[Recheck watcher ctx cancellation under conn.mu in onWGDisconnected

onWGDisconnected only checked conn.ctx (the engine-scoped context), never
the watcher's own context. disableWgWatcherIfNeeded cancels the wgWatcherCtx,
not conn.ctx, so a disabled watcher's timeout callback did not see the
cancellation.

handshakeCheck runs lock-free, so between the ctx check in periodicHandshakeCheck
and acquiring conn.mu a fast disconnect/reconnect can slip in: the stale watcher
then acquires the lock and tears down the *new*, healthy connection based on the
old timeout, forcing the guard into an unnecessary reconnect (flap).

Recheck watcherCtx.Err() under conn.mu so a superseded watcher exits without
touching the connection that replaced it.

.github/workflows/agent-network-e2e.yml

@@ -1,10 +1,10 @@
 name: Agent Network E2E
 
 on:
-  push:
-    branches:
-      - main
-  pull_request:
+  # Nightly at 03:00 UTC, plus on demand from the Actions tab.
+  schedule:
+    - cron: "0 3 * * *"
+  workflow_dispatch:
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
@@ -13,7 +13,6 @@ concurrency:
 jobs:
   e2e:
     name: Agent Network E2E
-    if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
     runs-on: ubuntu-latest
     timeout-minutes: 45
     steps:

.github/workflows/golang-test-linux.yml

@@ -158,7 +158,7 @@ jobs:
         run: git --no-pager diff --exit-code
 
       - name: Test
-        run: CGO_ENABLED=1 GOARCH=${{ matrix.arch }} CI=true go test -coverprofile=coverage.txt -tags devcert -timeout 10m -p 1 $(go list ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined)
+        run: CGO_ENABLED=1 GOARCH=${{ matrix.arch }} CI=true go test -coverprofile=coverage.txt -tags 'devcert privileged' -exec 'sudo --preserve-env=CI,CGO_ENABLED' -timeout 10m -p 1 $(go list ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined -e /client/testutil/privileged)
 
       - name: Upload coverage reports to Codecov
         if: matrix.arch == 'amd64'

.github/workflows/release.yml

@@ -293,8 +293,11 @@ jobs:
           ${{ steps.goreleaser.outputs.artifacts }}
           JSON
 
+          # dockers_v2 artifacts have no top-level goarch field, so match the
+          # per-platform -amd64 tag suffix instead; it works for both the old
+          # dockers and the new dockers_v2 image naming.
           mapfile -t src_images < <(
-            jq -r '.[] | select(.type == "Docker Image") | select(.goarch == "amd64") | .name | select(startswith("ghcr.io/"))' /tmp/goreleaser-artifacts.json
+            jq -r '.[] | select(.type == "Docker Image") | .name | select(startswith("ghcr.io/") and endswith("-amd64"))' /tmp/goreleaser-artifacts.json
           )
 
           for src in "${src_images[@]}"; do

.gitignore

@@ -1,3 +1,4 @@
+.claude
 .idea
 .run
 *.iml

client/android/env_list.go

@@ -10,7 +10,7 @@ var (
 	EnvKeyNBForceRelay = peer.EnvKeyNBForceRelay
 
 	// EnvKeyNBLazyConn Exported for Android java client to configure lazy connection
-	EnvKeyNBLazyConn = lazyconn.EnvEnableLazyConn
+	EnvKeyNBLazyConn = lazyconn.EnvLazyConn
 
 	// EnvKeyNBInactivityThreshold Exported for Android java client to configure connection inactivity threshold
 	EnvKeyNBInactivityThreshold = lazyconn.EnvInactivityThreshold

client/cmd/root.go

@@ -71,12 +71,14 @@ var (
 	extraIFaceBlackList     []string
 	anonymizeFlag           bool
 	dnsRouteInterval        time.Duration
-	lazyConnEnabled         bool
-	mtu                     uint16
-	profilesDisabled        bool
-	updateSettingsDisabled  bool
-	captureEnabled          bool
-	networksDisabled        bool
+	// lazyConnEnabled is the parse target for the deprecated --enable-lazy-connection
+	// flag. The flag is inert; the value is no longer read (use NB_LAZY_CONN instead).
+	lazyConnEnabled        bool
+	mtu                    uint16
+	profilesDisabled       bool
+	updateSettingsDisabled bool
+	captureEnabled         bool
+	networksDisabled       bool
 
 	rootCmd = &cobra.Command{
 		Use:          "netbird",
@@ -210,7 +212,8 @@ func init() {
 	upCmd.PersistentFlags().BoolVar(&rosenpassEnabled, enableRosenpassFlag, false, "[Experimental] Enable Rosenpass feature. If enabled, the connection will be post-quantum secured via Rosenpass.")
 	upCmd.PersistentFlags().BoolVar(&rosenpassPermissive, rosenpassPermissiveFlag, false, "[Experimental] Enable Rosenpass in permissive mode to allow this peer to accept WireGuard connections without requiring Rosenpass functionality from peers that do not have Rosenpass enabled.")
 	upCmd.PersistentFlags().BoolVar(&autoConnectDisabled, disableAutoConnectFlag, false, "Disables auto-connect feature. If enabled, then the client won't connect automatically when the service starts.")
-	upCmd.PersistentFlags().BoolVar(&lazyConnEnabled, enableLazyConnectionFlag, false, "[Experimental] Enable the lazy connection feature. If enabled, the client will establish connections on-demand. Note: this setting may be overridden by management configuration.")
+	upCmd.PersistentFlags().BoolVar(&lazyConnEnabled, enableLazyConnectionFlag, false, "Deprecated: no longer used. Lazy connections are controlled by the server and the NB_LAZY_CONN environment variable.")
+	_ = upCmd.PersistentFlags().MarkDeprecated(enableLazyConnectionFlag, "no longer used; lazy connections are controlled by the server and the NB_LAZY_CONN environment variable")
 
 }
 

client/cmd/up.go

@@ -479,10 +479,6 @@ func setupSetConfigReq(customDNSAddressConverted []byte, cmd *cobra.Command, pro
 		req.DisableIpv6 = &disableIPv6
 	}
 
-	if cmd.Flag(enableLazyConnectionFlag).Changed {
-		req.LazyConnectionEnabled = &lazyConnEnabled
-	}
-
 	return &req
 }
 
@@ -600,9 +596,6 @@ func setupConfig(customDNSAddressConverted []byte, cmd *cobra.Command, configFil
 		ic.DisableIPv6 = &disableIPv6
 	}
 
-	if cmd.Flag(enableLazyConnectionFlag).Changed {
-		ic.LazyConnectionEnabled = &lazyConnEnabled
-	}
 	return &ic, nil
 }
 
@@ -718,9 +711,6 @@ func setupLoginRequest(providedSetupKey string, customDNSAddressConverted []byte
 		loginRequest.DisableIpv6 = &disableIPv6
 	}
 
-	if cmd.Flag(enableLazyConnectionFlag).Changed {
-		loginRequest.LazyConnectionEnabled = &lazyConnEnabled
-	}
 	return &loginRequest, nil
 }
 

client/iface/configurer/kernel_unix.go

@@ -17,12 +17,15 @@ import (
 
 type KernelConfigurer struct {
 	deviceName string
+	statsCache *statsCache
 }
 
 func NewKernelConfigurer(deviceName string) *KernelConfigurer {
-	return &KernelConfigurer{
+	c := &KernelConfigurer{
 		deviceName: deviceName,
 	}
+	c.statsCache = newStatsCache(statsCacheTTL, c.fetchStats)
+	return c
 }
 
 func (c *KernelConfigurer) ConfigureInterface(privateKey string, port int) error {
@@ -246,12 +249,6 @@ func (c *KernelConfigurer) configure(config wgtypes.Config) error {
 		}
 	}()
 
-	// validate if device with name exists
-	_, err = wg.Device(c.deviceName)
-	if err != nil {
-		return err
-	}
-
 	return wg.ConfigureDevice(c.deviceName, config)
 }
 
@@ -300,6 +297,14 @@ func (c *KernelConfigurer) FullStats() (*Stats, error) {
 }
 
 func (c *KernelConfigurer) GetStats() (map[string]WGStats, error) {
+	return c.statsCache.get()
+}
+
+func (c *KernelConfigurer) LastActivities() map[string]monotime.Time {
+	return nil
+}
+
+func (c *KernelConfigurer) fetchStats() (map[string]WGStats, error) {
 	stats := make(map[string]WGStats)
 	wg, err := wgctrl.New()
 	if err != nil {
@@ -326,7 +331,3 @@ func (c *KernelConfigurer) GetStats() (map[string]WGStats, error) {
 	}
 	return stats, nil
 }
-
-func (c *KernelConfigurer) LastActivities() map[string]monotime.Time {
-	return nil
-}

client/iface/configurer/stats_cache.go

@@ -0,0 +1,52 @@
+package configurer
+
+import (
+	"sync"
+	"time"
+
+	"golang.org/x/sync/singleflight"
+)
+
+const statsCacheTTL = 1 * time.Second
+
+type statsCache struct {
+	ttl   time.Duration
+	fetch func() (map[string]WGStats, error)
+
+	mu       sync.RWMutex
+	value    map[string]WGStats
+	expireAt time.Time
+
+	sf singleflight.Group
+}
+
+func newStatsCache(ttl time.Duration, fetch func() (map[string]WGStats, error)) *statsCache {
+	return &statsCache{ttl: ttl, fetch: fetch}
+}
+
+func (c *statsCache) get() (map[string]WGStats, error) {
+	c.mu.RLock()
+	if c.value != nil && time.Now().Before(c.expireAt) {
+		value := c.value
+		c.mu.RUnlock()
+		return value, nil
+	}
+	c.mu.RUnlock()
+
+	value, err, _ := c.sf.Do("stats", func() (interface{}, error) {
+		res, err := c.fetch()
+		if err != nil {
+			return nil, err
+		}
+
+		c.mu.Lock()
+		c.value = res
+		c.expireAt = time.Now().Add(c.ttl)
+		c.mu.Unlock()
+		return res, nil
+	})
+	if err != nil {
+		return nil, err
+	}
+	return value.(map[string]WGStats), nil
+}

client/iface/configurer/stats_cache_test.go

@@ -0,0 +1,70 @@
+package configurer
+
+import (
+	"errors"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestStatsCache_CachesWithinTTL(t *testing.T) {
+	var calls atomic.Int64
+	c := newStatsCache(50*time.Millisecond, func() (map[string]WGStats, error) {
+		calls.Add(1)
+		return map[string]WGStats{"p": {}}, nil
+	})
+
+	for i := 0; i < 10; i++ {
+		_, err := c.get()
+		require.NoError(t, err)
+	}
+	require.Equal(t, int64(1), calls.Load(), "within TTL only one underlying fetch")
+
+	time.Sleep(60 * time.Millisecond)
+	_, err := c.get()
+	require.NoError(t, err)
+	require.Equal(t, int64(2), calls.Load(), "after TTL expiry a fresh fetch happens")
+}
+
+func TestStatsCache_SingleFlight(t *testing.T) {
+	var calls atomic.Int64
+	release := make(chan struct{})
+	c := newStatsCache(time.Minute, func() (map[string]WGStats, error) {
+		calls.Add(1)
+		<-release
+		return map[string]WGStats{}, nil
+	})
+
+	const n = 50
+	var wg sync.WaitGroup
+	wg.Add(n)
+	for i := 0; i < n; i++ {
+		go func() {
+			defer wg.Done()
+			_, _ = c.get()
+		}()
+	}
+	time.Sleep(20 * time.Millisecond)
+	close(release)
+	wg.Wait()
+
+	require.Equal(t, int64(1), calls.Load(), "concurrent misses collapse into one fetch")
+}
+
+func TestStatsCache_ErrorNotCached(t *testing.T) {
+	var calls atomic.Int64
+	wantErr := errors.New("dump failed")
+	c := newStatsCache(time.Minute, func() (map[string]WGStats, error) {
+		calls.Add(1)
+		return nil, wantErr
+	})
+
+	_, err := c.get()
+	require.ErrorIs(t, err, wantErr)
+	_, err = c.get()
+	require.ErrorIs(t, err, wantErr)
+	require.Equal(t, int64(2), calls.Load(), "errors are not cached; each call retries")
+}

client/iface/configurer/usp.go

@@ -40,6 +40,7 @@ type WGUSPConfigurer struct {
 	device           *device.Device
 	deviceName       string
 	activityRecorder *bind.ActivityRecorder
+	statsCache       *statsCache
 
 	uapiListener net.Listener
 }
@@ -50,16 +51,19 @@ func NewUSPConfigurer(device *device.Device, deviceName string, activityRecorder
 		deviceName:       deviceName,
 		activityRecorder: activityRecorder,
 	}
+	wgCfg.statsCache = newStatsCache(statsCacheTTL, wgCfg.fetchStats)
 	wgCfg.startUAPI()
 	return wgCfg
 }
 
 func NewUSPConfigurerNoUAPI(device *device.Device, deviceName string, activityRecorder *bind.ActivityRecorder) *WGUSPConfigurer {
-	return &WGUSPConfigurer{
+	wgCfg := &WGUSPConfigurer{
 		device:           device,
 		deviceName:       deviceName,
 		activityRecorder: activityRecorder,
 	}
+	wgCfg.statsCache = newStatsCache(statsCacheTTL, wgCfg.fetchStats)
+	return wgCfg
 }
 
 func (c *WGUSPConfigurer) ConfigureInterface(privateKey string, port int) error {
@@ -348,6 +352,10 @@ func (t *WGUSPConfigurer) Close() {
 }
 
 func (t *WGUSPConfigurer) GetStats() (map[string]WGStats, error) {
+	return t.statsCache.get()
+}
+
+func (t *WGUSPConfigurer) fetchStats() (map[string]WGStats, error) {
 	ipc, err := t.device.IpcGet()
 	if err != nil {
 		return nil, fmt.Errorf("ipc get: %w", err)

client/internal/auth/auth.go

@@ -322,7 +322,6 @@ func (a *Auth) setSystemInfoFlags(info *system.Info) {
 		a.config.BlockLANAccess,
 		a.config.BlockInbound,
 		a.config.DisableIPv6,
-		a.config.LazyConnectionEnabled,
 		a.config.EnableSSHRoot,
 		a.config.EnableSSHSFTP,
 		a.config.EnableSSHLocalPortForwarding,

client/internal/conn_mgr.go

@@ -16,6 +16,16 @@ import (
 	"github.com/netbirdio/netbird/route"
 )
 
+// lazyForce is the resolved local decision for lazy connections, layered above the
+// management feature flag. lazyForceNone defers to management.
+type lazyForce int
+
+const (
+	lazyForceNone lazyForce = iota
+	lazyForceOn
+	lazyForceOff
+)
+
 // ConnMgr coordinates both lazy connections (established on-demand) and permanent peer connections.
 //
 // The connection manager is responsible for:
@@ -28,7 +38,7 @@ type ConnMgr struct {
 	peerStore        *peerstore.Store
 	statusRecorder   *peer.Status
 	iface            lazyconn.WGIface
-	enabledLocally   bool
+	force            lazyForce
 	rosenpassEnabled bool
 
 	lazyConnMgr *manager.Manager
@@ -43,28 +53,34 @@ func NewConnMgr(engineConfig *EngineConfig, statusRecorder *peer.Status, peerSto
 		peerStore:        peerStore,
 		statusRecorder:   statusRecorder,
 		iface:            iface,
+		force:            resolveLazyForce(engineConfig.LazyConnection),
 		rosenpassEnabled: engineConfig.RosenpassEnabled,
 	}
-	if engineConfig.LazyConnectionEnabled || lazyconn.IsLazyConnEnabledByEnv() {
-		e.enabledLocally = true
-	}
 	return e
 }
 
-// Start initializes the connection manager and starts the lazy connection manager if enabled by env var or cmd line option.
+// Start initializes the connection manager. It starts the lazy connection manager when a
+// local override forces it on; with no local override it waits for the management feature flag.
 func (e *ConnMgr) Start(ctx context.Context) {
 	if e.lazyConnMgr != nil {
 		log.Errorf("lazy connection manager is already started")
 		return
 	}
 
-	if !e.enabledLocally {
-		log.Infof("lazy connection manager is disabled")
+	switch e.force {
+	case lazyForceOff:
+		log.Infof("lazy connection manager is disabled by local override (%s or MDM policy)", lazyconn.EnvLazyConn)
+		e.statusRecorder.UpdateLazyConnection(false)
+		return
+	case lazyForceNone:
+		log.Infof("lazy connection manager is managed by the management feature flag")
+		e.statusRecorder.UpdateLazyConnection(false)
 		return
 	}
 
 	if e.rosenpassEnabled {
 		log.Warnf("rosenpass connection manager is enabled, lazy connection manager will not be started")
+		e.statusRecorder.UpdateLazyConnection(false)
 		return
 	}
 
@@ -76,8 +92,8 @@ func (e *ConnMgr) Start(ctx context.Context) {
 // If enabled, it initializes the lazy connection manager and start it. Do not need to call Start() again.
 // If disabled, then it closes the lazy connection manager and open the connections to all peers.
 func (e *ConnMgr) UpdatedRemoteFeatureFlag(ctx context.Context, enabled bool) error {
-	// do not disable lazy connection manager if it was enabled by env var
-	if e.enabledLocally {
+	// a local override (NB_LAZY_CONN or local config) takes precedence over management
+	if e.force != lazyForceNone {
 		return nil
 	}
 
@@ -89,6 +105,7 @@ func (e *ConnMgr) UpdatedRemoteFeatureFlag(ctx context.Context, enabled bool) er
 
 		if e.rosenpassEnabled {
 			log.Infof("rosenpass connection manager is enabled, lazy connection manager will not be started")
+			e.statusRecorder.UpdateLazyConnection(false)
 			return nil
 		}
 
@@ -98,6 +115,7 @@ func (e *ConnMgr) UpdatedRemoteFeatureFlag(ctx context.Context, enabled bool) er
 		return e.addPeersToLazyConnManager()
 	} else {
 		if e.lazyConnMgr == nil {
+			e.statusRecorder.UpdateLazyConnection(false)
 			return nil
 		}
 		log.Infof("lazy connection manager is disabled by management feature flag")
@@ -309,6 +327,25 @@ func (e *ConnMgr) isStartedWithLazyMgr() bool {
 	return e.lazyConnMgr != nil && e.lazyCtxCancel != nil
 }
 
+// resolveLazyForce determines the local override. NB_LAZY_CONN takes precedence; when it
+// is unset the MDM policy override (mdmState) applies. Either wins in both directions over
+// the management feature flag; StateUnset for both defers to management.
+func resolveLazyForce(mdmState lazyconn.State) lazyForce {
+	state := lazyconn.EnvState()
+	if state == lazyconn.StateUnset {
+		state = mdmState
+	}
+
+	switch state {
+	case lazyconn.StateOn:
+		return lazyForceOn
+	case lazyconn.StateOff:
+		return lazyForceOff
+	default:
+		return lazyForceNone
+	}
+}
+
 func inactivityThresholdEnv() *time.Duration {
 	envValue := os.Getenv(lazyconn.EnvInactivityThreshold)
 	if envValue == "" {

client/internal/conn_mgr_test.go

@@ -0,0 +1,40 @@
+package internal
+
+import (
+	"os"
+	"testing"
+
+	"github.com/netbirdio/netbird/client/internal/lazyconn"
+)
+
+func TestResolveLazyForce(t *testing.T) {
+	tests := []struct {
+		name   string
+		env    string
+		envSet bool
+		mdm    lazyconn.State
+		want   lazyForce
+	}{
+		{name: "env unset, mdm unset -> defer to management", mdm: lazyconn.StateUnset, want: lazyForceNone},
+		{name: "env on -> force on", env: "on", envSet: true, mdm: lazyconn.StateUnset, want: lazyForceOn},
+		{name: "env off -> force off", env: "off", envSet: true, mdm: lazyconn.StateUnset, want: lazyForceOff},
+		{name: "env unset, mdm on -> force on", mdm: lazyconn.StateOn, want: lazyForceOn},
+		{name: "env unset, mdm off -> force off", mdm: lazyconn.StateOff, want: lazyForceOff},
+		{name: "env on beats mdm off", env: "on", envSet: true, mdm: lazyconn.StateOff, want: lazyForceOn},
+		{name: "env off beats mdm on", env: "off", envSet: true, mdm: lazyconn.StateOn, want: lazyForceOff},
+		{name: "unrecognized env, mdm on -> mdm wins", env: "auto", envSet: true, mdm: lazyconn.StateOn, want: lazyForceOn},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Setenv(lazyconn.EnvLazyConn, tt.env)
+			if !tt.envSet {
+				os.Unsetenv(lazyconn.EnvLazyConn)
+			}
+
+			if got := resolveLazyForce(tt.mdm); got != tt.want {
+				t.Fatalf("resolveLazyForce(%v) = %v, want %v", tt.mdm, got, tt.want)
+			}
+		})
+	}
+}

client/internal/connect.go

@@ -27,6 +27,7 @@ import (
 	"github.com/netbirdio/netbird/client/iface/device"
 	"github.com/netbirdio/netbird/client/iface/netstack"
 	"github.com/netbirdio/netbird/client/internal/dns"
+	"github.com/netbirdio/netbird/client/internal/lazyconn"
 	"github.com/netbirdio/netbird/client/internal/listener"
 	"github.com/netbirdio/netbird/client/internal/metrics"
 	"github.com/netbirdio/netbird/client/internal/peer"
@@ -601,7 +602,7 @@ func createEngineConfig(key wgtypes.Key, config *profilemanager.Config, peerConf
 		BlockInbound:        config.BlockInbound,
 		DisableIPv6:         config.DisableIPv6,
 
-		LazyConnectionEnabled: config.LazyConnectionEnabled,
+		LazyConnection: lazyconn.ParseState(config.LazyConnection),
 
 		MTU:     selectMTU(config.MTU, peerConfig.Mtu),
 		LogPath: logPath,
@@ -675,7 +676,6 @@ func loginToManagement(ctx context.Context, client mgm.Client, pubSSHKey []byte,
 		config.BlockLANAccess,
 		config.BlockInbound,
 		config.DisableIPv6,
-		config.LazyConnectionEnabled,
 		config.EnableSSHRoot,
 		config.EnableSSHSFTP,
 		config.EnableSSHLocalPortForwarding,

client/internal/debug/debug.go

@@ -681,7 +681,7 @@ func (g *BundleGenerator) addCommonConfigFields(configContent *strings.Builder)
 		configContent.WriteString(fmt.Sprintf("ClientCertKeyPath: %s\n", g.internalConfig.ClientCertKeyPath))
 	}
 
-	configContent.WriteString(fmt.Sprintf("LazyConnectionEnabled: %v\n", g.internalConfig.LazyConnectionEnabled))
+	configContent.WriteString(fmt.Sprintf("LazyConnection: %q\n", g.internalConfig.LazyConnection))
 	configContent.WriteString(fmt.Sprintf("MTU: %d\n", g.internalConfig.MTU))
 }
 

client/internal/debug/debug_test.go

@@ -885,7 +885,7 @@ func TestAddConfig_AllFieldsCovered(t *testing.T) {
 		DNSRouteInterval:              5 * time.Second,
 		ClientCertPath:                "/tmp/cert",
 		ClientCertKeyPath:             "/tmp/key",
-		LazyConnectionEnabled:         true,
+		LazyConnection:                "on",
 		MTU:                           1280,
 	}
 

client/internal/engine.go

@@ -40,6 +40,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/dnsfwd"
 	"github.com/netbirdio/netbird/client/internal/expose"
 	"github.com/netbirdio/netbird/client/internal/ingressgw"
+	"github.com/netbirdio/netbird/client/internal/lazyconn"
 	"github.com/netbirdio/netbird/client/internal/metrics"
 	"github.com/netbirdio/netbird/client/internal/netflow"
 	nftypes "github.com/netbirdio/netbird/client/internal/netflow/types"
@@ -147,7 +148,9 @@ type EngineConfig struct {
 	BlockInbound        bool
 	DisableIPv6         bool
 
-	LazyConnectionEnabled bool
+	// LazyConnection is the MDM-sourced lazy-connection override; StateUnset defers to
+	// the env var and management feature flag.
+	LazyConnection lazyconn.State
 
 	MTU uint16
 
@@ -1130,7 +1133,6 @@ func (e *Engine) applyInfoFlags(info *system.Info) {
 		e.config.BlockLANAccess,
 		e.config.BlockInbound,
 		e.config.DisableIPv6,
-		e.config.LazyConnectionEnabled,
 		e.config.EnableSSHRoot,
 		e.config.EnableSSHSFTP,
 		e.config.EnableSSHLocalPortForwarding,
@@ -1999,7 +2001,6 @@ func (e *Engine) readInitialSettings() ([]*route.Route, *nbdns.Config, bool, err
 		e.config.BlockLANAccess,
 		e.config.BlockInbound,
 		e.config.DisableIPv6,
-		e.config.LazyConnectionEnabled,
 		e.config.EnableSSHRoot,
 		e.config.EnableSSHSFTP,
 		e.config.EnableSSHLocalPortForwarding,

client/internal/lazyconn/env.go

@@ -3,24 +3,57 @@ package lazyconn
 import (
 	"os"
 	"strconv"
+	"strings"
 
 	log "github.com/sirupsen/logrus"
 )
 
 const (
-	EnvEnableLazyConn      = "NB_ENABLE_EXPERIMENTAL_LAZY_CONN"
+	EnvLazyConn            = "NB_LAZY_CONN"
 	EnvInactivityThreshold = "NB_LAZY_CONN_INACTIVITY_THRESHOLD"
 )
 
-func IsLazyConnEnabledByEnv() bool {
-	val := os.Getenv(EnvEnableLazyConn)
-	if val == "" {
-		return false
-	}
-	enabled, err := strconv.ParseBool(val)
-	if err != nil {
-		log.Warnf("failed to parse %s: %v", EnvEnableLazyConn, err)
-		return false
-	}
-	return enabled
+// State is the tri-state local override for lazy connections read from the environment.
+type State int
+
+const (
+	// StateUnset means no local override; defer to the management feature flag.
+	StateUnset State = iota
+	// StateOn forces lazy connections on, overriding management.
+	StateOn
+	// StateOff forces lazy connections off, overriding management.
+	StateOff
+)
+
+// EnvState reads NB_LAZY_CONN and returns the local override state.
+func EnvState() State {
+	return ParseState(os.Getenv(EnvLazyConn))
+}
+
+// ParseState interprets a lazy-connection override value (from the environment or an MDM
+// policy). It accepts the on/off aliases plus any value strconv.ParseBool understands
+// (true/false/1/0). An empty or unrecognized value returns StateUnset so that the
+// management feature flag remains in control.
+func ParseState(raw string) State {
+	if raw == "" {
+		return StateUnset
+	}
+
+	normalized := strings.ToLower(strings.TrimSpace(raw))
+	switch normalized {
+	case "on":
+		return StateOn
+	case "off":
+		return StateOff
+	}
+
+	enabled, err := strconv.ParseBool(normalized)
+	if err != nil {
+		log.Warnf("failed to parse lazy connection value %q (from %s env or MDM policy): %v", raw, EnvLazyConn, err)
+		return StateUnset
+	}
+	if enabled {
+		return StateOn
+	}
+	return StateOff
 }

client/internal/lazyconn/env_test.go

@@ -0,0 +1,45 @@
+package lazyconn
+
+import (
+	"os"
+	"testing"
+)
+
+func TestEnvState(t *testing.T) {
+	tests := []struct {
+		value string
+		set   bool
+		want  State
+	}{
+		{set: false, want: StateUnset},
+		{value: "", set: true, want: StateUnset},
+		{value: "on", set: true, want: StateOn},
+		{value: "ON", set: true, want: StateOn},
+		{value: "true", set: true, want: StateOn},
+		{value: "1", set: true, want: StateOn},
+		{value: " on ", set: true, want: StateOn},
+		{value: "off", set: true, want: StateOff},
+		{value: "OFF", set: true, want: StateOff},
+		{value: "false", set: true, want: StateOff},
+		{value: "0", set: true, want: StateOff},
+		{value: "auto", set: true, want: StateUnset},
+		{value: "garbage", set: true, want: StateUnset},
+	}
+
+	for _, tt := range tests {
+		name := tt.value
+		if !tt.set {
+			name = "unset"
+		}
+		t.Run(name, func(t *testing.T) {
+			t.Setenv(EnvLazyConn, tt.value)
+			if !tt.set {
+				os.Unsetenv(EnvLazyConn)
+			}
+
+			if got := EnvState(); got != tt.want {
+				t.Fatalf("EnvState() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}

client/internal/netflow/logger/logger.go

@@ -27,7 +27,7 @@ type Logger struct {
 	wgIfaceNetV6       netip.Prefix
 	dnsCollection      atomic.Bool
 	exitNodeCollection atomic.Bool
-	Store              types.AggregatingStore
+	Store              types.Store
 }
 
 func New(statusRecorder *peer.Status, wgIfaceIPNet, wgIfaceIPNetV6 netip.Prefix) *Logger {
@@ -35,7 +35,7 @@ func New(statusRecorder *peer.Status, wgIfaceIPNet, wgIfaceIPNetV6 netip.Prefix)
 		statusRecorder: statusRecorder,
 		wgIfaceNet:     wgIfaceIPNet,
 		wgIfaceNetV6:   wgIfaceIPNetV6,
-		Store:          store.NewAggregatingMemoryStore(),
+		Store:          store.NewMemoryStore(),
 	}
 }
 
@@ -125,10 +125,6 @@ func (l *Logger) stop() {
 	l.mux.Unlock()
 }
 
-func (l *Logger) ResetAggregationWindow() types.FlowEventAggregator {
-	return l.Store.ResetAggregationWindow()
-}
-
 func (l *Logger) GetEvents() []*types.Event {
 	return l.Store.GetEvents()
 }

client/internal/netflow/manager.go

@@ -9,14 +9,12 @@ import (
 	"sync"
 	"time"
 
-	"github.com/cenkalti/backoff/v4"
 	"github.com/google/uuid"
 	log "github.com/sirupsen/logrus"
 	"google.golang.org/protobuf/types/known/timestamppb"
 
 	"github.com/netbirdio/netbird/client/internal/netflow/conntrack"
 	"github.com/netbirdio/netbird/client/internal/netflow/logger"
-	"github.com/netbirdio/netbird/client/internal/netflow/store"
 	nftypes "github.com/netbirdio/netbird/client/internal/netflow/types"
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/flow/client"
@@ -25,16 +23,14 @@ import (
 
 // Manager handles netflow tracking and logging
 type Manager struct {
-	mux               sync.Mutex
-	shutdownWg        sync.WaitGroup
-	logger            nftypes.FlowLogger
-	flowConfig        *nftypes.FlowConfig
-	conntrack         nftypes.ConnTracker
-	receiverClient    *client.GRPCClient
-	eventsWithoutAcks nftypes.Store
-	publicKey         []byte
-	cancel            context.CancelFunc
-	retryInterval     time.Duration
+	mux            sync.Mutex
+	shutdownWg     sync.WaitGroup
+	logger         nftypes.FlowLogger
+	flowConfig     *nftypes.FlowConfig
+	conntrack      nftypes.ConnTracker
+	receiverClient *client.GRPCClient
+	publicKey      []byte
+	cancel         context.CancelFunc
 }
 
 // NewManager creates a new netflow manager
@@ -52,11 +48,9 @@ func NewManager(iface nftypes.IFaceMapper, publicKey []byte, statusRecorder *pee
 	}
 
 	return &Manager{
-		logger:            flowLogger,
-		conntrack:         ct,
-		publicKey:         publicKey,
-		retryInterval:     time.Second,
-		eventsWithoutAcks: store.NewMemoryStore(),
+		logger:    flowLogger,
+		conntrack: ct,
+		publicKey: publicKey,
 	}
 }
 
@@ -72,7 +66,6 @@ func (m *Manager) needsNewClient(previous *nftypes.FlowConfig) bool {
 }
 
 // enableFlow starts components for flow tracking
-// must be called under m.mux lock
 func (m *Manager) enableFlow(previous *nftypes.FlowConfig) error {
 	// first make sender ready so events don't pile up
 	if m.needsNewClient(previous) {
@@ -92,7 +85,6 @@ func (m *Manager) enableFlow(previous *nftypes.FlowConfig) error {
 	return nil
 }
 
-// must be called under m.mux lock
 func (m *Manager) resetClient() error {
 	if m.receiverClient != nil {
 		if err := m.receiverClient.Close(); err != nil {
@@ -115,19 +107,14 @@ func (m *Manager) resetClient() error {
 	ctx, cancel := context.WithCancel(context.Background())
 	m.cancel = cancel
 
-	m.shutdownWg.Add(3)
-	flowConfigInterval := m.flowConfig.Interval
+	m.shutdownWg.Add(2)
 	go func() {
 		defer m.shutdownWg.Done()
-		m.receiveACKs(ctx, flowClient, flowConfigInterval)
+		m.receiveACKs(ctx, flowClient)
 	}()
 	go func() {
 		defer m.shutdownWg.Done()
-		m.startSender(ctx, flowConfigInterval)
-	}()
-	go func() {
-		defer m.shutdownWg.Done()
-		m.startRetries(ctx, flowConfigInterval)
+		m.startSender(ctx)
 	}()
 
 	return nil
@@ -211,8 +198,8 @@ func (m *Manager) GetLogger() nftypes.FlowLogger {
 	return m.logger
 }
 
-func (m *Manager) startSender(ctx context.Context, flowConfigInterval time.Duration) {
-	ticker := time.NewTicker(flowConfigInterval)
+func (m *Manager) startSender(ctx context.Context) {
+	ticker := time.NewTicker(m.flowConfig.Interval)
 	defer ticker.Stop()
 
 	for {
@@ -220,29 +207,27 @@ func (m *Manager) startSender(ctx context.Context, flowConfigInterval time.Durat
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			collectedEvents := m.logger.ResetAggregationWindow()
-			events := collectedEvents.GetAggregatedEvents()
+			events := m.logger.GetEvents()
 			for _, event := range events {
-				m.eventsWithoutAcks.StoreEvent(event)
 				if err := m.send(event); err != nil {
 					log.Errorf("failed to send flow event to server: %v", err)
-				} else {
-					log.Tracef("sent flow event: %s", event.ID)
+					continue
 				}
+				log.Tracef("sent flow event: %s", event.ID)
 			}
 		}
 	}
 }
 
-func (m *Manager) receiveACKs(ctx context.Context, client *client.GRPCClient, flowConfigInterval time.Duration) {
-	err := client.Receive(ctx, flowConfigInterval, func(ack *proto.FlowEventAck) error {
+func (m *Manager) receiveACKs(ctx context.Context, client *client.GRPCClient) {
+	err := client.Receive(ctx, m.flowConfig.Interval, func(ack *proto.FlowEventAck) error {
 		id, err := uuid.FromBytes(ack.EventId)
 		if err != nil {
 			log.Warnf("failed to convert ack event id to uuid: %v", err)
 			return nil
 		}
 		log.Tracef("received flow event ack: %s", id)
-		m.eventsWithoutAcks.DeleteEvents([]uuid.UUID{id})
+		m.logger.DeleteEvents([]uuid.UUID{id})
 		return nil
 	})
 
@@ -251,51 +236,6 @@ func (m *Manager) receiveACKs(ctx context.Context, client *client.GRPCClient, fl
 	}
 }
 
-// We effectively never drop events (see MaxInterval), which makes eventsWithoutAcks unbounded.
-// We may want to limit the max size of the store, and start dropping oldest events when the threshold is reached.
-func (m *Manager) startRetries(ctx context.Context, flowConfigInterval time.Duration) {
-	timer := time.NewTimer(m.retryInterval)
-	retryBackoff := backoff.WithContext(&backoff.ExponentialBackOff{
-		InitialInterval:     1 * time.Second,
-		RandomizationFactor: 0.5,
-		Multiplier:          1.7,
-		MaxInterval:         flowConfigInterval / 2,
-		MaxElapsedTime:      3 * 30 * 24 * time.Hour, // 3 months
-		Stop:                backoff.Stop,
-		Clock:               backoff.SystemClock,
-	}, ctx)
-	defer timer.Stop()
-
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case <-timer.C:
-			resetBackoff := true
-			for _, e := range m.eventsWithoutAcks.GetEvents() {
-				if e.Timestamp.Add(time.Second).After(time.Now()) {
-					// grace period on retries to avoid early retries
-					// do not retry if the event is less than 1 sec old
-					continue
-				}
-				if err := m.send(e); err != nil {
-					if nextBackoff := retryBackoff.NextBackOff(); nextBackoff != backoff.Stop {
-						timer = time.NewTimer(nextBackoff)
-						resetBackoff = false
-					} else {
-						resetBackoff = true // we exhausted retries, reset retry loop
-					}
-					break
-				}
-			}
-			if resetBackoff { // use regular retry interval in absence of network errors
-				retryBackoff.Reset()
-				timer = time.NewTimer(m.retryInterval)
-			}
-		}
-	}
-}
-
 func (m *Manager) send(event *nftypes.Event) error {
 	m.mux.Lock()
 	client := m.receiverClient
@@ -310,11 +250,9 @@ func (m *Manager) send(event *nftypes.Event) error {
 
 func toProtoEvent(publicKey []byte, event *nftypes.Event) *proto.FlowEvent {
 	protoEvent := &proto.FlowEvent{
-		EventId:     event.ID[:],
-		Timestamp:   timestamppb.New(event.Timestamp),
-		PublicKey:   publicKey,
-		WindowStart: timestamppb.New(event.WindowStart),
-		WindowEnd:   timestamppb.New(event.WindowEnd),
+		EventId:   event.ID[:],
+		Timestamp: timestamppb.New(event.Timestamp),
+		PublicKey: publicKey,
 		FlowFields: &proto.FlowFields{
 			FlowId:           event.FlowID[:],
 			RuleId:           event.RuleID,
@@ -329,9 +267,6 @@ func toProtoEvent(publicKey []byte, event *nftypes.Event) *proto.FlowEvent {
 			TxBytes:          event.TxBytes,
 			SourceResourceId: event.SourceResourceID,
 			DestResourceId:   event.DestResourceID,
-			NumOfStarts:      event.NumOfStarts,
-			NumOfEnds:        event.NumOfEnds,
-			NumOfDrops:       event.NumOfDrops,
 		},
 	}
 

client/internal/netflow/manager_integration_test.go

@@ -1,291 +0,0 @@
-package netflow
-
-import (
-	"context"
-	"errors"
-	"fmt"
-	"net"
-	"net/netip"
-	"slices"
-	"testing"
-	"time"
-
-	"github.com/google/uuid"
-	"github.com/netbirdio/netbird/client/iface/wgaddr"
-	"github.com/netbirdio/netbird/client/internal/netflow/types"
-	"github.com/netbirdio/netbird/flow/proto"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"google.golang.org/grpc"
-)
-
-type testServer struct {
-	proto.UnimplementedFlowServiceServer
-	events         chan *proto.FlowEvent
-	acks           chan *proto.FlowEventAck
-	grpcSrv        *grpc.Server
-	addr           string
-	handlerDone    chan struct{} // signaled each time Events() exits
-	handlerStarted chan struct{} // signaled each time Events() begins
-}
-
-func newTestServer(t *testing.T) *testServer {
-	listener, err := net.Listen("tcp", "127.0.0.1:0")
-	require.NoError(t, err)
-
-	s := &testServer{
-		events:         make(chan *proto.FlowEvent, 100),
-		acks:           make(chan *proto.FlowEventAck, 100),
-		grpcSrv:        grpc.NewServer(),
-		addr:           listener.Addr().String(),
-		handlerDone:    make(chan struct{}, 10),
-		handlerStarted: make(chan struct{}, 10),
-	}
-
-	proto.RegisterFlowServiceServer(s.grpcSrv, s)
-
-	go func() {
-		if err := s.grpcSrv.Serve(listener); err != nil && !errors.Is(err, grpc.ErrServerStopped) {
-			t.Logf("server error: %v", err)
-		}
-	}()
-
-	t.Cleanup(func() {
-		s.grpcSrv.Stop()
-	})
-
-	return s
-}
-
-func (s *testServer) Events(stream proto.FlowService_EventsServer) error {
-	defer func() {
-		select {
-		case s.handlerDone <- struct{}{}:
-		default:
-		}
-	}()
-
-	err := stream.Send(&proto.FlowEventAck{IsInitiator: true})
-	if err != nil {
-		return err
-	}
-
-	select {
-	case s.handlerStarted <- struct{}{}:
-	default:
-	}
-
-	ctx, cancel := context.WithCancel(stream.Context())
-	defer cancel()
-
-	go func() {
-		defer cancel()
-		for {
-			event, err := stream.Recv()
-			if err != nil {
-				return
-			}
-
-			if !event.IsInitiator {
-				select {
-				case s.events <- event:
-				case <-ctx.Done():
-					return
-				}
-			}
-		}
-	}()
-
-	for {
-		select {
-		case ack := <-s.acks:
-			if err := stream.Send(ack); err != nil {
-				return err
-			}
-		case <-ctx.Done():
-			return ctx.Err()
-		}
-	}
-}
-
-func TestSendEventReceiveAck(t *testing.T) {
-	_, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-	t.Cleanup(cancel)
-
-	server := newTestServer(t)
-	manager := createManager(t, server.addr, 60*time.Second) // set high to prevent retries in this test
-	defer manager.Close()
-
-	assert.Eventually(t, func() bool {
-		select {
-		case <-server.handlerStarted:
-			return true
-		default:
-			return false
-		}
-	}, 3*time.Second, 100*time.Millisecond)
-
-	event1 := types.EventFields{
-		FlowID:    uuid.New(),
-		Type:      types.TypeStart,
-		Direction: types.Ingress,
-		DestIP:    ipAddr("172.16.1.2"),
-		DestPort:  2345,
-		Protocol:  6,
-	}
-	manager.logger.StoreEvent(event1)
-	event2 := types.EventFields{
-		FlowID:    uuid.New(),
-		Type:      types.TypeStart,
-		Direction: types.Ingress,
-		DestIP:    ipAddr("172.16.1.1"),
-		DestPort:  1234,
-		Protocol:  6,
-	}
-	manager.logger.StoreEvent(event2)
-
-	// verify the server received logged events
-	serverSideEvents := make([]*proto.FlowEvent, 0)
-	assert.Eventually(t, func() bool {
-		select {
-		case event := <-server.events:
-			serverSideEvents = append(serverSideEvents, event)
-			if len(serverSideEvents) == 2 {
-				return true
-			}
-		default:
-			if len(serverSideEvents) == 2 {
-				return true
-			}
-		}
-		return false
-	}, 5*time.Second, 100*time.Millisecond)
-
-	serverSideFlowIds := make([]uuid.UUID, 0, 2)
-	slices.Values(serverSideEvents)(func(e *proto.FlowEvent) bool {
-		id, err := uuid.FromBytes(e.FlowFields.FlowId)
-		assert.NoError(t, err)
-		serverSideFlowIds = append(serverSideFlowIds, id)
-		return true
-	})
-	assert.ElementsMatch(t, []uuid.UUID{event1.FlowID, event2.FlowID}, serverSideFlowIds)
-
-	// verify the manager tracks un-acked events
-	unackedEvents := manager.eventsWithoutAcks.GetEvents()
-	assert.Len(t, unackedEvents, 2)
-	flowIds := make([]uuid.UUID, 0)
-	slices.Values(unackedEvents)(func(e *types.Event) bool {
-		flowIds = append(flowIds, e.FlowID)
-		return true
-	})
-	assert.ElementsMatch(t, flowIds, []uuid.UUID{event1.FlowID, event2.FlowID})
-}
-
-// verify handling of retries:
-//   - unacked events are retried
-//   - when acks arrive, events are removed from the un-acked event tracker
-func TestRetryEvents(t *testing.T) {
-	_, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-	t.Cleanup(cancel)
-
-	server := newTestServer(t)
-	manager := createManager(t, server.addr, time.Second) // set low to start retries sooner
-	defer manager.Close()
-
-	assert.Eventually(t, func() bool {
-		select {
-		case <-server.handlerStarted:
-			return true
-		default:
-			return false
-		}
-	}, 3*time.Second, 100*time.Millisecond)
-
-	event1 := types.EventFields{
-		FlowID:    uuid.New(),
-		Type:      types.TypeStart,
-		Direction: types.Ingress,
-		DestIP:    ipAddr("172.16.1.2"),
-		DestPort:  2345,
-		Protocol:  6,
-	}
-	manager.logger.StoreEvent(event1)
-	event2 := types.EventFields{
-		FlowID:    uuid.New(),
-		Type:      types.TypeStart,
-		Direction: types.Ingress,
-		DestIP:    ipAddr("172.16.1.1"),
-		DestPort:  1234,
-		Protocol:  6,
-	}
-	manager.logger.StoreEvent(event2)
-
-	// verify the server received retries of logged events
-	serverSideEvents := make([]*proto.FlowEvent, 0)
-	func() {
-		c := time.After(2500 * time.Millisecond)
-		for {
-			select {
-			case event := <-server.events:
-				serverSideEvents = append(serverSideEvents, event)
-			case <-c:
-				return
-			}
-		}
-	}()
-	assert.True(t, len(serverSideEvents) > 2) // must see retries
-
-	uniqueServerSideEvents := make(map[uuid.UUID]*proto.FlowEvent)
-	slices.Values(serverSideEvents)(func(e *proto.FlowEvent) bool {
-		id, err := uuid.FromBytes(e.FlowFields.FlowId)
-		assert.NoError(t, err)
-		uniqueServerSideEvents[id] = e
-		return true
-	})
-	assert.Contains(t, uniqueServerSideEvents, event1.FlowID)
-	assert.Contains(t, uniqueServerSideEvents, event2.FlowID)
-
-	// ack events
-	server.acks <- &proto.FlowEventAck{EventId: uniqueServerSideEvents[event1.FlowID].EventId}
-	server.acks <- &proto.FlowEventAck{EventId: uniqueServerSideEvents[event2.FlowID].EventId}
-
-	assert.EventuallyWithT(t, func(c *assert.CollectT) {
-		unackedEvents := manager.eventsWithoutAcks.GetEvents()
-		assert.Empty(c, unackedEvents)
-
-	}, 3*time.Second, 100*time.Millisecond)
-}
-
-func createManager(t *testing.T, serverAddr string, retryInterval time.Duration) *Manager {
-	t.Helper()
-
-	mockIFace := &mockIFaceMapper{
-		address: wgaddr.Address{
-			Network: netip.MustParsePrefix("192.168.1.1/32"),
-		},
-		isUserspaceBind: true,
-	}
-
-	publicKey := []byte("test-public-key")
-	manager := NewManager(mockIFace, publicKey, nil)
-	manager.retryInterval = retryInterval
-
-	initialConfig := &types.FlowConfig{
-		Enabled:        true,
-		URL:            fmt.Sprintf("http://%s", serverAddr),
-		TokenPayload:   "initial-payload",
-		TokenSignature: "initial-signature",
-		Interval:       500 * time.Millisecond,
-	}
-
-	err := manager.Update(initialConfig)
-	require.NoError(t, err)
-
-	return manager
-}
-
-func ipAddr(a string) netip.Addr {
-	addr, _ := netip.ParseAddr(a)
-	return addr
-}

client/internal/netflow/store/event_aggregation_test.go

@@ -1,362 +0,0 @@
-package store
-
-import (
-	"math/rand"
-	"net/netip"
-	"testing"
-	"time"
-
-	"github.com/google/uuid"
-	"github.com/netbirdio/netbird/client/internal/netflow/types"
-	"github.com/stretchr/testify/assert"
-)
-
-var random = rand.New(rand.NewSource(time.Now().UnixNano()))
-
-func TestFlowAggregation(t *testing.T) {
-	var protocols = []types.Protocol{types.ICMP, types.ICMPv6, types.TCP, types.UDP}
-	var tests = []struct {
-		description string
-		addresses   [][]netip.Addr
-		dstPort     uint16
-		eventTypes  []types.Type
-	}{
-		{
-			description: "start and stop",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeStart, types.TypeEnd},
-		},
-		{
-			description: "start and drop",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeStart, types.TypeDrop},
-		},
-		{
-			description: "start only",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeStart},
-		},
-		{
-			description: "drop only",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeDrop},
-		}}
-
-	for _, protocol := range protocols {
-		for _, tt := range tests {
-			t.Run(tt.description+" "+protocol.String(), func(t *testing.T) {
-				store := NewAggregatingMemoryStore()
-				store.WindowEnd = time.Now().Add(5 * time.Second)
-
-				allExpected := make([]*types.Event, 0)
-
-				for _, srcAndDst := range tt.addresses {
-					inEvents, expected := generateEvents(srcAndDst[0], srcAndDst[1], tt.dstPort, tt.eventTypes, protocol, types.Ingress, 0, store.WindowStart, store.WindowEnd)
-					for _, e := range inEvents {
-						store.StoreEvent(e)
-					}
-					allExpected = append(allExpected, expected)
-				}
-
-				events := store.GetAggregatedEvents()
-				assert.ElementsMatch(t, events, allExpected)
-			})
-		}
-	}
-}
-
-func TestIcmpEventAggregation(t *testing.T) {
-	var protocols = []types.Protocol{types.ICMP, types.ICMPv6}
-	var icmpTypes = []uint8{1, 2, 3}
-
-	var tests = []struct {
-		description string
-		addresses   [][]netip.Addr
-		eventTypes  []types.Type
-	}{
-		{
-			description: "start and stop",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}},
-			eventTypes:  []types.Type{types.TypeStart, types.TypeEnd},
-		},
-		{
-			description: "start and drop",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}},
-			eventTypes:  []types.Type{types.TypeStart, types.TypeDrop},
-		},
-		{
-			description: "start only",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}},
-			eventTypes:  []types.Type{types.TypeStart},
-		},
-		{
-			description: "drop only",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}},
-			eventTypes:  []types.Type{types.TypeDrop},
-		}}
-
-	for _, protocol := range protocols {
-		for _, tt := range tests {
-			t.Run(tt.description+" "+protocol.String(), func(t *testing.T) {
-				store := NewAggregatingMemoryStore()
-				store.WindowEnd = time.Now().Add(5 * time.Second)
-
-				allExpected := make([]*types.Event, 0)
-				for _, icmpType := range icmpTypes {
-					events, expected := generateEvents(tt.addresses[0][0], tt.addresses[0][1], 0, tt.eventTypes, protocol, types.Ingress, icmpType, store.WindowStart, store.WindowEnd)
-					for _, e := range events {
-						store.StoreEvent(e)
-					}
-					allExpected = append(allExpected, expected)
-				}
-				aggregatedEvents := store.GetAggregatedEvents()
-				assert.Len(t, aggregatedEvents, len(allExpected))
-				assert.ElementsMatch(t, aggregatedEvents, allExpected)
-			})
-		}
-	}
-}
-
-func TestFlowAggregationOfUnknownProtocols(t *testing.T) {
-	var tests = []struct {
-		description string
-		addresses   [][]netip.Addr
-		dstPort     uint16
-		eventTypes  []types.Type
-	}{
-		{
-			description: "start and stop",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeStart, types.TypeEnd},
-		},
-		{
-			description: "start and drop",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeStart, types.TypeDrop},
-		},
-		{
-			description: "start only",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeStart},
-		},
-		{
-			description: "drop only",
-			addresses:   [][]netip.Addr{{netip.MustParseAddr("1.1.1.1"), netip.MustParseAddr("2.2.2.2")}, {netip.MustParseAddr("3.3.3.3"), netip.MustParseAddr("2.2.2.2")}},
-			dstPort:     uint16(random.Uint32() >> 16),
-			eventTypes:  []types.Type{types.TypeDrop},
-		}}
-
-	for _, tt := range tests {
-		t.Run(tt.description+" "+types.ProtocolUnknown.String(), func(t *testing.T) {
-			store := NewAggregatingMemoryStore()
-			store.WindowEnd = time.Now().Add(5 * time.Second)
-
-			allExpected := make([]*types.Event, 0)
-
-			for _, srcAndDst := range tt.addresses {
-				inEvents, expected := generateEventsForUnknownProtocol(srcAndDst[0], srcAndDst[1], tt.dstPort, tt.eventTypes, types.ProtocolUnknown, types.Ingress, store.WindowStart, store.WindowEnd)
-				for _, e := range inEvents {
-					store.StoreEvent(e)
-				}
-				allExpected = append(allExpected, expected...)
-			}
-
-			events := store.GetAggregatedEvents()
-			assert.ElementsMatch(t, events, allExpected)
-		})
-	}
-}
-
-func TestResetAggregationWindow(t *testing.T) {
-	store := NewAggregatingMemoryStore()
-	store.StoreEvent(&types.Event{
-		ID:        uuid.New(),
-		Timestamp: time.Now(),
-		EventFields: types.EventFields{
-			FlowID:           uuid.New(),
-			Type:             types.TypeStart,
-			Protocol:         types.TCP,
-			RuleID:           []byte("rule-id-1"),
-			Direction:        types.Ingress,
-			SourceIP:         netip.MustParseAddr("1.1.1.1"),
-			SourcePort:       1234,
-			DestIP:           netip.MustParseAddr("2.2.2.2"),
-			DestPort:         5678,
-			SourceResourceID: []byte("source-resource-id"),
-			DestResourceID:   []byte("dest-resource-id"),
-			RxPackets:        random.Uint64(),
-			TxPackets:        random.Uint64(),
-			RxBytes:          random.Uint64(),
-			TxBytes:          random.Uint64(),
-		},
-	})
-
-	reset := store.ResetAggregationWindow()
-	previousEvents, ok := reset.(*AggregatingMemory)
-	assert.True(t, ok)
-	assert.NotEqual(t, previousEvents.WindowStart, store.WindowStart)
-	assert.Equal(t, previousEvents.WindowEnd, store.WindowStart)
-	assert.NotEmpty(t, previousEvents.events)
-	assert.Empty(t, store.events)
-}
-
-func generateEvents(srcIp, dstIp netip.Addr, dstPort uint16, eventTypes []types.Type, protocol types.Protocol,
-	direction types.Direction, icmpType uint8, windowStart, windowEnd time.Time) ([]*types.Event, *types.Event) {
-	var rxPackets, txPackets, rxBytes, txBytes uint64
-	inEvents := make([]*types.Event, 0)
-	ts := time.Now()
-	flowId := uuid.New()
-	srcPort := uint16(random.Uint32() >> 16)
-
-	for idx, eventType := range eventTypes {
-		e := &types.Event{
-			ID:        uuid.New(),
-			Timestamp: ts.Add(time.Duration(idx) * time.Second),
-			EventFields: types.EventFields{
-				FlowID:           flowId,
-				Type:             eventType,
-				Protocol:         protocol,
-				RuleID:           []byte("rule-id-1"),
-				Direction:        direction,
-				SourceIP:         srcIp,
-				SourcePort:       srcPort,
-				DestIP:           dstIp,
-				DestPort:         dstPort,
-				SourceResourceID: []byte("source-resource-id"),
-				DestResourceID:   []byte("dest-resource-id"),
-				RxPackets:        random.Uint64(),
-				TxPackets:        random.Uint64(),
-				RxBytes:          random.Uint64(),
-				TxBytes:          random.Uint64(),
-			}}
-		rxBytes += e.RxBytes
-		txBytes += e.TxBytes
-		rxPackets += e.RxPackets
-		txPackets += e.TxPackets
-		inEvents = append(inEvents, e)
-		if protocol == types.ICMP || protocol == types.ICMPv6 {
-			e.ICMPType = icmpType
-		}
-	}
-
-	var start, end, drop uint64
-	for _, eventType := range eventTypes {
-		switch eventType {
-		case types.TypeStart:
-			start += 1
-		case types.TypeDrop:
-			drop += 1
-		case types.TypeEnd:
-			end += 1
-		}
-	}
-	aggregatedEvent := &types.Event{
-		ID:          inEvents[0].ID,
-		Timestamp:   inEvents[0].Timestamp,
-		WindowStart: windowStart,
-		WindowEnd:   windowEnd,
-		EventFields: types.EventFields{
-			FlowID:           flowId,
-			Type:             types.TypeUnknown,
-			Protocol:         inEvents[0].Protocol,
-			RuleID:           []byte("rule-id-1"),
-			Direction:        inEvents[0].Direction,
-			SourceIP:         srcIp,
-			SourcePort:       srcPort,
-			DestIP:           dstIp,
-			DestPort:         dstPort,
-			SourceResourceID: []byte("source-resource-id"),
-			DestResourceID:   []byte("dest-resource-id"),
-			RxPackets:        rxPackets,
-			TxPackets:        txPackets,
-			RxBytes:          rxBytes,
-			TxBytes:          txBytes,
-			NumOfStarts:      start,
-			NumOfEnds:        end,
-			NumOfDrops:       drop,
-		}}
-	if protocol == types.ICMP || protocol == types.ICMPv6 {
-		aggregatedEvent.ICMPType = icmpType
-	}
-
-	return inEvents, aggregatedEvent
-}
-
-func generateEventsForUnknownProtocol(srcIp, dstIp netip.Addr, dstPort uint16, eventTypes []types.Type, protocol types.Protocol,
-	direction types.Direction, windowStart, windowEnd time.Time) ([]*types.Event, []*types.Event) {
-	inEvents := make([]*types.Event, 0)
-	expectedEvents := make([]*types.Event, 0)
-
-	ts := time.Now()
-	flowId := uuid.New()
-	srcPort := uint16(random.Uint32() >> 16)
-
-	for idx, eventType := range eventTypes {
-		e := &types.Event{
-			ID:        uuid.New(),
-			Timestamp: ts.Add(time.Duration(idx) * time.Second),
-			EventFields: types.EventFields{
-				FlowID:           flowId,
-				Type:             eventType,
-				Protocol:         protocol,
-				RuleID:           []byte("rule-id-1"),
-				Direction:        direction,
-				SourceIP:         srcIp,
-				SourcePort:       srcPort,
-				DestIP:           dstIp,
-				DestPort:         dstPort,
-				SourceResourceID: []byte("source-resource-id"),
-				DestResourceID:   []byte("dest-resource-id"),
-				RxPackets:        random.Uint64(),
-				TxPackets:        random.Uint64(),
-				RxBytes:          random.Uint64(),
-				TxBytes:          random.Uint64(),
-			}}
-		inEvents = append(inEvents, e)
-
-		var start, end, drop uint64
-		switch eventType {
-		case types.TypeStart:
-			start = 1
-		case types.TypeDrop:
-			drop = 1
-		case types.TypeEnd:
-			end = 1
-		}
-
-		expectedEvents = append(expectedEvents, &types.Event{
-			ID:          e.ID,
-			Timestamp:   e.Timestamp,
-			WindowStart: windowStart,
-			WindowEnd:   windowEnd,
-			EventFields: types.EventFields{
-				FlowID:           flowId,
-				Type:             types.TypeUnknown,
-				Protocol:         e.Protocol,
-				RuleID:           []byte("rule-id-1"),
-				Direction:        e.Direction,
-				SourceIP:         srcIp,
-				SourcePort:       srcPort,
-				DestIP:           dstIp,
-				DestPort:         dstPort,
-				SourceResourceID: []byte("source-resource-id"),
-				DestResourceID:   []byte("dest-resource-id"),
-				RxPackets:        e.RxPackets,
-				TxPackets:        e.TxPackets,
-				RxBytes:          e.RxBytes,
-				TxBytes:          e.TxBytes,
-				NumOfStarts:      start,
-				NumOfEnds:        end,
-				NumOfDrops:       drop,
-			}})
-	}
-
-	return inEvents, expectedEvents
-}

client/internal/netflow/store/memory.go

@@ -1,15 +1,10 @@
 package store
 
 import (
-	"maps"
-	"math/rand"
-	v2 "math/rand/v2"
-	"net/netip"
-	"slices"
 	"sync"
-	"time"
 
 	"github.com/google/uuid"
+
 	"github.com/netbirdio/netbird/client/internal/netflow/types"
 )
 
@@ -24,13 +19,6 @@ type Memory struct {
 	events map[uuid.UUID]*types.Event
 }
 
-type AggregatingMemory struct {
-	Memory
-	WindowStart time.Time
-	WindowEnd   time.Time
-	rnd         *v2.PCG
-}
-
 func (m *Memory) StoreEvent(event *types.Event) {
 	m.mux.Lock()
 	defer m.mux.Unlock()
@@ -60,95 +48,3 @@ func (m *Memory) DeleteEvents(ids []uuid.UUID) {
 		delete(m.events, id)
 	}
 }
-
-func NewAggregatingMemoryStore() *AggregatingMemory {
-	return &AggregatingMemory{WindowStart: time.Now(), Memory: Memory{events: make(map[uuid.UUID]*types.Event)}, rnd: v2.NewPCG(rand.Uint64(), rand.Uint64())}
-}
-
-func (am *AggregatingMemory) ResetAggregationWindow() types.FlowEventAggregator {
-	am.mux.Lock()
-	defer am.mux.Unlock()
-
-	now := time.Now()
-	toret := AggregatingMemory{WindowStart: am.WindowStart, WindowEnd: now, Memory: Memory{events: am.events}, rnd: v2.NewPCG(rand.Uint64(), rand.Uint64())}
-
-	am.events = make(map[uuid.UUID]*types.Event)
-	am.WindowStart = now
-
-	return &toret
-}
-
-type aggregationKey struct {
-	srcAddr   netip.Addr
-	destAddr  netip.Addr
-	destPort  uint16
-	direction int
-	protocol  uint8
-	icmpType  uint8
-	unique    uint64 // used to prevent aggregation on non icmp/udp/tcp events
-}
-
-func (am *AggregatingMemory) GetAggregatedEvents() []*types.Event {
-	am.mux.Lock()
-	defer am.mux.Unlock()
-
-	aggregated := make(map[aggregationKey]*types.Event)
-	for _, v := range am.events {
-		lookupKey := aggregationKey{srcAddr: v.SourceIP, destAddr: v.DestIP, destPort: v.DestPort, direction: int(v.Direction), protocol: uint8(v.Protocol), icmpType: v.ICMPType}
-		if _, ok := aggregated[lookupKey]; !ok {
-			event := v.Clone()
-
-			switch event.Type {
-			case types.TypeStart:
-				event.NumOfStarts += 1
-			case types.TypeDrop:
-				event.NumOfDrops += 1
-			case types.TypeEnd:
-				event.NumOfEnds += 1
-			}
-			event.Type = types.TypeUnknown
-
-			// Please note that ICMPCode field isn't propagated by the manager (see flow/proto/flow.pb.go, FlowFields struct)
-			// so the field value in an icmp event in the "aggregated" doesn't matter
-
-			event.WindowStart = am.WindowStart
-			event.WindowEnd = am.WindowEnd
-
-			if event.Protocol != types.ICMP && event.Protocol != types.ICMPv6 && event.Protocol != types.UDP && event.Protocol != types.TCP {
-				lookupKey.unique = am.rnd.Uint64() // to make the lookup key unique so we don't aggregate on it
-			}
-
-			aggregated[lookupKey] = event
-			continue
-		}
-
-		aggregatedEvent := aggregated[lookupKey]
-		if aggregatedEvent.Protocol != types.ICMP && aggregatedEvent.Protocol != types.ICMPv6 && aggregatedEvent.Protocol != types.UDP && aggregatedEvent.Protocol != types.TCP {
-			continue // we don't aggregate this type of events; shouldn't ever get here
-		}
-
-		// track the number of connections, duration?, open and close events?
-		aggregatedEvent.RxBytes += v.RxBytes
-		aggregatedEvent.RxPackets += v.RxPackets
-		aggregatedEvent.TxBytes += v.TxBytes
-		aggregatedEvent.TxPackets += v.TxPackets
-		switch v.Type {
-		case types.TypeStart:
-			aggregatedEvent.NumOfStarts += 1
-		case types.TypeDrop:
-			aggregatedEvent.NumOfDrops += 1
-		case types.TypeEnd:
-			aggregatedEvent.NumOfEnds += 1
-		}
-		if aggregatedEvent.Timestamp.Compare(v.Timestamp) > 0 {
-			aggregatedEvent.Timestamp = v.Timestamp
-			aggregatedEvent.ID = v.ID
-			aggregatedEvent.SourcePort = v.SourcePort
-		}
-		if len(aggregatedEvent.RuleID) == 0 && len(v.RuleID) != 0 {
-			aggregatedEvent.RuleID = slices.Clone(v.RuleID)
-		}
-	}
-
-	return slices.Collect(maps.Values(aggregated)) // could return an iterator instead here
-}

client/internal/netflow/types/types.go

@@ -2,7 +2,6 @@ package types
 
 import (
 	"net/netip"
-	"slices"
 	"strconv"
 	"time"
 
@@ -70,10 +69,8 @@ const (
 )
 
 type Event struct {
-	ID          uuid.UUID
-	Timestamp   time.Time
-	WindowStart time.Time
-	WindowEnd   time.Time
+	ID        uuid.UUID
+	Timestamp time.Time
 	EventFields
 }
 
@@ -95,17 +92,6 @@ type EventFields struct {
 	TxPackets        uint64
 	RxBytes          uint64
 	TxBytes          uint64
-	NumOfStarts      uint64
-	NumOfEnds        uint64
-	NumOfDrops       uint64
-}
-
-func (e *Event) Clone() *Event {
-	toret := *e
-	toret.RuleID = slices.Clone(e.RuleID)
-	toret.SourceResourceID = slices.Clone(e.SourceResourceID)
-	toret.DestResourceID = slices.Clone(e.DestResourceID)
-	return &toret
 }
 
 type FlowConfig struct {
@@ -128,15 +114,13 @@ type FlowManager interface {
 	GetLogger() FlowLogger
 }
 
-type FlowEventAggregator interface {
-	ResetAggregationWindow() FlowEventAggregator
-	GetAggregatedEvents() []*Event
-}
-
 type FlowLogger interface {
-	ResetAggregationWindow() FlowEventAggregator
 	// StoreEvent stores a flow event
 	StoreEvent(flowEvent EventFields)
+	// GetEvents returns all stored events
+	GetEvents() []*Event
+	// DeleteEvents deletes events from the store
+	DeleteEvents([]uuid.UUID)
 	// Close closes the logger
 	Close()
 	// Enable enables the flow logger receiver
@@ -156,11 +140,6 @@ type Store interface {
 	Close()
 }
 
-type AggregatingStore interface {
-	FlowEventAggregator
-	Store
-}
-
 // ConnTracker defines the interface for connection tracking functionality
 type ConnTracker interface {
 	// Start begins tracking connections by listening for conntrack events.

client/internal/peer/conn.go

@@ -195,7 +195,6 @@ func NewConn(config ConnConfig, services ServiceDependencies) (*Conn, error) {
 		statusICE:          worker.NewAtomicStatus(),
 		dumpState:          dumpState,
 		endpointUpdater:    NewEndpointUpdater(connLog, config.WgConfig, isController(config)),
-		wgWatcher:          NewWGWatcher(connLog, config.WgConfig.WgInterface, config.Key, dumpState),
 		metricsRecorder:    services.MetricsRecorder,
 	}
 
@@ -663,11 +662,16 @@ func (conn *Conn) onGuardEvent() {
 	}
 }
 
-func (conn *Conn) onWGDisconnected() {
+// onWGDisconnected is invoked by the watcher goroutine when a handshake timeout is detected.
+// watcherCtx is the context of the watcher that fired: the timeout check runs lock-free, so by
+// the time we acquire conn.mu the watcher may have been cancelled (disabled) and a new connection
+// (and watcher) may already be in place. Re-checking watcherCtx under the lock prevents a stale
+// watcher from tearing down the connection that superseded it.
+func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
-	if conn.ctx.Err() != nil {
+	if conn.ctx.Err() != nil || watcherCtx.Err() != nil {
 		return
 	}
 
@@ -802,23 +806,44 @@ func (conn *Conn) isConnectedOnAllWay() (status guard.ConnStatus) {
 	})
 }
 
+// enableWgWatcherIfNeeded starts a fresh watcher for the current connection. A new WGWatcher
+// instance is created per attempt (rather than reusing one) so its lifecycle is bound entirely
+// to conn.mu: enable/disable can never race against an old watcher goroutine's shutdown, which
+// was the source of the "watcher silently fails to restart on a fast reconnect" bug. Caller must
+// hold conn.mu.
 func (conn *Conn) enableWgWatcherIfNeeded(enabledTime time.Time) {
-	if !conn.wgWatcher.IsEnabled() {
-		wgWatcherCtx, wgWatcherCancel := context.WithCancel(conn.ctx)
-		conn.wgWatcherCancel = wgWatcherCancel
-		conn.wgWatcherWg.Add(1)
-		go func() {
-			defer conn.wgWatcherWg.Done()
-			conn.wgWatcher.EnableWgWatcher(wgWatcherCtx, enabledTime, conn.onWGDisconnected, conn.onWGHandshakeSuccess)
-		}()
+	if conn.wgWatcher != nil {
+		// a watcher is already running for the current connection
+		return
 	}
+
+	watcher := NewWGWatcher(conn.Log, conn.config.WgConfig.WgInterface, conn.config.Key, conn.dumpState)
+	watcher.PrepareInitialHandshake()
+
+	wgWatcherCtx, wgWatcherCancel := context.WithCancel(conn.ctx)
+	conn.wgWatcher = watcher
+	conn.wgWatcherCancel = wgWatcherCancel
+
+	conn.wgWatcherWg.Add(1)
+	go func() {
+		defer conn.wgWatcherWg.Done()
+		onDisconnected := func() { conn.onWGDisconnected(wgWatcherCtx) }
+		watcher.EnableWgWatcher(wgWatcherCtx, enabledTime, onDisconnected, conn.onWGHandshakeSuccess)
+	}()
 }
 
+// disableWgWatcherIfNeeded stops and drops the current watcher once no transport is active. It
+// only signals the watcher goroutine (cancel) and clears the reference; it never waits for the
+// goroutine to exit, because the watcher's own timeout path reentrantly calls back here under
+// conn.mu (via onWGDisconnected), so blocking would deadlock. The cancelled goroutine drains
+// harmlessly. Caller must hold conn.mu.
 func (conn *Conn) disableWgWatcherIfNeeded() {
-	if conn.currentConnPriority == conntype.None && conn.wgWatcherCancel != nil {
-		conn.wgWatcherCancel()
-		conn.wgWatcherCancel = nil
+	if conn.currentConnPriority != conntype.None || conn.wgWatcher == nil {
+		return
 	}
+	conn.wgWatcherCancel()
+	conn.wgWatcher = nil
+	conn.wgWatcherCancel = nil
 }
 
 func (conn *Conn) newProxy(remoteConn net.Conn) (wgproxy.Proxy, error) {
@@ -841,7 +866,9 @@ func (conn *Conn) resetEndpoint() {
 		return
 	}
 	conn.Log.Infof("reset wg endpoint")
-	conn.wgWatcher.Reset()
+	if conn.wgWatcher != nil {
+		conn.wgWatcher.Reset()
+	}
 	if err := conn.endpointUpdater.RemoveEndpointAddress(); err != nil {
 		conn.Log.Warnf("failed to remove endpoint address before update: %v", err)
 	}

client/internal/peer/wg_watcher.go

@@ -3,7 +3,6 @@ package peer
 import (
 	"context"
 	"fmt"
-	"sync"
 	"time"
 
 	log "github.com/sirupsen/logrus"
@@ -24,14 +23,17 @@ type WGInterfaceStater interface {
 	GetStats() (map[string]configurer.WGStats, error)
 }
 
+// WGWatcher is single-shot: create one instance per connection attempt, run it once via
+// EnableWgWatcher, then discard it. Lifecycle (start/stop) is owned by Conn under conn.mu,
+// so the watcher itself keeps no "enabled" state to go stale on a fast disconnect/reconnect.
 type WGWatcher struct {
 	log           *log.Entry
 	wgIfaceStater WGInterfaceStater
 	peerKey       string
 	stateDump     *stateDump
 
-	enabled   bool
-	muEnabled sync.RWMutex
+	// initialHandshake is not thread-safe; never call PrepareInitialHandshake and EnableWgWatcher concurrently.
+	initialHandshake time.Time
 
 	resetCh chan struct{}
 }
@@ -46,36 +48,21 @@ func NewWGWatcher(log *log.Entry, wgIfaceStater WGInterfaceStater, peerKey strin
 	}
 }
 
-// EnableWgWatcher starts the WireGuard watcher. If it is already enabled, it will return immediately and do nothing.
-// The watcher runs until ctx is cancelled. Caller is responsible for context lifecycle management.
-func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time)) {
-	w.muEnabled.Lock()
-	if w.enabled {
-		w.muEnabled.Unlock()
-		return
-	}
-
+// PrepareInitialHandshake reads the peer's current WireGuard handshake time. It must be
+// called before the peer is (re)configured on the WireGuard interface, so the captured
+// baseline reflects the state prior to this connection attempt instead of racing with
+// that configuration.
+func (w *WGWatcher) PrepareInitialHandshake() {
 	w.log.Debugf("enable WireGuard watcher")
-	w.enabled = true
-	w.muEnabled.Unlock()
-
-	initialHandshake, err := w.wgState()
-	if err != nil {
-		w.log.Warnf("failed to read initial wg stats: %v", err)
-	}
-
-	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, enabledTime, initialHandshake)
-
-	w.muEnabled.Lock()
-	w.enabled = false
-	w.muEnabled.Unlock()
+	handshake, _ := w.wgState()
+	w.initialHandshake = handshake
 }
 
-// IsEnabled returns true if the WireGuard watcher is currently enabled
-func (w *WGWatcher) IsEnabled() bool {
-	w.muEnabled.RLock()
-	defer w.muEnabled.RUnlock()
-	return w.enabled
+// EnableWgWatcher runs the WireGuard watcher loop using the handshake baseline captured by
+// PrepareInitialHandshake. The watcher runs until ctx is cancelled. Caller is responsible
+// for context lifecycle management.
+func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time)) {
+	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, enabledTime, w.initialHandshake)
 }
 
 // Reset signals the watcher that the WireGuard peer has been reset and a new
@@ -101,13 +88,17 @@ func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn
 		case <-timer.C:
 			handshake, ok := w.handshakeCheck(lastHandshake)
 			if !ok {
+				// early ctx cancel check return
+				if ctx.Err() != nil {
+					return
+				}
 				onDisconnectedFn()
 				return
 			}
 			if lastHandshake.IsZero() {
 				elapsed := calcElapsed(enabledTime, *handshake)
 				w.log.Infof("first wg handshake detected within: %.2fsec, (%s)", elapsed, handshake)
-				if onHandshakeSuccessFn != nil {
+				if onHandshakeSuccessFn != nil && ctx.Err() == nil {
 					onHandshakeSuccessFn(*handshake)
 				}
 			}
@@ -142,9 +133,9 @@ func (w *WGWatcher) handshakeCheck(lastHandshake time.Time) (*time.Time, bool) {
 
 	w.log.Tracef("previous handshake, handshake: %v, %v", lastHandshake, handshake)
 
-	// the current know handshake did not change
+	// the current known handshake did not change
 	if handshake.Equal(lastHandshake) {
-		w.log.Warnf("WireGuard handshake timed out: %v", handshake)
+		w.log.Warnf("WireGuard handshake not updated: %v", handshake)
 		return nil, false
 	}
 

client/internal/peer/wg_watcher_test.go

@@ -34,6 +34,8 @@ func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
+	watcher.PrepareInitialHandshake()
+
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
 		mlog.Infof("onDisconnectedFn")
@@ -62,6 +64,8 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	watcher := NewWGWatcher(mlog, mocWgIface, "", newStateDump("peer", mlog, &Status{}))
 
 	ctx, cancel := context.WithCancel(context.Background())
+	watcher.PrepareInitialHandshake()
+
 	wg := &sync.WaitGroup{}
 	wg.Add(1)
 	go func() {
@@ -76,6 +80,8 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	ctx, cancel = context.WithCancel(context.Background())
 	defer cancel()
 
+	watcher.PrepareInitialHandshake()
+
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
 		onDisconnected <- struct{}{}

client/internal/profilemanager/config.go

@@ -101,8 +101,6 @@ type ConfigInput struct {
 
 	DNSLabels domain.List
 
-	LazyConnectionEnabled *bool
-
 	MTU *uint16
 }
 
@@ -180,7 +178,9 @@ type Config struct {
 
 	ClientCertKeyPair *tls.Certificate `json:"-"`
 
-	LazyConnectionEnabled bool
+	// LazyConnection is the MDM-managed lazy-connection override ("on"/"off"/"").
+	// Runtime-only: re-derived from MDM policy on each load, never persisted.
+	LazyConnection string `json:"-"`
 
 	MTU uint16
 
@@ -632,12 +632,6 @@ func (config *Config) apply(input ConfigInput) (updated bool, err error) {
 		updated = true
 	}
 
-	if input.LazyConnectionEnabled != nil && *input.LazyConnectionEnabled != config.LazyConnectionEnabled {
-		log.Infof("switching lazy connection to %t", *input.LazyConnectionEnabled)
-		config.LazyConnectionEnabled = *input.LazyConnectionEnabled
-		updated = true
-	}
-
 	if input.MTU != nil && *input.MTU != config.MTU {
 		log.Infof("updating MTU to %d (old value %d)", *input.MTU, config.MTU)
 		config.MTU = *input.MTU
@@ -728,6 +722,15 @@ func (config *Config) applyMDMPolicy(policy *mdm.Policy) {
 			log.Warnf("MDM wireguard port %d out of range [1,65535]; keeping previous value", v)
 		}
 	}
+
+	if v, ok := policy.GetBool(mdm.KeyLazyConnection); ok {
+		state := "off"
+		if v {
+			state = "on"
+		}
+		config.LazyConnection = state
+		logApplied(mdm.KeyLazyConnection, state)
+	}
 }
 
 // parseURL parses and validates the URL for the named service. The URL

client/internal/profilemanager/config_mdm_test.go

@@ -130,6 +130,37 @@ func TestApply_MDMBoolKeysOverrideOnDiskValue(t *testing.T) {
 	assert.True(t, cfg.Policy().HasKey(mdm.KeyRosenpassEnabled))
 }
 
+func TestApply_MDMLazyConnection(t *testing.T) {
+	cases := []struct {
+		name string
+		raw  any
+		want string
+	}{
+		{"native true", true, "on"},
+		{"native false", false, "off"},
+		{"string on", "on", "on"},
+		{"string off", "off", "off"},
+		{"string yes", "yes", "on"},
+		{"string no", "no", "off"},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			withMDMPolicy(t, mdm.NewPolicy(map[string]any{
+				mdm.KeyLazyConnection: c.raw,
+			}))
+
+			cfg, err := UpdateOrCreateConfig(ConfigInput{
+				ConfigPath: filepath.Join(t.TempDir(), "config.json"),
+			})
+			require.NoError(t, err)
+			require.NotNil(t, cfg)
+
+			assert.Equal(t, c.want, cfg.LazyConnection)
+			assert.True(t, cfg.Policy().HasKey(mdm.KeyLazyConnection))
+		})
+	}
+}
+
 func TestApply_MDMPreSharedKeyRedactionSentinelRejected(t *testing.T) {
 	const maskSentinel = "**********"
 

client/internal/routemanager/exit_node_selection_test.go

@@ -0,0 +1,191 @@
+package routemanager
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/routeselector"
+	"github.com/netbirdio/netbird/route"
+)
+
+func newExitNodeTestManager() *DefaultManager {
+	return &DefaultManager{routeSelector: routeselector.NewRouteSelector()}
+}
+
+func exitRoute(netID, peer string, skipAutoApply bool) *route.Route {
+	return &route.Route{
+		NetID:         route.NetID(netID),
+		Network:       netip.MustParsePrefix("0.0.0.0/0"),
+		Peer:          peer,
+		SkipAutoApply: skipAutoApply,
+	}
+}
+
+func TestPickPreferredExitNode(t *testing.T) {
+	tests := []struct {
+		name string
+		info exitNodeInfo
+		want route.NetID
+	}{
+		{
+			name: "persisted user selection wins over management",
+			info: exitNodeInfo{
+				allIDs:               []route.NetID{"a", "b", "c"},
+				userSelected:         []route.NetID{"b"},
+				selectedByManagement: []route.NetID{"a"},
+			},
+			want: "b",
+		},
+		{
+			name: "multiple user-selected self-heal to deterministic min",
+			info: exitNodeInfo{
+				allIDs:       []route.NetID{"a", "b", "c"},
+				userSelected: []route.NetID{"c", "a"},
+			},
+			want: "a",
+		},
+		{
+			name: "explicit opt-out keeps none",
+			info: exitNodeInfo{
+				allIDs:         []route.NetID{"a", "b"},
+				userDeselected: []route.NetID{"a", "b"},
+			},
+			want: "",
+		},
+		{
+			name: "fresh defaults to management auto-apply pick",
+			info: exitNodeInfo{
+				allIDs:               []route.NetID{"a", "b", "c"},
+				selectedByManagement: []route.NetID{"b"},
+			},
+			want: "b",
+		},
+		{
+			name: "no user pick and no management auto-apply selects none",
+			info: exitNodeInfo{
+				allIDs: []route.NetID{"c", "a", "b"},
+			},
+			want: "",
+		},
+		{
+			name: "user-deselect does not block a management auto-apply sibling",
+			info: exitNodeInfo{
+				allIDs:               []route.NetID{"a", "b"},
+				userDeselected:       []route.NetID{"a"},
+				selectedByManagement: []route.NetID{"b"},
+			},
+			want: "b",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, pickPreferredExitNode(tt.info), "preferred exit node")
+		})
+	}
+}
+
+func TestEnforceSingleExitNode(t *testing.T) {
+	m := newExitNodeTestManager()
+	all := []route.NetID{"a", "b", "c"}
+
+	m.enforceSingleExitNode("b", all)
+	assert.False(t, m.routeSelector.IsSelected("a"), "a should be deselected")
+	assert.True(t, m.routeSelector.IsSelected("b"), "b should be the only selected exit node")
+	assert.False(t, m.routeSelector.IsSelected("c"), "c should be deselected")
+
+	// Switching the preferred node moves the single selection.
+	m.enforceSingleExitNode("c", all)
+	assert.False(t, m.routeSelector.IsSelected("a"), "a stays deselected")
+	assert.False(t, m.routeSelector.IsSelected("b"), "b should now be deselected")
+	assert.True(t, m.routeSelector.IsSelected("c"), "c should now be selected")
+
+	// Empty preferred turns every exit node off.
+	m.enforceSingleExitNode("", all)
+	for _, id := range all {
+		assert.False(t, m.routeSelector.IsSelected(id), "no exit node should be selected")
+	}
+}
+
+func TestEnforceSingleExitNode_RespectsDeselectAll(t *testing.T) {
+	m := newExitNodeTestManager()
+	m.routeSelector.DeselectAllRoutes()
+
+	m.enforceSingleExitNode("b", []route.NetID{"a", "b"})
+
+	assert.True(t, m.routeSelector.IsDeselectAll(), "global deselect-all must stay in effect")
+	assert.False(t, m.routeSelector.IsSelected("b"), "no exit node should be forced on while deselect-all is set")
+}
+
+func TestUpdateRouteSelectorFromManagement_FreshSelectsOne(t *testing.T) {
+	m := newExitNodeTestManager()
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0":    {exitRoute("exitA", "p1", false)},
+		"exitB|0.0.0.0/0":    {exitRoute("exitB", "p2", false)},
+		"lan|192.168.1.0/24": {{NetID: "lan", Network: netip.MustParsePrefix("192.168.1.0/24"), Peer: "p3"}},
+		"exitC|0.0.0.0/0":    {exitRoute("exitC", "p4", false)},
+	}
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	// Exactly one exit node (the deterministic first) is selected.
+	assert.True(t, m.routeSelector.IsSelected("exitA"), "exitA is the deterministic default")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "exitB must not also be selected")
+	assert.False(t, m.routeSelector.IsSelected("exitC"), "exitC must not also be selected")
+	// Non-exit routes are left at their default-on state.
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit route selection is untouched")
+}
+
+func TestUpdateRouteSelectorFromManagement_HonorsPersistedPick(t *testing.T) {
+	m := newExitNodeTestManager()
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0": {exitRoute("exitA", "p1", false)},
+		"exitB|0.0.0.0/0": {exitRoute("exitB", "p2", false)},
+	}
+	all := []route.NetID{"exitA", "exitB"}
+
+	// Simulate the state the runtime select path leaves behind: exactly one
+	// exit node explicitly selected, its sibling deselected.
+	require.NoError(t, m.routeSelector.SelectRoutes([]route.NetID{"exitB"}, true, all))
+	require.NoError(t, m.routeSelector.DeselectRoutes([]route.NetID{"exitA"}, all))
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	assert.True(t, m.routeSelector.IsSelected("exitB"), "persisted pick must stay selected")
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "the other exit node stays deselected")
+}
+
+func TestUpdateRouteSelectorFromManagement_OptOutKeepsNone(t *testing.T) {
+	m := newExitNodeTestManager()
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0": {exitRoute("exitA", "p1", false)},
+		"exitB|0.0.0.0/0": {exitRoute("exitB", "p2", false)},
+	}
+	all := []route.NetID{"exitA", "exitB"}
+
+	// User deselected exit nodes and selected none.
+	require.NoError(t, m.routeSelector.DeselectRoutes(all, all))
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "opt-out keeps exitA off")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "opt-out keeps exitB off")
+}
+
+func TestUpdateRouteSelectorFromManagement_NoAutoApplySelectsNone(t *testing.T) {
+	m := newExitNodeTestManager()
+	// SkipAutoApply=true: management offers the exit nodes but doesn't request
+	// auto-activation, so none should be selected until the user picks one.
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0": {exitRoute("exitA", "p1", true)},
+		"exitB|0.0.0.0/0": {exitRoute("exitB", "p2", true)},
+	}
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "no auto-apply keeps exitA off")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "no auto-apply keeps exitB off")
+}

client/internal/routemanager/manager.go

@@ -701,7 +701,13 @@ func resolveURLsToIPs(urls []string) []net.IP {
 	return ips
 }
 
-// updateRouteSelectorFromManagement updates the route selector based on the isSelected status from the management server
+// updateRouteSelectorFromManagement reconciles exit-node selection on every
+// network map: it keeps at most one exit node selected — the user's persisted
+// pick, else whatever management marks for auto-apply (SkipAutoApply=false),
+// else none. We never auto-activate an exit node the map doesn't request; it
+// stays off until the user picks it. Exit nodes are mutually exclusive, but the
+// RouteSelector stores routes with default-on semantics, so without this every
+// available exit node would report selected at once.
 func (m *DefaultManager) updateRouteSelectorFromManagement(clientRoutes route.HAMap) {
 	m.mirrorV6ExitPairSelections(clientRoutes)
 
@@ -712,13 +718,14 @@ func (m *DefaultManager) updateRouteSelectorFromManagement(clientRoutes route.HA
 		return
 	}
 
-	exitNodeInfo := m.collectExitNodeInfo(clientRoutes)
-	if len(exitNodeInfo.allIDs) == 0 {
+	info := m.collectExitNodeInfo(clientRoutes)
+	if len(info.allIDs) == 0 {
 		return
 	}
 
-	m.updateExitNodeSelections(exitNodeInfo)
-	m.logExitNodeUpdate(exitNodeInfo)
+	preferred := pickPreferredExitNode(info)
+	m.enforceSingleExitNode(preferred, info.allIDs)
+	m.logExitNodeUpdate(info, preferred)
 }
 
 // mirrorV6ExitPairSelections keeps every synthesized "-v6" exit route's selection
@@ -746,6 +753,10 @@ type exitNodeInfo struct {
 	userDeselected       []route.NetID
 }
 
+// collectExitNodeInfo categorises the available exit nodes by their persisted
+// selection state. It keys on the base (v4) NetID and skips the synthesized
+// "-v6" partner, which inherits its base's selection through the RouteSelector
+// — counting it separately would double-count the pair.
 func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeInfo {
 	var info exitNodeInfo
 
@@ -755,6 +766,9 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 		}
 
 		netID := haID.NetID()
+		if strings.HasSuffix(string(netID), route.V6ExitSuffix) {
+			continue
+		}
 		info.allIDs = append(info.allIDs, netID)
 
 		if m.routeSelector.HasUserSelectionForRoute(netID) {
@@ -791,45 +805,52 @@ func (m *DefaultManager) checkManagementSelection(routes []*route.Route, netID r
 	}
 }
 
-func (m *DefaultManager) updateExitNodeSelections(info exitNodeInfo) {
-	routesToDeselect := m.getRoutesToDeselect(info.allIDs)
-	m.deselectExitNodes(routesToDeselect)
-	m.selectExitNodesByManagement(info.selectedByManagement, info.allIDs)
+// pickPreferredExitNode chooses the single exit node to keep selected. In order:
+//   - a persisted user selection wins (deterministic if several survive from
+//     legacy state, so the set self-heals down to one);
+//   - otherwise activate only what management marks for auto-apply
+//     (SkipAutoApply=false); the lexicographically first if it marks several.
+//
+// Returns "" when neither holds — we never force an arbitrary exit node on. A
+// route the map doesn't auto-apply stays off until the user selects it.
+// info.userDeselected is informational only: an explicit deselect simply keeps
+// that route out of both lists above, so it can't be picked.
+func pickPreferredExitNode(info exitNodeInfo) route.NetID {
+	if len(info.userSelected) > 0 {
+		return minNetID(info.userSelected)
+	}
+	if len(info.selectedByManagement) > 0 {
+		return minNetID(info.selectedByManagement)
+	}
+	return ""
 }
 
-func (m *DefaultManager) getRoutesToDeselect(allIDs []route.NetID) []route.NetID {
-	var routesToDeselect []route.NetID
-	for _, netID := range allIDs {
-		if !m.routeSelector.HasUserSelectionForRoute(netID) {
-			routesToDeselect = append(routesToDeselect, netID)
+// enforceSingleExitNode makes preferred the only selected exit node: every other
+// available exit node is deselected and preferred (if any) is selected, without
+// disturbing non-exit route selections. The whole reconciliation runs under a
+// single RouteSelector lock (SetExclusiveExitNode) so a concurrent deselect-all
+// cannot interleave and get undone; a global deselect-all is left untouched so
+// the user's "all off" stays in effect.
+func (m *DefaultManager) enforceSingleExitNode(preferred route.NetID, allIDs []route.NetID) {
+	m.routeSelector.SetExclusiveExitNode(preferred, allIDs)
+}
+
+func (m *DefaultManager) logExitNodeUpdate(info exitNodeInfo, preferred route.NetID) {
+	log.Debugf("Exit node selection: %d available, preferred=%q (%d user-selected, %d user-deselected, %d management-selected)",
+		len(info.allIDs), preferred, len(info.userSelected), len(info.userDeselected), len(info.selectedByManagement))
+}
+
+// minNetID returns the lexicographically smallest NetID, for a deterministic
+// default pick that stays stable across restarts.
+func minNetID(ids []route.NetID) route.NetID {
+	if len(ids) == 0 {
+		return ""
+	}
+	best := ids[0]
+	for _, id := range ids[1:] {
+		if id < best {
+			best = id
 		}
 	}
-	return routesToDeselect
-}
-
-func (m *DefaultManager) deselectExitNodes(routesToDeselect []route.NetID) {
-	if len(routesToDeselect) == 0 {
-		return
-	}
-
-	err := m.routeSelector.DeselectRoutes(routesToDeselect, routesToDeselect)
-	if err != nil {
-		log.Warnf("Failed to deselect exit nodes: %v", err)
-	}
-}
-
-func (m *DefaultManager) selectExitNodesByManagement(selectedByManagement []route.NetID, allIDs []route.NetID) {
-	if len(selectedByManagement) == 0 {
-		return
-	}
-
-	err := m.routeSelector.SelectRoutes(selectedByManagement, true, allIDs)
-	if err != nil {
-		log.Warnf("Failed to select exit nodes: %v", err)
-	}
-}
-
-func (m *DefaultManager) logExitNodeUpdate(info exitNodeInfo) {
-	log.Debugf("Updated route selector: %d exit nodes available, %d selected by management, %d user-selected, %d user-deselected",
-		len(info.allIDs), len(info.selectedByManagement), len(info.userSelected), len(info.userDeselected))
+	return best
 }

client/internal/routeselector/routeselector.go

@@ -115,7 +115,38 @@ func (rs *RouteSelector) DeselectAllRoutes() {
 	clear(rs.selectedRoutes)
 }
 
-// IsDeselectAll reports whether the user has explicitly deselected all routes.
+// SetExclusiveExitNode atomically makes preferred the only selected exit node
+// among exitIDs: every other ID in exitIDs is deselected and preferred (when
+// non-empty) is selected, all under a single lock. Holding the lock across the
+// whole reconciliation prevents a concurrent DeselectAllRoutes from interleaving
+// between the deselect and select steps and being silently undone. A global
+// deselect-all is left untouched so the user's "all off" stays in effect;
+// non-exit routes are never referenced, so their selection is preserved.
+func (rs *RouteSelector) SetExclusiveExitNode(preferred route.NetID, exitIDs []route.NetID) {
+	rs.mu.Lock()
+	defer rs.mu.Unlock()
+
+	if rs.deselectAll {
+		return
+	}
+
+	for _, id := range exitIDs {
+		if id == preferred {
+			continue
+		}
+		rs.deselectedRoutes[id] = struct{}{}
+		delete(rs.selectedRoutes, id)
+	}
+
+	if preferred != "" {
+		delete(rs.deselectedRoutes, preferred)
+		rs.selectedRoutes[preferred] = struct{}{}
+	}
+}
+
+// IsDeselectAll reports whether the global "deselect all" flag is set, i.e. the
+// user explicitly disabled every route. Callers enforcing per-route invariants
+// (e.g. single exit node) should leave the selection untouched when it is.
 func (rs *RouteSelector) IsDeselectAll() bool {
 	rs.mu.RLock()
 	defer rs.mu.RUnlock()

client/ios/NetBirdSDK/env_list.go

@@ -38,7 +38,7 @@ func GetEnvKeyNBForceRelay() string {
 
 // GetEnvKeyNBLazyConn Exports the environment variable for the iOS client
 func GetEnvKeyNBLazyConn() string {
-	return lazyconn.EnvEnableLazyConn
+	return lazyconn.EnvLazyConn
 }
 
 // GetEnvKeyNBInactivityThreshold Exports the environment variable for the iOS client

client/mdm/canonical_loaders.go

@@ -27,6 +27,7 @@ var allKeys = []string{
 	KeyWireguardPort,
 	KeySplitTunnelMode,
 	KeySplitTunnelApps,
+	KeyLazyConnection,
 }
 
 // canonicalKey maps the lowercase form of a managed-config value name to

client/mdm/policy.go

@@ -11,6 +11,7 @@ package mdm
 import (
 	"sort"
 	"strconv"
+	"strings"
 
 	log "github.com/sirupsen/logrus"
 )
@@ -41,6 +42,11 @@ const (
 	// construction — only one mode can be set at a time.
 	KeySplitTunnelMode = "splitTunnelMode"
 	KeySplitTunnelApps = "splitTunnelApps"
+
+	// KeyLazyConnection forces the lazy-connection feature on or off, overriding
+	// the management feature flag. Read as a bool (native bool, or on/off,
+	// true/false, 1/0, yes/no); absent = defer to management.
+	KeyLazyConnection = "lazyConnection"
 )
 
 // Split-tunnel mode literals (KeySplitTunnelMode values).
@@ -62,12 +68,13 @@ var boolStringLiterals = map[string]bool{
 	"true":  true,
 	"1":     true,
 	"yes":   true,
+	"on":    true,
 	"false": false,
 	"0":     false,
 	"no":    false,
+	"off":   false,
 }
 
-
 // Policy holds MDM-managed settings read from the platform source. A nil or
 // empty Policy means no enforcement is active.
 type Policy struct {
@@ -150,7 +157,8 @@ func (p *Policy) GetString(key string) (string, bool) {
 }
 
 // GetBool returns the managed value for key coerced to bool, and whether the
-// key was set. Accepts native bool and string literals "true"/"false"/"1"/"0".
+// key was set. Accepts native bool and string literals (true/false, 1/0,
+// yes/no, on/off), case-insensitively and trimmed of surrounding whitespace.
 func (p *Policy) GetBool(key string) (bool, bool) {
 	if p == nil {
 		return false, false
@@ -163,7 +171,7 @@ func (p *Policy) GetBool(key string) (bool, bool) {
 	case bool:
 		return t, true
 	case string:
-		b, known := boolStringLiterals[t]
+		b, known := boolStringLiterals[strings.ToLower(strings.TrimSpace(t))]
 		return b, known
 	case int:
 		return t != 0, true

client/mdm/policy_test.go

@@ -31,8 +31,8 @@ func TestPolicy_Empty(t *testing.T) {
 
 func TestPolicy_HasKey(t *testing.T) {
 	p := NewPolicy(map[string]any{
-		KeyManagementURL:    "https://corp.example.com",
-		KeyDisableProfiles:  true,
+		KeyManagementURL:   "https://corp.example.com",
+		KeyDisableProfiles: true,
 	})
 	assert.False(t, p.IsEmpty())
 	assert.True(t, p.HasKey(KeyManagementURL))
@@ -53,8 +53,8 @@ func TestPolicy_ManagedKeysSorted(t *testing.T) {
 func TestPolicy_GetString(t *testing.T) {
 	p := NewPolicy(map[string]any{
 		KeyManagementURL:   "https://corp.example.com",
-		KeyDisableProfiles: true,            // wrong type for GetString
-		KeyPreSharedKey:        "",              // empty rejected
+		KeyDisableProfiles: true, // wrong type for GetString
+		KeyPreSharedKey:    "",   // empty rejected
 	})
 	v, ok := p.GetString(KeyManagementURL)
 	assert.True(t, ok)
@@ -85,6 +85,11 @@ func TestPolicy_GetBool(t *testing.T) {
 		{"string 0", "0", false, true},
 		{"string yes", "yes", true, true},
 		{"string no", "no", false, true},
+		{"string on", "on", true, true},
+		{"string off", "off", false, true},
+		{"mixed case On", "On", true, true},
+		{"upper TRUE", "TRUE", true, true},
+		{"padded yes", "  yes  ", true, true},
 		{"int nonzero", 1, true, true},
 		{"int zero", 0, false, true},
 		{"int64 nonzero", int64(2), true, true},

client/server/mdm.go

@@ -152,7 +152,6 @@ func (s *Server) restartEngineForMDMLocked() error {
 	s.config = config
 	s.statusRecorder.UpdateManagementAddress(config.ManagementURL.String())
 	s.statusRecorder.UpdateRosenpass(config.RosenpassEnabled, config.RosenpassPermissive)
-	s.statusRecorder.UpdateLazyConnection(config.LazyConnectionEnabled)
 
 	ctx, cancel := context.WithCancel(s.rootCtx)
 	s.actCancel = cancel
@@ -305,7 +304,6 @@ func setConfigRequestHasConfigOverrides(msg *proto.SetConfigRequest) bool {
 		msg.DisableFirewall != nil ||
 		msg.BlockLanAccess != nil ||
 		msg.DisableNotifications != nil ||
-		msg.LazyConnectionEnabled != nil ||
 		msg.BlockInbound != nil ||
 		msg.DisableIpv6 != nil ||
 		msg.EnableSSHRoot != nil ||
@@ -348,7 +346,6 @@ func loginRequestHasConfigOverrides(msg *proto.LoginRequest) bool {
 		msg.BlockLanAccess != nil ||
 		msg.DisableNotifications != nil ||
 		len(msg.DnsLabels) > 0 || msg.CleanDNSLabels ||
-		msg.LazyConnectionEnabled != nil ||
 		msg.BlockInbound != nil
 }
 

client/server/server.go

@@ -214,7 +214,6 @@ func (s *Server) Start() error {
 
 	s.statusRecorder.UpdateManagementAddress(config.ManagementURL.String())
 	s.statusRecorder.UpdateRosenpass(config.RosenpassEnabled, config.RosenpassPermissive)
-	s.statusRecorder.UpdateLazyConnection(config.LazyConnectionEnabled)
 
 	if s.sessionWatcher == nil {
 		s.sessionWatcher = internal.NewSessionWatcher(s.rootCtx, s.statusRecorder)
@@ -463,7 +462,6 @@ func (s *Server) setConfigInputFromRequest(msg *proto.SetConfigRequest) (profile
 	config.DisableFirewall = msg.DisableFirewall
 	config.BlockLANAccess = msg.BlockLanAccess
 	config.DisableNotifications = msg.DisableNotifications
-	config.LazyConnectionEnabled = msg.LazyConnectionEnabled
 	config.BlockInbound = msg.BlockInbound
 	config.DisableIPv6 = msg.DisableIpv6
 	config.EnableSSHRoot = msg.EnableSSHRoot
@@ -1647,7 +1645,6 @@ func (s *Server) GetConfig(ctx context.Context, req *proto.GetConfigRequest) (*p
 		ServerSSHAllowed:              *cfg.ServerSSHAllowed,
 		RosenpassEnabled:              cfg.RosenpassEnabled,
 		RosenpassPermissive:           cfg.RosenpassPermissive,
-		LazyConnectionEnabled:         cfg.LazyConnectionEnabled,
 		BlockInbound:                  cfg.BlockInbound,
 		DisableNotifications:          disableNotifications,
 		NetworkMonitor:                networkMonitor,

client/server/setconfig_test.go

@@ -69,43 +69,41 @@ func TestSetConfig_AllFieldsSaved(t *testing.T) {
 	disableFirewall := true
 	blockLANAccess := true
 	disableNotifications := true
-	lazyConnectionEnabled := true
 	blockInbound := true
 	disableIPv6 := true
 	mtu := int64(1280)
 	sshJWTCacheTTL := int32(300)
 
 	req := &proto.SetConfigRequest{
-		ProfileName:           profName,
-		Username:              currUser.Username,
-		ManagementUrl:         "https://new-api.netbird.io:443",
-		AdminURL:              "https://new-admin.netbird.io",
-		RosenpassEnabled:      &rosenpassEnabled,
-		RosenpassPermissive:   &rosenpassPermissive,
-		ServerSSHAllowed:      &serverSSHAllowed,
-		InterfaceName:         &interfaceName,
-		WireguardPort:         &wireguardPort,
-		OptionalPreSharedKey:  &preSharedKey,
-		DisableAutoConnect:    &disableAutoConnect,
-		NetworkMonitor:        &networkMonitor,
-		DisableClientRoutes:   &disableClientRoutes,
-		DisableServerRoutes:   &disableServerRoutes,
-		DisableDns:            &disableDNS,
-		DisableFirewall:       &disableFirewall,
-		BlockLanAccess:        &blockLANAccess,
-		DisableNotifications:  &disableNotifications,
-		LazyConnectionEnabled: &lazyConnectionEnabled,
-		BlockInbound:          &blockInbound,
-		DisableIpv6:           &disableIPv6,
-		NatExternalIPs:        []string{"1.2.3.4", "5.6.7.8"},
-		CleanNATExternalIPs:   false,
-		CustomDNSAddress:      []byte("1.1.1.1:53"),
-		ExtraIFaceBlacklist:   []string{"eth1", "eth2"},
-		DnsLabels:             []string{"label1", "label2"},
-		CleanDNSLabels:        false,
-		DnsRouteInterval:      durationpb.New(2 * time.Minute),
-		Mtu:                   &mtu,
-		SshJWTCacheTTL:        &sshJWTCacheTTL,
+		ProfileName:          profName,
+		Username:             currUser.Username,
+		ManagementUrl:        "https://new-api.netbird.io:443",
+		AdminURL:             "https://new-admin.netbird.io",
+		RosenpassEnabled:     &rosenpassEnabled,
+		RosenpassPermissive:  &rosenpassPermissive,
+		ServerSSHAllowed:     &serverSSHAllowed,
+		InterfaceName:        &interfaceName,
+		WireguardPort:        &wireguardPort,
+		OptionalPreSharedKey: &preSharedKey,
+		DisableAutoConnect:   &disableAutoConnect,
+		NetworkMonitor:       &networkMonitor,
+		DisableClientRoutes:  &disableClientRoutes,
+		DisableServerRoutes:  &disableServerRoutes,
+		DisableDns:           &disableDNS,
+		DisableFirewall:      &disableFirewall,
+		BlockLanAccess:       &blockLANAccess,
+		DisableNotifications: &disableNotifications,
+		BlockInbound:         &blockInbound,
+		DisableIpv6:          &disableIPv6,
+		NatExternalIPs:       []string{"1.2.3.4", "5.6.7.8"},
+		CleanNATExternalIPs:  false,
+		CustomDNSAddress:     []byte("1.1.1.1:53"),
+		ExtraIFaceBlacklist:  []string{"eth1", "eth2"},
+		DnsLabels:            []string{"label1", "label2"},
+		CleanDNSLabels:       false,
+		DnsRouteInterval:     durationpb.New(2 * time.Minute),
+		Mtu:                  &mtu,
+		SshJWTCacheTTL:       &sshJWTCacheTTL,
 	}
 
 	_, err = s.SetConfig(ctx, req)
@@ -140,7 +138,6 @@ func TestSetConfig_AllFieldsSaved(t *testing.T) {
 	require.Equal(t, blockLANAccess, cfg.BlockLANAccess)
 	require.NotNil(t, cfg.DisableNotifications)
 	require.Equal(t, disableNotifications, *cfg.DisableNotifications)
-	require.Equal(t, lazyConnectionEnabled, cfg.LazyConnectionEnabled)
 	require.Equal(t, blockInbound, cfg.BlockInbound)
 	require.Equal(t, disableIPv6, cfg.DisableIPv6)
 	require.Equal(t, []string{"1.2.3.4", "5.6.7.8"}, cfg.NATExternalIPs)
@@ -164,13 +161,14 @@ func verifyAllFieldsCovered(t *testing.T, req *proto.SetConfigRequest) {
 	t.Helper()
 
 	metadataFields := map[string]bool{
-		"state":               true, // protobuf internal
-		"sizeCache":           true, // protobuf internal
-		"unknownFields":       true, // protobuf internal
-		"Username":            true, // metadata
-		"ProfileName":         true, // metadata
-		"CleanNATExternalIPs": true, // control flag for clearing
-		"CleanDNSLabels":      true, // control flag for clearing
+		"state":                 true, // protobuf internal
+		"sizeCache":             true, // protobuf internal
+		"unknownFields":         true, // protobuf internal
+		"Username":              true, // metadata
+		"ProfileName":           true, // metadata
+		"CleanNATExternalIPs":   true, // control flag for clearing
+		"CleanDNSLabels":        true, // control flag for clearing
+		"LazyConnectionEnabled": true, // deprecated: proto field retained for compat, no longer applied
 	}
 
 	expectedFields := map[string]bool{
@@ -190,7 +188,6 @@ func verifyAllFieldsCovered(t *testing.T, req *proto.SetConfigRequest) {
 		"DisableFirewall":               true,
 		"BlockLanAccess":                true,
 		"DisableNotifications":          true,
-		"LazyConnectionEnabled":         true,
 		"BlockInbound":                  true,
 		"DisableIpv6":                   true,
 		"NatExternalIPs":                true,
@@ -252,7 +249,6 @@ func TestCLIFlags_MappedToSetConfig(t *testing.T) {
 		"block-lan-access":                  "BlockLanAccess",
 		"block-inbound":                     "BlockInbound",
 		"disable-ipv6":                      "DisableIpv6",
-		"enable-lazy-connection":            "LazyConnectionEnabled",
 		"external-ip-map":                   "NatExternalIPs",
 		"dns-resolver-address":              "CustomDNSAddress",
 		"extra-iface-blacklist":             "ExtraIFaceBlacklist",
@@ -269,7 +265,8 @@ func TestCLIFlags_MappedToSetConfig(t *testing.T) {
 
 	// SetConfigRequest fields that don't have CLI flags (settable only via UI or other means).
 	fieldsWithoutCLIFlags := map[string]bool{
-		"DisableNotifications": true, // Only settable via UI
+		"DisableNotifications":  true, // Only settable via UI
+		"LazyConnectionEnabled": true, // deprecated: no longer settable (managed by server + NB_LAZY_CONN)
 	}
 
 	// Get all SetConfigRequest fields to verify our map is complete.

client/system/info.go

@@ -74,8 +74,6 @@ type Info struct {
 	BlockInbound        bool
 	DisableIPv6         bool
 
-	LazyConnectionEnabled bool
-
 	EnableSSHRoot                 bool
 	EnableSSHSFTP                 bool
 	EnableSSHLocalPortForwarding  bool
@@ -87,7 +85,7 @@ func (i *Info) SetFlags(
 	rosenpassEnabled, rosenpassPermissive bool,
 	serverSSHAllowed *bool,
 	disableClientRoutes, disableServerRoutes,
-	disableDNS, disableFirewall, blockLANAccess, blockInbound, disableIPv6, lazyConnectionEnabled bool,
+	disableDNS, disableFirewall, blockLANAccess, blockInbound, disableIPv6 bool,
 	enableSSHRoot, enableSSHSFTP, enableSSHLocalPortForwarding, enableSSHRemotePortForwarding *bool,
 	disableSSHAuth *bool,
 ) {
@@ -105,8 +103,6 @@ func (i *Info) SetFlags(
 	i.BlockInbound = blockInbound
 	i.DisableIPv6 = disableIPv6
 
-	i.LazyConnectionEnabled = lazyConnectionEnabled
-
 	if enableSSHRoot != nil {
 		i.EnableSSHRoot = *enableSSHRoot
 	}

client/ui/client_ui.go

@@ -266,7 +266,6 @@ type serviceClient struct {
 	mAllowSSH          *systray.MenuItem
 	mAutoConnect       *systray.MenuItem
 	mEnableRosenpass   *systray.MenuItem
-	mLazyConnEnabled   *systray.MenuItem
 	mBlockInbound      *systray.MenuItem
 	mNotifications     *systray.MenuItem
 	mAdvancedSettings  *systray.MenuItem
@@ -336,11 +335,11 @@ type serviceClient struct {
 	// mNetworks + mExitNode submenu items. Combines features.DisableNetworks
 	// AND s.connected — both must be true for the menus to be active.
 	// Zero value (false) matches the Disable() call at AddMenuItem time.
-	networksMenuEnabled  bool
-	showNetworks         bool
-	wNetworks            fyne.Window
-	wProfiles            fyne.Window
-	wQuickActions        fyne.Window
+	networksMenuEnabled bool
+	showNetworks        bool
+	wNetworks           fyne.Window
+	wProfiles           fyne.Window
+	wQuickActions       fyne.Window
 
 	eventManager *event.Manager
 
@@ -1094,7 +1093,6 @@ func (s *serviceClient) onTrayReady() {
 	s.mAllowSSH = s.mSettings.AddSubMenuItemCheckbox("Allow SSH", allowSSHMenuDescr, false)
 	s.mAutoConnect = s.mSettings.AddSubMenuItemCheckbox("Connect on Startup", autoConnectMenuDescr, false)
 	s.mEnableRosenpass = s.mSettings.AddSubMenuItemCheckbox("Enable Quantum-Resistance", quantumResistanceMenuDescr, false)
-	s.mLazyConnEnabled = s.mSettings.AddSubMenuItemCheckbox("Enable Lazy Connections", lazyConnMenuDescr, false)
 	s.mBlockInbound = s.mSettings.AddSubMenuItemCheckbox("Block Inbound Connections", blockInboundMenuDescr, false)
 	s.mNotifications = s.mSettings.AddSubMenuItemCheckbox("Notifications", notificationsMenuDescr, false)
 	s.mSettings.AddSeparator()
@@ -1578,7 +1576,6 @@ func protoConfigToConfig(cfg *proto.GetConfigResponse) *profilemanager.Config {
 	config.RosenpassEnabled = cfg.RosenpassEnabled
 	config.RosenpassPermissive = cfg.RosenpassPermissive
 	config.DisableNotifications = &cfg.DisableNotifications
-	config.LazyConnectionEnabled = cfg.LazyConnectionEnabled
 	config.BlockInbound = cfg.BlockInbound
 	config.NetworkMonitor = &cfg.NetworkMonitor
 	config.DisableDNS = cfg.DisableDns
@@ -1682,12 +1679,6 @@ func (s *serviceClient) loadSettings() {
 		s.mEnableRosenpass.Uncheck()
 	}
 
-	if cfg.LazyConnectionEnabled {
-		s.mLazyConnEnabled.Check()
-	} else {
-		s.mLazyConnEnabled.Uncheck()
-	}
-
 	if cfg.BlockInbound {
 		s.mBlockInbound.Check()
 	} else {
@@ -1833,7 +1824,6 @@ func (s *serviceClient) updateConfig() error {
 	disableAutoStart := !s.mAutoConnect.Checked()
 	sshAllowed := s.mAllowSSH.Checked()
 	rosenpassEnabled := s.mEnableRosenpass.Checked()
-	lazyConnectionEnabled := s.mLazyConnEnabled.Checked()
 	blockInbound := s.mBlockInbound.Checked()
 	notificationsDisabled := !s.mNotifications.Checked()
 
@@ -1856,14 +1846,13 @@ func (s *serviceClient) updateConfig() error {
 	}
 
 	req := proto.SetConfigRequest{
-		ProfileName:           activeProf.ID.String(),
-		Username:              currUser.Username,
-		DisableAutoConnect:    &disableAutoStart,
-		ServerSSHAllowed:      &sshAllowed,
-		RosenpassEnabled:      &rosenpassEnabled,
-		LazyConnectionEnabled: &lazyConnectionEnabled,
-		BlockInbound:          &blockInbound,
-		DisableNotifications:  &notificationsDisabled,
+		ProfileName:          activeProf.ID.String(),
+		Username:             currUser.Username,
+		DisableAutoConnect:   &disableAutoStart,
+		ServerSSHAllowed:     &sshAllowed,
+		RosenpassEnabled:     &rosenpassEnabled,
+		BlockInbound:         &blockInbound,
+		DisableNotifications: &notificationsDisabled,
 	}
 
 	if _, err := conn.SetConfig(s.ctx, &req); err != nil {

client/ui/const.go

@@ -4,7 +4,6 @@ const (
 	allowSSHMenuDescr          = "Allow SSH connections"
 	autoConnectMenuDescr       = "Connect automatically when the service starts"
 	quantumResistanceMenuDescr = "Enable post-quantum security via Rosenpass"
-	lazyConnMenuDescr          = "[Experimental] Enable lazy connections"
 	blockInboundMenuDescr      = "Block inbound connections to the local machine and routed networks"
 	notificationsMenuDescr     = "Enable notifications"
 	advancedSettingsMenuDescr  = "Advanced settings of the application"

client/ui/event_handler.go

@@ -43,8 +43,6 @@ func (h *eventHandler) listen(ctx context.Context) {
 			h.handleAutoConnectClick()
 		case <-h.client.mEnableRosenpass.ClickedCh:
 			h.handleRosenpassClick()
-		case <-h.client.mLazyConnEnabled.ClickedCh:
-			h.handleLazyConnectionClick()
 		case <-h.client.mBlockInbound.ClickedCh:
 			h.handleBlockInboundClick()
 		case <-h.client.mAdvancedSettings.ClickedCh:
@@ -152,15 +150,6 @@ func (h *eventHandler) handleRosenpassClick() {
 	}
 }
 
-func (h *eventHandler) handleLazyConnectionClick() {
-	h.toggleCheckbox(h.client.mLazyConnEnabled)
-	if err := h.updateConfigWithErr(); err != nil {
-		h.toggleCheckbox(h.client.mLazyConnEnabled) // revert checkbox state on error
-		log.Errorf("failed to update config: %v", err)
-		h.client.notifier.Send("Error", "Failed to update lazy connection settings")
-	}
-}
-
 func (h *eventHandler) handleBlockInboundClick() {
 	h.toggleCheckbox(h.client.mBlockInbound)
 	if err := h.updateConfigWithErr(); err != nil {

e2e/agentnetwork/skiptls_test.go

@@ -0,0 +1,140 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestProviderSkipTLSVerification proves skip_tls_verification is per-provider:
+// two providers share one self-signed upstream, one skipping TLS verification
+// and one not. The skip=true provider's chat reaches the upstream and returns
+// 200; the skip=false provider's chat fails at the TLS handshake — same
+// upstream, opposite outcome. This is the behaviour a target-level flag could
+// not give, since all of an account's providers share one synthesised target.
+func TestProviderSkipTLSVerification(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	up, err := harness.StartFakeUpstream(ctx, srv)
+	require.NoError(t, err, "start self-signed upstream")
+	t.Cleanup(func() { _ = up.Terminate(context.Background()) })
+
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-skiptls"})
+	require.NoError(t, err, "create group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-skiptls-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grp.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	const (
+		insecureModel = "insecure-model"
+		secureModel   = "secure-model"
+	)
+
+	// Two providers on the SAME self-signed upstream, distinguished only by their
+	// skip_tls_verification and a unique model string so the router picks each
+	// unambiguously.
+	newReq := func(name, model string, skip bool) api.AgentNetworkProviderRequest {
+		key := "sk-dummy-e2e"
+		return api.AgentNetworkProviderRequest{
+			Name:                name,
+			ProviderId:          "openai_api",
+			UpstreamUrl:         up.URL,
+			ApiKey:              &key,
+			Enabled:             ptr(true),
+			SkipTlsVerification: ptr(skip),
+			Models: &[]api.AgentNetworkProviderModel{
+				{Id: model, InputPer1k: 0.001, OutputPer1k: 0.002},
+			},
+		}
+	}
+
+	// First create bootstraps the account cluster.
+	insecureReq := newReq("skip-tls", insecureModel, true)
+	insecureReq.BootstrapCluster = ptr(harness.AgentNetworkCluster)
+	insecureProv, err := srv.CreateProvider(ctx, insecureReq)
+	require.NoError(t, err, "create skip-tls provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), insecureProv.Id) })
+	require.True(t, insecureProv.SkipTlsVerification, "response must echo skip_tls_verification=true")
+
+	secureProv, err := srv.CreateProvider(ctx, newReq("verify-tls", secureModel, false))
+	require.NoError(t, err, "create verify-tls provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), secureProv.Id) })
+	require.False(t, secureProv.SkipTlsVerification, "response must echo skip_tls_verification=false")
+
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-skiptls-allow",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: []string{insecureProv.Id, secureProv.Id},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-skiptls-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+
+	// Positive: skip=true reaches the self-signed upstream. Retry to absorb
+	// tunnel/DNS jitter on the first call; success also proves the path works.
+	var code int
+	var body string
+	deadline := time.Now().Add(90 * time.Second)
+	for time.Now().Before(deadline) {
+		c, b, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, insecureModel, "Reply with exactly: pong", "e2e-skiptls-insecure")
+		if cerr == nil {
+			code, body = c, b
+			if code == 200 {
+				break
+			}
+		}
+		time.Sleep(5 * time.Second)
+	}
+	require.Equal(t, 200, code,
+		"skip_tls_verification=true must reach the self-signed upstream; body: %s\n=== upstream logs ===\n%s\n=== proxy logs ===\n%s",
+		body, up.Logs(context.Background()), px.Logs(context.Background()))
+
+	// Negative: skip=false must fail the TLS handshake to the SAME upstream. The
+	// path is already proven working, so a non-200 here is the cert rejection.
+	secureCode, secureBody, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, secureModel, "Reply with exactly: pong", "e2e-skiptls-secure")
+	require.NoError(t, cerr, "the chat call itself must complete (proxy returns an error status, not a transport error)")
+	require.NotEqual(t, 200, secureCode,
+		"skip_tls_verification=false must NOT reach the self-signed upstream; got %d, body: %s", secureCode, secureBody)
+	require.GreaterOrEqual(t, secureCode, 500,
+		"a TLS verification failure should surface as a 5xx from the proxy; got %d, body: %s", secureCode, secureBody)
+}

e2e/agentnetwork/vllm_test.go

@@ -0,0 +1,171 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestVLLMProvider proves the proxy supports a self-hosted vLLM backend. vLLM is
+// OpenAI-compatible, so it uses the "vllm" catalog entry (KindCustom) and is
+// reached over plain HTTP — no TLS anywhere on the path:
+//
+//	client --tunnel--> netbird proxy --http--> vllm (:8000, OpenAI-compatible)
+//
+// The mock vLLM server answers /v1/chat/completions with an OpenAI-shaped
+// completion carrying a non-zero usage block. The test asserts the chat returns
+// 200 with the completion, that the request is recorded in the access log by its
+// session id, and that vLLM's usage block is metered into a consumption row —
+// which together prove request routing, response parsing, and token accounting
+// all work for a self-hosted OpenAI-compatible provider.
+//
+// It needs no external credentials (the mock ignores auth), so it always runs.
+func TestVLLMProvider(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock vLLM server")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-vllm"})
+	require.NoError(t, err, "create group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-vllm-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grp.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	// vLLM provider pointed at the mock over plain HTTP. The mock ignores auth,
+	// so a dummy key satisfies the "Bearer ${API_KEY}" template. The served model
+	// is enumerated so the router dispatches this model string to this provider.
+	dummyKey := "sk-vllm-e2e"
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:             "vllm",
+		ProviderId:       "vllm",
+		UpstreamUrl:      vllm.URL,
+		ApiKey:           &dummyKey,
+		Enabled:          ptr(true),
+		BootstrapCluster: ptr(harness.AgentNetworkCluster),
+		Models: &[]api.AgentNetworkProviderModel{
+			{Id: harness.VLLMModel, InputPer1k: 0.001, OutputPer1k: 0.002},
+		},
+	})
+	require.NoError(t, err, "create vllm provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	// Token limit far above the handful of tokens this test drives, so it never
+	// blocks but switches on usage metering — the switch that makes consumption
+	// rows get recorded.
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-vllm-allow",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: []string{prov.Id},
+		Limits: &api.AgentNetworkPolicyLimits{
+			TokenLimit: api.AgentNetworkPolicyTokenLimit{
+				Enabled:       true,
+				GroupCap:      10_000_000,
+				UserCap:       10_000_000,
+				WindowSeconds: 60,
+			},
+		},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-vllm-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+
+	before, _ := srv.ListAccessLogs(ctx)
+	sessionID := "e2e-session-vllm"
+
+	// Retry to absorb tunnel/DNS jitter on the first call.
+	var code int
+	var body string
+	deadline := time.Now().Add(90 * time.Second)
+	for time.Now().Before(deadline) {
+		c, b, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, harness.VLLMModel, "Reply with exactly: pong", sessionID)
+		if cerr == nil {
+			code, body = c, b
+			if code == 200 {
+				break
+			}
+		}
+		time.Sleep(5 * time.Second)
+	}
+	require.Equal(t, 200, code,
+		"chat through the vLLM provider must return 200; body: %s\n=== vllm logs ===\n%s\n=== proxy logs ===\n%s",
+		body, vllm.Logs(context.Background()), px.Logs(context.Background()))
+	require.True(t, strings.Contains(body, "chat.completion"),
+		"body should be an OpenAI-compatible chat completion; got: %s", body)
+
+	// The request must surface as an access-log row carrying our session id.
+	require.Eventually(t, func() bool {
+		logs, lerr := srv.ListAccessLogs(ctx)
+		return lerr == nil && logs.TotalRecords > before.TotalRecords
+	}, 30*time.Second, 2*time.Second, "an access-log row should be ingested for the vLLM provider")
+
+	require.Eventually(t, func() bool {
+		logs, lerr := srv.ListAccessLogs(ctx)
+		if lerr != nil {
+			return false
+		}
+		for _, r := range logs.Data {
+			if r.SessionId != nil && *r.SessionId == sessionID {
+				return true
+			}
+		}
+		return false
+	}, 30*time.Second, 2*time.Second, "session id %q must be recorded in an access-log row", sessionID)
+
+	// vLLM's usage block (prompt_tokens=11, completion_tokens=2) must be parsed
+	// and metered into a consumption row with positive token counts.
+	require.Eventually(t, func() bool {
+		rows, lerr := srv.ListConsumption(ctx)
+		if lerr != nil {
+			return false
+		}
+		for _, r := range rows {
+			if r.TokensInput > 0 && r.TokensOutput > 0 {
+				return true
+			}
+		}
+		return false
+	}, 60*time.Second, 3*time.Second, "vLLM usage must be metered into a consumption row")
+}

e2e/harness/client.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"io"
 	"os/exec"
+	"strconv"
 	"strings"
 	"time"
 
@@ -108,9 +109,48 @@ func (cl *Client) WaitConnected(ctx context.Context, timeout time.Duration) erro
 	return cl.pollStatus(ctx, timeout, "Management: Connected")
 }
 
-// WaitProxyPeer polls until the client sees the proxy peer connected (1/1).
+// WaitProxyPeer polls until the client sees at least one connected peer — the
+// proxy serving the agent-network endpoint. It requires ">=1 connected" rather
+// than an exact "1/1" because proxy peers from earlier tests linger in the
+// account as disconnected (each proxy container registers a fresh WireGuard key
+// and the peer is not removed on teardown), so the count is e.g. "1/2". Only the
+// live proxy can be connected, and the caller's subsequent chat is the real
+// end-to-end assertion.
 func (cl *Client) WaitProxyPeer(ctx context.Context, timeout time.Duration) error {
-	return cl.pollStatus(ctx, timeout, "1/1 Connected")
+	deadline := time.Now().Add(timeout)
+	var last string
+	for time.Now().Before(deadline) {
+		out, _ := cl.Status(ctx)
+		last = out
+		if connectedPeers(out) >= 1 {
+			return nil
+		}
+		time.Sleep(3 * time.Second)
+	}
+	return fmt.Errorf("timed out waiting for a connected proxy peer; last status:\n%s", last)
+}
+
+// connectedPeers parses the "Peers count: X/Y Connected" line from `netbird
+// status` and returns X (the connected count), or 0 when absent/unparseable.
+func connectedPeers(status string) int {
+	for _, line := range strings.Split(status, "\n") {
+		line = strings.TrimSpace(line)
+		rest, ok := strings.CutPrefix(line, "Peers count:")
+		if !ok {
+			continue
+		}
+		rest = strings.TrimSpace(rest)
+		slash := strings.IndexByte(rest, '/')
+		if slash <= 0 {
+			return 0
+		}
+		n, err := strconv.Atoi(strings.TrimSpace(rest[:slash]))
+		if err != nil {
+			return 0
+		}
+		return n
+	}
+	return 0
 }
 
 func (cl *Client) pollStatus(ctx context.Context, timeout time.Duration, want string) error {

e2e/harness/upstream.go

@@ -0,0 +1,107 @@
+//go:build e2e
+
+package harness
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+
+	"github.com/docker/docker/api/types/container"
+	"github.com/testcontainers/testcontainers-go"
+	"github.com/testcontainers/testcontainers-go/wait"
+)
+
+const (
+	fakeUpstreamImage = "nginx:alpine"
+	fakeUpstreamAlias = "fakeupstream"
+	fakeUpstreamPort  = "443/tcp"
+)
+
+// fakeUpstreamNginxConf serves a canned OpenAI-shaped chat completion for any
+// path over a self-signed certificate, so the proxy reaches it only when the
+// provider opts into skipping TLS verification.
+const fakeUpstreamNginxConf = `pid /tmp/nginx.pid;
+events {}
+http {
+  server {
+    listen 443 ssl;
+    ssl_certificate     /certs/tls.crt;
+    ssl_certificate_key /certs/tls.key;
+    location / {
+      default_type application/json;
+      return 200 '{"id":"chatcmpl-e2e","object":"chat.completion","choices":[{"index":0,"message":{"role":"assistant","content":"pong"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}';
+    }
+  }
+}
+`
+
+// FakeUpstream is a self-signed HTTPS server on the combined server's network,
+// used to exercise provider skip_tls_verification: a proxy that verifies the
+// certificate rejects it, one that skips verification reaches it.
+type FakeUpstream struct {
+	container testcontainers.Container
+	workDir   string
+	// URL is the upstream URL providers point at (https://<alias>).
+	URL string
+}
+
+// StartFakeUpstream runs the self-signed upstream on the shared network.
+func StartFakeUpstream(ctx context.Context, c *Combined) (*FakeUpstream, error) {
+	workDir, err := os.MkdirTemp("/tmp", "nb-e2e-upstream-*")
+	if err != nil {
+		return nil, fmt.Errorf("create upstream work dir: %w", err)
+	}
+	// Widen so the (non-root worker) nginx container can traverse the bind mount.
+	if err := os.Chmod(workDir, 0o755); err != nil { //nolint:gosec // throwaway e2e cert dir
+		return nil, fmt.Errorf("chmod upstream dir: %w", err)
+	}
+	if err := writeSelfSignedCert(workDir, []string{fakeUpstreamAlias}); err != nil {
+		return nil, err
+	}
+	if err := os.WriteFile(filepath.Join(workDir, "nginx.conf"), []byte(fakeUpstreamNginxConf), 0o644); err != nil { //nolint:gosec // non-secret e2e config
+		return nil, fmt.Errorf("write nginx conf: %w", err)
+	}
+
+	req := testcontainers.ContainerRequest{
+		Image:          fakeUpstreamImage,
+		ExposedPorts:   []string{fakeUpstreamPort},
+		Networks:       []string{c.network.Name},
+		NetworkAliases: map[string][]string{c.network.Name: {fakeUpstreamAlias}},
+		Cmd:            []string{"nginx", "-c", "/certs/nginx.conf", "-g", "daemon off;"},
+		HostConfigModifier: func(hc *container.HostConfig) {
+			hc.Binds = append(hc.Binds, workDir+":/certs:ro")
+		},
+		WaitingFor: wait.ForListeningPort(fakeUpstreamPort).WithStartupTimeout(60 * time.Second),
+	}
+
+	ctr, err := testcontainers.GenericContainer(ctx, testcontainers.GenericContainerRequest{
+		ContainerRequest: req,
+		Started:          true,
+	})
+	if err != nil {
+		_ = os.RemoveAll(workDir)
+		return nil, fmt.Errorf("start fake upstream container: %w", err)
+	}
+
+	return &FakeUpstream{container: ctr, workDir: workDir, URL: "https://" + fakeUpstreamAlias}, nil
+}
+
+// Logs returns the upstream container logs, for diagnostics on failure.
+func (u *FakeUpstream) Logs(ctx context.Context) string {
+	return containerLogs(ctx, u.container)
+}
+
+// Terminate stops the upstream container and cleans its work dir.
+func (u *FakeUpstream) Terminate(ctx context.Context) error {
+	var err error
+	if u.container != nil {
+		err = u.container.Terminate(ctx)
+	}
+	if u.workDir != "" {
+		_ = os.RemoveAll(u.workDir)
+	}
+	return err
+}

e2e/harness/vllm.go

@@ -0,0 +1,113 @@
+//go:build e2e
+
+package harness
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+
+	"github.com/docker/docker/api/types/container"
+	"github.com/testcontainers/testcontainers-go"
+	"github.com/testcontainers/testcontainers-go/wait"
+)
+
+const (
+	vllmImage = "nginx:alpine"
+	vllmAlias = "vllm"
+	vllmPort  = "8000/tcp"
+	// VLLMModel is the served model id the mock advertises and echoes back. It
+	// matches a real small model commonly served by vLLM so the provider's
+	// enumerated model and the client's request line up.
+	VLLMModel = "Qwen/Qwen2.5-0.5B-Instruct"
+)
+
+// vllmNginxConf emulates a vLLM OpenAI-compatible server over plain HTTP (vLLM's
+// default: no TLS, port 8000). It answers /v1/models with a one-model list and
+// any chat/completions path with a canned OpenAI-shaped chat completion carrying
+// a non-zero usage block, so the proxy's OpenAI parser records real token
+// consumption. Running actual vLLM in CI is infeasible (GPU + multi-GB model
+// download), so this stands in for the wire contract the proxy depends on.
+const vllmNginxConf = `pid /tmp/nginx.pid;
+events {}
+http {
+  server {
+    listen 8000;
+    location = /v1/models {
+      default_type application/json;
+      return 200 '{"object":"list","data":[{"id":"Qwen/Qwen2.5-0.5B-Instruct","object":"model","owned_by":"vllm"}]}';
+    }
+    location / {
+      default_type application/json;
+      return 200 '{"id":"chatcmpl-e2e-vllm","object":"chat.completion","created":1700000000,"model":"Qwen/Qwen2.5-0.5B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"pong"},"finish_reason":"stop"}],"usage":{"prompt_tokens":11,"completion_tokens":2,"total_tokens":13}}';
+    }
+  }
+}
+`
+
+// VLLM is a mock vLLM OpenAI-compatible server on the combined server's network,
+// reachable at http://vllm:8000. A "vllm" provider points at it to exercise the
+// proxy's support for self-hosted OpenAI-compatible backends.
+type VLLM struct {
+	container testcontainers.Container
+	workDir   string
+	// URL is the upstream URL the vllm provider points at (http://<alias>:8000).
+	URL string
+}
+
+// StartVLLM runs the mock vLLM server on the shared network over plain HTTP.
+func StartVLLM(ctx context.Context, c *Combined) (*VLLM, error) {
+	workDir, err := os.MkdirTemp("/tmp", "nb-e2e-vllm-*")
+	if err != nil {
+		return nil, fmt.Errorf("create vllm work dir: %w", err)
+	}
+	// Widen so the (non-root worker) nginx container can traverse the bind mount.
+	if err := os.Chmod(workDir, 0o755); err != nil { //nolint:gosec // throwaway e2e config dir
+		return nil, fmt.Errorf("chmod vllm dir: %w", err)
+	}
+	if err := os.WriteFile(filepath.Join(workDir, "nginx.conf"), []byte(vllmNginxConf), 0o644); err != nil { //nolint:gosec // non-secret e2e config
+		return nil, fmt.Errorf("write nginx conf: %w", err)
+	}
+
+	req := testcontainers.ContainerRequest{
+		Image:          vllmImage,
+		ExposedPorts:   []string{vllmPort},
+		Networks:       []string{c.network.Name},
+		NetworkAliases: map[string][]string{c.network.Name: {vllmAlias}},
+		Cmd:            []string{"nginx", "-c", "/conf/nginx.conf", "-g", "daemon off;"},
+		HostConfigModifier: func(hc *container.HostConfig) {
+			hc.Binds = append(hc.Binds, workDir+":/conf:ro")
+		},
+		WaitingFor: wait.ForListeningPort(vllmPort).WithStartupTimeout(60 * time.Second),
+	}
+
+	ctr, err := testcontainers.GenericContainer(ctx, testcontainers.GenericContainerRequest{
+		ContainerRequest: req,
+		Started:          true,
+	})
+	if err != nil {
+		_ = os.RemoveAll(workDir)
+		return nil, fmt.Errorf("start vllm container: %w", err)
+	}
+
+	return &VLLM{container: ctr, workDir: workDir, URL: "http://" + vllmAlias + ":8000"}, nil
+}
+
+// Logs returns the vLLM container logs, for diagnostics on failure.
+func (v *VLLM) Logs(ctx context.Context) string {
+	return containerLogs(ctx, v.container)
+}
+
+// Terminate stops the vLLM container and cleans its work dir.
+func (v *VLLM) Terminate(ctx context.Context) error {
+	var err error
+	if v.container != nil {
+		err = v.container.Terminate(ctx)
+	}
+	if v.workDir != "" {
+		_ = os.RemoveAll(v.workDir)
+	}
+	return err
+}

flow/client/client.go

@@ -109,7 +109,7 @@ func (c *GRPCClient) Close() error {
 func (c *GRPCClient) Send(event *proto.FlowEvent) error {
 	c.mu.Lock()
 	stream := c.stream
-	defer c.mu.Unlock() // stream.Send() is not safe to call concurrently from multiple goroutines
+	c.mu.Unlock()
 
 	if stream == nil {
 		return errors.New("stream not initialized")

flow/proto/flow.pb.go

@@ -134,11 +134,9 @@ type FlowEvent struct {
 	// When the event occurred
 	Timestamp *timestamppb.Timestamp `protobuf:"bytes,2,opt,name=timestamp,proto3" json:"timestamp,omitempty"`
 	// Public key of the sending peer
-	PublicKey   []byte                 `protobuf:"bytes,3,opt,name=public_key,json=publicKey,proto3" json:"public_key,omitempty"`
-	FlowFields  *FlowFields            `protobuf:"bytes,4,opt,name=flow_fields,json=flowFields,proto3" json:"flow_fields,omitempty"`
-	IsInitiator bool                   `protobuf:"varint,5,opt,name=isInitiator,proto3" json:"isInitiator,omitempty"`
-	WindowStart *timestamppb.Timestamp `protobuf:"bytes,6,opt,name=window_start,json=windowStart,proto3" json:"window_start,omitempty"`
-	WindowEnd   *timestamppb.Timestamp `protobuf:"bytes,7,opt,name=window_end,json=windowEnd,proto3" json:"window_end,omitempty"`
+	PublicKey   []byte      `protobuf:"bytes,3,opt,name=public_key,json=publicKey,proto3" json:"public_key,omitempty"`
+	FlowFields  *FlowFields `protobuf:"bytes,4,opt,name=flow_fields,json=flowFields,proto3" json:"flow_fields,omitempty"`
+	IsInitiator bool        `protobuf:"varint,5,opt,name=isInitiator,proto3" json:"isInitiator,omitempty"`
 }
 
 func (x *FlowEvent) Reset() {
@@ -208,20 +206,6 @@ func (x *FlowEvent) GetIsInitiator() bool {
 	return false
 }
 
-func (x *FlowEvent) GetWindowStart() *timestamppb.Timestamp {
-	if x != nil {
-		return x.WindowStart
-	}
-	return nil
-}
-
-func (x *FlowEvent) GetWindowEnd() *timestamppb.Timestamp {
-	if x != nil {
-		return x.WindowEnd
-	}
-	return nil
-}
-
 type FlowEventAck struct {
 	state         protoimpl.MessageState
 	sizeCache     protoimpl.SizeCache
@@ -300,6 +284,7 @@ type FlowFields struct {
 	// Layer 4 -specific information
 	//
 	// Types that are assignable to ConnectionInfo:
+	//
 	//	*FlowFields_PortInfo
 	//	*FlowFields_IcmpInfo
 	ConnectionInfo isFlowFields_ConnectionInfo `protobuf_oneof:"connection_info"`
@@ -312,9 +297,6 @@ type FlowFields struct {
 	// Resource ID
 	SourceResourceId []byte `protobuf:"bytes,14,opt,name=source_resource_id,json=sourceResourceId,proto3" json:"source_resource_id,omitempty"`
 	DestResourceId   []byte `protobuf:"bytes,15,opt,name=dest_resource_id,json=destResourceId,proto3" json:"dest_resource_id,omitempty"`
-	NumOfStarts      uint64 `protobuf:"varint,16,opt,name=num_of_starts,json=numOfStarts,proto3" json:"num_of_starts,omitempty"`
-	NumOfEnds        uint64 `protobuf:"varint,17,opt,name=num_of_ends,json=numOfEnds,proto3" json:"num_of_ends,omitempty"`
-	NumOfDrops       uint64 `protobuf:"varint,18,opt,name=num_of_drops,json=numOfDrops,proto3" json:"num_of_drops,omitempty"`
 }
 
 func (x *FlowFields) Reset() {
@@ -461,27 +443,6 @@ func (x *FlowFields) GetDestResourceId() []byte {
 	return nil
 }
 
-func (x *FlowFields) GetNumOfStarts() uint64 {
-	if x != nil {
-		return x.NumOfStarts
-	}
-	return 0
-}
-
-func (x *FlowFields) GetNumOfEnds() uint64 {
-	if x != nil {
-		return x.NumOfEnds
-	}
-	return 0
-}
-
-func (x *FlowFields) GetNumOfDrops() uint64 {
-	if x != nil {
-		return x.NumOfDrops
-	}
-	return 0
-}
-
 type isFlowFields_ConnectionInfo interface {
 	isFlowFields_ConnectionInfo()
 }
@@ -618,7 +579,7 @@ var file_flow_proto_rawDesc = []byte{
 	0x0a, 0x0a, 0x66, 0x6c, 0x6f, 0x77, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x12, 0x04, 0x66, 0x6c,
 	0x6f, 0x77, 0x1a, 0x1f, 0x67, 0x6f, 0x6f, 0x67, 0x6c, 0x65, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f,
 	0x62, 0x75, 0x66, 0x2f, 0x74, 0x69, 0x6d, 0x65, 0x73, 0x74, 0x61, 0x6d, 0x70, 0x2e, 0x70, 0x72,
-	0x6f, 0x74, 0x6f, 0x22, 0xce, 0x02, 0x0a, 0x09, 0x46, 0x6c, 0x6f, 0x77, 0x45, 0x76, 0x65, 0x6e,
+	0x6f, 0x74, 0x6f, 0x22, 0xd4, 0x01, 0x0a, 0x09, 0x46, 0x6c, 0x6f, 0x77, 0x45, 0x76, 0x65, 0x6e,
 	0x74, 0x12, 0x19, 0x0a, 0x08, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x5f, 0x69, 0x64, 0x18, 0x01, 0x20,
 	0x01, 0x28, 0x0c, 0x52, 0x07, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x49, 0x64, 0x12, 0x38, 0x0a, 0x09,
 	0x74, 0x69, 0x6d, 0x65, 0x73, 0x74, 0x61, 0x6d, 0x70, 0x18, 0x02, 0x20, 0x01, 0x28, 0x0b, 0x32,
@@ -631,59 +592,45 @@ var file_flow_proto_rawDesc = []byte{
 	0x77, 0x2e, 0x46, 0x6c, 0x6f, 0x77, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x52, 0x0a, 0x66, 0x6c,
 	0x6f, 0x77, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x12, 0x20, 0x0a, 0x0b, 0x69, 0x73, 0x49, 0x6e,
 	0x69, 0x74, 0x69, 0x61, 0x74, 0x6f, 0x72, 0x18, 0x05, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0b, 0x69,
-	0x73, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x74, 0x6f, 0x72, 0x12, 0x3d, 0x0a, 0x0c, 0x77, 0x69,
-	0x6e, 0x64, 0x6f, 0x77, 0x5f, 0x73, 0x74, 0x61, 0x72, 0x74, 0x18, 0x06, 0x20, 0x01, 0x28, 0x0b,
-	0x32, 0x1a, 0x2e, 0x67, 0x6f, 0x6f, 0x67, 0x6c, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62,
-	0x75, 0x66, 0x2e, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x74, 0x61, 0x6d, 0x70, 0x52, 0x0b, 0x77, 0x69,
-	0x6e, 0x64, 0x6f, 0x77, 0x53, 0x74, 0x61, 0x72, 0x74, 0x12, 0x39, 0x0a, 0x0a, 0x77, 0x69, 0x6e,
-	0x64, 0x6f, 0x77, 0x5f, 0x65, 0x6e, 0x64, 0x18, 0x07, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x1a, 0x2e,
-	0x67, 0x6f, 0x6f, 0x67, 0x6c, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62, 0x75, 0x66, 0x2e,
-	0x54, 0x69, 0x6d, 0x65, 0x73, 0x74, 0x61, 0x6d, 0x70, 0x52, 0x09, 0x77, 0x69, 0x6e, 0x64, 0x6f,
-	0x77, 0x45, 0x6e, 0x64, 0x22, 0x4b, 0x0a, 0x0c, 0x46, 0x6c, 0x6f, 0x77, 0x45, 0x76, 0x65, 0x6e,
-	0x74, 0x41, 0x63, 0x6b, 0x12, 0x19, 0x0a, 0x08, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x5f, 0x69, 0x64,
-	0x18, 0x01, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x07, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x49, 0x64, 0x12,
-	0x20, 0x0a, 0x0b, 0x69, 0x73, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x74, 0x6f, 0x72, 0x18, 0x02,
-	0x20, 0x01, 0x28, 0x08, 0x52, 0x0b, 0x69, 0x73, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x74, 0x6f,
-	0x72, 0x22, 0x82, 0x05, 0x0a, 0x0a, 0x46, 0x6c, 0x6f, 0x77, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x73,
-	0x12, 0x17, 0x0a, 0x07, 0x66, 0x6c, 0x6f, 0x77, 0x5f, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28,
-	0x0c, 0x52, 0x06, 0x66, 0x6c, 0x6f, 0x77, 0x49, 0x64, 0x12, 0x1e, 0x0a, 0x04, 0x74, 0x79, 0x70,
-	0x65, 0x18, 0x02, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x0a, 0x2e, 0x66, 0x6c, 0x6f, 0x77, 0x2e, 0x54,
-	0x79, 0x70, 0x65, 0x52, 0x04, 0x74, 0x79, 0x70, 0x65, 0x12, 0x17, 0x0a, 0x07, 0x72, 0x75, 0x6c,
-	0x65, 0x5f, 0x69, 0x64, 0x18, 0x03, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x06, 0x72, 0x75, 0x6c, 0x65,
-	0x49, 0x64, 0x12, 0x2d, 0x0a, 0x09, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x18,
-	0x04, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x0f, 0x2e, 0x66, 0x6c, 0x6f, 0x77, 0x2e, 0x44, 0x69, 0x72,
-	0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x09, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x69, 0x6f,
-	0x6e, 0x12, 0x1a, 0x0a, 0x08, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x63, 0x6f, 0x6c, 0x18, 0x05, 0x20,
-	0x01, 0x28, 0x0d, 0x52, 0x08, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x63, 0x6f, 0x6c, 0x12, 0x1b, 0x0a,
-	0x09, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x69, 0x70, 0x18, 0x06, 0x20, 0x01, 0x28, 0x0c,
-	0x52, 0x08, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x49, 0x70, 0x12, 0x17, 0x0a, 0x07, 0x64, 0x65,
-	0x73, 0x74, 0x5f, 0x69, 0x70, 0x18, 0x07, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x06, 0x64, 0x65, 0x73,
-	0x74, 0x49, 0x70, 0x12, 0x2d, 0x0a, 0x09, 0x70, 0x6f, 0x72, 0x74, 0x5f, 0x69, 0x6e, 0x66, 0x6f,
-	0x18, 0x08, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x0e, 0x2e, 0x66, 0x6c, 0x6f, 0x77, 0x2e, 0x50, 0x6f,
-	0x72, 0x74, 0x49, 0x6e, 0x66, 0x6f, 0x48, 0x00, 0x52, 0x08, 0x70, 0x6f, 0x72, 0x74, 0x49, 0x6e,
-	0x66, 0x6f, 0x12, 0x2d, 0x0a, 0x09, 0x69, 0x63, 0x6d, 0x70, 0x5f, 0x69, 0x6e, 0x66, 0x6f, 0x18,
-	0x09, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x0e, 0x2e, 0x66, 0x6c, 0x6f, 0x77, 0x2e, 0x49, 0x43, 0x4d,
-	0x50, 0x49, 0x6e, 0x66, 0x6f, 0x48, 0x00, 0x52, 0x08, 0x69, 0x63, 0x6d, 0x70, 0x49, 0x6e, 0x66,
-	0x6f, 0x12, 0x1d, 0x0a, 0x0a, 0x72, 0x78, 0x5f, 0x70, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x18,
-	0x0a, 0x20, 0x01, 0x28, 0x04, 0x52, 0x09, 0x72, 0x78, 0x50, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73,
-	0x12, 0x1d, 0x0a, 0x0a, 0x74, 0x78, 0x5f, 0x70, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x18, 0x0b,
-	0x20, 0x01, 0x28, 0x04, 0x52, 0x09, 0x74, 0x78, 0x50, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x12,
-	0x19, 0x0a, 0x08, 0x72, 0x78, 0x5f, 0x62, 0x79, 0x74, 0x65, 0x73, 0x18, 0x0c, 0x20, 0x01, 0x28,
-	0x04, 0x52, 0x07, 0x72, 0x78, 0x42, 0x79, 0x74, 0x65, 0x73, 0x12, 0x19, 0x0a, 0x08, 0x74, 0x78,
-	0x5f, 0x62, 0x79, 0x74, 0x65, 0x73, 0x18, 0x0d, 0x20, 0x01, 0x28, 0x04, 0x52, 0x07, 0x74, 0x78,
-	0x42, 0x79, 0x74, 0x65, 0x73, 0x12, 0x2c, 0x0a, 0x12, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f,
-	0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x69, 0x64, 0x18, 0x0e, 0x20, 0x01, 0x28,
-	0x0c, 0x52, 0x10, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x52, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63,
-	0x65, 0x49, 0x64, 0x12, 0x28, 0x0a, 0x10, 0x64, 0x65, 0x73, 0x74, 0x5f, 0x72, 0x65, 0x73, 0x6f,
-	0x75, 0x72, 0x63, 0x65, 0x5f, 0x69, 0x64, 0x18, 0x0f, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x0e, 0x64,
-	0x65, 0x73, 0x74, 0x52, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x49, 0x64, 0x12, 0x22, 0x0a,
-	0x0d, 0x6e, 0x75, 0x6d, 0x5f, 0x6f, 0x66, 0x5f, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x18, 0x10,
-	0x20, 0x01, 0x28, 0x04, 0x52, 0x0b, 0x6e, 0x75, 0x6d, 0x4f, 0x66, 0x53, 0x74, 0x61, 0x72, 0x74,
-	0x73, 0x12, 0x1e, 0x0a, 0x0b, 0x6e, 0x75, 0x6d, 0x5f, 0x6f, 0x66, 0x5f, 0x65, 0x6e, 0x64, 0x73,
-	0x18, 0x11, 0x20, 0x01, 0x28, 0x04, 0x52, 0x09, 0x6e, 0x75, 0x6d, 0x4f, 0x66, 0x45, 0x6e, 0x64,
-	0x73, 0x12, 0x20, 0x0a, 0x0c, 0x6e, 0x75, 0x6d, 0x5f, 0x6f, 0x66, 0x5f, 0x64, 0x72, 0x6f, 0x70,
-	0x73, 0x18, 0x12, 0x20, 0x01, 0x28, 0x04, 0x52, 0x0a, 0x6e, 0x75, 0x6d, 0x4f, 0x66, 0x44, 0x72,
-	0x6f, 0x70, 0x73, 0x42, 0x11, 0x0a, 0x0f, 0x63, 0x6f, 0x6e, 0x6e, 0x65, 0x63, 0x74, 0x69, 0x6f,
+	0x73, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x74, 0x6f, 0x72, 0x22, 0x4b, 0x0a, 0x0c, 0x46, 0x6c,
+	0x6f, 0x77, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x41, 0x63, 0x6b, 0x12, 0x19, 0x0a, 0x08, 0x65, 0x76,
+	0x65, 0x6e, 0x74, 0x5f, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x07, 0x65, 0x76,
+	0x65, 0x6e, 0x74, 0x49, 0x64, 0x12, 0x20, 0x0a, 0x0b, 0x69, 0x73, 0x49, 0x6e, 0x69, 0x74, 0x69,
+	0x61, 0x74, 0x6f, 0x72, 0x18, 0x02, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0b, 0x69, 0x73, 0x49, 0x6e,
+	0x69, 0x74, 0x69, 0x61, 0x74, 0x6f, 0x72, 0x22, 0x9c, 0x04, 0x0a, 0x0a, 0x46, 0x6c, 0x6f, 0x77,
+	0x46, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x12, 0x17, 0x0a, 0x07, 0x66, 0x6c, 0x6f, 0x77, 0x5f, 0x69,
+	0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x06, 0x66, 0x6c, 0x6f, 0x77, 0x49, 0x64, 0x12,
+	0x1e, 0x0a, 0x04, 0x74, 0x79, 0x70, 0x65, 0x18, 0x02, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x0a, 0x2e,
+	0x66, 0x6c, 0x6f, 0x77, 0x2e, 0x54, 0x79, 0x70, 0x65, 0x52, 0x04, 0x74, 0x79, 0x70, 0x65, 0x12,
+	0x17, 0x0a, 0x07, 0x72, 0x75, 0x6c, 0x65, 0x5f, 0x69, 0x64, 0x18, 0x03, 0x20, 0x01, 0x28, 0x0c,
+	0x52, 0x06, 0x72, 0x75, 0x6c, 0x65, 0x49, 0x64, 0x12, 0x2d, 0x0a, 0x09, 0x64, 0x69, 0x72, 0x65,
+	0x63, 0x74, 0x69, 0x6f, 0x6e, 0x18, 0x04, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x0f, 0x2e, 0x66, 0x6c,
+	0x6f, 0x77, 0x2e, 0x44, 0x69, 0x72, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x09, 0x64, 0x69,
+	0x72, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x12, 0x1a, 0x0a, 0x08, 0x70, 0x72, 0x6f, 0x74, 0x6f,
+	0x63, 0x6f, 0x6c, 0x18, 0x05, 0x20, 0x01, 0x28, 0x0d, 0x52, 0x08, 0x70, 0x72, 0x6f, 0x74, 0x6f,
+	0x63, 0x6f, 0x6c, 0x12, 0x1b, 0x0a, 0x09, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x69, 0x70,
+	0x18, 0x06, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x08, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x49, 0x70,
+	0x12, 0x17, 0x0a, 0x07, 0x64, 0x65, 0x73, 0x74, 0x5f, 0x69, 0x70, 0x18, 0x07, 0x20, 0x01, 0x28,
+	0x0c, 0x52, 0x06, 0x64, 0x65, 0x73, 0x74, 0x49, 0x70, 0x12, 0x2d, 0x0a, 0x09, 0x70, 0x6f, 0x72,
+	0x74, 0x5f, 0x69, 0x6e, 0x66, 0x6f, 0x18, 0x08, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x0e, 0x2e, 0x66,
+	0x6c, 0x6f, 0x77, 0x2e, 0x50, 0x6f, 0x72, 0x74, 0x49, 0x6e, 0x66, 0x6f, 0x48, 0x00, 0x52, 0x08,
+	0x70, 0x6f, 0x72, 0x74, 0x49, 0x6e, 0x66, 0x6f, 0x12, 0x2d, 0x0a, 0x09, 0x69, 0x63, 0x6d, 0x70,
+	0x5f, 0x69, 0x6e, 0x66, 0x6f, 0x18, 0x09, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x0e, 0x2e, 0x66, 0x6c,
+	0x6f, 0x77, 0x2e, 0x49, 0x43, 0x4d, 0x50, 0x49, 0x6e, 0x66, 0x6f, 0x48, 0x00, 0x52, 0x08, 0x69,
+	0x63, 0x6d, 0x70, 0x49, 0x6e, 0x66, 0x6f, 0x12, 0x1d, 0x0a, 0x0a, 0x72, 0x78, 0x5f, 0x70, 0x61,
+	0x63, 0x6b, 0x65, 0x74, 0x73, 0x18, 0x0a, 0x20, 0x01, 0x28, 0x04, 0x52, 0x09, 0x72, 0x78, 0x50,
+	0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x12, 0x1d, 0x0a, 0x0a, 0x74, 0x78, 0x5f, 0x70, 0x61, 0x63,
+	0x6b, 0x65, 0x74, 0x73, 0x18, 0x0b, 0x20, 0x01, 0x28, 0x04, 0x52, 0x09, 0x74, 0x78, 0x50, 0x61,
+	0x63, 0x6b, 0x65, 0x74, 0x73, 0x12, 0x19, 0x0a, 0x08, 0x72, 0x78, 0x5f, 0x62, 0x79, 0x74, 0x65,
+	0x73, 0x18, 0x0c, 0x20, 0x01, 0x28, 0x04, 0x52, 0x07, 0x72, 0x78, 0x42, 0x79, 0x74, 0x65, 0x73,
+	0x12, 0x19, 0x0a, 0x08, 0x74, 0x78, 0x5f, 0x62, 0x79, 0x74, 0x65, 0x73, 0x18, 0x0d, 0x20, 0x01,
+	0x28, 0x04, 0x52, 0x07, 0x74, 0x78, 0x42, 0x79, 0x74, 0x65, 0x73, 0x12, 0x2c, 0x0a, 0x12, 0x73,
+	0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x69,
+	0x64, 0x18, 0x0e, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x10, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x52,
+	0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x49, 0x64, 0x12, 0x28, 0x0a, 0x10, 0x64, 0x65, 0x73,
+	0x74, 0x5f, 0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x69, 0x64, 0x18, 0x0f, 0x20,
+	0x01, 0x28, 0x0c, 0x52, 0x0e, 0x64, 0x65, 0x73, 0x74, 0x52, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63,
+	0x65, 0x49, 0x64, 0x42, 0x11, 0x0a, 0x0f, 0x63, 0x6f, 0x6e, 0x6e, 0x65, 0x63, 0x74, 0x69, 0x6f,
 	0x6e, 0x5f, 0x69, 0x6e, 0x66, 0x6f, 0x22, 0x48, 0x0a, 0x08, 0x50, 0x6f, 0x72, 0x74, 0x49, 0x6e,
 	0x66, 0x6f, 0x12, 0x1f, 0x0a, 0x0b, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x5f, 0x70, 0x6f, 0x72,
 	0x74, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0d, 0x52, 0x0a, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x50,
@@ -736,19 +683,17 @@ var file_flow_proto_goTypes = []interface{}{
 var file_flow_proto_depIdxs = []int32{
 	7, // 0: flow.FlowEvent.timestamp:type_name -> google.protobuf.Timestamp
 	4, // 1: flow.FlowEvent.flow_fields:type_name -> flow.FlowFields
-	7, // 2: flow.FlowEvent.window_start:type_name -> google.protobuf.Timestamp
-	7, // 3: flow.FlowEvent.window_end:type_name -> google.protobuf.Timestamp
-	0, // 4: flow.FlowFields.type:type_name -> flow.Type
-	1, // 5: flow.FlowFields.direction:type_name -> flow.Direction
-	5, // 6: flow.FlowFields.port_info:type_name -> flow.PortInfo
-	6, // 7: flow.FlowFields.icmp_info:type_name -> flow.ICMPInfo
-	2, // 8: flow.FlowService.Events:input_type -> flow.FlowEvent
-	3, // 9: flow.FlowService.Events:output_type -> flow.FlowEventAck
-	9, // [9:10] is the sub-list for method output_type
-	8, // [8:9] is the sub-list for method input_type
-	8, // [8:8] is the sub-list for extension type_name
-	8, // [8:8] is the sub-list for extension extendee
-	0, // [0:8] is the sub-list for field type_name
+	0, // 2: flow.FlowFields.type:type_name -> flow.Type
+	1, // 3: flow.FlowFields.direction:type_name -> flow.Direction
+	5, // 4: flow.FlowFields.port_info:type_name -> flow.PortInfo
+	6, // 5: flow.FlowFields.icmp_info:type_name -> flow.ICMPInfo
+	2, // 6: flow.FlowService.Events:input_type -> flow.FlowEvent
+	3, // 7: flow.FlowService.Events:output_type -> flow.FlowEventAck
+	7, // [7:8] is the sub-list for method output_type
+	6, // [6:7] is the sub-list for method input_type
+	6, // [6:6] is the sub-list for extension type_name
+	6, // [6:6] is the sub-list for extension extendee
+	0, // [0:6] is the sub-list for field type_name
 }
 
 func init() { file_flow_proto_init() }

flow/proto/flow.proto

@@ -24,9 +24,6 @@ message FlowEvent {
   FlowFields flow_fields = 4;
 
   bool isInitiator = 5;
-
-  google.protobuf.Timestamp window_start = 6;
-  google.protobuf.Timestamp window_end = 7;
 }
 
 message FlowEventAck {
@@ -78,9 +75,6 @@ message FlowFields {
   bytes source_resource_id = 14;
   bytes dest_resource_id = 15;
 
-  uint64 num_of_starts = 16;
-  uint64 num_of_ends = 17;
-  uint64 num_of_drops = 18;
 }
 
 // Flow event types

flow/proto/flow_grpc.pb.go

@@ -1,8 +1,4 @@
 // Code generated by protoc-gen-go-grpc. DO NOT EDIT.
-// versions:
-// - protoc-gen-go-grpc v1.6.1
-// - protoc             v3.21.9
-// source: flow.proto
 
 package proto
 
@@ -15,19 +11,15 @@ import (
 
 // This is a compile-time assertion to ensure that this generated file
 // is compatible with the grpc package it is being compiled against.
-// Requires gRPC-Go v1.64.0 or later.
-const _ = grpc.SupportPackageIsVersion9
-
-const (
-	FlowService_Events_FullMethodName = "/flow.FlowService/Events"
-)
+// Requires gRPC-Go v1.32.0 or later.
+const _ = grpc.SupportPackageIsVersion7
 
 // FlowServiceClient is the client API for FlowService service.
 //
 // For semantics around ctx use and closing/ending streaming RPCs, please refer to https://pkg.go.dev/google.golang.org/grpc/?tab=doc#ClientConn.NewStream.
 type FlowServiceClient interface {
 	// Client to receiver streams of events and acknowledgements
-	Events(ctx context.Context, opts ...grpc.CallOption) (grpc.BidiStreamingClient[FlowEvent, FlowEventAck], error)
+	Events(ctx context.Context, opts ...grpc.CallOption) (FlowService_EventsClient, error)
 }
 
 type flowServiceClient struct {
@@ -38,40 +30,54 @@ func NewFlowServiceClient(cc grpc.ClientConnInterface) FlowServiceClient {
 	return &flowServiceClient{cc}
 }
 
-func (c *flowServiceClient) Events(ctx context.Context, opts ...grpc.CallOption) (grpc.BidiStreamingClient[FlowEvent, FlowEventAck], error) {
-	cOpts := append([]grpc.CallOption{grpc.StaticMethod()}, opts...)
-	stream, err := c.cc.NewStream(ctx, &FlowService_ServiceDesc.Streams[0], FlowService_Events_FullMethodName, cOpts...)
+func (c *flowServiceClient) Events(ctx context.Context, opts ...grpc.CallOption) (FlowService_EventsClient, error) {
+	stream, err := c.cc.NewStream(ctx, &FlowService_ServiceDesc.Streams[0], "/flow.FlowService/Events", opts...)
 	if err != nil {
 		return nil, err
 	}
-	x := &grpc.GenericClientStream[FlowEvent, FlowEventAck]{ClientStream: stream}
+	x := &flowServiceEventsClient{stream}
 	return x, nil
 }
 
-// This type alias is provided for backwards compatibility with existing code that references the prior non-generic stream type by name.
-type FlowService_EventsClient = grpc.BidiStreamingClient[FlowEvent, FlowEventAck]
+type FlowService_EventsClient interface {
+	Send(*FlowEvent) error
+	Recv() (*FlowEventAck, error)
+	grpc.ClientStream
+}
+
+type flowServiceEventsClient struct {
+	grpc.ClientStream
+}
+
+func (x *flowServiceEventsClient) Send(m *FlowEvent) error {
+	return x.ClientStream.SendMsg(m)
+}
+
+func (x *flowServiceEventsClient) Recv() (*FlowEventAck, error) {
+	m := new(FlowEventAck)
+	if err := x.ClientStream.RecvMsg(m); err != nil {
+		return nil, err
+	}
+	return m, nil
+}
 
 // FlowServiceServer is the server API for FlowService service.
 // All implementations must embed UnimplementedFlowServiceServer
-// for forward compatibility.
+// for forward compatibility
 type FlowServiceServer interface {
 	// Client to receiver streams of events and acknowledgements
-	Events(grpc.BidiStreamingServer[FlowEvent, FlowEventAck]) error
+	Events(FlowService_EventsServer) error
 	mustEmbedUnimplementedFlowServiceServer()
 }
 
-// UnimplementedFlowServiceServer must be embedded to have
-// forward compatible implementations.
-//
-// NOTE: this should be embedded by value instead of pointer to avoid a nil
-// pointer dereference when methods are called.
-type UnimplementedFlowServiceServer struct{}
+// UnimplementedFlowServiceServer must be embedded to have forward compatible implementations.
+type UnimplementedFlowServiceServer struct {
+}
 
-func (UnimplementedFlowServiceServer) Events(grpc.BidiStreamingServer[FlowEvent, FlowEventAck]) error {
-	return status.Error(codes.Unimplemented, "method Events not implemented")
+func (UnimplementedFlowServiceServer) Events(FlowService_EventsServer) error {
+	return status.Errorf(codes.Unimplemented, "method Events not implemented")
 }
 func (UnimplementedFlowServiceServer) mustEmbedUnimplementedFlowServiceServer() {}
-func (UnimplementedFlowServiceServer) testEmbeddedByValue()                     {}
 
 // UnsafeFlowServiceServer may be embedded to opt out of forward compatibility for this service.
 // Use of this interface is not recommended, as added methods to FlowServiceServer will
@@ -81,22 +87,34 @@ type UnsafeFlowServiceServer interface {
 }
 
 func RegisterFlowServiceServer(s grpc.ServiceRegistrar, srv FlowServiceServer) {
-	// If the following call panics, it indicates UnimplementedFlowServiceServer was
-	// embedded by pointer and is nil.  This will cause panics if an
-	// unimplemented method is ever invoked, so we test this at initialization
-	// time to prevent it from happening at runtime later due to I/O.
-	if t, ok := srv.(interface{ testEmbeddedByValue() }); ok {
-		t.testEmbeddedByValue()
-	}
 	s.RegisterService(&FlowService_ServiceDesc, srv)
 }
 
 func _FlowService_Events_Handler(srv interface{}, stream grpc.ServerStream) error {
-	return srv.(FlowServiceServer).Events(&grpc.GenericServerStream[FlowEvent, FlowEventAck]{ServerStream: stream})
+	return srv.(FlowServiceServer).Events(&flowServiceEventsServer{stream})
 }
 
-// This type alias is provided for backwards compatibility with existing code that references the prior non-generic stream type by name.
-type FlowService_EventsServer = grpc.BidiStreamingServer[FlowEvent, FlowEventAck]
+type FlowService_EventsServer interface {
+	Send(*FlowEventAck) error
+	Recv() (*FlowEvent, error)
+	grpc.ServerStream
+}
+
+type flowServiceEventsServer struct {
+	grpc.ServerStream
+}
+
+func (x *flowServiceEventsServer) Send(m *FlowEventAck) error {
+	return x.ServerStream.SendMsg(m)
+}
+
+func (x *flowServiceEventsServer) Recv() (*FlowEvent, error) {
+	m := new(FlowEvent)
+	if err := x.ServerStream.RecvMsg(m); err != nil {
+		return nil, err
+	}
+	return m, nil
+}
 
 // FlowService_ServiceDesc is the grpc.ServiceDesc for FlowService service.
 // It's only intended for direct use with grpc.RegisterService,

infrastructure_files/getting-started.sh

@@ -351,11 +351,6 @@ initialize_default_values() {
   NETBIRD_STUN_PORT=3478
 
   # Docker images
-  # Record whether the operator explicitly pinned the server/proxy images via
-  # env vars, so the agent-network preset can pick its own defaults without
-  # clobbering an explicit override.
-  NETBIRD_SERVER_IMAGE_EXPLICIT=${NETBIRD_SERVER_IMAGE:+true}
-  NETBIRD_PROXY_IMAGE_EXPLICIT=${NETBIRD_PROXY_IMAGE:+true}
   DASHBOARD_IMAGE=${DASHBOARD_IMAGE:-"netbirdio/dashboard:latest"}
   # Combined server replaces separate signal, relay, and management containers
   NETBIRD_SERVER_IMAGE=${NETBIRD_SERVER_IMAGE:-"netbirdio/netbird-server:latest"}
@@ -415,15 +410,6 @@ apply_agent_network_preset() {
   ENABLE_PROXY="true"
   ENABLE_CROWDSEC="false"
 
-  # Agent-network ships dedicated server/proxy images. Honor an explicit
-  # env override; otherwise pin the agent-network builds.
-  if [[ "${NETBIRD_SERVER_IMAGE_EXPLICIT}" != "true" ]]; then
-    NETBIRD_SERVER_IMAGE="netbirdio/netbird-server:0.74.0-rc.2"
-  fi
-  if [[ "${NETBIRD_PROXY_IMAGE_EXPLICIT}" != "true" ]]; then
-    NETBIRD_PROXY_IMAGE="netbirdio/reverse-proxy:0.74.0-rc.2"
-  fi
-
   if [[ -n "${NETBIRD_LETSENCRYPT_EMAIL}" ]]; then
     TRAEFIK_ACME_EMAIL="${NETBIRD_LETSENCRYPT_EMAIL}"
   else

management/internals/modules/agentnetwork/catalog/catalog.go

@@ -627,6 +627,21 @@ var providers = []Provider{
 		},
 		Models: []Model{},
 	},
+	{
+		// vLLM is an OpenAI-compatible self-hosted server. It behaves like
+		// the generic custom entry; it gets its own catalog id purely so it
+		// surfaces as a named "vLLM" choice in the provider picker.
+		ID:                 "vllm",
+		Kind:               KindCustom,
+		Name:               "vLLM",
+		Description:        "Self-hosted vLLM (OpenAI-compatible)",
+		DefaultHost:        "",
+		AuthHeaderName:     "Authorization",
+		AuthHeaderTemplate: "Bearer ${API_KEY}",
+		DefaultContentType: "application/json",
+		BrandColor:         "#30A2FF",
+		Models:             []Model{},
+	},
 	{
 		ID:                 "custom",
 		Kind:               KindCustom,

management/internals/modules/agentnetwork/synthesizer.go

@@ -366,6 +366,10 @@ type routerProviderRoute struct {
 	// + refreshes the OAuth token at request time instead of injecting a static
 	// AuthHeaderValue.
 	GCPServiceAccountKeyB64 string `json:"gcp_sa_key_b64,omitempty"`
+	// SkipTLSVerify disables upstream TLS certificate verification when the
+	// proxy dials this provider's upstream. For self-hosted / internal gateways
+	// behind a private or self-signed certificate.
+	SkipTLSVerify bool `json:"skip_tls_verify,omitempty"`
 }
 
 // indexProviderGroups walks the enabled policies and returns, per
@@ -450,6 +454,7 @@ func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]
 			Vertex:                  catalog.IsVertexPathStyle(p.ProviderID),
 			Bedrock:                 catalog.IsBedrockPathStyle(p.ProviderID),
 			GCPServiceAccountKeyB64: gcpSAKeyB64,
+			SkipTLSVerify:           p.SkipTLSVerification,
 		})
 	}
 	out, err := json.Marshal(cfg)

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -1057,6 +1057,41 @@ func TestSynthesizeServices_UpstreamURLPath_FlowsToRouter(t *testing.T) {
 		"upstream path must be carried so the router can disambiguate same-model providers; trailing slash trimmed for stable string-prefix matching")
 }
 
+func TestSynthesizeServices_SkipTLSVerification_FlowsToRouter(t *testing.T) {
+	ctx := context.Background()
+	ctrl := gomock.NewController(t)
+	defer ctrl.Finish()
+	mockStore := store.NewMockStore(ctrl)
+
+	// A provider fronting a self-hosted / internal gateway opts into skipping
+	// upstream TLS verification; the synthesiser must carry it into the router
+	// route so the proxy dials that upstream insecurely.
+	provider := newSynthTestProvider()
+	provider.SkipTLSVerification = true
+	policy := newSynthTestPolicy(provider.ID, "grp-eng", "")
+
+	expectSynthBaseInputs(mockStore, ctx, newSynthTestSettings(),
+		[]*types.Provider{provider},
+		[]*types.Policy{policy},
+		[]*types.Guardrail{})
+
+	services, err := SynthesizeServices(ctx, mockStore, testAccountID)
+	require.NoError(t, err)
+	require.Len(t, services, 1)
+
+	mws := services[0].Targets[0].Options.Middlewares
+	var routerCfg routerConfig
+	for _, m := range mws {
+		if m.ID == middlewareIDLLMRouter {
+			require.NoError(t, json.Unmarshal(m.ConfigJSON, &routerCfg))
+			break
+		}
+	}
+	require.Len(t, routerCfg.Providers, 1)
+	assert.True(t, routerCfg.Providers[0].SkipTLSVerify,
+		"provider skip_tls_verification must flow into the router route")
+}
+
 func TestSynthesizeServices_UnknownProviderID_FailsClosed(t *testing.T) {
 	ctx := context.Background()
 	ctrl := gomock.NewController(t)

management/internals/modules/agentnetwork/types/provider.go

@@ -46,6 +46,11 @@ type Provider struct {
 	// Empty means all catalog models are allowed at catalog prices.
 	Models  []ProviderModel `gorm:"serializer:json"`
 	Enabled bool
+	// SkipTLSVerification disables upstream TLS certificate verification for
+	// this provider's URL. For self-hosted / internal gateways fronted by a
+	// private or self-signed certificate. The synthesiser propagates it into
+	// the router route so the proxy dials that provider's upstream insecurely.
+	SkipTLSVerification bool `gorm:"column:skip_tls_verification"`
 	// SessionPrivateKey + SessionPublicKey are the ed25519 keypair the
 	// synthesised reverse-proxy service uses to sign / verify session
 	// JWTs after a successful OIDC handshake. Generated once on
@@ -129,6 +134,9 @@ func (p *Provider) FromAPIRequest(req *api.AgentNetworkProviderRequest) {
 	if req.Enabled != nil {
 		p.Enabled = *req.Enabled
 	}
+	if req.SkipTlsVerification != nil {
+		p.SkipTLSVerification = *req.SkipTlsVerification
+	}
 	// Identity-header overrides for catalogs flagged Customizable.
 	// nil pointer = "field omitted on the wire" → leave the stored
 	// value untouched (per the openapi description). Empty string is
@@ -155,14 +163,15 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 	created := p.CreatedAt
 	updated := p.UpdatedAt
 	resp := &api.AgentNetworkProvider{
-		Id:          p.ID,
-		ProviderId:  p.ProviderID,
-		Name:        p.Name,
-		UpstreamUrl: p.UpstreamURL,
-		Models:      models,
-		Enabled:     p.Enabled,
-		CreatedAt:   &created,
-		UpdatedAt:   &updated,
+		Id:                  p.ID,
+		ProviderId:          p.ProviderID,
+		Name:                p.Name,
+		UpstreamUrl:         p.UpstreamURL,
+		Models:              models,
+		Enabled:             p.Enabled,
+		SkipTlsVerification: p.SkipTLSVerification,
+		CreatedAt:           &created,
+		UpdatedAt:           &updated,
 	}
 	if len(p.ExtraValues) > 0 {
 		out := make(map[string]string, len(p.ExtraValues))

management/internals/modules/agentnetwork/types/provider_test.go

@@ -0,0 +1,44 @@
+package types
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestProvider_SkipTLSVerification_RoundTrip covers the request→provider→
+// response mapping of skip_tls_verification, including the update semantics
+// (nil pointer preserves, explicit false clears).
+func TestProvider_SkipTLSVerification_RoundTrip(t *testing.T) {
+	enable := true
+	disable := false
+
+	base := func() *api.AgentNetworkProviderRequest {
+		return &api.AgentNetworkProviderRequest{
+			ProviderId:  "openai_api",
+			Name:        "internal",
+			UpstreamUrl: "https://gw.internal",
+		}
+	}
+
+	p := NewProvider("acc-1")
+
+	req := base()
+	req.SkipTlsVerification = &enable
+	p.FromAPIRequest(req)
+	assert.True(t, p.SkipTLSVerification, "create with skip_tls_verification=true must set the field")
+	assert.True(t, p.ToAPIResponse().SkipTlsVerification, "response must surface skip_tls_verification")
+
+	// Omitting the field on update leaves the stored value untouched.
+	p.FromAPIRequest(base())
+	assert.True(t, p.SkipTLSVerification, "omitting skip_tls_verification on update must preserve it")
+
+	// Explicit false clears it.
+	req = base()
+	req.SkipTlsVerification = &disable
+	p.FromAPIRequest(req)
+	assert.False(t, p.SkipTLSVerification, "explicit false must clear skip_tls_verification")
+	assert.False(t, p.ToAPIResponse().SkipTlsVerification, "response must reflect the cleared value")
+}

management/internals/shared/grpc/conversion.go

@@ -47,16 +47,13 @@ func init() {
 	precomputedDeprecatedRemotePeersConstraint = constraint
 }
 
+// toNetbirdConfig converts the server configuration to the wire representation. It returns
+// nil when no server config is set (the fan-out network-map path) because clients treat any
+// non-nil config as authoritative: a config without a relay section is interpreted as relay
+// disabled and wipes the clients' relay URLs.
 func toNetbirdConfig(config *nbconfig.Config, turnCredentials *Token, relayToken *Token, extraSettings *types.ExtraSettings, settings *types.Settings) *proto.NetbirdConfig {
 	if config == nil {
-		if settings == nil {
-			return nil
-		}
-		return &proto.NetbirdConfig{
-			Metrics: &proto.MetricsConfig{
-				Enabled: settings.MetricsPushEnabled,
-			},
-		}
+		return nil
 	}
 
 	var stuns []*proto.HostConfig

management/internals/shared/grpc/conversion_test.go

@@ -8,11 +8,13 @@ import (
 	"time"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map"
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map/controller/cache"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+	"github.com/netbirdio/netbird/management/server/types"
 )
 
 func TestToProtocolDNSConfigWithCache(t *testing.T) {
@@ -263,3 +265,39 @@ func TestEncodeSessionExpiresAt(t *testing.T) {
 		assert.True(t, got.AsTime().Equal(deadline))
 	})
 }
+
+// TestToNetbirdConfig_RelayInvariant guards against the v0.74.0 relay-wipe regression.
+// Clients treat any non-nil NetbirdConfig as authoritative and interpret a missing relay
+// section as relay disabled, wiping their relay URLs. toNetbirdConfig must therefore
+// return nil when no server config is set (the fan-out network-map path) instead of a
+// partial config, and a result built from a relay-enabled config must carry the relay
+// section.
+func TestToNetbirdConfig_RelayInvariant(t *testing.T) {
+	settings := &types.Settings{MetricsPushEnabled: true}
+
+	t.Run("nil server config returns nil config", func(t *testing.T) {
+		nbCfg := toNetbirdConfig(nil, nil, nil, nil, settings)
+		assert.Nil(t, nbCfg, "fan-out updates must not carry a partial NetbirdConfig even when settings are present")
+	})
+
+	t.Run("relay-enabled config carries relay section", func(t *testing.T) {
+		cfg := &nbconfig.Config{
+			Stuns: []*nbconfig.Host{{Proto: nbconfig.UDP, URI: "stun:stun.example.com:3478"}},
+			TURNConfig: &nbconfig.TURNConfig{
+				Turns: []*nbconfig.Host{{Proto: nbconfig.UDP, URI: "turn:turn.example.com:3478", Username: "user", Password: "pass"}},
+			},
+			Relay:  &nbconfig.Relay{Addresses: []string{"rels://relay.example.com:443"}},
+			Signal: &nbconfig.Host{Proto: nbconfig.HTTP, URI: "signal.example.com:10000"},
+		}
+		relayToken := &Token{Payload: "token-payload", Signature: "token-signature"}
+
+		nbCfg := toNetbirdConfig(cfg, nil, relayToken, nil, settings)
+		require.NotNil(t, nbCfg)
+		require.NotNil(t, nbCfg.Relay, "non-nil NetbirdConfig must include the relay section")
+		assert.Equal(t, cfg.Relay.Addresses, nbCfg.Relay.Urls, "relay URLs should match the server config")
+		assert.Equal(t, relayToken.Payload, nbCfg.Relay.TokenPayload, "relay token payload should be set")
+		assert.Equal(t, relayToken.Signature, nbCfg.Relay.TokenSignature, "relay token signature should be set")
+		require.NotNil(t, nbCfg.Metrics)
+		assert.True(t, nbCfg.Metrics.Enabled, "metrics flag should carry the settings value")
+	})
+}

management/server/account.go

@@ -2057,6 +2057,7 @@ func newAccountWithId(ctx context.Context, accountID, userID, domain, email, nam
 			Extra: &types.ExtraSettings{
 				UserApprovalRequired: true,
 			},
+			LazyConnectionEnabled: true,
 		},
 		Onboarding: types.AccountOnboarding{
 			OnboardingFlowPending: true,

management/server/peer_test.go

@@ -1048,11 +1048,7 @@ func testUpdateAccountPeers(t *testing.T) {
 
 			for _, channel := range peerChannels {
 				update := <-channel
-				assert.NotNil(t, update.Update.NetbirdConfig)
-				assert.Nil(t, update.Update.NetbirdConfig.Stuns)
-				assert.Nil(t, update.Update.NetbirdConfig.Turns)
-				assert.Nil(t, update.Update.NetbirdConfig.Signal)
-				assert.Nil(t, update.Update.NetbirdConfig.Relay)
+				assert.Nil(t, update.Update.NetbirdConfig, "fan-out updates must not carry a NetbirdConfig; clients treat a config without relay as relay disabled and wipe their relay URLs")
 				assert.Equal(t, tc.peers, len(update.Update.NetworkMap.RemotePeers))
 				assert.Equal(t, tc.peers*2, len(update.Update.NetworkMap.FirewallRules))
 			}

proxy/internal/middleware/builtin/llm_router/factory.go

@@ -59,6 +59,10 @@ type ProviderRoute struct {
 	// (instead of the static AuthHeaderValue) — so the gateway holds a durable
 	// Vertex credential rather than a 1-hour token.
 	GCPServiceAccountKeyB64 string `json:"gcp_sa_key_b64,omitempty"`
+	// SkipTLSVerify disables upstream TLS certificate verification when dialing
+	// this route's upstream. For self-hosted / internal gateways behind a
+	// private or self-signed certificate.
+	SkipTLSVerify bool `json:"skip_tls_verify,omitempty"`
 }
 
 // Config is the on-wire configuration accepted by the factory. An

proxy/internal/middleware/builtin/llm_router/middleware.go

@@ -615,8 +615,9 @@ func (m *Middleware) allowWithRoute(route ProviderRoute, userGroups []string) *m
 		// path is silently dropped and the gateway returns a 4xx for
 		// the malformed URL. Empty value leaves the original
 		// target's path untouched.
-		Path:         route.UpstreamPath,
-		StripHeaders: append([]string(nil), strippedAuthHeaders...),
+		Path:          route.UpstreamPath,
+		StripHeaders:  append([]string(nil), strippedAuthHeaders...),
+		SkipTLSVerify: route.SkipTLSVerify,
 	}
 	authValue := route.AuthHeaderValue
 	if route.GCPServiceAccountKeyB64 != "" {

proxy/internal/middleware/builtin/llm_router/middleware_test.go

@@ -107,6 +107,41 @@ func TestRouter_HappyPath(t *testing.T) {
 	assert.Equal(t, "allow", dec, "decision metadata must be allow on a match")
 }
 
+func TestRouter_SkipTLSVerifyPropagates(t *testing.T) {
+	base := ProviderRoute{
+		ID:              "internal-gw",
+		Models:          []string{"gpt-4o"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "gateway.internal",
+		AuthHeaderName:  "Authorization",
+		AuthHeaderValue: "Bearer sk-test-123",
+	}
+
+	t.Run("enabled", func(t *testing.T) {
+		route := base
+		route.SkipTLSVerify = true
+		mw := New(Config{Providers: []ProviderRoute{route}})
+
+		out, err := mw.Invoke(context.Background(), newInputWithModel("gpt-4o"))
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations, "matched route must emit mutations")
+		require.NotNil(t, out.Mutations.RewriteUpstream, "matched route must emit upstream rewrite")
+		assert.True(t, out.Mutations.RewriteUpstream.SkipTLSVerify,
+			"skip_tls_verify on the route must ride on the upstream rewrite")
+	})
+
+	t.Run("default off", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{base}})
+
+		out, err := mw.Invoke(context.Background(), newInputWithModel("gpt-4o"))
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations.RewriteUpstream, "matched route must emit upstream rewrite")
+		assert.False(t, out.Mutations.RewriteUpstream.SkipTLSVerify,
+			"skip_tls_verify must default to false when the route does not set it")
+	})
+}
+
 func TestRouter_MissingModel(t *testing.T) {
 	mw := New(Config{Providers: []ProviderRoute{{
 		ID:              "openai-prod",

proxy/internal/middleware/types.go

@@ -243,6 +243,10 @@ type UpstreamRewrite struct {
 	StripPathPrefix string
 	AuthHeader      *AuthHeader
 	StripHeaders    []string
+	// SkipTLSVerify, when true, makes the proxy dial the rewritten upstream
+	// without verifying its TLS certificate. Set by llm_router from the
+	// provider's skip_tls_verification for self-hosted / internal gateways.
+	SkipTLSVerify bool
 }
 
 // AuthHeader is a single name/value pair the proxy injects on the

proxy/internal/proxy/reverseproxy.go

@@ -346,6 +346,11 @@ func (p *ReverseProxy) forwardUpstream(respWriter http.ResponseWriter, r *http.R
 		r.Host = effectiveURL.Host
 		applyUpstreamHeaders(r, upstreamRewrite)
 		stripUpstreamPathPrefix(r, upstreamRewrite.StripPathPrefix)
+		// A router-selected route (e.g. agent-network provider) can opt into
+		// skipping upstream TLS verification per its provider config.
+		if upstreamRewrite.SkipTLSVerify {
+			ctx = roundtrip.WithSkipTLSVerify(ctx)
+		}
 	}
 
 	rp := &httputil.ReverseProxy{

shared/management/client/grpc.go

@@ -33,10 +33,15 @@ const ConnectTimeout = 10 * time.Second
 const healthCheckTimeout = 5 * time.Second
 
 const (
-	// EnvMaxRecvMsgSize overrides the default gRPC max receive message size (4 MB)
+	// EnvMaxRecvMsgSize overrides the default gRPC max receive message size
 	// for the management client connection. Value is in bytes.
 	EnvMaxRecvMsgSize = "NB_MANAGEMENT_GRPC_MAX_MSG_SIZE"
 
+	// defaultMaxRecvMsgSize is the max gRPC receive message size used for the
+	// management client connection when EnvMaxRecvMsgSize is unset or invalid.
+	// It overrides the gRPC library default of 4 MB.
+	defaultMaxRecvMsgSize = 1024 * 1024 * 16
+
 	errMsgMgmtPublicKey    = "failed getting Management Service public key: %s"
 	errMsgNoMgmtConnection = "no connection to management"
 )
@@ -84,22 +89,22 @@ type ExposeResponse struct {
 }
 
 // MaxRecvMsgSize returns the configured max gRPC receive message size from
-// the environment, or 0 if unset (which uses the gRPC default of 4 MB).
+// the environment, or defaultMaxRecvMsgSize (16 MB) if unset or invalid.
 func MaxRecvMsgSize() int {
 	val := os.Getenv(EnvMaxRecvMsgSize)
 	if val == "" {
-		return 0
+		return defaultMaxRecvMsgSize
 	}
 
 	size, err := strconv.Atoi(val)
 	if err != nil {
 		log.Warnf("invalid %s value %q, using default: %v", EnvMaxRecvMsgSize, val, err)
-		return 0
+		return defaultMaxRecvMsgSize
 	}
 
 	if size <= 0 {
 		log.Warnf("invalid %s value %d, must be positive, using default", EnvMaxRecvMsgSize, size)
-		return 0
+		return defaultMaxRecvMsgSize
 	}
 
 	return size
@@ -536,7 +541,7 @@ func (c *GrpcClient) IsHealthy() bool {
 	ctx, cancel := context.WithTimeout(c.ctx, healthCheckTimeout)
 	defer cancel()
 
-	_, err := c.realClient.GetServerKey(ctx, &proto.Empty{})
+	_, err := c.realClient.IsHealthy(ctx, &proto.Empty{})
 	if err != nil {
 		c.notifyDisconnected(err)
 		log.Warnf("health check returned: %s", err)
@@ -1030,8 +1035,6 @@ func infoToMetaData(info *system.Info) *proto.PeerSystemMeta {
 			BlockLANAccess:      info.BlockLANAccess,
 			BlockInbound:        info.BlockInbound,
 			DisableIPv6:         info.DisableIPv6,
-
-			LazyConnectionEnabled: info.LazyConnectionEnabled,
 		},
 
 		Capabilities: peerCapabilities(*info),

shared/management/client/grpc_test.go

@@ -21,11 +21,11 @@ func TestMaxRecvMsgSize(t *testing.T) {
 		envValue string
 		expected int
 	}{
-		{name: "unset returns 0", envValue: "", expected: 0},
+		{name: "unset returns default", envValue: "", expected: defaultMaxRecvMsgSize},
 		{name: "valid value", envValue: "10485760", expected: 10485760},
-		{name: "non-numeric returns 0", envValue: "abc", expected: 0},
-		{name: "negative returns 0", envValue: "-1", expected: 0},
-		{name: "zero returns 0", envValue: "0", expected: 0},
+		{name: "non-numeric returns default", envValue: "abc", expected: defaultMaxRecvMsgSize},
+		{name: "negative returns default", envValue: "-1", expected: defaultMaxRecvMsgSize},
+		{name: "zero returns default", envValue: "0", expected: defaultMaxRecvMsgSize},
 	}
 
 	for _, tt := range tests {

shared/management/http/api/openapi.yml

@@ -2769,28 +2769,6 @@ components:
           type: integer
           description: "Number of packets transmitted."
           example: 5
-        num_of_starts:
-          type: integer
-          description: "Number of start events."
-          example: 3
-        num_of_ends:
-          type: integer
-          description: "Number of end events."
-          example: 4
-        num_of_drops:
-          type: integer
-          description: "Number of drop events."
-          example: 5
-        window_start:
-          type: string
-          format: date-time
-          description: Timestamp of the start of the aggregation window.
-          example: 2025-03-20T16:23:58.125397Z
-        window_end:
-          type: string
-          format: date-time
-          description: Timestamp of the end of the aggregation window.
-          example: 2025-03-20T16:23:58.125397Z
         events:
           type: array
           description: "List of events that are correlated to this flow (e.g., start, end)."
@@ -2812,11 +2790,6 @@ components:
         - rx_packets
         - tx_bytes
         - tx_packets
-        - num_of_starts
-        - num_of_ends
-        - num_of_drops
-        - window_start
-        - window_end
         - events
     NetworkTrafficEventsResponse:
       type: object
@@ -5146,6 +5119,10 @@ components:
           type: boolean
           description: Whether the provider is enabled.
           example: true
+        skip_tls_verification:
+          type: boolean
+          description: Whether upstream TLS certificate verification is skipped when the proxy dials this provider's URL. Intended for self-hosted / internal gateways behind a private or self-signed certificate.
+          example: false
         created_at:
           type: string
           format: date-time
@@ -5165,6 +5142,7 @@ components:
         - upstream_url
         - models
         - enabled
+        - skip_tls_verification
         - created_at
         - updated_at
     AgentNetworkProviderRequest:
@@ -5217,6 +5195,10 @@ components:
           type: boolean
           description: Whether the provider is enabled. Defaults to true on create.
           example: true
+        skip_tls_verification:
+          type: boolean
+          description: Skip upstream TLS certificate verification when the proxy dials this provider's URL. For self-hosted / internal gateways behind a private or self-signed certificate. Defaults to false. When omitted on update, the stored value is left unchanged.
+          example: false
       required:
         - provider_id
         - name

shared/management/http/api/types.gen.go

@@ -2224,6 +2224,9 @@ type AgentNetworkProvider struct {
 	// ProviderId Catalog identifier for the upstream AI provider (e.g. openai_api, anthropic_api, azure_openai_api, bedrock_api, vertex_ai_api, mistral_api, custom).
 	ProviderId string `json:"provider_id"`
 
+	// SkipTlsVerification Whether upstream TLS certificate verification is skipped when the proxy dials this provider's URL. Intended for self-hosted / internal gateways behind a private or self-signed certificate.
+	SkipTlsVerification bool `json:"skip_tls_verification"`
+
 	// UpdatedAt Timestamp when the provider was last updated.
 	UpdatedAt *time.Time `json:"updated_at,omitempty"`
 
@@ -2272,6 +2275,9 @@ type AgentNetworkProviderRequest struct {
 	// ProviderId Catalog identifier for the upstream AI provider (e.g. openai_api, anthropic_api, azure_openai_api, bedrock_api, vertex_ai_api, mistral_api, custom).
 	ProviderId string `json:"provider_id"`
 
+	// SkipTlsVerification Skip upstream TLS certificate verification when the proxy dials this provider's URL. For self-hosted / internal gateways behind a private or self-signed certificate. Defaults to false. When omitted on update, the stored value is left unchanged.
+	SkipTlsVerification *bool `json:"skip_tls_verification,omitempty"`
+
 	// UpstreamUrl Full upstream URL (with scheme) that NetBird forwards traffic to.
 	UpstreamUrl string `json:"upstream_url"`
 }
@@ -3703,18 +3709,9 @@ type NetworkTrafficEvent struct {
 	Events []NetworkTrafficSubEvent `json:"events"`
 
 	// FlowId FlowID is the ID of the connection flow. Not unique because it can be the same for multiple events (e.g., start and end of the connection).
-	FlowId string             `json:"flow_id"`
-	Icmp   NetworkTrafficICMP `json:"icmp"`
-
-	// NumOfDrops Number of drop events.
-	NumOfDrops int `json:"num_of_drops"`
-
-	// NumOfEnds Number of end events.
-	NumOfEnds int `json:"num_of_ends"`
-
-	// NumOfStarts Number of start events.
-	NumOfStarts int                  `json:"num_of_starts"`
-	Policy      NetworkTrafficPolicy `json:"policy"`
+	FlowId string               `json:"flow_id"`
+	Icmp   NetworkTrafficICMP   `json:"icmp"`
+	Policy NetworkTrafficPolicy `json:"policy"`
 
 	// Protocol Protocol is the protocol of the traffic (e.g. 1 = ICMP, 6 = TCP, 17 = UDP, etc.).
 	Protocol int `json:"protocol"`
@@ -3735,12 +3732,6 @@ type NetworkTrafficEvent struct {
 	// TxPackets Number of packets transmitted.
 	TxPackets int                `json:"tx_packets"`
 	User      NetworkTrafficUser `json:"user"`
-
-	// WindowEnd Timestamp of the end of the aggregation window.
-	WindowEnd time.Time `json:"window_end"`
-
-	// WindowStart Timestamp of the start of the aggregation window.
-	WindowStart time.Time `json:"window_start"`
 }
 
 // NetworkTrafficEventsResponse defines model for NetworkTrafficEventsResponse.