+
+	diskAlert, err := beszelTests.CreateRecord(hub, "alerts", map[string]any{
+		"name":   "Disk",
+		"system": systemRecord.Id,
+		"user":   user.Id,
+		"value":  80, // threshold: 80%
+		"min":    2,  // requires historical averaging
+	})
+	require.NoError(t, err)
+
+	am := hub.GetAlertManager()
+	now := time.Now().UTC()
+
+	poolHigh := map[string]*system.ZfsPool{
+		"tank": {Total: 1000, Used: 920}, // 92% - above threshold
+	}
+
+	recordTimes := []time.Duration{
+		-180 * time.Second,
+		-90 * time.Second,
+		-60 * time.Second,
+		-30 * time.Second,
+	}
+
+	for _, offset := range recordTimes {
+		stats := system.Stats{
+			DiskPct:  30, // root disk at 30% - below threshold
+			ZfsPools: poolHigh,
+		}
+		statsJSON, _ := json.Marshal(stats)
+
+		recordTime := now.Add(offset)
+		record, err := beszelTests.CreateRecord(hub, "system_stats", map[string]any{
+			"system": systemRecord.Id,
+			"type":   "1m",
+			"stats":  string(statsJSON),
+		})
+		require.NoError(t, err)
+		record.SetRaw("created", recordTime.Format(types.DefaultDateLayout))
+		err = hub.SaveNoValidate(record)
+		require.NoError(t, err)
+	}
+
+	combinedDataHigh := &system.CombinedData{
+		Stats: system.Stats{
+			DiskPct:  30,
+			ZfsPools: poolHigh,
+		},
+		Info: system.Info{
+			DiskPct: 30,
+		},
+	}
+
+	systemRecord.Set("updated", now)
+	err = hub.SaveNoValidate(systemRecord)
+	require.NoError(t, err)
+
+	err = am.HandleSystemAlerts(systemRecord, combinedDataHigh)
+	require.NoError(t, err)
+
+	time.Sleep(20 * time.Millisecond)
+
+	diskAlert, err = hub.FindFirstRecordByFilter("alerts", "id={:id}", dbx.Params{"id": diskAlert.Id})
+	require.NoError(t, err)
+	assert.True(t, diskAlert.GetBool("triggered"),
+		"Alert should be triggered when ZFS pool average (92%%) exceeds threshold (80%%)")
+
+	// --- Resolution: pool drops to 50%, alert should resolve ---
+
+	poolLow := map[string]*system.ZfsPool{
+		"tank": {Total: 1000, Used: 500}, // 50% - below threshold
+	}
+
+	newNow := now.Add(2 * time.Minute)
+	for _, offset := range recordTimes {
+		stats := system.Stats{
+			DiskPct:  30,
+			ZfsPools: poolLow,
+		}
+		statsJSON, _ := json.Marshal(stats)
+
+		recordTime := newNow.Add(offset)
+		record, err := beszelTests.CreateRecord(hub, "system_stats", map[string]any{
+			"system": systemRecord.Id,
+			"type":   "1m",
+			"stats":  string(statsJSON),
+		})
+		require.NoError(t, err)
+		record.SetRaw("created", recordTime.Format(types.DefaultDateLayout))
+		err = hub.SaveNoValidate(record)
+		require.NoError(t, err)
+	}
+
+	combinedDataLow := &system.CombinedData{
+		Stats: system.Stats{
+			DiskPct:  30,
+			ZfsPools: poolLow,
+		},
+		Info: system.Info{
+			DiskPct: 30,
+		},
+	}
+
+	systemRecord.Set("updated", newNow)
+	err = hub.SaveNoValidate(systemRecord)
+	require.NoError(t, err)
+
+	err = am.HandleSystemAlerts(systemRecord, combinedDataLow)
+	require.NoError(t, err)
+
+	time.Sleep(20 * time.Millisecond)
+
+	diskAlert, err = hub.FindFirstRecordByFilter("alerts", "id={:id}", dbx.Params{"id": diskAlert.Id})
+	require.NoError(t, err)
+	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

@@ -0,0 +1,15 @@
+//go:build testing
+
+package alerts
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestZfsDiskAlertKeyIsNamespaced(t *testing.T) {
+	assert.Equal(t, "zfs:tank", 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

@@ -0,0 +1,292 @@
+//go:build testing
+
+package alerts_test
+
+import (
+	"testing"
+	"time"
+
+	beszelTests "github.com/henrygd/beszel/internal/tests"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestZfsPoolAlertOnlineToDegraded(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "ONLINE",
+	})
+	assert.NoError(t, err)
+
+	// Re-fetch so PocketBase tracks original values
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+
+	pool.Set("health", "DEGRADED")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+
+	time.Sleep(50 * time.Millisecond)
+
+	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, "Storage pool DEGRADED on test-system")
+	assert.Contains(t, lastMessage.Subject, "tank")
+	assert.Contains(t, lastMessage.Text, "ONLINE to DEGRADED")
+}
+
+func TestZfsPoolAlertDegradedToFaulted(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "rpool",
+		"health": "DEGRADED",
+	})
+	assert.NoError(t, err)
+
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+
+	pool.Set("health", "FAULTED")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+
+	time.Sleep(50 * time.Millisecond)
+
+	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, "Storage pool FAULTED on test-system")
+}
+
+func TestZfsPoolAlertNoAlertOnRecovery(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "DEGRADED",
+	})
+	assert.NoError(t, err)
+
+	// Trigger a worsening alert first
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+	pool.Set("health", "FAULTED")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+	time.Sleep(50 * time.Millisecond)
+	assert.EqualValues(t, 2, hub.TestMailer.TotalSend(), "expected alerts for initial DEGRADED state and DEGRADED -> FAULTED")
+
+	// Recovery back to ONLINE must not send a new alert
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+	pool.Set("health", "ONLINE")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+	time.Sleep(50 * time.Millisecond)
+	assert.EqualValues(t, 2, hub.TestMailer.TotalSend(), "recovery should not send a new alert")
+
+	// And the open history entry should have been resolved
+	history, err := hub.FindRecordsByFilter("alerts_history", "alert_id={:alert_id}", "", 0, 0, map[string]any{"alert_id": pool.Id})
+	assert.NoError(t, err)
+	requireHistoryResolved(t, history)
+}
+
+func TestZfsPoolAlertUnknownHealthDoesNotResolve(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	require.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "DEGRADED",
+	})
+	require.NoError(t, err)
+	time.Sleep(50 * time.Millisecond)
+
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	require.NoError(t, err)
+	pool.Set("health", "")
+	require.NoError(t, hub.Save(pool))
+	time.Sleep(50 * time.Millisecond)
+
+	history, err := hub.FindRecordsByFilter("alerts_history", "alert_id={:alert_id} && resolved=null", "", 0, 0, map[string]any{"alert_id": pool.Id})
+	require.NoError(t, err)
+	require.Len(t, history, 1, "unknown health must not resolve an active alert")
+}
+
+func TestZfsPoolAlertUnknownToFaulted(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "",
+	})
+	assert.NoError(t, err)
+
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+
+	pool.Set("health", "FAULTED")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+
+	time.Sleep(50 * time.Millisecond)
+
+	assert.EqualValues(t, 1, hub.TestMailer.TotalSend(), "should alert when a previously unknown pool becomes FAULTED")
+}
+
+func TestZfsPoolAlertOnInitialUnhealthyState(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	require.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "DEGRADED",
+	})
+	require.NoError(t, err)
+	time.Sleep(50 * time.Millisecond)
+
+	require.EqualValues(t, 1, hub.TestMailer.TotalSend())
+	assert.Contains(t, hub.TestMailer.LastMessage().Text, "first observed as DEGRADED")
+
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	require.NoError(t, err)
+	require.NoError(t, hub.Save(pool))
+	time.Sleep(50 * time.Millisecond)
+	assert.EqualValues(t, 1, hub.TestMailer.TotalSend(), "unchanged unhealthy health must not duplicate alerts")
+}
+
+func TestZfsPoolAlertWritesHistory(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "ONLINE",
+	})
+	assert.NoError(t, err)
+
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+	pool.Set("health", "FAULTED")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+
+	time.Sleep(50 * time.Millisecond)
+
+	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, "Storage Pool: tank", history[0].GetString("name"))
+	assert.Equal(t, system.Id, history[0].GetString("system"))
+}
+
+func TestZfsPoolAlertResolvedOnRecordDelete(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	pool, err := beszelTests.CreateRecord(hub, "zfs_pools", map[string]any{
+		"system": system.Id,
+		"name":   "tank",
+		"health": "ONLINE",
+	})
+	assert.NoError(t, err)
+
+	// Trigger an alert so an open history entry exists.
+	pool, err = hub.FindRecordById("zfs_pools", pool.Id)
+	assert.NoError(t, err)
+	pool.Set("health", "FAULTED")
+	err = hub.Save(pool)
+	assert.NoError(t, err)
+	time.Sleep(50 * time.Millisecond)
+
+	history, err := hub.FindRecordsByFilter("alerts_history", "alert_id={:alert_id} && resolved=null", "", 0, 0, map[string]any{"alert_id": pool.Id})
+	assert.NoError(t, err)
+	require.Len(t, history, 1, "expected one open history entry")
+
+	// Deleting the pool record must resolve the open entry.
+	err = hub.Delete(pool)
+	assert.NoError(t, err)
+	time.Sleep(50 * time.Millisecond)
+
+	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)
+	requireHistoryResolved(t, history)
+}
+
+func requireHistoryResolved(t *testing.T, history []*core.Record) {
+	t.Helper()
+	for _, record := range history {
+		assert.False(t, record.GetDateTime("resolved").Time().IsZero(), "expected history entry to be resolved")
+	}
+}

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

@@ -22,6 +22,10 @@ const (
 	GetSmartData
 	// Request detailed systemd service info from agent
 	GetSystemdInfo
+	// Request ZFS detail data from agent
+	GetZfsData
+	// Sync network monitor configuration to agent
+	SyncNetworkMonitors
 	// Add new actions here...
 )
 
@@ -64,6 +68,10 @@ type DataRequestOptions struct {
 	IncludeDetails bool   `cbor:"1,keyasint"`
 }
 
+type ZfsDataRequest struct {
+	Force bool `cbor:"0,keyasint,omitempty"`
+}
+
 type ContainerLogsRequest struct {
 	ContainerID string `cbor:"0,keyasint"`
 }

internal/dockerfile_agent

@@ -8,6 +8,8 @@ RUN go mod download
 # Copy source files
 COPY . ./
 
+RUN apk add --no-cache ca-certificates && update-ca-certificates
+
 # Build
 ARG TARGETOS TARGETARCH
 RUN CGO_ENABLED=0 GOGC=75 GOOS=$TARGETOS GOARCH=$TARGETARCH go build -ldflags "-w -s" -o /agent ./internal/cmd/agent
@@ -19,6 +21,7 @@ RUN rm -rf /tmp/*
 # --------------------------
 FROM scratch
 COPY --from=builder /agent /agent
+COPY --from=builder /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/
 
 # this is so we don't need to create the /tmp directory in the scratch container
 COPY --from=builder /tmp /tmp
@@ -29,4 +32,4 @@ COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
 # Ensure data persistence across container recreations
 VOLUME ["/var/lib/beszel-agent"]
 
-ENTRYPOINT ["/agent"]
+ENTRYPOINT ["/agent"]

internal/dockerfile_agent_alpine

@@ -23,7 +23,7 @@ COPY --from=builder /agent /agent
 # AMD GPU name lookup (used by agent on Linux when /usr/share/libdrm/amdgpu.ids is read)
 COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
 
-RUN apk add --no-cache smartmontools
+RUN apk add --no-cache smartmontools zfs
 
 # Ensure data persistence across container recreations
 VOLUME ["/var/lib/beszel-agent"]

internal/dockerfile_agent_intel

@@ -20,9 +20,9 @@ FROM alpine:3.23
 
 COPY --from=builder /agent /agent
 
-RUN apk add --no-cache -X https://dl-cdn.alpinelinux.org/alpine/edge/testing igt-gpu-tools nvtop smartmontools
+RUN apk add --no-cache -X https://dl-cdn.alpinelinux.org/alpine/edge/testing igt-gpu-tools nvtop smartmontools zfs
 
 # Ensure data persistence across container recreations
 VOLUME ["/var/lib/beszel-agent"]
 
-ENTRYPOINT ["/agent"]
+ENTRYPOINT ["/agent"]

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"]