revert component types

## Describe your changes

Restore the management gRPC client's backoff reset after a stream
connects.

This is a regression: PR #4935 originally added this reset; it was then
lost in
commit `58daa674e` during the `withMgmtStream` refactor (Sync/Job
unification).

Compared to before the reset was done AFTER receiveEvents, whereas 
now it's done before. 

Now should cover for Sync and Job (which didn't exist in #4935. Hold for
long living stream that breaks.

Reset prevents two things

1. long living conns going wrong from waiting a long backoff time window
before to drive reconn
2. past MaxElapsedTime (3months) long lived conns failures from being
treated as "unrecoverable" (hence triggering an full restart of the
engine)

## Issue ticket number and link

Internal support case (customer agents lost data-plane connectivity
during a
management maintenance/release window). No public issue. Regression
introduced
in commit `58daa674e`, which removed the `backOff.Reset()` added by PR
#4935.

## Stack

<!-- branch-stack -->

### Checklist
- [x] Is it a bug fix
- [ ] Is a typo/documentation fix
- [ ] Is a feature enhancement
- [ ] It is a refactor
- [ ] Created tests that fail without the change (if possible)

> By submitting this pull request, you confirm that you have read and
agree to the terms of the [Contributor License
Agreement](https://github.com/netbirdio/netbird/blob/main/CONTRIBUTOR_LICENSE_AGREEMENT.md).

## Documentation
Select exactly one:

- [ ] I added/updated documentation for this change
- [x] Documentation is **not needed** for this change (explain why)

Internal reconnection-behavior fix in the management gRPC client. No
public
NetBird CLI, API, or configuration surface changes.

### Docs PR URL (required if "docs added" is checked)
Paste the PR link from https://github.com/netbirdio/docs here:

N/A

<!-- codesmith:footer -->
---
<a
href="https://app.blacksmith.sh/netbirdio/codesmith/netbird/pr/6883"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/view-with-codesmith-dark-v2.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/view-with-codesmith-light-v2.svg"><img
alt="View with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/view-with-codesmith-dark-v2.svg"></picture></a>
<a
href="https://backend.blacksmith.sh/track/enable-autofix?expires=1787484622&installation_model_id=427504&pr_number=6883&repository=netbirdio%2Fnetbird&return_to=https%3A%2F%2Fgithub.com%2Fnetbirdio%2Fnetbird%2Fpull%2F6883&signature=525b08edbcb94ee6f4ae4226fb1fe8ddb829caa7d4489e8efa639770af9dccdf"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-light.svg"><img
alt="Autofix with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"></picture></a>
<sup>Need help on this PR? Tag <code>@codesmith-bot</code> with what you
need. Autofix is disabled.</sup>

<!-- codesmith:autofix:disabled -->
<!-- /codesmith:footer -->

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **Bug Fixes**
  * Improved management stream retry behavior.
* Retry delays now reset after a stream is successfully established,
helping subsequent connection attempts recover more quickly.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

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

@@ -51,6 +51,9 @@ jobs:
           # token (and URL, for gateways) is unset, so partial coverage is fine.
           OPENAI_TOKEN: ${{ secrets.E2E_OPENAI_TOKEN }}
           ANTHROPIC_TOKEN: ${{ secrets.E2E_ANTHROPIC_TOKEN }}
+          # Moonshot AI platform key (platform.kimi.ai); drives both Kimi wire
+          # shapes (OpenAI /v1 and Anthropic /anthropic) through kimi_api.
+          KIMI_TOKEN: ${{ secrets.E2E_KIMI_TOKEN }}
           VERCEL_URL: ${{ secrets.E2E_VERCEL_URL }}
           VERCEL_TOKEN: ${{ secrets.E2E_VERCEL_TOKEN }}
           OPENROUTER_URL: ${{ secrets.E2E_OPENROUTER_URL }}

.github/workflows/frontend-ui.yml

@@ -86,7 +86,7 @@ jobs:
             ${{ runner.os }}-pnpm-
 
       - name: Install dependencies
-        run: pnpm install --frozen-lockfile
+        run: pnpm install --frozen-lockfile --ignore-scripts
 
       - name: Generate Wails bindings
         run: pnpm run bindings

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

@@ -166,7 +166,7 @@ jobs:
         # resolve; the grep then drops the broken package by path. Without -e,
         # go list aborts with empty stdout and `go test` falls back to the repo
         # root, which has no Go files.
-        run: CGO_ENABLED=1 GOARCH=${{ matrix.arch }} CI=true go test -coverprofile=coverage.txt -tags devcert -timeout 10m -p 1 $(go list -e ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined -e /client/ui)
+        run: CGO_ENABLED=1 GOARCH=${{ 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 -e ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined -e /client/ui -e /client/testutil/privileged)
 
       - name: Upload coverage reports to Codecov
         if: matrix.arch == 'amd64'

.github/workflows/golangci-lint.yml

@@ -45,7 +45,7 @@ jobs:
             display_name: Linux
     name: ${{ matrix.display_name }}
     runs-on: ${{ matrix.os }}
-    timeout-minutes: 15
+    timeout-minutes: 25
     steps:
       - name: Checkout code
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
@@ -79,4 +79,4 @@ jobs:
           skip-cache: true
           skip-save-cache: true
           cache-invalidation-interval: 0
-          args: --timeout=12m
+          args: --timeout=20m

.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

agent-network/README.md

@@ -1,16 +1,47 @@
 # NetBird Agent Network
 
-Agent Network is NetBird's access control layer for AI agents and the people who run
-them. It gives every agent a real identity, tied to your identity provider (IdP), and
-governs what it can reach — the LLM APIs and AI gateways it can call, and the internal
-resources it can access. Traffic flows only over the encrypted NetBird tunnel, scoped by
-policy, with no API keys to leak.
+Agent Network is NetBird's access control layer for AI agents and the people who run them. 
+It gives every agent a real identity, tied to an identity provider (IdP), and governs what it can reach: LLM APIs and 
+AI gateways it can call, and the internal resources it can access. Traffic flows only over the encrypted NetBird tunnel, 
+scoped by policy, with no API keys or other credentials to leak. It also gives you control over cost and token usage.
 
-> **Beta.** Agent Network is open source and can be self-hosted on your own
-> infrastructure.
+Because every LLM request passes through an
+identity-aware proxy, you can:
+
+- **Set spending and rate limits** per agent, per user, or per team — with hard caps
+  that stop requests once a budget is reached.
+- **Restrict models and providers** so agents can only call approved (and cost-appropriate)
+  endpoints, keeping expensive models off-limits unless explicitly allowed.
+- **Attribute usage** by tracking token consumption and cost per identity, group, or cost center so every
+  request is tied back to the agent and person responsible.
+- **Reuse your existing AI gateway** — point the proxy at a gateway you already run,
+  keeping its routing and config in place while it adds identity on top, so you skip
+  API key distribution.
+
+https://github.com/user-attachments/assets/44d18286-d8ab-49f8-a457-98ccd66f3268
+
+> **Beta.** Agent Network is in beta, but it's stable and already running in
+> production environments. It's fully open source and can be self-hosted on your own
+> infrastructure, with no vendor lock-in and no data leaving your environment.
 
 ## How it works
 
+Say you have a simple use case: your Engineering or IT team needs access to Claude Code or Codex, and you want visibility into usage plus the ability to enforce budgets.
+How can you do that without creating a dedicated API key for every team?
+
+With Agent Network you get a private endpoint inside your network, for example: https://mirror.netbird.ai
+Teams configure their agents to point to that endpoint instead of using individual API keys directly. 
+
+This endpoint is only reachable when users are connected to your NetBird network and authenticated through your IdP. Otherwise, it is not accessible from the public internet.
+You can then use this private endpoint to configure your AI agents, whether that is Claude Code, Codex, or another tool.
+
+## Quickstart
+
+Full step-by-step setup:
+**https://docs.netbird.io/agent-network/quickstart**
+
+## Architecture 
+
 Agent Network is built on two existing NetBird capabilities:
 
 - **Overlay network** — the encrypted WireGuard mesh between peers.
@@ -22,6 +53,9 @@ LLM traffic is routed through the proxy's identity-aware pipeline, while interna
 resources (databases, internal APIs, self-hosted models) are reached directly over
 peer-to-peer WireGuard tunnels, governed by the same identities and access policies.
 
+<img width="4720" height="2218" alt="image" src="https://github.com/user-attachments/assets/1afa5da1-4b82-4f8a-a7a8-f417efadf1eb" />
+
+
 ## Where the code lives
 
 There is no separate "agent-network" service — it reuses the reverse-proxy and management

client/android/client.go

@@ -247,6 +247,9 @@ func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool) (strin
 		deps.SyncResponse = resp
 
 		if e := cc.Engine(); e != nil {
+			deps.RefreshStatus = func() {
+				e.RunHealthProbes(context.Background(), true)
+			}
 			if cm := e.GetClientMetrics(); cm != nil {
 				deps.ClientMetrics = cm
 			}

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/login.go

@@ -17,7 +17,9 @@ import (
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/auth"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	nbnet "github.com/netbirdio/netbird/client/net"
 	"github.com/netbirdio/netbird/client/proto"
+	"github.com/netbirdio/netbird/client/server"
 	"github.com/netbirdio/netbird/client/system"
 	"github.com/netbirdio/netbird/util"
 )
@@ -331,6 +333,14 @@ func doForegroundLogin(ctx context.Context, cmd *cobra.Command, setupKey string,
 		return fmt.Errorf("read config file %s: %v", configFilePath, err)
 	}
 
+	// Mirror runInForegroundMode: recover residual state (DNS, firewall,
+	// ssh config, legacy routing) from a previous unclean shutdown and
+	// enable advanced routing before dialing management.
+	if err := server.RestoreResidualState(ctx, profilemanager.NewServiceManager(configFilePath).GetStatePath()); err != nil {
+		log.Warnf("failed to restore residual state: %v", err)
+	}
+	nbnet.Init()
+
 	err = foregroundLogin(ctx, cmd, config, setupKey, activeProf.ID)
 	if err != nil {
 		return fmt.Errorf("foreground login failed: %v", err)

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/service.go

@@ -5,6 +5,7 @@ package cmd
 import (
 	"context"
 	"fmt"
+	"net/http"
 	"runtime"
 	"strings"
 	"sync"
@@ -22,15 +23,21 @@ var serviceCmd = &cobra.Command{
 	Short: "Manage the NetBird daemon service",
 }
 
+const defaultJSONSocket = "unix:///var/run/netbird-http.sock"
+
 var (
-	serviceName    string
-	serviceEnvVars []string
+	serviceName      string
+	serviceEnvVars   []string
+	jsonSocket       string
+	enableJSONSocket bool
 )
 
 type program struct {
 	ctx              context.Context
 	cancel           context.CancelFunc
 	serv             *grpc.Server
+	jsonServ         *http.Server
+	jsonServMu       sync.Mutex
 	serverInstance   *server.Server
 	serverInstanceMu sync.Mutex
 }
@@ -46,6 +53,8 @@ func init() {
 	serviceCmd.PersistentFlags().BoolVar(&updateSettingsDisabled, "disable-update-settings", false, "Disables update settings feature. If enabled, the client will not be able to change or edit any settings. To persist this setting, use: netbird service install --disable-update-settings")
 	serviceCmd.PersistentFlags().BoolVar(&captureEnabled, "enable-capture", false, "Enables packet capture via 'netbird debug capture'. To persist, use: netbird service install --enable-capture")
 	serviceCmd.PersistentFlags().BoolVar(&networksDisabled, "disable-networks", false, "Disables network selection. If enabled, the client will not allow listing, selecting, or deselecting networks. To persist, use: netbird service install --disable-networks")
+	serviceCmd.PersistentFlags().BoolVar(&enableJSONSocket, "enable-json-socket", false, "Enables the HTTP/JSON API socket served by grpc-gateway. To persist, use: netbird service install --enable-json-socket")
+	serviceCmd.PersistentFlags().StringVar(&jsonSocket, "json-socket", defaultJSONSocket, "HTTP/JSON API socket address [unix|tcp]://[path|host:port]. Requires --enable-json-socket to serve. To persist, use: netbird service install --enable-json-socket --json-socket")
 
 	rootCmd.PersistentFlags().StringVarP(&serviceName, "service", "s", defaultServiceName, "Netbird system service name")
 	serviceEnvDesc := `Sets extra environment variables for the service. ` +

client/cmd/service_controller.go

@@ -5,9 +5,6 @@ package cmd
 import (
 	"context"
 	"fmt"
-	"net"
-	"os"
-	"strings"
 	"time"
 
 	"github.com/kardianos/service"
@@ -22,41 +19,56 @@ import (
 	"github.com/netbirdio/netbird/util"
 )
 
+func validateJSONSocketFlags() error {
+	if serviceCmd.PersistentFlags().Changed("json-socket") && !enableJSONSocket {
+		return fmt.Errorf("--json-socket requires --enable-json-socket to configure the daemon JSON gateway")
+	}
+	return nil
+}
+
 func (p *program) Start(svc service.Service) error {
 	// Start should not block. Do the actual work async.
 	log.Info("starting NetBird service") //nolint
 
+	if err := validateJSONSocketFlags(); err != nil {
+		return err
+	}
+
 	// Collect static system and platform information
 	system.UpdateStaticInfoAsync()
 
 	// in any case, even if configuration does not exists we run daemon to serve CLI gRPC API.
 	p.serv = grpc.NewServer()
 
-	split := strings.Split(daemonAddr, "://")
-	switch split[0] {
-	case "unix":
-		// cleanup failed close
-		stat, err := os.Stat(split[1])
-		if err == nil && !stat.IsDir() {
-			if err := os.Remove(split[1]); err != nil {
-				log.Debugf("remove socket file: %v", err)
-			}
-		}
-	case "tcp":
-	default:
-		return fmt.Errorf("unsupported daemon address protocol: %v", split[0])
-	}
-
-	listen, err := net.Listen(split[0], split[1])
+	daemonListener, err := listenOnAddress(daemonAddr)
 	if err != nil {
 		return fmt.Errorf("listen daemon interface: %w", err)
 	}
-	go func() {
-		defer listen.Close()
 
-		if split[0] == "unix" {
-			if err := os.Chmod(split[1], 0666); err != nil {
-				log.Errorf("failed setting daemon permissions: %v", split[1])
+	var jsonListener *socketListener
+	if enableJSONSocket {
+		jsonListener, err = listenOnAddress(jsonSocket)
+		if err != nil {
+			_ = daemonListener.Close()
+			return fmt.Errorf("listen daemon JSON interface: %w", err)
+		}
+	} else {
+		removeStaleUnixSocketForAddress(jsonSocket)
+	}
+
+	go func() {
+		defer daemonListener.Close()
+		if jsonListener != nil {
+			defer jsonListener.Close()
+		}
+
+		if err := daemonListener.chmodUnixSocket("daemon"); err != nil {
+			log.Error(err)
+			return
+		}
+		if jsonListener != nil {
+			if err := jsonListener.chmodUnixSocket("daemon JSON"); err != nil {
+				log.Error(err)
 				return
 			}
 		}
@@ -71,8 +83,16 @@ func (p *program) Start(svc service.Service) error {
 		p.serverInstance = serverInstance
 		p.serverInstanceMu.Unlock()
 
-		log.Printf("started daemon server: %v", split[1])
-		if err := p.serv.Serve(listen); err != nil {
+		if jsonListener != nil {
+			if err := p.startJSONGateway(jsonListener, daemonAddr); err != nil {
+				log.Fatalf("failed to start daemon JSON server: %v", err)
+			}
+		} else {
+			log.Debug("daemon JSON socket disabled")
+		}
+
+		log.Printf("started daemon server: %v", daemonListener.address)
+		if err := p.serv.Serve(daemonListener.Listener); err != nil {
 			log.Errorf("failed to serve daemon requests: %v", err)
 		}
 	}()
@@ -92,6 +112,20 @@ func (p *program) Stop(srv service.Service) error {
 
 	p.cancel()
 
+	p.jsonServMu.Lock()
+	jsonServ := p.jsonServ
+	p.jsonServMu.Unlock()
+	if jsonServ != nil {
+		shutdownCtx, shutdownCancel := context.WithTimeout(context.Background(), 2*time.Second)
+		if err := jsonServ.Shutdown(shutdownCtx); err != nil {
+			log.Errorf("failed to stop daemon JSON server gracefully: %v", err)
+			if err := jsonServ.Close(); err != nil {
+				log.Errorf("failed to close daemon JSON server: %v", err)
+			}
+		}
+		shutdownCancel()
+	}
+
 	if p.serv != nil {
 		p.serv.Stop()
 	}
@@ -148,6 +182,9 @@ var runCmd = &cobra.Command{
 		if err != nil {
 			return err
 		}
+		if err := validateJSONSocketFlags(); err != nil {
+			return err
+		}
 
 		return s.Run()
 	},
@@ -162,6 +199,9 @@ var startCmd = &cobra.Command{
 		if err != nil {
 			return err
 		}
+		if err := validateJSONSocketFlags(); err != nil {
+			return err
+		}
 
 		if err := s.Start(); err != nil {
 			return fmt.Errorf("start service: %w", err)
@@ -198,6 +238,9 @@ var restartCmd = &cobra.Command{
 		if err != nil {
 			return err
 		}
+		if err := validateJSONSocketFlags(); err != nil {
+			return err
+		}
 
 		if err := s.Restart(); err != nil {
 			return fmt.Errorf("restart service: %w", err)

client/cmd/service_installer.go

@@ -67,6 +67,10 @@ func buildServiceArguments() []string {
 		args = append(args, "--disable-networks")
 	}
 
+	if enableJSONSocket {
+		args = append(args, "--enable-json-socket", "--json-socket", jsonSocket)
+	}
+
 	return args
 }
 
@@ -106,6 +110,10 @@ func configurePlatformSpecificSettings(svcConfig *service.Config) error {
 
 // Create fully configured service config for install/reconfigure
 func createServiceConfigForInstall() (*service.Config, error) {
+	if err := validateJSONSocketFlags(); err != nil {
+		return nil, err
+	}
+
 	svcConfig, err := newSVCConfig()
 	if err != nil {
 		return nil, fmt.Errorf("create service config: %w", err)

client/cmd/service_json_gateway.go

@@ -0,0 +1,52 @@
+//go:build !ios && !android
+
+package cmd
+
+import (
+	"context"
+	"errors"
+	"net"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/grpc-ecosystem/grpc-gateway/v2/runtime"
+	log "github.com/sirupsen/logrus"
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/credentials/insecure"
+
+	"github.com/netbirdio/netbird/client/proto"
+)
+
+func grpcGatewayEndpoint(addr string) string {
+	return strings.TrimPrefix(addr, "tcp://")
+}
+
+func (p *program) startJSONGateway(jsonListener *socketListener, daemonEndpoint string) error {
+	mux := runtime.NewServeMux()
+	opts := []grpc.DialOption{grpc.WithTransportCredentials(insecure.NewCredentials())}
+	if err := proto.RegisterDaemonServiceHandlerFromEndpoint(p.ctx, mux, grpcGatewayEndpoint(daemonEndpoint), opts); err != nil {
+		return err
+	}
+
+	jsonServer := &http.Server{
+		Handler:           mux,
+		ReadHeaderTimeout: 5 * time.Second,
+		BaseContext: func(net.Listener) context.Context {
+			return p.ctx
+		},
+	}
+
+	p.jsonServMu.Lock()
+	p.jsonServ = jsonServer
+	p.jsonServMu.Unlock()
+
+	go func() {
+		log.Printf("started daemon JSON server: %v", jsonListener.address)
+		if err := jsonServer.Serve(jsonListener.Listener); err != nil && !errors.Is(err, http.ErrServerClosed) {
+			log.Errorf("failed to serve daemon JSON requests: %v", err)
+		}
+	}()
+
+	return nil
+}

client/cmd/service_json_socket_test.go

@@ -0,0 +1,176 @@
+//go:build !ios && !android
+
+package cmd
+
+import (
+	"net"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+
+	"github.com/spf13/cobra"
+	"github.com/spf13/pflag"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func preserveJSONSocketTestState(t *testing.T) {
+	t.Helper()
+
+	origJSONSocket := jsonSocket
+	origEnableJSONSocket := enableJSONSocket
+	origChanged := map[string]bool{}
+	serviceCmd.PersistentFlags().VisitAll(func(flag *pflag.Flag) {
+		origChanged[flag.Name] = flag.Changed
+	})
+
+	t.Cleanup(func() {
+		jsonSocket = origJSONSocket
+		enableJSONSocket = origEnableJSONSocket
+		serviceCmd.PersistentFlags().VisitAll(func(flag *pflag.Flag) {
+			flag.Changed = origChanged[flag.Name]
+		})
+	})
+}
+
+func TestJSONSocketFlagsArePositiveEnableOnly(t *testing.T) {
+	assert.NotNil(t, serviceCmd.PersistentFlags().Lookup("enable-json-socket"))
+	assert.NotNil(t, serviceCmd.PersistentFlags().Lookup("json-socket"))
+	assert.Nil(t, serviceCmd.PersistentFlags().Lookup("disable-json-socket"))
+	assert.Equal(t, "false", serviceCmd.PersistentFlags().Lookup("enable-json-socket").DefValue)
+}
+
+func TestBuildServiceArgumentsDefaultDisablesJSONSocket(t *testing.T) {
+	preserveJSONSocketTestState(t)
+
+	enableJSONSocket = false
+	jsonSocket = "tcp://127.0.0.1:8080"
+
+	args := buildServiceArguments()
+
+	assert.NotContains(t, args, "--enable-json-socket")
+	assert.NotContains(t, args, "--json-socket")
+}
+
+func TestBuildServiceArgumentsIncludesJSONSocketWhenEnabled(t *testing.T) {
+	preserveJSONSocketTestState(t)
+
+	enableJSONSocket = true
+	jsonSocket = "tcp://127.0.0.1:8080"
+
+	args := buildServiceArguments()
+
+	enableIndex := indexOfArg(args, "--enable-json-socket")
+	jsonIndex := indexOfArg(args, "--json-socket")
+	require.NotEqual(t, -1, enableIndex)
+	require.NotEqual(t, -1, jsonIndex)
+	require.Less(t, enableIndex, jsonIndex)
+	require.Less(t, jsonIndex+1, len(args))
+	assert.Equal(t, "tcp://127.0.0.1:8080", args[jsonIndex+1])
+}
+
+func TestJSONSocketWithoutEnableValidation(t *testing.T) {
+	preserveJSONSocketTestState(t)
+
+	enableJSONSocket = false
+	require.NoError(t, serviceCmd.PersistentFlags().Set("json-socket", "tcp://127.0.0.1:8080"))
+
+	err := validateJSONSocketFlags()
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "--enable-json-socket")
+}
+
+func TestJSONSocketWithEnableValidation(t *testing.T) {
+	preserveJSONSocketTestState(t)
+
+	require.NoError(t, serviceCmd.PersistentFlags().Set("enable-json-socket", "true"))
+	require.NoError(t, serviceCmd.PersistentFlags().Set("json-socket", "tcp://127.0.0.1:8080"))
+
+	assert.NoError(t, validateJSONSocketFlags())
+}
+
+func TestJSONSocketServiceParamsPersistEnableAndAddress(t *testing.T) {
+	preserveJSONSocketTestState(t)
+	serviceCmd.PersistentFlags().VisitAll(func(flag *pflag.Flag) {
+		flag.Changed = false
+	})
+
+	enableJSONSocket = true
+	jsonSocket = "tcp://127.0.0.1:8080"
+
+	params := currentServiceParams()
+	require.True(t, params.EnableJSONSocket)
+	require.Equal(t, "tcp://127.0.0.1:8080", params.JSONSocket)
+
+	enableJSONSocket = false
+	jsonSocket = defaultJSONSocket
+	applyServiceParams(testServiceEnvCommand(), params)
+
+	assert.True(t, enableJSONSocket)
+	assert.Equal(t, "tcp://127.0.0.1:8080", jsonSocket)
+}
+
+func TestRemoveStaleUnixSocketDoesNotRemoveRegularFile(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "netbird-http.sock")
+	require.NoError(t, os.WriteFile(path, []byte("not a socket"), 0600))
+
+	removeStaleUnixSocket(path)
+
+	data, err := os.ReadFile(path)
+	require.NoError(t, err)
+	assert.Equal(t, []byte("not a socket"), data)
+}
+
+func TestRemoveStaleUnixSocketRemovesSocket(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("unix sockets are not available on Windows")
+	}
+
+	path := filepath.Join(t.TempDir(), "netbird-http.sock")
+	addr := &net.UnixAddr{Name: path, Net: "unix"}
+	listener, err := net.ListenUnix("unix", addr)
+	require.NoError(t, err)
+	listener.SetUnlinkOnClose(false)
+	require.NoError(t, listener.Close())
+
+	_, err = os.Lstat(path)
+	require.NoError(t, err, "test setup must leave a stale Unix socket path")
+
+	removeStaleUnixSocket(path)
+
+	_, err = os.Lstat(path)
+	assert.True(t, os.IsNotExist(err), "expected stale Unix socket to be removed, got %v", err)
+}
+
+func TestRemoveStaleUnixSocketDoesNotRemoveLiveSocket(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("unix sockets are not available on Windows")
+	}
+
+	path := filepath.Join(t.TempDir(), "netbird-http.sock")
+	listener, err := net.Listen("unix", path)
+	require.NoError(t, err)
+	defer listener.Close()
+
+	removeStaleUnixSocket(path)
+
+	_, err = os.Lstat(path)
+	assert.NoError(t, err, "expected live Unix socket to be preserved")
+}
+
+func testServiceEnvCommand() *cobra.Command {
+	cmd := &cobra.Command{}
+	cmd.Flags().StringSlice("service-env", nil, "")
+	return cmd
+}
+
+func indexOfArg(args []string, arg string) int {
+	for i, candidate := range args {
+		if candidate == arg {
+			return i
+		}
+	}
+	return -1
+}

client/cmd/service_params.go

@@ -23,6 +23,7 @@ const serviceParamsFile = "service.json"
 type serviceParams struct {
 	LogLevel              string            `json:"log_level"`
 	DaemonAddr            string            `json:"daemon_addr"`
+	JSONSocket            string            `json:"json_socket"`
 	ManagementURL         string            `json:"management_url,omitempty"`
 	ConfigPath            string            `json:"config_path,omitempty"`
 	LogFiles              []string          `json:"log_files,omitempty"`
@@ -30,6 +31,7 @@ type serviceParams struct {
 	DisableUpdateSettings bool              `json:"disable_update_settings,omitempty"`
 	EnableCapture         bool              `json:"enable_capture,omitempty"`
 	DisableNetworks       bool              `json:"disable_networks,omitempty"`
+	EnableJSONSocket      bool              `json:"enable_json_socket,omitempty"`
 	ServiceEnvVars        map[string]string `json:"service_env_vars,omitempty"`
 }
 
@@ -75,6 +77,7 @@ func currentServiceParams() *serviceParams {
 	params := &serviceParams{
 		LogLevel:              logLevel,
 		DaemonAddr:            daemonAddr,
+		JSONSocket:            jsonSocket,
 		ManagementURL:         managementURL,
 		ConfigPath:            configPath,
 		LogFiles:              logFiles,
@@ -82,6 +85,7 @@ func currentServiceParams() *serviceParams {
 		DisableUpdateSettings: updateSettingsDisabled,
 		EnableCapture:         captureEnabled,
 		DisableNetworks:       networksDisabled,
+		EnableJSONSocket:      enableJSONSocket,
 	}
 
 	if len(serviceEnvVars) > 0 {
@@ -113,9 +117,8 @@ func applyServiceParams(cmd *cobra.Command, params *serviceParams) {
 		return
 	}
 
-	// For fields with non-empty defaults (log-level, daemon-addr), keep the
-	// != "" guard so that an older service.json missing the field doesn't
-	// clobber the default with an empty string.
+	// For fields with non-empty defaults, keep the != "" guard so that an older
+	// service.json missing the field doesn't clobber the default with an empty string.
 	if !rootCmd.PersistentFlags().Changed("log-level") && params.LogLevel != "" {
 		logLevel = params.LogLevel
 	}
@@ -124,6 +127,14 @@ func applyServiceParams(cmd *cobra.Command, params *serviceParams) {
 		daemonAddr = params.DaemonAddr
 	}
 
+	if !serviceCmd.PersistentFlags().Changed("json-socket") && params.JSONSocket != "" {
+		jsonSocket = params.JSONSocket
+	}
+
+	if !serviceCmd.PersistentFlags().Changed("enable-json-socket") {
+		enableJSONSocket = params.EnableJSONSocket
+	}
+
 	// For optional fields where empty means "use default", always apply so
 	// that an explicit clear (--management-url "") persists across reinstalls.
 	if !rootCmd.PersistentFlags().Changed("management-url") {

client/cmd/service_params_test.go

@@ -41,6 +41,8 @@ func TestSaveAndLoadServiceParams(t *testing.T) {
 	params := &serviceParams{
 		LogLevel:              "debug",
 		DaemonAddr:            "unix:///var/run/netbird.sock",
+		JSONSocket:            "tcp://127.0.0.1:8080",
+		EnableJSONSocket:      true,
 		ManagementURL:         "https://my.server.com",
 		ConfigPath:            "/etc/netbird/config.json",
 		LogFiles:              []string{"/var/log/netbird/client.log", "console"},
@@ -63,6 +65,8 @@ func TestSaveAndLoadServiceParams(t *testing.T) {
 
 	assert.Equal(t, params.LogLevel, loaded.LogLevel)
 	assert.Equal(t, params.DaemonAddr, loaded.DaemonAddr)
+	assert.Equal(t, params.JSONSocket, loaded.JSONSocket)
+	assert.Equal(t, params.EnableJSONSocket, loaded.EnableJSONSocket)
 	assert.Equal(t, params.ManagementURL, loaded.ManagementURL)
 	assert.Equal(t, params.ConfigPath, loaded.ConfigPath)
 	assert.Equal(t, params.LogFiles, loaded.LogFiles)
@@ -101,6 +105,8 @@ func TestLoadServiceParams_InvalidJSON(t *testing.T) {
 func TestCurrentServiceParams(t *testing.T) {
 	origLogLevel := logLevel
 	origDaemonAddr := daemonAddr
+	origJSONSocket := jsonSocket
+	origEnableJSONSocket := enableJSONSocket
 	origManagementURL := managementURL
 	origConfigPath := configPath
 	origLogFiles := logFiles
@@ -110,6 +116,8 @@ func TestCurrentServiceParams(t *testing.T) {
 	t.Cleanup(func() {
 		logLevel = origLogLevel
 		daemonAddr = origDaemonAddr
+		jsonSocket = origJSONSocket
+		enableJSONSocket = origEnableJSONSocket
 		managementURL = origManagementURL
 		configPath = origConfigPath
 		logFiles = origLogFiles
@@ -120,6 +128,8 @@ func TestCurrentServiceParams(t *testing.T) {
 
 	logLevel = "trace"
 	daemonAddr = "tcp://127.0.0.1:9999"
+	jsonSocket = "tcp://127.0.0.1:8080"
+	enableJSONSocket = true
 	managementURL = "https://mgmt.example.com"
 	configPath = "/tmp/test-config.json"
 	logFiles = []string{"/tmp/test.log"}
@@ -131,6 +141,8 @@ func TestCurrentServiceParams(t *testing.T) {
 
 	assert.Equal(t, "trace", params.LogLevel)
 	assert.Equal(t, "tcp://127.0.0.1:9999", params.DaemonAddr)
+	assert.Equal(t, "tcp://127.0.0.1:8080", params.JSONSocket)
+	assert.True(t, params.EnableJSONSocket)
 	assert.Equal(t, "https://mgmt.example.com", params.ManagementURL)
 	assert.Equal(t, "/tmp/test-config.json", params.ConfigPath)
 	assert.Equal(t, []string{"/tmp/test.log"}, params.LogFiles)
@@ -142,6 +154,8 @@ func TestCurrentServiceParams(t *testing.T) {
 func TestApplyServiceParams_OnlyUnchangedFlags(t *testing.T) {
 	origLogLevel := logLevel
 	origDaemonAddr := daemonAddr
+	origJSONSocket := jsonSocket
+	origEnableJSONSocket := enableJSONSocket
 	origManagementURL := managementURL
 	origConfigPath := configPath
 	origLogFiles := logFiles
@@ -151,6 +165,8 @@ func TestApplyServiceParams_OnlyUnchangedFlags(t *testing.T) {
 	t.Cleanup(func() {
 		logLevel = origLogLevel
 		daemonAddr = origDaemonAddr
+		jsonSocket = origJSONSocket
+		enableJSONSocket = origEnableJSONSocket
 		managementURL = origManagementURL
 		configPath = origConfigPath
 		logFiles = origLogFiles
@@ -162,6 +178,8 @@ func TestApplyServiceParams_OnlyUnchangedFlags(t *testing.T) {
 	// Reset all flags to defaults.
 	logLevel = "info"
 	daemonAddr = "unix:///var/run/netbird.sock"
+	jsonSocket = defaultJSONSocket
+	enableJSONSocket = false
 	managementURL = ""
 	configPath = "/etc/netbird/config.json"
 	logFiles = []string{"/var/log/netbird/client.log"}
@@ -184,6 +202,8 @@ func TestApplyServiceParams_OnlyUnchangedFlags(t *testing.T) {
 	saved := &serviceParams{
 		LogLevel:              "debug",
 		DaemonAddr:            "tcp://127.0.0.1:5555",
+		JSONSocket:            "tcp://127.0.0.1:8080",
+		EnableJSONSocket:      true,
 		ManagementURL:         "https://saved.example.com",
 		ConfigPath:            "/saved/config.json",
 		LogFiles:              []string{"/saved/client.log"},
@@ -201,6 +221,8 @@ func TestApplyServiceParams_OnlyUnchangedFlags(t *testing.T) {
 
 	// All other fields were not Changed, so they should use saved values.
 	assert.Equal(t, "tcp://127.0.0.1:5555", daemonAddr)
+	assert.Equal(t, "tcp://127.0.0.1:8080", jsonSocket)
+	assert.True(t, enableJSONSocket)
 	assert.Equal(t, "https://saved.example.com", managementURL)
 	assert.Equal(t, "/saved/config.json", configPath)
 	assert.Equal(t, []string{"/saved/client.log"}, logFiles)
@@ -212,14 +234,17 @@ func TestApplyServiceParams_OnlyUnchangedFlags(t *testing.T) {
 func TestApplyServiceParams_BooleanRevertToFalse(t *testing.T) {
 	origProfilesDisabled := profilesDisabled
 	origUpdateSettingsDisabled := updateSettingsDisabled
+	origEnableJSONSocket := enableJSONSocket
 	t.Cleanup(func() {
 		profilesDisabled = origProfilesDisabled
 		updateSettingsDisabled = origUpdateSettingsDisabled
+		enableJSONSocket = origEnableJSONSocket
 	})
 
 	// Simulate current state where booleans are true (e.g. set by previous install).
 	profilesDisabled = true
 	updateSettingsDisabled = true
+	enableJSONSocket = true
 
 	// Reset Changed state so flags appear unset.
 	serviceCmd.PersistentFlags().VisitAll(func(f *pflag.Flag) {
@@ -238,6 +263,7 @@ func TestApplyServiceParams_BooleanRevertToFalse(t *testing.T) {
 
 	assert.False(t, profilesDisabled, "saved false should override current true")
 	assert.False(t, updateSettingsDisabled, "saved false should override current true")
+	assert.False(t, enableJSONSocket, "saved false should override current true")
 }
 
 func TestApplyServiceParams_ClearManagementURL(t *testing.T) {
@@ -530,6 +556,7 @@ func fieldToGlobalVar(field string) string {
 	m := map[string]string{
 		"LogLevel":              "logLevel",
 		"DaemonAddr":            "daemonAddr",
+		"JSONSocket":            "jsonSocket",
 		"ManagementURL":         "managementURL",
 		"ConfigPath":            "configPath",
 		"LogFiles":              "logFiles",
@@ -537,6 +564,7 @@ func fieldToGlobalVar(field string) string {
 		"DisableUpdateSettings": "updateSettingsDisabled",
 		"EnableCapture":         "captureEnabled",
 		"DisableNetworks":       "networksDisabled",
+		"EnableJSONSocket":      "enableJSONSocket",
 		"ServiceEnvVars":        "serviceEnvVars",
 	}
 	if v, ok := m[field]; ok {

client/cmd/service_socket.go

@@ -0,0 +1,111 @@
+//go:build !ios && !android
+
+package cmd
+
+import (
+	"errors"
+	"fmt"
+	"net"
+	"os"
+	"strings"
+	"syscall"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+type socketListener struct {
+	net.Listener
+	network string
+	address string
+}
+
+func listenOnAddress(addr string) (*socketListener, error) {
+	network, address, err := parseListenAddress(addr)
+	if err != nil {
+		return nil, err
+	}
+
+	if network == "unix" {
+		removeStaleUnixSocket(address)
+	}
+
+	listener, err := net.Listen(network, address)
+	if err != nil {
+		return nil, err
+	}
+
+	return &socketListener{Listener: listener, network: network, address: address}, nil
+}
+
+func parseListenAddress(addr string) (string, string, error) {
+	network, address, ok := strings.Cut(addr, "://")
+	if !ok || network == "" || address == "" {
+		return "", "", fmt.Errorf("address must be in [unix|tcp]://[path|host:port] format: %q", addr)
+	}
+
+	switch network {
+	case "unix", "tcp":
+		return network, address, nil
+	default:
+		return "", "", fmt.Errorf("unsupported daemon address protocol: %v", network)
+	}
+}
+
+func removeStaleUnixSocket(path string) {
+	stat, err := os.Lstat(path)
+	if err != nil {
+		if !os.IsNotExist(err) {
+			log.Debugf("stat socket file: %v", err)
+		}
+		return
+	}
+
+	if stat.Mode()&os.ModeSocket == 0 {
+		return
+	}
+
+	if !isStaleUnixSocket(path) {
+		return
+	}
+
+	if err := os.Remove(path); err != nil {
+		log.Debugf("remove socket file: %v", err)
+	}
+}
+
+func isStaleUnixSocket(path string) bool {
+	conn, err := net.DialTimeout("unix", path, 100*time.Millisecond)
+	if err == nil {
+		if closeErr := conn.Close(); closeErr != nil {
+			log.Debugf("close unix socket probe: %v", closeErr)
+		}
+		return false
+	}
+
+	if os.IsNotExist(err) || os.IsPermission(err) || os.IsTimeout(err) {
+		log.Debugf("not removing unix socket %s after probe error: %v", path, err)
+		return false
+	}
+
+	return errors.Is(err, syscall.ECONNREFUSED)
+}
+
+func removeStaleUnixSocketForAddress(addr string) {
+	network, address, err := parseListenAddress(addr)
+	if err != nil || network != "unix" {
+		return
+	}
+	removeStaleUnixSocket(address)
+}
+
+func (l *socketListener) chmodUnixSocket(description string) error {
+	if l == nil || l.network != "unix" {
+		return nil
+	}
+
+	if err := os.Chmod(l.address, 0666); err != nil {
+		return fmt.Errorf("failed setting %s permissions for %s: %w", description, l.address, err)
+	}
+	return nil
+}

client/cmd/up.go

@@ -22,6 +22,8 @@ import (
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/proto"
+	nbnet "github.com/netbirdio/netbird/client/net"
+	"github.com/netbirdio/netbird/client/server"
 	"github.com/netbirdio/netbird/client/system"
 	"github.com/netbirdio/netbird/shared/management/domain"
 	"github.com/netbirdio/netbird/util"
@@ -229,6 +231,24 @@ func runInForegroundMode(ctx context.Context, cmd *cobra.Command, activeProf *pr
 
 	_, _ = profilemanager.UpdateOldManagementURL(ctx, config, configFilePath)
 
+	// Restore residual state left by a previous run that did not shut down
+	// cleanly, mirroring what the daemon does before connecting: it recovers
+	// DNS config (a stale resolv.conf takeover can make the management
+	// hostname unresolvable), firewall rules, ssh config and legacy routing.
+	// Route cleanup itself happens at engine start; nbnet.Init() below lets
+	// the management dial bypass a leftover fwmark rule until then.
+	// Foreground mode is particularly exposed in containers: a crashed
+	// container restarts inside the same (pod) network namespace, so stale
+	// state survives while the process does not.
+	if err := server.RestoreResidualState(ctx, profilemanager.NewServiceManager(configPath).GetStatePath()); err != nil {
+		log.Warnf("failed to restore residual state: %v", err)
+	}
+
+	// Enable advanced routing (as the daemon does on startup) so the
+	// management dial bypasses a leftover fwmark rule instead of being
+	// shunted into a stale routing table.
+	nbnet.Init()
+
 	err = foregroundLogin(ctx, cmd, config, providedSetupKey, activeProf.ID)
 	if err != nil {
 		return fmt.Errorf("foreground login failed: %v", err)
@@ -479,10 +499,6 @@ func setupSetConfigReq(customDNSAddressConverted []byte, cmd *cobra.Command, pro
 		req.DisableIpv6 = &disableIPv6
 	}
 
-	if cmd.Flag(enableLazyConnectionFlag).Changed {
-		req.LazyConnectionEnabled = &lazyConnEnabled
-	}
-
 	return &req
 }
 
@@ -600,9 +616,6 @@ func setupConfig(customDNSAddressConverted []byte, cmd *cobra.Command, configFil
 		ic.DisableIPv6 = &disableIPv6
 	}
 
-	if cmd.Flag(enableLazyConnectionFlag).Changed {
-		ic.LazyConnectionEnabled = &lazyConnEnabled
-	}
 	return &ic, nil
 }
 
@@ -718,9 +731,6 @@ func setupLoginRequest(providedSetupKey string, customDNSAddressConverted []byte
 		loginRequest.DisableIpv6 = &disableIPv6
 	}
 
-	if cmd.Flag(enableLazyConnectionFlag).Changed {
-		loginRequest.LazyConnectionEnabled = &lazyConnEnabled
-	}
 	return &loginRequest, nil
 }
 

client/firewall/uspfilter/filter.go

@@ -121,6 +121,7 @@ type Manager struct {
 	udpTracker     *conntrack.UDPTracker
 	icmpTracker    *conntrack.ICMPTracker
 	tcpTracker     *conntrack.TCPTracker
+	fragments      *fragmentTracker
 	forwarder      atomic.Pointer[forwarder.Forwarder]
 	pendingCapture atomic.Pointer[forwarder.PacketCapture]
 	logger         *nblog.Logger
@@ -183,6 +184,41 @@ func (d *decoder) decodePacket(data []byte) error {
 	}
 }
 
+// decodeTransport decodes the transport header of a first fragment (which
+// gopacket leaves undecoded) into the decoder and appends its layer type to
+// decoded, so the ACL pipeline can evaluate it like a normal packet. It returns
+// false if the protocol is unsupported or the header is truncated.
+func (d *decoder) decodeTransport(proto layers.IPProtocol, payload []byte) bool {
+	var l4 gopacket.DecodingLayer
+	var layerType gopacket.LayerType
+	var minLen int
+	switch proto {
+	case layers.IPProtocolTCP:
+		l4, layerType, minLen = &d.tcp, layers.LayerTypeTCP, 20
+	case layers.IPProtocolUDP:
+		l4, layerType, minLen = &d.udp, layers.LayerTypeUDP, 8
+	case layers.IPProtocolICMPv4:
+		l4, layerType, minLen = &d.icmp4, layers.LayerTypeICMPv4, 8
+	case layers.IPProtocolICMPv6:
+		l4, layerType, minLen = &d.icmp6, layers.LayerTypeICMPv6, 8
+	default:
+		return false
+	}
+
+	// Reject a fragment too small to hold the full transport header before
+	// decoding: it can't be ACL-evaluated (tiny-fragment attack), and skipping
+	// the decode avoids gopacket allocating an error on the drop path.
+	if len(payload) < minLen {
+		return false
+	}
+
+	if err := l4.DecodeFromBytes(payload, gopacket.NilDecodeFeedback); err != nil {
+		return false
+	}
+	d.decoded = append(d.decoded, layerType)
+	return true
+}
+
 // Create userspace firewall manager constructor
 func Create(iface common.IFaceMapper, disableServerRoutes bool, flowLogger nftypes.FlowLogger, mtu uint16) (*Manager, error) {
 	return create(iface, nil, disableServerRoutes, flowLogger, mtu)
@@ -286,6 +322,8 @@ func create(iface common.IFaceMapper, nativeFirewall firewall.Manager, disableSe
 	if err := m.localipmanager.UpdateLocalIPs(iface); err != nil {
 		return nil, fmt.Errorf("update local IPs: %w", err)
 	}
+	m.fragments = newFragmentTracker(m.logger)
+
 	if disableConntrack {
 		log.Info("conntrack is disabled")
 	} else {
@@ -299,6 +337,7 @@ func create(iface common.IFaceMapper, nativeFirewall firewall.Manager, disableSe
 		}
 	}
 	if err := iface.SetFilter(m); err != nil {
+		m.fragments.Close()
 		return nil, fmt.Errorf("set filter: %w", err)
 	}
 	return m, nil
@@ -694,6 +733,10 @@ func (m *Manager) resetState() {
 		m.tcpTracker.Close()
 	}
 
+	if m.fragments != nil {
+		m.fragments.Close()
+	}
+
 	if fwder := m.forwarder.Load(); fwder != nil {
 		fwder.SetCapture(nil)
 		fwder.Stop()
@@ -1046,19 +1089,20 @@ func (m *Manager) filterInbound(packetData []byte, size int) bool {
 		return true
 	}
 
-	// TODO: pass fragments of routed packets to forwarder
+	// gopacket does not decode the transport header of any IP fragment, so
+	// fragments take a dedicated path: the first fragment's header is decoded
+	// and ACL-evaluated here, and the remaining fragments inherit its verdict.
 	if fragment {
-		if m.logger.Enabled(nblog.LevelTrace) {
-			if d.decoded[0] == layers.LayerTypeIPv4 {
-				m.logger.Trace4("packet is a fragment: src=%v dst=%v id=%v flags=%v",
-					srcIP, dstIP, d.ip4.Id, d.ip4.Flags)
-			} else {
-				m.logger.Trace2("packet is an IPv6 fragment: src=%v dst=%v", srcIP, dstIP)
-			}
-		}
-		return false
+		return m.filterInboundFragment(d, srcIP, dstIP, size)
 	}
 
+	return m.filterInboundDecoded(d, srcIP, dstIP, packetData, size)
+}
+
+// filterInboundDecoded runs the ACL, DNAT and conntrack pipeline on a fully
+// decoded (non-fragment) inbound packet. It returns true if the packet should
+// be dropped.
+func (m *Manager) filterInboundDecoded(d *decoder, srcIP, dstIP netip.Addr, packetData []byte, size int) bool {
 	// TODO: optimize port DNAT by caching matched rules in conntrack
 	if translated := m.translateInboundPortDNAT(packetData, d, srcIP, dstIP); translated {
 		// Re-decode after port DNAT translation to update port information
@@ -1089,33 +1133,226 @@ func (m *Manager) filterInbound(packetData []byte, size int) bool {
 	return m.handleRoutedTraffic(d, srcIP, dstIP, packetData, size)
 }
 
+// fragmentMeta holds the reassembly identity and layout of an IP fragment,
+// extracted uniformly for IPv4 and IPv6.
+type fragmentMeta struct {
+	key fragmentKey
+	// offset is the fragment offset in 8-byte units (zero for the first
+	// fragment).
+	offset uint16
+	// moreFragments is the More Fragments bit. A first fragment with it unset is
+	// an IPv6 atomic fragment (a complete datagram, RFC 6946): it has no trailing
+	// fragments to inherit a verdict, so it must not be recorded.
+	moreFragments bool
+	proto         layers.IPProtocol
+	// l4payload is the fragmentable payload of this fragment. For the first
+	// fragment it starts with the transport header.
+	l4payload []byte
+	// headerEndOctets is the first fragment's payload length in 8-byte units:
+	// the smallest offset a trailing fragment may start at without overlapping
+	// the inspected transport header.
+	headerEndOctets uint16
+}
+
+// fragmentMetadata extracts the fragment identity and layout from a decoded IP
+// fragment. It returns false for fragments it can't interpret (e.g. an IPv6
+// fragment header shorter than 8 bytes), which are then dropped.
+func fragmentMetadata(d *decoder, srcIP, dstIP netip.Addr) (fragmentMeta, bool) {
+	switch d.decoded[0] {
+	case layers.LayerTypeIPv4:
+		payload := d.ip4.Payload
+		return fragmentMeta{
+			key:             fragmentKey{srcIP: srcIP, dstIP: dstIP, id: uint32(d.ip4.Id), proto: uint8(d.ip4.Protocol)},
+			offset:          d.ip4.FragOffset,
+			moreFragments:   d.ip4.Flags&layers.IPv4MoreFragments != 0,
+			proto:           d.ip4.Protocol,
+			l4payload:       payload,
+			headerEndOctets: octets(len(payload)),
+		}, true
+
+	case layers.LayerTypeIPv6:
+		// IPv6 fragment extension header: 8 bytes, followed by the fragmentable
+		// payload. Layout: next header (1), reserved (1), offset+flags (2), id (4).
+		payload := d.ip6.Payload
+		if len(payload) < 8 {
+			return fragmentMeta{}, false
+		}
+		nextHeader := layers.IPProtocol(payload[0])
+		offsetFlags := binary.BigEndian.Uint16(payload[2:4])
+		id := binary.BigEndian.Uint32(payload[4:8])
+		l4 := payload[8:]
+		return fragmentMeta{
+			key:             fragmentKey{srcIP: srcIP, dstIP: dstIP, id: id, proto: uint8(nextHeader)},
+			offset:          offsetFlags >> 3,
+			moreFragments:   offsetFlags&1 != 0,
+			proto:           nextHeader,
+			l4payload:       l4,
+			headerEndOctets: octets(len(l4)),
+		}, true
+
+	default:
+		return fragmentMeta{}, false
+	}
+}
+
+// octets rounds a byte length up to whole 8-byte units, the granularity of the
+// IP fragment offset field.
+func octets(nbytes int) uint16 {
+	return uint16((nbytes + 7) / 8)
+}
+
+// filterInboundFragment decides the fate of an IP fragment. gopacket stops
+// decoding at the network layer for every fragment, so the first fragment's
+// transport header is decoded and ACL-evaluated here and its verdict recorded;
+// the remaining (headerless) fragments inherit that verdict. Anything that
+// cannot be tied to an allowed, non-overlapping first fragment is dropped.
+func (m *Manager) filterInboundFragment(d *decoder, srcIP, dstIP netip.Addr, size int) bool {
+	meta, ok := fragmentMetadata(d, srcIP, dstIP)
+	if !ok {
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace2("dropping unsupported fragment: src=%v dst=%v", srcIP, dstIP)
+		}
+		return true
+	}
+
+	if meta.offset != 0 {
+		return m.filterTrailingFragment(meta, srcIP, dstIP)
+	}
+
+	// A new first fragment supersedes any recorded verdict for this datagram, so
+	// a re-sent or overlapping offset-zero fragment can't inherit the old one.
+	m.fragments.poison(meta.key)
+
+	// First fragment: decode its transport header so the ACL can evaluate it. A
+	// decode failure means the fragment is too small to hold the full transport
+	// header (RFC 1858 §3 tiny-fragment attack); it can't be evaluated, so drop it.
+	if !d.decodeTransport(meta.proto, meta.l4payload) {
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace3("dropping first fragment without full L4 header: src=%v dst=%v id=%v",
+				srcIP, dstIP, meta.key.id)
+		}
+		return true
+	}
+
+	return m.filterFirstFragment(d, meta, srcIP, dstIP, size)
+}
+
+// filterTrailingFragment applies a recorded first-fragment verdict to a
+// non-first fragment.
+func (m *Manager) filterTrailingFragment(meta fragmentMeta, srcIP, dstIP netip.Addr) bool {
+	switch m.fragments.verdict(meta.key, meta.offset) {
+	case fragmentAllow:
+		return false
+	case fragmentOverlap:
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace3("dropping overlapping fragment rewriting inspected header: src=%v dst=%v id=%v",
+				srcIP, dstIP, meta.key.id)
+		}
+		return true
+	default:
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace3("dropping fragment with no allowed first fragment: src=%v dst=%v id=%v",
+				srcIP, dstIP, meta.key.id)
+		}
+		return true
+	}
+}
+
+// filterFirstFragment runs the verdict part of the inbound pipeline on a first
+// fragment with its transport header decoded. It mirrors filterInboundDecoded
+// but skips DNAT (port rewriting on fragments is unsupported) and forwarder
+// injection (fragments are left to the stack to reassemble, not forwarded).
+// Allowed fragments have their verdict recorded so the datagram's trailing
+// fragments inherit it.
+func (m *Manager) filterFirstFragment(d *decoder, meta fragmentMeta, srcIP, dstIP netip.Addr, size int) bool {
+	if m.stateful && m.isValidTrackedConnection(d, srcIP, dstIP, size) {
+		m.recordFirstFragment(meta)
+		return false
+	}
+
+	if m.localipmanager.IsLocalIP(dstIP) {
+		ruleID, blocked := m.peerACLsBlock(srcIP, d, nil)
+		if blocked {
+			m.storeDropFlow("Dropping local first fragment (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+				d, srcIP, dstIP, ruleID, size)
+			return true
+		}
+		m.trackInbound(d, srcIP, dstIP, ruleID, size)
+		m.recordFirstFragment(meta)
+		return false
+	}
+
+	if !m.routingEnabled.Load() {
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace2("Dropping routed fragment (routing disabled): src=%s dst=%s", srcIP, dstIP)
+		}
+		return true
+	}
+	if m.nativeRouter.Load() {
+		m.trackInbound(d, srcIP, dstIP, nil, size)
+		m.recordFirstFragment(meta)
+		return false
+	}
+
+	// TODO: pass fragments of routed packets to the forwarder; until then
+	// allowed routed fragments go to the native stack.
+	srcPort, dstPort := getPortsFromPacket(d)
+	ruleID, pass := m.routeACLsPass(srcIP, dstIP, d.decoded[1], srcPort, dstPort)
+	if !pass {
+		m.storeDropFlow("Dropping routed first fragment (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+			d, srcIP, dstIP, ruleID, size)
+		return true
+	}
+
+	m.recordFirstFragment(meta)
+	return false
+}
+
+// recordFirstFragment caches an allowed first fragment's verdict for its
+// trailing fragments to inherit. Atomic fragments (no More Fragments bit) are
+// complete datagrams with no trailing fragments, so they are not cached and
+// cannot exhaust the verdict table.
+func (m *Manager) recordFirstFragment(meta fragmentMeta) {
+	if !meta.moreFragments {
+		return
+	}
+	m.fragments.recordAllowed(meta.key, meta.headerEndOctets)
+}
+
+// storeDropFlow logs and records a netflow drop event for an inbound packet
+// denied by the ACLs. msg is the trace format taking rule id, protocol, source
+// and destination.
+func (m *Manager) storeDropFlow(msg string, d *decoder, srcIP, dstIP netip.Addr, ruleID []byte, size int) {
+	pnum := getProtocolFromPacket(d)
+	srcPort, dstPort := getPortsFromPacket(d)
+
+	if m.logger.Enabled(nblog.LevelTrace) {
+		m.logger.Trace6(msg, ruleID, pnum, srcIP, srcPort, dstIP, dstPort)
+	}
+
+	m.flowLogger.StoreEvent(nftypes.EventFields{
+		FlowID:     uuid.New(),
+		Type:       nftypes.TypeDrop,
+		RuleID:     ruleID,
+		Direction:  nftypes.Ingress,
+		Protocol:   pnum,
+		SourceIP:   srcIP,
+		DestIP:     dstIP,
+		SourcePort: srcPort,
+		DestPort:   dstPort,
+		// TODO: icmp type/code
+		RxPackets: 1,
+		RxBytes:   uint64(size),
+	})
+}
+
 // handleLocalTraffic handles local traffic.
 // If it returns true, the packet should be dropped.
 func (m *Manager) handleLocalTraffic(d *decoder, srcIP, dstIP netip.Addr, packetData []byte, size int) bool {
 	ruleID, blocked := m.peerACLsBlock(srcIP, d, packetData)
 	if blocked {
-		pnum := getProtocolFromPacket(d)
-		srcPort, dstPort := getPortsFromPacket(d)
-
-		if m.logger.Enabled(nblog.LevelTrace) {
-			m.logger.Trace6("Dropping local packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
-				ruleID, pnum, srcIP, srcPort, dstIP, dstPort)
-		}
-
-		m.flowLogger.StoreEvent(nftypes.EventFields{
-			FlowID:     uuid.New(),
-			Type:       nftypes.TypeDrop,
-			RuleID:     ruleID,
-			Direction:  nftypes.Ingress,
-			Protocol:   pnum,
-			SourceIP:   srcIP,
-			DestIP:     dstIP,
-			SourcePort: srcPort,
-			DestPort:   dstPort,
-			// TODO: icmp type/code
-			RxPackets: 1,
-			RxBytes:   uint64(size),
-		})
+		m.storeDropFlow("Dropping local packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+			d, srcIP, dstIP, ruleID, size)
 		return true
 	}
 
@@ -1168,27 +1405,8 @@ func (m *Manager) handleRoutedTraffic(d *decoder, srcIP, dstIP netip.Addr, packe
 
 	ruleID, pass := m.routeACLsPass(srcIP, dstIP, protoLayer, srcPort, dstPort)
 	if !pass {
-		proto := getProtocolFromPacket(d)
-
-		if m.logger.Enabled(nblog.LevelTrace) {
-			m.logger.Trace6("Dropping routed packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
-				ruleID, proto, srcIP, srcPort, dstIP, dstPort)
-		}
-
-		m.flowLogger.StoreEvent(nftypes.EventFields{
-			FlowID:     uuid.New(),
-			Type:       nftypes.TypeDrop,
-			RuleID:     ruleID,
-			Direction:  nftypes.Ingress,
-			Protocol:   proto,
-			SourceIP:   srcIP,
-			DestIP:     dstIP,
-			SourcePort: srcPort,
-			DestPort:   dstPort,
-			// TODO: icmp type/code
-			RxPackets: 1,
-			RxBytes:   uint64(size),
-		})
+		m.storeDropFlow("Dropping routed packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+			d, srcIP, dstIP, ruleID, size)
 		return true
 	}
 

client/firewall/uspfilter/forwarder/forwarder.go

@@ -5,7 +5,9 @@ import (
 	"fmt"
 	"net"
 	"net/netip"
+	"os"
 	"runtime"
+	"strconv"
 	"sync"
 	"time"
 
@@ -31,6 +33,11 @@ const (
 	defaultMaxInFlight   = 1024
 	iosReceiveWindow     = 16384
 	iosMaxInFlight       = 256
+
+	// envForceTCPRACK overrides the platform default for gVisor's RACK loss
+	// detection. Set to a truthy value to force RACK on, or a falsy value to
+	// force it off, on any platform.
+	envForceTCPRACK = "NB_FORCE_TCP_RACK"
 )
 
 type Forwarder struct {
@@ -152,6 +159,8 @@ func New(iface common.IFaceMapper, logger *nblog.Logger, flowLogger nftypes.Flow
 		maxInFlight = iosMaxInFlight
 	}
 
+	configureTCPRecovery(s)
+
 	tcpForwarder := tcp.NewForwarder(s, receiveWindow, maxInFlight, f.handleTCP)
 	s.SetTransportProtocolHandler(tcp.ProtocolNumber, tcpForwarder.HandlePacket)
 
@@ -466,3 +475,31 @@ func probeRawICMP(network, addr string, logger *nblog.Logger) bool {
 	logger.Debug1("forwarder: raw %s socket access available", network)
 	return true
 }
+
+// configureTCPRecovery disables gVisor's RACK loss detection on Windows, where
+// it interacts poorly with the host and collapses throughput on routed TCP
+// connections (gVisor issue #9778). Other platforms keep the default. The
+// EnvForceTCPRACK environment variable overrides the platform default.
+func configureTCPRecovery(s *stack.Stack) {
+	disableRACK := runtime.GOOS == "windows"
+
+	if val := os.Getenv(envForceTCPRACK); val != "" {
+		force, err := strconv.ParseBool(val)
+		if err != nil {
+			log.Warnf("parse %s: %v", envForceTCPRACK, err)
+		} else {
+			disableRACK = !force
+		}
+	}
+
+	if !disableRACK {
+		return
+	}
+
+	opt := tcpip.TCPRecovery(0)
+	if err := s.SetTransportProtocolOption(tcp.ProtocolNumber, &opt); err != nil {
+		log.Warnf("disable TCP RACK loss detection: %v", err)
+		return
+	}
+	log.Info("forwarder: TCP RACK loss detection disabled")
+}

client/firewall/uspfilter/fragment.go

@@ -0,0 +1,204 @@
+package uspfilter
+
+import (
+	"context"
+	"net/netip"
+	"os"
+	"strconv"
+	"sync"
+	"time"
+
+	nblog "github.com/netbirdio/netbird/client/firewall/uspfilter/log"
+)
+
+const (
+	// defaultFragmentTimeout bounds how long a first-fragment verdict is kept
+	// while the remaining fragments arrive. It mirrors the Linux IP reassembly
+	// timeout (net.ipv4.ipfrag_time).
+	defaultFragmentTimeout = 30 * time.Second
+	// fragmentCleanupInterval is how often expired verdicts are purged.
+	fragmentCleanupInterval = 10 * time.Second
+	// defaultMaxFragmentEntries caps the number of concurrently tracked
+	// fragmented datagrams. The table stays bounded because each datagram is a
+	// single small entry regardless of how many fragments it is split into, and
+	// the 13-bit IPv4 fragment-offset field limits any datagram to 64 KiB.
+	defaultMaxFragmentEntries = 16384
+
+	// EnvFragmentMaxEntries overrides defaultMaxFragmentEntries.
+	EnvFragmentMaxEntries = "NB_FRAGMENT_MAX_ENTRIES"
+)
+
+// fragmentVerdict is the decision for a trailing (headerless) fragment.
+type fragmentVerdict int
+
+const (
+	// fragmentDeny drops the fragment: no allowed first fragment is on record.
+	fragmentDeny fragmentVerdict = iota
+	// fragmentAllow passes the fragment: it belongs to an allowed datagram and
+	// does not overlap the already-inspected transport header.
+	fragmentAllow
+	// fragmentOverlap drops the fragment and poisons its datagram: it overlaps
+	// the transport header the ACL inspected (RFC 1858 §4, RFC 3128; RFC 5722
+	// requires discarding the whole datagram on overlap for IPv6).
+	fragmentOverlap
+)
+
+// fragmentKey identifies a fragmented datagram. It matches the RFC 791 / RFC
+// 8200 reassembly key: source, destination, protocol and identification. The id
+// is 32-bit to hold both the IPv4 (16-bit) and IPv6 (32-bit) identification.
+type fragmentKey struct {
+	srcIP netip.Addr
+	dstIP netip.Addr
+	id    uint32
+	proto uint8
+}
+
+// fragmentEntry records the verdict of an allowed first fragment.
+type fragmentEntry struct {
+	// headerEndOctets is the offset, in 8-byte units, at which the first
+	// fragment's payload ended. A trailing fragment starting before this
+	// overlaps bytes the ACL already inspected and is rejected.
+	headerEndOctets uint16
+	// recordedAt is when the first fragment was accepted. The verdict expires a
+	// fixed timeout later and is not refreshed, mirroring the kernel reassembly
+	// timer so a trailing-fragment flood can't keep a datagram alive.
+	recordedAt time.Time
+}
+
+// fragmentTracker records the ACL verdict of a datagram's first fragment so the
+// remaining fragments, which carry no L4 header, can inherit the decision
+// without reassembling the datagram. Only allowed first fragments are stored;
+// anything that cannot be tied to an allowed, non-overlapping first fragment is
+// dropped (fail closed).
+type fragmentTracker struct {
+	logger  *nblog.Logger
+	mutex   sync.Mutex
+	entries map[fragmentKey]fragmentEntry
+	timeout time.Duration
+	// maxEntries caps the table; atCapacity dedups the capacity warning until
+	// the table drains below the cap again.
+	maxEntries    int
+	atCapacity    bool
+	cleanupTicker *time.Ticker
+	cancel        context.CancelFunc
+}
+
+func newFragmentTracker(logger *nblog.Logger) *fragmentTracker {
+	ctx, cancel := context.WithCancel(context.Background())
+	t := &fragmentTracker{
+		logger:        logger,
+		entries:       make(map[fragmentKey]fragmentEntry),
+		timeout:       defaultFragmentTimeout,
+		maxEntries:    fragmentMaxEntries(logger),
+		cleanupTicker: time.NewTicker(fragmentCleanupInterval),
+		cancel:        cancel,
+	}
+	go t.cleanupRoutine(ctx)
+	return t
+}
+
+func fragmentMaxEntries(logger *nblog.Logger) int {
+	v := os.Getenv(EnvFragmentMaxEntries)
+	if v == "" {
+		return defaultMaxFragmentEntries
+	}
+	n, err := strconv.Atoi(v)
+	if err != nil || n <= 0 {
+		logger.Warn2("invalid %s=%q, using default", EnvFragmentMaxEntries, v)
+		return defaultMaxFragmentEntries
+	}
+	return n
+}
+
+// recordAllowed stores the verdict of an allowed first fragment. headerEndOctets
+// is the first fragment's payload length in 8-byte units. When the table is full
+// the record is dropped, which fails closed: the datagram's trailing fragments
+// will be denied.
+func (t *fragmentTracker) recordAllowed(key fragmentKey, headerEndOctets uint16) {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+
+	if t.entries == nil {
+		return
+	}
+	if _, ok := t.entries[key]; !ok && len(t.entries) >= t.maxEntries {
+		if !t.atCapacity {
+			t.atCapacity = true
+			t.logger.Warn2("fragment verdict table at capacity (%d/%d): trailing fragments of new datagrams will be dropped",
+				len(t.entries), t.maxEntries)
+		}
+		return
+	}
+	t.entries[key] = fragmentEntry{
+		headerEndOctets: headerEndOctets,
+		recordedAt:      time.Now(),
+	}
+}
+
+// poison drops any recorded verdict for a datagram, so its later fragments are
+// denied until a new allowed first fragment is recorded. Called on every
+// offset-zero fragment to defeat offset-zero overlap rewrites (RFC 3128).
+func (t *fragmentTracker) poison(key fragmentKey) {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+	delete(t.entries, key)
+}
+
+// verdict decides the fate of a trailing fragment at fragOffsetOctets (the IPv4
+// fragment offset, in 8-byte units). A fragment overlapping the inspected
+// header poisons the datagram: the entry is removed so all further fragments of
+// that datagram are denied too.
+func (t *fragmentTracker) verdict(key fragmentKey, fragOffsetOctets uint16) fragmentVerdict {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+
+	entry, ok := t.entries[key]
+	if !ok {
+		return fragmentDeny
+	}
+	if time.Since(entry.recordedAt) > t.timeout {
+		delete(t.entries, key)
+		return fragmentDeny
+	}
+	if fragOffsetOctets < entry.headerEndOctets {
+		delete(t.entries, key)
+		return fragmentOverlap
+	}
+	return fragmentAllow
+}
+
+func (t *fragmentTracker) cleanupRoutine(ctx context.Context) {
+	defer t.cleanupTicker.Stop()
+	for {
+		select {
+		case <-t.cleanupTicker.C:
+			t.cleanup()
+		case <-ctx.Done():
+			return
+		}
+	}
+}
+
+func (t *fragmentTracker) cleanup() {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+
+	for key, entry := range t.entries {
+		if time.Since(entry.recordedAt) > t.timeout {
+			delete(t.entries, key)
+		}
+	}
+
+	if len(t.entries) < t.maxEntries {
+		t.atCapacity = false
+	}
+}
+
+// Close stops the cleanup routine and releases resources.
+func (t *fragmentTracker) Close() {
+	t.cancel()
+
+	t.mutex.Lock()
+	t.entries = nil
+	t.mutex.Unlock()
+}

client/firewall/uspfilter/fragment_bench_test.go

@@ -0,0 +1,115 @@
+package uspfilter
+
+import (
+	"encoding/binary"
+	"testing"
+)
+
+// benchFilterInbound drives filterInbound over a fixed packet in a tight loop.
+// Packets are built once, outside the timed region, so the benchmark measures
+// only pipeline cost, which is what an attacker can amplify.
+func benchFilterInbound(b *testing.B, pkt []byte) {
+	b.Helper()
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkt)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		m := benchManager
+		m.filterInbound(pkt, len(pkt))
+	}
+}
+
+// benchManager is a package-level manager reused across fragment benchmarks so
+// setup cost stays out of the timed region.
+var benchManager *Manager
+
+func setupBenchManager(b *testing.B) *Manager {
+	b.Helper()
+	m := newFragmentTestManager(b)
+	allowUDP(b, m, 8080)
+	// Disable conntrack so the allowed-first-fragment path measures transport
+	// decode + ACL every iteration instead of matching the connection tracked
+	// on the first iteration.
+	m.stateful = false
+	benchManager = m
+	return m
+}
+
+// BenchmarkInbound_NormalPacket is the baseline: a full, non-fragmented UDP
+// packet that passes the ACL. Fragment paths should stay comparable to this.
+func BenchmarkInbound_NormalPacket(b *testing.B) {
+	setupBenchManager(b)
+	pkt := normalUDPPacket(b, 8080, 32)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_FirstFragmentAllowed measures the first-fragment path:
+// transport decode + ACL evaluation + verdict record.
+func BenchmarkInbound_FirstFragmentAllowed(b *testing.B) {
+	setupBenchManager(b)
+	pkt := firstFragmentUDP(b, 0x2000, 8080, 32)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TrailingFragmentAllowed measures the common trailing-fragment
+// path: a single map lookup after the first fragment is on record.
+func BenchmarkInbound_TrailingFragmentAllowed(b *testing.B) {
+	m := setupBenchManager(b)
+	first := firstFragmentUDP(b, 0x3000, 8080, 32)
+	m.filterInbound(first, len(first))
+	pkt := trailingFragment(b, 0x3000, 5, false, 24)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TrailingFragmentNoFirst is the primary DoS vector: an
+// attacker floods trailing fragments with no first fragment on record. Each is
+// a map miss and must be cheap.
+func BenchmarkInbound_TrailingFragmentNoFirst(b *testing.B) {
+	setupBenchManager(b)
+	pkt := trailingFragment(b, 0x4000, 185, false, 40)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TinyFirstFragment measures the tiny-fragment drop path: a
+// first fragment too small to decode a transport header.
+func BenchmarkInbound_TinyFirstFragment(b *testing.B) {
+	setupBenchManager(b)
+	pkt := trailingFragment(b, 0x5000, 0, true, 4)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TrailingFragmentDistinctIDs is the worst case for the
+// verdict table: an attacker varies the datagram id on every packet so no first
+// fragment ever matches. Verdict lookups always miss and nothing is recorded,
+// so the table cannot grow. Each iteration rewrites the id field in place.
+func BenchmarkInbound_TrailingFragmentDistinctIDs(b *testing.B) {
+	setupBenchManager(b)
+	pkt := trailingFragment(b, 0x6000, 185, false, 40)
+	m := benchManager
+
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkt)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		// IPv4 identification field is at bytes 4:6.
+		binary.BigEndian.PutUint16(pkt[4:6], uint16(i))
+		m.filterInbound(pkt, len(pkt))
+	}
+}
+
+// BenchmarkInbound_FirstFragmentDistinctIDs measures sustained first-fragment
+// pressure with distinct ids: transport decode + ACL + verdict insert until the
+// table caps, exercising the map growth and capacity guard.
+func BenchmarkInbound_FirstFragmentDistinctIDs(b *testing.B) {
+	setupBenchManager(b)
+	pkt := firstFragmentUDP(b, 0x7000, 8080, 32)
+	m := benchManager
+
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkt)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		binary.BigEndian.PutUint16(pkt[4:6], uint16(i))
+		m.filterInbound(pkt, len(pkt))
+	}
+}

client/firewall/uspfilter/fragment_test.go

@@ -0,0 +1,554 @@
+package uspfilter
+
+import (
+	"encoding/binary"
+	"net"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/google/gopacket"
+	"github.com/google/gopacket/layers"
+	"github.com/stretchr/testify/require"
+
+	fw "github.com/netbirdio/netbird/client/firewall/manager"
+	nbiface "github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/device"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+)
+
+const (
+	fragTestSrc   = "100.10.0.1"
+	fragTestDst   = "100.10.0.100"
+	fragTestSrcV6 = "fd00::1"
+	fragTestDstV6 = "fd00::100"
+)
+
+func newFragmentTestManager(tb testing.TB) *Manager {
+	tb.Helper()
+
+	ifaceMock := &IFaceMock{
+		SetFilterFunc: func(device.PacketFilter) error { return nil },
+		AddressFunc: func() wgaddr.Address {
+			return wgaddr.Address{
+				IP:      netip.MustParseAddr(fragTestDst),
+				Network: netip.MustParsePrefix("100.10.0.0/16"),
+				IPv6:    netip.MustParseAddr(fragTestDstV6),
+				IPv6Net: netip.MustParsePrefix("fd00::/64"),
+			}
+		},
+	}
+
+	m, err := Create(ifaceMock, false, flowLogger, nbiface.DefaultMTU)
+	require.NoError(tb, err)
+	require.NoError(tb, m.UpdateLocalIPs())
+	tb.Cleanup(func() { require.NoError(tb, m.Close(nil)) })
+	return m
+}
+
+// firstFragmentUDPTo builds the first fragment of a fragmented UDP datagram to
+// the given destination: it carries the full UDP header plus payloadLen bytes
+// of data, with the More Fragments flag set and offset zero.
+func firstFragmentUDPTo(tb testing.TB, dst string, id uint16, dstPort uint16, payloadLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       id,
+		Protocol: layers.IPProtocolUDP,
+		SrcIP:    net.ParseIP(fragTestSrc),
+		DstIP:    net.ParseIP(dst),
+		Flags:    layers.IPv4MoreFragments,
+	}
+	udp := &layers.UDP{SrcPort: 40000, DstPort: layers.UDPPort(dstPort)}
+	require.NoError(tb, udp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, udp, gopacket.Payload(make([]byte, payloadLen))))
+	return buf.Bytes()
+}
+
+func firstFragmentUDP(tb testing.TB, id uint16, dstPort uint16, payloadLen int) []byte {
+	tb.Helper()
+	return firstFragmentUDPTo(tb, fragTestDst, id, dstPort, payloadLen)
+}
+
+// firstFragmentTCP builds the first fragment of a fragmented TCP datagram: the
+// full 20-byte TCP header plus 12 bytes of data, with the More Fragments flag
+// set and offset zero.
+func firstFragmentTCP(tb testing.TB, id uint16, dstPort uint16) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       id,
+		Protocol: layers.IPProtocolTCP,
+		SrcIP:    net.ParseIP(fragTestSrc),
+		DstIP:    net.ParseIP(fragTestDst),
+		Flags:    layers.IPv4MoreFragments,
+	}
+	tcp := &layers.TCP{SrcPort: 40000, DstPort: layers.TCPPort(dstPort), SYN: true, Window: 64240}
+	require.NoError(tb, tcp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, tcp, gopacket.Payload(make([]byte, 12))))
+	return buf.Bytes()
+}
+
+// trailingFragmentTo builds a non-first fragment to the given destination: an
+// IPv4 header at the given fragment offset (in 8-byte units) carrying raw
+// payload and no L4 header.
+func trailingFragmentTo(tb testing.TB, dst string, proto layers.IPProtocol, id uint16, fragOffsetOctets uint16, moreFragments bool, payloadLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:    4,
+		TTL:        64,
+		Id:         id,
+		Protocol:   proto,
+		SrcIP:      net.ParseIP(fragTestSrc),
+		DstIP:      net.ParseIP(dst),
+		FragOffset: fragOffsetOctets,
+	}
+	if moreFragments {
+		ip.Flags = layers.IPv4MoreFragments
+	}
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, gopacket.Payload(make([]byte, payloadLen))))
+	return buf.Bytes()
+}
+
+func trailingFragment(tb testing.TB, id uint16, fragOffsetOctets uint16, moreFragments bool, payloadLen int) []byte {
+	tb.Helper()
+	return trailingFragmentTo(tb, fragTestDst, layers.IPProtocolUDP, id, fragOffsetOctets, moreFragments, payloadLen)
+}
+
+// outboundUDPPacket builds a complete outbound UDP packet from the local
+// address, used to establish conntrack state for reply-direction tests.
+func outboundUDPPacket(tb testing.TB, srcPort, dstPort uint16) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       1,
+		Protocol: layers.IPProtocolUDP,
+		SrcIP:    net.ParseIP(fragTestDst),
+		DstIP:    net.ParseIP(fragTestSrc),
+	}
+	udp := &layers.UDP{SrcPort: layers.UDPPort(srcPort), DstPort: layers.UDPPort(dstPort)}
+	require.NoError(tb, udp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, udp, gopacket.Payload(make([]byte, 16))))
+	return buf.Bytes()
+}
+
+// normalUDPPacket builds a complete, non-fragmented UDP packet for baseline
+// comparisons against the fragment paths.
+func normalUDPPacket(tb testing.TB, dstPort uint16, payloadLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       1,
+		Protocol: layers.IPProtocolUDP,
+		SrcIP:    net.ParseIP(fragTestSrc),
+		DstIP:    net.ParseIP(fragTestDst),
+	}
+	udp := &layers.UDP{SrcPort: 40000, DstPort: layers.UDPPort(dstPort)}
+	require.NoError(tb, udp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, udp, gopacket.Payload(make([]byte, payloadLen))))
+	return buf.Bytes()
+}
+
+func allowUDP(tb testing.TB, m *Manager, dstPort uint16) {
+	tb.Helper()
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolUDP, nil,
+		&fw.Port{Values: []uint16{dstPort}}, fw.ActionAccept, "")
+	require.NoError(tb, err)
+}
+
+// TestFragment_TrailingWithoutFirstDropped is the core bypass repro: a trailing
+// fragment with no allowed first fragment on record must be dropped. Before the
+// fix, filterInbound returned false (allow) for any fragment.
+func TestFragment_TrailingWithoutFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+
+	frag := trailingFragment(t, 0x1234, 185, false, 40)
+	require.True(t, m.filterInbound(frag, len(frag)),
+		"trailing fragment without an allowed first fragment must be dropped")
+}
+
+// TestFragment_AllowedFirstPassesTrailing verifies that once a first fragment
+// passes the ACL, its trailing fragments inherit the allow verdict.
+func TestFragment_AllowedFirstPassesTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	// First fragment: UDP header (8) + 32 payload = 40 octets -> headerEnd = 5.
+	first := firstFragmentUDP(t, 0x2222, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"allowed first fragment should pass and be recorded")
+
+	trailing := trailingFragment(t, 0x2222, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed datagram should pass")
+}
+
+// TestFragment_DeniedFirstDropsTrailing verifies that a first fragment blocked
+// by the ACL leaves no verdict, so its trailing fragments are dropped.
+func TestFragment_DeniedFirstDropsTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	// No accept rule: local traffic defaults to deny.
+
+	first := firstFragmentUDP(t, 0x3333, 9999, 32)
+	require.True(t, m.filterInbound(first, len(first)),
+		"first fragment to a blocked port should be dropped by the ACL")
+
+	trailing := trailingFragment(t, 0x3333, 5, false, 24)
+	require.True(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of a denied datagram must be dropped")
+}
+
+// TestFragment_OverlappingHeaderDropped covers the RFC 1858 §4 / RFC 3128
+// overlapping-fragment rewrite: a trailing fragment starting inside the range
+// the ACL already inspected is dropped and poisons the datagram. TCP is used so
+// the overlap lands on real header bytes (the flags at byte 13).
+func TestFragment_OverlappingHeaderDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolTCP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	// First fragment: TCP header (20) + 12 data = 32 bytes -> headerEnd = 4 octets.
+	first := firstFragmentTCP(t, 0x4444, 8080)
+	require.False(t, m.filterInbound(first, len(first)))
+
+	// Overlapping fragment at offset 1 (byte 8) falls inside the inspected TCP
+	// header, so it could rewrite the flags or port on reassembly.
+	overlap := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x4444, 1, true, 32)
+	require.True(t, m.filterInbound(overlap, len(overlap)),
+		"fragment overlapping the inspected header must be dropped")
+
+	// The datagram is now poisoned: a later, non-overlapping fragment is also
+	// dropped because the verdict was removed.
+	later := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x4444, 4, false, 24)
+	require.True(t, m.filterInbound(later, len(later)),
+		"fragments after an overlap must be dropped (datagram poisoned)")
+}
+
+// TestFragment_OffsetZeroOverlapPoisons covers the RFC 3128 offset-zero rewrite:
+// an allowed first fragment followed by a denied offset-zero fragment for the
+// same datagram must not leave the earlier allow verdict in place.
+func TestFragment_OffsetZeroOverlapPoisons(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	allowed := firstFragmentUDP(t, 0x5A5A, 8080, 32)
+	require.False(t, m.filterInbound(allowed, len(allowed)),
+		"allowed first fragment should pass and be recorded")
+
+	// A second offset-zero fragment to a denied port supersedes the datagram's
+	// verdict; it is dropped and must not leave the allow in place.
+	denied := firstFragmentUDP(t, 0x5A5A, 9999, 32)
+	require.True(t, m.filterInbound(denied, len(denied)),
+		"denied offset-zero fragment must be dropped")
+
+	trailing := trailingFragment(t, 0x5A5A, 5, false, 24)
+	require.True(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment must be denied after the datagram was poisoned")
+}
+
+// TestFragment_TinyFirstDropped covers the tiny-fragment attack: a first
+// fragment too small to contain the full transport header can't be
+// ACL-evaluated and must be dropped.
+func TestFragment_TinyFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	// IPv4 header + 4 raw bytes, MF set, offset 0: too small for the 8-byte UDP
+	// header, so it decodes to L3 only.
+	tiny := trailingFragment(t, 0x5555, 0, true, 4)
+	require.True(t, m.filterInbound(tiny, len(tiny)),
+		"tiny first fragment without a full L4 header must be dropped")
+}
+
+// TestFragment_TCPFirstFragment verifies the TCP arm of the transport decode: a
+// first fragment carrying the full 20-byte TCP header is ACL-evaluated and its
+// trailing fragments inherit the verdict.
+func TestFragment_TCPFirstFragment(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolTCP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	// TCP header (20) + 12 data = 32 bytes -> headerEnd = 4 octets.
+	first := firstFragmentTCP(t, 0x6666, 8080)
+	require.False(t, m.filterInbound(first, len(first)),
+		"allowed TCP first fragment should pass and be recorded")
+
+	trailing := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x6666, 4, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed TCP datagram should pass")
+}
+
+// TestFragment_TCPTinyFirstDropped verifies the TCP minimum header length: 12
+// bytes would satisfy a UDP header but falls short of the 20-byte TCP header.
+func TestFragment_TCPTinyFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolTCP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	tiny := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x7777, 0, true, 12)
+	require.True(t, m.filterInbound(tiny, len(tiny)),
+		"first fragment shorter than the TCP header must be dropped")
+}
+
+// TestFragment_ConntrackAllowsFirstFragment verifies the conntrack branch: reply
+// fragments of an outbound-established UDP flow pass without any inbound rule.
+func TestFragment_ConntrackAllowsFirstFragment(t *testing.T) {
+	m := newFragmentTestManager(t)
+
+	out := outboundUDPPacket(t, 12345, 40000)
+	require.False(t, m.filterOutbound(out, len(out)))
+
+	first := firstFragmentUDP(t, 0x8888, 12345, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"reply first fragment should pass via conntrack")
+
+	trailing := trailingFragment(t, 0x8888, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of a tracked flow should pass")
+}
+
+// TestFragment_RoutingDisabledDropsFragment verifies routed first fragments are
+// dropped when routing is disabled.
+func TestFragment_RoutingDisabledDropsFragment(t *testing.T) {
+	m := newFragmentTestManager(t)
+	m.routingEnabled.Store(false)
+
+	first := firstFragmentUDPTo(t, "198.51.100.10", 0x9999, 8080, 32)
+	require.True(t, m.filterInbound(first, len(first)),
+		"routed first fragment must be dropped when routing is disabled")
+}
+
+// TestFragment_RouteACL verifies the route-ACL branch: fragments to a non-local
+// destination follow the route rules, allowed datagrams pass their trailing
+// fragments and denied ones don't.
+func TestFragment_RouteACL(t *testing.T) {
+	m := newFragmentTestManager(t)
+	m.routingEnabled.Store(true)
+	m.nativeRouter.Store(false)
+
+	_, err := m.AddRouteFiltering(
+		[]byte("rt-1"),
+		[]netip.Prefix{netip.MustParsePrefix("100.10.0.0/16")},
+		fw.Network{Prefix: netip.MustParsePrefix("198.51.100.0/24")},
+		fw.ProtocolUDP,
+		nil,
+		&fw.Port{Values: []uint16{8080}},
+		fw.ActionAccept,
+	)
+	require.NoError(t, err)
+
+	first := firstFragmentUDPTo(t, "198.51.100.10", 0xAAAA, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"route-ACL-allowed first fragment should pass")
+	trailing := trailingFragmentTo(t, "198.51.100.10", layers.IPProtocolUDP, 0xAAAA, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed routed datagram should pass")
+
+	denied := firstFragmentUDPTo(t, "198.51.100.10", 0xBBBB, 9999, 32)
+	require.True(t, m.filterInbound(denied, len(denied)),
+		"route-ACL-denied first fragment must be dropped")
+	deniedTrailing := trailingFragmentTo(t, "198.51.100.10", layers.IPProtocolUDP, 0xBBBB, 5, false, 24)
+	require.True(t, m.filterInbound(deniedTrailing, len(deniedTrailing)),
+		"trailing fragment of a denied routed datagram must be dropped")
+}
+
+// TestFragment_ExpiredVerdictDropsTrailing verifies a verdict older than the
+// tracker timeout no longer admits trailing fragments.
+func TestFragment_ExpiredVerdictDropsTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	first := firstFragmentUDP(t, 0xCCCC, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)))
+
+	m.fragments.mutex.Lock()
+	for key, entry := range m.fragments.entries {
+		entry.recordedAt = time.Now().Add(-defaultFragmentTimeout - time.Second)
+		m.fragments.entries[key] = entry
+	}
+	m.fragments.mutex.Unlock()
+
+	trailing := trailingFragment(t, 0xCCCC, 5, false, 24)
+	require.True(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment after verdict expiry must be dropped")
+}
+
+// TestFragment_CapacityFailsClosed verifies the table cap: at capacity, new
+// datagram verdicts are not recorded (their trailing fragments are dropped)
+// while already-recorded datagrams keep working.
+func TestFragment_CapacityFailsClosed(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	m.fragments.mutex.Lock()
+	m.fragments.maxEntries = 1
+	m.fragments.mutex.Unlock()
+
+	first1 := firstFragmentUDP(t, 0x0101, 8080, 32)
+	require.False(t, m.filterInbound(first1, len(first1)))
+
+	first2 := firstFragmentUDP(t, 0x0202, 8080, 32)
+	require.False(t, m.filterInbound(first2, len(first2)),
+		"first fragment itself still passes at capacity")
+
+	trailing2 := trailingFragment(t, 0x0202, 5, false, 24)
+	require.True(t, m.filterInbound(trailing2, len(trailing2)),
+		"trailing fragment of an unrecorded datagram must be dropped at capacity")
+
+	trailing1 := trailingFragment(t, 0x0101, 5, false, 24)
+	require.False(t, m.filterInbound(trailing1, len(trailing1)),
+		"already-recorded datagram should keep passing at capacity")
+}
+
+// v6FragmentHeader builds the 8-byte IPv6 fragment extension header for the
+// given inner protocol, offset (8-byte units), More Fragments bit and id.
+func v6FragmentHeader(proto layers.IPProtocol, offsetOctets uint16, moreFragments bool, id uint32) []byte {
+	offsetFlags := offsetOctets << 3
+	if moreFragments {
+		offsetFlags |= 1
+	}
+	hdr := make([]byte, 8)
+	hdr[0] = uint8(proto)
+	binary.BigEndian.PutUint16(hdr[2:4], offsetFlags)
+	binary.BigEndian.PutUint32(hdr[4:8], id)
+	return hdr
+}
+
+func v6UDPHeader(dstPort uint16, dataLen int) []byte {
+	hdr := make([]byte, 8)
+	binary.BigEndian.PutUint16(hdr[0:2], 40000)
+	binary.BigEndian.PutUint16(hdr[2:4], dstPort)
+	binary.BigEndian.PutUint16(hdr[4:6], uint16(8+dataLen))
+	return hdr
+}
+
+// firstFragmentUDPv6 builds the first fragment of a fragmented IPv6 UDP
+// datagram: fragment header (offset 0, More Fragments set) + full UDP header +
+// data.
+func firstFragmentUDPv6(tb testing.TB, id uint32, dstPort uint16, dataLen int) []byte {
+	tb.Helper()
+	return fragmentUDPv6(tb, id, dstPort, dataLen, true)
+}
+
+// fragmentUDPv6 builds an offset-zero IPv6 UDP fragment. With moreFragments
+// false it is an atomic fragment (a complete datagram, RFC 6946).
+func fragmentUDPv6(tb testing.TB, id uint32, dstPort uint16, dataLen int, moreFragments bool) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv6{
+		Version:    6,
+		NextHeader: layers.IPProtocolIPv6Fragment,
+		HopLimit:   64,
+		SrcIP:      net.ParseIP(fragTestSrcV6),
+		DstIP:      net.ParseIP(fragTestDstV6),
+	}
+	payload := append(v6FragmentHeader(layers.IPProtocolUDP, 0, moreFragments, id), v6UDPHeader(dstPort, dataLen)...)
+	payload = append(payload, make([]byte, dataLen)...)
+
+	buf := gopacket.NewSerializeBuffer()
+	require.NoError(tb, gopacket.SerializeLayers(buf, gopacket.SerializeOptions{FixLengths: true}, ip, gopacket.Payload(payload)))
+	return buf.Bytes()
+}
+
+// trailingFragmentV6 builds a non-first IPv6 fragment: fragment header at the
+// given offset carrying raw data and no transport header.
+func trailingFragmentV6(tb testing.TB, id uint32, offsetOctets uint16, moreFragments bool, dataLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv6{
+		Version:    6,
+		NextHeader: layers.IPProtocolIPv6Fragment,
+		HopLimit:   64,
+		SrcIP:      net.ParseIP(fragTestSrcV6),
+		DstIP:      net.ParseIP(fragTestDstV6),
+	}
+	payload := append(v6FragmentHeader(layers.IPProtocolUDP, offsetOctets, moreFragments, id), make([]byte, dataLen)...)
+
+	buf := gopacket.NewSerializeBuffer()
+	require.NoError(tb, gopacket.SerializeLayers(buf, gopacket.SerializeOptions{FixLengths: true}, ip, gopacket.Payload(payload)))
+	return buf.Bytes()
+}
+
+// TestFragmentV6_TrailingWithoutFirstDropped verifies the IPv6 bypass is closed:
+// a trailing fragment with no allowed first fragment is dropped.
+func TestFragmentV6_TrailingWithoutFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+
+	frag := trailingFragmentV6(t, 0xAABBCCDD, 100, false, 40)
+	require.True(t, m.filterInbound(frag, len(frag)),
+		"IPv6 trailing fragment without an allowed first fragment must be dropped")
+}
+
+// TestFragmentV6_AllowedFirstPassesTrailing verifies IPv6 fragments are
+// evaluated like IPv4: an allowed first fragment lets its trailing fragments
+// through.
+func TestFragmentV6_AllowedFirstPassesTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrcV6), fw.ProtocolUDP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	// First fragment: UDP header (8) + 32 data = 40 octets -> headerEnd = 5.
+	first := firstFragmentUDPv6(t, 0xAABBCCDD, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"allowed IPv6 first fragment should pass and be recorded")
+
+	trailing := trailingFragmentV6(t, 0xAABBCCDD, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed IPv6 datagram should pass")
+}
+
+// TestFragmentV6_AtomicNotCached verifies an IPv6 atomic fragment (fragment
+// header with offset 0 and no More Fragments, a complete datagram per RFC 6946)
+// is evaluated but not recorded, so a flood of allowed atomic fragments can't
+// exhaust the verdict table.
+func TestFragmentV6_AtomicNotCached(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrcV6), fw.ProtocolUDP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	atomic := fragmentUDPv6(t, 0xA70301C, 8080, 16, false)
+	require.False(t, m.filterInbound(atomic, len(atomic)),
+		"allowed IPv6 atomic fragment should pass")
+
+	m.fragments.mutex.Lock()
+	n := len(m.fragments.entries)
+	m.fragments.mutex.Unlock()
+	require.Zero(t, n, "atomic fragment must not create a verdict entry")
+
+	// A genuine fragmented datagram (More Fragments set) is still recorded.
+	first := fragmentUDPv6(t, 0xBEEF, 8080, 32, true)
+	require.False(t, m.filterInbound(first, len(first)))
+	m.fragments.mutex.Lock()
+	n = len(m.fragments.entries)
+	m.fragments.mutex.Unlock()
+	require.Equal(t, 1, n, "genuine first fragment must record a verdict")
+}

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/iface/iface_test.go

@@ -569,17 +569,17 @@ func Test_ConnectPeers(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	// todo: investigate why in some tests execution we need 30s
+	// The peers use userspace WireGuard (stdnet transport). A tight busy-loop
+	// here starves the wireguard-go goroutines that process the handshake, so
+	// poll on a ticker instead and yield the CPU between checks. WireGuard also
+	// only retries a lost handshake initiation every REKEY_TIMEOUT (5s), which
+	// is why the overall wait can occasionally stretch to tens of seconds.
 	timeout := 30 * time.Second
 	timeoutChannel := time.After(timeout)
+	ticker := time.NewTicker(500 * time.Millisecond)
+	defer ticker.Stop()
 
 	for {
-		select {
-		case <-timeoutChannel:
-			t.Fatalf("waiting for peer handshake timeout after %s", timeout.String())
-		default:
-		}
-
 		peer, gpErr := getPeer(peer1ifaceName, peer2Key.PublicKey().String())
 		if gpErr != nil {
 			t.Fatal(gpErr)
@@ -588,6 +588,12 @@ func Test_ConnectPeers(t *testing.T) {
 			t.Log("peers successfully handshake")
 			break
 		}
+
+		select {
+		case <-timeoutChannel:
+			t.Fatalf("waiting for peer handshake timeout after %s", timeout.String())
+		case <-ticker.C:
+		}
 	}
 
 }

client/iface/netstack/env.go

@@ -3,14 +3,31 @@
 package netstack
 
 import (
-	"fmt"
+	"net"
 	"os"
 	"strconv"
 
 	log "github.com/sirupsen/logrus"
 )
 
-const EnvUseNetstackMode = "NB_USE_NETSTACK_MODE"
+const (
+	EnvUseNetstackMode = "NB_USE_NETSTACK_MODE"
+
+	// EnvSocks5ListenerPort overrides the port the SOCKS5 proxy listens on.
+	EnvSocks5ListenerPort = "NB_SOCKS5_LISTENER_PORT"
+
+	// EnvSocks5ListenerAddress overrides the host/IP the SOCKS5 proxy binds to.
+	// The proxy is a bridge for local host applications into the userspace
+	// WireGuard netstack, so it binds to loopback by default. Override this only
+	// when the proxy must be reachable from other hosts (e.g. a container
+	// gateway); doing so exposes an unauthenticated SOCKS5 proxy on that
+	// address.
+	EnvSocks5ListenerAddress = "NB_SOCKS5_LISTENER_ADDRESS"
+
+	// defaultSocks5Host is the loopback address the SOCKS5 proxy binds to unless
+	// overridden via EnvSocks5ListenerAddress.
+	defaultSocks5Host = "127.0.0.1"
+)
 
 // IsEnabled todo: move these function to cmd layer
 func IsEnabled() bool {
@@ -18,24 +35,40 @@ func IsEnabled() bool {
 }
 
 func ListenAddr() string {
-	sPort := os.Getenv("NB_SOCKS5_LISTENER_PORT")
+	return net.JoinHostPort(listenHost(), strconv.Itoa(listenPort()))
+}
+
+// listenHost returns the host/IP the SOCKS5 proxy binds to. It defaults to
+// loopback and only honors EnvSocks5ListenerAddress when it holds a valid IP.
+func listenHost() string {
+	addr := os.Getenv(EnvSocks5ListenerAddress)
+	if addr == "" {
+		return defaultSocks5Host
+	}
+	if net.ParseIP(addr) == nil {
+		log.Warnf("invalid socks5 listener address %q, falling back to default: %s", addr, defaultSocks5Host)
+		return defaultSocks5Host
+	}
+	return addr
+}
+
+// listenPort returns the port the SOCKS5 proxy binds to, defaulting to
+// DefaultSocks5Port when EnvSocks5ListenerPort is unset or invalid.
+func listenPort() int {
+	sPort := os.Getenv(EnvSocks5ListenerPort)
 	if sPort == "" {
-		return listenAddr(DefaultSocks5Port)
+		return DefaultSocks5Port
 	}
 
 	port, err := strconv.Atoi(sPort)
 	if err != nil {
 		log.Warnf("invalid socks5 listener port, unable to convert it to int, falling back to default: %d", DefaultSocks5Port)
-		return listenAddr(DefaultSocks5Port)
+		return DefaultSocks5Port
 	}
 	if port < 1 || port > 65535 {
 		log.Warnf("invalid socks5 listener port, it should be in the range 1-65535, falling back to default: %d", DefaultSocks5Port)
-		return listenAddr(DefaultSocks5Port)
+		return DefaultSocks5Port
 	}
 
-	return listenAddr(port)
-}
-
-func listenAddr(port int) string {
-	return fmt.Sprintf("0.0.0.0:%d", port)
+	return port
 }

client/iface/netstack/env_test.go

@@ -0,0 +1,63 @@
+//go:build !js
+
+package netstack
+
+import (
+	"net"
+	"strconv"
+	"testing"
+)
+
+func TestListenAddr_DefaultsToLoopback(t *testing.T) {
+	// No env overrides: must bind loopback, never all interfaces.
+	got := ListenAddr()
+	want := net.JoinHostPort("127.0.0.1", strconv.Itoa(DefaultSocks5Port))
+	if got != want {
+		t.Fatalf("ListenAddr() = %q, want %q", got, want)
+	}
+}
+
+func TestListenAddr_AddressOverride(t *testing.T) {
+	tests := []struct {
+		name string
+		env  string
+		want string
+	}{
+		{name: "valid override honored", env: "0.0.0.0", want: "0.0.0.0"},
+		{name: "valid specific ip honored", env: "10.0.0.5", want: "10.0.0.5"},
+		{name: "ipv6 loopback bracketed", env: "::1", want: "::1"},
+		{name: "invalid falls back to loopback", env: "not-an-ip", want: "127.0.0.1"},
+		{name: "empty falls back to loopback", env: "", want: "127.0.0.1"},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(EnvSocks5ListenerAddress, tc.env)
+			want := net.JoinHostPort(tc.want, strconv.Itoa(DefaultSocks5Port))
+			if got := ListenAddr(); got != want {
+				t.Fatalf("ListenAddr() = %q, want %q", got, want)
+			}
+		})
+	}
+}
+
+func TestListenAddr_PortOverride(t *testing.T) {
+	tests := []struct {
+		name string
+		env  string
+		want int
+	}{
+		{name: "valid port honored", env: "1081", want: 1081},
+		{name: "non-numeric falls back", env: "abc", want: DefaultSocks5Port},
+		{name: "out of range falls back", env: "70000", want: DefaultSocks5Port},
+		{name: "zero falls back", env: "0", want: DefaultSocks5Port},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(EnvSocks5ListenerPort, tc.env)
+			want := net.JoinHostPort("127.0.0.1", strconv.Itoa(tc.want))
+			if got := ListenAddr(); got != want {
+				t.Fatalf("ListenAddr() = %q, want %q", got, want)
+			}
+		})
+	}
+}

client/internal/auth/auth.go

@@ -351,7 +351,7 @@ func (a *Auth) setSystemInfoFlags(info *system.Info) {
 		a.config.BlockLANAccess,
 		a.config.BlockInbound,
 		a.config.DisableIPv6,
-		a.config.LazyConnectionEnabled,
+		a.config.SyncMessageVersion,
 		a.config.EnableSSHRoot,
 		a.config.EnableSSHSFTP,
 		a.config.EnableSSHLocalPortForwarding,

client/internal/auth/device_flow.go

@@ -259,12 +259,18 @@ func (d *DeviceAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowIn
 	ticker := time.NewTicker(interval)
 	defer ticker.Stop()
 
+	log.Infof("device flow: waiting for user authorization, polling token endpoint every %s, code expires in %s", interval, timeout)
+
+	start := time.Now()
+	polls := 0
+
 	for {
 		select {
 		case <-waitCtx.Done():
 			return TokenInfo{}, waitCtx.Err()
 		case <-ticker.C:
 
+			polls++
 			tokenResponse, err := d.requestToken(info)
 			if err != nil {
 				return TokenInfo{}, fmt.Errorf("parsing token response failed with error: %v", err)
@@ -272,10 +278,12 @@ func (d *DeviceAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowIn
 
 			if tokenResponse.Error != "" {
 				if tokenResponse.Error == "authorization_pending" {
+					log.Tracef("device flow: authorization still pending after poll %d", polls)
 					continue
 				} else if tokenResponse.Error == "slow_down" {
 					interval += (3 * time.Second)
 					ticker.Reset(interval)
+					log.Infof("device flow: IdP requested slow_down, polling interval increased to %s", interval)
 					continue
 				}
 
@@ -291,11 +299,12 @@ func (d *DeviceAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowIn
 				UseIDToken:   d.providerConfig.UseIDToken,
 			}
 
-			err = isValidAccessToken(tokenInfo.GetTokenToUse(), d.providerConfig.Audience)
+			err = validateTokenAudience(tokenInfo.GetTokenToUse(), d.providerConfig.Audience)
 			if err != nil {
 				return TokenInfo{}, fmt.Errorf("validate access token failed with error: %v", err)
 			}
 
+			log.Infof("device flow: user authorization confirmed after %d polls in %s", polls, time.Since(start).Round(time.Second))
 			return tokenInfo, err
 		}
 	}

client/internal/auth/pkce_flow.go

@@ -188,6 +188,8 @@ func (p *PKCEAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowInfo
 	waitCtx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
 
+	log.Infof("pkce flow: waiting for authorization callback on %s, timeout %s", p.oAuthConfig.RedirectURL, timeout)
+
 	tokenChan := make(chan *oauth2.Token, 1)
 	errChan := make(chan error, 1)
 
@@ -221,6 +223,7 @@ func (p *PKCEAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowInfo
 func (p *PKCEAuthorizationFlow) startServer(server *http.Server, tokenChan chan<- *oauth2.Token, errChan chan<- error) {
 	mux := http.NewServeMux()
 	mux.HandleFunc("/", func(w http.ResponseWriter, req *http.Request) {
+		log.Infof("pkce flow: received authorization callback from IdP")
 		cert := p.providerConfig.ClientCertPair
 		if cert != nil {
 			tr := &http.Transport{
@@ -271,11 +274,18 @@ func (p *PKCEAuthorizationFlow) handleRequest(req *http.Request) (*oauth2.Token,
 		return nil, fmt.Errorf("authentication failed: missing code")
 	}
 
-	return p.oAuthConfig.Exchange(
+	exchangeStart := time.Now()
+	token, err := p.oAuthConfig.Exchange(
 		req.Context(),
 		code,
 		oauth2.SetAuthURLParam("code_verifier", p.codeVerifier),
 	)
+	if err != nil {
+		return nil, err
+	}
+
+	log.Infof("pkce flow: authorization code exchanged for token in %s", time.Since(exchangeStart).Round(time.Millisecond))
+	return token, nil
 }
 
 func (p *PKCEAuthorizationFlow) parseOAuthToken(token *oauth2.Token) (TokenInfo, error) {
@@ -296,7 +306,7 @@ func (p *PKCEAuthorizationFlow) parseOAuthToken(token *oauth2.Token) (TokenInfo,
 		audience = p.providerConfig.ClientID
 	}
 
-	if err := isValidAccessToken(tokenInfo.GetTokenToUse(), audience); err != nil {
+	if err := validateTokenAudience(tokenInfo.GetTokenToUse(), audience); err != nil {
 		return TokenInfo{}, fmt.Errorf("authentication failed: invalid access token - %w", err)
 	}
 
@@ -310,6 +320,11 @@ func (p *PKCEAuthorizationFlow) parseOAuthToken(token *oauth2.Token) (TokenInfo,
 	return tokenInfo, nil
 }
 
+// parseEmailFromIDToken extracts the email (or name) claim from an ID token
+// without verifying its signature. The value is best-effort and used only as a
+// UX convenience (login hint prefill and display); it never drives an
+// authorization decision. The authoritative identity is established server-side
+// from the signature-verified token.
 func parseEmailFromIDToken(token string) (string, error) {
 	parts := strings.Split(token, ".")
 	if len(parts) < 2 {

client/internal/auth/sessionwatch/watcher.go

@@ -24,11 +24,7 @@ import (
 )
 
 const (
-	// Skew tolerates a small clock difference between the management
-	// server and this peer before treating a deadline as "in the past".
-	// Slightly above typical NTP drift; tight enough that the UI doesn't
-	// paint a stale expiry as if it were valid.
-	Skew = 30 * time.Second
+	maxPastHorizon = 30 * 24 * time.Hour
 
 	// maxDeadlineHorizon caps how far in the future an accepted deadline
 	// can sit. A timestamp beyond this is almost certainly a protocol
@@ -57,7 +53,7 @@ var (
 	ErrDeadlineTooFarFuture = errors.New("session deadline too far in the future")
 
 	// ErrDeadlineInPast is returned by Update when the supplied deadline
-	// is more than Skew in the past.
+	// is more than maxPastHorizon in the past.
 	ErrDeadlineInPast = errors.New("session deadline in the past")
 )
 
@@ -66,15 +62,14 @@ var (
 // for deadline change/clear, PublishEvent for the two warnings); tests pass
 // a fake recorder so the same surface is observable without an engine.
 //
-// The watcher is the single owner of the deadline propagated to the
-// recorder: every set, clear, sanity-check rejection and Close routes the
-// value through SetSessionExpiresAt, so the SubscribeStatus snapshot the UI
-// reads can never drift from the watcher's timer state. (SetSessionExpiresAt
-// fans out its own state-change notification, so no separate notify is
-// needed.) The recorder is server-scoped and outlives this engine-scoped
-// watcher — without the Close-time clear a teardown (Down, or the Down+Up of
-// a profile switch) would leave the next session showing the previous one's
-// stale "expires in" value.
+// While the watcher runs, it owns the deadline propagated to the recorder:
+// every set, clear and sanity-check rejection routes the value through
+// SetSessionExpiresAt, so the SubscribeStatus snapshot the UI reads can
+// never drift from the watcher's timer state. (SetSessionExpiresAt fans
+// out its own state-change notification, so no separate notify is needed.)
+// The recorder is server-scoped and outlives this engine-scoped watcher;
+// Close deliberately leaves the recorder value in place so transient engine
+// restarts don't blank it — the client run loop clears it on real teardown.
 //
 // PublishEvent's signature mirrors peer.Status.PublishEvent: the watcher
 // composes the metadata internally so the wire format (MetaSession*) is
@@ -135,10 +130,13 @@ func NewWithLeads(lead, final time.Duration, recorder StatusRecorder) *Watcher {
 // was disabled).
 //
 // Same-value updates are no-ops. A different non-zero value cancels any
-// pending timer, resets the "already fired" guard, and arms a new one.
+// pending timer, resets the "already fired" guards, and — when the
+// deadline lies in the future — arms fresh warning timers. A deadline
+// already in the past (within maxPastHorizon) is recorded as-is with no
+// timers: the session has expired and consumers render it that way.
 //
 // Returns one of the sentinel Err* values when the deadline fails the
-// sanity checks (pre-epoch, far future, or in the past beyond Skew).
+// sanity checks (pre-epoch, far future, or past beyond maxPastHorizon).
 // In every error case the watcher first clears its state so it stays
 // consistent with what the caller will push into its other sinks (e.g.
 // applySessionDeadline forces a zero deadline into the status recorder
@@ -163,7 +161,7 @@ func (w *Watcher) Update(deadline time.Time) error {
 	case deadline.After(now.Add(maxDeadlineHorizon)):
 		w.clearLocked()
 		return fmt.Errorf("%w: %v", ErrDeadlineTooFarFuture, deadline)
-	case deadline.Before(now.Add(-Skew)):
+	case deadline.Before(now.Add(-maxPastHorizon)):
 		w.clearLocked()
 		return fmt.Errorf("%w: %v (now=%v)", ErrDeadlineInPast, deadline, now)
 	}
@@ -183,7 +181,9 @@ func (w *Watcher) Update(deadline time.Time) error {
 	w.finalFiredAt = time.Time{}
 	w.dismissedAt = time.Time{}
 
-	w.armTimerLocked(deadline)
+	if deadline.After(now) {
+		w.armTimerLocked(deadline)
+	}
 	recorder := w.recorder
 	w.mu.Unlock()
 	if recorder != nil {
@@ -227,30 +227,25 @@ func (w *Watcher) Dismiss() {
 	log.Infof("auth session final-warning dismissed for deadline %s", w.current.Format(time.RFC3339))
 }
 
-// Close stops any pending timer and drops the deadline on the status
-// recorder. Update calls after Close are ignored. Clearing the recorder
-// here is what keeps a teardown (Down, or the Down+Up of a profile switch)
-// from leaving the next session showing this one's stale "expires in"
-// value — the recorder is server-scoped and outlives this engine-scoped
-// watcher, so nothing else drops the anchor on teardown.
+// Close stops any pending timer. Update calls after Close are ignored.
+// The recorder keeps its deadline: the watcher is engine-scoped and closes
+// on every engine restart (network change, sleep/wake, stream errors)
+// while the SSO deadline stays valid across those, so clearing here would
+// blank the UI's "expires in" row on every transient reconnect. The
+// client run loop clears the server-scoped recorder when it exits for
+// real (Down, profile switch, permanent login failure).
 func (w *Watcher) Close() {
 	w.mu.Lock()
+	defer w.mu.Unlock()
 	if w.closed {
-		w.mu.Unlock()
 		return
 	}
 	w.closed = true
 	w.stopTimerLocked()
-	hadDeadline := !w.current.IsZero()
 	w.current = time.Time{}
 	w.firedAt = time.Time{}
 	w.finalFiredAt = time.Time{}
 	w.dismissedAt = time.Time{}
-	recorder := w.recorder
-	w.mu.Unlock()
-	if recorder != nil && hadDeadline {
-		recorder.SetSessionExpiresAt(time.Time{})
-	}
 }
 
 // clearLocked drops the tracked deadline and notifies the recorder so

client/internal/auth/sessionwatch/watcher_test.go

@@ -224,11 +224,13 @@ func TestNewDeadlineCancelsPriorTimer(t *testing.T) {
 
 func TestRefreshAfterFireArmsNewWarning(t *testing.T) {
 	r := &fakeRecorder{}
-	lead := 30 * time.Millisecond
+	lead := 150 * time.Millisecond
 	w := newWatcher(lead, r)
 	defer w.Close()
 
-	first := time.Now().Add(50 * time.Millisecond)
+	// Warning fires ~20ms in; the deadline itself stays 150ms away so the
+	// replacement below lands well before it.
+	first := time.Now().Add(170 * time.Millisecond)
 	_ = w.Update(first)
 
 	// Wait for stateChange + warning of the first cycle.
@@ -306,7 +308,29 @@ func TestUpdateRejectsTooFarFuture(t *testing.T) {
 	}
 }
 
-func TestUpdateInPastClearsDeadline(t *testing.T) {
+func TestUpdateRecentPastRecordedAsExpired(t *testing.T) {
+	r := &fakeRecorder{}
+	w := newWatcher(50*time.Millisecond, r)
+	defer w.Close()
+
+	d := time.Now().Add(-1 * time.Hour)
+	if err := w.Update(d); err != nil {
+		t.Fatalf("recent-past Update should succeed, got %v", err)
+	}
+	if !w.Deadline().Equal(d) {
+		t.Fatalf("expected deadline to be recorded, got %v want %v", w.Deadline(), d)
+	}
+	if got := r.deadline(); !got.Equal(d) {
+		t.Fatalf("recorder deadline = %v, want %v", got, d)
+	}
+
+	time.Sleep(80 * time.Millisecond)
+	if n := countWhere(r.snapshot(), func(e event) bool { return e.kind == publish }); n != 0 {
+		t.Fatalf("no warning events may fire for an already-past deadline, got %+v", r.snapshot())
+	}
+}
+
+func TestUpdateAncientPastRejected(t *testing.T) {
 	r := &fakeRecorder{}
 	w := newWatcher(50*time.Millisecond, r)
 	defer w.Close()
@@ -318,12 +342,12 @@ func TestUpdateInPastClearsDeadline(t *testing.T) {
 	// Drain the stateChange from the seed.
 	waitForEvents(t, r, 1)
 
-	err := w.Update(time.Now().Add(-1 * time.Hour))
+	err := w.Update(time.Now().Add(-31 * 24 * time.Hour))
 	if !errors.Is(err, ErrDeadlineInPast) {
 		t.Fatalf("want ErrDeadlineInPast, got %v", err)
 	}
 	if !w.Deadline().IsZero() {
-		t.Fatalf("in-past update must clear the deadline, got %v", w.Deadline())
+		t.Fatalf("rejected ancient-past update must clear the deadline, got %v", w.Deadline())
 	}
 	events := waitForEvents(t, r, 2)
 	if events[1].kind != stateChange {
@@ -331,39 +355,25 @@ func TestUpdateInPastClearsDeadline(t *testing.T) {
 	}
 }
 
-func TestUpdateWithinSkewAccepted(t *testing.T) {
-	r := &fakeRecorder{}
-	w := newWatcher(50*time.Millisecond, r)
-	defer w.Close()
-
-	// 5 seconds in the past is within the 30s Skew tolerance — accept it.
-	d := time.Now().Add(-5 * time.Second)
-	if err := w.Update(d); err != nil {
-		t.Fatalf("within-skew Update should succeed, got %v", err)
-	}
-	if !w.Deadline().Equal(d) {
-		t.Fatalf("expected deadline to be applied, got %v want %v", w.Deadline(), d)
-	}
-}
-
 func TestCloseSilencesUpdates(t *testing.T) {
 	r := &fakeRecorder{}
 	w := newWatcher(50*time.Millisecond, r)
 	w.Close()
 
-	_ = w.Update(time.Now().Add(time.Hour))
-
-	time.Sleep(20 * time.Millisecond)
+	if err := w.Update(time.Now().Add(time.Hour)); err != nil {
+		t.Fatalf("Update after Close: want nil, got %v", err)
+	}
 	if got := r.snapshot(); len(got) != 0 {
 		t.Fatalf("expected no events after Close, got %+v", got)
 	}
 }
 
-// TestCloseClearsRecorderDeadline pins the profile-switch fix: a watcher
-// holding a live deadline must zero the recorder on Close so the next
-// engine's watcher (and the UI reading the shared server-scoped recorder)
-// doesn't start out showing the previous session's stale "expires in".
-func TestCloseClearsRecorderDeadline(t *testing.T) {
+// TestCloseKeepsRecorderDeadline pins the reconnect-flap fix: the watcher
+// closes on every engine restart (network change, sleep/wake) while the
+// SSO deadline stays valid across those, so Close must leave the
+// server-scoped recorder's value in place. The client run loop clears the
+// recorder when it exits for real.
+func TestCloseKeepsRecorderDeadline(t *testing.T) {
 	r := &fakeRecorder{}
 	w := newWatcher(time.Hour, r)
 
@@ -377,8 +387,8 @@ func TestCloseClearsRecorderDeadline(t *testing.T) {
 
 	w.Close()
 
-	if got := r.deadline(); !got.IsZero() {
-		t.Fatalf("recorder deadline after Close = %v, want zero", got)
+	if got := r.deadline(); !got.Equal(d) {
+		t.Fatalf("recorder deadline after Close = %v, want %v", got, d)
 	}
 }
 

client/internal/auth/util.go

@@ -20,14 +20,26 @@ func randomBytesInHex(count int) (string, error) {
 	return hex.EncodeToString(buf), nil
 }
 
-// isValidAccessToken is a simple validation of the access token
-func isValidAccessToken(token string, audience string) error {
+// validateTokenAudience checks that the token is a well-formed JWT whose
+// audience claim matches the expected audience.
+//
+// It does NOT verify the token's cryptographic signature and therefore must not
+// be treated as an authenticity check. The token is obtained by the client
+// directly from the IdP token endpoint over TLS, and its signature is verified
+// server-side by the management server against the IdP's JWKS
+// (see shared/auth/jwt/validator.go). This function is only a client-side
+// sanity check that the returned token targets the expected audience.
+func validateTokenAudience(token string, audience string) error {
 	if token == "" {
 		return fmt.Errorf("token received is empty")
 	}
 
-	encodedClaims := strings.Split(token, ".")[1]
-	claimsString, err := base64.RawURLEncoding.DecodeString(encodedClaims)
+	parts := strings.Split(token, ".")
+	if len(parts) != 3 {
+		return fmt.Errorf("token is not a well-formed JWT")
+	}
+
+	claimsString, err := base64.RawURLEncoding.DecodeString(parts[1])
 	if err != nil {
 		return err
 	}

client/internal/auth/util_test.go

@@ -0,0 +1,108 @@
+package auth
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"testing"
+)
+
+// makeJWT builds an unsigned JWT-shaped string (header.payload.signature) with
+// the given claims payload. The signature part is arbitrary because
+// validateTokenAudience intentionally does not verify it.
+func makeJWT(t *testing.T, claims map[string]interface{}) string {
+	t.Helper()
+	header := base64.RawURLEncoding.EncodeToString([]byte(`{"alg":"RS256","typ":"JWT"}`))
+	payloadBytes, err := json.Marshal(claims)
+	if err != nil {
+		t.Fatalf("marshal claims: %v", err)
+	}
+	payload := base64.RawURLEncoding.EncodeToString(payloadBytes)
+	return header + "." + payload + ".unverified-signature"
+}
+
+func TestValidateTokenAudience(t *testing.T) {
+	tests := []struct {
+		name     string
+		token    string
+		audience string
+		wantErr  bool
+	}{
+		{
+			name:     "empty token",
+			token:    "",
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "not a JWT - no dots",
+			token:    "notajwt",
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "not a JWT - two parts only",
+			token:    "header.payload",
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "matching string audience",
+			token:    makeJWT(t, map[string]interface{}{"aud": "netbird"}),
+			audience: "netbird",
+			wantErr:  false,
+		},
+		{
+			name:     "mismatching string audience",
+			token:    makeJWT(t, map[string]interface{}{"aud": "other"}),
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "matching audience in array",
+			token:    makeJWT(t, map[string]interface{}{"aud": []interface{}{"other", "netbird"}}),
+			audience: "netbird",
+			wantErr:  false,
+		},
+		{
+			name:     "mismatching audience array",
+			token:    makeJWT(t, map[string]interface{}{"aud": []interface{}{"a", "b"}}),
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "missing audience claim",
+			token:    makeJWT(t, map[string]interface{}{"sub": "user"}),
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "invalid base64 payload",
+			token:    "header.!!!not-base64!!!.sig",
+			audience: "netbird",
+			wantErr:  true,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			err := validateTokenAudience(tc.token, tc.audience)
+			if tc.wantErr && err == nil {
+				t.Fatalf("expected error, got nil")
+			}
+			if !tc.wantErr && err != nil {
+				t.Fatalf("expected no error, got %v", err)
+			}
+		})
+	}
+}
+
+// TestValidateTokenAudienceNoPanic guards the regression where a non-empty
+// token without the JWT dot structure caused an index-out-of-range panic.
+func TestValidateTokenAudienceNoPanic(t *testing.T) {
+	inputs := []string{"a", ".", "a.", "aaaa", "no-dots-here"}
+	for _, in := range inputs {
+		if err := validateTokenAudience(in, "netbird"); err == nil {
+			t.Fatalf("expected error for malformed token %q, got nil", in)
+		}
+	}
+}

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:
@@ -24,47 +34,69 @@ import (
 // - Handling connection establishment based on peer signaling
 //
 // The implementation is not thread-safe; it is protected by engine.syncMsgMux.
+// The only exception is ActivatePeer, which is safe for concurrent use so the
+// DNS warm-up path can call it without contending on the engine mutex.
 type ConnMgr struct {
 	peerStore        *peerstore.Store
 	statusRecorder   *peer.Status
 	iface            lazyconn.WGIface
-	enabledLocally   bool
+	force            lazyForce
 	rosenpassEnabled bool
 
 	lazyConnMgr *manager.Manager
+	// lazyConnMgrMu guards the lazyConnMgr pointer for readers outside the
+	// engine loop (ActivatePeer). Writers hold it in addition to
+	// engine.syncMsgMux; all other reads stay under engine.syncMsgMux only.
+	lazyConnMgrMu sync.RWMutex
+
+	// reconcileRoutedIPs re-applies a peer's routed allowed IPs after its lazy wake endpoint is
+	// (re)armed (Mode A at arm time). Injected by the engine; nil disables the reconcile.
+	reconcileRoutedIPs func(peerKey string) error
 
 	wg            sync.WaitGroup
 	lazyCtx       context.Context
 	lazyCtxCancel context.CancelFunc
 }
 
+// SetRoutedIPsReconciler injects the callback used to re-apply a peer's routed allowed IPs when
+// its lazy wake endpoint is (re)armed. Must be called before the lazy manager starts.
+func (e *ConnMgr) SetRoutedIPsReconciler(fn func(peerKey string) error) {
+	e.reconcileRoutedIPs = fn
+}
+
 func NewConnMgr(engineConfig *EngineConfig, statusRecorder *peer.Status, peerStore *peerstore.Store, iface lazyconn.WGIface) *ConnMgr {
 	e := &ConnMgr{
 		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 +108,8 @@ 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
 	}
 
@@ -89,15 +121,17 @@ 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
 		}
 
-		log.Warnf("lazy connection manager is enabled by management feature flag")
+		log.Infof("lazy connection manager is enabled by the management feature flag")
 		e.initLazyManager(ctx)
 		e.statusRecorder.UpdateLazyConnection(true)
 		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")
@@ -107,37 +141,6 @@ func (e *ConnMgr) UpdatedRemoteFeatureFlag(ctx context.Context, enabled bool) er
 	}
 }
 
-// SetLocalLazyConn applies a local lazy connection override (UI / CLI / env).
-// While enabledLocally is true, UpdatedRemoteFeatureFlag (management sync) is a
-// no-op, so the local setting wins until it is turned off again.
-func (e *ConnMgr) SetLocalLazyConn(ctx context.Context, enabled bool) error {
-	e.enabledLocally = enabled
-
-	if enabled {
-		if e.lazyConnMgr != nil {
-			return nil
-		}
-
-		if e.rosenpassEnabled {
-			log.Warnf("rosenpass connection manager is enabled, lazy connection manager will not be started")
-			return nil
-		}
-
-		log.Infof("lazy connection manager is enabled locally")
-		e.initLazyManager(ctx)
-		e.statusRecorder.UpdateLazyConnection(true)
-		return e.addPeersToLazyConnManager()
-	}
-
-	if e.lazyConnMgr == nil {
-		return nil
-	}
-	log.Infof("lazy connection manager is disabled locally")
-	e.closeManager(ctx)
-	e.statusRecorder.UpdateLazyConnection(false)
-	return nil
-}
-
 // UpdateRouteHAMap updates the route HA mappings in the lazy connection manager
 func (e *ConnMgr) UpdateRouteHAMap(haMap route.HAMap) {
 	if !e.isStartedWithLazyMgr() {
@@ -251,12 +254,20 @@ func (e *ConnMgr) RemovePeerConn(peerKey string) {
 	conn.Log.Infof("removed peer from lazy conn manager")
 }
 
+// ActivatePeer wakes an idle lazy connection. Unlike the rest of ConnMgr it is
+// safe for concurrent use: the lazy manager pointer is read under lazyConnMgrMu
+// and the manager itself is internally synchronized, so callers outside the
+// engine loop (DNS warm-up) do not need engine.syncMsgMux.
 func (e *ConnMgr) ActivatePeer(ctx context.Context, conn *peer.Conn) {
-	if !e.isStartedWithLazyMgr() {
+	e.lazyConnMgrMu.RLock()
+	lazyConnMgr := e.lazyConnMgr
+	started := lazyConnMgr != nil && e.lazyCtxCancel != nil
+	e.lazyConnMgrMu.RUnlock()
+	if !started {
 		return
 	}
 
-	if found := e.lazyConnMgr.ActivatePeer(conn.GetKey()); found {
+	if found := lazyConnMgr.ActivatePeer(conn.GetKey()); found {
 		if err := conn.Open(ctx); err != nil {
 			conn.Log.Errorf("failed to open connection: %v", err)
 		}
@@ -281,16 +292,22 @@ func (e *ConnMgr) Close() {
 
 	e.lazyCtxCancel()
 	e.wg.Wait()
+
+	e.lazyConnMgrMu.Lock()
 	e.lazyConnMgr = nil
+	e.lazyConnMgrMu.Unlock()
 }
 
 func (e *ConnMgr) initLazyManager(engineCtx context.Context) {
 	cfg := manager.Config{
 		InactivityThreshold: inactivityThresholdEnv(),
+		ReconcileAllowedIPs: e.reconcileRoutedIPs,
 	}
-	e.lazyConnMgr = manager.NewManager(cfg, engineCtx, e.peerStore, e.iface)
 
+	e.lazyConnMgrMu.Lock()
+	e.lazyConnMgr = manager.NewManager(cfg, engineCtx, e.peerStore, e.iface)
 	e.lazyCtx, e.lazyCtxCancel = context.WithCancel(engineCtx)
+	e.lazyConnMgrMu.Unlock()
 
 	e.wg.Add(1)
 	go func() {
@@ -329,7 +346,10 @@ func (e *ConnMgr) closeManager(ctx context.Context) {
 
 	e.lazyCtxCancel()
 	e.wg.Wait()
+
+	e.lazyConnMgrMu.Lock()
 	e.lazyConnMgr = nil
+	e.lazyConnMgrMu.Unlock()
 
 	for _, peerID := range e.peerStore.PeersPubKey() {
 		e.peerStore.PeerConnOpen(ctx, peerID)
@@ -340,6 +360,25 @@ 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 +1,106 @@
+package internal
+
+import (
+	"context"
+	"net"
+	"net/netip"
+	"os"
+	"sync"
+	"testing"
+	"time"
+
+	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
+
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+	"github.com/netbirdio/netbird/client/internal/lazyconn"
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+	"github.com/netbirdio/netbird/monotime"
+)
+
+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)
+			}
+		})
+	}
+}
+
+type mockLazyWGIface struct{}
+
+func (mockLazyWGIface) RemovePeer(string) error { return nil }
+func (mockLazyWGIface) UpdatePeer(string, []netip.Prefix, time.Duration, *net.UDPAddr, *wgtypes.Key) error {
+	return nil
+}
+func (mockLazyWGIface) IsUserspaceBind() bool                    { return false }
+func (mockLazyWGIface) Address() wgaddr.Address                  { return wgaddr.Address{} }
+func (mockLazyWGIface) LastActivities() map[string]monotime.Time { return nil }
+func (mockLazyWGIface) MTU() uint16                              { return 1280 }
+
+// TestConnMgr_ActivatePeerConcurrentWithLifecycle exercises ActivatePeer from
+// non-engine goroutines (the DNS warm-up path) racing the manager lifecycle,
+// which stays on the engine loop. Run with -race: it fails if ActivatePeer
+// still requires engine.syncMsgMux for safety.
+func TestConnMgr_ActivatePeerConcurrentWithLifecycle(t *testing.T) {
+	t.Setenv(lazyconn.EnvLazyConn, "on")
+
+	status := peer.NewRecorder("https://mgm")
+	store := peerstore.NewConnStore()
+	connMgr := NewConnMgr(&EngineConfig{}, status, store, mockLazyWGIface{})
+
+	conn := newTestPeerConn(t, "peerA")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	connMgr.Start(ctx)
+
+	done := make(chan struct{})
+	var wg sync.WaitGroup
+	for range 4 {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for {
+				select {
+				case <-done:
+					return
+				default:
+					connMgr.ActivatePeer(ctx, conn)
+				}
+			}
+		}()
+	}
+
+	// Let the activators spin against the started manager, then tear it down
+	// underneath them and let them spin against the stopped manager.
+	time.Sleep(100 * time.Millisecond)
+	connMgr.Close()
+	time.Sleep(50 * time.Millisecond)
+
+	close(done)
+	wg.Wait()
+}

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"
@@ -256,7 +257,10 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 		log.Errorf("failed to clean up temporary installer file: %v", err)
 	}
 
-	defer c.statusRecorder.ClientStop()
+	defer func() {
+		c.statusRecorder.SetSessionExpiresAt(time.Time{})
+		c.statusRecorder.ClientStop()
+	}()
 	operation := func() error {
 		// if context cancelled we not start new backoff cycle
 		if c.ctx.Err() != nil {
@@ -617,8 +621,9 @@ func createEngineConfig(key wgtypes.Key, config *profilemanager.Config, peerConf
 		BlockLANAccess:      config.BlockLANAccess,
 		BlockInbound:        config.BlockInbound,
 		DisableIPv6:         config.DisableIPv6,
+		SyncMessageVersion:  config.SyncMessageVersion,
 
-		LazyConnectionEnabled: config.LazyConnectionEnabled,
+		LazyConnection: lazyconn.ParseState(config.LazyConnection),
 
 		MTU:     selectMTU(config.MTU, peerConfig.Mtu),
 		LogPath: logPath,
@@ -692,7 +697,7 @@ func loginToManagement(ctx context.Context, client mgm.Client, pubSSHKey []byte,
 		config.BlockLANAccess,
 		config.BlockInbound,
 		config.DisableIPv6,
-		config.LazyConnectionEnabled,
+		config.SyncMessageVersion,
 		config.EnableSSHRoot,
 		config.EnableSSHSFTP,
 		config.EnableSSHLocalPortForwarding,

client/internal/debug/debug.go

@@ -480,7 +480,6 @@ func (g *BundleGenerator) addStatus() error {
 
 		fullStatus := g.statusRecorder.GetFullStatus()
 		protoFullStatus := nbstatus.ToProtoFullStatus(fullStatus)
-		protoFullStatus.Events = g.statusRecorder.GetEventHistory()
 		overview := nbstatus.ConvertToStatusOutputOverview(protoFullStatus, nbstatus.ConvertOptions{
 			Anonymize:     g.anonymize,
 			ProfileName:   profName,
@@ -677,6 +676,7 @@ func (g *BundleGenerator) addCommonConfigFields(configContent *strings.Builder)
 	configContent.WriteString(fmt.Sprintf("BlockLANAccess: %v\n", g.internalConfig.BlockLANAccess))
 	configContent.WriteString(fmt.Sprintf("BlockInbound: %v\n", g.internalConfig.BlockInbound))
 	configContent.WriteString(fmt.Sprintf("DisableIPv6: %v\n", g.internalConfig.DisableIPv6))
+	configContent.WriteString(fmt.Sprintf("SyncMessageVersion: %v\n", g.internalConfig.SyncMessageVersion))
 
 	if g.internalConfig.DisableNotifications != nil {
 		configContent.WriteString(fmt.Sprintf("DisableNotifications: %v\n", *g.internalConfig.DisableNotifications))
@@ -695,7 +695,7 @@ 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

@@ -885,8 +885,10 @@ func TestAddConfig_AllFieldsCovered(t *testing.T) {
 		DNSRouteInterval:              5 * time.Second,
 		ClientCertPath:                "/tmp/cert",
 		ClientCertKeyPath:             "/tmp/key",
-		LazyConnectionEnabled:         true,
+		LazyConnection:                "on",
 		MTU:                           1280,
+		DisableIPv6:                   true,
+		SyncMessageVersion:            func(v int) *int { return &v }(1),
 	}
 
 	for _, anonymize := range []bool{false, true} {

client/internal/dns/local/local.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"net"
 	"net/netip"
+	"os"
 	"slices"
 	"strings"
 	"sync"
@@ -36,7 +37,43 @@ type resolver interface {
 // record is left alone (it points at something outside our mesh, e.g.
 // a non-peer upstream).
 type PeerConnectivity interface {
-	IsConnectedByIP(ip string) (known, connected bool)
+	IsConnectedByIP(ip netip.Addr) (known, connected bool)
+}
+
+// PeerActivator wakes lazy-connection peers on demand. The local resolver calls
+// it with the tunnel IPs an answer points at, so a peer that is idle (lazily
+// disconnected) starts connecting at DNS-resolution time rather than racing the
+// client's first request packet. nil disables warm-up.
+type PeerActivator interface {
+	// ActivatePeersByIP triggers wake-up for the peer(s) owning addrs and blocks
+	// until one is connected or ctx (a short per-query budget) expires. It is a
+	// fast no-op for unknown or already-connected addresses.
+	ActivatePeersByIP(ctx context.Context, addrs []netip.Addr)
+}
+
+const (
+	defaultLazyWarmupTimeout = 2 * time.Second
+	envLazyWarmupTimeout     = "NB_DNS_LAZY_WARMUP_TIMEOUT"
+)
+
+// lazyWarmupTimeoutFromEnv returns the per-query budget for waking a
+// lazy-connection peer a DNS answer points at. Tunable via
+// NB_DNS_LAZY_WARMUP_TIMEOUT (a Go duration). Parsed once at construction time.
+func lazyWarmupTimeoutFromEnv() time.Duration {
+	v := os.Getenv(envLazyWarmupTimeout)
+	if v == "" {
+		return defaultLazyWarmupTimeout
+	}
+	d, err := time.ParseDuration(v)
+	if err != nil {
+		log.Warnf("invalid %s value %q, using default %s: %v", envLazyWarmupTimeout, v, defaultLazyWarmupTimeout, err)
+		return defaultLazyWarmupTimeout
+	}
+	if d <= 0 {
+		log.Warnf("non-positive %s value %q, using default %s", envLazyWarmupTimeout, v, defaultLazyWarmupTimeout)
+		return defaultLazyWarmupTimeout
+	}
+	return d
 }
 
 type Resolver struct {
@@ -51,6 +88,12 @@ type Resolver struct {
 	// filter and preserves the legacy "return whatever is registered"
 	// behaviour for callers that never wire a status source.
 	peerConn PeerConnectivity
+	// peerActivator, when non-nil, is called at resolution time to warm the
+	// lazy connection to the peer(s) an answer points at. nil disables warm-up.
+	peerActivator PeerActivator
+	// warmupTimeout is the per-query budget for the lazy-connection warm-up
+	// wait, resolved from the environment once at construction time.
+	warmupTimeout time.Duration
 
 	ctx    context.Context
 	cancel context.CancelFunc
@@ -59,11 +102,12 @@ type Resolver struct {
 func NewResolver() *Resolver {
 	ctx, cancel := context.WithCancel(context.Background())
 	return &Resolver{
-		records: make(map[dns.Question][]dns.RR),
-		domains: make(map[domain.Domain]struct{}),
-		zones:   make(map[domain.Domain]bool),
-		ctx:     ctx,
-		cancel:  cancel,
+		records:       make(map[dns.Question][]dns.RR),
+		domains:       make(map[domain.Domain]struct{}),
+		zones:         make(map[domain.Domain]bool),
+		warmupTimeout: lazyWarmupTimeoutFromEnv(),
+		ctx:           ctx,
+		cancel:        cancel,
 	}
 }
 
@@ -76,6 +120,14 @@ func (d *Resolver) SetPeerConnectivity(p PeerConnectivity) {
 	d.peerConn = p
 }
 
+// SetPeerActivator wires the DNS-time lazy-connection warm-up. Pass nil to
+// disable. Safe to call multiple times; the latest value wins.
+func (d *Resolver) SetPeerActivator(a PeerActivator) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	d.peerActivator = a
+}
+
 func (d *Resolver) MatchSubdomains() bool {
 	return true
 }
@@ -122,6 +174,9 @@ func (d *Resolver) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	replyMessage.RecursionAvailable = true
 
 	result := d.lookupRecords(logger, question)
+	// Warm before filtering: activation flips a lazily-idle target to connected,
+	// which then lets it survive the disconnected-peer filter below.
+	d.warmLazyPeers(question, result.records)
 	result.records = d.filterDisconnectedPeerAnswers(logger, question, result.records)
 	replyMessage.Authoritative = !result.hasExternalData
 	replyMessage.Answer = result.records
@@ -495,8 +550,8 @@ func (d *Resolver) filterDisconnectedPeerAnswers(logger *log.Entry, question dns
 	kept := make([]dns.RR, 0, len(records))
 	var dropped int
 	for _, rr := range records {
-		ip := extractRecordIP(rr)
-		if ip == "" {
+		ip, ok := extractRecordAddr(rr)
+		if !ok {
 			kept = append(kept, rr)
 			continue
 		}
@@ -518,22 +573,57 @@ func (d *Resolver) filterDisconnectedPeerAnswers(logger *log.Entry, question dns
 	return kept
 }
 
-// extractRecordIP returns the dotted-decimal / colon-hex IP carried by
-// an A or AAAA record, or "" for any other record type.
-func extractRecordIP(rr dns.RR) string {
+// warmLazyPeers triggers lazy-connection wake-up for the peers a resolved
+// answer points at and waits briefly for one to connect, so the caller's first
+// request doesn't race the connection establishment. Warm-up is scoped to
+// match-only (non-authoritative) zones — the synthesized private-service zones
+// and user-created zones whose records point at specific peers. The account's
+// peer zone is authoritative, so plain peer-name lookups never trigger warm-up;
+// otherwise resolving any peer's name would wake its idle connection, defeating
+// laziness mesh-wide. No-op when no activator is wired (lazy connections
+// disabled) or the answer carries no peer IPs.
+func (d *Resolver) warmLazyPeers(question dns.Question, records []dns.RR) {
+	if len(records) < 2 {
+		return
+	}
+	d.mu.RLock()
+	activator := d.peerActivator
+	var nonAuth, found bool
+	if activator != nil {
+		nonAuth, found = d.findZone(question.Name)
+	}
+	d.mu.RUnlock()
+	if activator == nil || !found || !nonAuth {
+		return
+	}
+
+	var addrs []netip.Addr
+	for _, rr := range records {
+		if addr, ok := extractRecordAddr(rr); ok {
+			addrs = append(addrs, addr)
+		}
+	}
+	if len(addrs) == 0 {
+		return
+	}
+
+	ctx, cancel := context.WithTimeout(d.ctx, d.warmupTimeout)
+	defer cancel()
+	activator.ActivatePeersByIP(ctx, addrs)
+}
+
+// extractRecordAddr returns the IP address carried by an A or AAAA record.
+// ok is false for any other record type or a record with no address.
+func extractRecordAddr(rr dns.RR) (netip.Addr, bool) {
 	switch r := rr.(type) {
 	case *dns.A:
-		if r.A == nil {
-			return ""
-		}
-		return r.A.String()
+		addr, ok := netip.AddrFromSlice(r.A)
+		return addr.Unmap(), ok
 	case *dns.AAAA:
-		if r.AAAA == nil {
-			return ""
-		}
-		return r.AAAA.String()
+		addr, ok := netip.AddrFromSlice(r.AAAA)
+		return addr.Unmap(), ok
 	}
-	return ""
+	return netip.Addr{}, false
 }
 
 // Update replaces all zones and their records

client/internal/dns/local/local_test.go

@@ -37,8 +37,8 @@ type mockPeerConnectivity struct {
 	byIP map[string]struct{ known, connected bool }
 }
 
-func (m mockPeerConnectivity) IsConnectedByIP(ip string) (known, connected bool) {
-	v, ok := m.byIP[ip]
+func (m mockPeerConnectivity) IsConnectedByIP(ip netip.Addr) (known, connected bool) {
+	v, ok := m.byIP[ip.String()]
 	if !ok {
 		return false, false
 	}

client/internal/dns/local/warmup_test.go

@@ -0,0 +1,204 @@
+package local
+
+import (
+	"context"
+	"net"
+	"net/netip"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/miekg/dns"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/dns/test"
+	nbdns "github.com/netbirdio/netbird/dns"
+)
+
+// recordingActivator records the addresses it was asked to warm and returns
+// immediately, so ServeDNS is not blocked by the test.
+type recordingActivator struct {
+	mu     sync.Mutex
+	called bool
+	addrs  []netip.Addr
+}
+
+func (r *recordingActivator) ActivatePeersByIP(_ context.Context, addrs []netip.Addr) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.called = true
+	r.addrs = append(r.addrs, addrs...)
+}
+
+func serveA(t *testing.T, resolver *Resolver, name string) *dns.Msg {
+	t.Helper()
+	var resp *dns.Msg
+	w := &test.MockResponseWriter{WriteMsgFunc: func(m *dns.Msg) error { resp = m; return nil }}
+	resolver.ServeDNS(w, new(dns.Msg).SetQuestion(name, dns.TypeA))
+	return resp
+}
+
+// serviceZone registers rec in a match-only (non-authoritative) zone, the shape
+// the synthesized private-service zones arrive in.
+func serviceZone(t *testing.T, resolver *Resolver, zone string, records ...nbdns.SimpleRecord) {
+	t.Helper()
+	resolver.Update([]nbdns.CustomZone{{
+		Domain:           zone,
+		Records:          records,
+		NonAuthoritative: true,
+	}})
+}
+
+func TestLocalResolver_WarmsLazyPeerOnResolve(t *testing.T) {
+	// Warm-up fires only for multi-record answers (the HA / round-robin shape of
+	// the synthesized private-service zones), so register two peer targets.
+	const name = "svc.proxy.netbird.cloud."
+	recs := []nbdns.SimpleRecord{
+		{Name: name, Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"},
+		{Name: name, Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.8"},
+	}
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud", recs...)
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	resp := serveA(t, resolver, name)
+	require.NotNil(t, resp, "resolver must answer")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A records")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.True(t, act.called, "activator must be invoked for a multi-record service-zone answer")
+	assert.Contains(t, act.addrs, netip.MustParseAddr("100.64.0.7"), "activator must receive the first peer IP")
+	assert.Contains(t, act.addrs, netip.MustParseAddr("100.64.0.8"), "activator must receive the second peer IP")
+}
+
+func TestLocalResolver_NoWarmupForSingleRecord(t *testing.T) {
+	// A single-record answer does not trigger warm-up; the resolver only warms
+	// multi-record answers.
+	rec := nbdns.SimpleRecord{Name: "svc.proxy.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"}
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud", rec)
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	resp := serveA(t, resolver, rec.Name)
+	require.NotNil(t, resp, "resolver must answer")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A record")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.False(t, act.called, "activator must not be invoked for a single-record answer")
+}
+
+func TestLocalResolver_NoActivatorNoWarmup(t *testing.T) {
+	// With no activator wired the resolver behaves exactly as before.
+	rec := nbdns.SimpleRecord{Name: "svc.proxy.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"}
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud", rec)
+
+	resp := serveA(t, resolver, rec.Name)
+	require.NotNil(t, resp, "resolver must still answer without an activator")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A record")
+}
+
+func TestLocalResolver_NoWarmupForMissingRecord(t *testing.T) {
+	// A query that resolves to nothing must not invoke the activator (no IPs).
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud",
+		nbdns.SimpleRecord{Name: "svc.proxy.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"})
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	serveA(t, resolver, "absent.proxy.netbird.cloud.")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.False(t, act.called, "activator must not be invoked when there is no answer")
+}
+
+func TestLocalResolver_NoWarmupInAuthoritativeZone(t *testing.T) {
+	// The account's peer zone is authoritative; resolving a peer's name there
+	// must not wake its lazy connection — warm-up is scoped to match-only
+	// (non-authoritative) zones such as the synthesized private-service zones.
+	// Use a multi-record answer so the authoritative-zone scoping is the only
+	// reason warm-up is skipped, not the single-record guard.
+	const name = "peer.netbird.cloud."
+	recs := []nbdns.SimpleRecord{
+		{Name: name, Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.9"},
+		{Name: name, Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.10"},
+	}
+	resolver := NewResolver()
+	resolver.Update([]nbdns.CustomZone{{
+		Domain:  "netbird.cloud",
+		Records: recs,
+	}})
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	resp := serveA(t, resolver, name)
+	require.NotNil(t, resp, "resolver must answer")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A records")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.False(t, act.called, "activator must not be invoked for authoritative-zone answers")
+}
+
+func TestLazyWarmupTimeoutFromEnv(t *testing.T) {
+	tests := []struct {
+		name   string
+		value  string
+		envSet bool
+		want   time.Duration
+	}{
+		{name: "unset uses default", want: defaultLazyWarmupTimeout},
+		{name: "valid overrides", value: "5s", envSet: true, want: 5 * time.Second},
+		{name: "invalid falls back", value: "not-a-duration", envSet: true, want: defaultLazyWarmupTimeout},
+		{name: "negative falls back", value: "-1s", envSet: true, want: defaultLazyWarmupTimeout},
+		{name: "zero falls back", value: "0s", envSet: true, want: defaultLazyWarmupTimeout},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if tt.envSet {
+				t.Setenv(envLazyWarmupTimeout, tt.value)
+			}
+			assert.Equal(t, tt.want, lazyWarmupTimeoutFromEnv())
+			assert.Equal(t, tt.want, NewResolver().warmupTimeout, "constructor must resolve the timeout once")
+		})
+	}
+}
+
+func TestExtractRecordAddr(t *testing.T) {
+	t.Run("A record yields unmapped v4", func(t *testing.T) {
+		// net.ParseIP returns the 16-byte v4-in-v6 form, the same shape
+		// miekg/dns stores after parsing an A record; the extracted address
+		// must compare equal to a plain v4 netip.Addr.
+		addr, ok := extractRecordAddr(&dns.A{A: net.ParseIP("100.64.0.7")})
+		require.True(t, ok)
+		assert.True(t, addr.Is4())
+		assert.Equal(t, netip.MustParseAddr("100.64.0.7"), addr)
+	})
+
+	t.Run("AAAA record yields v6", func(t *testing.T) {
+		addr, ok := extractRecordAddr(&dns.AAAA{AAAA: net.ParseIP("fd00::1")})
+		require.True(t, ok)
+		assert.Equal(t, netip.MustParseAddr("fd00::1"), addr)
+	})
+
+	t.Run("A record without address", func(t *testing.T) {
+		_, ok := extractRecordAddr(&dns.A{})
+		assert.False(t, ok)
+	})
+
+	t.Run("non-address record", func(t *testing.T) {
+		_, ok := extractRecordAddr(&dns.CNAME{Target: "target.netbird.cloud."})
+		assert.False(t, ok)
+	})
+}

client/internal/dns/mock_server.go

@@ -8,6 +8,7 @@ import (
 	"github.com/miekg/dns"
 
 	dnsconfig "github.com/netbirdio/netbird/client/internal/dns/config"
+	"github.com/netbirdio/netbird/client/internal/dns/local"
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/route"
 	"github.com/netbirdio/netbird/shared/management/domain"
@@ -92,6 +93,11 @@ func (m *MockServer) SetFirewall(Firewall) {
 	// Mock implementation - no-op
 }
 
+// SetPeerActivator mock implementation of SetPeerActivator from Server interface
+func (m *MockServer) SetPeerActivator(local.PeerActivator) {
+	// Mock implementation - no-op
+}
+
 // BeginBatch mock implementation of BeginBatch from Server interface
 func (m *MockServer) BeginBatch() {
 	// Mock implementation - no-op

client/internal/dns/server.go

@@ -82,6 +82,7 @@ type Server interface {
 	PopulateManagementDomain(mgmtURL *url.URL) error
 	SetRouteSources(selected, active func() route.HAMap)
 	SetFirewall(Firewall)
+	SetPeerActivator(local.PeerActivator)
 }
 
 type nsGroupsByDomain struct {
@@ -491,6 +492,13 @@ func (s *DefaultServer) SetFirewall(fw Firewall) {
 	}
 }
 
+// SetPeerActivator wires the DNS-time lazy-connection warm-up on the local
+// resolver. Injected after the connection manager exists (it does not at
+// DNS-server construction time). Pass nil to disable.
+func (s *DefaultServer) SetPeerActivator(a local.PeerActivator) {
+	s.localResolver.SetPeerActivator(a)
+}
+
 // Stop stops the server
 func (s *DefaultServer) Stop() {
 	s.ctxCancel()
@@ -1435,11 +1443,11 @@ type localPeerConnectivity struct {
 
 // IsConnectedByIP looks the IP up in the peerstore and surfaces both
 // the known and connected bits. Used by Resolver.filterDisconnectedPeerAnswers.
-func (l localPeerConnectivity) IsConnectedByIP(ip string) (known, connected bool) {
+func (l localPeerConnectivity) IsConnectedByIP(ip netip.Addr) (known, connected bool) {
 	if l.status == nil {
 		return false, false
 	}
-	state, ok := l.status.PeerStateByIP(ip)
+	state, ok := l.status.PeerStateByIP(ip.String())
 	if !ok {
 		return false, false
 	}

client/internal/dns/service_listener.go

@@ -292,18 +292,16 @@ func (s *serviceViaListener) generateFreePort() (uint16, error) {
 		return customPort, nil
 	}
 
-	udpAddr := net.UDPAddrFromAddrPort(netip.MustParseAddrPort("0.0.0.0:0"))
-	probeListener, err := net.ListenUDP("udp", udpAddr)
+	probeListener, err := net.ListenUDP("udp4", &net.UDPAddr{})
 	if err != nil {
 		log.Debugf("failed to bind random port for DNS: %s", err)
 		return 0, err
 	}
 
-	addrPort := netip.MustParseAddrPort(probeListener.LocalAddr().String()) // might panic if address is incorrect
-	err = probeListener.Close()
-	if err != nil {
+	port := uint16(probeListener.LocalAddr().(*net.UDPAddr).Port)
+	if err = probeListener.Close(); err != nil {
 		log.Debugf("failed to free up DNS port: %s", err)
 		return 0, err
 	}
-	return addrPort.Port(), nil
+	return port, nil
 }

client/internal/dns_peer_activator.go

@@ -0,0 +1,76 @@
+package internal
+
+import (
+	"context"
+	"net/netip"
+	"time"
+
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+)
+
+const dnsActivationPollInterval = 50 * time.Millisecond
+
+// dnsPeerActivator wakes lazy-connection peers from the DNS resolution path. It
+// implements dns/local.PeerActivator. DNS queries run on their own goroutines,
+// so it only touches state that is safe for concurrent use — ConnMgr.ActivatePeer,
+// peerstore.Store and peer.Status — and never takes the engine's syncMsgMux,
+// keeping DNS resolution from contending with network-map processing.
+type dnsPeerActivator struct {
+	connMgr   *ConnMgr
+	peerStore *peerstore.Store
+	status    *peer.Status
+	// ctx is the engine's long-lived context. The connection dial is tied to it
+	// (not the per-query DNS wait budget) so a handshake that outlasts the wait
+	// still completes in the background rather than being cancelled at the deadline.
+	ctx context.Context
+}
+
+// ActivatePeersByIP triggers wake-up for the peer(s) owning addrs and waits
+// until one is connected or ctx (the per-query DNS wait budget) expires.
+// Activation itself is tied to the engine's long-lived context so the dial
+// survives a wait that times out. Unknown or already-connected addresses are
+// skipped, so the steady-state (warm) path adds no latency.
+func (a *dnsPeerActivator) ActivatePeersByIP(ctx context.Context, addrs []netip.Addr) {
+	if a == nil || a.connMgr == nil {
+		return
+	}
+
+	var pending []string
+	for _, addr := range addrs {
+		ip := addr.String()
+		st, ok := a.status.PeerStateByIP(ip)
+		if !ok || st.ConnStatus == peer.StatusConnected {
+			continue
+		}
+		conn, ok := a.peerStore.PeerConn(st.PubKey)
+		if !ok {
+			continue
+		}
+		a.connMgr.ActivatePeer(a.ctx, conn)
+		pending = append(pending, ip)
+	}
+
+	if len(pending) == 0 {
+		return
+	}
+	a.waitConnected(ctx, pending)
+}
+
+// waitConnected blocks until any of ips reports a connected peer or ctx expires.
+func (a *dnsPeerActivator) waitConnected(ctx context.Context, ips []string) {
+	ticker := time.NewTicker(dnsActivationPollInterval)
+	defer ticker.Stop()
+	for {
+		for _, ip := range ips {
+			if st, ok := a.status.PeerStateByIP(ip); ok && st.ConnStatus == peer.StatusConnected {
+				return
+			}
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+	}
+}

client/internal/dns_peer_activator_test.go

@@ -0,0 +1,129 @@
+package internal
+
+import (
+	"context"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+)
+
+func newTestPeerConn(t *testing.T, key string) *peer.Conn {
+	t.Helper()
+	conn, err := peer.NewConn(peer.ConnConfig{
+		Key:      key,
+		LocalKey: "local",
+		WgConfig: peer.WgConfig{
+			AllowedIps: []netip.Prefix{netip.MustParsePrefix("100.64.0.1/32")},
+		},
+	}, peer.ServiceDependencies{})
+	require.NoError(t, err)
+	return conn
+}
+
+func newTestDNSPeerActivator(t *testing.T) (*dnsPeerActivator, *peer.Status, *peerstore.Store) {
+	t.Helper()
+	status := peer.NewRecorder("https://mgm")
+	store := peerstore.NewConnStore()
+	// ConnMgr without Start: the lazy manager is nil, so ActivatePeer is a
+	// no-op — these tests exercise the activator's skip/wait logic.
+	connMgr := NewConnMgr(&EngineConfig{}, status, store, nil)
+	return &dnsPeerActivator{
+		connMgr:   connMgr,
+		peerStore: store,
+		status:    status,
+		ctx:       context.Background(),
+	}, status, store
+}
+
+func TestDNSPeerActivator_NilSafe(t *testing.T) {
+	var a *dnsPeerActivator
+	a.ActivatePeersByIP(context.Background(), []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+}
+
+// TestDNSPeerActivator_SkipsUnknownAndConnectedPeers verifies the steady-state
+// (warm) path adds no latency: already-connected and unknown addresses never
+// enter the wait loop.
+func TestDNSPeerActivator_SkipsUnknownAndConnectedPeers(t *testing.T) {
+	a, status, store := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", "fd00::1"))
+	require.NoError(t, status.UpdatePeerState(peer.State{PubKey: "peerA", ConnStatus: peer.StatusConnected}))
+	store.AddPeerConn("peerA", newTestPeerConn(t, "peerA"))
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{
+		netip.MustParseAddr("100.64.0.1"),  // known, connected -> skipped
+		netip.MustParseAddr("fd00::1"),     // known via IPv6, connected -> skipped
+		netip.MustParseAddr("100.64.0.99"), // unknown -> skipped
+	})
+	require.Less(t, time.Since(start), time.Second, "no pending peer must mean no wait")
+}
+
+// TestDNSPeerActivator_WaitsForPendingPeerToConnect verifies the wait loop
+// returns as soon as a pending peer reports connected, well before the
+// per-query budget expires.
+func TestDNSPeerActivator_WaitsForPendingPeerToConnect(t *testing.T) {
+	a, status, store := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", ""))
+	store.AddPeerConn("peerA", newTestPeerConn(t, "peerA"))
+
+	go func() {
+		time.Sleep(150 * time.Millisecond)
+		_ = status.UpdatePeerState(peer.State{PubKey: "peerA", ConnStatus: peer.StatusConnected})
+	}()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+	elapsed := time.Since(start)
+
+	require.GreaterOrEqual(t, elapsed, 100*time.Millisecond, "must wait for the pending peer")
+	require.Less(t, elapsed, 5*time.Second, "must return on connect, not at the deadline")
+}
+
+// TestDNSPeerActivator_ReturnsAtBudgetWhenPeerStaysIdle verifies a peer that
+// never connects releases the DNS response at the per-query budget instead of
+// blocking it indefinitely.
+func TestDNSPeerActivator_ReturnsAtBudgetWhenPeerStaysIdle(t *testing.T) {
+	a, status, store := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", ""))
+	store.AddPeerConn("peerA", newTestPeerConn(t, "peerA"))
+
+	ctx, cancel := context.WithTimeout(context.Background(), 300*time.Millisecond)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+	elapsed := time.Since(start)
+
+	require.GreaterOrEqual(t, elapsed, 250*time.Millisecond, "must wait out the budget for a pending peer")
+	require.Less(t, elapsed, 5*time.Second, "must not block past the budget")
+}
+
+// TestDNSPeerActivator_NoWaitWithoutPeerConn verifies a known-but-idle peer
+// with no connection object in the store is not waited on: there is nothing to
+// activate, so waiting could only ever time out.
+func TestDNSPeerActivator_NoWaitWithoutPeerConn(t *testing.T) {
+	a, status, _ := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", ""))
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+	require.Less(t, time.Since(start), time.Second, "peer without a conn must not be waited on")
+}

client/internal/ebpf/ebpf/src/dns_fwd.c

@@ -52,11 +52,14 @@ int xdp_dns_fwd(struct iphdr  *ip, struct udphdr *udp) {
 
     if (udp->dest == GENERAL_DNS_PORT && ip->daddr == dns_ip) {
         udp->dest = dns_port;
+        // Clear the now-stale checksum; zero means "not computed" for IPv4.
+        udp->check = 0;
         return XDP_PASS;
     }
 
     if (udp->source == dns_port && ip->saddr == dns_ip) {
         udp->source = GENERAL_DNS_PORT;
+        udp->check = 0;
         return XDP_PASS;
     }
 

client/internal/ebpf/ebpf/src/wg_proxy.c

@@ -50,5 +50,11 @@ int xdp_wg_proxy(struct iphdr  *ip, struct udphdr *udp) {
     __be16 new_dst_port = htons(proxy_port);
     udp->dest = new_dst_port;
     udp->source = new_src_port;
+
+    // The ports are covered by the UDP checksum. This is an IPv4 loopback hop
+    // and the payload is already integrity-protected, so clear the checksum (a
+    // zero UDP checksum means "not computed" for IPv4) rather than leave a
+    // stale value the kernel would drop as UDP_CSUM.
+    udp->check = 0;
 	return XDP_PASS;
 }

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"
@@ -63,7 +64,10 @@ import (
 	"github.com/netbirdio/netbird/route"
 	mgm "github.com/netbirdio/netbird/shared/management/client"
 	"github.com/netbirdio/netbird/shared/management/domain"
+	sharedgrpc "github.com/netbirdio/netbird/shared/management/grpc"
+	nbnetworkmap "github.com/netbirdio/netbird/shared/management/networkmap"
 	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
+	types "github.com/netbirdio/netbird/shared/management/types"
 	"github.com/netbirdio/netbird/shared/netiputil"
 	auth "github.com/netbirdio/netbird/shared/relay/auth/hmac"
 	relayClient "github.com/netbirdio/netbird/shared/relay/client"
@@ -146,8 +150,11 @@ type EngineConfig struct {
 	BlockLANAccess      bool
 	BlockInbound        bool
 	DisableIPv6         bool
+	SyncMessageVersion  *int
 
-	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
 
@@ -217,6 +224,13 @@ type Engine struct {
 	// networkSerial is the latest CurrentSerial (state ID) of the network sent by the Management service
 	networkSerial uint64
 
+	// latestComponents is the most-recent NetworkMapComponents decoded from
+	// a NetworkMapEnvelope (capability=3 peers only). Held alongside the
+	// NetworkMap that Calculate() produced from it so future incremental
+	// updates have a base to apply changes against. nil for legacy-format
+	// peers. Guarded by syncMsgMux.
+	latestComponents *types.NetworkMapComponents
+
 	networkMonitor *networkmonitor.NetworkMonitor
 
 	sshServer sshServer
@@ -548,7 +562,7 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 		} else {
 			log.Infof("running rosenpass in strict mode")
 		}
-		e.rpManager, err = rosenpass.NewManager(e.config.PreSharedKey, e.config.WgIfaceName)
+		e.rpManager, err = rosenpass.NewManager(e.config.PreSharedKey, e.config.WgIfaceName, publicKey)
 		if err != nil {
 			return fmt.Errorf("create rosenpass manager: %w", err)
 		}
@@ -649,8 +663,24 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 	iceCfg := e.createICEConfig()
 
 	e.connMgr = NewConnMgr(e.config, e.statusRecorder, e.peerStore, wgIface)
+	e.connMgr.SetRoutedIPsReconciler(func(peerKey string) error {
+		if e.routeManager == nil {
+			return nil
+		}
+		return e.routeManager.ReconcilePeerAllowedIPs(peerKey)
+	})
 	e.connMgr.Start(e.ctx)
 
+	// Wire DNS-time lazy-connection warm-up now that the connection manager
+	// exists (it does not at DNS-server construction time). A DNS answer that
+	// points at an idle peer then wakes it before the client's first request.
+	e.dnsServer.SetPeerActivator(&dnsPeerActivator{
+		connMgr:   e.connMgr,
+		peerStore: e.peerStore,
+		status:    e.statusRecorder,
+		ctx:       e.ctx,
+	})
+
 	e.srWatcher = guard.NewSRWatcher(e.signal, e.relayManager, e.mobileDep.IFaceDiscover, iceCfg)
 	e.srWatcher.Start(peer.IsForceRelayed())
 
@@ -960,8 +990,12 @@ func (e *Engine) handleSync(update *mgmProto.SyncResponse) error {
 
 	e.ApplySessionDeadline(update.GetSessionExpiresAt())
 
-	if update.NetworkMap != nil && update.NetworkMap.PeerConfig != nil {
-		e.handleAutoUpdateVersion(update.NetworkMap.PeerConfig.AutoUpdate)
+	// Envelope sync responses carry PeerConfig at the top level; legacy
+	// NetworkMap syncs carry it under NetworkMap.PeerConfig.
+	if pc := update.GetPeerConfig(); pc != nil {
+		e.handleAutoUpdateVersion(pc.GetAutoUpdate())
+	} else if nm := update.GetNetworkMap(); nm != nil && nm.GetPeerConfig() != nil {
+		e.handleAutoUpdateVersion(nm.GetPeerConfig().GetAutoUpdate())
 	}
 
 	done := e.phase("netbird_config")
@@ -971,12 +1005,47 @@ func (e *Engine) handleSync(update *mgmProto.SyncResponse) error {
 		return err
 	}
 
+	// Decode the network map from either the components envelope or the
+	// legacy proto.NetworkMap before the posture-check gating below, so the
+	// "is there a network map" decision covers both wire shapes.
+	var (
+		nm         *mgmProto.NetworkMap
+		components *types.NetworkMapComponents
+	)
+	if version := update.GetVersion(); version == int32(sharedgrpc.ComponentNetworkMap) {
+		// Components-format peer: decode the envelope back to typed
+		// components, run Calculate() locally, and convert to the wire
+		// NetworkMap shape the rest of the engine consumes. Components are
+		// retained so future incremental updates can apply deltas instead
+		// of doing a full reconstruction.
+		envelope := update.GetNetworkMapEnvelope()
+		if envelope == nil {
+			return fmt.Errorf("received a SyncReponse indicating use of components network map, but components are missing")
+		}
+
+		localKey := e.config.WgPrivateKey.PublicKey().String()
+		dnsName := ""
+		if pc := update.GetPeerConfig(); pc != nil {
+			// PeerConfig.Fqdn = "<dns_label>.<dns_domain>" — extract the
+			// shared domain by stripping the peer's own label prefix. Falls
+			// back to empty if the FQDN doesn't have the expected shape.
+			dnsName = extractDNSDomainFromFQDN(pc.GetFqdn())
+		}
+		result, err := nbnetworkmap.EnvelopeToNetworkMap(e.ctx, envelope, localKey, dnsName)
+		if err != nil {
+			return fmt.Errorf("decode network map envelope: %w", err)
+		}
+		nm = result.NetworkMap
+		components = result.Components
+	} else {
+		nm = update.GetNetworkMap()
+	}
+
 	// Posture checks are bound to the network map presence:
 	//   NetworkMap != nil, checks present -> apply the received checks
 	//   NetworkMap != nil, checks nil     -> posture checks were removed, clear them
 	//   NetworkMap == nil                 -> config-only update (e.g. relay token rotation),
 	//                                        leave the previously applied checks untouched
-	nm := update.GetNetworkMap()
 	if nm == nil {
 		return nil
 	}
@@ -989,6 +1058,14 @@ func (e *Engine) handleSync(update *mgmProto.SyncResponse) error {
 	}
 
 	done = e.phase("persist")
+	// Only retain the components view when the server sent the envelope
+	// path. A legacy proto.NetworkMap means components == nil; writing it
+	// here would clobber a previously-cached snapshot, breaking the
+	// incremental-delta base on a future envelope sync.
+	if components != nil {
+		e.latestComponents = components
+	}
+
 	e.persistSyncResponse(update)
 	done()
 
@@ -1002,6 +1079,19 @@ func (e *Engine) handleSync(update *mgmProto.SyncResponse) error {
 	return nil
 }
 
+// extractDNSDomainFromFQDN returns the trailing dotted domain part of the
+// receiving peer's FQDN — the same value the management server fills as
+// dnsName when it builds the legacy NetworkMap. "peer42.netbird.cloud" →
+// "netbird.cloud". An empty string is returned for unrecognized formats.
+func extractDNSDomainFromFQDN(fqdn string) string {
+	for i := 0; i < len(fqdn); i++ {
+		if fqdn[i] == '.' && i+1 < len(fqdn) {
+			return fqdn[i+1:]
+		}
+	}
+	return ""
+}
+
 // updateNetbirdConfig applies the management-provided NetBird configuration:
 // STUN/TURN and relay servers, flow logging and DNS settings. A nil config is a no-op,
 // which is the case for sync updates carrying only a network map.
@@ -1161,7 +1251,7 @@ func (e *Engine) applyInfoFlags(info *system.Info) {
 		e.config.BlockLANAccess,
 		e.config.BlockInbound,
 		e.config.DisableIPv6,
-		e.config.LazyConnectionEnabled,
+		e.config.SyncMessageVersion,
 		e.config.EnableSSHRoot,
 		e.config.EnableSSHSFTP,
 		e.config.EnableSSHLocalPortForwarding,
@@ -2030,7 +2120,7 @@ func (e *Engine) readInitialSettings() ([]*route.Route, *nbdns.Config, bool, err
 		e.config.BlockLANAccess,
 		e.config.BlockInbound,
 		e.config.DisableIPv6,
-		e.config.LazyConnectionEnabled,
+		e.config.SyncMessageVersion,
 		e.config.EnableSSHRoot,
 		e.config.EnableSSHSFTP,
 		e.config.EnableSSHLocalPortForwarding,
@@ -2604,13 +2694,14 @@ func (e *Engine) updateForwardRules(rules []*mgmProto.ForwardingRule) ([]firewal
 
 func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, peers []*mgmProto.RemotePeerConfig) map[string]bool {
 	excludedPeers := make(map[string]bool)
+
+	// Ingress forward targets: inbound forwarded traffic is initiated remotely and
+	// cannot wake a lazy connection, so the peer routing the target must stay
+	// permanently connected. AllowedIPs are already parsed on the peer conn, so
+	// reuse those typed prefixes instead of re-parsing the network map strings.
 	for _, r := range rules {
-		ip := r.TranslatedAddress
 		for _, p := range peers {
-			for _, allowedIP := range p.GetAllowedIps() {
-				if allowedIP != ip.String() {
-					continue
-				}
+			if e.peerRoutesAddr(p, r.TranslatedAddress) {
 				log.Infof("exclude forwarder peer from lazy connection: %s", p.GetWgPubKey())
 				excludedPeers[p.GetWgPubKey()] = true
 			}
@@ -2620,6 +2711,27 @@ func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, peers
 	return excludedPeers
 }
 
+// peerRoutesAddr reports whether the peer is a router for addr, matched against
+// the peer's already-parsed AllowedIPs from the store (the same typed value the
+// lazy manager consumes) rather than re-parsing the network map strings.
+func (e *Engine) peerRoutesAddr(p *mgmProto.RemotePeerConfig, addr netip.Addr) bool {
+	prefixes, ok := e.peerStore.AllowedIPs(p.GetWgPubKey())
+	if !ok {
+		return false
+	}
+	return prefixesContain(prefixes, addr)
+}
+
+// prefixesContain reports whether addr falls within any of the prefixes.
+func prefixesContain(prefixes []netip.Prefix, addr netip.Addr) bool {
+	for _, prefix := range prefixes {
+		if prefix.Contains(addr) {
+			return true
+		}
+	}
+	return false
+}
+
 // isChecksEqual checks if two slices of checks are equal.
 func isChecksEqual(checks1, checks2 []*mgmProto.Checks) bool {
 	normalize := func(checks []*mgmProto.Checks) []string {

client/internal/engine_lazy_exclude_test.go

@@ -0,0 +1,87 @@
+package internal
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+
+	firewallManager "github.com/netbirdio/netbird/client/firewall/manager"
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
+)
+
+func TestPrefixesContain(t *testing.T) {
+	tests := []struct {
+		name     string
+		prefixes []string
+		addr     string
+		want     bool
+	}{
+		{name: "own overlay /32 matches", prefixes: []string{"100.110.8.145/32"}, addr: "100.110.8.145", want: true},
+		{name: "addr inside routed subnet", prefixes: []string{"10.121.0.0/16"}, addr: "10.121.208.4", want: true},
+		{name: "addr outside subnet", prefixes: []string{"10.121.0.0/16"}, addr: "10.122.0.1", want: false},
+		{name: "different /32", prefixes: []string{"100.110.8.145/32"}, addr: "100.110.8.146", want: false},
+		{name: "ipv6 /128 matches", prefixes: []string{"fd00::1/128"}, addr: "fd00::1", want: true},
+		{name: "no prefixes", prefixes: nil, addr: "10.121.208.4", want: false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			prefixes := make([]netip.Prefix, 0, len(tt.prefixes))
+			for _, p := range tt.prefixes {
+				prefixes = append(prefixes, netip.MustParsePrefix(p))
+			}
+			require.Equal(t, tt.want, prefixesContain(prefixes, netip.MustParseAddr(tt.addr)))
+		})
+	}
+}
+
+// TestToExcludedLazyPeers_ForwardTarget guards a regression: the forward-target
+// peer (the peer routing a ForwardRule.TranslatedAddress) must be excluded from
+// lazy connections, matched via the peer's already-parsed AllowedIPs.
+func TestToExcludedLazyPeers_ForwardTarget(t *testing.T) {
+	const targetPeerKey = "cccccccccccccccccccccccccccccccccccccccccc0="
+	const otherPeerKey = "dddddddddddddddddddddddddddddddddddddddddd0="
+
+	store := peerstore.NewConnStore()
+	store.AddPeerConn(targetPeerKey, newTestConn(t, targetPeerKey, "100.110.8.145/32"))
+	store.AddPeerConn(otherPeerKey, newTestConn(t, otherPeerKey, "100.110.9.10/32"))
+
+	e := &Engine{peerStore: store}
+
+	peers := []*mgmProto.RemotePeerConfig{
+		{WgPubKey: targetPeerKey, AllowedIps: []string{"100.110.8.145/32"}},
+		{WgPubKey: otherPeerKey, AllowedIps: []string{"100.110.9.10/32"}},
+	}
+	rules := []firewallManager.ForwardRule{
+		{TranslatedAddress: netip.MustParseAddr("100.110.8.145")},
+	}
+
+	excluded := e.toExcludedLazyPeers(rules, peers)
+
+	require.True(t, excluded[targetPeerKey], "forward-target peer must be excluded from lazy connections")
+	require.False(t, excluded[otherPeerKey], "non-target peer must not be excluded")
+	require.Len(t, excluded, 1)
+}
+
+func TestToExcludedLazyPeers_NoRules(t *testing.T) {
+	e := &Engine{peerStore: peerstore.NewConnStore()}
+
+	peers := []*mgmProto.RemotePeerConfig{
+		{WgPubKey: "peer-a", AllowedIps: []string{"100.110.8.145/32"}},
+	}
+
+	require.Empty(t, e.toExcludedLazyPeers(nil, peers))
+}
+
+func newTestConn(t *testing.T, key, allowedIP string) *peer.Conn {
+	t.Helper()
+	conn, err := peer.NewConn(peer.ConnConfig{
+		Key:      key,
+		WgConfig: peer.WgConfig{AllowedIps: []netip.Prefix{netip.MustParsePrefix(allowedIP)}},
+	}, peer.ServiceDependencies{})
+	require.NoError(t, err)
+	return conn
+}

client/internal/engine_lazyconn.go

@@ -1,19 +0,0 @@
-package internal
-
-import (
-	"errors"
-)
-
-// SetLazyConnEnabled applies a local lazy connection override to the running
-// engine. It pins the setting like an env/CLI flag, so a later management sync
-// cannot override it. syncMsgMux guards ConnMgr, which is not thread-safe.
-func (e *Engine) SetLazyConnEnabled(enabled bool) error {
-	e.syncMsgMux.Lock()
-	defer e.syncMsgMux.Unlock()
-
-	if e.connMgr == nil {
-		return errors.New("connection manager is not initialised")
-	}
-
-	return e.connMgr.SetLocalLazyConn(e.ctx, enabled)
-}

client/internal/engine_session_deadline_test.go

@@ -75,4 +75,14 @@ func TestApplySessionDeadline_ThreeState(t *testing.T) {
 		require.True(t, e.statusRecorder.GetSessionExpiresAt().IsZero(),
 			"invalid timestamp must clear the deadline")
 	})
+
+	t.Run("recently expired timestamp stays visible as expired", func(t *testing.T) {
+		e := newEngine()
+		expired := time.Now().Add(-5 * time.Minute).UTC().Truncate(time.Second)
+
+		e.ApplySessionDeadline(timestamppb.New(expired))
+
+		require.True(t, e.statusRecorder.GetSessionExpiresAt().Equal(expired),
+			"recently-expired deadline must stay on the recorder so consumers render it as expired")
+	})
 }

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/lazyconn/manager/manager.go

@@ -29,6 +29,11 @@ type managedPeer struct {
 
 type Config struct {
 	InactivityThreshold *time.Duration
+	// ReconcileAllowedIPs re-applies a peer's routed allowed IPs after its wake endpoint is
+	// armed. The activity listener creates the wake peer with the overlay /32 only; without the
+	// routed prefixes WireGuard would not steer subnet-bound traffic to the wake endpoint, so an
+	// idle routing peer could never be woken by that traffic. Optional; nil disables the reconcile.
+	ReconcileAllowedIPs func(peerKey string) error
 }
 
 // Manager manages lazy connections
@@ -56,6 +61,9 @@ type Manager struct {
 	peerToHAGroups map[string][]route.HAUniqueID // peer ID -> HA groups they belong to
 	haGroupToPeers map[route.HAUniqueID][]string // HA group -> peer IDs in the group
 	routesMu       sync.RWMutex
+
+	// reconcileAllowedIPs re-applies a peer's routed allowed IPs after its wake endpoint is armed.
+	reconcileAllowedIPs func(peerKey string) error
 }
 
 // NewManager creates a new lazy connection manager
@@ -73,6 +81,7 @@ func NewManager(config Config, engineCtx context.Context, peerStore *peerstore.S
 		activityManager:      activity.NewManager(wgIface),
 		peerToHAGroups:       make(map[string][]route.HAUniqueID),
 		haGroupToPeers:       make(map[route.HAUniqueID][]string),
+		reconcileAllowedIPs:  config.ReconcileAllowedIPs,
 	}
 
 	if wgIface.IsUserspaceBind() {
@@ -201,7 +210,7 @@ func (m *Manager) AddPeer(peerCfg lazyconn.PeerConfig) (bool, error) {
 		return false, nil
 	}
 
-	if err := m.activityManager.MonitorPeerActivity(peerCfg); err != nil {
+	if err := m.armActivityListener(peerCfg); err != nil {
 		return false, err
 	}
 
@@ -288,7 +297,7 @@ func (m *Manager) DeactivatePeer(peerID peerid.ConnID) {
 
 	m.inactivityManager.RemovePeer(mp.peerCfg.PublicKey)
 
-	if err := m.activityManager.MonitorPeerActivity(*mp.peerCfg); err != nil {
+	if err := m.armActivityListener(*mp.peerCfg); err != nil {
 		mp.peerCfg.Log.Errorf("failed to create activity monitor: %v", err)
 		return
 	}
@@ -465,6 +474,31 @@ func (m *Manager) close() {
 }
 
 // shouldDeferIdleForHA checks if peer should stay connected due to HA group requirements
+// armRoutedAllowedIPs re-applies the peer's routed allowed IPs onto its freshly armed wake
+// endpoint. The activity listener creates the wake peer with the overlay /32 only, so without
+// this the routed prefixes would be missing and traffic to a routed subnet could not wake the
+// idle routing peer. It is a no-op when no reconciler is configured.
+// armActivityListener (re)arms the peer's wake endpoint via the activity manager and then
+// re-applies its routed allowed IPs, so traffic to a routed subnet can wake an idle routing
+// peer. The routed prefixes must be re-applied after the wake endpoint exists because the
+// listener creates it with the overlay /32 only.
+func (m *Manager) armActivityListener(peerCfg lazyconn.PeerConfig) error {
+	if err := m.activityManager.MonitorPeerActivity(peerCfg); err != nil {
+		return err
+	}
+	m.armRoutedAllowedIPs(&peerCfg)
+	return nil
+}
+
+func (m *Manager) armRoutedAllowedIPs(peerCfg *lazyconn.PeerConfig) {
+	if m.reconcileAllowedIPs == nil {
+		return
+	}
+	if err := m.reconcileAllowedIPs(peerCfg.PublicKey); err != nil {
+		peerCfg.Log.Errorf("failed to reconcile routed allowed IPs on wake endpoint: %v", err)
+	}
+}
+
 func (m *Manager) shouldDeferIdleForHA(inactivePeers map[string]struct{}, peerID string) bool {
 	m.routesMu.RLock()
 	defer m.routesMu.RUnlock()
@@ -577,7 +611,7 @@ func (m *Manager) onPeerInactivityTimedOut(peerIDs map[string]struct{}) {
 
 		mp.peerCfg.Log.Infof("start activity monitor")
 
-		if err := m.activityManager.MonitorPeerActivity(*mp.peerCfg); err != nil {
+		if err := m.armActivityListener(*mp.peerCfg); err != nil {
 			mp.peerCfg.Log.Errorf("failed to create activity monitor: %v", err)
 			continue
 		}

client/internal/netflow/logger/logger.go

@@ -27,7 +27,7 @@ type Logger struct {
 	wgIfaceNetV6       netip.Prefix
 	dnsCollection      atomic.Bool
 	exitNodeCollection atomic.Bool
-	Store              types.Store
+	Store              types.AggregatingStore
 }
 
 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.NewMemoryStore(),
+		Store:          store.NewAggregatingMemoryStore(),
 	}
 }
 
@@ -125,6 +125,10 @@ 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

@@ -9,12 +9,14 @@ 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"
@@ -23,14 +25,16 @@ 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
-	publicKey      []byte
-	cancel         context.CancelFunc
+	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
 }
 
 // NewManager creates a new netflow manager
@@ -48,9 +52,11 @@ func NewManager(iface nftypes.IFaceMapper, publicKey []byte, statusRecorder *pee
 	}
 
 	return &Manager{
-		logger:    flowLogger,
-		conntrack: ct,
-		publicKey: publicKey,
+		logger:            flowLogger,
+		conntrack:         ct,
+		publicKey:         publicKey,
+		retryInterval:     time.Second,
+		eventsWithoutAcks: store.NewMemoryStore(),
 	}
 }
 
@@ -66,6 +72,7 @@ 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) {
@@ -85,6 +92,7 @@ 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 {
@@ -107,14 +115,19 @@ func (m *Manager) resetClient() error {
 	ctx, cancel := context.WithCancel(context.Background())
 	m.cancel = cancel
 
-	m.shutdownWg.Add(2)
+	m.shutdownWg.Add(3)
+	flowConfigInterval := m.flowConfig.Interval
 	go func() {
 		defer m.shutdownWg.Done()
-		m.receiveACKs(ctx, flowClient)
+		m.receiveACKs(ctx, flowClient, flowConfigInterval)
 	}()
 	go func() {
 		defer m.shutdownWg.Done()
-		m.startSender(ctx)
+		m.startSender(ctx, flowConfigInterval)
+	}()
+	go func() {
+		defer m.shutdownWg.Done()
+		m.startRetries(ctx, flowConfigInterval)
 	}()
 
 	return nil
@@ -198,8 +211,8 @@ func (m *Manager) GetLogger() nftypes.FlowLogger {
 	return m.logger
 }
 
-func (m *Manager) startSender(ctx context.Context) {
-	ticker := time.NewTicker(m.flowConfig.Interval)
+func (m *Manager) startSender(ctx context.Context, flowConfigInterval time.Duration) {
+	ticker := time.NewTicker(flowConfigInterval)
 	defer ticker.Stop()
 
 	for {
@@ -207,27 +220,29 @@ func (m *Manager) startSender(ctx context.Context) {
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			events := m.logger.GetEvents()
+			collectedEvents := m.logger.ResetAggregationWindow()
+			events := collectedEvents.GetAggregatedEvents()
 			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)
-					continue
+				} else {
+					log.Tracef("sent flow event: %s", event.ID)
 				}
-				log.Tracef("sent flow event: %s", event.ID)
 			}
 		}
 	}
 }
 
-func (m *Manager) receiveACKs(ctx context.Context, client *client.GRPCClient) {
-	err := client.Receive(ctx, m.flowConfig.Interval, func(ack *proto.FlowEventAck) error {
+func (m *Manager) receiveACKs(ctx context.Context, client *client.GRPCClient, flowConfigInterval time.Duration) {
+	err := client.Receive(ctx, flowConfigInterval, 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.logger.DeleteEvents([]uuid.UUID{id})
+		m.eventsWithoutAcks.DeleteEvents([]uuid.UUID{id})
 		return nil
 	})
 
@@ -236,6 +251,51 @@ func (m *Manager) receiveACKs(ctx context.Context, client *client.GRPCClient) {
 	}
 }
 
+// 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
@@ -250,9 +310,11 @@ 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,
+		EventId:     event.ID[:],
+		Timestamp:   timestamppb.New(event.Timestamp),
+		PublicKey:   publicKey,
+		WindowStart: timestamppb.New(event.WindowStart),
+		WindowEnd:   timestamppb.New(event.WindowEnd),
 		FlowFields: &proto.FlowFields{
 			FlowId:           event.FlowID[:],
 			RuleId:           event.RuleID,
@@ -267,6 +329,9 @@ 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

@@ -0,0 +1,291 @@
+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

@@ -0,0 +1,365 @@
+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) {
+	now := time.Now()
+	nowFunc := func() time.Time { return now }
+	store := NewAggregatingMemoryStoreWithTimeFunc(nowFunc)
+	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(),
+		},
+	})
+
+	now = now.Add(1 * time.Second)
+	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

@@ -1,10 +1,15 @@
 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"
 )
 
@@ -19,6 +24,14 @@ type Memory struct {
 	events map[uuid.UUID]*types.Event
 }
 
+type AggregatingMemory struct {
+	Memory
+	WindowStart time.Time
+	WindowEnd   time.Time
+	rnd         *v2.PCG
+	nowFunc     func() time.Time
+}
+
 func (m *Memory) StoreEvent(event *types.Event) {
 	m.mux.Lock()
 	defer m.mux.Unlock()
@@ -48,3 +61,104 @@ func (m *Memory) DeleteEvents(ids []uuid.UUID) {
 		delete(m.events, id)
 	}
 }
+
+func NewAggregatingMemoryStore() *AggregatingMemory {
+	return NewAggregatingMemoryStoreWithTimeFunc(defaultNowFunc)
+}
+
+// used in tests when deterministic (less random) time intervals are required
+func NewAggregatingMemoryStoreWithTimeFunc(nowFunc func() time.Time) *AggregatingMemory {
+	return &AggregatingMemory{WindowStart: nowFunc(), Memory: Memory{events: make(map[uuid.UUID]*types.Event)}, nowFunc: nowFunc, rnd: v2.NewPCG(rand.Uint64(), rand.Uint64())}
+}
+
+func (am *AggregatingMemory) ResetAggregationWindow() types.FlowEventAggregator {
+	am.mux.Lock()
+	defer am.mux.Unlock()
+
+	now := am.nowFunc()
+	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
+}
+
+func defaultNowFunc() time.Time {
+	return time.Now()
+}

client/internal/netflow/types/types.go

@@ -2,6 +2,7 @@ package types
 
 import (
 	"net/netip"
+	"slices"
 	"strconv"
 	"time"
 
@@ -69,8 +70,10 @@ const (
 )
 
 type Event struct {
-	ID        uuid.UUID
-	Timestamp time.Time
+	ID          uuid.UUID
+	Timestamp   time.Time
+	WindowStart time.Time
+	WindowEnd   time.Time
 	EventFields
 }
 
@@ -92,6 +95,17 @@ 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 {
@@ -114,13 +128,15 @@ 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
@@ -140,6 +156,11 @@ 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

@@ -30,6 +30,11 @@ import (
 	relayClient "github.com/netbirdio/netbird/shared/relay/client"
 )
 
+// wgTimeoutEscalationThreshold is the number of consecutive WireGuard
+// handshake timeouts after which the rosenpass state for the peer is
+// considered desynced and gets reset.
+const wgTimeoutEscalationThreshold = 3
+
 // MetricsRecorder is an interface for recording peer connection metrics
 type MetricsRecorder interface {
 	RecordConnectionStages(
@@ -118,6 +123,9 @@ type Conn struct {
 	wgWatcher       *WGWatcher
 	wgWatcherWg     sync.WaitGroup
 	wgWatcherCancel context.CancelFunc
+	// wgTimeouts counts consecutive WireGuard handshake timeouts without a
+	// successful handshake in between. Guarded by mu.
+	wgTimeouts int
 
 	// used to store the remote Rosenpass key for Relayed connection in case of connection update from ice
 	rosenpassRemoteKey []byte
@@ -195,7 +203,6 @@ 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 +670,12 @@ func (conn *Conn) onGuardEvent() {
 	}
 }
 
-func (conn *Conn) onWGDisconnected() {
+func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
-	if conn.ctx.Err() != nil {
+	// watcherCtx guards against a stale watcher tearing down a connection that already superseded it.
+	if conn.ctx.Err() != nil || watcherCtx.Err() != nil {
 		return
 	}
 
@@ -683,6 +691,29 @@ func (conn *Conn) onWGDisconnected() {
 	default:
 		conn.Log.Debugf("No active connection to close on WG timeout")
 	}
+
+	conn.escalateWGTimeoutLocked()
+}
+
+// escalateWGTimeoutLocked resets the peer's rosenpass state after repeated
+// handshake timeouts. With rosenpass enabled, persistent timeouts mean the
+// preshared keys have desynced; the renewal exchange runs over the dead
+// tunnel and cannot resync them. Reporting the peer disconnected drops its
+// rosenpass state, so the next connection configuration programs the
+// rendezvous key and the tunnel can bootstrap again. Callers must hold mu.
+func (conn *Conn) escalateWGTimeoutLocked() {
+	if conn.config.RosenpassConfig.PubKey == nil {
+		return
+	}
+
+	conn.wgTimeouts++
+	if conn.wgTimeouts < wgTimeoutEscalationThreshold || conn.onDisconnected == nil {
+		return
+	}
+	conn.wgTimeouts = 0
+
+	conn.Log.Warnf("%d consecutive WireGuard handshake timeouts, resetting rosenpass state for peer", wgTimeoutEscalationThreshold)
+	conn.onDisconnected(conn.config.WgConfig.RemoteKey)
 }
 
 func (conn *Conn) updateRelayStatus(relayServerAddr string, rosenpassPubKey []byte, updateTime time.Time) {
@@ -802,25 +833,39 @@ func (conn *Conn) isConnectedOnAllWay() (status guard.ConnStatus) {
 	})
 }
 
+// enableWgWatcherIfNeeded starts a fresh watcher instance per connection attempt, so its
+// lifecycle stays bound to conn.mu and enable/disable can't race an old goroutine's shutdown.
+// Caller must hold conn.mu.
 func (conn *Conn) enableWgWatcherIfNeeded(enabledTime time.Time) {
-	if !conn.wgWatcher.PrepareInitialHandshake() {
+	if conn.wgWatcher != nil {
 		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()
-		conn.wgWatcher.EnableWgWatcher(wgWatcherCtx, enabledTime, conn.onWGDisconnected, conn.onWGHandshakeSuccess)
+		onDisconnected := func() { conn.onWGDisconnected(wgWatcherCtx) }
+		watcher.EnableWgWatcher(wgWatcherCtx, enabledTime, onDisconnected, conn.onWGHandshakeSuccess, conn.onWGCheckSuccess)
 	}()
 }
 
+// disableWgWatcherIfNeeded cancels and drops the watcher once no transport is active. It never
+// waits for the goroutine: the timeout path reentrantly calls back here under conn.mu, so
+// blocking would deadlock. 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) {
@@ -843,7 +888,9 @@ 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)
 	}
@@ -892,6 +939,15 @@ func (conn *Conn) onWGHandshakeSuccess(when time.Time) {
 	conn.recordConnectionMetrics()
 }
 
+// onWGCheckSuccess is called for every watcher check that observed a fresh
+// handshake, including handshakes of connections that were already up when
+// the watcher started.
+func (conn *Conn) onWGCheckSuccess() {
+	conn.mu.Lock()
+	conn.wgTimeouts = 0
+	conn.mu.Unlock()
+}
+
 // recordConnectionMetrics records connection stage timestamps as metrics
 func (conn *Conn) recordConnectionMetrics() {
 	if conn.metricsRecorder == nil {

client/internal/peer/conn_test.go

@@ -7,6 +7,7 @@ import (
 	"testing"
 	"time"
 
+	log "github.com/sirupsen/logrus"
 	"github.com/stretchr/testify/assert"
 
 	"github.com/netbirdio/netbird/client/iface"
@@ -304,3 +305,84 @@ func TestConn_presharedKey_RosenpassManaged(t *testing.T) {
 		t.Fatalf("expected non-nil presharedKey before Rosenpass manages PSK")
 	}
 }
+
+func newWGTimeoutTestConn(rosenpassEnabled bool, disconnected *[]string) *Conn {
+	cfg := ConnConfig{
+		Key:      "LLHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU=",
+		LocalKey: "RRHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU=",
+		WgConfig: WgConfig{RemoteKey: "LLHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU="},
+	}
+	if rosenpassEnabled {
+		cfg.RosenpassConfig = RosenpassConfig{PubKey: []byte("dummykey")}
+	}
+
+	conn := &Conn{
+		ctx:           context.Background(),
+		config:        cfg,
+		Log:           log.WithField("peer", cfg.Key),
+		metricsStages: &MetricsStages{},
+	}
+	conn.SetOnDisconnected(func(remotePeer string) {
+		*disconnected = append(*disconnected, remotePeer)
+	})
+	return conn
+}
+
+// TestConn_onWGDisconnected_EscalatesToRosenpassReset: repeated handshake
+// timeouts with rosenpass enabled mean the preshared keys have desynced. The
+// renewal exchange runs over the dead tunnel and cannot resync them, so after
+// wgTimeoutEscalationThreshold consecutive timeouts the conn must report the
+// peer disconnected, dropping its rosenpass state so the next configuration
+// programs the rendezvous key.
+func TestConn_onWGDisconnected_EscalatesToRosenpassReset(t *testing.T) {
+	var disconnected []string
+	conn := newWGTimeoutTestConn(true, &disconnected)
+
+	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
+		conn.onWGDisconnected(conn.ctx)
+	}
+	assert.Empty(t, disconnected, "escalation must not fire below the threshold")
+
+	conn.onWGDisconnected(conn.ctx)
+	assert.Equal(t, []string{conn.config.WgConfig.RemoteKey}, disconnected,
+		"reaching the threshold must report the peer disconnected once")
+
+	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
+		conn.onWGDisconnected(conn.ctx)
+	}
+	assert.Len(t, disconnected, 1, "escalation must restart counting after firing")
+
+	conn.onWGDisconnected(conn.ctx)
+	assert.Len(t, disconnected, 2, "continued timeouts must escalate again")
+}
+
+// TestConn_onWGDisconnected_CheckSuccessResetsEscalation: a successful
+// handshake between timeouts means the tunnel recovered; the counter must
+// start over.
+func TestConn_onWGDisconnected_CheckSuccessResetsEscalation(t *testing.T) {
+	var disconnected []string
+	conn := newWGTimeoutTestConn(true, &disconnected)
+
+	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
+		conn.onWGDisconnected(conn.ctx)
+	}
+	conn.onWGCheckSuccess()
+
+	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
+		conn.onWGDisconnected(conn.ctx)
+	}
+	assert.Empty(t, disconnected, "handshake success must reset the timeout count")
+}
+
+// TestConn_onWGDisconnected_NoEscalationWithoutRosenpass: without rosenpass
+// there is no per-peer key state to reset; repeated timeouts must not report
+// disconnects.
+func TestConn_onWGDisconnected_NoEscalationWithoutRosenpass(t *testing.T) {
+	var disconnected []string
+	conn := newWGTimeoutTestConn(false, &disconnected)
+
+	for i := 0; i < wgTimeoutEscalationThreshold*3; i++ {
+		conn.onWGDisconnected(conn.ctx)
+	}
+	assert.Empty(t, disconnected, "escalation must be limited to rosenpass connections")
+}

client/internal/peer/guard/guard.go

@@ -85,7 +85,11 @@ func (g *Guard) reconnectLoopWithRetry(ctx context.Context, callback func()) {
 	defer g.srWatcher.RemoveListener(srReconnectedChan)
 
 	ticker := g.initialTicker(ctx)
-	defer ticker.Stop()
+	defer func() {
+		// If backoff.Ticker.send is blocked, context.Done will not close the Ticker goroutine.
+		// We have to explicitly call Stop, even if we use backoff.WithContext.
+		ticker.Stop()
+	}()
 
 	tickerChannel := ticker.C
 

client/internal/peer/guard/guard_leak_test.go

@@ -0,0 +1,92 @@
+package guard
+
+import (
+	"context"
+	"runtime"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/peer/ice"
+)
+
+func newTestGuard(status connStatusFunc) *Guard {
+	srw := NewSRWatcher(nil, nil, nil, ice.Config{})
+	return NewGuard(log.WithField("test", "guard"), status, 50*time.Millisecond, srw)
+}
+
+// countBackoffTickerGoroutines returns how many goroutines are currently sitting
+// in backoff/v4.(*Ticker).run (a ticker goroutine that has not exited).
+func countBackoffTickerGoroutines() int {
+	buf := make([]byte, 1<<25) // 32MB
+	n := runtime.Stack(buf, true)
+	return strings.Count(string(buf[:n]), "backoff/v4.(*Ticker).run")
+}
+
+// TestGuard_ReconnectTicker_NoGoroutineLeakOnShutdown reproduces a observed
+// leak: after a shutdown burst, ticker run/send goroutines stay parked
+// forever even though every reconnect loop has exited.
+func TestGuard_ReconnectTicker_NoGoroutineLeakOnShutdown(t *testing.T) {
+	before := countBackoffTickerGoroutines()
+
+	const peers = 6000
+	cancels := make([]context.CancelFunc, 0, peers)
+	var wg sync.WaitGroup
+
+	// A status check slower than the tick cadence. This models the real
+	// isConnectedOnAllWay/callback doing work: while the loop is busy in the
+	// handler, the ticker fires the next tick and parks in send(), because
+	// send() never selects on ctx.
+	slowStatus := func() ConnStatus {
+		time.Sleep(70 * time.Millisecond)
+		return ConnStatusConnected
+	}
+
+	for range peers {
+		g := newTestGuard(slowStatus)
+		ctx, cancel := context.WithCancel(context.Background())
+		cancels = append(cancels, cancel)
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			g.Start(ctx, func() {})
+		}()
+		// Force the live ticker to be a newReconnectTicker.
+		g.SetRelayedConnDisconnected()
+	}
+
+	// Let the replacement tickers get past their 800ms initial interval, so
+	// many are parked in send() waiting on the (slow) consumer when we tear
+	// everything down.
+	time.Sleep(1500 * time.Millisecond)
+
+	// Shutdown burst: cancel every peer at once, like engine teardown.
+	for _, c := range cancels {
+		c()
+	}
+
+	// Every reconnect loop must return
+	waitCh := make(chan struct{})
+	go func() { wg.Wait(); close(waitCh) }()
+	select {
+	case <-waitCh:
+	case <-time.After(30 * time.Second):
+		t.Fatal("not all reconnect loops returned after ctx cancel")
+	}
+
+	// Give any correctly-stopped ticker goroutines time to unwind.
+	for range 50 {
+		runtime.Gosched()
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	leaked := countBackoffTickerGoroutines() - before
+	t.Logf("backoff Ticker.run goroutines still parked after teardown of %d peers: %d", peers, leaked)
+	if leaked > 0 {
+		t.Errorf("LEAK: %d backoff ticker goroutines parked after all reconnect loops exited "+
+			"(defer ticker.Stop() stops the initial ticker, not the live replacement)", leaked)
+	}
+}

client/internal/peer/status.go

@@ -813,19 +813,14 @@ func (d *Status) SetSessionExpiresAt(deadline time.Time) {
 }
 
 // GetSessionExpiresAt returns the most recently recorded SSO session deadline,
-// or the zero value when no deadline is tracked. A deadline that has already
-// slipped into the past reports as "none": once the session has expired it is
-// no longer a meaningful countdown, and the sessionwatch.Watcher does not
-// arm a timer at the deadline itself to clear it (only the two pre-expiry
-// warnings). Without this guard the UI would keep painting a stale
-// "expires in …" against a moment that has passed until the next login,
-// extend, or teardown rewrote the value.
+// or the zero value when no deadline is tracked. A deadline in the past is
+// returned as-is: it means the session has expired, and consumers (tray row,
+// CLI status) render it as "expired" rather than hiding it — masking it as
+// "none" would blank the UI at the exact moment it should say the session
+// ended.
 func (d *Status) GetSessionExpiresAt() time.Time {
 	d.mux.Lock()
 	defer d.mux.Unlock()
-	if !d.sessionExpiresAt.IsZero() && d.sessionExpiresAt.Before(time.Now()) {
-		return time.Time{}
-	}
 	return d.sessionExpiresAt
 }
 

client/internal/peer/wg_watcher.go

@@ -3,7 +3,6 @@ package peer
 import (
 	"context"
 	"fmt"
-	"sync"
 	"time"
 
 	log "github.com/sirupsen/logrus"
@@ -24,14 +23,14 @@ type WGInterfaceStater interface {
 	GetStats() (map[string]configurer.WGStats, error)
 }
 
+// WGWatcher is single-shot: one instance per connection attempt, run once, then discarded.
+// Lifecycle is owned by Conn under conn.mu, so it keeps no "enabled" state to go stale.
 type WGWatcher struct {
 	log           *log.Entry
 	wgIfaceStater WGInterfaceStater
 	peerKey       string
 	stateDump     *stateDump
 
-	enabled   bool
-	muEnabled sync.Mutex
 	// initialHandshake is not thread-safe; never call PrepareInitialHandshake and EnableWgWatcher concurrently.
 	initialHandshake time.Time
 
@@ -48,36 +47,23 @@ func NewWGWatcher(log *log.Entry, wgIfaceStater WGInterfaceStater, peerKey strin
 	}
 }
 
-// PrepareInitialHandshake reserves the watcher and 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. Returns ok=false if the watcher is already
-// running, in which case EnableWgWatcher must not be called.
-func (w *WGWatcher) PrepareInitialHandshake() (ok bool) {
-	w.muEnabled.Lock()
-	if w.enabled {
-		w.muEnabled.Unlock()
-		return false
-	}
-
+// 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()
-
 	handshake, _ := w.wgState()
 	w.initialHandshake = handshake
-	return true
 }
 
 // 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)
-
-	w.muEnabled.Lock()
-	w.enabled = false
-	w.muEnabled.Unlock()
+// for context lifecycle management. onHandshakeSuccessFn is called only for the first
+// handshake observed by this run, onCheckSuccessFn for every check that observed a fresh
+// handshake, including the first.
+func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time), onCheckSuccessFn func()) {
+	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, onCheckSuccessFn, enabledTime, w.initialHandshake)
 }
 
 // Reset signals the watcher that the WireGuard peer has been reset and a new
@@ -90,7 +76,7 @@ func (w *WGWatcher) Reset() {
 }
 
 // wgStateCheck help to check the state of the WireGuard handshake and relay connection
-func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time), enabledTime time.Time, initialHandshake time.Time) {
+func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time), onCheckSuccessFn func(), enabledTime time.Time, initialHandshake time.Time) {
 	w.log.Infof("WireGuard watcher started")
 
 	timer := time.NewTimer(wgHandshakeOvertime)
@@ -103,6 +89,7 @@ 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
 				}
@@ -117,6 +104,10 @@ func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn
 				}
 			}
 
+			if onCheckSuccessFn != nil && ctx.Err() == nil {
+				onCheckSuccessFn()
+			}
+
 			lastHandshake = *handshake
 
 			resetTime := time.Until(handshake.Add(checkPeriod))
@@ -147,9 +138,9 @@ 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

@@ -7,7 +7,6 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
-	"github.com/stretchr/testify/require"
 
 	"github.com/netbirdio/netbird/client/iface/configurer"
 )
@@ -24,6 +23,72 @@ func (m *MocWgIface) disconnect() {
 	m.stop = true
 }
 
+type mockHandshakeStats struct {
+	mu        sync.Mutex
+	handshake time.Time
+}
+
+func (m *mockHandshakeStats) GetStats() (map[string]configurer.WGStats, error) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return map[string]configurer.WGStats{"": {LastHandshake: m.handshake}}, nil
+}
+
+func (m *mockHandshakeStats) advance() {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	m.handshake = time.Now()
+}
+
+// TestWGWatcher_CheckSuccessCallback: onCheckSuccessFn must fire for a fresh
+// handshake even when the watcher started with an existing handshake baseline,
+// the case where onHandshakeSuccessFn stays silent.
+func TestWGWatcher_CheckSuccessCallback(t *testing.T) {
+	// checkPeriod bounds how stale a handshake may be before the watcher treats it
+	// as a suspended-machine timeout. The first check fires after wgHandshakeOvertime,
+	// so keep checkPeriod well above any scheduling jitter to avoid a false timeout
+	// converting the expected success into a disconnect on a loaded runner.
+	checkPeriod = 1 * time.Minute
+	wgHandshakeOvertime = 1 * time.Second
+
+	mlog := log.WithField("peer", "tet")
+	// Use an old baseline so advance() yields a strictly newer handshake even on
+	// platforms with coarse clock resolution (Windows), where two time.Now() calls
+	// microseconds apart can return the same instant and read as a timed-out handshake.
+	stats := &mockHandshakeStats{handshake: time.Now().Add(-time.Hour)}
+	watcher := NewWGWatcher(mlog, stats, "", newStateDump("peer", mlog, &Status{}))
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	watcher.PrepareInitialHandshake()
+
+	firstHandshake := make(chan struct{}, 1)
+	checkSuccess := make(chan struct{}, 1)
+	go watcher.EnableWgWatcher(ctx, time.Now(), func() {}, func(when time.Time) {
+		firstHandshake <- struct{}{}
+	}, func() {
+		select {
+		case checkSuccess <- struct{}{}:
+		default:
+		}
+	})
+
+	stats.advance()
+
+	select {
+	case <-checkSuccess:
+	case <-time.After(10 * time.Second):
+		t.Errorf("timeout waiting for check success callback")
+	}
+
+	select {
+	case <-firstHandshake:
+		t.Errorf("first-handshake callback must not fire for a non-zero baseline")
+	default:
+	}
+}
+
 func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 	checkPeriod = 5 * time.Second
 	wgHandshakeOvertime = 1 * time.Second
@@ -35,8 +100,7 @@ func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
+	watcher.PrepareInitialHandshake()
 
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
@@ -44,7 +108,7 @@ func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 		onDisconnected <- struct{}{}
 	}, func(when time.Time) {
 		mlog.Infof("onHandshakeSuccess: %v", when)
-	})
+	}, nil)
 
 	// wait for initial reading
 	time.Sleep(2 * time.Second)
@@ -66,14 +130,13 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	watcher := NewWGWatcher(mlog, mocWgIface, "", newStateDump("peer", mlog, &Status{}))
 
 	ctx, cancel := context.WithCancel(context.Background())
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
+	watcher.PrepareInitialHandshake()
 
 	wg := &sync.WaitGroup{}
 	wg.Add(1)
 	go func() {
 		defer wg.Done()
-		watcher.EnableWgWatcher(ctx, time.Now(), func() {}, func(when time.Time) {})
+		watcher.EnableWgWatcher(ctx, time.Now(), func() {}, func(when time.Time) {}, nil)
 	}()
 	cancel()
 
@@ -83,13 +146,12 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	ctx, cancel = context.WithCancel(context.Background())
 	defer cancel()
 
-	ok = watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should be re-enabled after the previous run stopped")
+	watcher.PrepareInitialHandshake()
 
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
 		onDisconnected <- struct{}{}
-	}, func(when time.Time) {})
+	}, func(when time.Time) {}, nil)
 
 	time.Sleep(2 * time.Second)
 	mocWgIface.disconnect()

client/internal/profilemanager/config.go

@@ -96,13 +96,12 @@ type ConfigInput struct {
 	BlockLANAccess      *bool
 	BlockInbound        *bool
 	DisableIPv6         *bool
+	SyncMessageVersion  *int
 
 	DisableNotifications *bool
 
 	DNSLabels domain.List
 
-	LazyConnectionEnabled *bool
-
 	MTU *uint16
 }
 
@@ -139,6 +138,7 @@ type Config struct {
 	BlockLANAccess      bool
 	BlockInbound        bool
 	DisableIPv6         bool
+	SyncMessageVersion  *int
 
 	DisableNotifications *bool
 
@@ -180,7 +180,9 @@ 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
 
@@ -587,6 +589,12 @@ func (config *Config) apply(input ConfigInput) (updated bool, err error) {
 		updated = true
 	}
 
+	if input.SyncMessageVersion != nil && *input.SyncMessageVersion != *config.SyncMessageVersion {
+		log.Infof("setting SyncMessageVersion to %v", *input.SyncMessageVersion)
+		*config.SyncMessageVersion = *input.SyncMessageVersion
+		updated = true
+	}
+
 	if input.DisableNotifications != nil && (config.DisableNotifications == nil || *input.DisableNotifications != *config.DisableNotifications) {
 		if *input.DisableNotifications {
 			log.Infof("disabling notifications")
@@ -632,12 +640,6 @@ 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 +730,15 @@ 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/profilemanager/service.go

@@ -11,6 +11,7 @@ import (
 	"runtime"
 	"sort"
 	"strings"
+	"syscall"
 
 	log "github.com/sirupsen/logrus"
 
@@ -439,7 +440,11 @@ func (s *ServiceManager) GetStatePath() string {
 
 	activeProf, err := s.GetActiveProfileState()
 	if err != nil {
-		log.Warnf("failed to get active profile state: %v", err)
+		if errors.Is(err, syscall.ENOSYS) {
+			log.Debugf("active profile state unavailable on this platform: %v", err)
+		} else {
+			log.Warnf("failed to get active profile state: %v", err)
+		}
 		return defaultStatePath
 	}
 

client/internal/rosenpass/manager.go

@@ -39,6 +39,7 @@ type rpServer interface {
 
 type Manager struct {
 	ifaceName    string
+	localWgKey   wgtypes.Key
 	spk          []byte
 	ssk          []byte
 	rpKeyHash    string
@@ -51,8 +52,9 @@ type Manager struct {
 	wgIface      PresharedKeySetter
 }
 
-// NewManager creates a new Rosenpass manager
-func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string) (*Manager, error) {
+// NewManager creates a new Rosenpass manager. localWgKey is the local
+// WireGuard public key, used to derive the per-peer rendezvous key.
+func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string, localWgKey wgtypes.Key) (*Manager, error) {
 	public, secret, err := rp.GenerateKeyPair()
 	if err != nil {
 		return nil, err
@@ -62,6 +64,7 @@ func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string) (*Manager, error)
 	log.Tracef("generated new rosenpass key pair with public key %s", rpKeyHash)
 	return &Manager{
 		ifaceName:    wgIfaceName,
+		localWgKey:   localWgKey,
 		rpKeyHash:    rpKeyHash,
 		spk:          public,
 		ssk:          secret,
@@ -73,7 +76,7 @@ func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string) (*Manager, error)
 		// nil receiver in addPeer -> m.rpWgHandler.AddPeer. generateConfig will
 		// replace it with a fresh handler on each Run() to clear stale peer
 		// state from previous engine sessions.
-		rpWgHandler: NewNetbirdHandler(),
+		rpWgHandler: NewNetbirdHandler((*[32]byte)(preSharedKey), localWgKey),
 		lock:        sync.Mutex{},
 	}, nil
 }
@@ -161,7 +164,7 @@ func (m *Manager) generateConfig() (rp.Config, error) {
 	cfg.Peers = []rp.PeerConfig{}
 
 	m.lock.Lock()
-	m.rpWgHandler = NewNetbirdHandler()
+	m.rpWgHandler = NewNetbirdHandler(m.preSharedKey, m.localWgKey)
 	if m.wgIface != nil {
 		m.rpWgHandler.SetInterface(m.wgIface)
 	}

client/internal/rosenpass/manager_test.go

@@ -85,7 +85,7 @@ func newTestManager(spkFirstByte byte, mock *mockServer) *Manager {
 		ssk:         make([]byte, 32),
 		rpKeyHash:   "test-hash",
 		rpPeerIDs:   make(map[string]*rp.PeerID),
-		rpWgHandler: NewNetbirdHandler(),
+		rpWgHandler: NewNetbirdHandler(nil, wgtypes.Key{0x01}),
 		server:      mock,
 	}
 }
@@ -255,7 +255,7 @@ func TestAddPeer_NilServer_ReturnsErrorNoCrash(t *testing.T) {
 // issue #4341 cannot occur in the window between NewManager and Run().
 func TestNewManager_PreInitializesHandler(t *testing.T) {
 	psk := wgtypes.Key{}
-	m, err := NewManager(&psk, "wt0")
+	m, err := NewManager(&psk, "wt0", wgtypes.Key{0x01})
 	require.NoError(t, err)
 	require.NotNil(t, m.rpWgHandler, "rpWgHandler must be initialized in NewManager")
 }
@@ -329,10 +329,10 @@ func TestIsPresharedKeyInitialized_AddedButNotHandshaken_ReturnsFalse(t *testing
 	require.False(t, m.IsPresharedKeyInitialized(wgKey))
 }
 
-// --- NetbirdHandler.outputKey ----------------------------------------------
+// --- NetbirdHandler.applyKey ----------------------------------------------
 
-func TestHandler_OutputKey_FirstCallUsesUpdateOnlyFalse(t *testing.T) {
-	h := NewNetbirdHandler()
+func TestHandler_ApplyKey_FirstCallUsesUpdateOnlyFalse(t *testing.T) {
+	h := NewNetbirdHandler(nil, wgtypes.Key{0x01})
 	iface := &mockIface{}
 	h.SetInterface(iface)
 
@@ -348,8 +348,8 @@ func TestHandler_OutputKey_FirstCallUsesUpdateOnlyFalse(t *testing.T) {
 	require.Equal(t, wgKey.String(), iface.calls[0].peerKey)
 }
 
-func TestHandler_OutputKey_SubsequentCallsUseUpdateOnlyTrue(t *testing.T) {
-	h := NewNetbirdHandler()
+func TestHandler_ApplyKey_SubsequentCallsUseUpdateOnlyTrue(t *testing.T) {
+	h := NewNetbirdHandler(nil, wgtypes.Key{0x01})
 	iface := &mockIface{}
 	h.SetInterface(iface)
 
@@ -364,8 +364,8 @@ func TestHandler_OutputKey_SubsequentCallsUseUpdateOnlyTrue(t *testing.T) {
 	require.True(t, iface.calls[1].updateOnly, "subsequent rotations must use updateOnly=true")
 }
 
-func TestHandler_OutputKey_NilInterface_NoCrashNoCall(t *testing.T) {
-	h := NewNetbirdHandler()
+func TestHandler_ApplyKey_NilInterface_NoCrashNoCall(t *testing.T) {
+	h := NewNetbirdHandler(nil, wgtypes.Key{0x01})
 	// no SetInterface — iface remains nil
 	pid := rp.PeerID{0x03}
 	h.AddPeer(pid, "wt0", rp.Key(wgtypes.Key{}))
@@ -374,8 +374,8 @@ func TestHandler_OutputKey_NilInterface_NoCrashNoCall(t *testing.T) {
 	h.HandshakeCompleted(pid, rp.Key{})
 }
 
-func TestHandler_OutputKey_UnknownPeer_NoCall(t *testing.T) {
-	h := NewNetbirdHandler()
+func TestHandler_ApplyKey_UnknownPeer_NoCall(t *testing.T) {
+	h := NewNetbirdHandler(nil, wgtypes.Key{0x01})
 	iface := &mockIface{}
 	h.SetInterface(iface)
 
@@ -384,7 +384,7 @@ func TestHandler_OutputKey_UnknownPeer_NoCall(t *testing.T) {
 }
 
 func TestHandler_RemovePeer_ClearsInitializedState(t *testing.T) {
-	h := NewNetbirdHandler()
+	h := NewNetbirdHandler(nil, wgtypes.Key{0x01})
 	iface := &mockIface{}
 	h.SetInterface(iface)
 
@@ -398,7 +398,7 @@ func TestHandler_RemovePeer_ClearsInitializedState(t *testing.T) {
 }
 
 func TestHandler_SetInterfaceAfterAddPeer_StillReceivesKey(t *testing.T) {
-	h := NewNetbirdHandler()
+	h := NewNetbirdHandler(nil, wgtypes.Key{0x01})
 	pid := rp.PeerID{0x05}
 	wgKey := wgtypes.Key{0xEE}
 	h.AddPeer(pid, "wt0", rp.Key(wgKey))

client/internal/rosenpass/netbird_handler.go

@@ -18,19 +18,34 @@ type PresharedKeySetter interface {
 type wireGuardPeer struct {
 	Interface string
 	PublicKey rp.Key
+	// initialized is true once a completed exchange has set a
+	// Rosenpass-managed PSK for this peer.
+	initialized bool
+	// chainKey is the key output by the last completed exchange, advanced by
+	// one ratchet step on expiry. Nil until the first exchange completes and
+	// after the peer has fallen back to the rendezvous key.
+	chainKey *wgtypes.Key
+	// expiries counts failed renewals since the last completed exchange.
+	expiries int
 }
 
 type NetbirdHandler struct {
-	mu               sync.Mutex
-	iface            PresharedKeySetter
-	peers            map[rp.PeerID]wireGuardPeer
-	initializedPeers map[rp.PeerID]bool
+	mu    sync.Mutex
+	iface PresharedKeySetter
+	// preSharedKey is the account-level preshared key, used as the rendezvous
+	// key when set. Nil means the deterministic seed key is used instead.
+	preSharedKey *[32]byte
+	// localWgKey is the local WireGuard public key, one of the two inputs to
+	// the deterministic seed key.
+	localWgKey wgtypes.Key
+	peers      map[rp.PeerID]*wireGuardPeer
 }
 
-func NewNetbirdHandler() *NetbirdHandler {
+func NewNetbirdHandler(preSharedKey *[32]byte, localWgKey wgtypes.Key) *NetbirdHandler {
 	return &NetbirdHandler{
-		peers:            map[rp.PeerID]wireGuardPeer{},
-		initializedPeers: map[rp.PeerID]bool{},
+		preSharedKey: preSharedKey,
+		localWgKey:   localWgKey,
+		peers:        map[rp.PeerID]*wireGuardPeer{},
 	}
 }
 
@@ -42,10 +57,16 @@ func (h *NetbirdHandler) SetInterface(iface PresharedKeySetter) {
 	h.iface = iface
 }
 
+// AddPeer registers a peer with the handler. Re-adding a known peer (every
+// reconnection does) keeps its key recovery state.
 func (h *NetbirdHandler) AddPeer(pid rp.PeerID, intf string, pk rp.Key) {
 	h.mu.Lock()
 	defer h.mu.Unlock()
-	h.peers[pid] = wireGuardPeer{
+	if existing, ok := h.peers[pid]; ok && existing.PublicKey == pk {
+		existing.Interface = intf
+		return
+	}
+	h.peers[pid] = &wireGuardPeer{
 		Interface: intf,
 		PublicKey: pk,
 	}
@@ -55,7 +76,6 @@ func (h *NetbirdHandler) RemovePeer(pid rp.PeerID) {
 	h.mu.Lock()
 	defer h.mu.Unlock()
 	delete(h.peers, pid)
-	delete(h.initializedPeers, pid)
 }
 
 // IsPeerInitialized returns true if Rosenpass has completed a handshake
@@ -63,50 +83,120 @@ func (h *NetbirdHandler) RemovePeer(pid rp.PeerID) {
 func (h *NetbirdHandler) IsPeerInitialized(pid rp.PeerID) bool {
 	h.mu.Lock()
 	defer h.mu.Unlock()
-	return h.initializedPeers[pid]
+	peer, ok := h.peers[pid]
+	return ok && peer.initialized
 }
 
+// HandshakeCompleted programs the freshly exchanged output key and resets the
+// peer's key recovery state.
 func (h *NetbirdHandler) HandshakeCompleted(pid rp.PeerID, key rp.Key) {
-	h.outputKey(rp.KeyOutputReasonStale, pid, key)
-}
+	psk := wgtypes.Key(key)
 
-func (h *NetbirdHandler) HandshakeExpired(pid rp.PeerID) {
-	key, _ := rp.GeneratePresharedKey()
-	h.outputKey(rp.KeyOutputReasonStale, pid, key)
-}
-
-func (h *NetbirdHandler) outputKey(_ rp.KeyOutputReason, pid rp.PeerID, psk rp.Key) {
 	h.mu.Lock()
-	iface := h.iface
-	wg, ok := h.peers[pid]
-	isInitialized := h.initializedPeers[pid]
-	h.mu.Unlock()
+	defer h.mu.Unlock()
 
-	if iface == nil {
-		log.Warn("rosenpass: interface not set, cannot update preshared key")
+	peer, ok := h.peers[pid]
+	if !ok {
 		return
 	}
+	if peer.expiries > 0 {
+		log.Infof("rosenpass exchange completed for peer %s after %d expired renewals", wgtypes.Key(peer.PublicKey), peer.expiries)
+	}
+	// chainKey tracks the shared exchange output regardless of the local write
+	// outcome, so both ends still converge on the next expiry.
+	peer.chainKey = &psk
+	peer.expiries = 0
+	if !h.applyKeyLocked(pid, psk, peer.initialized) {
+		return
+	}
+	peer.initialized = true
+}
 
+// HandshakeExpired replaces the expired key. The renewal exchange runs over
+// the tunnel keyed by the PSK itself, so the replacement must be derivable on
+// both ends without communication: the first expiry ratchets the last shared
+// key forward, repeated expiries (and expiries without a completed exchange)
+// fall back to the rendezvous key and drop the peer out of the initialized
+// state so connection reconfigurations reprogram the rendezvous key as well.
+func (h *NetbirdHandler) HandshakeExpired(pid rp.PeerID) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+
+	peer, ok := h.peers[pid]
 	if !ok {
 		return
 	}
 
-	peerKey := wgtypes.Key(wg.PublicKey).String()
-	pskKey := wgtypes.Key(psk)
+	peer.expiries++
 
-	// Use updateOnly=true for later rotations (peer already has Rosenpass PSK)
-	// Use updateOnly=false for first rotation (peer has original/empty PSK)
-	if err := iface.SetPresharedKey(peerKey, pskKey, isInitialized); err != nil {
+	var psk wgtypes.Key
+	if peer.chainKey != nil && peer.expiries == 1 {
+		log.Infof("rosenpass key for peer %s expired without renewal, advancing to ratcheted key", wgtypes.Key(peer.PublicKey))
+		psk = RatchetKey(*peer.chainKey)
+		peer.chainKey = &psk
+	} else {
+		rendezvous, err := h.rendezvousKey(peer)
+		if err != nil {
+			// Fail closed: without a rendezvous key the expired key must
+			// still be rotated out, even if the replacement is unusable.
+			log.Errorf("failed to derive rendezvous key, replacing expired key with a random one: %v", err)
+			h.applyRandomKeyLocked(pid)
+			return
+		}
+		log.Warnf("rosenpass key for peer %s expired %d times without renewal, falling back to the rendezvous key", wgtypes.Key(peer.PublicKey), peer.expiries)
+		psk = rendezvous
+		peer.chainKey = nil
+		peer.initialized = false
+	}
+
+	h.applyKeyLocked(pid, psk, true)
+}
+
+// rendezvousKey returns the key both ends converge on without communication:
+// the account-level preshared key when configured, the deterministic seed key
+// otherwise. It mirrors the key that peer connections program when Rosenpass
+// does not manage the peer yet.
+func (h *NetbirdHandler) rendezvousKey(peer *wireGuardPeer) (wgtypes.Key, error) {
+	if h.preSharedKey != nil {
+		return *h.preSharedKey, nil
+	}
+
+	seed, err := DeterministicSeedKey(h.localWgKey.String(), wgtypes.Key(peer.PublicKey).String())
+	if err != nil {
+		return wgtypes.Key{}, err
+	}
+	return *seed, nil
+}
+
+// applyKeyLocked writes the preshared key for the peer to the WireGuard
+// interface and reports whether the write succeeded. Callers must hold h.mu
+// for the whole state-mutation-plus-write so that a concurrent completion and
+// expiry cannot reorder their writes relative to the in-memory chain key.
+func (h *NetbirdHandler) applyKeyLocked(pid rp.PeerID, psk wgtypes.Key, updateOnly bool) bool {
+	peer, ok := h.peers[pid]
+	if !ok {
+		return false
+	}
+
+	if h.iface == nil {
+		log.Warn("rosenpass: interface not set, cannot update preshared key")
+		return false
+	}
+
+	peerKey := wgtypes.Key(peer.PublicKey).String()
+	if err := h.iface.SetPresharedKey(peerKey, psk, updateOnly); err != nil {
 		log.Errorf("Failed to apply rosenpass key: %v", err)
+		return false
+	}
+
+	return true
+}
+
+func (h *NetbirdHandler) applyRandomKeyLocked(pid rp.PeerID) {
+	key, err := rp.GeneratePresharedKey()
+	if err != nil {
+		log.Errorf("failed to generate random preshared key: %v", err)
 		return
 	}
-
-	// Mark peer as isInitialized after the successful first rotation
-	if !isInitialized {
-		h.mu.Lock()
-		if _, exists := h.peers[pid]; exists {
-			h.initializedPeers[pid] = true
-		}
-		h.mu.Unlock()
-	}
+	h.applyKeyLocked(pid, wgtypes.Key(key), true)
 }

client/internal/rosenpass/netbird_handler_test.go

@@ -0,0 +1,250 @@
+package rosenpass
+
+import (
+	"testing"
+
+	rp "cunicu.li/go-rosenpass"
+	"github.com/stretchr/testify/require"
+	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
+)
+
+// handlerTestLink wires two NetbirdHandlers as the two ends of a single
+// tunnel: handler A manages the rosenpass peer B and vice versa, the way two
+// NetBird clients see each other.
+type handlerTestLink struct {
+	handlerA, handlerB *NetbirdHandler
+	ifaceA, ifaceB     *mockIface
+	pidA, pidB         rp.PeerID
+	wgKeyA, wgKeyB     wgtypes.Key
+}
+
+func newHandlerTestLink(t *testing.T, preSharedKey *[32]byte) *handlerTestLink {
+	t.Helper()
+
+	link := &handlerTestLink{
+		ifaceA: &mockIface{},
+		ifaceB: &mockIface{},
+	}
+	link.pidA[0] = 0xaa
+	link.pidB[0] = 0xbb
+	link.wgKeyA[31] = 1
+	link.wgKeyB[31] = 2
+
+	link.handlerA = NewNetbirdHandler(preSharedKey, link.wgKeyA)
+	link.handlerB = NewNetbirdHandler(preSharedKey, link.wgKeyB)
+
+	link.handlerA.SetInterface(link.ifaceA)
+	link.handlerB.SetInterface(link.ifaceB)
+
+	link.handlerA.AddPeer(link.pidB, "wt0", rp.Key(link.wgKeyB))
+	link.handlerB.AddPeer(link.pidA, "wt0", rp.Key(link.wgKeyA))
+
+	return link
+}
+
+// complete simulates a completed rosenpass exchange: both ends derive the
+// same output key.
+func (l *handlerTestLink) complete(osk rp.Key) {
+	l.handlerA.HandshakeCompleted(l.pidB, osk)
+	l.handlerB.HandshakeCompleted(l.pidA, osk)
+}
+
+// expire simulates a failed key renewal on both ends.
+func (l *handlerTestLink) expire() {
+	l.handlerA.HandshakeExpired(l.pidB)
+	l.handlerB.HandshakeExpired(l.pidA)
+}
+
+func lastPSK(t *testing.T, m *mockIface) wgtypes.Key {
+	t.Helper()
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	require.NotEmpty(t, m.calls, "expected at least one SetPresharedKey call")
+	return m.calls[len(m.calls)-1].psk
+}
+
+func TestHandshakeCompleted_SetsKeyAndInitializes(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	require.Equal(t, wgtypes.Key(osk), lastPSK(t, link.ifaceA), "completed exchange must program the osk")
+	require.False(t, link.ifaceA.calls[0].updateOnly, "first rotation must not be update-only")
+	require.True(t, link.handlerA.IsPeerInitialized(link.pidB), "peer must be initialized after first completed exchange")
+
+	link.complete(osk)
+	require.True(t, link.ifaceA.calls[1].updateOnly, "later rotations must be update-only")
+}
+
+// TestHandshakeExpired_BothSidesConverge encodes the core recovery invariant:
+// rosenpass renewals run over the tunnel that the PSK itself keys, so when a
+// renewal fails on both ends, both ends must fall back to the same key or the
+// tunnel can never handshake again.
+func TestHandshakeExpired_BothSidesConverge(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	link.expire()
+	keyA := lastPSK(t, link.ifaceA)
+	keyB := lastPSK(t, link.ifaceB)
+	require.NotEqual(t, wgtypes.Key(osk), keyA, "expired key must be rotated out")
+	require.Equal(t, keyA, keyB, "both ends must converge on the same key after expiry")
+
+	link.expire()
+	require.Equal(t, lastPSK(t, link.ifaceA), lastPSK(t, link.ifaceB),
+		"both ends must still converge after repeated expiries")
+}
+
+// TestHandshakeExpired_ExpiryWithoutCompletionConverges covers the bootstrap
+// case: the initial exchange never completed (the tunnel ran on the rendezvous
+// key), so an expiry must not replace the working key with an unrecoverable
+// one on either end.
+func TestHandshakeExpired_ExpiryWithoutCompletionConverges(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	link.expire()
+	require.Equal(t, lastPSK(t, link.ifaceA), lastPSK(t, link.ifaceB),
+		"both ends must converge when the exchange never completed")
+}
+
+// TestHandshakeExpired_RepeatedExpiryClearsInitialized: once renewals keep
+// failing, the peer must drop out of the initialized state so the next
+// connection reconfiguration reprograms the rendezvous key instead of
+// preserving a poisoned rosenpass-managed key.
+func TestHandshakeExpired_RepeatedExpiryClearsInitialized(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	link.expire()
+	link.expire()
+
+	require.False(t, link.handlerA.IsPeerInitialized(link.pidB),
+		"repeated expiries must clear the initialized state")
+	require.False(t, link.handlerB.IsPeerInitialized(link.pidA),
+		"repeated expiries must clear the initialized state")
+}
+
+// TestHandshakeCompleted_AfterExpiryRecovers: a completed exchange after a
+// desync must fully reset the recovery state.
+func TestHandshakeCompleted_AfterExpiryRecovers(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk1, osk2 rp.Key
+	osk1[0] = 1
+	osk2[0] = 2
+
+	link.complete(osk1)
+	link.expire()
+	link.expire()
+
+	link.complete(osk2)
+	require.Equal(t, wgtypes.Key(osk2), lastPSK(t, link.ifaceA), "new exchange must program the fresh osk")
+	require.True(t, link.handlerA.IsPeerInitialized(link.pidB), "peer must be initialized again after recovery")
+
+	link.expire()
+	require.Equal(t, lastPSK(t, link.ifaceA), lastPSK(t, link.ifaceB),
+		"recovered link must converge again on the next expiry")
+	require.NotEqual(t, wgtypes.Key(osk2), lastPSK(t, link.ifaceA), "expired key must be rotated out")
+}
+
+// TestHandshakeExpired_FirstExpiryRatchetsLastKey: the first expiry must
+// derive the replacement from the last shared key, so an attacker who only
+// blocks the renewal exchange gains nothing over the previous key.
+func TestHandshakeExpired_FirstExpiryRatchetsLastKey(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	link.expire()
+	require.Equal(t, RatchetKey(wgtypes.Key(osk)), lastPSK(t, link.ifaceA),
+		"first expiry must program the ratcheted key")
+	require.True(t, link.handlerA.IsPeerInitialized(link.pidB),
+		"ratchet step must keep the peer initialized so reconfigurations preserve the key")
+}
+
+// TestHandshakeExpired_RepeatedExpiryFallsBackToSeed: once the ratchet key
+// also fails, both ends must land on the same key that peer connections
+// program for uninitialized peers, so a reconnect completes the recovery.
+func TestHandshakeExpired_RepeatedExpiryFallsBackToSeed(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	link.expire()
+	link.expire()
+
+	seed, err := DeterministicSeedKey(link.wgKeyA.String(), link.wgKeyB.String())
+	require.NoError(t, err)
+	require.Equal(t, *seed, lastPSK(t, link.ifaceA), "repeated expiry must fall back to the seed key")
+	require.Equal(t, *seed, lastPSK(t, link.ifaceB), "repeated expiry must fall back to the seed key")
+}
+
+// TestHandshakeExpired_ConfiguredPSKUsedAsRendezvous: with an account-level
+// preshared key configured, the fallback must be that key, matching what peer
+// connections program for uninitialized peers.
+func TestHandshakeExpired_ConfiguredPSKUsedAsRendezvous(t *testing.T) {
+	psk := &[32]byte{0x77}
+	link := newHandlerTestLink(t, psk)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	link.expire()
+	link.expire()
+
+	require.Equal(t, wgtypes.Key(*psk), lastPSK(t, link.ifaceA),
+		"fallback must be the configured preshared key")
+	require.Equal(t, wgtypes.Key(*psk), lastPSK(t, link.ifaceB),
+		"fallback must be the configured preshared key on both ends")
+}
+
+// TestHandshakeExpired_ExpiryWritesAreUpdateOnly: expiry replacements must
+// never create a WireGuard peer that connection management has removed.
+func TestHandshakeExpired_ExpiryWritesAreUpdateOnly(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+
+	link.expire()
+	link.expire()
+
+	for _, call := range link.ifaceA.calls[1:] {
+		require.True(t, call.updateOnly, "expiry writes must be update-only")
+	}
+}
+
+// TestAddPeer_ReAddKeepsRecoveryState: reconnections re-add the peer on every
+// OnConnected; that must not reset the expiry chain state.
+func TestAddPeer_ReAddKeepsRecoveryState(t *testing.T) {
+	link := newHandlerTestLink(t, nil)
+
+	var osk rp.Key
+	osk[0] = 0x42
+	link.complete(osk)
+	link.expire()
+
+	link.handlerA.AddPeer(link.pidB, "wt0", rp.Key(link.wgKeyB))
+	require.True(t, link.handlerA.IsPeerInitialized(link.pidB),
+		"re-adding a known peer must keep its state")
+
+	link.expire()
+	seed, err := DeterministicSeedKey(link.wgKeyA.String(), link.wgKeyB.String())
+	require.NoError(t, err)
+	require.Equal(t, *seed, lastPSK(t, link.ifaceA),
+		"second expiry after re-add must continue to the seed fallback")
+}

client/internal/rosenpass/seed.go

@@ -1,11 +1,28 @@
 package rosenpass
 
 import (
+	"crypto/sha256"
 	"fmt"
 
 	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
 )
 
+// ratchetLabel domain-separates the expiry ratchet from other uses of the
+// rosenpass output key.
+const ratchetLabel = "netbird-rosenpass-expiry-ratchet"
+
+// RatchetKey derives the successor preshared key from the previous Rosenpass
+// output key. When a key expires without a completed renewal, both peers
+// advance their last shared key by one ratchet step: the expired key is
+// rotated out while both ends still converge on an identical, non-public
+// replacement without communicating.
+func RatchetKey(prev wgtypes.Key) wgtypes.Key {
+	input := make([]byte, 0, len(ratchetLabel)+len(prev))
+	input = append(input, ratchetLabel...)
+	input = append(input, prev[:]...)
+	return sha256.Sum256(input)
+}
+
 // DeterministicSeedKey derives a 32-byte WireGuard preshared key from a pair
 // of peer public keys. Both peers, given the same key pair, produce the same
 // output regardless of which side runs the function: the inputs are ordered

client/internal/routemanager/dynamic/route.go

@@ -185,7 +185,7 @@ func (r *Route) startResolver(ctx context.Context) {
 }
 
 func (r *Route) update(ctx context.Context) error {
-	resolved, err := r.resolveDomains()
+	resolved, err := r.resolveDomains(ctx)
 	if err != nil {
 		if len(resolved) == 0 {
 			return fmt.Errorf("resolve domains: %w", err)
@@ -199,9 +199,9 @@ func (r *Route) update(ctx context.Context) error {
 	return nil
 }
 
-func (r *Route) resolveDomains() (domainMap, error) {
+func (r *Route) resolveDomains(ctx context.Context) (domainMap, error) {
 	results := make(chan resolveResult)
-	go r.resolve(results)
+	go r.resolve(ctx, results)
 
 	resolved := domainMap{}
 	var merr *multierror.Error
@@ -217,7 +217,7 @@ func (r *Route) resolveDomains() (domainMap, error) {
 	return resolved, nberrors.FormatErrorOrNil(merr)
 }
 
-func (r *Route) resolve(results chan resolveResult) {
+func (r *Route) resolve(ctx context.Context, results chan resolveResult) {
 	var wg sync.WaitGroup
 
 	for _, d := range r.route.Domains {
@@ -225,10 +225,10 @@ func (r *Route) resolve(results chan resolveResult) {
 		go func(domain domain.Domain) {
 			defer wg.Done()
 
-			ips, err := r.getIPsFromResolver(domain)
+			ips, err := r.getIPsFromResolver(ctx, domain)
 			if err != nil {
 				log.Tracef("Failed to resolve domain %s with private resolver: %v", domain.SafeString(), err)
-				ips, err = net.LookupIP(domain.PunycodeString())
+				ips, err = lookupHostIPs(ctx, domain)
 				if err != nil {
 					results <- resolveResult{domain: domain, err: fmt.Errorf("resolve d %s: %w", domain.SafeString(), err)}
 					return
@@ -364,6 +364,20 @@ func determinePrefixChanges(oldPrefixes, newPrefixes []netip.Prefix) (toAdd, toR
 	return
 }
 
+// lookupHostIPs resolves d via the system resolver, honoring ctx cancellation.
+func lookupHostIPs(ctx context.Context, d domain.Domain) ([]net.IP, error) {
+	addrs, err := net.DefaultResolver.LookupIPAddr(ctx, d.PunycodeString())
+	if err != nil {
+		return nil, err
+	}
+
+	ips := make([]net.IP, 0, len(addrs))
+	for _, addr := range addrs {
+		ips = append(ips, addr.IP)
+	}
+	return ips, nil
+}
+
 func combinePrefixes(oldPrefixes, removedPrefixes, addedPrefixes []netip.Prefix) []netip.Prefix {
 	prefixSet := make(map[netip.Prefix]struct{})
 	for _, prefix := range oldPrefixes {

client/internal/routemanager/dynamic/route_generic.go

@@ -3,11 +3,12 @@
 package dynamic
 
 import (
+	"context"
 	"net"
 
 	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
-func (r *Route) getIPsFromResolver(domain domain.Domain) ([]net.IP, error) {
-	return net.LookupIP(domain.PunycodeString())
+func (r *Route) getIPsFromResolver(ctx context.Context, domain domain.Domain) ([]net.IP, error) {
+	return lookupHostIPs(ctx, domain)
 }

client/internal/routemanager/dynamic/route_ios.go

@@ -3,6 +3,7 @@
 package dynamic
 
 import (
+	"context"
 	"fmt"
 	"net"
 	"time"
@@ -16,7 +17,7 @@ import (
 
 const dialTimeout = 10 * time.Second
 
-func (r *Route) getIPsFromResolver(domain domain.Domain) ([]net.IP, error) {
+func (r *Route) getIPsFromResolver(ctx context.Context, domain domain.Domain) ([]net.IP, error) {
 	privateClient, err := nbdns.GetClientPrivate(r.wgInterface, r.resolverAddr.Addr(), dialTimeout)
 	if err != nil {
 		return nil, fmt.Errorf("error while creating private client: %s", err)
@@ -32,7 +33,7 @@ func (r *Route) getIPsFromResolver(domain domain.Domain) ([]net.IP, error) {
 		msg := new(dns.Msg)
 		msg.SetQuestion(fqdn, qtype)
 
-		response, _, err := nbdns.ExchangeWithFallback(nil, privateClient, msg, r.resolverAddr.String())
+		response, _, err := nbdns.ExchangeWithFallback(ctx, privateClient, msg, r.resolverAddr.String())
 		if err != nil {
 			if queryErr == nil {
 				queryErr = fmt.Errorf("DNS query for %s (type %d) after %s: %w", domain.SafeString(), qtype, time.Since(startTime), err)

client/internal/routemanager/manager.go

@@ -12,6 +12,7 @@ import (
 	"strings"
 	"sync"
 	"sync/atomic"
+	"syscall"
 	"time"
 
 	"github.com/google/uuid"
@@ -60,6 +61,7 @@ type Manager interface {
 	InitialRouteRange() []string
 	SetFirewall(firewall.Manager) error
 	SetDNSForwarderPort(port uint16)
+	ReconcilePeerAllowedIPs(peerKey string) error
 	Stop(stateManager *statemanager.Manager)
 }
 
@@ -231,6 +233,30 @@ func (m *DefaultManager) setupRefCounters(useNoop bool) {
 	)
 }
 
+// ReconcilePeerAllowedIPs re-applies every routed allowed IP currently tracked for the peer
+// onto the WireGuard device. The allowed-IP refcounter only calls its AddFunc (which pushes to
+// the device) on a prefix's 0->1 transition, so a peer whose device entry was rebuilt without a
+// matching refcounter change — e.g. a lazy connection cycling through idle->wake, which recreates
+// the WireGuard peer with the overlay /32 only — ends up missing routed prefixes the refcounter
+// still considers installed, and nothing retries. Calling this when the peer's WireGuard entry is
+// (re)created restores convergence. It is add-only and idempotent: AddAllowedIP is update-only, so
+// prefixes are re-added to an existing peer and an absent peer is left untouched.
+func (m *DefaultManager) ReconcilePeerAllowedIPs(peerKey string) error {
+	if m.allowedIPsRefCounter == nil {
+		return nil
+	}
+
+	return m.allowedIPsRefCounter.ReapplyMatching(
+		func(out string) bool { return out == peerKey },
+		func(prefix netip.Prefix) error {
+			if err := m.wgInterface.AddAllowedIP(peerKey, prefix); err != nil {
+				return fmt.Errorf("add allowed IP %s for peer %s: %w", prefix, peerKey, err)
+			}
+			return nil
+		},
+	)
+}
+
 // Init sets up the routing
 func (m *DefaultManager) Init() error {
 	m.routeSelector = m.initSelector()
@@ -264,7 +290,11 @@ func (m *DefaultManager) initSelector() *routeselector.RouteSelector {
 
 	// restore selector state if it exists
 	if err := m.stateManager.LoadState(state); err != nil {
-		log.Warnf("failed to load state: %v", err)
+		if errors.Is(err, syscall.ENOSYS) {
+			log.Debugf("route selector state unavailable on this platform: %v", err)
+		} else {
+			log.Warnf("failed to load state: %v", err)
+		}
 		return routeselector.NewRouteSelector()
 	}
 

client/internal/routemanager/mock.go

@@ -112,6 +112,11 @@ func (m *MockManager) SetFirewall(firewall.Manager) error {
 func (m *MockManager) SetDNSForwarderPort(port uint16) {
 }
 
+// ReconcilePeerAllowedIPs mock implementation of ReconcilePeerAllowedIPs from Manager interface
+func (m *MockManager) ReconcilePeerAllowedIPs(peerKey string) error {
+	return nil
+}
+
 // Stop mock implementation of Stop from Manager interface
 func (m *MockManager) Stop(stateManager *statemanager.Manager) {
 	if m.StopFunc != nil {

client/internal/routemanager/reconcile_test.go

@@ -0,0 +1,90 @@
+//go:build !windows
+
+package routemanager
+
+import (
+	"net"
+	"net/netip"
+	"sync"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.zx2c4.com/wireguard/tun/netstack"
+
+	"github.com/netbirdio/netbird/client/iface/device"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+	"github.com/netbirdio/netbird/client/internal/routemanager/refcounter"
+)
+
+// reconcileWGMock is a minimal iface.WGIface that only records AddAllowedIP calls; every other
+// method is an inert stub because ReconcilePeerAllowedIPs exercises none of them.
+type reconcileWGMock struct {
+	mu   sync.Mutex
+	adds map[string][]netip.Prefix
+}
+
+func (m *reconcileWGMock) AddAllowedIP(peerKey string, allowedIP netip.Prefix) error {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if m.adds == nil {
+		m.adds = map[string][]netip.Prefix{}
+	}
+	m.adds[peerKey] = append(m.adds[peerKey], allowedIP)
+	return nil
+}
+
+func (m *reconcileWGMock) added(peerKey string) []netip.Prefix {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.adds[peerKey]
+}
+
+func (m *reconcileWGMock) RemoveAllowedIP(string, netip.Prefix) error { return nil }
+func (m *reconcileWGMock) Name() string                               { return "utun-test" }
+func (m *reconcileWGMock) Address() wgaddr.Address                    { return wgaddr.Address{} }
+func (m *reconcileWGMock) ToInterface() *net.Interface                { return nil }
+func (m *reconcileWGMock) IsUserspaceBind() bool                      { return false }
+func (m *reconcileWGMock) GetFilter() device.PacketFilter             { return nil }
+func (m *reconcileWGMock) GetDevice() *device.FilteredDevice          { return nil }
+func (m *reconcileWGMock) GetNet() *netstack.Net                      { return nil }
+
+// TestReconcilePeerAllowedIPs verifies the declarative reconcile re-applies every routed prefix
+// tracked for the peer (self-heal, independent of refcount level) and stays scoped to that peer.
+func TestReconcilePeerAllowedIPs(t *testing.T) {
+	wg := &reconcileWGMock{}
+	m := &DefaultManager{wgInterface: wg}
+	m.allowedIPsRefCounter = refcounter.New[netip.Prefix, string, string](
+		func(_ netip.Prefix, peerKey string) (string, error) { return peerKey, nil },
+		func(netip.Prefix, string) error { return nil },
+	)
+
+	peerA1 := netip.MustParsePrefix("10.0.0.0/24")
+	peerA2 := netip.MustParsePrefix("10.1.0.0/24")
+	peerB1 := netip.MustParsePrefix("10.2.0.0/24")
+
+	for prefix, peer := range map[netip.Prefix]string{peerA1: "peerA", peerA2: "peerA", peerB1: "peerB"} {
+		_, err := m.allowedIPsRefCounter.Increment(prefix, peer)
+		require.NoError(t, err)
+	}
+	// Extra reference: reconcile must still re-apply the prefix even though its refcount never
+	// hit 0 again (the exact case the plain incremental path skips).
+	_, err := m.allowedIPsRefCounter.Increment(peerA1, "peerA")
+	require.NoError(t, err)
+
+	require.NoError(t, m.ReconcilePeerAllowedIPs("peerA"))
+
+	assert.ElementsMatch(t, []netip.Prefix{peerA1, peerA2}, wg.added("peerA"),
+		"reconcile must re-apply all routed prefixes of the peer")
+	assert.Empty(t, wg.added("peerB"), "reconcile must not touch another peer's prefixes")
+}
+
+// TestReconcilePeerAllowedIPsNoCounter verifies reconcile is a safe no-op before the refcounter is
+// set up.
+func TestReconcilePeerAllowedIPsNoCounter(t *testing.T) {
+	wg := &reconcileWGMock{}
+	m := &DefaultManager{wgInterface: wg}
+
+	require.NoError(t, m.ReconcilePeerAllowedIPs("peerA"))
+	assert.Empty(t, wg.added("peerA"))
+}

client/internal/routemanager/refcounter/refcounter.go

@@ -94,6 +94,26 @@ func (rm *Counter[Key, I, O]) Get(key Key) (Ref[O], bool) {
 	return ref, ok
 }
 
+// ReapplyMatching calls apply for every key whose stored Out satisfies pred, holding the
+// counter lock for the whole pass. Running apply under the lock keeps it atomic with respect
+// to Increment/Decrement: a prefix dropped to zero is removed from the map (and had its
+// RemoveFunc called) before this pass observes it, so a stale key can never be re-applied.
+// pred and apply are invoked under the lock, so they must not call back into the counter.
+func (rm *Counter[Key, I, O]) ReapplyMatching(pred func(out O) bool, apply func(key Key) error) error {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	var merr *multierror.Error
+	for key, ref := range rm.refCountMap {
+		if pred(ref.Out) {
+			if err := apply(key); err != nil {
+				merr = multierror.Append(merr, err)
+			}
+		}
+	}
+	return nberrors.FormatErrorOrNil(merr)
+}
+
 // Increment increments the reference count for the given key.
 // If this is the first reference to the key, the AddFunc is called.
 func (rm *Counter[Key, I, O]) Increment(key Key, in I) (Ref[O], error) {

client/internal/routemanager/refcounter/refcounter_test.go

@@ -0,0 +1,47 @@
+package refcounter
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestReapplyMatching verifies ReapplyMatching invokes apply for exactly the keys whose stored
+// Out satisfies the predicate (no duplicates for multiply-referenced keys) — the primitive
+// ReconcilePeerAllowedIPs relies on to re-apply a single peer's routed prefixes.
+func TestReapplyMatching(t *testing.T) {
+	rc := New[netip.Prefix, string, string](
+		func(_ netip.Prefix, peerKey string) (string, error) { return peerKey, nil },
+		func(netip.Prefix, string) error { return nil },
+	)
+
+	peerA1 := netip.MustParsePrefix("10.0.0.0/24")
+	peerA2 := netip.MustParsePrefix("10.1.0.0/24")
+	peerB1 := netip.MustParsePrefix("10.2.0.0/24")
+
+	for prefix, peer := range map[netip.Prefix]string{peerA1: "peerA", peerA2: "peerA", peerB1: "peerB"} {
+		_, err := rc.Increment(prefix, peer)
+		require.NoError(t, err)
+	}
+	// a second reference must not make the key applied twice
+	_, err := rc.Increment(peerA1, "peerA")
+	require.NoError(t, err)
+
+	var applied []netip.Prefix
+	err = rc.ReapplyMatching(
+		func(out string) bool { return out == "peerA" },
+		func(key netip.Prefix) error { applied = append(applied, key); return nil },
+	)
+	require.NoError(t, err)
+	assert.ElementsMatch(t, []netip.Prefix{peerA1, peerA2}, applied)
+
+	var none []netip.Prefix
+	err = rc.ReapplyMatching(
+		func(out string) bool { return out == "missing" },
+		func(key netip.Prefix) error { none = append(none, key); return nil },
+	)
+	require.NoError(t, err)
+	assert.Empty(t, none)
+}

client/internal/statemanager/manager.go

@@ -1,6 +1,7 @@
 package statemanager
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -305,6 +306,11 @@ func (m *Manager) loadStateFile(deleteCorrupt bool) (map[string]json.RawMessage,
 
 	var rawStates map[string]json.RawMessage
 	if err := json.Unmarshal(data, &rawStates); err != nil {
+		if len(bytes.TrimSpace(data)) == 0 {
+			log.Warnf("state file %s is empty (%d bytes)", m.filePath, len(data))
+		} else {
+			log.Warnf("state file %s has malformed content (%d bytes)", m.filePath, len(data))
+		}
 		m.handleCorruptedState(deleteCorrupt)
 		return nil, fmt.Errorf("unmarshal states: %w", err)
 	}

client/ios/NetBirdSDK/client.go

@@ -233,6 +233,9 @@ func (c *Client) DebugBundle(anonymize bool) (string, error) {
 		deps.SyncResponse = resp
 
 		if e := cc.Engine(); e != nil {
+			deps.RefreshStatus = func() {
+				e.RunHealthProbes(context.Background(), true)
+			}
 			if cm := e.GetClientMetrics(); cm != nil {
 				deps.ClientMetrics = cm
 			}

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/ios/NetBirdSDK/login.go

@@ -44,10 +44,25 @@ type Auth struct {
 // NewAuth instantiate Auth struct and validate the management URL
 func NewAuth(cfgPath string, mgmURL string) (*Auth, error) {
 	inputCfg := profilemanager.ConfigInput{
+		ConfigPath:    cfgPath,
 		ManagementURL: mgmURL,
 	}
 
-	cfg, err := profilemanager.CreateInMemoryConfig(inputCfg)
+	// Load the existing config when a config file is already present so an
+	// interactive re-login reuses the peer's persisted WireGuard private key
+	// (and thus its identity) instead of generating a fresh one. Generating a
+	// new key registers a brand-new peer on the management server on every
+	// re-auth (named after the fallback hostname). Only fall back to a fresh
+	// in-memory config for the first-time login when no config file exists yet.
+	// DirectUpdateOrCreateConfig uses non-atomic writes so it also works inside
+	// the tvOS App Group sandbox where atomic temp-file+rename is blocked.
+	var cfg *profilemanager.Config
+	var err error
+	if cfgPath != "" {
+		cfg, err = profilemanager.DirectUpdateOrCreateConfig(inputCfg)
+	} else {
+		cfg, err = profilemanager.CreateInMemoryConfig(inputCfg)
+	}
 	if err != nil {
 		return nil, err
 	}

client/mdm/canonical_loaders.go

@@ -22,12 +22,14 @@ var allKeys = []string{
 	KeyDisableMetricsCollection,
 	KeyAllowServerSSH,
 	KeyDisableAutoConnect,
+	KeyDisableAutostart,
 	KeyPreSharedKey,
 	KeyRosenpassEnabled,
 	KeyRosenpassPermissive,
 	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"
 )
@@ -19,10 +20,10 @@ import (
 // names (lowerCamelCase) so the daemon can map a Policy key directly to a
 // configuration field.
 const (
-	KeyManagementURL            = "managementURL"
-	KeyDisableUpdateSettings    = "disableUpdateSettings"
-	KeyDisableProfiles          = "disableProfiles"
-	KeyDisableNetworks          = "disableNetworks"
+	KeyManagementURL         = "managementURL"
+	KeyDisableUpdateSettings = "disableUpdateSettings"
+	KeyDisableProfiles       = "disableProfiles"
+	KeyDisableNetworks       = "disableNetworks"
 	// KeyDisableAdvancedView gates the advanced-view section in the
 	// upcoming UI revision. UI-only: NOT stored on Config, not
 	// applied by applyMDMPolicy, not rejectable via SetConfig. The
@@ -36,10 +37,16 @@ const (
 	KeyDisableMetricsCollection = "disableMetricsCollection"
 	KeyAllowServerSSH           = "allowServerSSH"
 	KeyDisableAutoConnect       = "disableAutoConnect"
-	KeyPreSharedKey             = "preSharedKey"
-	KeyRosenpassEnabled         = "rosenpassEnabled"
-	KeyRosenpassPermissive      = "rosenpassPermissive"
-	KeyWireguardPort            = "wireguardPort"
+	// KeyDisableAutostart suppresses the GUI's fresh-install
+	// launch-on-login default and marks the Settings toggle as
+	// MDM-managed. UI-only: NOT stored on Config and not applied by
+	// applyMDMPolicy; the GUI reads it directly and it appears in
+	// GetConfigResponse.mDMManagedFields when set.
+	KeyDisableAutostart    = "disableAutostart"
+	KeyPreSharedKey        = "preSharedKey"
+	KeyRosenpassEnabled    = "rosenpassEnabled"
+	KeyRosenpassPermissive = "rosenpassPermissive"
+	KeyWireguardPort       = "wireguardPort"
 
 	// Split tunnel is modeled as a single conceptual policy with two
 	// registry/plist values. KeySplitTunnelMode is the discriminator
@@ -48,6 +55,11 @@ 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).
@@ -69,12 +81,13 @@ 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 {
@@ -157,7 +170,8 @@ 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
@@ -170,7 +184,7 @@ 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},