Renames SetRemotePort to SetRemoteAddr

Source OnDataPathRekeyed from the WGWatcher's per-handshake callback
(onWGCheckSuccess), which fires only on a fresh handshake, and OnDataPathDown
from the handshake-timeout path. A fresh handshake clocks the next chained
KEM exchange pushed over the data-path UDP transport.

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

@@ -5,6 +5,13 @@ on:
   schedule:
     - cron: "0 3 * * *"
   workflow_dispatch:
+    inputs:
+      bedrock_model:
+        description: >-
+          Bedrock inference-profile id to drive the matrix with, exactly as
+          AWS issues it. Leave empty for the Sonnet 4.6 default.
+        required: false
+        default: ""
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
@@ -51,6 +58,9 @@ 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 }}
@@ -59,6 +69,8 @@ jobs:
           CLOUDFLARE_TOKEN: ${{ secrets.E2E_CLOUDFLARE_TOKEN }}
           AWS_BEARER_TOKEN_BEDROCK: ${{ secrets.E2E_AWS_BEARER_TOKEN_BEDROCK }}
           AWS_REGION: ${{ secrets.E2E_AWS_REGION }}
+          # Bedrock model override: dispatch input wins, then the repo variable, else the test default.
+          AWS_BEDROCK_MODEL: ${{ inputs.bedrock_model || vars.E2E_AWS_BEDROCK_MODEL }}
           # Vertex (Anthropic-on-Vertex): SA + project required; region defaults
           # to "global", model to a pinned claude snapshot.
           GOOGLE_VERTEX_SA_BASE64: ${{ secrets.E2E_GOOGLE_VERTEX_SA_BASE64 }}

.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/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

.goreleaser.yaml

@@ -273,8 +273,8 @@ dockers_v2:
        - netbirdio/netbird
        - ghcr.io/netbirdio/netbird
      tags:
-       - "v{{ .Version }}-rootless"
-       - "{{ if eq .Env.SKIP_PUBLISH \"false\" }}latest{{ end }}"
+       - "{{ .Version }}-rootless"
+       - "{{ if eq .Env.SKIP_PUBLISH \"false\" }}rootless-latest{{ end }}"
      dockerfile: client/Dockerfile-rootless
      extra_files:
        - client/netbird-entrypoint.sh

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/iface/iface_test.go

@@ -464,6 +464,8 @@ func Test_RemovePeer(t *testing.T) {
 }
 
 func Test_ConnectPeers(t *testing.T) {
+	t.Setenv("NB_DISABLE_EBPF_WG_PROXY", "true")
+
 	peer1ifaceName := fmt.Sprintf("utun%d", WgIntNumber+400)
 	peer1wgIP := netip.MustParsePrefix("10.99.99.17/30")
 	peer1Key, _ := wgtypes.GeneratePrivateKey()
@@ -505,12 +507,8 @@ func Test_ConnectPeers(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	localIP, err := getLocalIP()
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	peer1endpoint, err := net.ResolveUDPAddr("udp", fmt.Sprintf("%s:%d", localIP, peer1wgPort))
+	localIP1 := "127.0.0.1"
+	peer1endpoint, err := net.ResolveUDPAddr("udp", fmt.Sprintf("%s:%d", localIP1, peer1wgPort))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -546,7 +544,8 @@ func Test_ConnectPeers(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	peer2endpoint, err := net.ResolveUDPAddr("udp", fmt.Sprintf("%s:%d", localIP, peer2wgPort))
+	localIP2 := "127.0.0.1"
+	peer2endpoint, err := net.ResolveUDPAddr("udp", fmt.Sprintf("%s:%d", localIP2, peer2wgPort))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -569,17 +568,17 @@ 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 +587,12 @@ 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:
+		}
 	}
 
 }
@@ -615,28 +620,3 @@ func getPeer(ifaceName, peerPubKey string) (wgtypes.Peer, error) {
 	}
 	return wgtypes.Peer{}, fmt.Errorf("peer not found")
 }
-
-func getLocalIP() (string, error) {
-	// Get all interfaces
-	addrs, err := net.InterfaceAddrs()
-	if err != nil {
-		return "", err
-	}
-
-	for _, addr := range addrs {
-		ipNet, ok := addr.(*net.IPNet)
-		if !ok {
-			continue
-		}
-		if ipNet.IP.IsLoopback() {
-			continue
-		}
-
-		if ipNet.IP.To4() == nil {
-			continue
-		}
-		return ipNet.IP.String(), nil
-	}
-
-	return "", fmt.Errorf("no local IP found")
-}

client/internal/auth/auth.go

@@ -351,6 +351,7 @@ func (a *Auth) setSystemInfoFlags(info *system.Info) {
 		a.config.BlockLANAccess,
 		a.config.BlockInbound,
 		a.config.DisableIPv6,
+		a.config.SyncMessageVersion,
 		a.config.EnableSSHRoot,
 		a.config.EnableSSHSFTP,
 		a.config.EnableSSHLocalPortForwarding,

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/conn_mgr.go

@@ -34,6 +34,8 @@ const (
 // - 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
@@ -42,12 +44,26 @@ type ConnMgr struct {
 	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,
@@ -238,12 +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)
 		}
@@ -268,16 +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() {
@@ -316,7 +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)

client/internal/conn_mgr_test.go

@@ -1,10 +1,21 @@
 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) {
@@ -38,3 +49,58 @@ func TestResolveLazyForce(t *testing.T) {
 		})
 	}
 }
+
+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

@@ -257,7 +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 {
@@ -618,6 +621,7 @@ func createEngineConfig(key wgtypes.Key, config *profilemanager.Config, peerConf
 		BlockLANAccess:      config.BlockLANAccess,
 		BlockInbound:        config.BlockInbound,
 		DisableIPv6:         config.DisableIPv6,
+		SyncMessageVersion:  config.SyncMessageVersion,
 
 		LazyConnection: lazyconn.ParseState(config.LazyConnection),
 
@@ -693,6 +697,7 @@ func loginToManagement(ctx context.Context, client mgm.Client, pubSSHKey []byte,
 		config.BlockLANAccess,
 		config.BlockInbound,
 		config.DisableIPv6,
+		config.SyncMessageVersion,
 		config.EnableSSHRoot,
 		config.EnableSSHSFTP,
 		config.EnableSSHLocalPortForwarding,

client/internal/debug/debug.go

@@ -676,6 +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))

client/internal/debug/debug_test.go

@@ -887,6 +887,8 @@ func TestAddConfig_AllFieldsCovered(t *testing.T) {
 		ClientCertKeyPath:             "/tmp/key",
 		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_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

@@ -50,6 +50,7 @@ import (
 	icemaker "github.com/netbirdio/netbird/client/internal/peer/ice"
 	"github.com/netbirdio/netbird/client/internal/peerstore"
 	"github.com/netbirdio/netbird/client/internal/portforward"
+	"github.com/netbirdio/netbird/client/internal/pqkem"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/internal/relay"
 	"github.com/netbirdio/netbird/client/internal/rosenpass"
@@ -64,7 +65,10 @@ 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"
@@ -147,6 +151,7 @@ type EngineConfig struct {
 	BlockLANAccess      bool
 	BlockInbound        bool
 	DisableIPv6         bool
+	SyncMessageVersion  *int
 
 	// LazyConnection is the MDM-sourced lazy-connection override; StateUnset defers to
 	// the env var and management feature flag.
@@ -193,6 +198,10 @@ type Engine struct {
 	// rpManager is a Rosenpass manager
 	rpManager *rosenpass.Manager
 
+	// pqkemManager runs the ML-KEM post-quantum PSK exchange (gated by NB_ENABLE_PQ_MLKEM).
+	// It owns the data-path transport and peer endpoint routing.
+	pqkemManager *pqkem.Manager
+
 	// syncMsgMux is used to guarantee sequential Management Service message processing
 	syncMsgMux *sync.Mutex
 
@@ -220,6 +229,13 @@ 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
@@ -640,6 +656,19 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 		e.rpManager.SetInterface(e.wgInterface)
 	}
 
+	// Start the ML-KEM PQ manager after the interface is up so its dedicated UDP
+	// transport can bind on the WG overlay IP.
+	if pqkem.Enabled() {
+		tr, pqErr := newPQTransport(e.config.WgAddr.IP)
+		if pqErr != nil {
+			log.Errorf("pqkem: transport bind failed, exchange disabled: %v", pqErr)
+		} else {
+			e.pqkemManager = pqkem.NewManager(pqkem.LocalID(publicKey.String()), pqCallbackHandler{wg: e.wgInterface}, pqkem.NewLogger())
+			e.pqkemManager.Start(tr)
+			log.Infof("pqkem: enabled (udp port %d on overlay %s)", e.pqkemManager.LocalPort(), e.config.WgAddr.IP)
+		}
+	}
+
 	// if inbound conns are blocked there is no need to create the ACL manager
 	if e.firewall != nil && !e.config.BlockInbound {
 		e.acl = acl.NewDefaultManager(e.firewall)
@@ -652,8 +681,24 @@ 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())
 
@@ -887,6 +932,10 @@ func (e *Engine) removePeer(peerKey string) error {
 
 	e.connMgr.RemovePeerConn(peerKey)
 
+	if e.pqkemManager != nil {
+		e.pqkemManager.RemovePeer(pqkem.RemoteID(peerKey))
+	}
+
 	err := e.statusRecorder.RemovePeer(peerKey)
 	if err != nil {
 		log.Warnf("received error when removing peer %s from status recorder: %v", peerKey, err)
@@ -963,8 +1012,12 @@ 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")
@@ -974,12 +1027,47 @@ 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
 	}
@@ -992,6 +1080,14 @@ 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()
 
@@ -1005,6 +1101,19 @@ 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.
@@ -1164,6 +1273,7 @@ func (e *Engine) applyInfoFlags(info *system.Info) {
 		e.config.BlockLANAccess,
 		e.config.BlockInbound,
 		e.config.DisableIPv6,
+		e.config.SyncMessageVersion,
 		e.config.EnableSSHRoot,
 		e.config.EnableSSHSFTP,
 		e.config.EnableSSHLocalPortForwarding,
@@ -1812,6 +1922,9 @@ func (e *Engine) createPeerConn(pubKey string, allowedIPs []netip.Prefix, agentV
 		},
 		ICEConfig: e.createICEConfig(),
 	}
+	if e.pqkemManager != nil {
+		config.PQ = pqHandshaker{mgr: e.pqkemManager}
+	}
 
 	serviceDependencies := peer.ServiceDependencies{
 		StatusRecorder:     e.statusRecorder,
@@ -1995,6 +2108,10 @@ func (e *Engine) close() {
 		_ = e.rpManager.Close()
 	}
 
+	if e.pqkemManager != nil {
+		e.pqkemManager.Stop()
+	}
+
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 	if err := e.portForwardManager.GracefullyStop(ctx); err != nil {
@@ -2032,6 +2149,7 @@ func (e *Engine) readInitialSettings() ([]*route.Route, *nbdns.Config, bool, err
 		e.config.BlockLANAccess,
 		e.config.BlockInbound,
 		e.config.DisableIPv6,
+		e.config.SyncMessageVersion,
 		e.config.EnableSSHRoot,
 		e.config.EnableSSHSFTP,
 		e.config.EnableSSHLocalPortForwarding,
@@ -2806,6 +2924,8 @@ func convertToOfferAnswer(msg *sProto.Message) (*peer.OfferAnswer, error) {
 		Version:         msg.GetBody().GetNetBirdVersion(),
 		RosenpassPubKey: rosenpassPubKey,
 		RosenpassAddr:   rosenpassAddr,
+		MlkemPayload:    msg.GetBody().GetMlkemPayload(),
+		MlkemPort:       int(msg.GetBody().GetMlkemPort()),
 		RelaySrvAddress: msg.GetBody().GetRelayServerAddress(),
 		RelaySrvIP:      relayIP,
 		SessionID:       sessionID,

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/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/peer/conn.go

@@ -74,6 +74,32 @@ type RosenpassConfig struct {
 	PermissiveMode bool
 }
 
+// PQHandshaker attaches post-quantum ML-KEM material to signalling offers/answers and
+// feeds received material back. It is implemented by the engine over the pqkem
+// manager and is nil when the PQ exchange is disabled. remoteKey is the peer's
+// WireGuard public key.
+type PQHandshaker interface {
+	// OfferPayload returns the KEM offer to embed in an outgoing offer (nil if this
+	// peer is not the KEM initiator) and the local PQ data-path port to announce.
+	OfferPayload(remoteKey string) (payload []byte, port int)
+	// AnswerPayload processes a received KEM offer (nil if absent) and returns the KEM
+	// answer to embed in the outgoing answer (nil if none) and the local PQ port.
+	AnswerPayload(remoteKey string, recvOffer []byte) (payload []byte, port int)
+	// OnAnswer feeds a received KEM answer (nil if absent).
+	OnAnswer(remoteKey string, recvAnswer []byte)
+	// PSK returns the peer's latest derived post-quantum PSK to program at WG
+	// peer-config time (the pull path). ok is false until one has been derived.
+	PSK(remoteKey string) (wgtypes.Key, bool)
+	// SetRemoteAddr registers the peer's data-path endpoint learned from signalling:
+	// its WG overlay IP with the advertised pq UDP port.
+	SetRemoteAddr(remoteKey string, addr netip.AddrPort)
+	// OnDataPathRekeyed signals a fresh WireGuard handshake for the peer; it clocks the
+	// next chained PSK rotation pushed over the data path.
+	OnDataPathRekeyed(remoteKey string)
+	// OnDataPathDown signals the peer's tunnel went down.
+	OnDataPathDown(remoteKey string)
+}
+
 // ConnConfig is a peer Connection configuration
 type ConnConfig struct {
 	// Key is a public key of a remote peer
@@ -91,6 +117,9 @@ type ConnConfig struct {
 
 	RosenpassConfig RosenpassConfig
 
+	// PQ carries post-quantum ML-KEM material on offers/answers; nil when disabled.
+	PQ PQHandshaker
+
 	// ICEConfig ICE protocol configuration
 	ICEConfig icemaker.Config
 }
@@ -681,6 +710,10 @@ func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 
 	conn.Log.Warnf("WireGuard handshake timeout detected, closing current connection")
 
+	if conn.config.PQ != nil {
+		conn.config.PQ.OnDataPathDown(conn.config.Key)
+	}
+
 	// Close the active connection based on current priority
 	switch conn.currentConnPriority {
 	case conntype.Relay:
@@ -946,6 +979,11 @@ func (conn *Conn) onWGCheckSuccess() {
 	conn.mu.Lock()
 	conn.wgTimeouts = 0
 	conn.mu.Unlock()
+
+	// A fresh WireGuard handshake is the clock for the post-quantum PSK rotation.
+	if conn.config.PQ != nil {
+		conn.config.PQ.OnDataPathRekeyed(conn.config.Key)
+	}
 }
 
 // recordConnectionMetrics records connection stage timestamps as metrics
@@ -987,6 +1025,15 @@ func (conn *Conn) AgentVersionString() string {
 }
 
 func (conn *Conn) presharedKey(remoteRosenpassKey []byte) *wgtypes.Key {
+	// Post-quantum: once the ML-KEM exchange has derived a PSK for this peer, program
+	// it here so the peer's next WireGuard handshake adopts it. Applied at peer-config
+	// time (bootstrap / reconnect); steady-state rotation is pushed separately.
+	if conn.config.PQ != nil {
+		if psk, ok := conn.config.PQ.PSK(conn.config.Key); ok {
+			return &psk
+		}
+	}
+
 	if conn.config.RosenpassConfig.PubKey == nil {
 		return conn.config.WgConfig.PreSharedKey
 	}

client/internal/peer/handshaker.go

@@ -39,6 +39,16 @@ type OfferAnswer struct {
 	// This value is the local Rosenpass server address when sending the message
 	RosenpassAddr string
 
+	// MlkemPayload carries the post-quantum X25519MLKEM768 handshake message
+	// (pqkem-framed offer on an OFFER, answer on an ANSWER) that seeds the
+	// WireGuard PSK. Opaque here — the pqkem library frames and parses it. Nil
+	// when the peer does not run the ML-KEM PQ exchange.
+	MlkemPayload []byte
+
+	// MlkemPort is the peer's ML-KEM PQ service UDP port (bound on its WG overlay
+	// IP) where data-path rekey messages are sent. Zero when not running the exchange.
+	MlkemPort int
+
 	// relay server address
 	RelaySrvAddress string
 	// RelaySrvIP is the IP the remote peer is connected to on its
@@ -120,6 +130,8 @@ func (h *Handshaker) Listen(ctx context.Context) {
 
 			h.updateRemoteICEState(&remoteOfferAnswer)
 
+			h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
+
 			if h.relayListener != nil {
 				h.relayListener.Notify(&remoteOfferAnswer)
 			}
@@ -128,7 +140,7 @@ func (h *Handshaker) Listen(ctx context.Context) {
 				h.iceListener(&remoteOfferAnswer)
 			}
 
-			if err := h.sendAnswer(); err != nil {
+			if err := h.sendAnswer(&remoteOfferAnswer); err != nil {
 				h.log.Errorf("failed to send remote offer confirmation: %s", err)
 				continue
 			}
@@ -142,6 +154,8 @@ func (h *Handshaker) Listen(ctx context.Context) {
 
 			h.updateRemoteICEState(&remoteOfferAnswer)
 
+			h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
+
 			if h.relayListener != nil {
 				h.relayListener.Notify(&remoteOfferAnswer)
 			}
@@ -149,6 +163,10 @@ func (h *Handshaker) Listen(ctx context.Context) {
 			if h.iceListener != nil && h.RemoteICESupported() {
 				h.iceListener(&remoteOfferAnswer)
 			}
+
+			if h.config.PQ != nil {
+				h.config.PQ.OnAnswer(h.config.Key, remoteOfferAnswer.MlkemPayload)
+			}
 		case <-ctx.Done():
 			h.log.Infof("stop listening for remote offers and answers")
 			return
@@ -156,6 +174,16 @@ func (h *Handshaker) Listen(ctx context.Context) {
 	}
 }
 
+// pqRegisterEndpoint feeds the post-quantum handshaker the peer's data-path endpoint
+// (its WG overlay IP plus the advertised pq UDP port) learned from a remote offer/answer.
+func (h *Handshaker) pqRegisterEndpoint(remotePort int) {
+	if h.config.PQ == nil || remotePort <= 0 || remotePort > 65535 || len(h.config.WgConfig.AllowedIps) == 0 {
+		return
+	}
+	addr := netip.AddrPortFrom(h.config.WgConfig.AllowedIps[0].Addr(), uint16(remotePort))
+	h.config.PQ.SetRemoteAddr(h.config.Key, addr)
+}
+
 func (h *Handshaker) SendOffer() error {
 	h.mu.Lock()
 	defer h.mu.Unlock()
@@ -195,13 +223,23 @@ func (h *Handshaker) sendOffer() error {
 	}
 
 	offer := h.buildOfferAnswer()
+	if h.config.PQ != nil {
+		offer.MlkemPayload, offer.MlkemPort = h.config.PQ.OfferPayload(h.config.Key)
+	}
 	h.log.Debugf("sending offer with serial: %s", offer.SessionIDString())
 
 	return h.signaler.SignalOffer(offer, h.config.Key)
 }
 
-func (h *Handshaker) sendAnswer() error {
+func (h *Handshaker) sendAnswer(remoteOffer *OfferAnswer) error {
 	answer := h.buildOfferAnswer()
+	if h.config.PQ != nil {
+		var recvOffer []byte
+		if remoteOffer != nil {
+			recvOffer = remoteOffer.MlkemPayload
+		}
+		answer.MlkemPayload, answer.MlkemPort = h.config.PQ.AnswerPayload(h.config.Key, recvOffer)
+	}
 	h.log.Debugf("sending answer with serial: %s", answer.SessionIDString())
 
 	return h.signaler.SignalAnswer(answer, h.config.Key)

client/internal/peer/signaler.go

@@ -63,6 +63,8 @@ func (s *Signaler) signalOfferAnswer(offerAnswer OfferAnswer, remoteKey string,
 		},
 		RosenpassPubKey: offerAnswer.RosenpassPubKey,
 		RosenpassAddr:   offerAnswer.RosenpassAddr,
+		MlkemPayload:    offerAnswer.MlkemPayload,
+		MlkemPort:       offerAnswer.MlkemPort,
 		RelaySrvAddress: offerAnswer.RelaySrvAddress,
 		RelaySrvIP:      offerAnswer.RelaySrvIP,
 		SessionID:       sessionIDBytes,

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/pqkem/bench_test.go

@@ -0,0 +1,59 @@
+package pqkem
+
+import (
+	"crypto/ecdh"
+	"crypto/mlkem"
+	"crypto/rand"
+	"testing"
+)
+
+func BenchmarkX25519Keygen(b *testing.B) {
+	c := ecdh.X25519()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		if _, err := c.GenerateKey(rand.Reader); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkX25519ECDH(b *testing.B) {
+	c := ecdh.X25519()
+	a, _ := c.GenerateKey(rand.Reader)
+	p, _ := c.GenerateKey(rand.Reader)
+	pub := p.PublicKey()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		if _, err := a.ECDH(pub); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkMLKEMKeygen(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		if _, err := mlkem.GenerateKey768(); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkMLKEMEncaps(b *testing.B) {
+	dk, _ := mlkem.GenerateKey768()
+	ek := dk.EncapsulationKey()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		_, _ = ek.Encapsulate()
+	}
+}
+
+func BenchmarkMLKEMDecaps(b *testing.B) {
+	dk, _ := mlkem.GenerateKey768()
+	_, ct := dk.EncapsulationKey().Encapsulate()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		if _, err := dk.Decapsulate(ct); err != nil {
+			b.Fatal(err)
+		}
+	}
+}

client/internal/pqkem/callbacks.go

@@ -0,0 +1,18 @@
+package pqkem
+
+// CallbackHandler is implemented by the host and invoked by the library. The
+// library only reports events; the host owns the reaction. Keeping this an
+// interface — rather than touching the transport or keying directly — is what lets
+// the KEM code be extracted as a standalone library.
+type CallbackHandler interface {
+	// OnNewPSKReady fires when a fresh post-quantum PSK has been derived for a peer
+	// and must be programmed into the consumer's secure channel. It is invoked at
+	// the commit point of each side: the initiator on receiving the answer, the
+	// responder on receiving the confirm.
+	OnNewPSKReady(remoteID RemoteID, psk PSK) error
+
+	// OnRekeyFailed fires when an exchange fails to converge within the allotted
+	// time. The host should tear the peer connection down so it re-establishes, and
+	// log a WARN. The library reports the event; it does not dictate the reaction.
+	OnRekeyFailed(remoteID RemoteID) error
+}

client/internal/pqkem/convergence.go

@@ -0,0 +1,224 @@
+package pqkem
+
+import (
+	"context"
+	"time"
+)
+
+// startExchange creates a fresh initiator exchange (acknowledging ackID, zero for a
+// bootstrap) and returns the framed offer for the caller to send — pushed over the
+// data path for a chained rekey, or handed to the host for signalling when viaSignal
+// is set. Any previous in-flight exchange for the peer is cancelled.
+func (m *Manager) startExchange(remoteID RemoteID, viaSignal bool, ackID ExchangeID) ([]byte, error) {
+	init, err := NewInitiator()
+	if err != nil {
+		return nil, err
+	}
+	id, err := newExchangeID()
+	if err != nil {
+		return nil, err
+	}
+	raw, err := (&OfferMsg{ExchangeID: id, AckID: ackID, KEMOffer: init.Offer()}).Encode()
+	if err != nil {
+		return nil, err
+	}
+
+	ctx, cancel := context.WithCancel(m.rootCtx)
+	m.mu.Lock()
+	if old := m.exchanges[remoteID]; old != nil && old.cancel != nil {
+		old.cancel()
+	}
+	m.exchanges[remoteID] = &exchangeCtl{
+		id:        id,
+		state:     stateAwaitingAnswer,
+		startedAt: time.Now(),
+		cancel:    cancel,
+		lastSent:  raw,
+		initiator: init,
+		viaSignal: viaSignal,
+	}
+	m.mu.Unlock()
+
+	m.wait.Add(1)
+	go m.initiatorLoop(ctx, remoteID, id)
+	return raw, nil
+}
+
+// processOffer (responder) first acknowledges the previous exchange the offer names
+// (that offer riding the data path under the freshly adopted key proves it worked),
+// then derives the PSK for the new offer, commits it optimistically, and returns the
+// framed answer. A duplicate offer returns the cached answer without re-deriving.
+func (m *Manager) processOffer(remoteID RemoteID, o *OfferMsg) ([]byte, error) {
+	if o.AckID != (ExchangeID{}) {
+		m.ackConverged(remoteID, o.AckID)
+	}
+
+	m.mu.Lock()
+	if ex := m.exchanges[remoteID]; ex != nil && ex.id == o.ExchangeID {
+		state, last := ex.state, ex.lastSent
+		m.mu.Unlock()
+		if state == stateReserved {
+			return nil, nil
+		}
+		return last, nil
+	}
+	// Reserve the slot so a concurrent duplicate offer bails.
+	m.exchanges[remoteID] = &exchangeCtl{id: o.ExchangeID, state: stateReserved, startedAt: time.Now()}
+	m.mu.Unlock()
+
+	answerBytes, psk, err := Respond(o.KEMOffer, m.binding(remoteID))
+	if err != nil {
+		return nil, err
+	}
+	raw, err := (&AnswerMsg{ExchangeID: o.ExchangeID, KEMAnswer: answerBytes}).Encode()
+	if err != nil {
+		return nil, err
+	}
+
+	m.mu.Lock()
+	ex := m.exchanges[remoteID]
+	if ex == nil || ex.id != o.ExchangeID {
+		m.mu.Unlock()
+		return nil, nil
+	}
+	ex.state = stateAwaitingAck
+	ex.lastSent = raw
+	ex.pendingPSK = psk
+	m.psks[remoteID] = psk
+	m.mu.Unlock()
+
+	// Commit optimistically so our data path can rekey to the new PSK.
+	if err := m.cbHandler.OnNewPSKReady(remoteID, psk); err != nil {
+		return nil, err
+	}
+	return raw, nil
+}
+
+// processAnswer (initiator) derives and commits the PSK and parks in
+// stateAwaitingRekey; the next offer (chained from OnDataPathRekeyed) will acknowledge
+// this exchange. Only valid in stateAwaitingAnswer; advancing the state under the
+// lock makes a concurrent/duplicate answer bail.
+func (m *Manager) processAnswer(remoteID RemoteID, a *AnswerMsg) error {
+	m.mu.Lock()
+	ex := m.exchanges[remoteID]
+	if ex == nil || ex.id != a.ExchangeID || ex.state != stateAwaitingAnswer {
+		m.mu.Unlock()
+		return nil
+	}
+	ex.state = stateAwaitingRekey
+	init := ex.initiator
+	ex.initiator = nil
+	m.mu.Unlock()
+
+	psk, err := init.Finish(a.KEMAnswer, m.binding(remoteID))
+	if err != nil {
+		return err
+	}
+
+	// The initiator has converged: the responder must have derived the key to answer.
+	m.mu.Lock()
+	m.established[remoteID] = true
+	m.failures[remoteID] = 0
+	m.psks[remoteID] = psk
+	m.mu.Unlock()
+
+	return m.cbHandler.OnNewPSKReady(remoteID, psk)
+}
+
+// ackConverged (responder) records convergence of the exchange named by ackID: a
+// later offer acknowledging it proves both sides operate on that exchange's key. Only
+// acts on a matching stateAwaitingAck exchange; anything else is ignored.
+func (m *Manager) ackConverged(remoteID RemoteID, ackID ExchangeID) {
+	m.mu.Lock()
+	ex := m.exchanges[remoteID]
+	if ex == nil || ex.id != ackID || ex.state != stateAwaitingAck {
+		m.mu.Unlock()
+		return
+	}
+	delete(m.exchanges, remoteID)
+	m.established[remoteID] = true
+	m.failures[remoteID] = 0
+	_ = time.Since(ex.startedAt) // convergence latency (metrics hook, later step)
+	m.mu.Unlock()
+}
+
+// initiatorLoop enforces the offer->answer convergence deadline and retransmits the
+// initiator's outstanding data-path offer while awaiting the answer (a
+// signalling-bootstrapped offer is retransmitted by the host, so it is not resent
+// here). Exhausting the deadline before the answer arrives is a failure. Once the
+// answer is in (state past awaitingAnswer) the loop exits: the next rotation is driven
+// by OnDataPathRekeyed, and the idle wait for it has no deadline.
+func (m *Manager) initiatorLoop(ctx context.Context, remoteID RemoteID, id ExchangeID) {
+	defer m.wait.Done()
+	t := time.NewTicker(m.retryInterval)
+	defer t.Stop()
+
+	attempts := 0
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+			m.mu.Lock()
+			ex := m.exchanges[remoteID]
+			if ex == nil || ex.id != id {
+				m.mu.Unlock()
+				return
+			}
+
+			switch ex.state {
+			case stateAwaitingAnswer:
+				if attempts >= m.maxRetries {
+					delete(m.exchanges, remoteID)
+					fail := m.registerFailureLocked(remoteID)
+					m.mu.Unlock()
+					m.raiseFailure(remoteID, fail)
+					return
+				}
+				viaSignal := ex.viaSignal
+				msg := ex.lastSent
+				attempts++
+				m.mu.Unlock()
+				if !viaSignal {
+					if err := m.pushDataPath(remoteID, msg); err != nil {
+						m.logger.Warn("pqkem: offer retransmit failed", "peer", remoteID, "err", err)
+					}
+				}
+
+			default:
+				// Past awaiting the answer (converged) or superseded: the loop's job
+				// is done. The next rotation is driven externally by OnDataPathRekeyed,
+				// so there is no deadline while idle-waiting for it (that wait can be
+				// as long as the transport's natural rekey interval).
+				m.mu.Unlock()
+				return
+			}
+		}
+	}
+}
+
+// registerFailureLocked applies policy B and reports whether OnRekeyFailed is due:
+// an initial exchange (peer never established) fails immediately; a rekey tolerates
+// up to maxRekeyFailures consecutive misses (we stay on the still-valid previous
+// PSK) before failing. Assumes m.mu is held.
+func (m *Manager) registerFailureLocked(remoteID RemoteID) bool {
+	if !m.established[remoteID] {
+		return true
+	}
+	m.failures[remoteID]++
+	if m.failures[remoteID] >= m.maxRekeyFailures {
+		m.failures[remoteID] = 0
+		return true
+	}
+	return false
+}
+
+func (m *Manager) raiseFailure(remoteID RemoteID, fail bool) {
+	if !fail {
+		m.logger.Warn("pqkem: rekey attempt timed out, will retry next cycle", "peer", remoteID)
+		return
+	}
+	if err := m.cbHandler.OnRekeyFailed(remoteID); err != nil {
+		m.logger.Error("pqkem: OnRekeyFailed handler error", "peer", remoteID, "err", err)
+	}
+}

client/internal/pqkem/convergence_test.go

@@ -0,0 +1,74 @@
+package pqkem
+
+import (
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+// dropTransport is a pqkem.Transport that silently discards everything.
+type dropTransport struct{}
+
+func (dropTransport) Send(netip.AddrPort, []byte) error { return nil }
+func (dropTransport) LocalPort() int                    { return 0 }
+func (dropTransport) Run(func(netip.AddrPort, []byte))  {}
+func (dropTransport) Close() error                      { return nil }
+
+func failedCount(f *fakeWG) int {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return len(f.failed)
+}
+
+func TestManager_InitialTimeoutFailsImmediately(t *testing.T) {
+	wg := newFakeWG()
+	d := NewManager("bbbb", wg, nil) // bbbb > aaaa -> initiator
+	d.Start(dropTransport{})
+	d.retryInterval = 5 * time.Millisecond
+	d.maxRetries = 3
+	defer d.Stop()
+
+	// Bootstrap offer is produced for signalling; no answer ever comes back -> the
+	// initial exchange fails fast.
+	offer, err := d.SignalOffer("aaaa")
+	require.NoError(t, err)
+	require.NotNil(t, offer)
+
+	require.Eventually(t, func() bool { return failedCount(wg) == 1 }, time.Second, 5*time.Millisecond)
+}
+
+func TestManager_RekeyToleratesKFailures(t *testing.T) {
+	dA, dB, _, wgB, lbB := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	// Tighten B's timings before any exchange loop spawns (the loop reads these
+	// fields, so writing them after a loop is running would race).
+	dB.retryInterval = 5 * time.Millisecond
+	dB.maxRetries = 2
+
+	// Establish: bootstrap + data-path-rekeyed so B becomes established and its data
+	// path is usable.
+	bootstrap(t, dA, dB)
+	dA.OnDataPathRekeyed("bbbb")
+	dB.OnDataPathRekeyed("aaaa")
+	require.NotEqual(t, PSK{}, wgB.psk("aaaa"))
+
+	// Drop B's outbound so rekeys can no longer converge.
+	lbB.drop.Store(true)
+
+	// K-1 data-path rekeys must NOT raise OnRekeyFailed.
+	for i := 0; i < DefaultMaxRekeyFailures-1; i++ {
+		_, err := dB.startExchange("aaaa", false, ExchangeID{})
+		require.NoError(t, err)
+		time.Sleep(50 * time.Millisecond)
+	}
+	require.Equal(t, 0, failedCount(wgB), "no failure before K attempts")
+
+	// The K-th failure raises it once.
+	_, err := dB.startExchange("aaaa", false, ExchangeID{})
+	require.NoError(t, err)
+	require.Eventually(t, func() bool { return failedCount(wgB) == 1 }, time.Second, 5*time.Millisecond)
+}

client/internal/pqkem/env.go

@@ -0,0 +1,59 @@
+package pqkem
+
+import (
+	"log/slog"
+	"os"
+	"strconv"
+	"strings"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// EnvEnabled is the environment variable that turns the ML-KEM post-quantum
+// exchange on for this client. Accepts on/off aliases plus anything
+// strconv.ParseBool understands (true/false/1/0).
+const EnvEnabled = "NB_ENABLE_PQ_MLKEM"
+
+// Enabled reports whether the ML-KEM PQ exchange is enabled via the environment.
+// An empty or unrecognized value is treated as disabled.
+func Enabled() bool {
+	raw := strings.ToLower(strings.TrimSpace(os.Getenv(EnvEnabled)))
+	switch raw {
+	case "":
+		return false
+	case "on":
+		return true
+	case "off":
+		return false
+	}
+	enabled, err := strconv.ParseBool(raw)
+	if err != nil {
+		log.Warnf("failed to parse %s value %q: %v", EnvEnabled, raw, err)
+		return false
+	}
+	return enabled
+}
+
+// EnvLogLevel overrides the ML-KEM manager's slog level (debug/info/warn/error).
+// Defaults to info.
+const EnvLogLevel = "NB_PQ_MLKEM_LOG_LEVEL"
+
+// NewLogger builds the slog logger for the ML-KEM manager: a text handler to stdout
+// at the level from EnvLogLevel. Mirrors the Rosenpass manager's logger setup so PQ
+// components log consistently.
+func NewLogger() *slog.Logger {
+	return slog.New(slog.NewTextHandler(os.Stdout, &slog.HandlerOptions{Level: logLevel()}))
+}
+
+func logLevel() slog.Level {
+	switch strings.ToLower(strings.TrimSpace(os.Getenv(EnvLogLevel))) {
+	case "debug":
+		return slog.LevelDebug
+	case "warn":
+		return slog.LevelWarn
+	case "error":
+		return slog.LevelError
+	default:
+		return slog.LevelInfo
+	}
+}

client/internal/pqkem/kem.go

@@ -0,0 +1,167 @@
+// Package pqkem is a spike (NET-1406) for a post-quantum pre-shared-key exchange
+// that could replace Rosenpass. It performs an X25519MLKEM768 hybrid key
+// encapsulation and derives a 32-byte pre-shared key (PSK).
+//
+// The exchange is a single round trip designed to ride the (already
+// authenticated) Signal offer/answer channel:
+//
+//	initiator --Offer(1216B)--> responder
+//	initiator <--Answer(1120B)-- responder
+//
+// Both sides then hold the same PSK, which is bound to the two peers' identities
+// (their peer identity keys) so the derived key cannot be transplanted
+// to a different peer pair even if the transport authentication were bypassed.
+//
+// Combiner note: this follows draft-ietf-tls-ecdhe-mlkem for X25519MLKEM768 — on
+// the wire ML-KEM ‖ X25519 (the draft deliberately reversed the share order for
+// this group), and ML-KEM_ss ‖ X25519_ss fed into the KDF. The spike uses SHA-256
+// (also binding the transcript and peer identities); a production version should
+// use HKDF — see TODO below.
+package pqkem
+
+import (
+	"crypto/ecdh"
+	"crypto/mlkem"
+	"crypto/rand"
+	"crypto/sha256"
+	"fmt"
+)
+
+const (
+	// OfferSize is the initiator message: ML-KEM-768 encapsulation key ‖ X25519 public key
+	// (share order per draft-ietf-tls-ecdhe-mlkem for X25519MLKEM768).
+	OfferSize = mlkem.EncapsulationKeySize768 + 32 // 1216
+	// AnswerSize is the responder message: ML-KEM-768 ciphertext ‖ X25519 public key.
+	AnswerSize = mlkem.CiphertextSize768 + 32 // 1120
+
+	pskLabel = "netbird-pq-psk-v1"
+)
+
+// PSK is the 32-byte derived pre-shared key handed to the consumer to key its channel.
+type PSK [32]byte
+
+// Binding identifies the peer pair the PSK is derived for. Callers set both
+// peer identity keys; the order does not matter (it is canonicalised).
+type Binding struct {
+	LocalID  []byte
+	RemoteID []byte
+}
+
+// Initiator holds the ephemeral secrets between Offer and Finish.
+type Initiator struct {
+	x25519  *ecdh.PrivateKey
+	mlkemDK *mlkem.DecapsulationKey768
+	offer   []byte
+}
+
+// NewInitiator generates the ephemeral X25519 + ML-KEM-768 keypairs.
+func NewInitiator() (*Initiator, error) {
+	x, err := ecdh.X25519().GenerateKey(rand.Reader)
+	if err != nil {
+		return nil, fmt.Errorf("x25519 keygen: %w", err)
+	}
+	dk, err := mlkem.GenerateKey768()
+	if err != nil {
+		return nil, fmt.Errorf("ml-kem keygen: %w", err)
+	}
+
+	offer := make([]byte, 0, OfferSize)
+	offer = append(offer, dk.EncapsulationKey().Bytes()...)
+	offer = append(offer, x.PublicKey().Bytes()...)
+
+	return &Initiator{x25519: x, mlkemDK: dk, offer: offer}, nil
+}
+
+// Offer returns the initiator message to send over Signal.
+func (i *Initiator) Offer() []byte {
+	return i.offer
+}
+
+// Finish consumes the responder's answer and derives the PSK.
+func (i *Initiator) Finish(answer []byte, b Binding) (PSK, error) {
+	if len(answer) != AnswerSize {
+		return PSK{}, fmt.Errorf("answer: got %d bytes, want %d", len(answer), AnswerSize)
+	}
+	ct := answer[:mlkem.CiphertextSize768]
+	peerX := answer[mlkem.CiphertextSize768:]
+
+	ssMLKEM, err := i.mlkemDK.Decapsulate(ct)
+	if err != nil {
+		return PSK{}, fmt.Errorf("ml-kem decapsulate: %w", err)
+	}
+	pub, err := ecdh.X25519().NewPublicKey(peerX)
+	if err != nil {
+		return PSK{}, fmt.Errorf("parse peer x25519: %w", err)
+	}
+	ssX, err := i.x25519.ECDH(pub)
+	if err != nil {
+		return PSK{}, fmt.Errorf("x25519 ecdh: %w", err)
+	}
+
+	return derivePSK(ssMLKEM, ssX, i.offer, answer, b), nil
+}
+
+// Respond consumes an initiator offer, produces the answer, and derives the PSK.
+func Respond(offer []byte, b Binding) (answer []byte, psk PSK, err error) {
+	if len(offer) != OfferSize {
+		return nil, PSK{}, fmt.Errorf("offer: got %d bytes, want %d", len(offer), OfferSize)
+	}
+	peerEK := offer[:mlkem.EncapsulationKeySize768]
+	peerX := offer[mlkem.EncapsulationKeySize768:]
+
+	ek, err := mlkem.NewEncapsulationKey768(peerEK)
+	if err != nil {
+		return nil, PSK{}, fmt.Errorf("parse peer ml-kem key: %w", err)
+	}
+	ssMLKEM, ct := ek.Encapsulate()
+
+	x, err := ecdh.X25519().GenerateKey(rand.Reader)
+	if err != nil {
+		return nil, PSK{}, fmt.Errorf("x25519 keygen: %w", err)
+	}
+	pub, err := ecdh.X25519().NewPublicKey(peerX)
+	if err != nil {
+		return nil, PSK{}, fmt.Errorf("parse peer x25519: %w", err)
+	}
+	ssX, err := x.ECDH(pub)
+	if err != nil {
+		return nil, PSK{}, fmt.Errorf("x25519 ecdh: %w", err)
+	}
+
+	answer = make([]byte, 0, AnswerSize)
+	answer = append(answer, ct...)
+	answer = append(answer, x.PublicKey().Bytes()...)
+
+	// derivePSK uses the same argument order on both sides; the responder's local
+	// binding is the mirror of the initiator's, canonicalised inside derivePSK.
+	return answer, derivePSK(ssMLKEM, ssX, offer, answer, b), nil
+}
+
+// derivePSK combines the two shared secrets and binds the result to the full
+// transcript (offer ‖ answer) and the canonicalised peer identities.
+//
+// TODO(NET-1406): replace the SHA-256 concat with the RFC HKDF combiner
+// (crypto/hkdf, Go 1.24+) and proper labels before this leaves spike status.
+func derivePSK(ssMLKEM, ssX, offer, answer []byte, b Binding) PSK {
+	lo, hi := canonicalPair(b.LocalID, b.RemoteID)
+
+	h := sha256.New()
+	h.Write([]byte(pskLabel))
+	h.Write(ssMLKEM)
+	h.Write(ssX)
+	h.Write(offer)
+	h.Write(answer)
+	h.Write(lo)
+	h.Write(hi)
+
+	var psk PSK
+	copy(psk[:], h.Sum(nil))
+	return psk
+}
+
+func canonicalPair(a, b []byte) (lo, hi []byte) {
+	if string(a) <= string(b) {
+		return a, b
+	}
+	return b, a
+}

client/internal/pqkem/kem_test.go

@@ -0,0 +1,89 @@
+package pqkem
+
+import (
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+var (
+	wgA = []byte("peer-A-wireguard-pubkey-32bytes!")
+	wgB = []byte("peer-B-wireguard-pubkey-32bytes!")
+)
+
+func TestExchange_DerivesMatchingPSK(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	require.Len(t, init.Offer(), OfferSize)
+
+	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+	require.Len(t, answer, AnswerSize)
+
+	pskA, err := init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
+	require.NoError(t, err)
+
+	require.Equal(t, pskB, pskA, "both sides must derive the same PSK")
+	require.NotEqual(t, PSK{}, pskA, "PSK must not be zero")
+}
+
+func TestExchange_PSKBoundToPeerIdentities(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	// responder computes with the honest pair...
+	_, pskHonest, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	// ...a second responder run with a different peer identity yields a different PSK,
+	// even though the KEM material would otherwise combine identically.
+	wgC := []byte("peer-C-wireguard-pubkey-32bytes!")
+	_, pskWrong, err := Respond(init.Offer(), Binding{LocalID: wgC, RemoteID: wgA})
+	require.NoError(t, err)
+
+	require.NotEqual(t, pskHonest, pskWrong, "PSK must be bound to the peer pair")
+}
+
+func TestExchange_RejectsMalformedMessages(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	_, _, err = Respond(init.Offer()[:10], Binding{})
+	require.Error(t, err)
+
+	_, err = init.Finish([]byte("too short"), Binding{})
+	require.Error(t, err)
+}
+
+// TestExchange_ReportSizesAndTiming is a spike measurement, not a pass/fail gate.
+// Run with: go test -run TestExchange_ReportSizesAndTiming -v ./client/internal/pqkem/
+func TestExchange_ReportSizesAndTiming(t *testing.T) {
+	const iters = 200
+
+	var tInit, tResp, tFinish time.Duration
+	for i := 0; i < iters; i++ {
+		s0 := time.Now()
+		init, err := NewInitiator()
+		require.NoError(t, err)
+		tInit += time.Since(s0)
+
+		s1 := time.Now()
+		answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+		require.NoError(t, err)
+		tResp += time.Since(s1)
+
+		s2 := time.Now()
+		_, err = init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
+		require.NoError(t, err)
+		tFinish += time.Since(s2)
+	}
+
+	t.Logf("wire sizes:   offer=%d B   answer=%d B   (Rosenpass static pubkey ~524160 B)", OfferSize, AnswerSize)
+	t.Logf("total on-wire per handshake: %d B  (~%.0fx smaller than RP static key)", OfferSize+AnswerSize, 524160.0/float64(OfferSize+AnswerSize))
+	t.Logf("avg NewInitiator (keygen): %s", tInit/iters)
+	t.Logf("avg Respond (encaps+dh):   %s", tResp/iters)
+	t.Logf("avg Finish (decaps+dh):    %s", tFinish/iters)
+	t.Logf("avg full handshake CPU:    %s", (tInit+tResp+tFinish)/iters)
+}

client/internal/pqkem/manager.go

@@ -0,0 +1,372 @@
+package pqkem
+
+import (
+	"context"
+	"crypto/rand"
+	"fmt"
+	"log/slog"
+	"net/netip"
+	"sync"
+	"time"
+)
+
+const (
+	// DefaultRetryInterval is how often the initiator retransmits its outstanding
+	// data-path offer while awaiting the answer.
+	DefaultRetryInterval = 2 * time.Second
+	// DefaultMaxRetries bounds how many ticks an exchange may run before it is
+	// declared failed. The convergence deadline is thus MaxRetries * RetryInterval.
+	DefaultMaxRetries = 10
+	// DefaultMaxRekeyFailures is how many consecutive rekey (non-initial) failures
+	// are tolerated before OnRekeyFailed. The initial exchange fails immediately.
+	DefaultMaxRekeyFailures = 3
+)
+
+// LocalID and RemoteID are peer identity keys (e.g. WireGuard public keys). They are
+// distinct types so the local and a remote identity cannot be mixed up.
+type (
+	LocalID  string
+	RemoteID string
+)
+
+// Transport is the data-path socket the Manager drives (the analogue of
+// go-rosenpass's Conn). It is a dumb mover of bytes to/from endpoints: the Manager
+// owns the remoteID<->endpoint routing and hands the transport a resolved endpoint
+// to Send, and reverse-resolves the source of each inbound datagram. Its lifecycle
+// belongs to the Manager (Run at Start, Close at Stop).
+type Transport interface {
+	// Send delivers msg to the given data-path endpoint.
+	Send(endpoint netip.AddrPort, msg []byte) error
+	// LocalPort is the bound local UDP port, announced to peers so they know where
+	// to send data-path messages.
+	LocalPort() int
+	// Run starts delivering inbound datagrams as (source endpoint, msg) to onInbound
+	// and returns immediately; it runs until Close.
+	Run(onInbound func(src netip.AddrPort, msg []byte))
+	// Close stops delivery and releases the socket.
+	Close() error
+}
+
+// exchangeState is the single source of truth for an exchange's role and phase.
+type exchangeState uint8
+
+const (
+	stateReserved       exchangeState = iota // responder: deriving the answer
+	stateAwaitingAnswer                      // initiator: offer sent, awaiting the answer
+	stateAwaitingRekey                       // initiator: PSK derived+set, awaiting OnDataPathRekeyed to chain the next offer
+	stateAwaitingAck                         // responder: answer sent, awaiting the next offer that acks this exchange
+)
+
+// exchangeCtl holds all state for one in-flight exchange with a peer, under the
+// Manager's single lock. state drives every decision. lastSent is the current
+// data-path retransmit payload (the offer, for the initiator). initiator is the
+// ephemeral handle used at Finish; pendingPSK is the responder's derived key.
+// viaSignal records that the offer went to the host for the signalling channel, so
+// the loop does not retransmit it on the data path. Only the initiator runs a
+// retransmit loop, so only it sets cancel.
+type exchangeCtl struct {
+	id         ExchangeID
+	state      exchangeState
+	startedAt  time.Time
+	cancel     context.CancelFunc
+	lastSent   []byte
+	initiator  *Initiator
+	pendingPSK PSK
+	viaSignal  bool
+}
+
+// Manager is the stateful orchestrator — the analogue of go-rosenpass's Server. It
+// drives the X25519MLKEM768 exchange, owns the peer endpoint routing and the data-path
+// transport, and surfaces the derived PSK and convergence to the host via
+// CallbackHandler. It is event-driven: the bootstrap is triggered by the host
+// (SignalOffer) and each rotation is clocked by OnDataPathRekeyed. The cryptography is
+// the pure kem.go primitives; all state lives here under one lock.
+type Manager struct {
+	localID   LocalID
+	cbHandler CallbackHandler
+	logger    *slog.Logger
+
+	retryInterval    time.Duration
+	maxRetries       int
+	maxRekeyFailures int
+
+	rootCtx    context.Context
+	rootCancel context.CancelFunc
+
+	mu          sync.Mutex
+	transport   Transport
+	exchanges   map[RemoteID]*exchangeCtl   // in-flight exchange per peer
+	established map[RemoteID]bool           // peer has completed at least one exchange
+	failures    map[RemoteID]int            // consecutive rekey failures per peer
+	psks        map[RemoteID]PSK            // latest derived PSK per peer (pulled at WG peer-config time)
+	peerAddrs   map[RemoteID]netip.AddrPort // remoteID -> data-path endpoint (send routing)
+	peersByAddr map[netip.AddrPort]RemoteID // reverse: source endpoint -> remoteID (inbound)
+	wait        sync.WaitGroup
+}
+
+// NewManager builds a manager for the local peer identified by its peer identity key
+// (used for the deterministic initiator role and the identity binding). A nil logger
+// falls back to slog.Default(). Install the data-path transport with Start.
+func NewManager(localID LocalID, h CallbackHandler, logger *slog.Logger) *Manager {
+	if logger == nil {
+		logger = slog.Default()
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	return &Manager{
+		localID:          localID,
+		cbHandler:        h,
+		logger:           logger,
+		retryInterval:    DefaultRetryInterval,
+		maxRetries:       DefaultMaxRetries,
+		maxRekeyFailures: DefaultMaxRekeyFailures,
+		rootCtx:          ctx,
+		rootCancel:       cancel,
+		exchanges:        make(map[RemoteID]*exchangeCtl),
+		established:      make(map[RemoteID]bool),
+		failures:         make(map[RemoteID]int),
+		psks:             make(map[RemoteID]PSK),
+		peerAddrs:        make(map[RemoteID]netip.AddrPort),
+		peersByAddr:      make(map[netip.AddrPort]RemoteID),
+	}
+}
+
+// Start installs the data-path transport and begins its inbound delivery. The Manager
+// owns it from here; Stop closes it. Start/Stop are the transport lifecycle pair.
+func (m *Manager) Start(t Transport) {
+	m.mu.Lock()
+	m.transport = t
+	m.mu.Unlock()
+	if t != nil {
+		t.Run(m.onDataPathInbound)
+	}
+}
+
+// LocalPort is the data-path transport's bound UDP port (0 if no transport), to be
+// announced to peers.
+func (m *Manager) LocalPort() int {
+	m.mu.Lock()
+	t := m.transport
+	m.mu.Unlock()
+	if t == nil {
+		return 0
+	}
+	return t.LocalPort()
+}
+
+// IsInitiator reports whether the local peer drives the exchange for this remote
+// peer. Roles are deterministic (lexicographic identity-key compare) so exactly one
+// side initiates, mirroring how Rosenpass picks its handshake initiator.
+func (m *Manager) IsInitiator(remoteID RemoteID) bool {
+	return string(m.localID) > string(remoteID)
+}
+
+// PSK returns the latest PSK derived for the peer, for the host to program at WG
+// peer-config time (the pull path). ok is false until an exchange has derived one.
+func (m *Manager) PSK(remoteID RemoteID) (PSK, bool) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	psk, ok := m.psks[remoteID]
+	return psk, ok
+}
+
+// AddPeer registers where a peer's data-path messages are sent and received: its
+// overlay endpoint (IP:port). Re-adding updates the endpoint.
+func (m *Manager) AddPeer(remoteID RemoteID, endpoint netip.AddrPort) {
+	if !endpoint.IsValid() {
+		return
+	}
+	m.mu.Lock()
+	if old, ok := m.peerAddrs[remoteID]; ok {
+		delete(m.peersByAddr, old)
+	}
+	m.peerAddrs[remoteID] = endpoint
+	m.peersByAddr[endpoint] = remoteID
+	m.mu.Unlock()
+}
+
+// RemovePeer stops any in-flight exchange for a peer and drops its state and routing.
+func (m *Manager) RemovePeer(remoteID RemoteID) {
+	m.mu.Lock()
+	if ex, ok := m.exchanges[remoteID]; ok {
+		if ex.cancel != nil {
+			ex.cancel()
+		}
+		delete(m.exchanges, remoteID)
+	}
+	delete(m.established, remoteID)
+	delete(m.failures, remoteID)
+	delete(m.psks, remoteID)
+	if ep, ok := m.peerAddrs[remoteID]; ok {
+		delete(m.peersByAddr, ep)
+		delete(m.peerAddrs, remoteID)
+	}
+	m.mu.Unlock()
+}
+
+// Stop cancels all in-flight exchanges, closes the transport, and waits for the
+// exchange goroutines to exit.
+func (m *Manager) Stop() {
+	m.rootCancel()
+	m.wait.Wait()
+	m.mu.Lock()
+	t := m.transport
+	m.transport = nil
+	m.exchanges = make(map[RemoteID]*exchangeCtl)
+	m.psks = make(map[RemoteID]PSK)
+	m.mu.Unlock()
+	if t != nil {
+		if err := t.Close(); err != nil {
+			m.logger.Warn("pqkem: closing data-path transport", "err", err)
+		}
+	}
+}
+
+// ---- Signalling channel (host-driven; rides the host's negotiation) ----
+
+// SignalOffer returns the KEM offer for the host to embed in its outgoing offer to
+// remoteID (bootstrap). It returns (nil, nil) when the local peer is not the
+// initiator. It is idempotent for an in-flight bootstrap: a repeat call returns the
+// same offer rather than starting a new exchange.
+func (m *Manager) SignalOffer(remoteID RemoteID) ([]byte, error) {
+	if !m.IsInitiator(remoteID) {
+		return nil, nil
+	}
+	m.mu.Lock()
+	if ex := m.exchanges[remoteID]; ex != nil && ex.viaSignal && ex.state == stateAwaitingAnswer {
+		last := ex.lastSent
+		m.mu.Unlock()
+		return last, nil
+	}
+	m.mu.Unlock()
+	// bootstrap offer acknowledges nothing (zero AckID).
+	return m.startExchange(remoteID, true, ExchangeID{})
+}
+
+// SignalOnOffer processes a KEM offer the host extracted from an incoming offer and
+// returns the KEM answer for the host to embed in its outgoing answer.
+func (m *Manager) SignalOnOffer(remoteID RemoteID, offer []byte) ([]byte, error) {
+	typ, msg, err := Decode(offer)
+	if err != nil {
+		return nil, fmt.Errorf("decode signal offer from %s: %w", remoteID, err)
+	}
+	if typ != MsgOffer {
+		return nil, fmt.Errorf("expected offer from %s, got type %d", remoteID, typ)
+	}
+	return m.processOffer(remoteID, msg.(*OfferMsg))
+}
+
+// SignalOnAnswer processes a KEM answer the host extracted from an incoming answer.
+// There is no reply: the next offer (over the data path) acknowledges this exchange.
+func (m *Manager) SignalOnAnswer(remoteID RemoteID, answer []byte) error {
+	typ, msg, err := Decode(answer)
+	if err != nil {
+		return fmt.Errorf("decode signal answer from %s: %w", remoteID, err)
+	}
+	if typ != MsgAnswer {
+		return fmt.Errorf("expected answer from %s, got type %d", remoteID, typ)
+	}
+	return m.processAnswer(remoteID, msg.(*AnswerMsg))
+}
+
+// ---- Data path ----
+
+// onDataPathInbound is the transport's inbound handler: it reverse-resolves the
+// source endpoint to a peer and dispatches. Unknown sources are dropped.
+func (m *Manager) onDataPathInbound(src netip.AddrPort, msg []byte) {
+	m.mu.Lock()
+	remoteID, ok := m.peersByAddr[src]
+	m.mu.Unlock()
+	if !ok {
+		return
+	}
+	if err := m.OnDataPathMessage(remoteID, msg); err != nil {
+		m.logger.Debug("pqkem: inbound", "peer", remoteID, "err", err)
+	}
+}
+
+// OnDataPathMessage handles a KEM message received over the data path from remoteID
+// and pushes any reply back over the data path.
+func (m *Manager) OnDataPathMessage(remoteID RemoteID, raw []byte) error {
+	typ, msg, err := Decode(raw)
+	if err != nil {
+		return fmt.Errorf("decode data-path msg from %s: %w", remoteID, err)
+	}
+	switch typ {
+	case MsgOffer:
+		answer, err := m.processOffer(remoteID, msg.(*OfferMsg))
+		if err != nil {
+			return err
+		}
+		if answer == nil {
+			return nil
+		}
+		return m.pushDataPath(remoteID, answer)
+	case MsgAnswer:
+		return m.processAnswer(remoteID, msg.(*AnswerMsg))
+	default:
+		return fmt.Errorf("unhandled data-path message type %d from %s", typ, remoteID)
+	}
+}
+
+// OnDataPathRekeyed notifies that the peer's data path is up and freshly keyed with
+// the latest PSK (fired on first establishment AND every rekey). If we are the
+// initiator that just derived a PSK, it chains the next exchange: a fresh offer over
+// the data path that acknowledges the just-completed one (its arrival under the new
+// key proves to the responder that the key works).
+func (m *Manager) OnDataPathRekeyed(remoteID RemoteID) {
+	m.mu.Lock()
+	ex := m.exchanges[remoteID]
+	chain := ex != nil && ex.state == stateAwaitingRekey
+	var ackID ExchangeID
+	if chain {
+		ackID = ex.id
+	}
+	m.mu.Unlock()
+
+	if !chain {
+		return
+	}
+	offer, err := m.startExchange(remoteID, false, ackID)
+	if err != nil {
+		m.logger.Error("pqkem: chain offer failed to start", "peer", remoteID, "err", err)
+		return
+	}
+	if err := m.pushDataPath(remoteID, offer); err != nil {
+		m.logger.Warn("pqkem: send chain offer failed", "peer", remoteID, "err", err)
+	}
+}
+
+// OnDataPathDown notifies that the peer's data path went down. Rotations resume once
+// the host re-bootstraps over signalling on reconnect; in-flight data-path sends will
+// simply fail until then. Reserved as an explicit hook.
+func (m *Manager) OnDataPathDown(remoteID RemoteID) {}
+
+// ---- internals ----
+
+// pushDataPath resolves the peer's endpoint and sends over the data-path transport,
+// erroring if the peer is unknown or no transport is set.
+func (m *Manager) pushDataPath(remoteID RemoteID, msg []byte) error {
+	m.mu.Lock()
+	ep, ok := m.peerAddrs[remoteID]
+	t := m.transport
+	m.mu.Unlock()
+	if !ok {
+		return fmt.Errorf("no data-path endpoint for peer %s", remoteID)
+	}
+	if t == nil {
+		return fmt.Errorf("no data-path transport")
+	}
+	return t.Send(ep, msg)
+}
+
+func (m *Manager) binding(remoteID RemoteID) Binding {
+	return Binding{LocalID: []byte(m.localID), RemoteID: []byte(remoteID)}
+}
+
+func newExchangeID() (ExchangeID, error) {
+	var id ExchangeID
+	if _, err := rand.Read(id[:]); err != nil {
+		return ExchangeID{}, fmt.Errorf("generate exchange id: %w", err)
+	}
+	return id, nil
+}

client/internal/pqkem/manager_test.go

@@ -0,0 +1,169 @@
+package pqkem
+
+import (
+	"fmt"
+	"net/netip"
+	"sync"
+	"sync/atomic"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+// netSwitch is an in-memory UDP fabric: transports register their endpoint and get
+// datagrams delivered to their inbound handler.
+type netSwitch struct {
+	mu sync.Mutex
+	h  map[netip.AddrPort]func(netip.AddrPort, []byte)
+}
+
+func newSwitch() *netSwitch {
+	return &netSwitch{h: map[netip.AddrPort]func(netip.AddrPort, []byte){}}
+}
+
+func (s *netSwitch) register(ep netip.AddrPort, fn func(netip.AddrPort, []byte)) {
+	s.mu.Lock()
+	s.h[ep] = fn
+	s.mu.Unlock()
+}
+
+func (s *netSwitch) deliver(dst, src netip.AddrPort, msg []byte) error {
+	s.mu.Lock()
+	fn := s.h[dst]
+	s.mu.Unlock()
+	if fn == nil {
+		return fmt.Errorf("no route to %s", dst)
+	}
+	fn(src, msg)
+	return nil
+}
+
+// loopback is an endpoint-based pqkem.Transport over a netSwitch, with a switchable
+// drop flag.
+type loopback struct {
+	ep   netip.AddrPort
+	sw   *netSwitch
+	drop atomic.Bool
+}
+
+func (l *loopback) Send(dst netip.AddrPort, msg []byte) error {
+	if l.drop.Load() {
+		return nil
+	}
+	return l.sw.deliver(dst, l.ep, append([]byte(nil), msg...))
+}
+
+func (l *loopback) LocalPort() int                             { return int(l.ep.Port()) }
+func (l *loopback) Run(onInbound func(netip.AddrPort, []byte)) { l.sw.register(l.ep, onInbound) }
+func (l *loopback) Close() error                               { return nil }
+
+type fakeWG struct {
+	mu     sync.Mutex
+	psks   map[RemoteID]PSK
+	failed []RemoteID
+}
+
+func newFakeWG() *fakeWG { return &fakeWG{psks: map[RemoteID]PSK{}} }
+
+func (f *fakeWG) OnNewPSKReady(remoteID RemoteID, psk PSK) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.psks[remoteID] = psk
+	return nil
+}
+
+func (f *fakeWG) OnRekeyFailed(remoteID RemoteID) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.failed = append(f.failed, remoteID)
+	return nil
+}
+
+func (f *fakeWG) psk(peer RemoteID) PSK {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.psks[peer]
+}
+
+var (
+	epA = netip.MustParseAddrPort("100.64.0.1:51833")
+	epB = netip.MustParseAddrPort("100.64.0.2:51833")
+)
+
+// pair builds two wired managers (B is the initiator, "bbbb" > "aaaa") sharing a
+// netSwitch, with each peer's data-path endpoint registered. lbB is B's loopback
+// (for toggling drop).
+func pair(t *testing.T) (dA, dB *Manager, wgA, wgB *fakeWG, lbB *loopback) {
+	t.Helper()
+	sw := newSwitch()
+	wgA = newFakeWG()
+	wgB = newFakeWG()
+	dA = NewManager("aaaa", wgA, nil)
+	dB = NewManager("bbbb", wgB, nil)
+	dA.Start(&loopback{ep: epA, sw: sw})
+	lbB = &loopback{ep: epB, sw: sw}
+	dB.Start(lbB)
+	dA.AddPeer("bbbb", epB)
+	dB.AddPeer("aaaa", epA)
+	return dA, dB, wgA, wgB, lbB
+}
+
+// bootstrap runs the signalling offer/answer (the test plays the host carrying bytes).
+func bootstrap(t *testing.T, dA, dB *Manager) {
+	t.Helper()
+	offer, err := dB.SignalOffer("aaaa")
+	require.NoError(t, err)
+	require.NotNil(t, offer)
+	answer, err := dA.SignalOnOffer("bbbb", offer)
+	require.NoError(t, err)
+	require.NotNil(t, answer)
+	require.NoError(t, dB.SignalOnAnswer("aaaa", answer))
+}
+
+func TestManager_BootstrapDerivesSamePSK(t *testing.T) {
+	dA, dB, wgA, wgB, _ := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	bootstrap(t, dA, dB)
+
+	pskA := wgA.psk("bbbb")
+	pskB := wgB.psk("aaaa")
+	require.NotEqual(t, PSK{}, pskA)
+	require.Equal(t, pskB, pskA, "both sides derive the same PSK from the bootstrap exchange")
+}
+
+func TestManager_ChainRotatesAndAcks(t *testing.T) {
+	dA, dB, wgA, wgB, _ := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	bootstrap(t, dA, dB)
+	psk1 := wgB.psk("aaaa")
+
+	// Data path up: B (initiator) chains the next offer over the data path, which
+	// rotates both to a fresh PSK and acknowledges A.
+	dA.OnDataPathRekeyed("bbbb")
+	dB.OnDataPathRekeyed("aaaa")
+
+	psk2A := wgA.psk("bbbb")
+	psk2B := wgB.psk("aaaa")
+	require.Equal(t, psk2B, psk2A, "both sides converge on the rotated PSK")
+	require.NotEqual(t, psk1, psk2B, "the chain rotated to a new PSK")
+}
+
+func TestManager_NonInitiatorReturnsNoOffer(t *testing.T) {
+	dA := NewManager("aaaa", newFakeWG(), nil)
+	defer dA.Stop()
+
+	offer, err := dA.SignalOffer("bbbb") // not the initiator vs "bbbb"
+	require.NoError(t, err)
+	require.Nil(t, offer)
+}
+
+func TestManager_StopIsIdempotent(t *testing.T) {
+	dA := NewManager("aaaa", newFakeWG(), nil)
+	dA.Start(&loopback{ep: epA, sw: newSwitch()})
+	dA.Stop()
+	dA.Stop() // must not panic or hang
+}

client/internal/pqkem/message.go

@@ -0,0 +1,121 @@
+package pqkem
+
+import (
+	"crypto/mlkem"
+	"fmt"
+)
+
+// Wire framing for the PQ-KEM exchange. Messages are self-contained, versioned,
+// transport-agnostic byte blobs: the same bytes ride the signalling channel
+// (initial bootstrap) or a data-tunnel packet (rekey). The library only ever sees
+// opaque []byte at the transport seam.
+//
+// Layout (all messages): [type:1][version:1][exchangeID:16][payload...]
+//
+// There is no confirm message: an exchange is acknowledged by the NEXT offer, which
+// carries the acked exchange's id (see OfferMsg.AckID) and — riding the data path
+// under the freshly adopted key — proves that key works.
+
+const (
+	// ProtocolVersion is bumped on any wire-incompatible change; a peer rejects
+	// messages it does not understand rather than misparsing them.
+	ProtocolVersion uint8 = 1
+
+	// ExchangeIDSize identifies one exchange so answers/acks correlate and stale
+	// messages are dropped.
+	ExchangeIDSize = 16
+
+	headerSize = 1 + 1 + ExchangeIDSize
+)
+
+// MsgType tags the two message kinds of the exchange.
+type MsgType uint8
+
+const (
+	MsgOffer MsgType = iota + 1
+	MsgAnswer
+)
+
+// ExchangeID is the per-exchange correlator. The zero value means "none" (an offer
+// that acknowledges nothing, i.e. the first exchange of a connection).
+type ExchangeID [ExchangeIDSize]byte
+
+// OfferMsg carries the initiator's public material (X25519 pub ‖ ML-KEM encap key)
+// and AckID, the id of the previous exchange this offer acknowledges (zero if none).
+type OfferMsg struct {
+	ExchangeID ExchangeID
+	AckID      ExchangeID
+	// KEMOffer is the raw Initiator.Offer() blob (OfferSize bytes).
+	KEMOffer []byte
+}
+
+// AnswerMsg carries the responder's reply (ML-KEM ciphertext ‖ X25519 pub) for the
+// round identified by ExchangeID.
+type AnswerMsg struct {
+	ExchangeID ExchangeID
+	// KEMAnswer is the raw Respond() answer blob (AnswerSize bytes).
+	KEMAnswer []byte
+}
+
+// Encode serialises the offer with its framed header (payload = AckID ‖ KEMOffer).
+func (m *OfferMsg) Encode() ([]byte, error) {
+	if len(m.KEMOffer) != OfferSize {
+		return nil, fmt.Errorf("offer payload: got %d, want %d", len(m.KEMOffer), OfferSize)
+	}
+	payload := make([]byte, 0, ExchangeIDSize+OfferSize)
+	payload = append(payload, m.AckID[:]...)
+	payload = append(payload, m.KEMOffer...)
+	return frame(MsgOffer, m.ExchangeID, payload), nil
+}
+
+// Encode serialises the answer with its framed header.
+func (m *AnswerMsg) Encode() ([]byte, error) {
+	if len(m.KEMAnswer) != AnswerSize {
+		return nil, fmt.Errorf("answer payload: got %d, want %d", len(m.KEMAnswer), AnswerSize)
+	}
+	return frame(MsgAnswer, m.ExchangeID, m.KEMAnswer), nil
+}
+
+// Decode parses a framed message into one of *OfferMsg / *AnswerMsg.
+func Decode(buf []byte) (MsgType, any, error) {
+	if len(buf) < headerSize {
+		return 0, nil, fmt.Errorf("message too short: %d bytes", len(buf))
+	}
+	typ := MsgType(buf[0])
+	if ver := buf[1]; ver != ProtocolVersion {
+		return typ, nil, fmt.Errorf("unsupported protocol version %d (want %d)", ver, ProtocolVersion)
+	}
+
+	var id ExchangeID
+	copy(id[:], buf[2:headerSize])
+	payload := buf[headerSize:]
+
+	switch typ {
+	case MsgOffer:
+		if len(payload) != ExchangeIDSize+OfferSize {
+			return typ, nil, fmt.Errorf("offer payload: got %d, want %d", len(payload), ExchangeIDSize+OfferSize)
+		}
+		var ack ExchangeID
+		copy(ack[:], payload[:ExchangeIDSize])
+		return typ, &OfferMsg{ExchangeID: id, AckID: ack, KEMOffer: payload[ExchangeIDSize:]}, nil
+	case MsgAnswer:
+		if len(payload) != AnswerSize {
+			return typ, nil, fmt.Errorf("answer payload: got %d, want %d", len(payload), AnswerSize)
+		}
+		return typ, &AnswerMsg{ExchangeID: id, KEMAnswer: payload}, nil
+	default:
+		return typ, nil, fmt.Errorf("unknown message type %d", typ)
+	}
+}
+
+func frame(typ MsgType, id ExchangeID, payload []byte) []byte {
+	buf := make([]byte, headerSize+len(payload))
+	buf[0] = byte(typ)
+	buf[1] = ProtocolVersion
+	copy(buf[2:], id[:])
+	copy(buf[headerSize:], payload)
+	return buf
+}
+
+// compile-time assurance the KEM blob sizes referenced here stay in sync with kem.go.
+var _ = [1]struct{}{}[OfferSize-(32+mlkem.EncapsulationKeySize768)]

client/internal/pqkem/message_test.go

@@ -0,0 +1,57 @@
+package pqkem
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestMessageRoundTrip(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+	answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	id := ExchangeID{1, 2, 3, 4}
+	ack := ExchangeID{9, 9, 9}
+
+	offBytes, err := (&OfferMsg{ExchangeID: id, AckID: ack, KEMOffer: init.Offer()}).Encode()
+	require.NoError(t, err)
+	typ, decoded, err := Decode(offBytes)
+	require.NoError(t, err)
+	require.Equal(t, MsgOffer, typ)
+	require.Equal(t, id, decoded.(*OfferMsg).ExchangeID)
+	require.Equal(t, ack, decoded.(*OfferMsg).AckID)
+	require.Equal(t, init.Offer(), decoded.(*OfferMsg).KEMOffer)
+
+	ansBytes, err := (&AnswerMsg{ExchangeID: id, KEMAnswer: answer}).Encode()
+	require.NoError(t, err)
+	typ, decoded, err = Decode(ansBytes)
+	require.NoError(t, err)
+	require.Equal(t, MsgAnswer, typ)
+	require.Equal(t, answer, decoded.(*AnswerMsg).KEMAnswer)
+}
+
+func TestDecodeRejects(t *testing.T) {
+	// too short
+	_, _, err := Decode([]byte{1, 1})
+	require.Error(t, err)
+
+	// wrong version
+	bad := make([]byte, headerSize+ExchangeIDSize+OfferSize)
+	bad[0] = byte(MsgOffer)
+	bad[1] = ProtocolVersion + 1
+	_, _, err = Decode(bad)
+	require.Error(t, err)
+
+	// unknown type
+	bad2 := make([]byte, headerSize)
+	bad2[0] = 99
+	bad2[1] = ProtocolVersion
+	_, _, err = Decode(bad2)
+	require.Error(t, err)
+
+	// offer with wrong payload size
+	_, err = (&OfferMsg{KEMOffer: []byte{1, 2, 3}}).Encode()
+	require.Error(t, err)
+}

client/internal/pqkem_adapter.go

@@ -0,0 +1,102 @@
+package internal
+
+import (
+	"net/netip"
+
+	log "github.com/sirupsen/logrus"
+	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
+
+	"github.com/netbirdio/netbird/client/internal/pqkem"
+)
+
+// pqPresharedKeySetter is the subset of the WireGuard interface the ML-KEM callback
+// needs: programming a peer's preshared key. *iface.WGIface satisfies it.
+type pqPresharedKeySetter interface {
+	SetPresharedKey(peerKey string, psk wgtypes.Key, updateOnly bool) error
+}
+
+// pqCallbackHandler programs the derived PQ PSK onto the WireGuard peer. It is the
+// engine-side implementation of pqkem.CallbackHandler.
+type pqCallbackHandler struct {
+	wg pqPresharedKeySetter
+}
+
+// OnNewPSKReady programs the freshly derived PSK for the peer (updateOnly: a no-op
+// if the peer is not present, mirroring Rosenpass). remoteID is the peer's WG pubkey.
+func (h pqCallbackHandler) OnNewPSKReady(remoteID pqkem.RemoteID, psk pqkem.PSK) error {
+	// updateOnly: applies to an already-configured peer (rotation). At bootstrap the
+	// peer is not configured yet, so this is a no-op there and the PSK is instead
+	// pulled at peer-config time (pqHandshaker.PSK / conn.presharedKey).
+	log.Debugf("pqkem: programming PSK for peer %s", remoteID)
+	return h.wg.SetPresharedKey(string(remoteID), wgtypes.Key(psk), true)
+}
+
+// OnRekeyFailed reports a failed PQ (re)key convergence.
+// TODO(NET-1406): tear the peer connection down / trigger ICE reconnect.
+func (h pqCallbackHandler) OnRekeyFailed(remoteID pqkem.RemoteID) error {
+	log.Warnf("pqkem: post-quantum rekey failed for peer %s", remoteID)
+	return nil
+}
+
+// pqHandshaker adapts the pqkem manager to peer.PQHandshaker (string peer keys),
+// wiring the host's signalling offers/answers to the KEM exchange.
+type pqHandshaker struct {
+	mgr *pqkem.Manager
+}
+
+func (p pqHandshaker) OfferPayload(remoteKey string) ([]byte, int) {
+	payload, err := p.mgr.SignalOffer(pqkem.RemoteID(remoteKey))
+	if err != nil {
+		log.Warnf("pqkem: build offer for %s: %v", remoteKey, err)
+	}
+	return payload, p.mgr.LocalPort()
+}
+
+func (p pqHandshaker) AnswerPayload(remoteKey string, recvOffer []byte) ([]byte, int) {
+	if len(recvOffer) == 0 {
+		return nil, p.mgr.LocalPort()
+	}
+	payload, err := p.mgr.SignalOnOffer(pqkem.RemoteID(remoteKey), recvOffer)
+	if err != nil {
+		log.Warnf("pqkem: build answer for %s: %v", remoteKey, err)
+	}
+	return payload, p.mgr.LocalPort()
+}
+
+func (p pqHandshaker) OnAnswer(remoteKey string, recvAnswer []byte) {
+	if len(recvAnswer) == 0 {
+		return
+	}
+	if err := p.mgr.SignalOnAnswer(pqkem.RemoteID(remoteKey), recvAnswer); err != nil {
+		log.Warnf("pqkem: process answer from %s: %v", remoteKey, err)
+	}
+}
+
+// PSK exposes the peer's derived PSK for the conn to program at WG peer-config time.
+func (p pqHandshaker) PSK(remoteKey string) (wgtypes.Key, bool) {
+	psk, ok := p.mgr.PSK(pqkem.RemoteID(remoteKey))
+	if !ok {
+		return wgtypes.Key{}, false
+	}
+	return wgtypes.Key(psk), true
+}
+
+// SetRemoteAddr registers the peer's data-path endpoint (overlay IP + pq UDP port)
+// learned from signalling. Sends only ever fire once the tunnel is up (clocked by
+// OnDataPathRekeyed), so registering here is safe even before connection-up.
+func (p pqHandshaker) SetRemoteAddr(remoteKey string, addr netip.AddrPort) {
+	if !addr.IsValid() || addr.Port() == 0 {
+		return
+	}
+	p.mgr.AddPeer(pqkem.RemoteID(remoteKey), addr)
+}
+
+// OnDataPathRekeyed clocks the next chained PSK rotation on a fresh WG handshake.
+func (p pqHandshaker) OnDataPathRekeyed(remoteKey string) {
+	p.mgr.OnDataPathRekeyed(pqkem.RemoteID(remoteKey))
+}
+
+// OnDataPathDown signals the peer's tunnel went down.
+func (p pqHandshaker) OnDataPathDown(remoteKey string) {
+	p.mgr.OnDataPathDown(pqkem.RemoteID(remoteKey))
+}

client/internal/pqkem_transport.go

@@ -0,0 +1,72 @@
+package internal
+
+import (
+	"fmt"
+	"net"
+	"net/netip"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// DefaultPort is the preferred UDP port for the ML-KEM data-path service, bound on
+// the WG overlay IP. Since each client owns a distinct overlay IP, this port is
+// almost always free, so it need not be announced (peers assume it). A peer only
+// announces Body.mlkemPort when a collision forced it onto a different port.
+const DefaultPort = 51833
+
+// pqTransport is the ML-KEM data-path transport: a dumb UDP socket bound on the WG
+// overlay IP. It implements pqkem.Transport — the manager owns the remoteID<->endpoint
+// routing and drives this socket's lifecycle (Run / Close).
+type pqTransport struct {
+	conn *net.UDPConn
+	port int
+}
+
+// newPQTransport binds a UDP socket on the WG overlay IP, preferring DefaultPort and
+// falling back to an OS-assigned ephemeral port if it is in use. Call it after the WG
+// interface is up so the overlay IP is assigned; when the bound port is not
+// DefaultPort it must be announced to peers via Body.mlkemPort.
+func newPQTransport(overlayIP netip.Addr) (*pqTransport, error) {
+	if !overlayIP.IsValid() {
+		return nil, fmt.Errorf("invalid overlay IP for pqkem transport")
+	}
+	ip := net.IP(overlayIP.AsSlice())
+	conn, err := net.ListenUDP("udp4", &net.UDPAddr{IP: ip, Port: DefaultPort})
+	if err != nil {
+		log.Debugf("pqkem: default port %d unavailable on %s (%v), using an ephemeral port", DefaultPort, overlayIP, err)
+		conn, err = net.ListenUDP("udp4", &net.UDPAddr{IP: ip, Port: 0})
+		if err != nil {
+			return nil, fmt.Errorf("bind pqkem udp on overlay %s: %w", overlayIP, err)
+		}
+	}
+	return &pqTransport{conn: conn, port: conn.LocalAddr().(*net.UDPAddr).Port}, nil
+}
+
+// Send implements pqkem.Transport.
+func (t *pqTransport) Send(endpoint netip.AddrPort, msg []byte) error {
+	_, err := t.conn.WriteToUDPAddrPort(msg, endpoint)
+	return err
+}
+
+// LocalPort implements pqkem.Transport.
+func (t *pqTransport) LocalPort() int { return t.port }
+
+// Run implements pqkem.Transport: the receive loop, delivering each datagram as
+// (source endpoint, msg). Exits when the socket is closed.
+func (t *pqTransport) Run(onInbound func(src netip.AddrPort, msg []byte)) {
+	go func() {
+		buf := make([]byte, 2048)
+		for {
+			n, src, err := t.conn.ReadFromUDPAddrPort(buf)
+			if err != nil {
+				return
+			}
+			msg := make([]byte, n)
+			copy(msg, buf[:n])
+			onInbound(src, msg)
+		}
+	}()
+}
+
+// Close implements pqkem.Transport.
+func (t *pqTransport) Close() error { return t.conn.Close() }

client/internal/profilemanager/config.go

@@ -96,6 +96,7 @@ type ConfigInput struct {
 	BlockLANAccess      *bool
 	BlockInbound        *bool
 	DisableIPv6         *bool
+	SyncMessageVersion  *int
 
 	DisableNotifications *bool
 
@@ -137,6 +138,7 @@ type Config struct {
 	BlockLANAccess      bool
 	BlockInbound        bool
 	DisableIPv6         bool
+	SyncMessageVersion  *int
 
 	DisableNotifications *bool
 
@@ -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")

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

@@ -61,6 +61,7 @@ type Manager interface {
 	InitialRouteRange() []string
 	SetFirewall(firewall.Manager) error
 	SetDNSForwarderPort(port uint16)
+	ReconcilePeerAllowedIPs(peerKey string) error
 	Stop(stateManager *statemanager.Manager)
 }
 
@@ -232,6 +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()

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/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/version.go

@@ -0,0 +1,12 @@
+//go:build ios
+
+package NetBirdSDK
+
+import "github.com/netbirdio/netbird/version"
+
+// GoClientVersion returns the NetBird Go client version that was baked into
+// the framework at compile time via
+// -ldflags "-X github.com/netbirdio/netbird/version.version=<version>".
+func GoClientVersion() string {
+	return version.NetbirdVersion()
+}

client/server/server.go

@@ -1081,7 +1081,10 @@ func (s *Server) Down(ctx context.Context, _ *proto.DownRequest) (*proto.DownRes
 
 	if err := s.cleanupConnection(); err != nil {
 		s.mutex.Unlock()
-		// todo review to update the status in case any type of error
+		if errors.Is(err, ErrServiceNotUp) {
+			log.Debugf("Down called while service not up: %v", err)
+			return nil, err
+		}
 		log.Errorf("failed to shut down properly: %v", err)
 		return nil, err
 	}
@@ -1154,7 +1157,7 @@ func (s *Server) cleanupConnection() error {
 	// making the run loop the sole owner of engine shutdown.
 	if engine != nil {
 		if err := engine.Stop(); err != nil {
-			return err
+			log.Errorf("failed to stop engine during cleanup: %v", err)
 		}
 	}
 

client/system/info.go

@@ -79,13 +79,15 @@ type Info struct {
 	EnableSSHLocalPortForwarding  bool
 	EnableSSHRemotePortForwarding bool
 	DisableSSHAuth                bool
+
+	SyncMessageVersion *int
 }
 
 func (i *Info) SetFlags(
 	rosenpassEnabled, rosenpassPermissive bool,
 	serverSSHAllowed *bool,
 	disableClientRoutes, disableServerRoutes,
-	disableDNS, disableFirewall, blockLANAccess, blockInbound, disableIPv6 bool,
+	disableDNS, disableFirewall, blockLANAccess, blockInbound, disableIPv6 bool, syncMessageVersion *int,
 	enableSSHRoot, enableSSHSFTP, enableSSHLocalPortForwarding, enableSSHRemotePortForwarding *bool,
 	disableSSHAuth *bool,
 ) {
@@ -103,6 +105,8 @@ func (i *Info) SetFlags(
 	i.BlockInbound = blockInbound
 	i.DisableIPv6 = disableIPv6
 
+	i.SyncMessageVersion = syncMessageVersion
+
 	if enableSSHRoot != nil {
 		i.EnableSSHRoot = *enableSSHRoot
 	}

client/ui/autostart_default.go

@@ -51,7 +51,7 @@ func autostartDisabledByMDM(policy *mdm.Policy) bool {
 // netbirdFootprintExists reports whether the machine already carries NetBird
 // daemon config or state, meaning this is not a genuinely fresh install. It is
 // the update-safety gate for the autostart default: upgrading users always
-// have a footprint, so an update can never trigger a login-item write.
+// have a footprint, so an update can never trigger a autostart entry write.
 func netbirdFootprintExists() bool {
 	candidates := []string{
 		profilemanager.DefaultConfigPath,
@@ -69,9 +69,23 @@ func netbirdFootprintExists() bool {
 // applyAutostartDefault runs the one-time launch-on-login default for genuinely
 // fresh installs. The autostartInitialized marker is persisted before any
 // enable attempt so a crash mid-flow degrades to "never enabled" instead of
-// retrying login-item writes on every launch. A user's later disable in
+// retrying autostart entry writes on every launch. A user's later disable in
 // Settings is never overridden: the marker guarantees at-most-once, ever.
 func applyAutostartDefault(ctx context.Context, autostart *services.Autostart, prefs *preferences.Store, prefsFileExisted bool) {
+	mdmDisabled := autostartDisabledByMDM(mdm.LoadPolicy())
+
+	if mdmDisabled {
+		if enabled, err := autostart.IsEnabled(ctx); err != nil {
+			log.Warnf("MDM disableAutostart: read autostart state: %v", err)
+		} else if enabled {
+			if err := autostart.SetEnabled(ctx, false); err != nil {
+				log.Warnf("MDM disableAutostart: force off failed: %v", err)
+			} else {
+				log.Info("MDM disableAutostart enforced: autostart turned off")
+			}
+		}
+	}
+
 	priorFootprint := netbirdFootprintExists() || prefsFileExisted
 
 	if prefs.Get().AutostartInitialized {
@@ -84,7 +98,7 @@ func applyAutostartDefault(ctx context.Context, autostart *services.Autostart, p
 
 	state := autostartDefaultState{
 		supported:    autostart.Supported(ctx),
-		mdmDisabled:  autostartDisabledByMDM(mdm.LoadPolicy()),
+		mdmDisabled:  mdmDisabled,
 		priorInstall: priorFootprint,
 	}
 	enable, reason := shouldEnableAutostartDefault(state)

client/ui/frontend/src/components/empty-state/DaemonOutdatedOverlay.tsx

@@ -1,10 +1,13 @@
+import { useEffect, useState } from "react";
 import { useTranslation } from "react-i18next";
 import { AlertTriangleIcon, DownloadIcon } from "lucide-react";
 import { Browser } from "@wailsio/runtime";
+import { Version } from "@bindings/services";
 import { Button } from "@/components/buttons/Button";
 import { useStatus } from "@/contexts/StatusContext.tsx";
 
 const RELEASES_URL = "https://github.com/netbirdio/netbird/releases/latest";
+const RC_RELEASES_URL = "https://pkgs.netbird.io/releases/rc";
 
 function openUrl(url: string) {
     Browser.OpenURL(url).catch(() => globalThis.open(url, "_blank"));
@@ -12,7 +15,26 @@ function openUrl(url: string) {
 
 export const DaemonOutdatedOverlay = () => {
     const { t } = useTranslation();
-    const { isDaemonOutdated } = useStatus();
+    const { status, isDaemonOutdated } = useStatus();
+
+    const [guiVersion, setGuiVersion] = useState<string>("-");
+    const clientVersion = status?.daemonVersion ?? "—";
+
+    const isRc = /-rc/i.test(guiVersion) || /-rc/i.test(clientVersion);
+    const downloadUrl = isRc ? RC_RELEASES_URL : RELEASES_URL;
+
+    useEffect(() => {
+        if (!isDaemonOutdated) return;
+        let cancelled = false;
+        Version.GUI()
+            .then((v) => {
+                if (!cancelled) setGuiVersion(v);
+            })
+            .catch((err) => console.error("[DaemonOutdatedOverlay] GUI version error", err));
+        return () => {
+            cancelled = true;
+        };
+    }, [isDaemonOutdated]);
 
     if (!isDaemonOutdated) return null;
 
@@ -38,10 +60,37 @@ export const DaemonOutdatedOverlay = () => {
                     <p className={"text-sm text-nb-gray-300"}>{t("daemon.outdated.description")}</p>
                 </div>
 
+                <div className={"flex flex-col items-center gap-0.5 text-center"}>
+                    <p className={"text-sm font-semibold text-nb-gray-100"}>
+                        {clientVersion === "development" ? (
+                            <span>
+                                {t("settings.about.clientName")}{" "}
+                                <span className={"font-mono text-yellow-400"}>
+                                    {t("settings.about.development")}
+                                </span>
+                            </span>
+                        ) : (
+                            t("settings.about.client", { version: clientVersion })
+                        )}
+                    </p>
+                    <p className={"text-sm font-medium text-nb-gray-250"}>
+                        {guiVersion === "development" ? (
+                            <span>
+                                {t("settings.about.guiName")}{" "}
+                                <span className={"font-mono text-yellow-400"}>
+                                    {t("settings.about.development")}
+                                </span>
+                            </span>
+                        ) : (
+                            t("settings.about.gui", { version: guiVersion })
+                        )}
+                    </p>
+                </div>
+
                 <div className={"wails-no-draggable"}>
-                    <Button variant={"primary"} size={"xs"} onClick={() => openUrl(RELEASES_URL)}>
+                    <Button variant={"primary"} size={"xs"} onClick={() => openUrl(downloadUrl)}>
                         <DownloadIcon size={14} />
-                        {t("update.card.getInstaller")}
+                        {t("daemon.outdated.download")}
                     </Button>
                 </div>
             </div>

client/ui/frontend/src/contexts/ProfileContext.tsx

@@ -28,6 +28,7 @@ type ProfileContextValue = {
     loaded: boolean;
     refresh: () => Promise<void>;
     switchProfile: (id: string) => Promise<void>;
+    switchProfileNoConnect: (id: string) => Promise<void>;
     addProfile: (name: string) => Promise<string>;
     removeProfile: (id: string) => Promise<void>;
     renameProfile: (id: string, newName: string) => Promise<void>;
@@ -112,6 +113,16 @@ export const ProfileProvider = ({ children }: { children: ReactNode }) => {
         [username, refresh],
     );
 
+    // Manage-profiles variant: switches without connecting, so the user can
+    // still adjust the management URL before bringing the connection up.
+    const switchProfileNoConnect = useCallback(
+        async (id: string) => {
+            await ProfileSwitcher.SwitchActiveNoConnect({ profileName: id, username });
+            await refresh();
+        },
+        [username, refresh],
+    );
+
     // addProfile creates a profile by display name and returns the
     // daemon-generated ID, so the caller can immediately address it by ID.
     const addProfile = useCallback(
@@ -158,6 +169,7 @@ export const ProfileProvider = ({ children }: { children: ReactNode }) => {
             loaded,
             refresh,
             switchProfile,
+            switchProfileNoConnect,
             addProfile,
             removeProfile,
             renameProfile,
@@ -171,6 +183,7 @@ export const ProfileProvider = ({ children }: { children: ReactNode }) => {
             loaded,
             refresh,
             switchProfile,
+            switchProfileNoConnect,
             addProfile,
             removeProfile,
             renameProfile,

client/ui/frontend/src/modules/profiles/ProfilesTab.tsx

@@ -45,7 +45,7 @@ export function ProfilesTab() {
         activeProfileId,
         loaded,
         username,
-        switchProfile,
+        switchProfileNoConnect,
         addProfile,
         removeProfile,
         renameProfile,
@@ -100,7 +100,7 @@ export function ProfilesTab() {
             confirmLabel: t("profile.switch.confirm"),
         });
         if (!ok) return;
-        await guarded(i18next.t("profile.error.switchTitle"), () => switchProfile(id));
+        await guarded(i18next.t("profile.error.switchTitle"), () => switchProfileNoConnect(id));
     };
 
     const handleDeregister = async (id: string, name: string) => {
@@ -129,14 +129,13 @@ export function ProfilesTab() {
         await guarded(i18next.t("profile.error.createTitle"), async () => {
             const id = await addProfile(name);
             // SetConfig is keyed by the new profile's ID, so it writes the
-            // not-yet-active profile. Write before switching so any reconnect
-            // targets the right deployment.
+            // not-yet-active profile before the switch makes it current.
             if (!isNetbirdCloud(managementUrl)) {
                 await SettingsSvc.SetConfig(
                     new SetConfigParams({ profileName: id, username, managementUrl }),
                 );
             }
-            await switchProfile(id);
+            await switchProfileNoConnect(id);
         });
     };
 

client/ui/frontend/src/modules/session/SessionExpirationDialog.tsx

@@ -73,6 +73,13 @@ export default function SessionExpirationDialog() {
 
         let offCancel: (() => void) | undefined;
 
+        // Return the dialog to its interactive state and dismiss the browser popup
+        const resetDialog = () => {
+            offCancel?.();
+            WindowManager.CloseBrowserLogin().catch(console.error);
+            setBusy(false);
+        };
+
         try {
             const start = await Session.RequestExtend({ hint: "" });
             const uri = start.verificationUriComplete || start.verificationUri;
@@ -105,25 +112,22 @@ export default function SessionExpirationDialog() {
             if (outcome.kind === "cancel") {
                 waitPromise.cancel?.();
                 waitPromise.catch(() => {});
+                resetDialog();
                 return;
             }
 
             // Another surface owns this flow; keep the dialog open to retry.
             if (outcome.result.preempted) {
+                resetDialog();
                 return;
             }
-
-            // Close before the popup so the restore can't flash this window back.
-            WindowManager.CloseSessionExpiration().catch(console.error);
+            WindowManager.CloseRenewFlow().catch(console.error);
         } catch (e) {
+            resetDialog();
             await errorDialog({
                 Title: t("sessionExpiration.extendFailedTitle"),
                 Message: formatErrorMessage(e),
             });
-        } finally {
-            offCancel?.();
-            WindowManager.CloseBrowserLogin().catch(console.error);
-            setBusy(false);
         }
     }, [busy, t]);
 
@@ -139,12 +143,11 @@ export default function SessionExpirationDialog() {
             });
             WindowManager.CloseSessionExpiration().catch(console.error);
         } catch (e) {
+            setBusy(false);
             await errorDialog({
                 Title: t("sessionExpiration.logoutFailedTitle"),
                 Message: formatErrorMessage(e),
             });
-        } finally {
-            setBusy(false);
         }
     }, [busy, t]);
 

client/ui/frontend/src/modules/welcome/WelcomeStepTray.tsx

@@ -22,6 +22,9 @@ type WelcomeStepTrayProps = {
 export function WelcomeStepTray({ onContinue }: Readonly<WelcomeStepTrayProps>) {
     const { t } = useTranslation();
     const trayScreenshot = trayScreenshotForOS();
+    // macOS has no tray — the icon sits in the menu bar, so the copy says so.
+    const titleKey = isMacOS() ? "welcome.titleMac" : "welcome.title";
+    const descriptionKey = isMacOS() ? "welcome.descriptionMac" : "welcome.description";
 
     return (
         <>
@@ -36,9 +39,9 @@ export function WelcomeStepTray({ onContinue }: Readonly<WelcomeStepTrayProps>)
 
             <div className={"flex w-full flex-col gap-1"}>
                 <DialogHeading id={"nb-welcome-title"} align={"left"}>
-                    {t("welcome.title")}
+                    {t(titleKey)}
                 </DialogHeading>
-                <DialogDescription align={"left"}>{t("welcome.description")}</DialogDescription>
+                <DialogDescription align={"left"}>{t(descriptionKey)}</DialogDescription>
             </div>
 
             <DialogActions>

client/ui/i18n/locales/de/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Suchen Sie NetBird in der Taskleiste"
     },
+    "welcome.titleMac": {
+        "message": "Suchen Sie NetBird in der Menüleiste"
+    },
     "welcome.description": {
         "message": "NetBird läuft in Ihrer Taskleiste. Klicken Sie auf das Symbol, um sich zu verbinden, Profile zu wechseln oder die Einstellungen zu öffnen."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird läuft in Ihrer Menüleiste. Klicken Sie auf das Symbol, um sich zu verbinden, Profile zu wechseln oder die Einstellungen zu öffnen."
+    },
     "welcome.continue": {
         "message": "Weiter"
     },
@@ -1293,10 +1299,13 @@
         "message": "Dokumentation"
     },
     "daemon.outdated.title": {
-        "message": "NetBird-Dienst ist veraltet"
+        "message": "NetBird Client ist veraltet"
     },
     "daemon.outdated.description": {
-        "message": "Aktualisieren Sie den NetBird-Dienst, um diese App zu verwenden."
+        "message": "Die neue GUI ist nicht mit Ihrem älteren Client kompatibel. Aktualisieren Sie Ihren Client, um die neue Anwendung zu verwenden."
+    },
+    "daemon.outdated.download": {
+        "message": "Neueste Version herunterladen"
     },
     "error.jwt_clock_skew": {
         "message": "Anmeldung fehlgeschlagen: Die Uhr dieses Geräts ist nicht mit dem Server synchron. Bitte synchronisieren Sie die Systemuhr und versuchen Sie es erneut."

client/ui/i18n/locales/en/common.json

@@ -1377,11 +1377,19 @@
     },
     "welcome.title": {
         "message": "Look for NetBird in your tray",
-        "description": "Heading on the first onboarding step, pointing the user to the tray icon. 'tray' = system tray / menu bar."
+        "description": "Heading on the first onboarding step, pointing the user to the tray icon. Shown on Windows and Linux; macOS uses welcome.titleMac."
+    },
+    "welcome.titleMac": {
+        "message": "Look for NetBird in your menu bar",
+        "description": "Heading on the first onboarding step on macOS, pointing the user to the menu bar icon. Use your language's Apple term for the macOS menu bar."
     },
     "welcome.description": {
         "message": "NetBird lives in your tray. Click the icon to connect, switch profiles, or open settings.",
-        "description": "Body of the first onboarding step explaining the tray icon."
+        "description": "Body of the first onboarding step explaining the tray icon. Shown on Windows and Linux; macOS uses welcome.descriptionMac."
+    },
+    "welcome.descriptionMac": {
+        "message": "NetBird lives in your menu bar. Click the icon to connect, switch profiles, or open settings.",
+        "description": "Body of the first onboarding step on macOS explaining the menu bar icon. Use your language's Apple term for the macOS menu bar."
     },
     "welcome.continue": {
         "message": "Continue",
@@ -1724,12 +1732,16 @@
         "description": "Documentation link on the daemon-unavailable overlay."
     },
     "daemon.outdated.title": {
-        "message": "NetBird Service Is Outdated",
-        "description": "Title of the overlay shown when the NetBird background service is too old to drive this UI."
+        "message": "NetBird Client Is Outdated",
+        "description": "Title of the overlay shown when the NetBird client (daemon) is too old to drive this UI."
     },
     "daemon.outdated.description": {
-        "message": "Update the NetBird service to use this app.",
-        "description": "Body of the daemon-outdated overlay telling the user to upgrade the service."
+        "message": "The new GUI isn't compatible with the older NetBird client. Update your client to use the new application.",
+        "description": "Body of the daemon-outdated overlay explaining that the GUI is newer than the client and the client must be updated."
+    },
+    "daemon.outdated.download": {
+        "message": "Download Latest",
+        "description": "Button on the daemon-outdated overlay that opens the download page for the latest release."
     },
     "error.jwt_clock_skew": {
         "message": "Sign-in failed: this device's clock is out of sync with the server. Please sync your system clock and try again.",

client/ui/i18n/locales/es/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Busque NetBird en su bandeja del sistema"
     },
+    "welcome.titleMac": {
+        "message": "Busque NetBird en su barra de menús"
+    },
     "welcome.description": {
         "message": "NetBird reside en su bandeja del sistema. Haga clic en el icono para conectarse, cambiar de perfil o abrir la configuración."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird reside en su barra de menús. Haga clic en el icono para conectarse, cambiar de perfil o abrir la configuración."
+    },
     "welcome.continue": {
         "message": "Continuar"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentación"
     },
     "daemon.outdated.title": {
-        "message": "El servicio de NetBird está desactualizado"
+        "message": "NetBird Client está desactualizado"
     },
     "daemon.outdated.description": {
-        "message": "Actualice el servicio de NetBird para usar esta aplicación."
+        "message": "La nueva GUI no es compatible con su cliente anterior. Actualice su cliente para usar la nueva aplicación."
+    },
+    "daemon.outdated.download": {
+        "message": "Descargar la última versión"
     },
     "error.jwt_clock_skew": {
         "message": "Error al iniciar sesión: el reloj de este dispositivo no está sincronizado con el servidor. Sincronice el reloj del sistema e inténtelo de nuevo."

client/ui/i18n/locales/fr/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Cherchez NetBird dans votre barre d’état système"
     },
+    "welcome.titleMac": {
+        "message": "Cherchez NetBird dans votre barre des menus"
+    },
     "welcome.description": {
         "message": "NetBird se trouve dans votre barre d’état système. Cliquez sur l’icône pour vous connecter, changer de profil ou ouvrir les paramètres."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird se trouve dans votre barre des menus. Cliquez sur l’icône pour vous connecter, changer de profil ou ouvrir les paramètres."
+    },
     "welcome.continue": {
         "message": "Continuer"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentation"
     },
     "daemon.outdated.title": {
-        "message": "Le service NetBird est obsolète"
+        "message": "Le Client NetBird est obsolète"
     },
     "daemon.outdated.description": {
-        "message": "Mettez à jour le service NetBird pour utiliser cette application."
+        "message": "La nouvelle GUI n'est pas compatible avec votre ancien client. Mettez à jour votre client pour utiliser la nouvelle application."
+    },
+    "daemon.outdated.download": {
+        "message": "Télécharger la dernière version"
     },
     "error.jwt_clock_skew": {
         "message": "Échec de la connexion : l’horloge de cet appareil n’est pas synchronisée avec le serveur. Veuillez synchroniser l’horloge de votre système et réessayer."

client/ui/i18n/locales/hu/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Keresse a NetBirdöt a tálcán"
     },
+    "welcome.titleMac": {
+        "message": "Keresse a NetBirdöt a menüsorban"
+    },
     "welcome.description": {
         "message": "A NetBird a tálcán fut. Kattintson az ikonra a csatlakozáshoz, profilváltáshoz vagy a beállítások megnyitásához."
     },
+    "welcome.descriptionMac": {
+        "message": "A NetBird a menüsorban fut. Kattintson az ikonra a csatlakozáshoz, profilváltáshoz vagy a beállítások megnyitásához."
+    },
     "welcome.continue": {
         "message": "Folytatás"
     },
@@ -1293,10 +1299,13 @@
         "message": "Dokumentáció"
     },
     "daemon.outdated.title": {
-        "message": "A NetBird szolgáltatás elavult"
+        "message": "A NetBird Kliens elavult"
     },
     "daemon.outdated.description": {
-        "message": "Frissítsd a NetBird szolgáltatást az alkalmazás használatához."
+        "message": "Az új GUI nem kompatibilis a régebbi klienseddel. Frissítsd a klienst az új alkalmazás használatához."
+    },
+    "daemon.outdated.download": {
+        "message": "Legújabb letöltése"
     },
     "error.jwt_clock_skew": {
         "message": "A bejelentkezés sikertelen: az eszköz órája eltér a szerverétől. Kérjük, szinkronizálja a rendszer óráját, majd próbálja újra."

client/ui/i18n/locales/it/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Cerchi NetBird nella tray"
     },
+    "welcome.titleMac": {
+        "message": "Cerchi NetBird nella barra dei menu"
+    },
     "welcome.description": {
         "message": "NetBird risiede nella tray. Clicchi sull'icona per connettersi, cambiare profilo o aprire le impostazioni."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird risiede nella barra dei menu. Clicchi sull'icona per connettersi, cambiare profilo o aprire le impostazioni."
+    },
     "welcome.continue": {
         "message": "Continua"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentazione"
     },
     "daemon.outdated.title": {
-        "message": "Il servizio NetBird è obsoleto"
+        "message": "NetBird Client è obsoleto"
     },
     "daemon.outdated.description": {
-        "message": "Aggiorna il servizio NetBird per usare questa app."
+        "message": "La nuova GUI non è compatibile con il tuo client precedente. Aggiorna il client per usare la nuova applicazione."
+    },
+    "daemon.outdated.download": {
+        "message": "Scarica l'ultima versione"
     },
     "error.jwt_clock_skew": {
         "message": "Accesso non riuscito: l'orologio di questo dispositivo non è sincronizzato con il server. Sincronizzi l'orologio di sistema e riprovi."

client/ui/i18n/locales/ja/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "トレイの NetBird を確認してください"
     },
+    "welcome.titleMac": {
+        "message": "メニューバーの NetBird を確認してください"
+    },
     "welcome.description": {
         "message": "NetBird はトレイに常駐します。アイコンをクリックして、接続、プロファイルの切り替え、設定を開くことができます。"
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird はメニューバーに常駐します。アイコンをクリックして、接続、プロファイルの切り替え、設定を開くことができます。"
+    },
     "welcome.continue": {
         "message": "続ける"
     },

client/ui/i18n/locales/pt/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Procure o NetBird na sua bandeja"
     },
+    "welcome.titleMac": {
+        "message": "Procure o NetBird na sua barra de menus"
+    },
     "welcome.description": {
         "message": "O NetBird fica na sua bandeja. Clique no ícone para conectar, alternar perfis ou abrir as configurações."
     },
+    "welcome.descriptionMac": {
+        "message": "O NetBird fica na sua barra de menus. Clique no ícone para conectar, alternar perfis ou abrir as configurações."
+    },
     "welcome.continue": {
         "message": "Continuar"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentação"
     },
     "daemon.outdated.title": {
-        "message": "O serviço NetBird está desatualizado"
+        "message": "O NetBird Client está desatualizado"
     },
     "daemon.outdated.description": {
-        "message": "Atualize o serviço NetBird para usar este aplicativo."
+        "message": "A nova GUI não é compatível com o seu cliente mais antigo. Atualize o seu cliente para usar o novo aplicativo."
+    },
+    "daemon.outdated.download": {
+        "message": "Baixar a versão mais recente"
     },
     "error.jwt_clock_skew": {
         "message": "Falha no login: o relógio deste dispositivo está fora de sincronia com o servidor. Sincronize o relógio do sistema e tente novamente."

client/ui/i18n/locales/ru/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Найдите NetBird в системном трее"
     },
+    "welcome.titleMac": {
+        "message": "Найдите NetBird в строке меню"
+    },
     "welcome.description": {
         "message": "NetBird находится в системном трее. Нажмите на значок, чтобы подключиться, переключить профиль или открыть настройки."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird находится в строке меню. Нажмите на значок, чтобы подключиться, переключить профиль или открыть настройки."
+    },
     "welcome.continue": {
         "message": "Продолжить"
     },
@@ -1293,10 +1299,13 @@
         "message": "Документация"
     },
     "daemon.outdated.title": {
-        "message": "Служба NetBird устарела"
+        "message": "Клиент NetBird устарел"
     },
     "daemon.outdated.description": {
-        "message": "Обновите службу NetBird, чтобы использовать это приложение."
+        "message": "Новый GUI несовместим с вашим более старым клиентом. Обновите клиент, чтобы использовать новое приложение."
+    },
+    "daemon.outdated.download": {
+        "message": "Скачать последнюю версию"
     },
     "error.jwt_clock_skew": {
         "message": "Не удалось войти: часы этого устройства рассинхронизированы с сервером. Синхронизируйте системные часы и повторите попытку."

client/ui/i18n/locales/zh-CN/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "在托盘中查找 NetBird"
     },
+    "welcome.titleMac": {
+        "message": "在菜单栏中查找 NetBird"
+    },
     "welcome.description": {
         "message": "NetBird 驻留在您的托盘中。点击图标即可连接、切换配置文件或打开设置。"
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird 驻留在您的菜单栏中。点击图标即可连接、切换配置文件或打开设置。"
+    },
     "welcome.continue": {
         "message": "继续"
     },
@@ -1293,10 +1299,13 @@
         "message": "文档"
     },
     "daemon.outdated.title": {
-        "message": "NetBird 服务版本过旧"
+        "message": "NetBird 客户端版本过旧"
     },
     "daemon.outdated.description": {
-        "message": "请更新 NetBird 服务以使用此应用。"
+        "message": "新版 GUI 与您较旧的客户端不兼容。请更新客户端以使用新应用。"
+    },
+    "daemon.outdated.download": {
+        "message": "下载最新版本"
     },
     "error.jwt_clock_skew": {
         "message": "登录失败：此设备的时钟与服务器不同步。请同步您的系统时钟后重试。"

client/ui/services/autostart.go

@@ -10,7 +10,7 @@ import (
 	"github.com/wailsapp/wails/v3/pkg/application"
 )
 
-// Autostart facade over Wails' AutostartManager. The OS login-item registration
+// Autostart facade over Wails' AutostartManager. The OS autostart entry registration
 // is the single source of truth; nothing is mirrored to preferences.
 type Autostart struct {
 	mgr *application.AutostartManager

client/ui/services/profileswitcher.go

@@ -12,13 +12,15 @@ import (
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 )
 
-// ProfileSwitcher holds the reconnect policy shared by the tray and React
-// frontend so both flip profiles identically. The policy keys off prevStatus
-// from DaemonFeed.Get at SwitchActive entry:
+// ProfileSwitcher holds the switch policy shared by the tray and React
+// frontend so both flip profiles identically. SwitchActive (plain selection:
+// header dropdown, tray submenu) always connects after the switch;
+// SwitchActiveNoConnect (manage-profiles screen) never does, so the user can
+// still adjust the management URL before connecting. prevStatus from
+// DaemonFeed.Get at entry only decides the teardown:
 //
-//	Connected/Connecting → Switch + Down + Up; optimistic Connecting paint.
-//	NeedsLogin/LoginFailed/SessionExpired → Switch + Down; clear stale error for re-login.
-//	Idle → Switch only.
+//	Connected/Connecting/NeedsLogin/LoginFailed/SessionExpired → Down first.
+//	Idle → no Down.
 type ProfileSwitcher struct {
 	profiles   *Profiles
 	connection *Connection
@@ -29,29 +31,40 @@ func NewProfileSwitcher(profiles *Profiles, connection *Connection, feed *Daemon
 	return &ProfileSwitcher{profiles: profiles, connection: connection, feed: feed}
 }
 
-// SwitchActive switches to the named profile applying the reconnect policy.
+// SwitchActive switches to the named profile and always connects afterwards.
 func (s *ProfileSwitcher) SwitchActive(ctx context.Context, p ProfileRef) error {
+	return s.switchActive(ctx, p, true)
+}
+
+// SwitchActiveNoConnect switches to the named profile without connecting,
+// tearing down any existing connection first.
+func (s *ProfileSwitcher) SwitchActiveNoConnect(ctx context.Context, p ProfileRef) error {
+	return s.switchActive(ctx, p, false)
+}
+
+func (s *ProfileSwitcher) switchActive(ctx context.Context, p ProfileRef, connect bool) error {
 	prevStatus := ""
-	if st, err := s.feed.Get(ctx); err == nil {
-		prevStatus = st.Status
-	} else {
-		log.Warnf("profileswitcher: get status: %v", err)
+	if s.feed != nil {
+		if st, err := s.feed.Get(ctx); err == nil {
+			prevStatus = st.Status
+		} else {
+			log.Warnf("profileswitcher: get status: %v", err)
+		}
 	}
 
-	wasActive := strings.EqualFold(prevStatus, StatusConnected) ||
-		strings.EqualFold(prevStatus, StatusConnecting)
-	needsDown := wasActive ||
+	needsDown := strings.EqualFold(prevStatus, StatusConnected) ||
+		strings.EqualFold(prevStatus, StatusConnecting) ||
 		strings.EqualFold(prevStatus, StatusNeedsLogin) ||
 		strings.EqualFold(prevStatus, StatusLoginFailed) ||
 		strings.EqualFold(prevStatus, StatusSessionExpired)
 
-	log.Infof("profileswitcher: switch profile=%q prevStatus=%q wasActive=%v needsDown=%v",
-		p.ProfileName, prevStatus, wasActive, needsDown)
+	log.Infof("profileswitcher: switch profile=%q prevStatus=%q connect=%v needsDown=%v",
+		p.ProfileName, prevStatus, connect, needsDown)
 
-	// Optimistic Connecting paint only when wasActive: those prevStatuses emit
-	// stale Connected + transient Idle pushes during Down that must be
-	// suppressed until Up resumes the stream (see DaemonFeed suppression table).
-	if wasActive {
+	// Optimistic Connecting paint plus stale-push suppression during Down (see
+	// DaemonFeed suppression table); also arms the login-watch that pops
+	// browser-login when the new profile turns out to need SSO.
+	if connect && s.feed != nil {
 		s.feed.BeginProfileSwitch()
 	}
 
@@ -76,9 +89,9 @@ func (s *ProfileSwitcher) SwitchActive(ctx context.Context, p ProfileRef) error
 		}
 	}
 
-	if wasActive {
+	if connect {
 		if err := s.connection.Up(ctx, UpParams(p)); err != nil {
-			return fmt.Errorf("reconnect %q: %w", p.ProfileName, err)
+			return fmt.Errorf("connect %q: %w", p.ProfileName, err)
 		}
 	}
 

client/ui/services/windowmanager.go

@@ -185,37 +185,38 @@ func (s *WindowManager) OpenBrowserLogin(uri string) {
 			startURL = "/#/dialog/browser-login?uri=" + url.QueryEscape(uri)
 		}
 		s.hideOtherWindowsLocked("browser-login")
-		// Prefer the main window's screen (multi-monitor); falls back to OS-default centering.
-		var screen *application.Screen
-		if s.mainWindow != nil {
-			if sc, err := s.mainWindow.GetScreen(); err == nil {
-				screen = sc
-			}
-		}
 		opts := DialogWindowOptions("browser-login", s.title("window.title.signIn"), startURL, s.linuxIcon)
 		// Not always-on-top: it would obscure the browser tab the user logs in through.
 		opts.AlwaysOnTop = false
 		opts.InitialPosition = application.WindowCentered
-		opts.Screen = screen
+		// Open on the active (where users cursor is) display, like the session-expiration dialog.
+		opts.Screen = s.getScreenBasedOnCursorPosition()
 		s.browserLogin = s.app.Window.NewWithOptions(opts)
 		bl := s.browserLogin
-		// Red-X close means cancel: emit the event so startLogin() tears down the SSO wait.
 		bl.OnWindowEvent(events.Common.WindowClosing, func(_ *application.WindowEvent) {
-			s.app.Event.Emit(EventBrowserLoginCancel)
 			s.mu.Lock()
-			s.browserLogin = nil
-			s.restoreHiddenWindowsLocked()
+			// Only a live user red-X still has this registered; programmatic closers
+			// nil s.browserLogin first and clean up themselves. Guarding here stops a
+			// stale close event from wiping a replacement popup's state.
+			userClosed := s.browserLogin == bl
+			if userClosed {
+				s.browserLogin = nil
+				s.restoreHiddenWindowsLocked()
+			}
 			s.mu.Unlock()
+			if userClosed {
+				s.app.Event.Emit(EventBrowserLoginCancel)
+			}
 		})
-		s.centerWhenReady(s.browserLogin)
+		s.centerOnCursorScreen(s.browserLogin)
 		return
 	}
 	if uri != "" {
 		s.browserLogin.SetURL("/#/dialog/browser-login?uri=" + url.QueryEscape(uri))
 	}
+	s.centerOnCursorScreen(s.browserLogin)
 	s.browserLogin.Show()
 	s.browserLogin.Focus()
-	s.centerWhenReady(s.browserLogin)
 }
 
 // BrowserLoginWindow returns the live SSO popup, or nil. While non-nil it is the
@@ -238,6 +239,15 @@ func (s *WindowManager) CloseBrowserLogin() {
 	s.mu.Lock()
 	w := s.browserLogin
 	s.browserLogin = nil
+	// The WindowClosing hook no-ops on a programmatic close, so restore here —
+	// but only if a popup was actually open. The frontend calls this even when no
+	// popup was ever shown (e.g. resetDialog() after an early RequestExtend failure,
+	// or connection.ts's catch path), and hiddenForLogin is shared with
+	// OpenInstallProgress, so an unconditional restore could re-show windows a
+	// still-running install-progress is hiding.
+	if w != nil {
+		s.restoreHiddenWindowsLocked()
+	}
 	s.mu.Unlock()
 	if w != nil {
 		w.Close()
@@ -279,6 +289,35 @@ func (s *WindowManager) CloseSessionExpiration() {
 	}
 }
 
+// CloseRenewFlow tears down the SSO session-renewal UI in a single call: it
+// closes the browser-login popup and the session-expiration window together.
+func (s *WindowManager) CloseRenewFlow() {
+	s.mu.Lock()
+	bl := s.browserLogin
+	se := s.sessionExpiration
+	s.browserLogin = nil
+	s.sessionExpiration = nil
+	if se != nil {
+		kept := s.hiddenForLogin[:0]
+		for _, w := range s.hiddenForLogin {
+			if w != se {
+				kept = append(kept, w)
+			}
+		}
+		s.hiddenForLogin = kept
+	}
+	s.restoreHiddenWindowsLocked()
+	s.mu.Unlock()
+
+	// Close after unlock so the re-entrant handlers can take s.mu.
+	if bl != nil {
+		bl.Close()
+	}
+	if se != nil {
+		se.Close()
+	}
+}
+
 // OpenInstallProgress shows the install-progress window and hides the rest for the duration
 // (restored on close). It owns its own result polling since the daemon restarts mid-install.
 func (s *WindowManager) OpenInstallProgress(version string) {

client/ui/tray.go

@@ -30,6 +30,8 @@ const (
 
 	statusError = "Error"
 
+	quitDownTimeout = 5 * time.Second
+
 	urlGitHubRepo     = "https://github.com/netbirdio/netbird"
 	urlGitHubReleases = "https://github.com/netbirdio/netbird/releases/latest"
 	urlDocs           = "https://docs.netbird.io"
@@ -315,8 +317,7 @@ func (t *Tray) relayoutMenu() {
 		if sessionDeadline.IsZero() {
 			t.sessionExpiresItem.SetHidden(true)
 		} else {
-			remaining := t.formatSessionRemaining(time.Until(sessionDeadline))
-			t.sessionExpiresItem.SetLabel(t.loc.T("tray.session.expiresIn", "remaining", remaining))
+			t.sessionExpiresItem.SetLabel(t.sessionRowLabel(sessionDeadline))
 			t.sessionExpiresItem.SetHidden(false)
 		}
 	}
@@ -446,11 +447,28 @@ func (t *Tray) buildMenu() *application.Menu {
 	menu.AddSeparator()
 	menu.Add(t.loc.T("tray.menu.quit")).
 		SetAccelerator("CmdOrCtrl+Q").
-		OnClick(func(*application.Context) { t.app.Quit() })
+		OnClick(func(*application.Context) { t.handleQuit() })
 
 	return menu
 }
 
+func (t *Tray) handleQuit() {
+	t.profileMu.Lock()
+	if t.switchCancel != nil {
+		t.switchCancel()
+		t.switchCancel = nil
+	}
+	t.profileMu.Unlock()
+	t.svc.DaemonFeed.CancelProfileSwitch()
+
+	ctx, cancel := context.WithTimeout(context.Background(), quitDownTimeout)
+	defer cancel()
+	if err := t.svc.Connection.Down(ctx); err != nil {
+		log.Errorf("disconnect on quit: %v", err)
+	}
+	t.app.Quit()
+}
+
 // handleConnect receives the clicked item from the buildMenu closure —
 // t.upItem is menuMu-guarded and must not be read here.
 func (t *Tray) handleConnect(upItem *application.MenuItem) {

client/ui/tray_session.go

@@ -64,11 +64,42 @@ func (t *Tray) applySessionExpiry(deadline *time.Time, connected bool) bool {
 	return changed
 }
 
-// runSessionExpiryTicker recomputes the "Expires in …" row label every 30s. Runs until process exit.
+// runSessionExpiryTicker recomputes the "Expires in …" row label until process exit.
+// The interval scales with the remaining time: coarse when the deadline is far off,
+// down to 10s in the final two minutes so the label doesn't lag the ceiling-rounded
+// countdown near expiry. The cached deadline is re-read every iteration, so an extend
+// or reconnect that moves it is picked up on the next tick.
 func (t *Tray) runSessionExpiryTicker() {
-	tk := time.NewTicker(30 * time.Second)
-	for range tk.C {
+	tm := time.NewTimer(sessionRefreshInterval(t.sessionRemaining()))
+	defer tm.Stop()
+	for range tm.C {
 		t.refreshSessionExpiresLabel()
+		tm.Reset(sessionRefreshInterval(t.sessionRemaining()))
+	}
+}
+
+// sessionRemaining returns the time left on the cached SSO deadline, or 0 when unknown.
+func (t *Tray) sessionRemaining() time.Duration {
+	t.sessionMu.Lock()
+	deadline := t.sessionExpiresAt
+	t.sessionMu.Unlock()
+	if deadline.IsZero() {
+		return 0
+	}
+	return time.Until(deadline)
+}
+
+// sessionRefreshInterval picks how long to wait before the next label recompute.
+func sessionRefreshInterval(remaining time.Duration) time.Duration {
+	switch {
+	case remaining <= 0:
+		return 30 * time.Second
+	case remaining <= 2*time.Minute:
+		return 10 * time.Second
+	case remaining <= time.Hour:
+		return 30 * time.Second
+	default:
+		return time.Minute
 	}
 }
 
@@ -87,30 +118,39 @@ func (t *Tray) refreshSessionExpiresLabel() {
 	if deadline.IsZero() {
 		return
 	}
-	remaining := t.formatSessionRemaining(time.Until(deadline))
-	item.SetLabel(t.loc.T("tray.session.expiresIn", "remaining", remaining))
+	item.SetLabel(t.sessionRowLabel(deadline))
+}
+
+func (t *Tray) sessionRowLabel(deadline time.Time) string {
+	remaining := time.Until(deadline)
+	if remaining <= 0 {
+		return t.loc.T("tray.status.sessionExpired")
+	}
+	return t.loc.T("tray.session.expiresIn", "remaining", t.formatSessionRemaining(remaining))
 }
 
 // formatSessionRemaining renders d as a localised long-form string picking the largest non-zero unit.
+// Each unit is rounded up so the label never claims less time than actually remains, matching the
+// upper-bound sense of the sub-minute "less than a minute" fragment.
 // Singular/plural keys are split per language for proper translation.
 func (t *Tray) formatSessionRemaining(d time.Duration) string {
 	switch {
 	case d < time.Minute:
 		return t.loc.T("tray.session.unit.lessThanMinute")
-	case d < time.Hour:
-		m := int(d / time.Minute)
+	case d <= 59*time.Minute:
+		m := ceilDiv(d, time.Minute)
 		if m == 1 {
 			return t.loc.T("tray.session.unit.minute")
 		}
 		return t.loc.T("tray.session.unit.minutes", "count", strconv.Itoa(m))
-	case d < 24*time.Hour:
-		h := int((d + 30*time.Minute) / time.Hour)
+	case d <= 23*time.Hour:
+		h := ceilDiv(d, time.Hour)
 		if h == 1 {
 			return t.loc.T("tray.session.unit.hour")
 		}
 		return t.loc.T("tray.session.unit.hours", "count", strconv.Itoa(h))
 	default:
-		days := int((d + 12*time.Hour) / (24 * time.Hour))
+		days := ceilDiv(d, 24*time.Hour)
 		if days == 1 {
 			return t.loc.T("tray.session.unit.day")
 		}
@@ -118,6 +158,11 @@ func (t *Tray) formatSessionRemaining(d time.Duration) string {
 	}
 }
 
+// ceilDiv divides d by unit rounding up, assuming d > 0.
+func ceilDiv(d, unit time.Duration) int {
+	return int((d + unit - time.Nanosecond) / unit)
+}
+
 // registerSessionWarningCategory wires the OS notification category and response handler for the expiry warning.
 // Errors are swallowed since the worst case is a plain notification without buttons.
 func (t *Tray) registerSessionWarningCategory() {
@@ -252,11 +297,9 @@ func (t *Tray) openSessionExpiration() {
 }
 
 // openSessionExtendFlow opens the SessionExpiration window seeded with the cached deadline's remaining time,
-// for the "Expires in …" tray row. No-ops when the deadline is unknown or elapsed.
+// for the "Expires in …" tray row. Once the deadline has elapsed the row reads "Session expired" and the
+// click routes to the login flow instead. No-op when the deadline is unknown.
 func (t *Tray) openSessionExtendFlow() {
-	if t.svc.WindowManager == nil {
-		return
-	}
 	t.sessionMu.Lock()
 	deadline := t.sessionExpiresAt
 	t.sessionMu.Unlock()
@@ -265,6 +308,14 @@ func (t *Tray) openSessionExtendFlow() {
 	}
 	seconds := int(time.Until(deadline).Seconds())
 	if seconds <= 0 {
+		if t.window != nil {
+			t.window.SetURL("/#/login")
+			t.window.Show()
+			t.window.Focus()
+		}
+		return
+	}
+	if t.svc.WindowManager == nil {
 		return
 	}
 	t.svc.WindowManager.OpenSessionExpiration(seconds)

combined/cmd/admin.go

@@ -0,0 +1,151 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/dexidp/dex/storage"
+	log "github.com/sirupsen/logrus"
+	"github.com/spf13/cobra"
+
+	"github.com/netbirdio/netbird/formatter/hook"
+	admincmd "github.com/netbirdio/netbird/management/cmd/admin"
+	tokencmd "github.com/netbirdio/netbird/management/cmd/token"
+	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+	"github.com/netbirdio/netbird/management/server/activity"
+	activitystore "github.com/netbirdio/netbird/management/server/activity/store"
+	"github.com/netbirdio/netbird/management/server/store"
+	"github.com/netbirdio/netbird/management/server/types"
+	"github.com/netbirdio/netbird/util"
+)
+
+// newAdminCommands creates the admin command tree with combined-specific resource openers.
+func newAdminCommands() *cobra.Command {
+	return admincmd.NewCommands(admincmd.Openers{
+		Resources: withAdminResources,
+		Store:     withAdminStoreOnly,
+		IDP:       withAdminIDPOnly,
+	})
+}
+
+func newLegacyTokenCommand() *cobra.Command {
+	cmd := tokencmd.NewCommands(tokencmd.StoreOpener(withAdminStoreOnly))
+	cmd.Deprecated = "use 'admin token' instead"
+	return cmd
+}
+
+// withAdminResources loads the combined YAML config, initializes stores, and calls fn.
+func withAdminResources(cmd *cobra.Command, fn func(ctx context.Context, resources admincmd.Resources) error) error {
+	return withAdminConfig(cmd, func(ctx context.Context, cfg *CombinedConfig) error {
+		mgmtConfig, err := adminManagementConfig(cfg)
+		if err != nil {
+			return err
+		}
+
+		managementStore, err := openAdminStore(ctx, cfg)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseStore(ctx, managementStore)
+
+		idpStorage, idpStorageFile, err := admincmd.OpenIDPStorage(mgmtConfig)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseIDPStorage(idpStorage)
+
+		eventStore, esErr := openAdminEventStore(ctx, cfg, mgmtConfig)
+		if esErr != nil {
+			_, _ = fmt.Fprintf(cmd.ErrOrStderr(), "Warning: audit events will not be recorded: %v\n", esErr)
+		}
+		if eventStore != nil {
+			defer func() {
+				if err := eventStore.Close(ctx); err != nil {
+					log.Debugf("close activity event store: %v", err)
+				}
+			}()
+		}
+
+		return fn(ctx, admincmd.Resources{Store: managementStore, IDPStorage: idpStorage, IDPStorageFile: idpStorageFile, EventStore: eventStore})
+	})
+}
+
+// withAdminStoreOnly opens only the management store for admin subcommands that do not
+// need embedded IdP storage.
+func withAdminStoreOnly(cmd *cobra.Command, fn func(ctx context.Context, s store.Store) error) error {
+	return withAdminConfig(cmd, func(ctx context.Context, cfg *CombinedConfig) error {
+		managementStore, err := openAdminStore(ctx, cfg)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseStore(ctx, managementStore)
+
+		return fn(ctx, managementStore)
+	})
+}
+
+func withAdminIDPOnly(cmd *cobra.Command, fn func(ctx context.Context, idpStorage storage.Storage, storageFile string) error) error {
+	return withAdminConfig(cmd, func(ctx context.Context, cfg *CombinedConfig) error {
+		mgmtConfig, err := adminManagementConfig(cfg)
+		if err != nil {
+			return err
+		}
+		idpStorage, idpStorageFile, err := admincmd.OpenIDPStorage(mgmtConfig)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseIDPStorage(idpStorage)
+
+		return fn(ctx, idpStorage, idpStorageFile)
+	})
+}
+
+func withAdminConfig(cmd *cobra.Command, fn func(ctx context.Context, cfg *CombinedConfig) error) error {
+	if err := util.InitLog("error", "console"); err != nil {
+		return fmt.Errorf("init log: %w", err)
+	}
+
+	ctx := context.WithValue(cmd.Context(), hook.ExecutionContextKey, hook.SystemSource) //nolint:staticcheck
+
+	cfg, err := LoadConfig(configPath)
+	if err != nil {
+		return fmt.Errorf("load config: %w", err)
+	}
+	cfg.ApplyAdminDefaults()
+	applyServerStoreEnv(cfg.Server.Store)
+
+	return fn(ctx, cfg)
+}
+
+func adminManagementConfig(cfg *CombinedConfig) (*nbconfig.Config, error) {
+	mgmtConfig, err := cfg.ToManagementConfig()
+	if err != nil {
+		return nil, fmt.Errorf("create management config: %w", err)
+	}
+	return mgmtConfig, nil
+}
+
+func openAdminStore(ctx context.Context, cfg *CombinedConfig) (store.Store, error) {
+	managementStore, err := store.NewStore(ctx, types.Engine(cfg.Management.Store.Engine), cfg.Management.DataDir, nil, true)
+	if err != nil {
+		return nil, fmt.Errorf("create store: %w", err)
+	}
+	return managementStore, nil
+}
+
+func openAdminEventStore(ctx context.Context, cfg *CombinedConfig, config *nbconfig.Config) (activity.Store, error) {
+	if config.DataStoreEncryptionKey == "" {
+		return nil, fmt.Errorf("data store encryption key is not configured")
+	}
+	if err := applyActivityStoreEnv(cfg.Server.ActivityStore); err != nil {
+		return nil, fmt.Errorf("configure activity event store: %w", err)
+	}
+	eventStore, err := activitystore.NewSqlStore(ctx, config.Datadir, config.DataStoreEncryptionKey)
+	if err != nil {
+		return nil, fmt.Errorf("open activity event store: %w", err)
+	}
+	if eventStore == nil {
+		return nil, fmt.Errorf("open activity event store: returned nil store")
+	}
+	return eventStore, nil
+}

combined/cmd/admin_config_test.go

@@ -0,0 +1,47 @@
+package cmd
+
+import (
+	"context"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+
+	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+)
+
+func TestApplyAdminDefaultsCopiesServerStoreWithoutExposedAddress(t *testing.T) {
+	cfg := DefaultConfig()
+	cfg.Server.ExposedAddress = ""
+	cfg.Server.DataDir = "/srv/netbird"
+	cfg.Server.Store = StoreConfig{
+		Engine: "postgres",
+		DSN:    "postgres://user:pass@example.com/netbird",
+	}
+
+	cfg.ApplyAdminDefaults()
+
+	require.Equal(t, "/srv/netbird", cfg.Management.DataDir)
+	require.Equal(t, "postgres", cfg.Management.Store.Engine)
+	require.Equal(t, cfg.Server.Store.DSN, cfg.Management.Store.DSN)
+}
+
+func TestOpenAdminEventStoreMissingEncryptionKeyReturnsNilInterface(t *testing.T) {
+	eventStore, err := openAdminEventStore(context.Background(), &CombinedConfig{}, &nbconfig.Config{})
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "encryption key")
+	require.Nil(t, eventStore)
+}
+
+func TestApplyServerStoreEnv(t *testing.T) {
+	t.Setenv("NB_STORE_ENGINE_POSTGRES_DSN", "")
+	t.Setenv("NB_STORE_ENGINE_MYSQL_DSN", "")
+	t.Setenv("NB_STORE_ENGINE_SQLITE_FILE", "")
+
+	applyServerStoreEnv(StoreConfig{Engine: "postgres", DSN: "postgres-dsn", File: "store.db"})
+	require.Equal(t, "postgres-dsn", os.Getenv("NB_STORE_ENGINE_POSTGRES_DSN"))
+	require.Equal(t, "store.db", os.Getenv("NB_STORE_ENGINE_SQLITE_FILE"))
+
+	applyServerStoreEnv(StoreConfig{Engine: "mysql", DSN: "mysql-dsn"})
+	require.Equal(t, "mysql-dsn", os.Getenv("NB_STORE_ENGINE_MYSQL_DSN"))
+}

combined/cmd/config.go

@@ -6,8 +6,7 @@ import (
 	"net"
 	"net/netip"
 	"os"
-	"path"
-	"path/filepath"
+	filePath "path/filepath"
 	"strings"
 	"time"
 
@@ -74,6 +73,9 @@ type ServerConfig struct {
 	ActivityStore           StoreConfig        `yaml:"activityStore"`
 	AuthStore               StoreConfig        `yaml:"authStore"`
 	ReverseProxy            ReverseProxyConfig `yaml:"reverseProxy"`
+
+	SupportedSyncMessageVersions           *int           `yaml:"supportedSyncMessageVersions,omitempty"`
+	PerAccountSupportedSyncMessageVersions map[string]int `yaml:"perAccountSupportedSyncMessageVersions,omitempty"`
 }
 
 // TLSConfig contains TLS/HTTPS settings
@@ -300,6 +302,19 @@ func (c *CombinedConfig) ApplySimplifiedDefaults() {
 	c.autoConfigureClientSettings(exposedProto, exposedHost, exposedHostPort, hasExternalStuns, hasExternalRelay, hasExternalSignal)
 }
 
+// ApplyAdminDefaults applies the management settings needed by admin commands even
+// when the full server config is invalid and ApplySimplifiedDefaults cannot run.
+func (c *CombinedConfig) ApplyAdminDefaults() {
+	if c.Management.DataDir == "" || c.Management.DataDir == "/var/lib/netbird/" {
+		c.Management.DataDir = c.Server.DataDir
+	}
+	if c.Management.Store.Engine == "" || c.Management.Store.Engine == "sqlite" {
+		if c.Server.Store.Engine != "" || c.Server.Store.File != "" || c.Server.Store.DSN != "" {
+			c.Management.Store = c.Server.Store
+		}
+	}
+}
+
 // applyRelayDefaults configures the relay service if no external relay is configured.
 func (c *CombinedConfig) applyRelayDefaults(exposedProto, exposedHostPort string, hasExternalRelay, hasExternalStuns bool) {
 	if hasExternalRelay {
@@ -577,11 +592,11 @@ func (c *CombinedConfig) buildEmbeddedIdPConfig(mgmt ManagementConfig) (*idp.Emb
 			return nil, fmt.Errorf("authStore.dsn is required when authStore.engine is postgres")
 		}
 	} else {
-		authStorageFile = path.Join(mgmt.DataDir, "idp.db")
+		authStorageFile = filePath.Join(mgmt.DataDir, "idp.db")
 		if c.Server.AuthStore.File != "" {
 			authStorageFile = c.Server.AuthStore.File
-			if !filepath.IsAbs(authStorageFile) {
-				authStorageFile = filepath.Join(mgmt.DataDir, authStorageFile)
+			if !filePath.IsAbs(authStorageFile) {
+				authStorageFile = filePath.Join(mgmt.DataDir, authStorageFile)
 			}
 		}
 	}
@@ -696,16 +711,18 @@ func (c *CombinedConfig) ToManagementConfig() (*nbconfig.Config, error) {
 	httpConfig.AuthCallbackURL = callbackURL + types.ProxyCallbackEndpointFull
 
 	return &nbconfig.Config{
-		Stuns:                  stuns,
-		Relay:                  relayConfig,
-		Signal:                 signalConfig,
-		Datadir:                mgmt.DataDir,
-		DataStoreEncryptionKey: mgmt.Store.EncryptionKey,
-		HttpConfig:             httpConfig,
-		StoreConfig:            storeConfig,
-		ReverseProxy:           reverseProxy,
-		DisableDefaultPolicy:   mgmt.DisableDefaultPolicy,
-		EmbeddedIdP:            embeddedIdP,
+		Stuns:                              stuns,
+		Relay:                              relayConfig,
+		Signal:                             signalConfig,
+		Datadir:                            mgmt.DataDir,
+		DataStoreEncryptionKey:             mgmt.Store.EncryptionKey,
+		HttpConfig:                         httpConfig,
+		StoreConfig:                        storeConfig,
+		ReverseProxy:                       reverseProxy,
+		DisableDefaultPolicy:               mgmt.DisableDefaultPolicy,
+		EmbeddedIdP:                        embeddedIdP,
+		HighestSupportedSyncMessageVersion: c.Server.SupportedSyncMessageVersions,
+		PerAccountHighestSupportedSyncMessageVersion: c.Server.PerAccountSupportedSyncMessageVersions,
 	}, nil
 }
 
@@ -729,7 +746,7 @@ func ApplyEmbeddedIdPConfig(ctx context.Context, cfg *nbconfig.Config, mgmtPort
 		cfg.EmbeddedIdP.Storage.Type = "sqlite3"
 	}
 	if cfg.EmbeddedIdP.Storage.Config.File == "" && cfg.Datadir != "" {
-		cfg.EmbeddedIdP.Storage.Config.File = path.Join(cfg.Datadir, "idp.db")
+		cfg.EmbeddedIdP.Storage.Config.File = filePath.Join(cfg.Datadir, "idp.db")
 	}
 
 	issuer := cfg.EmbeddedIdP.Issuer

combined/cmd/root.go

@@ -31,6 +31,7 @@ import (
 	relayServer "github.com/netbirdio/netbird/relay/server"
 	"github.com/netbirdio/netbird/relay/server/listener"
 	"github.com/netbirdio/netbird/relay/server/listener/ws"
+	syncgrpc "github.com/netbirdio/netbird/shared/management/grpc"
 	sharedMetrics "github.com/netbirdio/netbird/shared/metrics"
 	"github.com/netbirdio/netbird/shared/relay/auth"
 	"github.com/netbirdio/netbird/shared/signal/proto"
@@ -64,7 +65,8 @@ func init() {
 	rootCmd.PersistentFlags().StringVarP(&configPath, "config", "c", "", "path to YAML configuration file (required)")
 	_ = rootCmd.MarkPersistentFlagRequired("config")
 
-	rootCmd.AddCommand(newTokenCommands())
+	rootCmd.AddCommand(newAdminCommands())
+	rootCmd.AddCommand(newLegacyTokenCommand())
 }
 
 func RootCmd() *cobra.Command {
@@ -122,6 +124,37 @@ func execute(cmd *cobra.Command, _ []string) error {
 }
 
 // initializeConfig loads and validates the configuration, then initializes logging.
+func applyServerStoreEnv(storeConfig StoreConfig) {
+	if dsn := storeConfig.DSN; dsn != "" {
+		switch strings.ToLower(storeConfig.Engine) {
+		case "postgres":
+			os.Setenv("NB_STORE_ENGINE_POSTGRES_DSN", dsn)
+		case "mysql":
+			os.Setenv("NB_STORE_ENGINE_MYSQL_DSN", dsn)
+		}
+	}
+	if file := storeConfig.File; file != "" {
+		os.Setenv("NB_STORE_ENGINE_SQLITE_FILE", file)
+	}
+}
+
+func applyActivityStoreEnv(storeConfig StoreConfig) error {
+	if engine := storeConfig.Engine; engine != "" {
+		engineLower := strings.ToLower(engine)
+		if engineLower == "postgres" && storeConfig.DSN == "" {
+			return fmt.Errorf("activityStore.dsn is required when activityStore.engine is postgres")
+		}
+		os.Setenv("NB_ACTIVITY_EVENT_STORE_ENGINE", engineLower)
+		if dsn := storeConfig.DSN; dsn != "" {
+			os.Setenv("NB_ACTIVITY_EVENT_POSTGRES_DSN", dsn)
+		}
+	}
+	if file := storeConfig.File; file != "" {
+		os.Setenv("NB_ACTIVITY_EVENT_SQLITE_FILE", file)
+	}
+	return nil
+}
+
 func initializeConfig() error {
 	var err error
 	config, err = LoadConfig(configPath)
@@ -137,30 +170,10 @@ func initializeConfig() error {
 		return fmt.Errorf("failed to initialize log: %w", err)
 	}
 
-	if dsn := config.Server.Store.DSN; dsn != "" {
-		switch strings.ToLower(config.Server.Store.Engine) {
-		case "postgres":
-			os.Setenv("NB_STORE_ENGINE_POSTGRES_DSN", dsn)
-		case "mysql":
-			os.Setenv("NB_STORE_ENGINE_MYSQL_DSN", dsn)
-		}
-	}
-	if file := config.Server.Store.File; file != "" {
-		os.Setenv("NB_STORE_ENGINE_SQLITE_FILE", file)
-	}
+	applyServerStoreEnv(config.Server.Store)
 
-	if engine := config.Server.ActivityStore.Engine; engine != "" {
-		engineLower := strings.ToLower(engine)
-		if engineLower == "postgres" && config.Server.ActivityStore.DSN == "" {
-			return fmt.Errorf("activityStore.dsn is required when activityStore.engine is postgres")
-		}
-		os.Setenv("NB_ACTIVITY_EVENT_STORE_ENGINE", engineLower)
-		if dsn := config.Server.ActivityStore.DSN; dsn != "" {
-			os.Setenv("NB_ACTIVITY_EVENT_POSTGRES_DSN", dsn)
-		}
-	}
-	if file := config.Server.ActivityStore.File; file != "" {
-		os.Setenv("NB_ACTIVITY_EVENT_SQLITE_FILE", file)
+	if err := applyActivityStoreEnv(config.Server.ActivityStore); err != nil {
+		return err
 	}
 
 	log.Infof("Starting combined NetBird server")
@@ -505,6 +518,16 @@ func createManagementServer(cfg *CombinedConfig, mgmtConfig *nbconfig.Config) (m
 	}
 	mgmtPort, _ := strconv.Atoi(portStr)
 
+	if err := syncgrpc.ValidateSyncMessageVersion(mgmtConfig.HighestSupportedSyncMessageVersion); err != nil {
+		return nil, err
+	}
+
+	for accountId, version := range mgmtConfig.PerAccountHighestSupportedSyncMessageVersion {
+		if err := syncgrpc.ValidateSyncMessageVersion(&version); err != nil {
+			return nil, fmt.Errorf("unrecognized sync message version in perAccountSupportedSyncMessageVersions for account %s %w", accountId, err)
+		}
+	}
+
 	mgmtSrv := newServer(
 		&mgmtServer.Config{
 			NbConfig:                mgmtConfig,

combined/cmd/token.go

@@ -1,63 +0,0 @@
-package cmd
-
-import (
-	"context"
-	"fmt"
-	"os"
-	"strings"
-
-	log "github.com/sirupsen/logrus"
-	"github.com/spf13/cobra"
-
-	"github.com/netbirdio/netbird/formatter/hook"
-	tokencmd "github.com/netbirdio/netbird/management/cmd/token"
-	"github.com/netbirdio/netbird/management/server/store"
-	"github.com/netbirdio/netbird/management/server/types"
-	"github.com/netbirdio/netbird/util"
-)
-
-// newTokenCommands creates the token command tree with combined-specific store opener.
-func newTokenCommands() *cobra.Command {
-	return tokencmd.NewCommands(withTokenStore)
-}
-
-// withTokenStore loads the combined YAML config, initializes the store, and calls fn.
-func withTokenStore(cmd *cobra.Command, fn func(ctx context.Context, s store.Store) error) error {
-	if err := util.InitLog("error", "console"); err != nil {
-		return fmt.Errorf("init log: %w", err)
-	}
-
-	ctx := context.WithValue(cmd.Context(), hook.ExecutionContextKey, hook.SystemSource) //nolint:staticcheck
-
-	cfg, err := LoadConfig(configPath)
-	if err != nil {
-		return fmt.Errorf("load config: %w", err)
-	}
-
-	if dsn := cfg.Server.Store.DSN; dsn != "" {
-		switch strings.ToLower(cfg.Server.Store.Engine) {
-		case "postgres":
-			os.Setenv("NB_STORE_ENGINE_POSTGRES_DSN", dsn)
-		case "mysql":
-			os.Setenv("NB_STORE_ENGINE_MYSQL_DSN", dsn)
-		}
-	}
-	if file := cfg.Server.Store.File; file != "" {
-		os.Setenv("NB_STORE_ENGINE_SQLITE_FILE", file)
-	}
-
-	datadir := cfg.Management.DataDir
-	engine := types.Engine(cfg.Management.Store.Engine)
-
-	s, err := store.NewStore(ctx, engine, datadir, nil, true)
-	if err != nil {
-		return fmt.Errorf("create store: %w", err)
-	}
-	defer func() {
-		if err := s.Close(ctx); err != nil {
-			log.Debugf("close store: %v", err)
-		}
-	}()
-
-	return fn(ctx, s)
-}

dns/nameserver.go

@@ -53,6 +53,7 @@ type NameServerGroup struct {
 	ID string `gorm:"primaryKey"`
 	// AccountID is a reference to Account that this object belongs
 	AccountID string `gorm:"index"`
+	PublicID  string `json:"-"`
 	// Name group name
 	Name string
 	// Description group description

docs/io.netbird.client.plist

@@ -66,6 +66,9 @@
     <key>disableAutoConnect</key>
     <false/>
 
+    <key>disableAutostart</key>
+    <false/>
+
     <key>disableClientRoutes</key>
     <false/>
 

docs/netbird-macos.mobileconfig

@@ -103,6 +103,8 @@
                                 <!--
                                 <key>disableAutoConnect</key>
                                 <false/>
+                                <key>disableAutostart</key>
+                                <false/>
                                 <key>disableClientRoutes</key>
                                 <false/>
                                 <key>disableServerRoutes</key>

docs/netbird-macos.sh

@@ -58,6 +58,7 @@ preSharedKey="$NULL"                       # secret; redacted in log
 allowServerSSH='true'
 blockInbound="$NULL"
 disableAutoConnect="$NULL"
+disableAutostart="$NULL"
 disableClientRoutes="$NULL"
 disableServerRoutes="$NULL"
 disableMetricsCollection="$NULL"
@@ -155,6 +156,7 @@ main() {
   is_set "$allowServerSSH"            && emit_bool    allowServerSSH            "$allowServerSSH"
   is_set "$blockInbound"              && emit_bool    blockInbound              "$blockInbound"
   is_set "$disableAutoConnect"        && emit_bool    disableAutoConnect        "$disableAutoConnect"
+  is_set "$disableAutostart"          && emit_bool    disableAutostart          "$disableAutostart"
   is_set "$disableClientRoutes"       && emit_bool    disableClientRoutes       "$disableClientRoutes"
   is_set "$disableServerRoutes"       && emit_bool    disableServerRoutes       "$disableServerRoutes"
   is_set "$disableMetricsCollection"  && emit_bool    disableMetricsCollection  "$disableMetricsCollection"

docs/netbird.adml

@@ -24,6 +24,9 @@
       <string id="DisableAutoConnect_Name">Disable auto-connect</string>
       <string id="DisableAutoConnect_Help">When enabled, the NetBird tunnel does not auto-connect at daemon startup. Equivalent to --disable-auto-connect.</string>
 
+      <string id="DisableAutostart_Name">Disable autostart</string>
+      <string id="DisableAutostart_Help">When enabled, the NetBird GUI is prevented from registering itself as an OS autostart entry on fresh installs, and any existing OS autostart entry registration is removed on the next GUI launch (Windows Registry Run key, macOS Login Item, Linux .desktop). Once the admin lifts the policy, the setting stays off until the user re-enables it in Settings.</string>
+
       <string id="DisableClientRoutes_Name">Disable client routes</string>
       <string id="DisableClientRoutes_Help">When enabled, this client will not consume routes advertised by routing peers. Equivalent to --disable-client-routes.</string>
 

docs/netbird.admx

@@ -64,6 +64,18 @@
       <disabledValue><decimal value="0" /></disabledValue>
     </policy>
 
+    <policy name="DisableAutostart"
+            class="Machine"
+            displayName="$(string.DisableAutostart_Name)"
+            explainText="$(string.DisableAutostart_Help)"
+            key="Software\Policies\NetBird"
+            valueName="DisableAutostart">
+      <parentCategory ref="NetBird" />
+      <supportedOn ref="SUPPORTED_NetBird_All" />
+      <enabledValue><decimal value="1" /></enabledValue>
+      <disabledValue><decimal value="0" /></disabledValue>
+    </policy>
+
     <policy name="DisableClientRoutes"
             class="Machine"
             displayName="$(string.DisableClientRoutes_Name)"

e2e/agentnetwork/chat_test.go

@@ -9,25 +9,234 @@ import (
 	"testing"
 	"time"
 
+	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	"gorm.io/driver/sqlite"
+	"gorm.io/gorm"
 
 	"github.com/netbirdio/netbird/e2e/harness"
 	"github.com/netbirdio/netbird/shared/management/http/api"
 )
 
+// per1k is a model's published USD rates per 1k tokens. read is the prompt-cache read rate
+// (OpenAI: the cached-input discount rate); write is the cache-creation rate where one exists.
+type per1k struct{ in, out, read, write float64 }
+
+// publishedPer1k hardcodes the vendors' PUBLISHED rates for the models the live matrix can drive,
+// keyed by the normalized model id the proxy stamps. Deliberately independent of the proxy's
+// pricing table so a wrong embedded rate or a broken normalization fails the run.
+var publishedPer1k = map[string]per1k{
+	"gpt-4o-mini":                 {0.00015, 0.0006, 0.000075, 0},
+	"gpt-4o":                      {0.0025, 0.01, 0.00125, 0},
+	"claude-haiku-4-5":            {0.001, 0.005, 0.0001, 0.00125},
+	"claude-sonnet-4-5":           {0.003, 0.015, 0.0003, 0.00375},
+	"claude-sonnet-4-6":           {0.003, 0.015, 0.0003, 0.00375},
+	"kimi-k3":                     {0.003, 0.015, 0.0003, 0.003}, // no published write rate: bills at the input rate
+	"anthropic.claude-haiku-4-5":  {0.001, 0.005, 0.0001, 0.00125},
+	"anthropic.claude-sonnet-4-5": {0.003, 0.015, 0.0003, 0.00375},
+	"anthropic.claude-sonnet-4-6": {0.003, 0.015, 0.0003, 0.00375},
+}
+
+// rawCostVerificationSQL is the operator-facing double-check, run straight against the management
+// sqlite store: recompute each usage row's expected total and cache cost from its own persisted
+// token buckets and hardcoded published rates. OpenAI counts cached tokens as a subset of input;
+// Anthropic-shape providers count cache buckets additively.
+const rawCostVerificationSQL = `
+WITH rates(model, in_rate, out_rate, read_rate, write_rate) AS (
+  VALUES
+    ('gpt-4o-mini',                 0.00015, 0.0006, 0.000075, 0.0),
+    ('gpt-4o',                      0.0025,  0.01,   0.00125,  0.0),
+    ('claude-haiku-4-5',            0.001,   0.005,  0.0001,   0.00125),
+    ('claude-sonnet-4-5',           0.003,   0.015,  0.0003,   0.00375),
+    ('claude-sonnet-4-6',           0.003,   0.015,  0.0003,   0.00375),
+    ('kimi-k3',                     0.003,   0.015,  0.0003,   0.003),
+    ('anthropic.claude-haiku-4-5',  0.001,   0.005,  0.0001,   0.00125),
+    ('anthropic.claude-sonnet-4-5', 0.003,   0.015,  0.0003,   0.00375),
+    ('anthropic.claude-sonnet-4-6', 0.003,   0.015,  0.0003,   0.00375)
+)
+SELECT
+  u.provider,
+  u.model,
+  u.input_tokens,
+  u.output_tokens,
+  u.cached_input_tokens,
+  u.cache_creation_tokens,
+  u.input_cost_usd,
+  u.cached_input_cost_usd,
+  u.cache_creation_cost_usd,
+  u.output_cost_usd,
+  -- No cost_usd / cache_cost_usd columns are stored: both are derived from the
+  -- four per-bucket columns above, exactly as the API renders them.
+  (u.input_cost_usd + u.cached_input_cost_usd + u.cache_creation_cost_usd + u.output_cost_usd) AS cost_usd,
+  (u.cached_input_cost_usd + u.cache_creation_cost_usd) AS cache_cost_usd,
+  CASE WHEN u.provider = 'openai' THEN
+    (u.input_tokens - MIN(u.cached_input_tokens, u.input_tokens))*r.in_rate/1000.0
+  ELSE
+    u.input_tokens*r.in_rate/1000.0
+  END AS expected_input,
+  CASE WHEN u.provider = 'openai' THEN
+    MIN(u.cached_input_tokens, u.input_tokens)*r.read_rate/1000.0
+  ELSE
+    u.cached_input_tokens*r.read_rate/1000.0
+  END AS expected_cached_input,
+  CASE WHEN u.provider = 'openai' THEN
+    0.0
+  ELSE
+    u.cache_creation_tokens*r.write_rate/1000.0
+  END AS expected_cache_creation,
+  u.output_tokens*r.out_rate/1000.0 AS expected_output,
+  CASE WHEN u.provider = 'openai' THEN
+    (u.input_tokens - MIN(u.cached_input_tokens, u.input_tokens))*r.in_rate/1000.0
+      + MIN(u.cached_input_tokens, u.input_tokens)*r.read_rate/1000.0
+      + u.output_tokens*r.out_rate/1000.0
+  ELSE
+    u.input_tokens*r.in_rate/1000.0 + u.cached_input_tokens*r.read_rate/1000.0
+      + u.cache_creation_tokens*r.write_rate/1000.0 + u.output_tokens*r.out_rate/1000.0
+  END AS expected_total,
+  CASE WHEN u.provider = 'openai' THEN
+    MIN(u.cached_input_tokens, u.input_tokens)*r.read_rate/1000.0
+  ELSE
+    u.cached_input_tokens*r.read_rate/1000.0 + u.cache_creation_tokens*r.write_rate/1000.0
+  END AS expected_cache
+FROM agent_network_request_usage u
+JOIN rates r ON r.model = u.model
+ORDER BY u.timestamp`
+
+// verifyUsageRowsSQL re-checks every persisted usage row directly in the management sqlite store,
+// bypassing the API path — the same audit an operator can run on a production store.db.
+func verifyUsageRowsSQL(t *testing.T, srv *harness.Combined) {
+	t.Helper()
+
+	dbPath, err := srv.SnapshotStoreDB(t.TempDir())
+	require.NoError(t, err, "snapshot management sqlite store")
+
+	db, err := gorm.Open(sqlite.Open(dbPath), &gorm.Config{})
+	require.NoError(t, err, "open store snapshot")
+	sqlDB, err := db.DB()
+	require.NoError(t, err)
+	defer func() { _ = sqlDB.Close() }()
+
+	rows, err := db.Raw(rawCostVerificationSQL).Rows()
+	require.NoError(t, err, "run raw cost verification query")
+	defer func() { _ = rows.Close() }()
+
+	verified := 0
+	for rows.Next() {
+		var provider, model string
+		var inTok, outTok, readTok, writeTok int64
+		var inCost, cachedInCost, cacheCreateCost, outCost, cost, cacheCost float64
+		var wantInput, wantCachedInput, wantCacheCreation, wantOutput, wantTotal, wantCache float64
+		require.NoError(t, rows.Scan(&provider, &model, &inTok, &outTok, &readTok, &writeTok,
+			&inCost, &cachedInCost, &cacheCreateCost, &outCost, &cost, &cacheCost,
+			&wantInput, &wantCachedInput, &wantCacheCreation, &wantOutput, &wantTotal, &wantCache), "scan usage row")
+		t.Logf("[sql] %s/%s: in=%d out=%d cache_read=%d cache_write=%d stored in/cached/create/out=$%.6f/$%.6f/$%.6f/$%.6f total=$%.6f cache=$%.6f expected total=$%.6f cache=$%.6f",
+			provider, model, inTok, outTok, readTok, writeTok,
+			inCost, cachedInCost, cacheCreateCost, outCost, cost, cacheCost, wantTotal, wantCache)
+		assert.InDeltaf(t, wantInput, inCost, 1e-6, "stored input_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantCachedInput, cachedInCost, 1e-6, "stored cached_input_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantCacheCreation, cacheCreateCost, 1e-6, "stored cache_creation_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantOutput, outCost, 1e-6, "stored output_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantTotal, cost, 1e-6, "derived cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantCache, cacheCost, 1e-6, "derived cache_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, inCost+cachedInCost+cacheCreateCost+outCost, cost, 1e-9,
+			"stored buckets must sum to the derived cost_usd for %s/%s", provider, model)
+		verified++
+	}
+	require.NoError(t, rows.Err(), "iterate usage rows")
+	require.Positive(t, verified, "raw SQL check must cover at least one usage row")
+	t.Logf("[sql] verified %d usage rows in store.db against published rates", verified)
+
+	gwRows, err := db.Raw(`SELECT model,
+	  (input_cost_usd + cached_input_cost_usd + cache_creation_cost_usd + output_cost_usd) AS cost_usd
+	  FROM agent_network_request_usage WHERE model LIKE '%/%'`).Rows()
+	require.NoError(t, err, "query gateway-prefixed usage rows")
+	defer func() { _ = gwRows.Close() }()
+	for gwRows.Next() {
+		var model string
+		var cost float64
+		require.NoError(t, gwRows.Scan(&model, &cost), "scan gateway usage row")
+		t.Logf("[sql] gateway %s: stored=$%.6f (must be 0 — deliberately unpriced)", model, cost)
+		assert.Zerof(t, cost, "gateway-prefixed model %q must store cost 0, never a guessed rate", model)
+	}
+	require.NoError(t, gwRows.Err(), "iterate gateway usage rows")
+}
+
+// validateAccessLogCost recomputes a live access-log row's expected total and cache cost from the
+// published per-1k rates and the row's persisted token buckets, and asserts both stored values.
+// Gateway-prefixed model ids the proxy deliberately does not price must store cost 0.
+func validateAccessLogCost(t *testing.T, pc providerCase, row api.AgentNetworkAccessLog) {
+	t.Helper()
+	model := catalogModel(pc)
+	provider := ""
+	if row.Provider != nil {
+		provider = *row.Provider
+	}
+	t.Logf("[cost] %s: provider=%s model=%s in=%d out=%d total=%d cache_read=%d cache_write=%d cost=$%.6f cache_cost=$%.6f",
+		pc.name, provider, model, row.InputTokens, row.OutputTokens, row.TotalTokens,
+		row.CachedInputTokens, row.CacheCreationTokens, row.CostUsd, row.CacheCostUsd)
+
+	rates, known := publishedPer1k[model]
+	if !known {
+		if strings.Contains(model, "/") {
+			assert.Zerof(t, row.CostUsd, "gateway-prefixed model %q is not priced so the cost meter must skip (cost 0)", model)
+			return
+		}
+		t.Logf("[cost] %s: no published rate on file for model %q (env-overridden?); skipping cost validation", pc.name, model)
+		return
+	}
+
+	// input_tokens may legitimately be 0: Moonshot/Kimi reports fully cached prompts under the cache
+	// buckets only. Output and total must always be present on a priced row.
+	require.Positive(t, row.OutputTokens, "priced row must carry output tokens")
+	require.Positive(t, row.TotalTokens, "priced row must carry total tokens")
+
+	var wantInput, wantCachedInput, wantCacheCreation float64
+	if provider == "openai" {
+		cached := min(row.CachedInputTokens, row.InputTokens) // cached is a subset of input
+		wantInput = float64(row.InputTokens-cached) / 1000 * rates.in
+		wantCachedInput = float64(cached) / 1000 * rates.read
+		// OpenAI has no cache-write bucket; wantCacheCreation stays 0.
+	} else {
+		// Anthropic / Bedrock shape: cache buckets are additive to input_tokens.
+		wantInput = float64(row.InputTokens) / 1000 * rates.in
+		wantCachedInput = float64(row.CachedInputTokens) / 1000 * rates.read
+		wantCacheCreation = float64(row.CacheCreationTokens) / 1000 * rates.write
+	}
+	wantOutput := float64(row.OutputTokens) / 1000 * rates.out
+	wantCache := wantCachedInput + wantCacheCreation
+	wantTotal := wantInput + wantCache + wantOutput
+
+	t.Logf("[cost] %s: expecting input=$%.6f cached_input=$%.6f cache_creation=$%.6f output=$%.6f total=$%.6f cache=$%.6f from published rates",
+		pc.name, wantInput, wantCachedInput, wantCacheCreation, wantOutput, wantTotal, wantCache)
+	assert.InDeltaf(t, wantInput, row.InputCostUsd, 1e-6, "stored input_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantCachedInput, row.CachedInputCostUsd, 1e-6, "stored cached_input_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantCacheCreation, row.CacheCreationCostUsd, 1e-6, "stored cache_creation_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantOutput, row.OutputCostUsd, 1e-6, "stored output_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantTotal, row.CostUsd, 1e-6, "derived cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantCache, row.CacheCostUsd, 1e-6, "derived cache_cost_usd for %s (%s)", pc.name, model)
+
+	// The aggregates must be exactly the sum of the stored components, not an
+	// independently-computed figure that could drift from the breakdown.
+	assert.InDeltaf(t, row.InputCostUsd+row.CachedInputCostUsd+row.CacheCreationCostUsd+row.OutputCostUsd,
+		row.CostUsd, 1e-9, "stored buckets must sum to the derived cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, row.CachedInputCostUsd+row.CacheCreationCostUsd,
+		row.CacheCostUsd, 1e-9, "stored cache buckets must sum to the derived cache_cost_usd for %s (%s)", pc.name, model)
+}
+
 // providerCase is one entry in the live provider matrix. The same scenario runs
 // for every available provider; availability is keyed off env vars so the suite
 // covers whatever credentials are present (source ~/.llm-keys locally / set the
 // Actions secrets in CI).
 type providerCase struct {
-	name      string
-	catalogID string
-	upstream  string
-	apiKey    string
-	model     string // body model (chat/messages) or path model@version (vertex)
-	kind      string // harness.WireChat, harness.WireMessages, or harness.WireVertex
-	project   string // vertex only: GCP project for the rawPredict path
-	region    string // vertex only: GCP region for the rawPredict path
+	name       string
+	catalogID  string
+	upstream   string
+	apiKey     string
+	model      string // body model (chat/messages) or path model@version (vertex)
+	kind       string // harness.WireChat, harness.WireMessages, or harness.WireVertex
+	project    string // vertex only: GCP project for the rawPredict path
+	region     string // vertex only: GCP region for the rawPredict path
+	pathPrefix string // base-URL path prefix the agent carries (e.g. "/anthropic" for Kimi)
 }
 
 // availableProviders builds the matrix from the provider env vars that are set.
@@ -39,6 +248,24 @@ func availableProviders() []providerCase {
 	if k := os.Getenv("ANTHROPIC_TOKEN"); k != "" {
 		ps = append(ps, providerCase{name: "anthropic", catalogID: "anthropic_api", upstream: "https://api.anthropic.com", apiKey: k, model: "claude-haiku-4-5", kind: harness.WireMessages})
 	}
+	if k := os.Getenv("KIMI_TOKEN"); k != "" {
+		// Kimi (Moonshot AI) serves two body shapes from the same key: OpenAI
+		// Chat Completions on the bare host (/v1/...) and the Anthropic
+		// Messages API under the /anthropic path prefix (the endpoint
+		// Moonshot's Claude Code guide uses). The provider keeps the bare
+		// default upstream and the AGENT carries the /anthropic prefix in
+		// its base URL — exactly the documented Claude Code / Kimi CLI
+		// setup (ANTHROPIC_BASE_URL=https://<endpoint>/anthropic) — so one
+		// provider serves both shapes and the prefix rides through to
+		// Moonshot. Run the Anthropic shape, the flagship Claude Code path;
+		// the OpenAI wire shape is covered live by the other chat-shaped
+		// matrix providers, and Kimi-over-chat passed with kimi-k3 before
+		// the single-model constraint surfaced (run #73 on the kimi feature
+		// branch). The platform serves this account exactly ONE model —
+		// kimi-k3 (kimi-k2-thinking and even kimi-latest return
+		// resource_not_found_error on both surfaces).
+		ps = append(ps, providerCase{name: "kimi", catalogID: "kimi_api", upstream: "https://api.moonshot.ai", apiKey: k, model: "kimi-k3", kind: harness.WireMessages, pathPrefix: "/anthropic"})
+	}
 	if k, u := os.Getenv("VERCEL_TOKEN"), os.Getenv("VERCEL_URL"); k != "" && u != "" {
 		ps = append(ps, providerCase{name: "vercel", catalogID: "vercel_ai_gateway", upstream: u, apiKey: k, model: "openai/gpt-4o-mini", kind: harness.WireChat})
 	}
@@ -84,14 +311,27 @@ func availableProviders() []providerCase {
 		}
 	}
 
-	// Bedrock: path-routed, bearer auth. Model is a cross-region inference
-	// profile id (distinct string from the first-party Anthropic case).
+	// Bedrock: path-routed, bearer auth. Model is the FULL cross-region
+	// inference-profile id exactly as AWS issues it — region-family prefix
+	// plus the date/version suffix. A bare or wrong-region id makes Bedrock
+	// reject the request with "The provided model identifier is invalid"
+	// before any inference runs. The proxy normalizes this id to the catalog
+	// key (anthropic.claude-haiku-4-5) for routing/pricing/allowlists.
+	// Defaults pair eu-central-1 with the eu.* profile; AWS_REGION overrides
+	// the region and the prefix follows its family.
 	if k := os.Getenv("AWS_BEARER_TOKEN_BEDROCK"); k != "" {
 		region := os.Getenv("AWS_REGION")
 		if region == "" {
-			region = "us-east-1"
+			region = "eu-central-1"
 		}
-		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: "us.anthropic.claude-haiku-4-5", kind: harness.WireBedrock})
+		// A valid Bedrock inference-profile id, overridable per account (AWS_BEDROCK_MODEL, also the
+		// workflow's bedrock_model dispatch input). `global.` profiles work from any region. Defaults to
+		// Sonnet 4.6, whose id convention dropped the -YYYYMMDD-v1:0 suffix that Haiku 4.5 still carries.
+		model := os.Getenv("AWS_BEDROCK_MODEL")
+		if model == "" {
+			model = "global.anthropic.claude-sonnet-4-6"
+		}
+		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: model, kind: harness.WireBedrock})
 	}
 	return ps
 }
@@ -108,8 +348,16 @@ func providerRequest(pc providerCase) api.AgentNetworkProviderRequest {
 		Enabled:     ptr(true),
 	}
 	if pc.kind != harness.WireVertex {
+		// The router matches the normalized catalog id. Bedrock's request model
+		// travels as a region-prefixed inference-profile id in the URL path
+		// (us.anthropic...), which the router strips before matching, so register
+		// the normalized form here or routing fails as model_not_routable.
+		modelID := pc.model
+		if pc.kind == harness.WireBedrock {
+			modelID = catalogModel(pc)
+		}
 		req.Models = &[]api.AgentNetworkProviderModel{
-			{Id: pc.model, InputPer1k: 0.001, OutputPer1k: 0.002},
+			{Id: modelID, InputPer1k: 0.001, OutputPer1k: 0.002},
 		}
 	}
 	return req
@@ -201,11 +449,12 @@ func TestProvidersMatrix(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Probe first: the GET resolves the endpoint (DNS error fails) and its first packet wakes the lazy proxy peer, so WaitProxyPeer sees it connected; any HTTP status counts.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
 	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
 		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
 	}
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
 
 	for _, pc := range matrix {
 		pc := pc
@@ -216,6 +465,10 @@ func TestProvidersMatrix(t *testing.T) {
 			// session id and confirm the marker propagated end-to-end.
 			sessionID := "e2e-session-" + pc.name
 
+			// A long-form prompt so completions carry realistic token counts for cost validation;
+			// max_tokens in the harness bodies (2048) lets the full answer through.
+			const matrixPrompt = "explain GitHub workflow in 1000 words"
+
 			// Retry briefly to absorb tunnel/DNS jitter on the first call.
 			var code int
 			var body string
@@ -226,11 +479,11 @@ func TestProvidersMatrix(t *testing.T) {
 				var cerr error
 				switch pc.kind {
 				case harness.WireVertex:
-					c, b, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model, "Reply with exactly: pong", sessionID)
+					c, b, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model, matrixPrompt, sessionID)
 				case harness.WireBedrock:
-					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, "Reply with exactly: pong", sessionID)
+					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, matrixPrompt, sessionID)
 				default:
-					c, b, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
+					c, b, cerr = cl.ChatPrefixed(ctx, settings.Endpoint, proxyIP, pc.pathPrefix, pc.kind, pc.model, matrixPrompt, sessionID)
 				}
 				if cerr == nil {
 					code, body = c, b
@@ -249,6 +502,7 @@ func TestProvidersMatrix(t *testing.T) {
 
 			// The session id sent as x-session-id must round-trip into the
 			// access-log row for this provider.
+			var row api.AgentNetworkAccessLog
 			require.Eventually(t, func() bool {
 				logs, lerr := srv.ListAccessLogs(ctx)
 				if lerr != nil {
@@ -256,11 +510,15 @@ func TestProvidersMatrix(t *testing.T) {
 				}
 				for _, r := range logs.Data {
 					if r.SessionId != nil && *r.SessionId == sessionID {
+						row = r
 						return true
 					}
 				}
 				return false
 			}, 30*time.Second, 2*time.Second, "session id %q must be recorded in an access-log row for %s", sessionID, pc.name)
+
+			// Stored total and cache cost must match the published rates applied to the row's buckets.
+			validateAccessLogCost(t, pc, row)
 		})
 	}
 
@@ -281,4 +539,7 @@ func TestProvidersMatrix(t *testing.T) {
 		}
 		return false
 	}, 60*time.Second, 3*time.Second, "consumption must be recorded with positive token counts after live traffic")
+
+	// Final raw-SQL audit: bypass the API and re-verify every persisted usage row in the store.
+	verifyUsageRowsSQL(t, srv)
 }

e2e/agentnetwork/guardrail_test.go

@@ -4,6 +4,7 @@ package agentnetwork
 
 import (
 	"context"
+	"regexp"
 	"strings"
 	"testing"
 	"time"
@@ -15,13 +16,29 @@ import (
 	"github.com/netbirdio/netbird/shared/management/http/api"
 )
 
+// bedrockRegionPrefixes and bedrockVersionSuffix mirror the proxy's Bedrock
+// model normalization (region/inference-profile prefix + version suffix) so the
+// provider is registered under the same catalog key the router matches against.
+var (
+	bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
+	bedrockVersionSuffix  = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
+)
+
 // catalogModel returns the normalized catalog id the proxy stamps for a
-// path-routed provider's configured model — the form the guardrail allowlist is
-// compared against (region prefix / @version stripped).
+// path-routed provider's configured model — the form the router and guardrail
+// allowlist compare against (Bedrock region prefix + version stripped, Vertex
+// @version stripped).
 func catalogModel(pc providerCase) string {
 	switch pc.kind {
 	case harness.WireBedrock:
-		return strings.TrimPrefix(pc.model, "us.")
+		m := pc.model
+		for _, p := range bedrockRegionPrefixes {
+			if strings.HasPrefix(m, p) {
+				m = m[len(p):]
+				break
+			}
+		}
+		return bedrockVersionSuffix.ReplaceAllString(m, "")
 	case harness.WireVertex:
 		return strings.SplitN(pc.model, "@", 2)[0]
 	default:
@@ -35,7 +52,9 @@ func catalogModel(pc providerCase) string {
 func disallowedModel(pc providerCase) string {
 	switch pc.kind {
 	case harness.WireBedrock:
-		return "us.anthropic.claude-opus-4-8"
+		// Same profile prefix as the allowed model so only the model name
+		// differs; the guardrail must deny it before it reaches AWS.
+		return strings.SplitN(pc.model, ".", 2)[0] + ".anthropic.claude-opus-4-8"
 	case harness.WireVertex:
 		return "claude-opus-4-8@20250101"
 	default:
@@ -55,7 +74,7 @@ func sendModel(ctx context.Context, t *testing.T, cl *harness.Client, endpoint,
 	case harness.WireVertex:
 		code, _, err = cl.Vertex(ctx, endpoint, proxyIP, pc.project, pc.region, model, "Reply with exactly: pong", "")
 	default:
-		code, _, err = cl.Chat(ctx, endpoint, proxyIP, pc.kind, model, "Reply with exactly: pong", "")
+		code, _, err = cl.ChatPrefixed(ctx, endpoint, proxyIP, pc.pathPrefix, pc.kind, model, "Reply with exactly: pong", "")
 	}
 	require.NoError(t, err, "request must reach the proxy for %s", pc.name)
 	return code
@@ -147,11 +166,12 @@ func TestModelAllowlistEnforced(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Probe first: the GET resolves the endpoint (DNS error fails) and its first packet wakes the lazy proxy peer, so WaitProxyPeer sees it connected; any HTTP status counts.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
 	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
 		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
 	}
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
 
 	for _, pc := range providers {
 		pc := pc

e2e/agentnetwork/skiptls_test.go

@@ -104,11 +104,12 @@ func TestProviderSkipTLSVerification(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Probe first: the GET resolves the endpoint (DNS error fails) and its first packet wakes the lazy proxy peer, so WaitProxyPeer sees it connected; any HTTP status counts.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
 	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
 		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
 	}
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
 
 	// Positive: skip=true reaches the self-signed upstream. Retry to absorb
 	// tunnel/DNS jitter on the first call; success also proves the path works.

e2e/agentnetwork/vllm_test.go

@@ -106,11 +106,12 @@ func TestVLLMProvider(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Probe first: the GET resolves the endpoint (DNS error fails) and its first packet wakes the lazy proxy peer, so WaitProxyPeer sees it connected; any HTTP status counts.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
 	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
 		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
 	}
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
 
 	before, _ := srv.ListAccessLogs(ctx)
 	sessionID := "e2e-session-vllm"

e2e/harness/client.go

@@ -4,6 +4,7 @@ package harness
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"io"
 	"os/exec"
@@ -167,22 +168,54 @@ func (cl *Client) pollStatus(ctx context.Context, timeout time.Duration, want st
 	return fmt.Errorf("timed out waiting for %q; last status:\n%s", want, last)
 }
 
-// ResolveProxyIP resolves the agent-network endpoint to the proxy peer's
-// NetBird IP from inside the client (via magic DNS).
+const (
+	// curlExitCouldNotResolve is curl's exit code for a DNS resolution failure, distinct from connection-level failures.
+	curlExitCouldNotResolve = 6
+	// dnsProbeRetryWindow bounds DNS-failure retries: the synthesized zone lands a beat after management connects, so early NXDOMAIN is propagation; a zone still absent after this window is a real failure.
+	dnsProbeRetryWindow   = 30 * time.Second
+	dnsProbeRetryInterval = 2 * time.Second
+)
+
+// ResolveProxyIP GETs https://<endpoint>/ from the client's netns: any HTTP status proves DNS + tunnel and wakes the lazy proxy peer; only DNS failures retry, within dnsProbeRetryWindow. Returns the connected IP for --resolve pinning.
 func (cl *Client) ResolveProxyIP(ctx context.Context, endpoint string) (string, error) {
-	code, reader, err := cl.container.Exec(ctx, []string{"getent", "hosts", endpoint}, tcexec.Multiplexed())
-	if err != nil {
-		return "", err
+	args := []string{
+		"run", "--rm",
+		"--network", "container:" + cl.container.GetContainerID(),
+		curlImage,
+		"-ksS", "-o", "/dev/null",
+		"--connect-timeout", "30", "--max-time", "60",
+		"-w", "%{remote_ip}",
+		"https://" + endpoint + "/",
 	}
-	out, _ := io.ReadAll(reader)
-	if code != 0 {
-		return "", fmt.Errorf("getent hosts %s exited %d", endpoint, code)
+	deadline := time.Now().Add(dnsProbeRetryWindow)
+	for {
+		cmd := exec.CommandContext(ctx, "docker", args...)
+		var stdout, stderr strings.Builder
+		cmd.Stdout = &stdout
+		cmd.Stderr = &stderr
+		err := cmd.Run()
+		if err == nil {
+			ip := strings.TrimSpace(stdout.String())
+			if ip == "" {
+				return "", fmt.Errorf("got an HTTP response from %s but no remote IP", endpoint)
+			}
+			return ip, nil
+		}
+
+		var exitErr *exec.ExitError
+		if !errors.As(err, &exitErr) || exitErr.ExitCode() != curlExitCouldNotResolve {
+			return "", fmt.Errorf("no HTTP response from %s: %w (%s)", endpoint, err, strings.TrimSpace(stderr.String()))
+		}
+		dnsErr := fmt.Errorf("DNS resolution failed for %s: %s", endpoint, strings.TrimSpace(stderr.String()))
+		if time.Until(deadline) < dnsProbeRetryInterval {
+			return "", dnsErr
+		}
+		select {
+		case <-ctx.Done():
+			return "", fmt.Errorf("%w (%w)", dnsErr, ctx.Err())
+		case <-time.After(dnsProbeRetryInterval):
+		}
 	}
-	fields := strings.Fields(string(out))
-	if len(fields) == 0 {
-		return "", fmt.Errorf("no address for %s", endpoint)
-	}
-	return fields[0], nil
 }
 
 // Wire shapes for Chat.
@@ -206,18 +239,29 @@ const (
 // the wire shape: WireChat (OpenAI) or WireMessages (Anthropic). A non-empty
 // sessionID is sent as the universal x-session-id header the proxy records.
 func (cl *Client) Chat(ctx context.Context, endpoint, proxyIP, kind, model, prompt, sessionID string) (int, string, error) {
+	return cl.ChatPrefixed(ctx, endpoint, proxyIP, "", kind, model, prompt, sessionID)
+}
+
+// ChatPrefixed is Chat with a base-URL path prefix prepended to the wire
+// path, mirroring agents whose base URL carries a shape-selecting prefix that
+// rides through to the upstream — e.g. Claude Code against a Kimi provider
+// sets ANTHROPIC_BASE_URL=https://<endpoint>/anthropic so the proxy forwards
+// /anthropic/v1/messages to Moonshot's Anthropic surface while the provider's
+// upstream URL stays the bare https://api.moonshot.ai. Empty prefix is plain
+// Chat.
+func (cl *Client) ChatPrefixed(ctx context.Context, endpoint, proxyIP, pathPrefix, kind, model, prompt, sessionID string) (int, string, error) {
 	var path, body string
 	var headers []string
 	switch kind {
 	case WireMessages:
 		path = "/v1/messages"
 		headers = []string{"anthropic-version: 2023-06-01"}
-		body = fmt.Sprintf(`{"model":%q,"max_tokens":64,"messages":[{"role":"user","content":%q}]}`, model, prompt)
+		body = fmt.Sprintf(`{"model":%q,"max_tokens":2048,"messages":[{"role":"user","content":%q}]}`, model, prompt)
 	default:
 		path = "/v1/chat/completions"
 		body = fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":%q}]}`, model, prompt)
 	}
-	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(headers, sessionID))
+	return cl.post(ctx, endpoint, proxyIP, pathPrefix+path, body, withSessionID(headers, sessionID))
 }
 
 // Vertex issues an Anthropic-on-Vertex rawPredict POST over the tunnel. Unlike
@@ -227,7 +271,7 @@ func (cl *Client) Chat(ctx context.Context, endpoint, proxyIP, kind, model, prom
 // is sent as the universal x-session-id header the proxy records.
 func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region, model, prompt, sessionID string) (int, string, error) {
 	path := fmt.Sprintf("/v1/projects/%s/locations/%s/publishers/anthropic/models/%s:rawPredict", project, region, model)
-	body := fmt.Sprintf(`{"anthropic_version":"vertex-2023-10-16","max_tokens":64,"messages":[{"role":"user","content":%q}]}`, prompt)
+	body := fmt.Sprintf(`{"anthropic_version":"vertex-2023-10-16","max_tokens":2048,"messages":[{"role":"user","content":%q}]}`, prompt)
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
@@ -238,7 +282,7 @@ func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region
 // header the proxy records.
 func (cl *Client) Bedrock(ctx context.Context, endpoint, proxyIP, model, prompt, sessionID string) (int, string, error) {
 	path := "/model/" + model + "/invoke"
-	body := fmt.Sprintf(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":64,"messages":[{"role":"user","content":%q}]}`, prompt)
+	body := fmt.Sprintf(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":2048,"messages":[{"role":"user","content":%q}]}`, prompt)
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
@@ -297,7 +341,7 @@ func (cl *Client) Terminate(ctx context.Context) error {
 	return cl.container.Terminate(ctx)
 }
 
-// containerLogs reads up to 256 KiB of a container's logs for diagnostics.
+// containerLogs reads up to 4 MiB of a container's logs for diagnostics — enough for a whole provider-matrix run.
 func containerLogs(ctx context.Context, c testcontainers.Container) string {
 	if c == nil {
 		return ""
@@ -307,6 +351,6 @@ func containerLogs(ctx context.Context, c testcontainers.Container) string {
 		return fmt.Sprintf("<logs error: %v>", err)
 	}
 	defer r.Close()
-	b, _ := io.ReadAll(io.LimitReader(r, 256<<10))
+	b, _ := io.ReadAll(io.LimitReader(r, 4<<20))
 	return string(b)
 }

e2e/harness/combined.go

@@ -221,6 +221,29 @@ func (c *Combined) CreateProxyTokenCLI(ctx context.Context, name string) (string
 	return "", fmt.Errorf("token not found in CLI output: %s", string(out))
 }
 
+// SnapshotStoreDB copies the management sqlite store (with WAL/SHM sidecars) out of the bind-mounted
+// data dir into dstDir and returns the copy's path; reading a copy avoids locking against live writes.
+func (c *Combined) SnapshotStoreDB(dstDir string) (string, error) {
+	src := filepath.Join(c.workDir, "data", "store.db")
+	if _, err := os.Stat(src); err != nil {
+		return "", fmt.Errorf("management store not found at %s: %w", src, err)
+	}
+	dst := filepath.Join(dstDir, "store.db")
+	for _, suffix := range []string{"", "-wal", "-shm"} {
+		data, err := os.ReadFile(src + suffix)
+		if err != nil {
+			if os.IsNotExist(err) && suffix != "" {
+				continue // sidecar only exists in WAL mode
+			}
+			return "", fmt.Errorf("read %s: %w", src+suffix, err)
+		}
+		if err := os.WriteFile(dst+suffix, data, 0o600); err != nil {
+			return "", fmt.Errorf("write %s: %w", dst+suffix, err)
+		}
+	}
+	return dst, nil
+}
+
 // Logs returns the combined server container logs, for diagnostics.
 func (c *Combined) Logs(ctx context.Context) string {
 	return containerLogs(ctx, c.container)

go.mod

@@ -113,7 +113,7 @@ require (
 	github.com/ti-mo/conntrack v0.5.1
 	github.com/ti-mo/netfilter v0.5.2
 	github.com/vmihailenco/msgpack/v5 v5.4.1
-	github.com/wailsapp/wails/v3 v3.0.0-alpha2.111
+	github.com/wailsapp/wails/v3 v3.0.0-alpha2.117
 	github.com/yusufpapurcu/wmi v1.2.4
 	github.com/zcalusic/sysinfo v1.1.3
 	go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.67.0
@@ -303,7 +303,6 @@ require (
 	github.com/tklauser/numcpus v0.10.0 // indirect
 	github.com/vishvananda/netns v0.0.5 // indirect
 	github.com/vmihailenco/tagparser/v2 v2.0.0 // indirect
-	github.com/wailsapp/wails/webview2 v1.0.27 // indirect
 	github.com/wlynxg/anet v0.0.5 // indirect
 	github.com/x448/float16 v0.8.4 // indirect
 	github.com/zeebo/blake3 v0.2.3 // indirect

go.sum

@@ -660,10 +660,8 @@ github.com/vmihailenco/msgpack/v5 v5.4.1 h1:cQriyiUvjTwOHg8QZaPihLWeRAAVoCpE00IU
 github.com/vmihailenco/msgpack/v5 v5.4.1/go.mod h1:GaZTsDaehaPpQVyxrf5mtQlH+pc21PIudVV/E3rRQok=
 github.com/vmihailenco/tagparser/v2 v2.0.0 h1:y09buUbR+b5aycVFQs/g70pqKVZNBmxwAhO7/IwNM9g=
 github.com/vmihailenco/tagparser/v2 v2.0.0/go.mod h1:Wri+At7QHww0WTrCBeu4J6bNtoV6mEfg5OIWRZA9qds=
-github.com/wailsapp/wails/v3 v3.0.0-alpha2.111 h1:MKx1nOnhnDuEGrRBmtxLOJq1NERwailu2cI4BvzWhi4=
-github.com/wailsapp/wails/v3 v3.0.0-alpha2.111/go.mod h1:wrdvmyeCsB/K3YqJDoH8E3MwcN8NXAMnEFaDTW46w60=
-github.com/wailsapp/wails/webview2 v1.0.27 h1:wjgAi/I8BBZ7kUGU8um3XF3ILEfzr96Q2Q1G4GPjMns=
-github.com/wailsapp/wails/webview2 v1.0.27/go.mod h1:zdM4jcO1IaC61RiJL5F1BzgoqBHFIdacz8gPr5exr0o=
+github.com/wailsapp/wails/v3 v3.0.0-alpha2.117 h1:udyjqPG3AIgkod5QDR/WblCkpV8R86BFPSrsWxSyt5Y=
+github.com/wailsapp/wails/v3 v3.0.0-alpha2.117/go.mod h1:74WH2FScMsgucZvHHvv7eOefDXCm/CjuIxqhhZgPhKg=
 github.com/wlynxg/anet v0.0.5 h1:J3VJGi1gvo0JwZ/P1/Yc/8p63SoW98B5dHkYDmpgvvU=
 github.com/wlynxg/anet v0.0.5/go.mod h1:eay5PRQr7fIVAMbTbchTnO9gG65Hg/uYGdc7mguHxoA=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=

idp/dex/config.go

@@ -308,7 +308,7 @@ func (s *Storage) OpenStorage(logger *slog.Logger) (storage.Storage, error) {
 		if file == "" {
 			return nil, fmt.Errorf("sqlite3 storage requires 'file' config")
 		}
-		return (&sql.SQLite3{File: file}).Open(logger)
+		return newSQLite3(file).Open(logger)
 	case "postgres":
 		dsn, _ := s.Config["dsn"].(string)
 		if dsn == "" {