pqkem: recover from persistent rekey failure by re-bootstrapping over signal

OnRekeyFailed now re-runs the KEM bootstrap over Signal (conn.RequestReoffer ->
handshaker.SendOffer) instead of only logging: a fresh signalling offer starts a new
exchange that overwrites the stalled PSK on both sides, resyncing after a persistent
data-path desync. Chosen over a responder-side awaitingAck revert (which fights the
confirm-less ack timing) and a full tunnel teardown (heavier). The tunnel stays up on
the previous PSK meanwhile since Signal is independent of the broken data path.

.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/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/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

@@ -621,6 +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),
 
@@ -696,6 +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 +656,30 @@ 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 {
+			cbHandler := pqCallbackHandler{
+				wg: e.wgInterface,
+				// On a persistent rekey failure, re-bootstrap the KEM over Signal: a
+				// fresh signalling offer starts a new exchange that overwrites the
+				// stalled PSK on both sides, recovering from a data-path desync.
+				reoffer: func(remoteKey string) {
+					if conn, ok := e.peerStore.PeerConn(remoteKey); ok {
+						conn.RequestReoffer()
+					}
+				},
+			}
+			e.pqkemManager = pqkem.NewManager(pqkem.LocalID(publicKey.String()), cbHandler, 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 +692,24 @@ 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 +943,10 @@ 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 +1023,12 @@ 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 +1038,47 @@ 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 +1091,14 @@ 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 +1112,19 @@ 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 +1284,7 @@ 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 +1933,10 @@ func (e *Engine) createPeerConn(pubKey string, allowedIPs []netip.Prefix, agentV
 		},
 		ICEConfig: e.createICEConfig(),
 	}
+	if e.pqkemManager != nil {
+		config.PQ = pqHandshaker{mgr: e.pqkemManager}
+		config.PQStrict = pqkem.Strict()
+	}
 
 	serviceDependencies := peer.ServiceDependencies{
 		StatusRecorder:     e.statusRecorder,
@@ -1995,6 +2120,10 @@ 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 +2161,7 @@ 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 +2936,8 @@ 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/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

@@ -3,6 +3,7 @@ package peer
 import (
 	"context"
 	"fmt"
+	"math"
 	"net"
 	"net/netip"
 	"runtime"
@@ -26,6 +27,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/portforward"
 	"github.com/netbirdio/netbird/client/internal/rosenpass"
 	"github.com/netbirdio/netbird/client/internal/stdnet"
+	"github.com/netbirdio/netbird/monotime"
 	"github.com/netbirdio/netbird/route"
 	relayClient "github.com/netbirdio/netbird/shared/relay/client"
 )
@@ -74,6 +76,34 @@ 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. sinceActivity is how long
+	// ago the peer last exchanged real user data, so the rotation can be skipped for
+	// idle tunnels.
+	OnDataPathRekeyed(remoteKey string, sinceActivity time.Duration)
+	// 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 +121,12 @@ type ConnConfig struct {
 
 	RosenpassConfig RosenpassConfig
 
+	// PQ carries post-quantum ML-KEM material on offers/answers; nil when disabled.
+	PQ PQHandshaker
+	// PQStrict fails closed: block peer traffic until the ML-KEM PSK is established,
+	// instead of letting the tunnel come up classically and upgrading to PQ later.
+	PQStrict bool
+
 	// ICEConfig ICE protocol configuration
 	ICEConfig icemaker.Config
 }
@@ -149,6 +185,11 @@ type Conn struct {
 	// pendingFirstPacket is the lazyconn-captured handshake init, replayed once the real
 	// transport is up.
 	pendingFirstPacket []byte
+
+	// pqBlockingKey is a per-conn random sentinel PSK used in PQ strict mode to fail
+	// closed: it is programmed until the real ML-KEM PSK is derived, so no session can
+	// form on a non-PQ key. Per-conn random so two strict peers never match by chance.
+	pqBlockingKey *wgtypes.Key
 }
 
 // injectPendingFirstPacket replays the captured handshake through the proxy if present, else
@@ -206,6 +247,14 @@ func NewConn(config ConnConfig, services ServiceDependencies) (*Conn, error) {
 		metricsRecorder:    services.MetricsRecorder,
 	}
 
+	if config.PQ != nil && config.PQStrict {
+		if k, err := wgtypes.GenerateKey(); err != nil {
+			connLog.Errorf("pqkem: failed to generate strict-mode sentinel key, strict fail-closed disabled for this peer: %v", err)
+		} else {
+			conn.pqBlockingKey = &k
+		}
+	}
+
 	return conn, nil
 }
 
@@ -670,6 +719,22 @@ func (conn *Conn) onGuardEvent() {
 	}
 }
 
+// RequestReoffer sends a fresh signalling offer for the peer, re-running the
+// post-quantum bootstrap over Signal. Used to recover from a persistent data-path
+// rekey failure: a new exchange overwrites the stalled PSK on both sides. No-op if the
+// connection is not open yet.
+func (conn *Conn) RequestReoffer() {
+	conn.mu.Lock()
+	h := conn.handshaker
+	conn.mu.Unlock()
+	if h == nil {
+		return
+	}
+	if err := h.SendOffer(); err != nil {
+		conn.Log.Debugf("pqkem: recovery re-offer failed: %v", err)
+	}
+}
+
 func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
@@ -681,6 +746,10 @@ 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:
@@ -723,7 +792,7 @@ func (conn *Conn) updateRelayStatus(relayServerAddr string, rosenpassPubKey []by
 		ConnStatus:         conn.evalStatus(),
 		Relayed:            conn.isRelayed(),
 		RelayServerAddress: relayServerAddr,
-		RosenpassEnabled:   isRosenpassEnabled(rosenpassPubKey),
+		RosenpassEnabled:   conn.quantumResistant(rosenpassPubKey),
 	}
 
 	err := conn.statusRecorder.UpdatePeerRelayedState(peerState)
@@ -742,7 +811,7 @@ func (conn *Conn) updateIceState(iceConnInfo ICEConnInfo, updateTime time.Time)
 		RemoteIceCandidateType:     iceConnInfo.RemoteIceCandidateType,
 		LocalIceCandidateEndpoint:  iceConnInfo.LocalIceCandidateEndpoint,
 		RemoteIceCandidateEndpoint: iceConnInfo.RemoteIceCandidateEndpoint,
-		RosenpassEnabled:           isRosenpassEnabled(iceConnInfo.RosenpassPubKey),
+		RosenpassEnabled:           conn.quantumResistant(iceConnInfo.RosenpassPubKey),
 	}
 
 	err := conn.statusRecorder.UpdatePeerICEState(peerState)
@@ -946,6 +1015,35 @@ func (conn *Conn) onWGCheckSuccess() {
 	conn.mu.Lock()
 	conn.wgTimeouts = 0
 	conn.mu.Unlock()
+
+	// A fresh WireGuard handshake clocks the post-quantum PSK rotation. Pass how long
+	// ago the peer last exchanged real user data (keepalives excluded) so the pqkem
+	// manager can skip rotation on idle tunnels — rotating then would push data-path
+	// traffic that keeps the lazy connection artificially active.
+	if conn.config.PQ != nil {
+		conn.config.PQ.OnDataPathRekeyed(conn.config.Key, conn.dataActivityAge())
+	}
+}
+
+// dataActivityAge returns how long ago the peer last exchanged real user data
+// (WireGuard keepalives excluded), per the same LastActivities signal the
+// lazy-connection inactivity monitor uses. It reports a very large duration when no
+// activity has ever been recorded, so the peer is treated as idle.
+//
+// In kernel mode there is no per-peer data-activity signal (LastActivities is
+// userspace-only), so we cannot tell active from idle. We report zero — always
+// "active" — so PSK rotation is not disabled in kernel mode. Lazy back-to-idle is
+// already limited there; the eBPF WG-activity detection (future) will supply a real
+// signal that excludes handshake/pqkem traffic.
+func (conn *Conn) dataActivityAge() time.Duration {
+	if !conn.config.WgConfig.WgInterface.IsUserspaceBind() {
+		return 0
+	}
+	last, ok := conn.config.WgConfig.WgInterface.LastActivities()[conn.config.WgConfig.RemoteKey]
+	if !ok {
+		return time.Duration(math.MaxInt64)
+	}
+	return monotime.Since(last)
 }
 
 // recordConnectionMetrics records connection stage timestamps as metrics
@@ -987,6 +1085,23 @@ 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.PQStrict && conn.pqBlockingKey != nil {
+			// Fail closed: program a non-matching sentinel so no session forms on a
+			// non-PQ key until the ML-KEM exchange derives the real PSK (pushed via
+			// SetPresharedKey once it converges). "pending" — turns into a "stuck"
+			// warning from the manager if the exchange keeps failing (see raiseFailure).
+			conn.Log.Debugf("pqkem: strict mode — no PQ PSK yet, blocking peer traffic until the ML-KEM exchange converges")
+			return conn.pqBlockingKey
+		}
+	}
+
 	if conn.config.RosenpassConfig.PubKey == nil {
 		return conn.config.WgConfig.PreSharedKey
 	}
@@ -1026,6 +1141,21 @@ func isRosenpassEnabled(remoteRosenpassPubKey []byte) bool {
 	return remoteRosenpassPubKey != nil
 }
 
+// quantumResistant reports whether the peer's tunnel is post-quantum protected, for
+// the status "Quantum resistance" field: either Rosenpass (the remote advertised a
+// Rosenpass key) or the ML-KEM exchange (a PQ PSK has been derived for this peer).
+func (conn *Conn) quantumResistant(remoteRosenpassPubKey []byte) bool {
+	if isRosenpassEnabled(remoteRosenpassPubKey) {
+		return true
+	}
+	if conn.config.PQ != nil {
+		if _, ok := conn.config.PQ.PSK(conn.config.Key); ok {
+			return true
+		}
+	}
+	return false
+}
+
 func evalConnStatus(in connStatusInputs) guard.ConnStatus {
 	// "Relay up and needed" — the peer uses relay and the transport is connected.
 	relayUsedAndUp := in.peerUsesRelay && in.relayConnected

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
@@ -81,14 +91,20 @@ type Handshaker struct {
 
 func NewHandshaker(log *log.Entry, config ConnConfig, signaler *Signaler, ice *WorkerICE, relay *WorkerRelay, metricsStages *MetricsStages) *Handshaker {
 	h := &Handshaker{
-		log:            log,
-		config:         config,
-		signaler:       signaler,
-		ice:            ice,
-		relay:          relay,
-		metricsStages:  metricsStages,
-		remoteOffersCh: make(chan OfferAnswer),
-		remoteAnswerCh: make(chan OfferAnswer),
+		log:           log,
+		config:        config,
+		signaler:      signaler,
+		ice:           ice,
+		relay:         relay,
+		metricsStages: metricsStages,
+		// Buffered by 1: the single Listen goroutine can be busy handling an offer
+		// (sendAnswer does a blocking signal send) exactly when the matching answer
+		// arrives on the other channel. Unbuffered, that answer would hit the
+		// non-blocking send's default and be dropped — fatal for the post-quantum
+		// exchange, which needs the answer to converge. A 1-slot cushion lets it wait
+		// until Listen loops back, without ever blocking the signal receiver.
+		remoteOffersCh: make(chan OfferAnswer, 1),
+		remoteAnswerCh: make(chan OfferAnswer, 1),
 	}
 	// assume remote supports ICE until we learn otherwise from received offers
 	h.remoteICESupported.Store(ice != nil)
@@ -120,6 +136,8 @@ func (h *Handshaker) Listen(ctx context.Context) {
 
 			h.updateRemoteICEState(&remoteOfferAnswer)
 
+			h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
+
 			if h.relayListener != nil {
 				h.relayListener.Notify(&remoteOfferAnswer)
 			}
@@ -128,7 +146,7 @@ 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 +160,8 @@ func (h *Handshaker) Listen(ctx context.Context) {
 
 			h.updateRemoteICEState(&remoteOfferAnswer)
 
+			h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
+
 			if h.relayListener != nil {
 				h.relayListener.Notify(&remoteOfferAnswer)
 			}
@@ -149,6 +169,10 @@ 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 +180,16 @@ 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 +229,23 @@ 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/iface.go

@@ -10,6 +10,7 @@ import (
 	"github.com/netbirdio/netbird/client/iface/configurer"
 	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	"github.com/netbirdio/netbird/client/iface/wgproxy"
+	"github.com/netbirdio/netbird/monotime"
 )
 
 type WGIface interface {
@@ -19,4 +20,11 @@ type WGIface interface {
 	GetProxy() wgproxy.Proxy
 	Address() wgaddr.Address
 	RemoveEndpointAddress(key string) error
+	// LastActivities returns the last real-data activity time per peer (WireGuard
+	// keepalives excluded), used to gate post-quantum PSK rotation on active tunnels.
+	LastActivities() map[string]monotime.Time
+	// IsUserspaceBind reports whether WireGuard runs in userspace. Only there does
+	// LastActivities track per-peer data activity; in kernel mode it is unavailable,
+	// so PSK rotation cannot be gated on activity.
+	IsUserspaceBind() bool
 }

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/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 +1,271 @@
+package pqkem
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"time"
+)
+
+// idHex renders an exchange ID for logs.
+func idHex(id ExchangeID) string { return hex.EncodeToString(id[:]) }
+
+// pskFingerprint is a short, non-secret digest of a derived PSK: identical on both
+// peers iff they derived the same key. Logged instead of the raw PSK so debug logs
+// never carry the actual WireGuard preshared key.
+func pskFingerprint(psk PSK) string {
+	sum := sha256.Sum256(psk[:])
+	return hex.EncodeToString(sum[:8])
+}
+
+// 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)
+
+	via := "data-path"
+	if viaSignal {
+		via = "signal"
+	}
+	m.trace("pqkem: offer sent", "peer", remoteID, "exchange", idHex(id), "acks", idHex(ackID), "via", via)
+	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) {
+	m.trace("pqkem: offer received", "peer", remoteID, "exchange", idHex(o.ExchangeID), "acks", idHex(o.AckID))
+
+	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
+		}
+		m.trace("pqkem: duplicate offer, resending cached answer", "peer", remoteID, "exchange", idHex(o.ExchangeID))
+		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()
+		m.trace("pqkem: exchange superseded during respond, dropping answer", "peer", remoteID, "exchange", idHex(o.ExchangeID))
+		return nil, nil
+	}
+	ex.state = stateAwaitingAck
+	ex.lastSent = raw
+	ex.pendingPSK = psk
+	m.psks[remoteID] = psk
+	m.mu.Unlock()
+
+	m.trace("pqkem: new PSK derived", "peer", remoteID, "exchange", idHex(o.ExchangeID), "role", "responder", "psk_fp", pskFingerprint(psk))
+
+	// Commit optimistically so our data path can rekey to the new PSK.
+	if err := m.cbHandler.OnNewPSKReady(remoteID, psk); err != nil {
+		return nil, err
+	}
+	m.trace("pqkem: answer sent", "peer", remoteID, "exchange", idHex(o.ExchangeID))
+	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 {
+		haveID := "none"
+		if ex != nil {
+			haveID = idHex(ex.id)
+		}
+		m.mu.Unlock()
+		m.trace("pqkem: unexpected answer dropped (inconsistency)", "peer", remoteID, "answer_for", idHex(a.ExchangeID), "have_exchange", haveID)
+		return nil
+	}
+	ex.state = stateAwaitingRekey
+	init := ex.initiator
+	ex.initiator = nil
+	m.mu.Unlock()
+
+	m.trace("pqkem: answer received", "peer", remoteID, "exchange", idHex(a.ExchangeID))
+
+	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()
+
+	m.trace("pqkem: new PSK derived", "peer", remoteID, "exchange", idHex(a.ExchangeID), "role", "initiator", "psk_fp", pskFingerprint(psk))
+
+	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()
+		m.trace("pqkem: ack for unknown/mismatched exchange, ignored (inconsistency)", "peer", remoteID, "acks", idHex(ackID))
+		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()
+
+	m.trace("pqkem: previous exchange confirmed by ack", "peer", remoteID, "exchange", idHex(ackID))
+}
+
+// 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)
+					initial := !m.established[remoteID]
+					fail := m.registerFailureLocked(remoteID)
+					m.mu.Unlock()
+					m.raiseFailure(remoteID, fail, initial)
+					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
+}
+
+// raiseFailure reports a convergence failure. initial distinguishes a never-established
+// peer (bootstrap failed → no PQ PSK at all; in strict mode the peer stays blocked =
+// "stuck") from a rekey failure (a previous PSK is still in force and traffic continues).
+func (m *Manager) raiseFailure(remoteID RemoteID, fail, initial bool) {
+	if !fail {
+		m.logger.Warn("pqkem: rekey attempt timed out, will retry next cycle", "peer", remoteID)
+		return
+	}
+	if initial {
+		m.logger.Warn("pqkem: initial exchange failed — no PQ PSK established for peer (strict mode keeps the peer blocked until it converges)", "peer", remoteID)
+	} else {
+		m.logger.Warn("pqkem: rekey failed after retries — staying on the previous PSK", "peer", remoteID)
+	}
+	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", 0)
+	dB.OnDataPathRekeyed("aaaa", 0)
+	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 +1,138 @@
+package pqkem
+
+import (
+	"context"
+	"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
+}
+
+// EnvStrict enables strict (fail-closed) mode: block peer traffic until the ML-KEM
+// PSK has been established, instead of the default opportunistic behaviour that lets
+// the tunnel come up classically and upgrades to PQ once the exchange converges.
+const EnvStrict = "NB_PQ_MLKEM_STRICT"
+
+// Strict reports whether strict (fail-closed) mode is enabled via the environment.
+// An empty or unrecognized value is treated as disabled (opportunistic).
+func Strict() bool {
+	switch strings.ToLower(strings.TrimSpace(os.Getenv(EnvStrict))) {
+	case "on":
+		return true
+	case "", "off":
+		return false
+	}
+	enabled, err := strconv.ParseBool(strings.TrimSpace(os.Getenv(EnvStrict)))
+	if err != nil {
+		log.Warnf("failed to parse %s value %q: %v", EnvStrict, os.Getenv(EnvStrict), err)
+		return false
+	}
+	return enabled
+}
+
+// EnvLogLevel overrides the ML-KEM manager's slog level (trace/debug/info/warn/error).
+// Defaults to info. The verbose per-exchange lifecycle logs are emitted at trace.
+const EnvLogLevel = "NB_PQ_MLKEM_LOG_LEVEL"
+
+// LevelTrace is a custom slog level below Debug for the verbose per-exchange lifecycle
+// logs, so they stay off unless NB_PQ_MLKEM_LOG_LEVEL=trace (and the daemon log level
+// is trace, since the records are forwarded to logrus).
+const LevelTrace = slog.LevelDebug - 4
+
+// NewLogger builds the slog logger for the ML-KEM manager. It forwards records to
+// logrus so PQ logs land in the same sink as the rest of the daemon (console +
+// client.log) rather than stdout. Verbosity is gated by EnvLogLevel.
+func NewLogger() *slog.Logger {
+	return slog.New(slogToLogrus{})
+}
+
+func logLevel() slog.Level {
+	switch strings.ToLower(strings.TrimSpace(os.Getenv(EnvLogLevel))) {
+	case "trace":
+		return LevelTrace
+	case "debug":
+		return slog.LevelDebug
+	case "warn":
+		return slog.LevelWarn
+	case "error":
+		return slog.LevelError
+	default:
+		return slog.LevelInfo
+	}
+}
+
+// slogToLogrus is a slog.Handler that forwards records to logrus, so the ML-KEM
+// manager's logs go wherever the daemon's logrus is configured (console + client.log)
+// instead of stdout. Verbosity is gated by EnvLogLevel via logLevel().
+type slogToLogrus struct {
+	fields log.Fields
+}
+
+func (h slogToLogrus) Enabled(_ context.Context, level slog.Level) bool {
+	return level >= logLevel()
+}
+
+func (h slogToLogrus) Handle(_ context.Context, r slog.Record) error {
+	fields := make(log.Fields, len(h.fields)+r.NumAttrs())
+	for k, v := range h.fields {
+		fields[k] = v
+	}
+	r.Attrs(func(a slog.Attr) bool {
+		fields[a.Key] = a.Value.Any()
+		return true
+	})
+	entry := log.WithFields(fields)
+	switch {
+	case r.Level >= slog.LevelError:
+		entry.Error(r.Message)
+	case r.Level >= slog.LevelWarn:
+		entry.Warn(r.Message)
+	case r.Level >= slog.LevelInfo:
+		entry.Info(r.Message)
+	case r.Level >= slog.LevelDebug:
+		entry.Debug(r.Message)
+	default:
+		entry.Trace(r.Message)
+	}
+	return nil
+}
+
+func (h slogToLogrus) WithAttrs(attrs []slog.Attr) slog.Handler {
+	fields := make(log.Fields, len(h.fields)+len(attrs))
+	for k, v := range h.fields {
+		fields[k] = v
+	}
+	for _, a := range attrs {
+		fields[a.Key] = a.Value.Any()
+	}
+	return slogToLogrus{fields: fields}
+}
+
+func (h slogToLogrus) WithGroup(_ string) slog.Handler { return h }

client/internal/pqkem/kem.go

@@ -0,0 +1,178 @@
+// 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 as the KDF input. The PSK is derived with
+// HKDF-SHA256 over that hybrid secret, salted with a domain-separation label and
+// bound (via the HKDF info) to the full transcript and the canonicalised peer
+// identities.
+package pqkem
+
+import (
+	"crypto/ecdh"
+	"crypto/hkdf"
+	"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)
+}
+
+// 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.
+	psk, err = derivePSK(ssMLKEM, ssX, offer, answer, b)
+	if err != nil {
+		return nil, PSK{}, err
+	}
+	return answer, psk, nil
+}
+
+// derivePSK runs HKDF-SHA256 over the hybrid shared secret (ML-KEM_ss ‖ X25519_ss,
+// per draft-ietf-tls-ecdhe-mlkem), salted with the domain-separation label, and binds
+// the result — via the HKDF info — to the full transcript (offer ‖ answer) and the
+// canonicalised peer identities, so the PSK cannot be transplanted to another peer
+// pair or a different exchange.
+func derivePSK(ssMLKEM, ssX, offer, answer []byte, b Binding) (PSK, error) {
+	lo, hi := canonicalPair(b.LocalID, b.RemoteID)
+
+	ikm := make([]byte, 0, len(ssMLKEM)+len(ssX))
+	ikm = append(ikm, ssMLKEM...)
+	ikm = append(ikm, ssX...)
+
+	info := make([]byte, 0, len(offer)+len(answer)+len(lo)+len(hi))
+	info = append(info, offer...)
+	info = append(info, answer...)
+	info = append(info, lo...)
+	info = append(info, hi...)
+
+	var psk PSK
+	key, err := hkdf.Key(sha256.New, ikm, []byte(pskLabel), string(info), len(psk))
+	if err != nil {
+		return PSK{}, fmt.Errorf("hkdf derive psk: %w", err)
+	}
+	copy(psk[:], key)
+	return psk, nil
+}
+
+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 +1,399 @@
+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
+
+	// rotationActivityWindow gates rotation on recent real-data activity: a rekey
+	// clocks a rotation only if the peer exchanged user data within this window. It
+	// must stay shorter than the data path's rekey interval (WireGuard
+	// REKEY_AFTER_TIME ~120s) so the rotation's own traffic — which itself renews the
+	// activity signal — ages out before the next rekey, letting an idle tunnel stop
+	// rotating instead of self-sustaining.
+	rotationActivityWindow = 90 * time.Second
+)
+
+// 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
+}
+
+// trace logs at LevelTrace, the verbose per-exchange lifecycle level gated by
+// NB_PQ_MLKEM_LOG_LEVEL=trace.
+func (m *Manager) trace(msg string, args ...any) {
+	m.logger.Log(context.Background(), LevelTrace, msg, args...)
+}
+
+// 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.trace("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).
+// OnDataPathRekeyed clocks the next chained PSK rotation on a fresh data-path rekey.
+// sinceActivity is how long ago the peer last exchanged real user data; when it
+// exceeds rotationActivityWindow the tunnel is treated as idle and rotation is
+// skipped — an idle tunnel has nothing to protect, and rotating would emit data-path
+// traffic that keeps the peer artificially active (see conn.onWGCheckSuccess).
+func (m *Manager) OnDataPathRekeyed(remoteID RemoteID, sinceActivity time.Duration) {
+	if sinceActivity >= rotationActivityWindow {
+		m.trace("pqkem: peer idle, skipping data-path rotation", "peer", remoteID, "since_activity", sinceActivity)
+		return
+	}
+
+	m.mu.Lock()
+	ex := m.exchanges[remoteID]
+	chain := ex != nil && ex.state == stateAwaitingRekey
+	var ackID ExchangeID
+	if chain {
+		ackID = ex.id
+	}
+	m.mu.Unlock()
+
+	m.trace("pqkem: data-path rekey signal", "peer", remoteID, "chaining", chain)
+	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)
+		return
+	}
+	m.trace("pqkem: chain offer sent over data path", "peer", remoteID)
+}
+
+// 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 +1,193 @@
+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", 0)
+	dB.OnDataPathRekeyed("aaaa", 0)
+
+	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_RotationSkippedWhenIdle(t *testing.T) {
+	dA, dB, wgA, wgB, _ := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	bootstrap(t, dA, dB)
+	psk1 := wgB.psk("aaaa")
+	require.NotEqual(t, PSK{}, psk1)
+
+	// Idle: the peer's last real-data activity is older than the window, so a rekey
+	// must NOT clock a rotation.
+	dA.OnDataPathRekeyed("bbbb", rotationActivityWindow)
+	dB.OnDataPathRekeyed("aaaa", rotationActivityWindow)
+	require.Equal(t, psk1, wgB.psk("aaaa"), "idle peer must not rotate the PSK")
+	require.Equal(t, psk1, wgA.psk("bbbb"), "idle peer must not rotate the PSK")
+
+	// Active: activity within the window clocks the rotation as usual.
+	dA.OnDataPathRekeyed("bbbb", rotationActivityWindow-1)
+	dB.OnDataPathRekeyed("aaaa", rotationActivityWindow-1)
+	psk2 := wgB.psk("aaaa")
+	require.NotEqual(t, psk1, psk2, "recent activity must clock a rotation")
+	require.Equal(t, psk2, wgA.psk("bbbb"), "both sides converge on the rotated 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 +1,114 @@
+package internal
+
+import (
+	"net/netip"
+	"time"
+
+	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
+	// reoffer re-bootstraps the KEM over Signal for a peer (a fresh signalling offer)
+	// to recover from a persistent data-path rekey failure. Nil disables recovery.
+	reoffer func(remoteKey string)
+}
+
+// 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.Tracef("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 and re-bootstraps the KEM over
+// Signal to recover: a fresh signalling offer starts a new exchange that overwrites the
+// stalled PSK on both sides, resyncing after a persistent data-path desync. The tunnel
+// stays up on the previous PSK meanwhile (the Signal channel is independent of the
+// broken data path).
+func (h pqCallbackHandler) OnRekeyFailed(remoteID pqkem.RemoteID) error {
+	log.Warnf("pqkem: post-quantum rekey failed for peer %s, re-bootstrapping over signal", remoteID)
+	if h.reoffer != nil {
+		h.reoffer(string(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.
+// sinceActivity is how long ago the peer last exchanged real user data; the manager
+// skips rotation for idle tunnels.
+func (p pqHandshaker) OnDataPathRekeyed(remoteKey string, sinceActivity time.Duration) {
+	p.mgr.OnDataPathRekeyed(pqkem.RemoteID(remoteKey), sinceActivity)
+}
+
+// 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/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/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/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
 	}

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

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)

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 == "" {

idp/dex/provider.go

@@ -20,7 +20,6 @@ import (
 	"github.com/dexidp/dex/server"
 	"github.com/dexidp/dex/server/signer"
 	"github.com/dexidp/dex/storage"
-	"github.com/dexidp/dex/storage/sql"
 	"github.com/go-jose/go-jose/v4"
 	"github.com/google/uuid"
 	"github.com/prometheus/client_golang/prometheus"
@@ -41,7 +40,7 @@ type Config struct {
 	GRPCAddr string
 }
 
-const localConnectorID = "local"
+const LocalConnectorID = "local"
 
 // Provider wraps a Dex server
 type Provider struct {
@@ -79,7 +78,7 @@ func NewProvider(ctx context.Context, config *Config) (*Provider, error) {
 
 	// Initialize SQLite storage
 	dbPath := filepath.Join(config.DataDir, "oidc.db")
-	sqliteConfig := &sql.SQLite3{File: dbPath}
+	sqliteConfig := newSQLite3(dbPath)
 	stor, err := sqliteConfig.Open(logger)
 	if err != nil {
 		return nil, fmt.Errorf("failed to open storage: %w", err)
@@ -495,18 +494,60 @@ func (p *Provider) Storage() storage.Storage {
 	return p.storage
 }
 
+// SetClientsMFAChain updates the MFAChain field on OAuth2 clients in Dex storage.
+// Pass a non-empty slice (e.g. []string{"default-totp"}) to enable MFA, or nil to disable it.
+func SetClientsMFAChain(ctx context.Context, st storage.Storage, clientIDs []string, mfaChain []string) error {
+	previousChains := make(map[string][]string, len(clientIDs))
+	for _, clientID := range clientIDs {
+		client, err := st.GetClient(ctx, clientID)
+		if err != nil {
+			return fmt.Errorf("failed to get client %s before MFA chain update: %w", clientID, err)
+		}
+		previousChains[clientID] = cloneMFAChain(client.MFAChain)
+	}
+
+	updatedClientIDs := make([]string, 0, len(clientIDs))
+	for _, clientID := range clientIDs {
+		if err := st.UpdateClient(ctx, clientID, func(old storage.Client) (storage.Client, error) {
+			old.MFAChain = cloneMFAChain(mfaChain)
+			return old, nil
+		}); err != nil {
+			if rollbackErr := rollbackClientsMFAChain(ctx, st, updatedClientIDs, previousChains); rollbackErr != nil {
+				return fmt.Errorf("failed to update MFA chain on client %s: %w (also failed to roll back previous MFA chains: %v)", clientID, err, rollbackErr)
+			}
+			return fmt.Errorf("failed to update MFA chain on client %s: %w", clientID, err)
+		}
+		updatedClientIDs = append(updatedClientIDs, clientID)
+	}
+	return nil
+}
+
+func rollbackClientsMFAChain(ctx context.Context, st storage.Storage, clientIDs []string, previousChains map[string][]string) error {
+	var rollbackErrs []error
+	for i := len(clientIDs) - 1; i >= 0; i-- {
+		clientID := clientIDs[i]
+		previousChain := cloneMFAChain(previousChains[clientID])
+		if err := st.UpdateClient(ctx, clientID, func(old storage.Client) (storage.Client, error) {
+			old.MFAChain = previousChain
+			return old, nil
+		}); err != nil {
+			rollbackErrs = append(rollbackErrs, fmt.Errorf("client %s: %w", clientID, err))
+		}
+	}
+	return errors.Join(rollbackErrs...)
+}
+
+func cloneMFAChain(chain []string) []string {
+	if chain == nil {
+		return nil
+	}
+	return append([]string(nil), chain...)
+}
+
 // SetClientsMFAChain updates the MFAChain field on the dashboard and CLI OAuth2 clients.
 // Pass a non-empty slice (e.g. []string{"default-totp"}) to enable MFA, or nil to disable it.
 func (p *Provider) SetClientsMFAChain(ctx context.Context, clientIDs []string, mfaChain []string) error {
-	for _, clientID := range clientIDs {
-		if err := p.storage.UpdateClient(ctx, clientID, func(old storage.Client) (storage.Client, error) {
-			old.MFAChain = mfaChain
-			return old, nil
-		}); err != nil {
-			return fmt.Errorf("failed to update MFA chain on client %s: %w", clientID, err)
-		}
-	}
-	return nil
+	return SetClientsMFAChain(ctx, p.storage, clientIDs, mfaChain)
 }
 
 // Handler returns the Dex server as an http.Handler for embedding in another server.
@@ -546,7 +587,7 @@ func (p *Provider) CreateUser(ctx context.Context, email, username, password str
 
 	// Encode the user ID in Dex's format: base64(protobuf{user_id, connector_id})
 	// This matches the format Dex uses in JWT tokens
-	encodedID := EncodeDexUserID(userID, localConnectorID)
+	encodedID := EncodeDexUserID(userID, LocalConnectorID)
 	return encodedID, nil
 }
 
@@ -625,7 +666,7 @@ func DecodeDexUserID(encodedID string) (userID, connectorID string, err error) {
 // local password connector.
 func IsLocalUserID(encodedID string) bool {
 	_, connectorID, err := DecodeDexUserID(encodedID)
-	return err == nil && connectorID == localConnectorID
+	return err == nil && connectorID == LocalConnectorID
 }
 
 // GetUser returns a user by email

idp/dex/provider_test.go

@@ -3,6 +3,8 @@ package dex
 import (
 	"context"
 	"encoding/json"
+	"errors"
+	"io"
 	"log/slog"
 	"net/http"
 	"net/http/httptest"
@@ -11,11 +13,44 @@ import (
 	"testing"
 
 	"github.com/dexidp/dex/storage"
+	"github.com/dexidp/dex/storage/memory"
 	sqllib "github.com/dexidp/dex/storage/sql"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
+type updateFailingStorage struct {
+	storage.Storage
+	failClientID string
+}
+
+func (s *updateFailingStorage) UpdateClient(ctx context.Context, id string, updater func(storage.Client) (storage.Client, error)) error {
+	if id == s.failClientID {
+		return errors.New("forced update failure")
+	}
+	return s.Storage.UpdateClient(ctx, id, updater)
+}
+
+func TestSetClientsMFAChainRollsBackUpdatedClients(t *testing.T) {
+	ctx := context.Background()
+	st := memory.New(slog.New(slog.NewTextHandler(io.Discard, nil)))
+
+	require.NoError(t, st.CreateClient(ctx, storage.Client{ID: "client-1", MFAChain: []string{"old-1"}}))
+	require.NoError(t, st.CreateClient(ctx, storage.Client{ID: "client-2", MFAChain: []string{"old-2"}}))
+
+	err := SetClientsMFAChain(ctx, &updateFailingStorage{Storage: st, failClientID: "client-2"}, []string{"client-1", "client-2"}, []string{"new"})
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "failed to update MFA chain on client client-2")
+
+	client1, err := st.GetClient(ctx, "client-1")
+	require.NoError(t, err)
+	require.Equal(t, []string{"old-1"}, client1.MFAChain)
+
+	client2, err := st.GetClient(ctx, "client-2")
+	require.NoError(t, err)
+	require.Equal(t, []string{"old-2"}, client2.MFAChain)
+}
+
 func TestUserCreationFlow(t *testing.T) {
 	ctx := context.Background()
 

idp/dex/sqlite_cgo.go

@@ -0,0 +1,15 @@
+//go:build cgo
+
+package dex
+
+import (
+	sql "github.com/dexidp/dex/storage/sql"
+)
+
+// newSQLite3 builds the dex SQLite3 config. CGO builds use the upstream
+// struct that takes a File path. Non-CGO builds get an empty stub whose
+// Open() returns the dex "SQLite not available" error — correct behaviour
+// for binaries that can't link sqlite3 (e.g. cross-compiled ARM targets).
+func newSQLite3(file string) *sql.SQLite3 {
+	return &sql.SQLite3{File: file}
+}

idp/dex/sqlite_nocgo.go

@@ -0,0 +1,15 @@
+//go:build !cgo
+
+package dex
+
+import (
+	sql "github.com/dexidp/dex/storage/sql"
+)
+
+// newSQLite3 for non-CGO builds. The dex SQLite3 stub has no fields and its
+// Open() returns an error documenting the missing CGO support — correct
+// behaviour for cross-compiled artefacts that never actually run the
+// embedded IdP. The `file` argument is ignored.
+func newSQLite3(_ string) *sql.SQLite3 {
+	return &sql.SQLite3{}
+}

infrastructure_files/getting-started-enterprise.sh

@@ -234,9 +234,6 @@ init_environment() {
 
   NETBIRD_LICENSE_KEY=$(read_secret "Enter license key (input hidden)")
 
-  GHCR_USERNAME="netbirdExtAccess1"
-  GHCR_TOKEN=$(read_secret "Enter GHCR token (input hidden)")
-
   POSTGRES_USER="netbird"
   POSTGRES_DB="netbird"
   POSTGRES_PASSWORD=$(rand_secret)
@@ -263,10 +260,6 @@ init_environment() {
   install -m 600 /dev/null config.yaml
   render_config_yaml >> config.yaml
 
-  echo "Logging in to ghcr.io ..."
-  printf '%s' "$GHCR_TOKEN" | docker login ghcr.io -u "$GHCR_USERNAME" --password-stdin
-  unset GHCR_TOKEN
-
   echo ""
   echo "Pulling images ..."
   $DOCKER_COMPOSE_COMMAND pull

infrastructure_files/getting-started.sh

@@ -556,7 +556,7 @@ start_services_and_show_instructions() {
       echo "Creating proxy access token..."
       # Use docker exec with bash to run the token command directly
       PROXY_TOKEN=$($DOCKER_COMPOSE_COMMAND exec -T netbird-server \
-        /go/bin/netbird-server token create --name "default-proxy" --config /etc/netbird/config.yaml 2>/dev/null | grep "^Token:" | awk '{print $2}')
+        /go/bin/netbird-server admin token create --name "default-proxy" --config /etc/netbird/config.yaml 2>/dev/null | grep "^Token:" | awk '{print $2}')
 
       if [[ -z "$PROXY_TOKEN" ]]; then
         echo "ERROR: Failed to create proxy token. Check netbird-server logs." > /dev/stderr

infrastructure_files/migrate-to-enterprise.sh

@@ -490,8 +490,6 @@ init_migration() {
   echo ""
   echo "Step 1: Image swap (community → Enterprise). License key required."
   NB_LICENSE_KEY=$(read_secret "  License key")
-  GHCR_USERNAME="netbirdExtAccess1"
-  GHCR_TOKEN=$(read_secret "  GHCR token (input hidden)")
 
   # Step 2 — optional
   echo ""
@@ -588,11 +586,6 @@ apply_changes() {
     fi
   } >> "$ENV_FILE"
 
-  echo ""
-  echo "Logging in to ghcr.io ..."
-  printf '%s' "$GHCR_TOKEN" | docker login ghcr.io -u "$GHCR_USERNAME" --password-stdin
-  unset GHCR_TOKEN
-
   echo ""
   echo "Pulling enterprise images ..."
   $DOCKER_COMPOSE_COMMAND pull

management/cmd/admin.go

@@ -0,0 +1,177 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+	"path"
+	"path/filepath"
+
+	"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/util"
+)
+
+var adminDatadir string
+
+// newAdminCommands creates the admin command tree with management-specific resource openers.
+func newAdminCommands() *cobra.Command {
+	cmd := admincmd.NewCommands(admincmd.Openers{
+		Resources: withAdminResources,
+		Store:     withAdminStoreOnly,
+		IDP:       withAdminIDPOnly,
+	})
+	cmd.PersistentFlags().StringVar(&adminDatadir, "datadir", "", "Override the data directory from config (used for store.db and the default idp.db)")
+	return cmd
+}
+
+func newLegacyTokenCommand() *cobra.Command {
+	cmd := tokencmd.NewCommands(tokencmd.StoreOpener(withAdminStoreOnly))
+	cmd.Deprecated = "use 'admin token' instead"
+	cmd.PersistentFlags().StringVar(&nbconfig.MgmtConfigPath, "config", defaultMgmtConfig, "Netbird config file location")
+	return cmd
+}
+
+// withAdminResources initializes logging, loads config, opens the management store
+// and embedded IdP storage, and calls fn.
+func withAdminResources(cmd *cobra.Command, fn func(ctx context.Context, resources admincmd.Resources) error) error {
+	return withAdminConfig(cmd, true, func(ctx context.Context, config *nbconfig.Config, datadir string) error {
+		managementStore, err := openAdminStore(ctx, config, datadir)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseStore(ctx, managementStore)
+
+		idpStorage, idpStorageFile, err := admincmd.OpenIDPStorage(config)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseIDPStorage(idpStorage)
+
+		eventStore, esErr := openAdminEventStore(ctx, config, datadir)
+		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, false, func(ctx context.Context, config *nbconfig.Config, datadir string) error {
+		managementStore, err := openAdminStore(ctx, config, datadir)
+		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, true, func(ctx context.Context, config *nbconfig.Config, _ string) error {
+		idpStorage, idpStorageFile, err := admincmd.OpenIDPStorage(config)
+		if err != nil {
+			return err
+		}
+		defer admincmd.CloseIDPStorage(idpStorage)
+
+		return fn(ctx, idpStorage, idpStorageFile)
+	})
+}
+
+func withAdminConfig(cmd *cobra.Command, applyIDPDefaults bool, fn func(ctx context.Context, config *nbconfig.Config, datadir string) 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
+
+	config, datadir, err := loadAdminMgmtConfig(ctx, applyIDPDefaults)
+	if err != nil {
+		return fmt.Errorf("load config: %w", err)
+	}
+
+	return fn(ctx, config, datadir)
+}
+
+func loadAdminMgmtConfig(ctx context.Context, applyIDPDefaults bool) (*nbconfig.Config, string, error) {
+	config := &nbconfig.Config{}
+	if _, err := util.ReadJsonWithEnvSub(nbconfig.MgmtConfigPath, config); err != nil {
+		return nil, "", err
+	}
+
+	if applyIDPDefaults {
+		if err := ApplyEmbeddedIdPConfig(ctx, config); err != nil {
+			return nil, "", err
+		}
+	}
+
+	datadir := config.Datadir
+	applyAdminDatadirOverride(config, &datadir)
+	return config, datadir, nil
+}
+
+func applyAdminDatadirOverride(config *nbconfig.Config, datadir *string) {
+	if adminDatadir == "" {
+		return
+	}
+
+	oldDatadir := *datadir
+	*datadir = adminDatadir
+	if config.EmbeddedIdP != nil && config.EmbeddedIdP.Storage.Type == "sqlite3" && isDefaultIDPStorageFile(config.EmbeddedIdP.Storage.Config.File, oldDatadir) {
+		config.EmbeddedIdP.Storage.Config.File = filepath.Join(*datadir, "idp.db")
+	}
+}
+
+func isDefaultIDPStorageFile(file, datadir string) bool {
+	if file == "" {
+		return true
+	}
+	defaultFile := filepath.Join(datadir, "idp.db")
+	legacyDefaultFile := path.Join(datadir, "idp.db")
+	legacySlashDefaultFile := path.Join(filepath.ToSlash(datadir), "idp.db")
+	return filepath.Clean(file) == filepath.Clean(defaultFile) ||
+		file == legacyDefaultFile ||
+		filepath.ToSlash(file) == legacySlashDefaultFile
+}
+
+func openAdminStore(ctx context.Context, config *nbconfig.Config, datadir string) (store.Store, error) {
+	managementStore, err := store.NewStore(ctx, config.StoreConfig.Engine, datadir, nil, true)
+	if err != nil {
+		return nil, fmt.Errorf("create store: %w", err)
+	}
+	return managementStore, nil
+}
+
+func openAdminEventStore(ctx context.Context, config *nbconfig.Config, datadir string) (activity.Store, error) {
+	if config.DataStoreEncryptionKey == "" {
+		return nil, fmt.Errorf("data store encryption key is not configured")
+	}
+	eventStore, err := activitystore.NewSqlStore(ctx, 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
+}

management/cmd/admin/admin.go

@@ -0,0 +1,577 @@
+// Package admincmd provides reusable cobra commands for self-hosted administrator helpers.
+// Both the management and combined binaries use these commands, each providing
+// their own opener to handle config loading and storage initialization.
+package admincmd
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/dexidp/dex/storage"
+	log "github.com/sirupsen/logrus"
+	"github.com/spf13/cobra"
+	"golang.org/x/crypto/bcrypt"
+
+	"github.com/netbirdio/netbird/formatter/hook"
+	nbdex "github.com/netbirdio/netbird/idp/dex"
+	"github.com/netbirdio/netbird/management/cmd/proxy"
+	"github.com/netbirdio/netbird/management/cmd/token"
+	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+	"github.com/netbirdio/netbird/management/server"
+	"github.com/netbirdio/netbird/management/server/activity"
+	"github.com/netbirdio/netbird/management/server/idp"
+	"github.com/netbirdio/netbird/management/server/store"
+	"github.com/netbirdio/netbird/management/server/types"
+)
+
+// Resources contains the storages required by the admin commands.
+type Resources struct {
+	Store          store.Store
+	IDPStorage     storage.Storage
+	IDPStorageFile string
+	EventStore     activity.Store
+}
+
+// Opener initializes command resources from the command context and calls fn.
+type Opener func(cmd *cobra.Command, fn func(ctx context.Context, resources Resources) error) error
+
+// StoreOpener initializes only the management store from the command context and calls fn.
+type StoreOpener func(cmd *cobra.Command, fn func(ctx context.Context, s store.Store) error) error
+
+// IDPOpener initializes only the embedded IdP storage from the command context and calls fn.
+type IDPOpener func(cmd *cobra.Command, fn func(ctx context.Context, idpStorage storage.Storage, storageFile string) error) error
+
+// Openers contains the resource openers needed by the admin command tree.
+type Openers struct {
+	Resources Opener
+	Store     StoreOpener
+	IDP       IDPOpener
+}
+
+type userSelector struct {
+	email  string
+	userID string
+}
+
+func (s userSelector) normalized() userSelector {
+	return userSelector{
+		email:  strings.TrimSpace(s.email),
+		userID: strings.TrimSpace(s.userID),
+	}
+}
+
+func (s userSelector) validate() error {
+	s = s.normalized()
+	if (s.email == "") == (s.userID == "") {
+		return fmt.Errorf("provide exactly one of --email or --user-id")
+	}
+	return nil
+}
+
+// NewCommands creates the admin command tree with the given resource openers.
+func NewCommands(openers Openers) *cobra.Command {
+	adminCmd := &cobra.Command{
+		Use:   "admin",
+		Short: "Self-hosted administrator helpers",
+		Long:  "Administrative helpers for self-hosted deployments using the embedded identity provider.",
+	}
+
+	userCmd := &cobra.Command{
+		Use:   "user",
+		Short: "Manage local embedded IdP users",
+	}
+
+	var passwordSelector userSelector
+	var password string
+	var passwordFile string
+	passwordCmd := &cobra.Command{
+		Use:     "change-password (--email email | --user-id id) (--password password | --password-file path)",
+		Aliases: []string{"set-password"},
+		Short:   "Change a local user's password",
+		Args:    cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			if err := passwordSelector.validate(); err != nil {
+				return err
+			}
+			newPassword, err := resolvePasswordInput(cmd, password, passwordFile)
+			if err != nil {
+				return err
+			}
+			return openers.IDP(cmd, func(ctx context.Context, idpStorage storage.Storage, storageFile string) error {
+				return runChangePassword(ctx, idpStorage, cmd.OutOrStdout(), passwordSelector, newPassword, storageFile)
+			})
+		},
+	}
+	addUserSelectorFlags(passwordCmd, &passwordSelector)
+	passwordCmd.Flags().StringVar(&password, "password", "", "New password for the user")
+	passwordCmd.Flags().StringVar(&passwordFile, "password-file", "", "Read new password from file ('-' for stdin)")
+
+	var resetSelector userSelector
+	resetMFACmd := &cobra.Command{
+		Use:   "reset-mfa (--email email | --user-id id)",
+		Short: "Reset a local user's MFA enrollment",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			if err := resetSelector.validate(); err != nil {
+				return err
+			}
+			return openers.IDP(cmd, func(ctx context.Context, idpStorage storage.Storage, storageFile string) error {
+				return runResetMFA(ctx, idpStorage, cmd.OutOrStdout(), resetSelector, storageFile)
+			})
+		},
+	}
+	addUserSelectorFlags(resetMFACmd, &resetSelector)
+
+	userCmd.AddCommand(passwordCmd, resetMFACmd)
+
+	mfaCmd := &cobra.Command{
+		Use:   "mfa",
+		Short: "Manage local MFA for embedded IdP users",
+	}
+
+	enableCmd := &cobra.Command{
+		Use:   "enable",
+		Short: "Enable MFA for local embedded IdP users",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			return openers.Resources(cmd, func(ctx context.Context, resources Resources) error {
+				return runSetMFAEnabled(ctx, resources, cmd.OutOrStdout(), true)
+			})
+		},
+	}
+
+	disableCmd := &cobra.Command{
+		Use:   "disable",
+		Short: "Disable MFA for local embedded IdP users",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			return openers.Resources(cmd, func(ctx context.Context, resources Resources) error {
+				return runSetMFAEnabled(ctx, resources, cmd.OutOrStdout(), false)
+			})
+		},
+	}
+
+	statusCmd := &cobra.Command{
+		Use:   "status",
+		Short: "Show local MFA status",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			return openers.Resources(cmd, func(ctx context.Context, resources Resources) error {
+				return runMFAStatus(ctx, resources, cmd.OutOrStdout())
+			})
+		},
+	}
+
+	mfaCmd.AddCommand(enableCmd, disableCmd, statusCmd)
+	adminCmd.AddCommand(userCmd, mfaCmd)
+	if openers.Store != nil {
+		adminCmd.AddCommand(tokencmd.NewCommands(tokencmd.StoreOpener(openers.Store)))
+		adminCmd.AddCommand(proxycmd.NewCommands(proxycmd.StoreOpener(openers.Store)))
+	}
+	return adminCmd
+}
+
+// OpenEmbeddedIDPStorage opens the Dex storage configured for the embedded IdP.
+func OpenEmbeddedIDPStorage(cfg *idp.EmbeddedIdPConfig) (storage.Storage, error) {
+	if cfg == nil || !cfg.Enabled {
+		return nil, fmt.Errorf("admin commands require the embedded IdP to be enabled")
+	}
+
+	yamlConfig, err := cfg.ToYAMLConfig()
+	if err != nil {
+		return nil, fmt.Errorf("build embedded IdP config: %w", err)
+	}
+
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+	st, err := yamlConfig.Storage.OpenStorage(logger)
+	if err != nil {
+		return nil, fmt.Errorf("open embedded IdP storage: %w", err)
+	}
+	return st, nil
+}
+
+// CloseStore closes the management store and logs cleanup errors at debug level.
+func CloseStore(ctx context.Context, s store.Store) {
+	if s == nil {
+		return
+	}
+	if err := s.Close(ctx); err != nil {
+		log.Debugf("close store: %v", err)
+	}
+}
+
+// OpenIDPStorage opens embedded IdP storage and returns its sqlite file path when applicable.
+func OpenIDPStorage(config *nbconfig.Config) (storage.Storage, string, error) {
+	if config == nil {
+		return nil, "", fmt.Errorf("management config is required")
+	}
+	idpStorage, err := OpenEmbeddedIDPStorage(config.EmbeddedIdP)
+	if err != nil {
+		return nil, "", err
+	}
+	return idpStorage, embeddedIDPStorageFile(config), nil
+}
+
+func embeddedIDPStorageFile(config *nbconfig.Config) string {
+	if config.EmbeddedIdP == nil || config.EmbeddedIdP.Storage.Type != "sqlite3" {
+		return ""
+	}
+	return config.EmbeddedIdP.Storage.Config.File
+}
+
+// CloseIDPStorage closes embedded IdP storage and logs cleanup errors at debug level.
+func CloseIDPStorage(s storage.Storage) {
+	if s == nil {
+		return
+	}
+	if err := s.Close(); err != nil {
+		log.Debugf("close embedded IdP storage: %v", err)
+	}
+}
+
+func addUserSelectorFlags(cmd *cobra.Command, selector *userSelector) {
+	cmd.Flags().StringVar(&selector.email, "email", "", "User email")
+	cmd.Flags().StringVar(&selector.userID, "user-id", "", "User ID")
+}
+
+func resolvePasswordInput(cmd *cobra.Command, password, passwordFile string) (string, error) {
+	if password != "" && passwordFile != "" {
+		return "", fmt.Errorf("provide only one of --password or --password-file")
+	}
+	if passwordFile == "" {
+		return password, nil
+	}
+
+	var data []byte
+	var err error
+	if passwordFile == "-" {
+		data, err = io.ReadAll(cmd.InOrStdin())
+	} else {
+		data, err = os.ReadFile(passwordFile)
+	}
+	if err != nil {
+		return "", fmt.Errorf("read password: %w", err)
+	}
+	return strings.TrimRight(string(data), "\r\n"), nil
+}
+
+func runChangePassword(ctx context.Context, idpStorage storage.Storage, w io.Writer, selector userSelector, password string, idpStorageFile string) error {
+	if idpStorage == nil {
+		return fmt.Errorf("embedded IdP storage is required")
+	}
+	selector = selector.normalized()
+	if err := selector.validate(); err != nil {
+		return err
+	}
+	if password == "" {
+		return fmt.Errorf("password is required")
+	}
+	if err := server.ValidatePassword(password); err != nil {
+		return fmt.Errorf("invalid password: %w", err)
+	}
+
+	user, err := findLocalUser(ctx, idpStorage, selector, idpStorageFile)
+	if err != nil {
+		return err
+	}
+
+	hash, err := bcrypt.GenerateFromPassword([]byte(password), bcrypt.DefaultCost)
+	if err != nil {
+		return fmt.Errorf("hash password: %w", err)
+	}
+
+	if err := idpStorage.UpdatePassword(ctx, user.Email, func(old storage.Password) (storage.Password, error) {
+		old.Hash = hash
+		return old, nil
+	}); err != nil {
+		return fmt.Errorf("update password for %s: %w", user.Email, err)
+	}
+
+	if err := deleteLocalAuthSession(ctx, idpStorage, user.UserID); err != nil {
+		return err
+	}
+
+	_, _ = fmt.Fprintf(w, "Password updated for %s.\n", user.Email)
+	return nil
+}
+
+func runResetMFA(ctx context.Context, idpStorage storage.Storage, w io.Writer, selector userSelector, idpStorageFile string) error {
+	if idpStorage == nil {
+		return fmt.Errorf("embedded IdP storage is required")
+	}
+	selector = selector.normalized()
+	if err := selector.validate(); err != nil {
+		return err
+	}
+
+	user, err := findLocalUser(ctx, idpStorage, selector, idpStorageFile)
+	if err != nil {
+		return err
+	}
+
+	reset := false
+	err = idpStorage.UpdateUserIdentity(ctx, user.UserID, idp.LocalConnectorID, func(old storage.UserIdentity) (storage.UserIdentity, error) {
+		reset = reset || len(old.MFASecrets) > 0 || len(old.WebAuthnCredentials) > 0
+		old.MFASecrets = map[string]*storage.MFASecret{}
+		old.WebAuthnCredentials = map[string][]storage.WebAuthnCredential{}
+		return old, nil
+	})
+	if errors.Is(err, storage.ErrNotFound) {
+		if err := deleteLocalAuthSession(ctx, idpStorage, user.UserID); err != nil {
+			return err
+		}
+		_, _ = fmt.Fprintf(w, "No MFA enrollment found for %s.\n", user.Email)
+		return nil
+	}
+	if err != nil {
+		return fmt.Errorf("reset MFA for %s: %w", user.Email, err)
+	}
+
+	if err := deleteLocalAuthSession(ctx, idpStorage, user.UserID); err != nil {
+		return err
+	}
+
+	if reset {
+		_, _ = fmt.Fprintf(w, "MFA reset for %s. The user will re-enroll at next login.\n", user.Email)
+	} else {
+		_, _ = fmt.Fprintf(w, "No MFA enrollment found for %s.\n", user.Email)
+	}
+	return nil
+}
+
+func runSetMFAEnabled(ctx context.Context, resources Resources, w io.Writer, enabled bool) error {
+	if resources.Store == nil {
+		return fmt.Errorf("management store is required")
+	}
+	if resources.IDPStorage == nil {
+		return fmt.Errorf("embedded IdP storage is required")
+	}
+
+	accountID, settings, err := getSingleAccountSettings(ctx, resources.Store)
+	if err != nil {
+		return err
+	}
+
+	oldEnabled := settings.LocalMfaEnabled
+	newSettings := settings.Copy()
+	newSettings.LocalMfaEnabled = enabled
+
+	if err := setIDPClientsMFA(ctx, resources.IDPStorage, enabled); err != nil {
+		return err
+	}
+
+	if err := resources.Store.SaveAccountSettings(ctx, accountID, newSettings); err != nil {
+		if rollbackErr := setIDPClientsMFA(ctx, resources.IDPStorage, oldEnabled); rollbackErr != nil {
+			return fmt.Errorf("save local MFA account setting: %w (also failed to roll back embedded IdP MFA state: %v)", err, rollbackErr)
+		}
+		return fmt.Errorf("save local MFA account setting: %w", err)
+	}
+
+	if err := storeMFAActivity(ctx, resources.EventStore, accountID, enabled); err != nil {
+		_, _ = fmt.Fprintf(w, "Warning: failed to record audit event: %v\n", err)
+	}
+
+	state := "disabled"
+	if enabled {
+		state = "enabled"
+	}
+	_, _ = fmt.Fprintf(w, "Local MFA %s.\n", state)
+	return nil
+}
+
+func runMFAStatus(ctx context.Context, resources Resources, w io.Writer) error {
+	if resources.Store == nil {
+		return fmt.Errorf("management store is required")
+	}
+	if resources.IDPStorage == nil {
+		return fmt.Errorf("embedded IdP storage is required")
+	}
+
+	_, settings, err := getSingleAccountSettings(ctx, resources.Store)
+	if err != nil {
+		return err
+	}
+	accountStatus := "disabled"
+	if settings.LocalMfaEnabled {
+		accountStatus = "enabled"
+	}
+
+	clientStatus, err := idpClientsMFAStatus(ctx, resources.IDPStorage)
+	if err != nil {
+		return err
+	}
+
+	_, _ = fmt.Fprintf(w, "Account setting: %s\n", accountStatus)
+	_, _ = fmt.Fprintf(w, "Embedded IdP clients: %s\n", clientStatus)
+	return nil
+}
+
+func getSingleAccountSettings(ctx context.Context, s store.Store) (string, *types.Settings, error) {
+	count, err := s.GetAccountsCounter(ctx)
+	if err != nil {
+		return "", nil, fmt.Errorf("count accounts: %w", err)
+	}
+	if count != 1 {
+		return "", nil, fmt.Errorf("expected exactly one account, got %d; local MFA is supported only in single-account embedded IdP deployments", count)
+	}
+
+	accountID, err := s.GetAnyAccountID(ctx)
+	if err != nil {
+		return "", nil, fmt.Errorf("get account ID: %w", err)
+	}
+
+	settings, err := s.GetAccountSettings(ctx, store.LockingStrengthNone, accountID)
+	if err != nil {
+		return "", nil, fmt.Errorf("get account settings: %w", err)
+	}
+	if settings == nil {
+		settings = &types.Settings{}
+	}
+	return accountID, settings, nil
+}
+
+func storeMFAActivity(ctx context.Context, eventStore activity.Store, accountID string, enabled bool) error {
+	if eventStore == nil {
+		return nil
+	}
+	event := activity.AccountLocalMfaDisabled
+	if enabled {
+		event = activity.AccountLocalMfaEnabled
+	}
+	_, err := eventStore.Save(ctx, &activity.Event{
+		Timestamp:   time.Now().UTC(),
+		Activity:    event,
+		InitiatorID: string(hook.SystemSource),
+		TargetID:    accountID,
+		AccountID:   accountID,
+	})
+	if err != nil {
+		return fmt.Errorf("save local MFA audit event: %w", err)
+	}
+	return nil
+}
+
+func findLocalUser(ctx context.Context, idpStorage storage.Storage, selector userSelector, idpStorageFile string) (storage.Password, error) {
+	selector = selector.normalized()
+	if err := selector.validate(); err != nil {
+		return storage.Password{}, err
+	}
+
+	if selector.email != "" {
+		user, err := idpStorage.GetPassword(ctx, selector.email)
+		if errors.Is(err, storage.ErrNotFound) {
+			if empty, listErr := localUsersEmpty(ctx, idpStorage); listErr != nil {
+				return storage.Password{}, listErr
+			} else if empty {
+				return storage.Password{}, noLocalUsersError(idpStorageFile)
+			}
+			return storage.Password{}, fmt.Errorf("local user with email %q not found", selector.email)
+		}
+		if err != nil {
+			return storage.Password{}, fmt.Errorf("get local user by email %q: %w", selector.email, err)
+		}
+		return user, nil
+	}
+
+	rawUserID := selector.userID
+	if decodedUserID, _, err := nbdex.DecodeDexUserID(selector.userID); err == nil && decodedUserID != "" {
+		rawUserID = decodedUserID
+	}
+
+	users, err := idpStorage.ListPasswords(ctx)
+	if err != nil {
+		return storage.Password{}, fmt.Errorf("list local users: %w", err)
+	}
+	for _, user := range users {
+		if user.UserID == rawUserID || user.UserID == selector.userID {
+			return user, nil
+		}
+	}
+
+	if len(users) == 0 {
+		return storage.Password{}, noLocalUsersError(idpStorageFile)
+	}
+
+	return storage.Password{}, fmt.Errorf("local user with ID %q not found", selector.userID)
+}
+
+func localUsersEmpty(ctx context.Context, idpStorage storage.Storage) (bool, error) {
+	users, err := idpStorage.ListPasswords(ctx)
+	if err != nil {
+		return false, fmt.Errorf("list local users: %w", err)
+	}
+	return len(users) == 0, nil
+}
+
+func noLocalUsersError(idpStorageFile string) error {
+	location := ""
+	if idpStorageFile != "" {
+		location = fmt.Sprintf(" (%s)", idpStorageFile)
+	}
+	return fmt.Errorf("no local users exist in the embedded IdP storage%s; the management server may never have started with this config, or --datadir points at the wrong location", location)
+}
+
+func deleteLocalAuthSession(ctx context.Context, idpStorage storage.Storage, userID string) error {
+	err := idpStorage.DeleteAuthSession(ctx, userID, idp.LocalConnectorID)
+	if err == nil || errors.Is(err, storage.ErrNotFound) {
+		return nil
+	}
+	return fmt.Errorf("delete local auth session for user %s: %w", userID, err)
+}
+
+func setIDPClientsMFA(ctx context.Context, idpStorage storage.Storage, enabled bool) error {
+	var mfaChain []string
+	if enabled {
+		mfaChain = []string{idp.DefaultTOTPAuthenticatorID}
+	}
+
+	clientIDs := []string{idp.StaticClientCLI, idp.StaticClientDashboard}
+	if err := nbdex.SetClientsMFAChain(ctx, idpStorage, clientIDs, mfaChain); err != nil {
+		if errors.Is(err, storage.ErrNotFound) {
+			return fmt.Errorf("embedded IdP client not found; start the management server once before toggling MFA: %w", err)
+		}
+		return fmt.Errorf("update MFA chain on embedded IdP clients: %w", err)
+	}
+	return nil
+}
+
+func idpClientsMFAStatus(ctx context.Context, idpStorage storage.Storage) (string, error) {
+	clientIDs := []string{idp.StaticClientCLI, idp.StaticClientDashboard}
+	enabledCount := 0
+	for _, clientID := range clientIDs {
+		client, err := idpStorage.GetClient(ctx, clientID)
+		if errors.Is(err, storage.ErrNotFound) {
+			return "unknown", fmt.Errorf("embedded IdP client %q not found", clientID)
+		}
+		if err != nil {
+			return "unknown", fmt.Errorf("get embedded IdP client %q: %w", clientID, err)
+		}
+		if hasAuthenticator(client.MFAChain, idp.DefaultTOTPAuthenticatorID) {
+			enabledCount++
+		}
+	}
+
+	switch enabledCount {
+	case 0:
+		return "disabled", nil
+	case len(clientIDs):
+		return "enabled", nil
+	default:
+		return "partially enabled", nil
+	}
+}
+
+func hasAuthenticator(chain []string, authenticatorID string) bool {
+	for _, id := range chain {
+		if id == authenticatorID {
+			return true
+		}
+	}
+	return false
+}

management/cmd/admin/admin_test.go

@@ -0,0 +1,250 @@
+package admincmd
+
+import (
+	"bytes"
+	"context"
+	"io"
+	"log/slog"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/dexidp/dex/storage"
+	"github.com/dexidp/dex/storage/memory"
+	"github.com/spf13/cobra"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/crypto/bcrypt"
+
+	nbdex "github.com/netbirdio/netbird/idp/dex"
+	"github.com/netbirdio/netbird/management/server/idp"
+	mgmtstore "github.com/netbirdio/netbird/management/server/store"
+	"github.com/netbirdio/netbird/management/server/types"
+)
+
+func newTestIDPStorage(t *testing.T) storage.Storage {
+	t.Helper()
+
+	st := memory.New(slog.New(slog.NewTextHandler(io.Discard, nil)))
+	hash, err := bcrypt.GenerateFromPassword([]byte("OldPass1!"), bcrypt.DefaultCost)
+	require.NoError(t, err)
+
+	require.NoError(t, st.CreatePassword(context.Background(), storage.Password{
+		Email:    "user@example.com",
+		Username: "User",
+		UserID:   "user-1",
+		Hash:     hash,
+	}))
+	require.NoError(t, st.CreateUserIdentity(context.Background(), storage.UserIdentity{
+		UserID:      "user-1",
+		ConnectorID: idp.LocalConnectorID,
+		MFASecrets: map[string]*storage.MFASecret{
+			idp.DefaultTOTPAuthenticatorID: {
+				AuthenticatorID: idp.DefaultTOTPAuthenticatorID,
+				Type:            "TOTP",
+				Secret:          "otpauth://totp/NetBird:user@example.com?secret=ABC",
+				Confirmed:       true,
+				CreatedAt:       time.Now(),
+			},
+		},
+		WebAuthnCredentials: map[string][]storage.WebAuthnCredential{
+			"webauthn": {{CredentialID: []byte("credential")}},
+		},
+	}))
+	require.NoError(t, st.CreateAuthSession(context.Background(), storage.AuthSession{
+		UserID:      "user-1",
+		ConnectorID: idp.LocalConnectorID,
+		Nonce:       "nonce",
+	}))
+	require.NoError(t, st.CreateClient(context.Background(), storage.Client{ID: idp.StaticClientCLI, Name: "CLI"}))
+	require.NoError(t, st.CreateClient(context.Background(), storage.Client{ID: idp.StaticClientDashboard, Name: "Dashboard"}))
+
+	return st
+}
+
+func TestRunChangePassword(t *testing.T) {
+	ctx := context.Background()
+	st := newTestIDPStorage(t)
+	var out bytes.Buffer
+
+	err := runChangePassword(ctx, st, &out, userSelector{email: "user@example.com"}, "NewPass1!", "")
+	require.NoError(t, err)
+	require.Contains(t, out.String(), "Password updated")
+
+	user, err := st.GetPassword(ctx, "user@example.com")
+	require.NoError(t, err)
+	require.NoError(t, bcrypt.CompareHashAndPassword(user.Hash, []byte("NewPass1!")))
+
+	_, err = st.GetAuthSession(ctx, "user-1", idp.LocalConnectorID)
+	require.ErrorIs(t, err, storage.ErrNotFound)
+}
+
+func TestRunChangePasswordValidatesPassword(t *testing.T) {
+	st := newTestIDPStorage(t)
+	err := runChangePassword(context.Background(), st, io.Discard, userSelector{email: "user@example.com"}, "short", "")
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "invalid password")
+}
+
+func TestRunResetMFA(t *testing.T) {
+	ctx := context.Background()
+	st := newTestIDPStorage(t)
+	var out bytes.Buffer
+
+	encodedUserID := nbdex.EncodeDexUserID("user-1", idp.LocalConnectorID)
+	err := runResetMFA(ctx, st, &out, userSelector{userID: encodedUserID}, "")
+	require.NoError(t, err)
+	require.Contains(t, out.String(), "MFA reset")
+
+	identity, err := st.GetUserIdentity(ctx, "user-1", idp.LocalConnectorID)
+	require.NoError(t, err)
+	require.Empty(t, identity.MFASecrets)
+	require.Empty(t, identity.WebAuthnCredentials)
+
+	_, err = st.GetAuthSession(ctx, "user-1", idp.LocalConnectorID)
+	require.ErrorIs(t, err, storage.ErrNotFound)
+}
+
+func TestRunResetMFAWithoutEnrollment(t *testing.T) {
+	ctx := context.Background()
+	st := newTestIDPStorage(t)
+	require.NoError(t, st.UpdateUserIdentity(ctx, "user-1", idp.LocalConnectorID, func(old storage.UserIdentity) (storage.UserIdentity, error) {
+		old.MFASecrets = nil
+		old.WebAuthnCredentials = nil
+		return old, nil
+	}))
+
+	var out bytes.Buffer
+	err := runResetMFA(ctx, st, &out, userSelector{email: "user@example.com"}, "")
+	require.NoError(t, err)
+	require.Contains(t, out.String(), "No MFA enrollment found")
+}
+
+func TestSetIDPClientsMFA(t *testing.T) {
+	ctx := context.Background()
+	st := newTestIDPStorage(t)
+
+	require.NoError(t, setIDPClientsMFA(ctx, st, true))
+	status, err := idpClientsMFAStatus(ctx, st)
+	require.NoError(t, err)
+	require.Equal(t, "enabled", status)
+
+	require.NoError(t, setIDPClientsMFA(ctx, st, false))
+	status, err = idpClientsMFAStatus(ctx, st)
+	require.NoError(t, err)
+	require.Equal(t, "disabled", status)
+}
+
+func newTestManagementStore(t *testing.T, localMFAEnabled bool) mgmtstore.Store {
+	t.Helper()
+	ctx := context.Background()
+	st, err := mgmtstore.NewStore(ctx, types.SqliteStoreEngine, t.TempDir(), nil, false)
+	require.NoError(t, err)
+	t.Cleanup(func() { require.NoError(t, st.Close(ctx)) })
+	require.NoError(t, st.SaveAccount(ctx, &types.Account{
+		Id:       "account-1",
+		Settings: &types.Settings{LocalMfaEnabled: localMFAEnabled},
+	}))
+	return st
+}
+
+func TestRunSetMFAEnabledDoesNotSaveWhenIDPUpdateFails(t *testing.T) {
+	ctx := context.Background()
+	managementStore := newTestManagementStore(t, false)
+	idpStorage := memory.New(slog.New(slog.NewTextHandler(io.Discard, nil)))
+
+	err := runSetMFAEnabled(ctx, Resources{Store: managementStore, IDPStorage: idpStorage}, io.Discard, true)
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "embedded IdP client")
+
+	settings, err := managementStore.GetAccountSettings(ctx, mgmtstore.LockingStrengthNone, "account-1")
+	require.NoError(t, err)
+	require.False(t, settings.LocalMfaEnabled)
+}
+
+func TestRunSetMFAEnabledUpdatesSettingsAfterIDP(t *testing.T) {
+	ctx := context.Background()
+	managementStore := newTestManagementStore(t, false)
+	idpStorage := newTestIDPStorage(t)
+
+	err := runSetMFAEnabled(ctx, Resources{Store: managementStore, IDPStorage: idpStorage}, io.Discard, true)
+	require.NoError(t, err)
+
+	settings, err := managementStore.GetAccountSettings(ctx, mgmtstore.LockingStrengthNone, "account-1")
+	require.NoError(t, err)
+	require.True(t, settings.LocalMfaEnabled)
+	clientStatus, err := idpClientsMFAStatus(ctx, idpStorage)
+	require.NoError(t, err)
+	require.Equal(t, "enabled", clientStatus)
+}
+
+func TestRunSetMFAEnabledSucceedsWithNilEventStore(t *testing.T) {
+	ctx := context.Background()
+	managementStore := newTestManagementStore(t, false)
+	idpStorage := newTestIDPStorage(t)
+	var out bytes.Buffer
+	var err error
+
+	require.NotPanics(t, func() {
+		err = runSetMFAEnabled(ctx, Resources{Store: managementStore, IDPStorage: idpStorage, EventStore: nil}, &out, true)
+	})
+	require.NoError(t, err)
+	require.Contains(t, out.String(), "Local MFA enabled")
+
+	settings, err := managementStore.GetAccountSettings(ctx, mgmtstore.LockingStrengthNone, "account-1")
+	require.NoError(t, err)
+	require.True(t, settings.LocalMfaEnabled)
+}
+
+func TestUserSelectorValidate(t *testing.T) {
+	require.NoError(t, userSelector{email: " user@example.com "}.validate())
+	require.NoError(t, userSelector{userID: "user-1"}.validate())
+	require.Error(t, userSelector{}.validate())
+	require.Error(t, userSelector{email: "user@example.com", userID: "user-1"}.validate())
+}
+
+func TestFindLocalUserNotFound(t *testing.T) {
+	st := newTestIDPStorage(t)
+	_, err := findLocalUser(context.Background(), st, userSelector{email: "missing@example.com"}, "")
+	require.Error(t, err)
+	require.True(t, strings.Contains(err.Error(), "not found"))
+}
+
+func TestFindLocalUserZeroUsersIncludesStoragePath(t *testing.T) {
+	st := memory.New(slog.New(slog.NewTextHandler(io.Discard, nil)))
+	_, err := findLocalUser(context.Background(), st, userSelector{email: "missing@example.com"}, "/var/lib/netbird/idp.db")
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "no local users exist")
+	require.Contains(t, err.Error(), "/var/lib/netbird/idp.db")
+}
+
+func TestUserCommandValidatesSelectorBeforeOpeningStorage(t *testing.T) {
+	opened := false
+	cmd := NewCommands(Openers{
+		IDP: func(cmd *cobra.Command, fn func(ctx context.Context, idpStorage storage.Storage, storageFile string) error) error {
+			opened = true
+			return nil
+		},
+	})
+	cmd.SetArgs([]string{"user", "change-password", "--password", "NewPass1!"})
+	cmd.SetOut(io.Discard)
+	cmd.SetErr(io.Discard)
+
+	err := cmd.Execute()
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "provide exactly one")
+	require.False(t, opened)
+}
+
+func TestResolvePasswordInputFromStdin(t *testing.T) {
+	cmd := &cobra.Command{}
+	cmd.SetIn(strings.NewReader("NewPass1!\n"))
+
+	password, err := resolvePasswordInput(cmd, "", "-")
+	require.NoError(t, err)
+	require.Equal(t, "NewPass1!", password)
+}
+
+func TestResolvePasswordInputRejectsMultipleSources(t *testing.T) {
+	_, err := resolvePasswordInput(&cobra.Command{}, "NewPass1!", "-")
+	require.Error(t, err)
+}

management/cmd/admin_config_test.go

@@ -0,0 +1,80 @@
+package cmd
+
+import (
+	"context"
+	"path"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+
+	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+	"github.com/netbirdio/netbird/management/server/idp"
+)
+
+func TestApplyAdminDatadirOverrideRelocatesDefaultIDPStorage(t *testing.T) {
+	oldDatadir := filepath.Join(t.TempDir(), "old")
+	newDatadir := filepath.Join(t.TempDir(), "new")
+
+	for _, defaultFile := range []string{
+		"",
+		filepath.Join(oldDatadir, "idp.db"),
+		path.Join(oldDatadir, "idp.db"),
+	} {
+		t.Run(defaultFile, func(t *testing.T) {
+			cfg := &nbconfig.Config{
+				EmbeddedIdP: &idp.EmbeddedIdPConfig{
+					Enabled: true,
+					Storage: idp.EmbeddedStorageConfig{
+						Type: "sqlite3",
+						Config: idp.EmbeddedStorageTypeConfig{
+							File: defaultFile,
+						},
+					},
+				},
+			}
+			datadir := oldDatadir
+			oldAdminDatadir := adminDatadir
+			adminDatadir = newDatadir
+			t.Cleanup(func() { adminDatadir = oldAdminDatadir })
+
+			applyAdminDatadirOverride(cfg, &datadir)
+
+			require.Equal(t, newDatadir, datadir)
+			require.Equal(t, filepath.Join(newDatadir, "idp.db"), cfg.EmbeddedIdP.Storage.Config.File)
+		})
+	}
+}
+
+func TestOpenAdminEventStoreMissingEncryptionKeyReturnsNilInterface(t *testing.T) {
+	eventStore, err := openAdminEventStore(context.Background(), &nbconfig.Config{}, t.TempDir())
+	require.Error(t, err)
+	require.Contains(t, err.Error(), "encryption key")
+	require.Nil(t, eventStore)
+}
+
+func TestApplyAdminDatadirOverrideKeepsExplicitIDPStorage(t *testing.T) {
+	oldDatadir := filepath.Join(t.TempDir(), "old")
+	newDatadir := filepath.Join(t.TempDir(), "new")
+	explicitFile := filepath.Join(t.TempDir(), "custom-idp.db")
+	cfg := &nbconfig.Config{
+		EmbeddedIdP: &idp.EmbeddedIdPConfig{
+			Enabled: true,
+			Storage: idp.EmbeddedStorageConfig{
+				Type: "sqlite3",
+				Config: idp.EmbeddedStorageTypeConfig{
+					File: explicitFile,
+				},
+			},
+		},
+	}
+	datadir := oldDatadir
+	oldAdminDatadir := adminDatadir
+	adminDatadir = newDatadir
+	t.Cleanup(func() { adminDatadir = oldAdminDatadir })
+
+	applyAdminDatadirOverride(cfg, &datadir)
+
+	require.Equal(t, newDatadir, datadir)
+	require.Equal(t, explicitFile, cfg.EmbeddedIdP.Storage.Config.File)
+}

management/cmd/management.go

@@ -13,6 +13,7 @@ import (
 	"os"
 	"os/signal"
 	"path"
+	"path/filepath"
 	"strings"
 	"syscall"
 
@@ -25,6 +26,7 @@ import (
 	"github.com/netbirdio/netbird/management/internals/server"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
 	nbdomain "github.com/netbirdio/netbird/shared/management/domain"
+	"github.com/netbirdio/netbird/shared/management/grpc"
 	"github.com/netbirdio/netbird/util"
 	"github.com/netbirdio/netbird/util/crypt"
 )
@@ -153,8 +155,20 @@ func LoadMgmtConfig(ctx context.Context, mgmtConfigPath string) (*nbconfig.Confi
 
 	ApplyCommandLineOverrides(loadedConfig)
 
+	err := grpc.ValidateSyncMessageVersion(loadedConfig.HighestSupportedSyncMessageVersion)
+	if err != nil {
+		return nil, err
+	}
+
+	for account, version := range loadedConfig.PerAccountHighestSupportedSyncMessageVersion {
+		err := grpc.ValidateSyncMessageVersion(&version)
+		if err != nil {
+			return nil, fmt.Errorf("unrecognized sync message version for account %s, %w", account, err)
+		}
+	}
+
 	// Apply EmbeddedIdP config to HttpConfig if embedded IdP is enabled
-	err := ApplyEmbeddedIdPConfig(ctx, loadedConfig)
+	err = ApplyEmbeddedIdPConfig(ctx, loadedConfig)
 	if err != nil {
 		return nil, err
 	}
@@ -209,7 +223,7 @@ func ApplyEmbeddedIdPConfig(ctx context.Context, cfg *nbconfig.Config) error {
 		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

management/cmd/management_test.go

@@ -4,6 +4,9 @@ import (
 	"context"
 	"os"
 	"testing"
+
+	"github.com/netbirdio/netbird/shared/management/grpc"
+	"github.com/stretchr/testify/assert"
 )
 
 const (
@@ -20,34 +23,49 @@ const (
 		"AuthAudience": "https://stageapp/",
 		"AuthIssuer": "https://something.eu.auth0.com/",
 		"OIDCConfigEndpoint": "https://something.eu.auth0.com/.well-known/openid-configuration"
+	  },
+	  "HighestSupportedSyncMessageVersion": 1,
+	  "PerAccountHighestSupportedSyncMessageVersion": {
+	    "1": 0,
+		"2": 1
 	  }
 	}`
 )
 
-func Test_loadMgmtConfig(t *testing.T) {
-	tmpFile, err := createConfig()
-	if err != nil {
-		t.Fatalf("failed to create config: %s", err)
-	}
+func Test_LoadMgmtConfig(t *testing.T) {
+	tmpFile, err := createConfig(exampleConfig)
+	assert.NoError(t, err)
 
 	cfg, err := LoadMgmtConfig(context.Background(), tmpFile)
-	if err != nil {
-		t.Fatalf("failed to load management config: %s", err)
-	}
-	if cfg.Relay == nil {
-		t.Fatalf("config is nil")
-	}
-	if len(cfg.Relay.Addresses) == 0 {
-		t.Fatalf("relay address is empty")
-	}
+	assert.NoError(t, err)
+	assert.NotEmpty(t, cfg.Relay)
+	assert.NotEmpty(t, cfg.Relay.Addresses)
+	assert.Equal(t, int(grpc.ComponentNetworkMap), *cfg.HighestSupportedSyncMessageVersion)
+	assert.Equal(t, map[string]int{"1": int(grpc.Base), "2": int(grpc.ComponentNetworkMap)}, cfg.PerAccountHighestSupportedSyncMessageVersion)
 }
 
-func createConfig() (string, error) {
+func Test_LoadMgmtConfig_Empty(t *testing.T) {
+	tmpFile, err := createConfig(`{
+		"HttpConfig": {
+			"AuthAudience": "https://stageapp/",
+			"AuthIssuer": "https://something.eu.auth0.com/",
+			"OIDCConfigEndpoint": "https://something.eu.auth0.com/.well-known/openid-configuration"
+		}
+	}`)
+	assert.NoError(t, err)
+
+	cfg, err := LoadMgmtConfig(context.Background(), tmpFile)
+	assert.NoError(t, err)
+	assert.Nil(t, cfg.HighestSupportedSyncMessageVersion)
+	assert.Nil(t, cfg.PerAccountHighestSupportedSyncMessageVersion)
+}
+
+func createConfig(config string) (string, error) {
 	tmpfile, err := os.CreateTemp("", "config.json")
 	if err != nil {
 		return "", err
 	}
-	_, err = tmpfile.Write([]byte(exampleConfig))
+	_, err = tmpfile.Write([]byte(config))
 	if err != nil {
 		return "", err
 	}

management/cmd/proxy/proxy.go

@@ -0,0 +1,141 @@
+// Package proxycmd provides reusable cobra commands for managing reverse proxy instances.
+// Both the management and combined binaries use these commands, each providing
+// their own StoreOpener to handle config loading and store initialization.
+package proxycmd
+
+import (
+	"bufio"
+	"context"
+	"fmt"
+	"io"
+	"strings"
+	"text/tabwriter"
+
+	"github.com/spf13/cobra"
+
+	rpproxy "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/proxy"
+	"github.com/netbirdio/netbird/management/server/store"
+)
+
+// StoreOpener initializes a store from the command context and calls fn.
+type StoreOpener func(cmd *cobra.Command, fn func(ctx context.Context, s store.Store) error) error
+
+const disconnectAllConfirmation = "disconnect all proxies"
+
+// NewCommands creates the proxy command tree with the given store opener.
+// Returns the parent "proxy" command with the disconnect-all subcommand.
+func NewCommands(opener StoreOpener) *cobra.Command {
+	var dryRun bool
+	var force bool
+
+	proxyCmd := &cobra.Command{
+		Use:   "proxy",
+		Short: "Manage reverse proxy instances",
+		Long:  "Commands for inspecting and repairing the reverse proxy instances registered with the management server.",
+	}
+
+	disconnectAllCmd := &cobra.Command{
+		Use:   "disconnect-all",
+		Short: "Force-mark all reverse proxy instances as disconnected",
+		Long: "Lists all reverse proxy instances and force-marks them as disconnected, regardless of their session state. " +
+			"Use this to repair stale connection state, e.g. after an unclean management server shutdown. " +
+			"By default, it asks for manual confirmation before changing state. Use --dry-run to preview without changing state, or --force to skip confirmation. " +
+			"Run during a maintenance window; affected live proxies may stay hidden until their next heartbeat or reconnect/re-register.",
+		Args: cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			return opener(cmd, func(ctx context.Context, s store.Store) error {
+				return runDisconnectAll(ctx, s, cmd.OutOrStdout(), cmd.InOrStdin(), dryRun, force)
+			})
+		},
+	}
+	disconnectAllCmd.Flags().BoolVar(&dryRun, "dry-run", false, "List reverse proxy instances that would be disconnected without changing state")
+	disconnectAllCmd.Flags().BoolVar(&force, "force", false, "Skip the confirmation prompt and apply the repair")
+
+	proxyCmd.AddCommand(disconnectAllCmd)
+	return proxyCmd
+}
+
+func runDisconnectAll(ctx context.Context, s store.Store, out io.Writer, in io.Reader, dryRun, force bool) error {
+	proxies, err := s.GetAllProxies(ctx)
+	if err != nil {
+		return fmt.Errorf("list proxies: %w", err)
+	}
+
+	if len(proxies) == 0 {
+		_, _ = fmt.Fprintln(out, "No reverse proxy instances found.")
+		return nil
+	}
+
+	toDisconnect := 0
+	w := tabwriter.NewWriter(out, 0, 0, 2, ' ', 0)
+	_, _ = fmt.Fprintln(w, "ID\tCLUSTER\tIP\tACCOUNT\tSTATUS\tLAST SEEN")
+	_, _ = fmt.Fprintln(w, "--\t-------\t--\t-------\t------\t---------")
+
+	for _, p := range proxies {
+		if p.Status != rpproxy.StatusDisconnected {
+			toDisconnect++
+		}
+
+		account := "-"
+		if p.AccountID != nil {
+			account = *p.AccountID
+		}
+
+		_, _ = fmt.Fprintf(w, "%s\t%s\t%s\t%s\t%s\t%s\n",
+			p.ID,
+			p.ClusterAddress,
+			p.IPAddress,
+			account,
+			p.Status,
+			p.LastSeen.Format("2006-01-02 15:04:05"),
+		)
+	}
+	if err := w.Flush(); err != nil {
+		return fmt.Errorf("write proxy list: %w", err)
+	}
+
+	if dryRun {
+		_, _ = fmt.Fprintf(out, "\nDry run: would force-mark %d of %d reverse proxy instance(s) as disconnected.\n", toDisconnect, len(proxies))
+		return nil
+	}
+
+	if !force {
+		confirmed, err := confirmDisconnectAll(out, in)
+		if err != nil {
+			return err
+		}
+		if !confirmed {
+			_, _ = fmt.Fprintln(out, "Aborted. No reverse proxy instances were changed.")
+			return nil
+		}
+	}
+
+	disconnected, err := s.DisconnectAllProxies(ctx)
+	if err != nil {
+		return fmt.Errorf("disconnect proxies: %w", err)
+	}
+
+	_, _ = fmt.Fprintf(out, "\nForce-marked %d of %d reverse proxy instance(s) as disconnected.\n", disconnected, len(proxies))
+	return nil
+}
+
+func confirmDisconnectAll(out io.Writer, in io.Reader) (bool, error) {
+	if in == nil {
+		in = strings.NewReader("")
+	}
+
+	_, _ = fmt.Fprintln(out, "\nWARNING: This command changes stored reverse proxy state for every non-disconnected instance.")
+	_, _ = fmt.Fprintln(out, "Run it during a maintenance window; affected live proxies may stay hidden until "+
+		"their next heartbeat or reconnect/re-register.")
+	_, _ = fmt.Fprintf(out, "Type %q to continue: ", disconnectAllConfirmation)
+
+	scanner := bufio.NewScanner(in)
+	if !scanner.Scan() {
+		if err := scanner.Err(); err != nil {
+			return false, fmt.Errorf("read confirmation: %w", err)
+		}
+		return false, nil
+	}
+
+	return strings.EqualFold(strings.TrimSpace(scanner.Text()), disconnectAllConfirmation), nil
+}

management/cmd/proxy/proxy_test.go

@@ -0,0 +1,180 @@
+package proxycmd
+
+import (
+	"bytes"
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/spf13/cobra"
+	"github.com/stretchr/testify/require"
+
+	rpproxy "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/proxy"
+	"github.com/netbirdio/netbird/management/server/store"
+)
+
+func newTestStore(t *testing.T) store.Store {
+	t.Helper()
+
+	s, cleanup, err := store.NewTestStoreFromSQL(context.Background(), "", t.TempDir())
+	require.NoError(t, err)
+	t.Cleanup(cleanup)
+
+	return s
+}
+
+func seedProxies(t *testing.T, ctx context.Context, s store.Store) {
+	t.Helper()
+
+	accountID := "account-1"
+	alreadyDisconnectedAt := time.Now().Add(-time.Hour)
+	seed := []*rpproxy.Proxy{
+		{
+			ID:             "proxy-1",
+			SessionID:      "session-1",
+			ClusterAddress: "cluster-a.example.com",
+			IPAddress:      "10.0.0.1",
+			LastSeen:       time.Now(),
+			Status:         rpproxy.StatusConnected,
+		},
+		{
+			ID:             "proxy-2",
+			SessionID:      "session-2",
+			ClusterAddress: "cluster-b.example.com",
+			IPAddress:      "10.0.0.2",
+			AccountID:      &accountID,
+			LastSeen:       time.Now(),
+			Status:         rpproxy.StatusConnected,
+		},
+		{
+			ID:             "proxy-3",
+			SessionID:      "session-3",
+			ClusterAddress: "cluster-a.example.com",
+			IPAddress:      "10.0.0.3",
+			LastSeen:       time.Now().Add(-time.Hour),
+			Status:         rpproxy.StatusDisconnected,
+			DisconnectedAt: &alreadyDisconnectedAt,
+		},
+	}
+	for _, p := range seed {
+		require.NoError(t, s.SaveProxy(ctx, p))
+	}
+}
+
+func proxiesByID(t *testing.T, ctx context.Context, s store.Store) map[string]*rpproxy.Proxy {
+	t.Helper()
+
+	proxies, err := s.GetAllProxies(ctx)
+	require.NoError(t, err)
+	require.Len(t, proxies, 3)
+
+	byID := make(map[string]*rpproxy.Proxy, len(proxies))
+	for _, p := range proxies {
+		byID[p.ID] = p
+	}
+	return byID
+}
+
+func TestRunDisconnectAllWithConfirmation(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	seedProxies(t, ctx, s)
+
+	var out bytes.Buffer
+	require.NoError(t, runDisconnectAll(ctx, s, &out, strings.NewReader(disconnectAllConfirmation+"\n"), false, false))
+
+	output := out.String()
+	require.Contains(t, output, "proxy-1")
+	require.Contains(t, output, "proxy-2")
+	require.Contains(t, output, "proxy-3")
+	require.Contains(t, output, "cluster-a.example.com")
+	require.Contains(t, output, "account-1")
+	require.Contains(t, output, "Type \"disconnect all proxies\" to continue")
+	require.Contains(t, output, "Force-marked 2 of 3 reverse proxy instance(s) as disconnected.")
+
+	for _, p := range proxiesByID(t, ctx, s) {
+		require.Equal(t, rpproxy.StatusDisconnected, p.Status, "proxy %s should be disconnected", p.ID)
+		require.NotNil(t, p.DisconnectedAt, "proxy %s should have a disconnected timestamp", p.ID)
+	}
+}
+
+func TestRunDisconnectAllForceSkipsConfirmation(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	seedProxies(t, ctx, s)
+
+	var out bytes.Buffer
+	require.NoError(t, runDisconnectAll(ctx, s, &out, strings.NewReader(""), false, true))
+
+	output := out.String()
+	require.NotContains(t, output, "Type \"disconnect all proxies\" to continue")
+	require.Contains(t, output, "Force-marked 2 of 3 reverse proxy instance(s) as disconnected.")
+}
+
+func TestRunDisconnectAllAbortLeavesProxiesUnchanged(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	seedProxies(t, ctx, s)
+
+	var out bytes.Buffer
+	require.NoError(t, runDisconnectAll(ctx, s, &out, strings.NewReader("no\n"), false, false))
+
+	output := out.String()
+	require.Contains(t, output, "Type \"disconnect all proxies\" to continue")
+	require.Contains(t, output, "Aborted. No reverse proxy instances were changed.")
+
+	byID := proxiesByID(t, ctx, s)
+	require.Equal(t, rpproxy.StatusConnected, byID["proxy-1"].Status)
+	require.Equal(t, rpproxy.StatusConnected, byID["proxy-2"].Status)
+	require.Equal(t, rpproxy.StatusDisconnected, byID["proxy-3"].Status)
+}
+
+func TestRunDisconnectAllDryRunLeavesProxiesUnchanged(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	seedProxies(t, ctx, s)
+
+	var out bytes.Buffer
+	require.NoError(t, runDisconnectAll(ctx, s, &out, strings.NewReader(""), true, false))
+
+	output := out.String()
+	require.Contains(t, output, "Dry run: would force-mark 2 of 3 reverse proxy instance(s) as disconnected.")
+	require.NotContains(t, output, "Type \"disconnect all proxies\" to continue")
+
+	byID := proxiesByID(t, ctx, s)
+	require.Equal(t, rpproxy.StatusConnected, byID["proxy-1"].Status)
+	require.Equal(t, rpproxy.StatusConnected, byID["proxy-2"].Status)
+	require.Equal(t, rpproxy.StatusDisconnected, byID["proxy-3"].Status)
+}
+
+func TestNewCommandsDisconnectAllDryRun(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+	seedProxies(t, ctx, s)
+
+	opened := false
+	cmd := NewCommands(func(cmd *cobra.Command, fn func(ctx context.Context, s store.Store) error) error {
+		opened = true
+		return fn(cmd.Context(), s)
+	})
+
+	var out bytes.Buffer
+	cmd.SetOut(&out)
+	cmd.SetErr(&out)
+	cmd.SetIn(strings.NewReader(""))
+	cmd.SetArgs([]string{"disconnect-all", "--dry-run"})
+
+	require.NoError(t, cmd.ExecuteContext(ctx))
+	require.True(t, opened)
+	require.Contains(t, out.String(), "Dry run: would force-mark 2 of 3 reverse proxy instance(s) as disconnected.")
+}
+
+func TestRunDisconnectAllEmpty(t *testing.T) {
+	ctx := context.Background()
+	s := newTestStore(t)
+
+	var out bytes.Buffer
+	require.NoError(t, runDisconnectAll(ctx, s, &out, strings.NewReader(""), false, false))
+	require.Contains(t, out.String(), "No reverse proxy instances found.")
+}

management/cmd/root.go

@@ -83,7 +83,8 @@ func init() {
 
 	rootCmd.AddCommand(migrationCmd)
 
-	tc := newTokenCommands()
-	tc.PersistentFlags().StringVar(&nbconfig.MgmtConfigPath, "config", defaultMgmtConfig, "Netbird config file location")
-	rootCmd.AddCommand(tc)
+	ac := newAdminCommands()
+	ac.PersistentFlags().StringVar(&nbconfig.MgmtConfigPath, "config", defaultMgmtConfig, "Netbird config file location")
+	rootCmd.AddCommand(ac)
+	rootCmd.AddCommand(newLegacyTokenCommand())
 }

management/cmd/token.go

@@ -1,55 +0,0 @@
-package cmd
-
-import (
-	"context"
-	"fmt"
-
-	log "github.com/sirupsen/logrus"
-	"github.com/spf13/cobra"
-
-	"github.com/netbirdio/netbird/formatter/hook"
-	tokencmd "github.com/netbirdio/netbird/management/cmd/token"
-	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
-	"github.com/netbirdio/netbird/management/server/store"
-	"github.com/netbirdio/netbird/util"
-)
-
-var tokenDatadir string
-
-// newTokenCommands creates the token command tree with management-specific store opener.
-func newTokenCommands() *cobra.Command {
-	cmd := tokencmd.NewCommands(withTokenStore)
-	cmd.PersistentFlags().StringVar(&tokenDatadir, "datadir", "", "Override the data directory from config (where store.db is located)")
-	return cmd
-}
-
-// withTokenStore initializes logging, loads config, opens 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
-
-	config, err := LoadMgmtConfig(ctx, nbconfig.MgmtConfigPath)
-	if err != nil {
-		return fmt.Errorf("load config: %w", err)
-	}
-
-	datadir := config.Datadir
-	if tokenDatadir != "" {
-		datadir = tokenDatadir
-	}
-
-	s, err := store.NewStore(ctx, config.StoreConfig.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)
-}

management/internals/controllers/network_map/controller/controller.go

@@ -29,6 +29,7 @@ import (
 	"github.com/netbirdio/netbird/management/server/store"
 	"github.com/netbirdio/netbird/management/server/telemetry"
 	"github.com/netbirdio/netbird/management/server/types"
+	sharedgrpc "github.com/netbirdio/netbird/shared/management/grpc"
 	"github.com/netbirdio/netbird/shared/management/proto"
 	"github.com/netbirdio/netbird/shared/management/status"
 	"github.com/netbirdio/netbird/util"
@@ -56,6 +57,10 @@ type Controller struct {
 	proxyController port_forwarding.Controller
 
 	integratedPeerValidator integrated_validator.IntegratedValidator
+
+	serverSupportedSyncMessageVersion sharedgrpc.SyncMessageVersion
+
+	perAccountServerSupportedSyncMessageVersions map[string]sharedgrpc.SyncMessageVersion
 }
 
 type bufferUpdate struct {
@@ -90,8 +95,10 @@ func NewController(ctx context.Context, store store.Store, metrics telemetry.App
 		dnsDomain:               dnsDomain,
 		config:                  config,
 
-		proxyController:       proxyController,
-		EphemeralPeersManager: ephemeralPeersManager,
+		proxyController:                              proxyController,
+		EphemeralPeersManager:                        ephemeralPeersManager,
+		serverSupportedSyncMessageVersion:            sharedgrpc.SyncMessageVersionFromConfig(config.HighestSupportedSyncMessageVersion),
+		perAccountServerSupportedSyncMessageVersions: sharedgrpc.SyncMessageVersionsFromMap(config.PerAccountHighestSupportedSyncMessageVersion),
 	}
 }
 
@@ -222,18 +229,53 @@ func (c *Controller) sendUpdateAccountPeers(ctx context.Context, accountID strin
 			c.metrics.CountCalcPostureChecksDuration(time.Since(start))
 			start = time.Now()
 
-			remotePeerNetworkMap := account.GetPeerNetworkMapFromComponents(ctx, p.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, c.accountManagerMetrics, groupIDToUserIDs)
+			peerGroups := account.GetPeerGroups(p.ID)
+			proxyNetworkMap := proxyNetworkMaps[p.ID]
+			var update *proto.SyncResponse
+
+			commonSyncMessageVersion := sharedgrpc.HighestCommonSyncMessageVersion(
+				c.perAccountOrGlobalSupportedSyncMessageVersions(accountID),
+				sharedgrpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion))
+
+			log.WithContext(ctx).
+				WithFields(log.Fields{
+					"sync_message_version":        commonSyncMessageVersion,
+					"server_sync_message_version": c.perAccountOrGlobalSupportedSyncMessageVersions(peer.AccountID),
+					"peer_sync_message_version":   sharedgrpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion),
+				}).Debug("common highest sync message version")
+
+			if commonSyncMessageVersion == sharedgrpc.ComponentNetworkMap {
+				components := account.GetPeerNetworkMapComponents(
+					ctx, p.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, groupIDToUserIDs)
+
+				c.metrics.CountCalcPeerNetworkMapDuration(time.Since(start))
+
+				start = time.Now()
+				// proxyNetworkMap rides the envelope as a ProxyPatch sidecar;
+				// the client merges it into Calculate()'s output the same
+				// way the legacy server did via NetworkMap.Merge.
+				update = grpc.ToComponentSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, p, nil, nil, components, proxyNetworkMap, dnsDomain, postureChecks, account.Settings, extraSetting, maps.Keys(peerGroups), dnsFwdPort)
+				c.metrics.CountToComponentSyncResponseDuration(time.Since(start))
+
+				c.peersUpdateManager.SendUpdate(ctx, p.ID, &network_map.UpdateMessage{
+					Update:      update,
+					MessageType: network_map.MessageTypeNetworkMap,
+				})
+
+				return
+			}
+
+			nmap := account.GetPeerNetworkMapFromComponents(
+				ctx, p.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, c.accountManagerMetrics, groupIDToUserIDs)
 
 			c.metrics.CountCalcPeerNetworkMapDuration(time.Since(start))
 
-			proxyNetworkMap, ok := proxyNetworkMaps[p.ID]
-			if ok {
-				remotePeerNetworkMap.Merge(proxyNetworkMap)
+			if proxyNetworkMap != nil {
+				nmap.Merge(proxyNetworkMap)
 			}
 
-			peerGroups := account.GetPeerGroups(p.ID)
 			start = time.Now()
-			update := grpc.ToSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, p, nil, nil, remotePeerNetworkMap, dnsDomain, postureChecks, dnsCache, account.Settings, extraSetting, maps.Keys(peerGroups), dnsFwdPort)
+			update = grpc.ToSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, p, nil, nil, nmap, dnsDomain, postureChecks, dnsCache, account.Settings, extraSetting, maps.Keys(peerGroups), dnsFwdPort)
 			c.metrics.CountToSyncResponseDuration(time.Since(start))
 
 			c.peersUpdateManager.SendUpdate(ctx, p.ID, &network_map.UpdateMessage{
@@ -251,6 +293,13 @@ func (c *Controller) sendUpdateAccountPeers(ctx context.Context, accountID strin
 	return nil
 }
 
+func (c *Controller) perAccountOrGlobalSupportedSyncMessageVersions(accountId string) sharedgrpc.SyncMessageVersion {
+	if perAccount, ok := c.perAccountServerSupportedSyncMessageVersions[accountId]; ok {
+		return perAccount
+	}
+	return c.serverSupportedSyncMessageVersion
+}
+
 // UpdatePeers updates all peers that belong to an account.
 // Should be called when changes have to be synced to peers.
 func (c *Controller) UpdateAccountPeers(ctx context.Context, accountID string, reason types.UpdateReason) error {
@@ -352,18 +401,53 @@ func (c *Controller) sendUpdateForAffectedPeers(ctx context.Context, accountID s
 			c.metrics.CountCalcPostureChecksDuration(time.Since(start))
 			start = time.Now()
 
-			remotePeerNetworkMap := account.GetPeerNetworkMapFromComponents(ctx, p.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, c.accountManagerMetrics, groupIDToUserIDs)
+			peerGroups := account.GetPeerGroups(p.ID)
+			proxyNetworkMap := proxyNetworkMaps[p.ID]
+			var update *proto.SyncResponse
+
+			commonSyncMessageVersion := sharedgrpc.HighestCommonSyncMessageVersion(
+				c.perAccountOrGlobalSupportedSyncMessageVersions(accountID),
+				sharedgrpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion))
+
+			log.WithContext(ctx).
+				WithFields(log.Fields{
+					"sync_message_version":        commonSyncMessageVersion,
+					"server_sync_message_version": c.perAccountOrGlobalSupportedSyncMessageVersions(peer.AccountID),
+					"peer_sync_message_version":   sharedgrpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion),
+				}).Debug("common highest sync message version")
+
+			if commonSyncMessageVersion == sharedgrpc.ComponentNetworkMap {
+				components := account.GetPeerNetworkMapComponents(
+					ctx, p.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, groupIDToUserIDs)
+
+				c.metrics.CountCalcPeerNetworkMapDuration(time.Since(start))
+
+				start = time.Now()
+				// proxyNetworkMap rides the envelope as a ProxyPatch sidecar;
+				// the client merges it into Calculate()'s output the same
+				// way the legacy server did via NetworkMap.Merge.
+				update = grpc.ToComponentSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, p, nil, nil, components, proxyNetworkMap, dnsDomain, postureChecks, account.Settings, extraSetting, maps.Keys(peerGroups), dnsFwdPort)
+				c.metrics.CountToComponentSyncResponseDuration(time.Since(start))
+
+				c.peersUpdateManager.SendUpdate(ctx, p.ID, &network_map.UpdateMessage{
+					Update:      update,
+					MessageType: network_map.MessageTypeNetworkMap,
+				})
+
+				return
+			}
+
+			nmap := account.GetPeerNetworkMapFromComponents(
+				ctx, p.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, c.accountManagerMetrics, groupIDToUserIDs)
 
 			c.metrics.CountCalcPeerNetworkMapDuration(time.Since(start))
 
-			proxyNetworkMap, ok := proxyNetworkMaps[p.ID]
-			if ok {
-				remotePeerNetworkMap.Merge(proxyNetworkMap)
+			if proxyNetworkMap != nil {
+				nmap.Merge(proxyNetworkMap)
 			}
 
-			peerGroups := account.GetPeerGroups(p.ID)
 			start = time.Now()
-			update := grpc.ToSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, p, nil, nil, remotePeerNetworkMap, dnsDomain, postureChecks, dnsCache, account.Settings, extraSetting, maps.Keys(peerGroups), dnsFwdPort)
+			update = grpc.ToSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, p, nil, nil, nmap, dnsDomain, postureChecks, dnsCache, account.Settings, extraSetting, maps.Keys(peerGroups), dnsFwdPort)
 			c.metrics.CountToSyncResponseDuration(time.Since(start))
 
 			c.peersUpdateManager.SendUpdate(ctx, p.ID, &network_map.UpdateMessage{
@@ -451,13 +535,7 @@ func (c *Controller) UpdateAccountPeer(ctx context.Context, accountId string, pe
 		return err
 	}
 
-	remotePeerNetworkMap := account.GetPeerNetworkMapFromComponents(ctx, peerId, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, c.accountManagerMetrics, groupIDToUserIDs)
-
-	proxyNetworkMap, ok := proxyNetworkMaps[peer.ID]
-	if ok {
-		remotePeerNetworkMap.Merge(proxyNetworkMap)
-	}
-
+	proxyNetworkMap := proxyNetworkMaps[peer.ID]
 	extraSettings, err := c.settingsManager.GetExtraSettings(ctx, peer.AccountID)
 	if err != nil {
 		return fmt.Errorf("failed to get extra settings: %v", err)
@@ -466,7 +544,45 @@ func (c *Controller) UpdateAccountPeer(ctx context.Context, accountId string, pe
 	peerGroups := account.GetPeerGroups(peerId)
 	dnsFwdPort := computeForwarderPort(maps.Values(account.Peers), network_map.DnsForwarderPortMinVersion)
 
-	update := grpc.ToSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, peer, nil, nil, remotePeerNetworkMap, dnsDomain, postureChecks, dnsCache, account.Settings, extraSettings, maps.Keys(peerGroups), dnsFwdPort)
+	var update *proto.SyncResponse
+
+	commonSyncMessageVersion := sharedgrpc.HighestCommonSyncMessageVersion(
+		c.perAccountOrGlobalSupportedSyncMessageVersions(accountId),
+		sharedgrpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion))
+
+	log.WithContext(ctx).
+		WithFields(log.Fields{
+			"sync_message_version":        commonSyncMessageVersion,
+			"server_sync_message_version": c.perAccountOrGlobalSupportedSyncMessageVersions(peer.AccountID),
+			"peer_sync_message_version":   sharedgrpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion),
+		}).Debug("common highest sync message version")
+
+	if commonSyncMessageVersion == sharedgrpc.ComponentNetworkMap {
+		components := account.GetPeerNetworkMapComponents(
+			ctx, peer.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, groupIDToUserIDs)
+
+		// proxyNetworkMap rides the envelope as a ProxyPatch sidecar;
+		// the client merges it into Calculate()'s output the same
+		// way the legacy server did via NetworkMap.Merge.
+		update = grpc.ToComponentSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, peer, nil, nil, components, proxyNetworkMap, dnsDomain, postureChecks, account.Settings, extraSettings, maps.Keys(peerGroups), dnsFwdPort)
+
+		c.peersUpdateManager.SendUpdate(ctx, peer.ID, &network_map.UpdateMessage{
+			Update:      update,
+			MessageType: network_map.MessageTypeNetworkMap,
+		})
+
+		return nil
+	}
+
+	nmap := account.GetPeerNetworkMapFromComponents(
+		ctx, peer.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, c.accountManagerMetrics, groupIDToUserIDs)
+
+	if proxyNetworkMap != nil {
+		nmap.Merge(proxyNetworkMap)
+	}
+
+	update = grpc.ToSyncResponse(ctx, nil, c.config.HttpConfig, c.config.DeviceAuthorizationFlow, peer, nil, nil, nmap, dnsDomain, postureChecks, dnsCache, account.Settings, extraSettings, maps.Keys(peerGroups), dnsFwdPort)
+
 	c.peersUpdateManager.SendUpdate(ctx, peer.ID, &network_map.UpdateMessage{
 		Update:      update,
 		MessageType: network_map.MessageTypeNetworkMap,
@@ -513,6 +629,65 @@ func (c *Controller) BufferUpdateAccountPeers(ctx context.Context, accountID str
 	return nil
 }
 
+// GetValidatedPeerWithComponents is the components-format counterpart of
+// GetValidatedPeerWithMap. It returns raw NetworkMapComponents for capable
+// peers along with the proxy NetworkMap fragment (BYOP / port-forwarding
+// data the legacy server folds in via NetworkMap.Merge). The gRPC layer
+// encodes both into the wire envelope. Callers must gate on capability
+// themselves before dispatching here — this method does NOT branch on it.
+func (c *Controller) GetValidatedPeerWithComponents(ctx context.Context, isRequiresApproval bool, accountID string, peer *nbpeer.Peer) (*nbpeer.Peer, *types.NetworkMapComponents, *types.NetworkMap, []*posture.Checks, int64, error) {
+	if isRequiresApproval {
+		network, err := c.repo.GetAccountNetwork(ctx, accountID)
+		if err != nil {
+			return nil, nil, nil, nil, 0, err
+		}
+		return peer, &types.NetworkMapComponents{Network: network.Copy()}, nil, nil, 0, nil
+	}
+
+	account, err := c.requestBuffer.GetAccountWithBackpressure(ctx, accountID)
+	if err != nil {
+		return nil, nil, nil, nil, 0, err
+	}
+
+	c.injectAllProxyPolicies(ctx, account)
+
+	approvedPeersMap, err := c.integratedPeerValidator.GetValidatedPeers(ctx, account.Id, maps.Values(account.Groups), maps.Values(account.Peers), account.Settings.Extra)
+	if err != nil {
+		return nil, nil, nil, nil, 0, err
+	}
+
+	postureChecks, err := c.getPeerPostureChecks(account, peer.ID)
+	if err != nil {
+		return nil, nil, nil, nil, 0, err
+	}
+
+	accountZones, err := c.repo.GetAccountZones(ctx, account.Id)
+	if err != nil {
+		return nil, nil, nil, nil, 0, err
+	}
+
+	// Fetch the proxy network map fragment for this peer alongside the
+	// components — same single-account-load path the streaming controller
+	// uses, so initial-sync delivers BYOP/forwarding patches synchronously
+	// instead of waiting for the next streaming push.
+	proxyNetworkMaps, err := c.proxyController.GetProxyNetworkMaps(ctx, account.Id, peer.ID, account.Peers)
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to get proxy network maps: %v", err)
+		return nil, nil, nil, nil, 0, err
+	}
+
+	dnsDomain := c.GetDNSDomain(account.Settings)
+	peersCustomZone := account.GetPeersCustomZone(ctx, dnsDomain)
+
+	resourcePolicies := account.GetResourcePoliciesMap()
+	routers := account.GetResourceRoutersMap()
+	groupIDToUserIDs := account.GetActiveGroupUsers()
+	components := account.GetPeerNetworkMapComponents(ctx, peer.ID, peersCustomZone, accountZones, approvedPeersMap, resourcePolicies, routers, groupIDToUserIDs)
+	dnsFwdPort := computeForwarderPort(maps.Values(account.Peers), network_map.DnsForwarderPortMinVersion)
+
+	return peer, components, proxyNetworkMaps[peer.ID], postureChecks, dnsFwdPort, nil
+}
+
 // BufferUpdateAffectedPeers accumulates peer IDs and flushes them after the buffer interval.
 func (c *Controller) BufferUpdateAffectedPeers(ctx context.Context, accountID string, peerIDs []string, reason types.UpdateReason) error {
 	if len(peerIDs) == 0 {

management/internals/controllers/network_map/interface.go

@@ -24,6 +24,7 @@ type Controller interface {
 	UpdateAccountPeer(ctx context.Context, accountId string, peerId string) error
 	BufferUpdateAccountPeers(ctx context.Context, accountID string, reason types.UpdateReason) error
 	GetValidatedPeerWithMap(ctx context.Context, isRequiresApproval bool, accountID string, peerID string) (*types.NetworkMap, []*posture.Checks, int64, error)
+	GetValidatedPeerWithComponents(ctx context.Context, isRequiresApproval bool, accountID string, p *nbpeer.Peer) (*nbpeer.Peer, *types.NetworkMapComponents, *types.NetworkMap, []*posture.Checks, int64, error)
 	GetDNSDomain(settings *types.Settings) string
 	StartWarmup(context.Context)
 	GetNetworkMap(ctx context.Context, peerID string) (*types.NetworkMap, error)

management/internals/controllers/network_map/interface_mock.go

@@ -1,9 +1,9 @@
 // Code generated by MockGen. DO NOT EDIT.
-// Source: management/internals/controllers/network_map/interface.go
+// Source: ./interface.go
 //
 // Generated by this command:
 //
-//	mockgen -package network_map -destination=management/internals/controllers/network_map/interface_mock.go -source=management/internals/controllers/network_map/interface.go -build_flags=-mod=mod
+//	mockgen -package network_map -destination=interface_mock.go -source=./interface.go -build_flags=-mod=mod
 //
 
 // Package network_map is a generated GoMock package.
@@ -126,8 +126,27 @@ func (mr *MockControllerMockRecorder) GetNetworkMap(ctx, peerID any) *gomock.Cal
 	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetNetworkMap", reflect.TypeOf((*MockController)(nil).GetNetworkMap), ctx, peerID)
 }
 
+// GetValidatedPeerWithComponents mocks base method.
+func (m *MockController) GetValidatedPeerWithComponents(ctx context.Context, isRequiresApproval bool, accountID string, p *peer.Peer) (*peer.Peer, *types.NetworkMapComponents, *types.NetworkMap, []*posture.Checks, int64, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "GetValidatedPeerWithComponents", ctx, isRequiresApproval, accountID, p)
+	ret0, _ := ret[0].(*peer.Peer)
+	ret1, _ := ret[1].(*types.NetworkMapComponents)
+	ret2, _ := ret[2].(*types.NetworkMap)
+	ret3, _ := ret[3].([]*posture.Checks)
+	ret4, _ := ret[4].(int64)
+	ret5, _ := ret[5].(error)
+	return ret0, ret1, ret2, ret3, ret4, ret5
+}
+
+// GetValidatedPeerWithComponents indicates an expected call of GetValidatedPeerWithComponents.
+func (mr *MockControllerMockRecorder) GetValidatedPeerWithComponents(ctx, isRequiresApproval, accountID, p any) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetValidatedPeerWithComponents", reflect.TypeOf((*MockController)(nil).GetValidatedPeerWithComponents), ctx, isRequiresApproval, accountID, p)
+}
+
 // GetValidatedPeerWithMap mocks base method.
-func (m *MockController) GetValidatedPeerWithMap(ctx context.Context, isRequiresApproval bool, accountID string, peerID string) (*types.NetworkMap, []*posture.Checks, int64, error) {
+func (m *MockController) GetValidatedPeerWithMap(ctx context.Context, isRequiresApproval bool, accountID, peerID string) (*types.NetworkMap, []*posture.Checks, int64, error) {
 	m.ctrl.T.Helper()
 	ret := m.ctrl.Call(m, "GetValidatedPeerWithMap", ctx, isRequiresApproval, accountID, peerID)
 	ret0, _ := ret[0].(*types.NetworkMap)
@@ -171,7 +190,7 @@ func (mr *MockControllerMockRecorder) OnPeerDisconnected(ctx, accountID, peerID
 }
 
 // OnPeersAdded mocks base method.
-func (m *MockController) OnPeersAdded(ctx context.Context, accountID string, peerIDs []string, affectedPeerIDs []string) error {
+func (m *MockController) OnPeersAdded(ctx context.Context, accountID string, peerIDs, affectedPeerIDs []string) error {
 	m.ctrl.T.Helper()
 	ret := m.ctrl.Call(m, "OnPeersAdded", ctx, accountID, peerIDs, affectedPeerIDs)
 	ret0, _ := ret[0].(error)
@@ -185,7 +204,7 @@ func (mr *MockControllerMockRecorder) OnPeersAdded(ctx, accountID, peerIDs, affe
 }
 
 // OnPeersDeleted mocks base method.
-func (m *MockController) OnPeersDeleted(ctx context.Context, accountID string, peerIDs []string, affectedPeerIDs []string) error {
+func (m *MockController) OnPeersDeleted(ctx context.Context, accountID string, peerIDs, affectedPeerIDs []string) error {
 	m.ctrl.T.Helper()
 	ret := m.ctrl.Call(m, "OnPeersDeleted", ctx, accountID, peerIDs, affectedPeerIDs)
 	ret0, _ := ret[0].(error)
@@ -199,7 +218,7 @@ func (mr *MockControllerMockRecorder) OnPeersDeleted(ctx, accountID, peerIDs, af
 }
 
 // OnPeersUpdated mocks base method.
-func (m *MockController) OnPeersUpdated(ctx context.Context, accountId string, peerIDs []string, affectedPeerIDs []string) error {
+func (m *MockController) OnPeersUpdated(ctx context.Context, accountId string, peerIDs, affectedPeerIDs []string) error {
 	m.ctrl.T.Helper()
 	ret := m.ctrl.Call(m, "OnPeersUpdated", ctx, accountId, peerIDs, affectedPeerIDs)
 	ret0, _ := ret[0].(error)

management/internals/modules/agentnetwork/accesslog_ingest.go

@@ -18,21 +18,26 @@ import (
 // contract between the proxy and management; management flattens them into
 // queryable columns. Keep in sync with the proxy side.
 const (
-	metaKeyProvider           = "llm.provider"
-	metaKeyModel              = "llm.model"
-	metaKeyResolvedProviderID = "llm.resolved_provider_id"
-	metaKeySelectedPolicyID   = "llm.selected_policy_id"
-	metaKeyPolicyDecision     = "llm_policy.decision"
-	metaKeyPolicyReason       = "llm_policy.reason"
-	metaKeyInputTokens        = "llm.input_tokens"  //nolint:gosec // metadata key name, not a credential
-	metaKeyOutputTokens       = "llm.output_tokens" //nolint:gosec // metadata key name, not a credential
-	metaKeyTotalTokens        = "llm.total_tokens"  //nolint:gosec // metadata key name, not a credential
-	metaKeyCostUSDTotal       = "cost.usd_total"
-	metaKeyStream             = "llm.stream"
-	metaKeySessionID          = "llm.session_id"
-	metaKeyAuthorisingGroups  = "llm.authorising_groups"
-	metaKeyRequestPrompt      = "llm.request_prompt"
-	metaKeyResponseCompletion = "llm.response_completion"
+	metaKeyProvider            = "llm.provider"
+	metaKeyModel               = "llm.model"
+	metaKeyResolvedProviderID  = "llm.resolved_provider_id"
+	metaKeySelectedPolicyID    = "llm.selected_policy_id"
+	metaKeyPolicyDecision      = "llm_policy.decision"
+	metaKeyPolicyReason        = "llm_policy.reason"
+	metaKeyInputTokens         = "llm.input_tokens"          //nolint:gosec // metadata key name, not a credential
+	metaKeyOutputTokens        = "llm.output_tokens"         //nolint:gosec // metadata key name, not a credential
+	metaKeyTotalTokens         = "llm.total_tokens"          //nolint:gosec // metadata key name, not a credential
+	metaKeyCachedInputTokens   = "llm.cached_input_tokens"   //nolint:gosec // metadata key name, not a credential
+	metaKeyCacheCreationTokens = "llm.cache_creation_tokens" //nolint:gosec // metadata key name, not a credential
+	metaKeyCostUSDInput        = "cost.usd_input"
+	metaKeyCostUSDCachedInput  = "cost.usd_cached_input"
+	metaKeyCostUSDCacheCreate  = "cost.usd_cache_creation"
+	metaKeyCostUSDOutput       = "cost.usd_output"
+	metaKeyStream              = "llm.stream"
+	metaKeySessionID           = "llm.session_id"
+	metaKeyAuthorisingGroups   = "llm.authorising_groups"
+	metaKeyRequestPrompt       = "llm.request_prompt"
+	metaKeyResponseCompletion  = "llm.response_completion"
 )
 
 // IngestAccessLog flattens the metadata-bearing reverse-proxy access-log entry
@@ -108,20 +113,25 @@ func flattenAccessLog(e *accesslogs.AccessLogEntry) (*types.AgentNetworkAccessLo
 		BytesUpload:   e.BytesUpload,
 		BytesDownload: e.BytesDownload,
 
-		Provider:           meta[metaKeyProvider],
-		Model:              meta[metaKeyModel],
-		SessionID:          meta[metaKeySessionID],
-		ResolvedProviderID: meta[metaKeyResolvedProviderID],
-		SelectedPolicyID:   meta[metaKeySelectedPolicyID],
-		Decision:           meta[metaKeyPolicyDecision],
-		DenyReason:         meta[metaKeyPolicyReason],
-		InputTokens:        parseMetaInt(meta, metaKeyInputTokens),
-		OutputTokens:       parseMetaInt(meta, metaKeyOutputTokens),
-		TotalTokens:        parseMetaInt(meta, metaKeyTotalTokens),
-		CostUSD:            parseMetaFloat(meta, metaKeyCostUSDTotal),
-		Stream:             parseMetaBool(meta, metaKeyStream),
-		RequestPrompt:      meta[metaKeyRequestPrompt],
-		ResponseCompletion: meta[metaKeyResponseCompletion],
+		Provider:             meta[metaKeyProvider],
+		Model:                meta[metaKeyModel],
+		SessionID:            meta[metaKeySessionID],
+		ResolvedProviderID:   meta[metaKeyResolvedProviderID],
+		SelectedPolicyID:     meta[metaKeySelectedPolicyID],
+		Decision:             meta[metaKeyPolicyDecision],
+		DenyReason:           meta[metaKeyPolicyReason],
+		InputTokens:          parseMetaInt(meta, metaKeyInputTokens),
+		OutputTokens:         parseMetaInt(meta, metaKeyOutputTokens),
+		TotalTokens:          parseMetaInt(meta, metaKeyTotalTokens),
+		CachedInputTokens:    parseMetaInt(meta, metaKeyCachedInputTokens),
+		CacheCreationTokens:  parseMetaInt(meta, metaKeyCacheCreationTokens),
+		InputCostUSD:         parseMetaFloat(meta, metaKeyCostUSDInput),
+		CachedInputCostUSD:   parseMetaFloat(meta, metaKeyCostUSDCachedInput),
+		CacheCreationCostUSD: parseMetaFloat(meta, metaKeyCostUSDCacheCreate),
+		OutputCostUSD:        parseMetaFloat(meta, metaKeyCostUSDOutput),
+		Stream:               parseMetaBool(meta, metaKeyStream),
+		RequestPrompt:        meta[metaKeyRequestPrompt],
+		ResponseCompletion:   meta[metaKeyResponseCompletion],
 	}
 
 	var groups []types.AgentNetworkAccessLogGroup
@@ -140,18 +150,23 @@ func flattenAccessLog(e *accesslogs.AccessLogEntry) (*types.AgentNetworkAccessLo
 // log's ID so the two correlate.
 func usageFromFlattenedLog(e *types.AgentNetworkAccessLog, groups []types.AgentNetworkAccessLogGroup) (*types.AgentNetworkUsage, []types.AgentNetworkUsageGroup) {
 	usage := &types.AgentNetworkUsage{
-		ID:                 e.ID,
-		AccountID:          e.AccountID,
-		Timestamp:          e.Timestamp,
-		UserID:             e.UserID,
-		ResolvedProviderID: e.ResolvedProviderID,
-		Provider:           e.Provider,
-		Model:              e.Model,
-		SessionID:          e.SessionID,
-		InputTokens:        e.InputTokens,
-		OutputTokens:       e.OutputTokens,
-		TotalTokens:        e.TotalTokens,
-		CostUSD:            e.CostUSD,
+		ID:                   e.ID,
+		AccountID:            e.AccountID,
+		Timestamp:            e.Timestamp,
+		UserID:               e.UserID,
+		ResolvedProviderID:   e.ResolvedProviderID,
+		Provider:             e.Provider,
+		Model:                e.Model,
+		SessionID:            e.SessionID,
+		InputTokens:          e.InputTokens,
+		OutputTokens:         e.OutputTokens,
+		TotalTokens:          e.TotalTokens,
+		CachedInputTokens:    e.CachedInputTokens,
+		CacheCreationTokens:  e.CacheCreationTokens,
+		InputCostUSD:         e.InputCostUSD,
+		CachedInputCostUSD:   e.CachedInputCostUSD,
+		CacheCreationCostUSD: e.CacheCreationCostUSD,
+		OutputCostUSD:        e.OutputCostUSD,
 	}
 
 	usageGroups := make([]types.AgentNetworkUsageGroup, 0, len(groups))

management/internals/modules/agentnetwork/accesslog_ingest_realstore_test.go

@@ -28,17 +28,22 @@ func newIngestTestEntry() *accesslogs.AccessLogEntry {
 		UserId:       "user-1",
 		AgentNetwork: true,
 		Metadata: map[string]string{
-			metaKeyProvider:           "openai",
-			metaKeyModel:              "gpt-5.4",
-			metaKeyResolvedProviderID: "prov-1",
-			metaKeySessionID:          "sess-1",
-			metaKeyInputTokens:        "100",
-			metaKeyOutputTokens:       "50",
-			metaKeyTotalTokens:        "150",
-			metaKeyCostUSDTotal:       "0.0123",
-			metaKeyStream:             "true",
-			metaKeyRequestPrompt:      "hello",
-			metaKeyResponseCompletion: "world",
+			metaKeyProvider:            "openai",
+			metaKeyModel:               "gpt-5.4",
+			metaKeyResolvedProviderID:  "prov-1",
+			metaKeySessionID:           "sess-1",
+			metaKeyInputTokens:         "100",
+			metaKeyOutputTokens:        "50",
+			metaKeyTotalTokens:         "1174",
+			metaKeyCachedInputTokens:   "256",
+			metaKeyCacheCreationTokens: "768",
+			metaKeyCostUSDInput:        "0.0071",
+			metaKeyCostUSDCachedInput:  "0.0009",
+			metaKeyCostUSDCacheCreate:  "0.0020",
+			metaKeyCostUSDOutput:       "0.0023",
+			metaKeyStream:              "true",
+			metaKeyRequestPrompt:       "hello",
+			metaKeyResponseCompletion:  "world",
 			// repeated id must be de-duplicated before the group rows insert.
 			metaKeyAuthorisingGroups: "grp-eng,grp-eng,grp-ops",
 		},
@@ -65,7 +70,19 @@ func TestIngestAccessLog_RealStore_LogCollectionOff(t *testing.T) {
 	require.Len(t, usage, 1, "usage row must be written even with log collection off")
 	assert.Equal(t, int64(100), usage[0].InputTokens, "input tokens must round-trip from metadata")
 	assert.Equal(t, int64(50), usage[0].OutputTokens, "output tokens must round-trip from metadata")
-	assert.InDelta(t, 0.0123, usage[0].CostUSD, 1e-9, "cost must round-trip from metadata")
+	assert.Equal(t, int64(256), usage[0].CachedInputTokens, "cache-read tokens must round-trip from metadata")
+	assert.Equal(t, int64(768), usage[0].CacheCreationTokens, "cache-write tokens must round-trip from metadata")
+	// The per-bucket breakdown is the only cost state stored, and must survive
+	// the write/read cycle as real columns — usage rows are the only cost
+	// record for accounts with log collection off, so a dropped column here
+	// loses the split permanently.
+	assert.InDelta(t, 0.0071, usage[0].InputCostUSD, 1e-9, "input cost must round-trip from metadata")
+	assert.InDelta(t, 0.0009, usage[0].CachedInputCostUSD, 1e-9, "cache-read cost must round-trip from metadata")
+	assert.InDelta(t, 0.0020, usage[0].CacheCreationCostUSD, 1e-9, "cache-write cost must round-trip from metadata")
+	assert.InDelta(t, 0.0023, usage[0].OutputCostUSD, 1e-9, "output cost must round-trip from metadata")
+	// Aggregates are derived from the stored columns, never stored themselves.
+	assert.InDelta(t, 0.0123, usage[0].TotalCostUSD(), 1e-9, "total is derived from the stored buckets")
+	assert.InDelta(t, 0.0029, usage[0].CacheCostUSD(), 1e-9, "cache cost is derived from the two cache buckets")
 
 	logs, _, err := s.GetAgentNetworkAccessLogs(ctx, store.LockingStrengthNone, testAccountID, types.AgentNetworkAccessLogFilter{})
 	require.NoError(t, err)
@@ -96,6 +113,14 @@ func TestIngestAccessLog_RealStore_LogCollectionOn(t *testing.T) {
 	require.Equal(t, int64(1), total, "exactly one access-log row expected")
 	require.Len(t, logs, 1, "full access-log row must be written when log collection is on")
 	assert.Equal(t, "gpt-5.4", logs[0].Model, "model must flatten from metadata")
+	assert.Equal(t, int64(256), logs[0].CachedInputTokens, "cache-read tokens must flatten from metadata")
+	assert.Equal(t, int64(768), logs[0].CacheCreationTokens, "cache-write tokens must flatten from metadata")
+	assert.InDelta(t, 0.0029, logs[0].CacheCostUSD(), 1e-9, "cache cost is derived from the two cache buckets")
+	assert.InDelta(t, 0.0123, logs[0].TotalCostUSD(), 1e-9, "total is derived from the stored buckets")
+	assert.InDelta(t, 0.0071, logs[0].InputCostUSD, 1e-9, "input cost must flatten from metadata")
+	assert.InDelta(t, 0.0009, logs[0].CachedInputCostUSD, 1e-9, "cache-read cost must flatten from metadata")
+	assert.InDelta(t, 0.0020, logs[0].CacheCreationCostUSD, 1e-9, "cache-write cost must flatten from metadata")
+	assert.InDelta(t, 0.0023, logs[0].OutputCostUSD, 1e-9, "output cost must flatten from metadata")
 	assert.Equal(t, "hello", logs[0].RequestPrompt, "prompt must be retained when log collection is on")
 	assert.Equal(t, "world", logs[0].ResponseCompletion, "completion must be retained when log collection is on")
 	assert.True(t, logs[0].Stream, "stream flag must flatten from metadata")

management/internals/modules/agentnetwork/accesslog_sessions_realstore_test.go

@@ -38,7 +38,7 @@ func accessLogRow(id, sessionID string, ts time.Time, opts ...func(*types.AgentN
 		InputTokens:        100,
 		OutputTokens:       50,
 		TotalTokens:        150,
-		CostUSD:            0.01,
+		InputCostUSD:       0.01,
 	}
 	for _, o := range opts {
 		o(e)
@@ -74,7 +74,7 @@ func withTokens(in, out, total int64, cost float64) func(*types.AgentNetworkAcce
 		e.InputTokens = in
 		e.OutputTokens = out
 		e.TotalTokens = total
-		e.CostUSD = cost
+		e.InputCostUSD = cost
 	}
 }
 
@@ -155,7 +155,7 @@ func TestAccessLogSessions_FoldAndAggregate(t *testing.T) {
 	assert.Equal(t, int64(310), a.InputTokens, "input tokens summed")
 	assert.Equal(t, int64(135), a.OutputTokens, "output tokens summed")
 	assert.Equal(t, int64(445), a.TotalTokens, "total tokens summed")
-	assert.InDelta(t, 0.031, a.CostUSD, 1e-9, "cost summed")
+	assert.InDelta(t, 0.031, a.TotalCostUSD(), 1e-9, "cost summed")
 	assert.Equal(t, "deny", a.Decision, "any deny makes the session a deny")
 	assert.ElementsMatch(t, []string{"openai", "anthropic"}, a.Providers, "distinct providers")
 	assert.ElementsMatch(t, []string{"gpt-5.4", "claude-haiku-4-5"}, a.Models, "distinct models")

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

@@ -420,6 +420,47 @@ var providers = []Provider{
 			{ID: "mistral-embed", Label: "Mistral Embed", InputPer1k: 0.0001, OutputPer1k: 0, ContextWindow: 8192},
 		},
 	},
+	{
+		ID:                 "kimi_api",
+		Kind:               KindProvider,
+		Name:               "Kimi (Moonshot AI) API",
+		Description:        "Kimi K3 / K2 models via the Moonshot AI platform",
+		DefaultHost:        "api.moonshot.ai",
+		AuthHeaderName:     "Authorization",
+		AuthHeaderTemplate: "Bearer ${API_KEY}",
+		DefaultContentType: "application/json",
+		BrandColor:         "#1A1A2E",
+		// ParserID empty on purpose: Moonshot serves two body shapes on
+		// the same host and key, and the proxy's URL sniffer dispatches
+		// both (same pattern as Bifrost). /v1/chat/completions matches
+		// OpenAIParser; the Anthropic-compatible endpoint the official
+		// Claude Code guide uses (/anthropic/v1/messages) contains
+		// "/v1/messages" and matches AnthropicParser. Pinning "openai"
+		// here would misparse the Claude Code path — the primary way
+		// teams consume Kimi for coding today. Both endpoints accept the
+		// same Moonshot key via Authorization: Bearer (Claude Code's
+		// ANTHROPIC_AUTH_TOKEN rides that header too).
+		//
+		// api.moonshot.ai is the international platform; mainland-China
+		// accounts live on api.moonshot.cn with separate billing —
+		// operators there override the host on the provider record. The
+		// kimi.com subscription coding endpoint (api.kimi.com/coding,
+		// model id "k3") is account-bound seat licensing rather than a
+		// meterable platform key, so it's deliberately not the default.
+		ParserID: "",
+		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
+		// $3/$15 per MTok with $0.30 cached input, flat across the 1M-token
+		// window. kimi-k3 is the ONLY model the platform serves newer
+		// accounts — K2-era ids (kimi-k2-thinking) and even the kimi-latest
+		// alias return resource_not_found_error, verified live 2026-07-21 —
+		// so it's the only catalog entry. Grandfathered accounts with K2
+		// access can still type those ids on the provider's model rows.
+		// The consumer app's "K3 Swarm Max" mode is not an API SKU, so it
+		// doesn't appear here.
+		Models: []Model{
+			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+		},
+	},
 	{
 		ID:                 "litellm_proxy",
 		Kind:               KindGateway,

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

@@ -41,8 +41,24 @@ type AgentNetworkAccessLog struct {
 	InputTokens        int64
 	OutputTokens       int64
 	TotalTokens        int64
-	CostUSD            float64
-	Stream             bool
+	// Prompt-cache buckets: read + write token counts.
+	CachedInputTokens   int64
+	CacheCreationTokens int64
+	// Per-bucket cost breakdown — one column per token bucket the provider
+	// bills separately. These four are the only cost state stored: the total
+	// and the cache portion are derived on read (TotalCostUSD / CacheCostUSD)
+	// rather than stored alongside, so a stored aggregate can never drift out
+	// of step with the components it summarises.
+	//
+	// default:0 matters on upgrade: these columns are ALTER TABLE ADD COLUMN
+	// on an existing table, and without it every historical row holds NULL —
+	// which a raw SUM()/scan into float64 can't read. The default backfills
+	// them as 0, so pre-upgrade rows report an unknown split, not an error.
+	InputCostUSD         float64 `gorm:"not null;default:0"`
+	CachedInputCostUSD   float64 `gorm:"not null;default:0"`
+	CacheCreationCostUSD float64 `gorm:"not null;default:0"`
+	OutputCostUSD        float64 `gorm:"not null;default:0"`
+	Stream               bool
 
 	// Prompt capture. Only populated when prompt collection is enabled
 	// (account master switch AND policy guardrail). Heavy free text.
@@ -60,19 +76,44 @@ type AgentNetworkAccessLog struct {
 // the reverse-proxy AccessLogEntry table.
 func (AgentNetworkAccessLog) TableName() string { return "agent_network_access_log" }
 
+// CostUSDSQLExpr is the SQL sum of the per-bucket cost columns — the total cost
+// of a row. Used wherever a query has to sort or aggregate on total cost now
+// that no cost_usd column is stored. Plain arithmetic over NOT NULL columns, so
+// it stays portable across SQLite and Postgres.
+const CostUSDSQLExpr = "(input_cost_usd + cached_input_cost_usd + cache_creation_cost_usd + output_cost_usd)"
+
+// TotalCostUSD is the request's total cost: the sum of the four per-bucket
+// costs. Derived rather than stored so it cannot disagree with the breakdown.
+func (a *AgentNetworkAccessLog) TotalCostUSD() float64 {
+	return a.InputCostUSD + a.CachedInputCostUSD + a.CacheCreationCostUSD + a.OutputCostUSD
+}
+
+// CacheCostUSD is the portion of the total billed for prompt-cache buckets:
+// cache reads plus cache writes.
+func (a *AgentNetworkAccessLog) CacheCostUSD() float64 {
+	return a.CachedInputCostUSD + a.CacheCreationCostUSD
+}
+
 // ToAPIResponse renders the flattened entry as the API representation.
 func (a *AgentNetworkAccessLog) ToAPIResponse() api.AgentNetworkAccessLog {
 	out := api.AgentNetworkAccessLog{
-		Id:           a.ID,
-		ServiceId:    a.ServiceID,
-		Timestamp:    a.Timestamp,
-		StatusCode:   a.StatusCode,
-		DurationMs:   int(a.Duration.Milliseconds()),
-		InputTokens:  a.InputTokens,
-		OutputTokens: a.OutputTokens,
-		TotalTokens:  a.TotalTokens,
-		CostUsd:      a.CostUSD,
-		Stream:       &a.Stream,
+		Id:                   a.ID,
+		ServiceId:            a.ServiceID,
+		Timestamp:            a.Timestamp,
+		StatusCode:           a.StatusCode,
+		DurationMs:           int(a.Duration.Milliseconds()),
+		InputTokens:          a.InputTokens,
+		OutputTokens:         a.OutputTokens,
+		TotalTokens:          a.TotalTokens,
+		CachedInputTokens:    a.CachedInputTokens,
+		CacheCreationTokens:  a.CacheCreationTokens,
+		InputCostUsd:         a.InputCostUSD,
+		CachedInputCostUsd:   a.CachedInputCostUSD,
+		CacheCreationCostUsd: a.CacheCreationCostUSD,
+		OutputCostUsd:        a.OutputCostUSD,
+		CostUsd:              a.TotalCostUSD(),
+		CacheCostUsd:         a.CacheCostUSD(),
+		Stream:               &a.Stream,
 	}
 
 	out.UserId = strPtr(a.UserID)
@@ -112,20 +153,36 @@ func strPtr(s string) *string {
 // summary plus its ordered entries. Assembled in Go from a page of entries — it
 // is not a stored table.
 type AgentNetworkAccessLogSession struct {
-	SessionID    string // empty for a session-less (singleton) request
-	UserID       string
-	GroupIDs     []string // union of the entries' authorising groups
-	StartedAt    time.Time
-	EndedAt      time.Time
-	RequestCount int
-	InputTokens  int64
-	OutputTokens int64
-	TotalTokens  int64
-	CostUSD      float64
-	Providers    []string // distinct vendors seen in the session
-	Models       []string // distinct models seen in the session
-	Decision     string   // "deny" if any entry was denied, else "allow"
-	Entries      []*AgentNetworkAccessLog
+	SessionID            string // empty for a session-less (singleton) request
+	UserID               string
+	GroupIDs             []string // union of the entries' authorising groups
+	StartedAt            time.Time
+	EndedAt              time.Time
+	RequestCount         int
+	InputTokens          int64
+	OutputTokens         int64
+	TotalTokens          int64
+	CachedInputTokens    int64
+	CacheCreationTokens  int64
+	InputCostUSD         float64
+	CachedInputCostUSD   float64
+	CacheCreationCostUSD float64
+	OutputCostUSD        float64
+	Providers            []string // distinct vendors seen in the session
+	Models               []string // distinct models seen in the session
+	Decision             string   // "deny" if any entry was denied, else "allow"
+	Entries              []*AgentNetworkAccessLog
+}
+
+// TotalCostUSD is the session's total cost: the sum of the four per-bucket
+// costs accumulated across its entries.
+func (sess *AgentNetworkAccessLogSession) TotalCostUSD() float64 {
+	return sess.InputCostUSD + sess.CachedInputCostUSD + sess.CacheCreationCostUSD + sess.OutputCostUSD
+}
+
+// CacheCostUSD is the session's prompt-cache spend: cache reads plus writes.
+func (sess *AgentNetworkAccessLogSession) CacheCostUSD() float64 {
+	return sess.CachedInputCostUSD + sess.CacheCreationCostUSD
 }
 
 // sessionKey is the grouping key for an entry: its session id, or — when the
@@ -205,7 +262,12 @@ func (sess *AgentNetworkAccessLogSession) foldEntry(sk *sessionSeen, e *AgentNet
 	sess.InputTokens += e.InputTokens
 	sess.OutputTokens += e.OutputTokens
 	sess.TotalTokens += e.TotalTokens
-	sess.CostUSD += e.CostUSD
+	sess.CachedInputTokens += e.CachedInputTokens
+	sess.CacheCreationTokens += e.CacheCreationTokens
+	sess.InputCostUSD += e.InputCostUSD
+	sess.CachedInputCostUSD += e.CachedInputCostUSD
+	sess.CacheCreationCostUSD += e.CacheCreationCostUSD
+	sess.OutputCostUSD += e.OutputCostUSD
 	if e.Timestamp.Before(sess.StartedAt) {
 		sess.StartedAt = e.Timestamp
 	}
@@ -248,15 +310,22 @@ func (sess *AgentNetworkAccessLogSession) ToAPIResponse() api.AgentNetworkAccess
 	}
 
 	out := api.AgentNetworkAccessLogSession{
-		StartedAt:    sess.StartedAt,
-		EndedAt:      sess.EndedAt,
-		RequestCount: sess.RequestCount,
-		InputTokens:  sess.InputTokens,
-		OutputTokens: sess.OutputTokens,
-		TotalTokens:  sess.TotalTokens,
-		CostUsd:      sess.CostUSD,
-		Decision:     sess.Decision,
-		Entries:      entries,
+		StartedAt:            sess.StartedAt,
+		EndedAt:              sess.EndedAt,
+		RequestCount:         sess.RequestCount,
+		InputTokens:          sess.InputTokens,
+		OutputTokens:         sess.OutputTokens,
+		TotalTokens:          sess.TotalTokens,
+		CachedInputTokens:    sess.CachedInputTokens,
+		CacheCreationTokens:  sess.CacheCreationTokens,
+		InputCostUsd:         sess.InputCostUSD,
+		CachedInputCostUsd:   sess.CachedInputCostUSD,
+		CacheCreationCostUsd: sess.CacheCreationCostUSD,
+		OutputCostUsd:        sess.OutputCostUSD,
+		CostUsd:              sess.TotalCostUSD(),
+		CacheCostUsd:         sess.CacheCostUSD(),
+		Decision:             sess.Decision,
+		Entries:              entries,
 	}
 	out.SessionId = strPtr(sess.SessionID)
 	out.UserId = strPtr(sess.UserID)

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

@@ -54,7 +54,7 @@ var accessLogSortFields = map[string]string{
 	"provider":     "provider",
 	"status_code":  "status_code",
 	"duration":     "duration",
-	"cost_usd":     "cost_usd",
+	"cost_usd":     CostUSDSQLExpr,
 	"total_tokens": "total_tokens",
 	"user_id":      "user_id",
 	"decision":     "decision",
@@ -70,7 +70,7 @@ var accessLogSortFields = map[string]string{
 var sessionSortExprs = map[string]string{ //nolint:gosec // G101 false positive: "total_tokens" sort key, not a credential
 	"timestamp":     "MAX(timestamp)",
 	"started_at":    "MIN(timestamp)",
-	"cost_usd":      "SUM(cost_usd)",
+	"cost_usd":      "SUM" + CostUSDSQLExpr,
 	"total_tokens":  "SUM(total_tokens)",
 	"duration":      "SUM(duration)",
 	"request_count": "COUNT(*)",

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

@@ -0,0 +1,124 @@
+package types
+
+import (
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// costRow builds an access-log entry carrying only a cost breakdown — the rest
+// of the row is irrelevant to the summation identities under test.
+func costRow(id, session string, ts time.Time, in, cachedIn, cacheCreate, out float64) *AgentNetworkAccessLog {
+	return &AgentNetworkAccessLog{
+		ID:                   id,
+		SessionID:            session,
+		Timestamp:            ts,
+		InputCostUSD:         in,
+		CachedInputCostUSD:   cachedIn,
+		CacheCreationCostUSD: cacheCreate,
+		OutputCostUSD:        out,
+	}
+}
+
+// TestAPIResponse_CostComponentsSumToAggregates is the contract a client adding
+// up an API response depends on: within a single rendered object, the four
+// per-bucket fields sum to cost_usd, and the two cache fields sum to
+// cache_cost_usd. Uses rates that are not exactly representable in binary
+// floating point, so the identity is checked against real arithmetic rather
+// than round numbers.
+func TestAPIResponse_CostComponentsSumToAggregates(t *testing.T) {
+	row := costRow("r1", "s1", time.Now(), 0.000768, 0.0002304, 0.00192, 0.003)
+
+	api := row.ToAPIResponse()
+	assert.InDelta(t, api.InputCostUsd+api.CachedInputCostUsd+api.CacheCreationCostUsd+api.OutputCostUsd,
+		api.CostUsd, 1e-12, "rendered buckets must sum to the rendered cost_usd")
+	assert.InDelta(t, api.CachedInputCostUsd+api.CacheCreationCostUsd, api.CacheCostUsd, 1e-12,
+		"rendered cache buckets must sum to the rendered cache_cost_usd")
+	assert.InDelta(t, 0.0059184, api.CostUsd, 1e-12, "total is the exact sum, not a separately rounded figure")
+	assert.InDelta(t, 0.0021504, api.CacheCostUsd, 1e-12, "cache cost is the exact sum of the two cache buckets")
+}
+
+// TestSessionSummary_SumsMatchSummedEntries proves a session summary equals the
+// sum of the entries it renders: a client that adds up the entries itself must
+// land on the same number the summary reports, per bucket and in total.
+func TestSessionSummary_SumsMatchSummedEntries(t *testing.T) {
+	base := time.Date(2026, 5, 5, 10, 0, 0, 0, time.UTC)
+	entries := []*AgentNetworkAccessLog{
+		costRow("r1", "s1", base, 0.000768, 0.0002304, 0.00192, 0.003),
+		costRow("r2", "s1", base.Add(time.Minute), 0.000625, 0.0009375, 0, 0.005),
+		costRow("r3", "s1", base.Add(2*time.Minute), 0.0000016, 0, 0, 0.0000032),
+	}
+
+	sessions := FoldAccessLogSessions([]string{"s1"}, entries)
+	require.Len(t, sessions, 1)
+	sess := sessions[0].ToAPIResponse()
+
+	var wantInput, wantCachedInput, wantCacheCreation, wantOutput float64
+	for _, e := range entries {
+		wantInput += e.InputCostUSD
+		wantCachedInput += e.CachedInputCostUSD
+		wantCacheCreation += e.CacheCreationCostUSD
+		wantOutput += e.OutputCostUSD
+	}
+
+	assert.InDelta(t, wantInput, sess.InputCostUsd, 1e-12, "session input cost is the sum of its entries")
+	assert.InDelta(t, wantCachedInput, sess.CachedInputCostUsd, 1e-12, "session cache-read cost is the sum of its entries")
+	assert.InDelta(t, wantCacheCreation, sess.CacheCreationCostUsd, 1e-12, "session cache-write cost is the sum of its entries")
+	assert.InDelta(t, wantOutput, sess.OutputCostUsd, 1e-12, "session output cost is the sum of its entries")
+	assert.InDelta(t, wantInput+wantCachedInput+wantCacheCreation+wantOutput, sess.CostUsd, 1e-12,
+		"session total equals the summed entry buckets")
+
+	// Summing the rendered entries must give the same answer as reading the
+	// summary — the property a UI relies on when it totals a table itself.
+	var fromEntries float64
+	for _, e := range sess.Entries {
+		fromEntries += e.CostUsd
+	}
+	assert.InDelta(t, sess.CostUsd, fromEntries, 1e-12, "summary total must match the summed rendered entries")
+
+	// The sub-microdollar row must still contribute; it would vanish under
+	// 6-decimal quantisation.
+	assert.Greater(t, sess.InputCostUsd, 0.001393, "small-cost rows must not be quantised away")
+}
+
+// TestUsageBuckets_SumsMatchSummedRows proves the same identity one level up:
+// a usage bucket equals the sum of the ledger rows folded into it, and the
+// buckets together equal the whole range.
+func TestUsageBuckets_SumsMatchSummedRows(t *testing.T) {
+	day1 := time.Date(2026, 5, 5, 9, 0, 0, 0, time.UTC)
+	day2 := time.Date(2026, 5, 6, 9, 0, 0, 0, time.UTC)
+	rows := []*AgentNetworkUsage{
+		{ID: "u1", Timestamp: day1, InputCostUSD: 0.000768, CachedInputCostUSD: 0.0002304, CacheCreationCostUSD: 0.00192, OutputCostUSD: 0.003},
+		{ID: "u2", Timestamp: day1.Add(time.Hour), InputCostUSD: 0.000625, CachedInputCostUSD: 0.0009375, OutputCostUSD: 0.005},
+		{ID: "u3", Timestamp: day2, InputCostUSD: 0.0000016, OutputCostUSD: 0.0000032},
+	}
+
+	buckets := AggregateUsageByGranularity(rows, UsageGranularityDay)
+	require.Len(t, buckets, 2, "two distinct days expected")
+
+	var total, cache float64
+	for _, b := range buckets {
+		api := b.ToAPIResponse()
+		assert.InDelta(t, api.InputCostUsd+api.CachedInputCostUsd+api.CacheCreationCostUsd+api.OutputCostUsd,
+			api.CostUsd, 1e-12, "each bucket's components must sum to its cost_usd")
+		total += api.CostUsd
+		cache += api.CacheCostUsd
+	}
+
+	var wantTotal, wantCache float64
+	for _, r := range rows {
+		wantTotal += r.TotalCostUSD()
+		wantCache += r.CacheCostUSD()
+	}
+	assert.InDelta(t, wantTotal, total, 1e-12, "buckets must sum to the total across all ledger rows")
+	assert.InDelta(t, wantCache, cache, 1e-12, "buckets must sum to the cache spend across all ledger rows")
+
+	// A month bucket over the same rows must total identically — regrouping
+	// changes the partition, never the sum.
+	monthly := AggregateUsageByGranularity(rows, UsageGranularityMonth)
+	require.Len(t, monthly, 1)
+	assert.InDelta(t, wantTotal, monthly[0].ToAPIResponse().CostUsd, 1e-12,
+		"re-bucketing at a different granularity must preserve the total")
+}

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

@@ -25,8 +25,19 @@ type AgentNetworkUsage struct {
 	InputTokens        int64
 	OutputTokens       int64
 	TotalTokens        int64
-	CostUSD            float64
-	CreatedAt          time.Time
+	// Prompt-cache buckets: read + write token counts.
+	CachedInputTokens   int64
+	CacheCreationTokens int64
+	// Per-bucket cost breakdown, mirroring AgentNetworkAccessLog — the only
+	// cost state stored; total and cache portion are derived on read. Kept on
+	// the usage ledger too so spend can be attributed per bucket even for
+	// accounts with log collection turned off. See AgentNetworkAccessLog for
+	// why the columns carry a zero default.
+	InputCostUSD         float64 `gorm:"not null;default:0"`
+	CachedInputCostUSD   float64 `gorm:"not null;default:0"`
+	CacheCreationCostUSD float64 `gorm:"not null;default:0"`
+	OutputCostUSD        float64 `gorm:"not null;default:0"`
+	CreatedAt            time.Time
 }
 
 // TableName keeps usage records in their own stripped table. Named
@@ -34,6 +45,17 @@ type AgentNetworkUsage struct {
 // agent_network_usage table in a shared database.
 func (AgentNetworkUsage) TableName() string { return "agent_network_request_usage" }
 
+// TotalCostUSD is the request's total cost: the sum of the four per-bucket
+// costs. Derived rather than stored so it cannot disagree with the breakdown.
+func (u *AgentNetworkUsage) TotalCostUSD() float64 {
+	return u.InputCostUSD + u.CachedInputCostUSD + u.CacheCreationCostUSD + u.OutputCostUSD
+}
+
+// CacheCostUSD is the portion of the total billed for prompt-cache buckets.
+func (u *AgentNetworkUsage) CacheCostUSD() float64 {
+	return u.CachedInputCostUSD + u.CacheCreationCostUSD
+}
+
 // AgentNetworkUsageGroup is the normalised many-to-many row linking a usage
 // record to one authorising group, mirroring AgentNetworkAccessLogGroup so the
 // usage overview can filter by group with a `group_id IN (...)` join.

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

@@ -33,21 +33,45 @@ func ParseUsageGranularity(s string) UsageGranularity {
 // AgentNetworkUsageBucket is one aggregated usage time bucket. PeriodStart is
 // the UTC start of the bucket as YYYY-MM-DD.
 type AgentNetworkUsageBucket struct {
-	PeriodStart  string
-	InputTokens  int64
-	OutputTokens int64
-	TotalTokens  int64
-	CostUSD      float64
+	PeriodStart          string
+	InputTokens          int64
+	OutputTokens         int64
+	TotalTokens          int64
+	CachedInputTokens    int64
+	CacheCreationTokens  int64
+	InputCostUSD         float64
+	CachedInputCostUSD   float64
+	CacheCreationCostUSD float64
+	OutputCostUSD        float64
+}
+
+// TotalCostUSD is the bucket's total spend: the sum of the four per-bucket
+// costs. Derived rather than accumulated separately so it cannot disagree with
+// the components.
+func (b *AgentNetworkUsageBucket) TotalCostUSD() float64 {
+	return b.InputCostUSD + b.CachedInputCostUSD + b.CacheCreationCostUSD + b.OutputCostUSD
+}
+
+// CacheCostUSD is the bucket's prompt-cache spend: cache reads plus writes.
+func (b *AgentNetworkUsageBucket) CacheCostUSD() float64 {
+	return b.CachedInputCostUSD + b.CacheCreationCostUSD
 }
 
 // ToAPIResponse renders the bucket as the API representation.
 func (b *AgentNetworkUsageBucket) ToAPIResponse() api.AgentNetworkUsageBucket {
 	return api.AgentNetworkUsageBucket{
-		PeriodStart:  b.PeriodStart,
-		InputTokens:  b.InputTokens,
-		OutputTokens: b.OutputTokens,
-		TotalTokens:  b.TotalTokens,
-		CostUsd:      b.CostUSD,
+		PeriodStart:          b.PeriodStart,
+		InputTokens:          b.InputTokens,
+		OutputTokens:         b.OutputTokens,
+		TotalTokens:          b.TotalTokens,
+		CachedInputTokens:    b.CachedInputTokens,
+		CacheCreationTokens:  b.CacheCreationTokens,
+		InputCostUsd:         b.InputCostUSD,
+		CachedInputCostUsd:   b.CachedInputCostUSD,
+		CacheCreationCostUsd: b.CacheCreationCostUSD,
+		OutputCostUsd:        b.OutputCostUSD,
+		CostUsd:              b.TotalCostUSD(),
+		CacheCostUsd:         b.CacheCostUSD(),
 	}
 }
 
@@ -84,7 +108,12 @@ func AggregateUsageByGranularity(rows []*AgentNetworkUsage, g UsageGranularity)
 		b.InputTokens += r.InputTokens
 		b.OutputTokens += r.OutputTokens
 		b.TotalTokens += r.TotalTokens
-		b.CostUSD += r.CostUSD
+		b.CachedInputTokens += r.CachedInputTokens
+		b.CacheCreationTokens += r.CacheCreationTokens
+		b.InputCostUSD += r.InputCostUSD
+		b.CachedInputCostUSD += r.CachedInputCostUSD
+		b.CacheCreationCostUSD += r.CacheCreationCostUSD
+		b.OutputCostUSD += r.OutputCostUSD
 	}
 
 	out := make([]*AgentNetworkUsageBucket, 0, len(byPeriod))

management/internals/server/config/config.go

@@ -61,6 +61,10 @@ type Config struct {
 	// EmbeddedIdP contains configuration for the embedded Dex OIDC provider.
 	// When set, Dex will be embedded in the management server and serve requests at /oauth2/
 	EmbeddedIdP *idp.EmbeddedIdPConfig
+
+	HighestSupportedSyncMessageVersion *int
+
+	PerAccountHighestSupportedSyncMessageVersion map[string]int
 }
 
 // GetAuthAudiences returns the audience from the http config and device authorization flow config

management/internals/shared/grpc/components_encoder.go

@@ -0,0 +1,766 @@
+package grpc
+
+import (
+	"encoding/base64"
+	"strconv"
+
+	nbdns "github.com/netbirdio/netbird/dns"
+	"github.com/netbirdio/netbird/management/server/types"
+	nbroute "github.com/netbirdio/netbird/route"
+	"github.com/netbirdio/netbird/shared/management/networkmap"
+	"github.com/netbirdio/netbird/shared/management/proto"
+)
+
+// wgKeyRawLen is the raw byte length of a WireGuard public key.
+const wgKeyRawLen = 32
+
+// ComponentsEnvelopeInput bundles the data the component-format encoder needs.
+// The envelope is fully self-contained — every field needed by the client's
+// local Calculate() comes from the components struct itself. The only
+// externally-supplied data is the receiving peer's PeerConfig (which is
+// computed alongside the components in the network_map controller and reused
+// from the legacy proto path) and the dns_domain string.
+type ComponentsEnvelopeInput struct {
+	Components       *types.NetworkMapComponents
+	PeerConfig       *proto.PeerConfig
+	DNSDomain        string
+	DNSForwarderPort int64
+	// UserIDClaim is the OIDC claim name the client should embed in
+	// SshAuth.UserIDClaim when reconstructing the NetworkMap. Empty value
+	// is OK — client treats empty as "no SshAuth to build".
+	UserIDClaim string
+	// ProxyPatch carries pre-expanded NetworkMap fragments injected by
+	// external controllers (BYOP/port-forwarding). Nil when no proxy data
+	// is present; encoder skips the field in that case.
+	ProxyPatch *proto.ProxyPatch
+}
+
+// EncodeNetworkMapEnvelope converts NetworkMapComponents into the component
+// wire envelope. The encoder is intentionally non-deterministic: it iterates
+// Go maps in their native (random) order. Indexes inside the envelope
+// (peer_indexes, source_group_ids, agent_version_idx, router_peer_indexes)
+// are self-consistent within a single encode, so the decoder reconstructs
+// the same typed objects regardless of emit order. Tests that need to
+// compare envelopes do so semantically via proto round-trip + canonicalize,
+// not byte-equal.
+//
+// Callers must NOT concatenate or merge envelopes from different encodes —
+// index spaces are local to a single envelope.
+func EncodeNetworkMapEnvelope(in ComponentsEnvelopeInput) *proto.NetworkMapEnvelope {
+	c := in.Components
+
+	// Graceful degrade when components is nil — matches the legacy path's
+	// behaviour for missing/unvalidated peers (return a NetworkMap with only
+	// Network populated). The receiver gets an envelope it can decode
+	// without crashing; AccountSettings stays non-nil so client-side
+	// dereferences are safe.
+	if c.IsEmpty() {
+		// Match legacy missing-peer minimum: a NetworkMap with only Network
+		// populated. The receiver gets enough to bootstrap (Network
+		// identifier, dns_domain, account_settings) and the peer itself.
+		return &proto.NetworkMapEnvelope{
+			Payload: &proto.NetworkMapEnvelope_Full{
+				Full: &proto.NetworkMapComponentsFull{
+					PeerConfig: in.PeerConfig,
+					// components.Peers always contains the target peer
+					Peers:            []*proto.PeerCompact{toPeerCompact(c.Peers[c.PeerID])},
+					DnsDomain:        in.DNSDomain,
+					DnsForwarderPort: in.DNSForwarderPort,
+					UserIdClaim:      in.UserIDClaim,
+					AccountSettings:  &proto.AccountSettingsCompact{},
+					ProxyPatch:       in.ProxyPatch,
+				},
+			},
+		}
+	}
+
+	// Phase 1: build dedup tables. Every routing peer (in c.RouterPeers) and
+	// every regular peer (in c.Peers) must be indexed before any encoder
+	// looks up indexes via e.peerOrder — otherwise routes / routers_map for
+	// peers that exist only in c.RouterPeers would silently lose their
+	// peer_index reference.
+	enc := newComponentEncoder(c)
+	enc.indexAllPeers()
+	routerIdxs := enc.indexRouterPeers(c.RouterPeers)
+
+	// Phase 2: gather every policy that any consumer references (peer-pair
+	// policies + resource-only policies) so encodeResourcePoliciesMap can
+	// translate every *Policy pointer to a wire index.
+	allPolicies := unionPolicies(c.Policies, c.ResourcePoliciesMap)
+	policies := enc.encodePolicies(allPolicies)
+
+	// Phase 3: emit. Order of struct field expressions no longer matters:
+	// every encoder either reads from the dedup tables or works on
+	// independent input.
+	full := &proto.NetworkMapComponentsFull{
+		Serial:              networkSerial(c.Network),
+		PeerConfig:          in.PeerConfig,
+		Network:             toAccountNetwork(c.Network),
+		AccountSettings:     toAccountSettingsCompact(c.AccountSettings),
+		DnsForwarderPort:    in.DNSForwarderPort,
+		UserIdClaim:         in.UserIDClaim,
+		ProxyPatch:          in.ProxyPatch,
+		DnsSettings:         enc.encodeDNSSettings(c.DNSSettings),
+		DnsDomain:           in.DNSDomain,
+		CustomZoneDomain:    c.CustomZoneDomain,
+		AgentVersions:       enc.agentVersions,
+		Peers:               enc.peers,
+		RouterPeerIndexes:   routerIdxs,
+		Policies:            policies,
+		Groups:              enc.encodeGroups(),
+		Routes:              enc.encodeRoutes(c.Routes),
+		NameserverGroups:    enc.encodeNameServerGroups(c.NameServerGroups),
+		AllDnsRecords:       encodeSimpleRecords(c.AllDNSRecords),
+		AccountZones:        encodeCustomZones(c.AccountZones),
+		NetworkResources:    enc.encodeNetworkResources(c.NetworkResources),
+		RoutersMap:          enc.encodeRoutersMap(c.RoutersMap),
+		ResourcePoliciesMap: enc.encodeResourcePoliciesMap(c.ResourcePoliciesMap),
+		GroupIdToUserIds:    enc.encodeGroupIDToUserIDs(c.GroupIDToUserIDs),
+		AllowedUserIds:      stringSetToSlice(c.AllowedUserIDs),
+		PostureFailedPeers:  enc.encodePostureFailedPeers(c.PostureFailedPeers),
+	}
+
+	return &proto.NetworkMapEnvelope{
+		Payload: &proto.NetworkMapEnvelope_Full{Full: full},
+	}
+}
+
+// networkSerial returns c.Network.CurrentSerial() with a nil guard. The
+// production path always populates c.Network, but the encoder is exported
+// and a hand-built components struct may omit it.
+func networkSerial(n *types.Network) uint64 {
+	if n == nil {
+		return 0
+	}
+	return n.CurrentSerial()
+}
+
+type componentEncoder struct {
+	components *types.NetworkMapComponents
+
+	peerOrder map[string]uint32
+	peers     []*proto.PeerCompact
+
+	agentVersionOrder map[string]uint32
+	agentVersions     []string
+}
+
+func newComponentEncoder(c *types.NetworkMapComponents) *componentEncoder {
+	return &componentEncoder{
+		components:        c,
+		peerOrder:         make(map[string]uint32, len(c.Peers)),
+		peers:             make([]*proto.PeerCompact, 0, len(c.Peers)),
+		agentVersionOrder: make(map[string]uint32),
+	}
+}
+
+func (e *componentEncoder) indexAllPeers() {
+	for _, p := range e.components.Peers {
+		if p == nil {
+			continue
+		}
+		e.appendPeer(p)
+	}
+}
+
+func (e *componentEncoder) appendPeer(p *types.ComponentPeer) uint32 {
+	if idx, ok := e.peerOrder[p.ID]; ok {
+		return idx
+	}
+	idx := uint32(len(e.peers))
+	e.peerOrder[p.ID] = idx
+	e.peers = append(e.peers, toPeerCompact(p))
+	return idx
+}
+
+// indexRouterPeers ensures every router peer is in the peer dedup table
+// (c.RouterPeers may contain peers not in c.Peers when validation rules drop
+// them) and returns their wire indexes for the RouterPeerIndexes field. Must
+// run before any encoder that resolves peer ids via e.peerOrder.
+func (e *componentEncoder) indexRouterPeers(routers map[string]*types.ComponentPeer) []uint32 {
+	if len(routers) == 0 {
+		return nil
+	}
+	out := make([]uint32, 0, len(routers))
+	for _, p := range routers {
+		if p == nil {
+			continue
+		}
+		out = append(out, e.appendPeer(p))
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeGroups() []*proto.GroupCompact {
+	if len(e.components.Groups) == 0 {
+		return nil
+	}
+
+	out := make([]*proto.GroupCompact, 0, len(e.components.Groups))
+	for _, g := range e.components.Groups {
+		peerIdxs := make([]uint32, 0, len(g.Peers))
+		for _, peerID := range g.Peers {
+			if idx, ok := e.peerOrder[peerID]; ok {
+				peerIdxs = append(peerIdxs, idx)
+			}
+		}
+		out = append(out, &proto.GroupCompact{
+			Id:          g.PublicID,
+			PeerIndexes: peerIdxs,
+			IsAll:       g.IsGroupAll(),
+		})
+	}
+	return out
+}
+
+// encodePolicies flattens Policy{Rules} → []PolicyCompact. Returns the wire
+// list and a map from policy pointer to the indexes of its emitted rules in
+// that list — used by encodeResourcePoliciesMap to translate
+// ResourcePoliciesMap[resourceID][]*Policy into wire-side indexes.
+func (e *componentEncoder) encodePolicies(policies []*types.Policy) []*proto.PolicyCompact {
+	if len(policies) == 0 {
+		return nil
+	}
+
+	out := make([]*proto.PolicyCompact, 0, len(policies))
+
+	for _, pol := range policies {
+		if !pol.Enabled {
+			continue
+		}
+		for _, r := range pol.Rules {
+			if r == nil || !r.Enabled {
+				continue
+			}
+			out = append(out, e.encodePolicyRule(pol, r))
+		}
+	}
+	return out
+}
+
+// encodePolicyRule maps a single PolicyRule under pol to a PolicyCompact entry.
+func (e *componentEncoder) encodePolicyRule(pol *types.Policy, r *types.PolicyRule) *proto.PolicyCompact {
+	return &proto.PolicyCompact{
+		Id:                    pol.PublicID,
+		Action:                networkmap.GetProtoAction(string(r.Action)),
+		Protocol:              networkmap.GetProtoProtocol(string(r.Protocol)),
+		Bidirectional:         r.Bidirectional,
+		Ports:                 portsToUint32(r.Ports),
+		PortRanges:            portRangesToProto(r.PortRanges),
+		SourceGroupIds:        e.groupPublicXids(r.Sources),
+		DestinationGroupIds:   e.groupPublicXids(r.Destinations),
+		AuthorizedUser:        r.AuthorizedUser,
+		AuthorizedGroups:      e.encodeAuthorizedGroups(r.AuthorizedGroups),
+		SourceResource:        e.resourceToProto(r.SourceResource),
+		DestinationResource:   e.resourceToProto(r.DestinationResource),
+		SourcePostureCheckIds: e.postureCheckSeqs(pol.SourcePostureChecks),
+	}
+}
+
+// groupPublicXids maps the xid group IDs in src to their public xids,
+// dropping any group with invalid public xid.
+func (e *componentEncoder) groupPublicXids(src []string) []string {
+	if len(src) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(src))
+	for _, gid := range src {
+		if id, ok := e.groupPublicXid(gid); ok {
+			out = append(out, id)
+		}
+	}
+	return out
+}
+
+// unionPolicies merges c.Policies with every policy referenced by
+// c.ResourcePoliciesMap, deduplicating by pointer identity. Resource-only
+// policies (relevant to a NetworkResource but not to peer-pair traffic)
+// only live in ResourcePoliciesMap; without this union step they'd be lost
+// from the wire and the client's resource-policy lookup would come back
+// empty.
+func unionPolicies(policies []*types.Policy, resourcePolicies map[string][]*types.Policy) []*types.Policy {
+	// Fast path: non-router peers have no resource-only policies, so the
+	// "union" is identical to `policies`. Skip the dedup map allocation.
+	if len(resourcePolicies) == 0 {
+		return policies
+	}
+	seen := make(map[string]struct{}, len(policies))
+	out := make([]*types.Policy, 0, len(policies))
+	for _, p := range policies {
+		if p == nil {
+			continue
+		}
+		if _, ok := seen[p.ID]; ok {
+			continue
+		}
+		seen[p.ID] = struct{}{}
+		out = append(out, p)
+	}
+	for _, list := range resourcePolicies {
+		for _, p := range list {
+			if p == nil {
+				continue
+			}
+			if _, ok := seen[p.ID]; ok {
+				continue
+			}
+			seen[p.ID] = struct{}{}
+			out = append(out, p)
+		}
+	}
+	return out
+}
+
+// encodeAuthorizedGroups translates rule.AuthorizedGroups (map keyed by
+// group xid → local-user names) to the wire form (map keyed by group
+// account_seq_id → UserNameList). Groups without a seq id are dropped —
+// matches how source/destination group references handle the same case.
+func (e *componentEncoder) encodeAuthorizedGroups(m map[string][]string) map[string]*proto.UserNameList {
+	if len(m) == 0 {
+		return nil
+	}
+	out := make(map[string]*proto.UserNameList, len(m))
+	for groupID, names := range m {
+		id, ok := e.groupPublicXid(groupID)
+		if !ok {
+			continue
+		}
+		out[id] = &proto.UserNameList{Names: names}
+	}
+	return out
+}
+
+func (e *componentEncoder) groupPublicXid(groupID string) (string, bool) {
+	g, ok := e.components.Groups[groupID]
+	if !ok {
+		return "", false
+	}
+	return g.PublicID, true
+}
+
+// resourceToProto translates types.Resource for the wire. For peer-typed
+// resources the peer id is converted to a peer index into the envelope's
+// peers array. For other resource types only the type string is shipped
+// today (Calculate's resource-typed rule path consults SourceResource only
+// for "peer" — other types fall through to group-based lookup).
+func (e *componentEncoder) resourceToProto(r types.Resource) *proto.ResourceCompact {
+	if r.ID == "" && r.Type == "" {
+		return nil
+	}
+	out := &proto.ResourceCompact{Type: string(r.Type)}
+	if r.Type == types.ResourceTypePeer && r.ID != "" {
+		if idx, ok := e.peerOrder[r.ID]; ok {
+			out.PeerIndexSet = true
+			out.PeerIndex = idx
+		}
+	}
+	return out
+}
+
+// postureCheckSeqs translates a slice of posture-check xids to their
+// public xids. Unresolvable xids are silently dropped — matches how group/peer
+// references handle the same case.
+func (e *componentEncoder) postureCheckSeqs(xids []string) []string {
+	if len(xids) == 0 || len(e.components.PostureCheckXIDToPublicID) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(xids))
+	for _, xid := range xids {
+		if seq, ok := e.components.PostureCheckXIDToPublicID[xid]; ok {
+			out = append(out, seq)
+		}
+	}
+	return out
+}
+
+// networkSeq translates a Network xid to its public id using
+// the NetworkMapComponents.NetworkXIDToPublicID lookup. Returns (0,false) when
+// the xid isn't known — callers decide whether to skip the parent record.
+func (e *componentEncoder) networkPublicId(xid string) (string, bool) {
+	if xid == "" {
+		return "", false
+	}
+	id, ok := e.components.NetworkXIDToPublicID[xid]
+	if !ok {
+		return "", false
+	}
+	return id, true
+}
+
+func (e *componentEncoder) encodeDNSSettings(s *types.DNSSettings) *proto.DNSSettingsCompact {
+	if s == nil || len(s.DisabledManagementGroups) == 0 {
+		return nil
+	}
+	out := &proto.DNSSettingsCompact{
+		DisabledManagementGroupIds: make([]string, 0, len(s.DisabledManagementGroups)),
+	}
+	for _, gid := range s.DisabledManagementGroups {
+		if id, ok := e.groupPublicXid(gid); ok {
+			out.DisabledManagementGroupIds = append(out.DisabledManagementGroupIds, id)
+		}
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeRoutes(routes []*nbroute.Route) []*proto.RouteRaw {
+	if len(routes) == 0 {
+		return nil
+	}
+	out := make([]*proto.RouteRaw, 0, len(routes))
+	for _, r := range routes {
+		if r == nil {
+			continue
+		}
+		rr := &proto.RouteRaw{
+			Id:                    r.PublicID,
+			NetId:                 string(r.NetID),
+			Description:           r.Description,
+			KeepRoute:             r.KeepRoute,
+			NetworkType:           int32(r.NetworkType),
+			Masquerade:            r.Masquerade,
+			Metric:                int32(r.Metric),
+			Enabled:               r.Enabled,
+			SkipAutoApply:         r.SkipAutoApply,
+			Domains:               r.Domains.ToPunycodeList(),
+			GroupIds:              e.groupPublicXids(r.Groups),
+			AccessControlGroupIds: e.groupPublicXids(r.AccessControlGroups),
+			PeerGroupIds:          e.groupPublicXids(r.PeerGroups),
+		}
+		if r.Network.IsValid() {
+			rr.NetworkCidr = r.Network.String()
+		}
+		if r.Peer != "" {
+			if idx, ok := e.peerOrder[r.Peer]; ok {
+				rr.PeerIndexSet = true
+				rr.PeerIndex = idx
+			}
+		}
+		out = append(out, rr)
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeNameServerGroups(nsgs []*nbdns.NameServerGroup) []*proto.NameServerGroupRaw {
+	if len(nsgs) == 0 {
+		return nil
+	}
+	out := make([]*proto.NameServerGroupRaw, 0, len(nsgs))
+	for _, nsg := range nsgs {
+		if nsg == nil {
+			continue
+		}
+		entry := &proto.NameServerGroupRaw{
+			Id:                   nsg.PublicID,
+			Nameservers:          encodeNameServers(nsg.NameServers),
+			GroupIds:             e.groupPublicXids(nsg.Groups),
+			Primary:              nsg.Primary,
+			Domains:              nsg.Domains,
+			Enabled:              nsg.Enabled,
+			SearchDomainsEnabled: nsg.SearchDomainsEnabled,
+		}
+		out = append(out, entry)
+	}
+	return out
+}
+
+func encodeNameServers(servers []nbdns.NameServer) []*proto.NameServer {
+	if len(servers) == 0 {
+		return nil
+	}
+	out := make([]*proto.NameServer, 0, len(servers))
+	for _, s := range servers {
+		out = append(out, &proto.NameServer{
+			IP:     s.IP.String(),
+			NSType: int64(s.NSType),
+			Port:   int64(s.Port),
+		})
+	}
+	return out
+}
+
+func encodeSimpleRecords(records []nbdns.SimpleRecord) []*proto.SimpleRecord {
+	if len(records) == 0 {
+		return nil
+	}
+	out := make([]*proto.SimpleRecord, 0, len(records))
+	for _, r := range records {
+		out = append(out, &proto.SimpleRecord{
+			Name:  r.Name,
+			Type:  int64(r.Type),
+			Class: r.Class,
+			TTL:   int64(r.TTL),
+			RData: r.RData,
+		})
+	}
+	return out
+}
+
+func encodeCustomZones(zones []nbdns.CustomZone) []*proto.CustomZone {
+	if len(zones) == 0 {
+		return nil
+	}
+	out := make([]*proto.CustomZone, 0, len(zones))
+	for _, z := range zones {
+		out = append(out, &proto.CustomZone{
+			Domain:               z.Domain,
+			Records:              encodeSimpleRecords(z.Records),
+			SearchDomainDisabled: z.SearchDomainDisabled,
+			NonAuthoritative:     z.NonAuthoritative,
+		})
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeNetworkResources(resources []*types.ComponentResource) []*proto.NetworkResourceRaw {
+	if len(resources) == 0 {
+		return nil
+	}
+	out := make([]*proto.NetworkResourceRaw, 0, len(resources))
+	for _, r := range resources {
+		if r == nil {
+			continue
+		}
+		entry := &proto.NetworkResourceRaw{
+			Id:          r.PublicID,
+			Name:        r.Name,
+			Description: r.Description,
+			Type:        string(r.Type),
+			Address:     r.Address,
+			DomainValue: r.Domain,
+			Enabled:     r.Enabled,
+		}
+		if id, ok := e.networkPublicId(r.NetworkID); ok {
+			entry.NetworkSeq = id
+		}
+		if r.Prefix.IsValid() {
+			entry.PrefixCidr = r.Prefix.String()
+		}
+		out = append(out, entry)
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeRoutersMap(routersMap map[string]map[string]*types.ComponentRouter) map[string]*proto.NetworkRouterList {
+	if len(routersMap) == 0 {
+		return nil
+	}
+	out := make(map[string]*proto.NetworkRouterList, len(routersMap))
+	for networkXID, routers := range routersMap {
+		if len(routers) == 0 {
+			continue
+		}
+		id, ok := e.networkPublicId(networkXID)
+		if !ok {
+			continue
+		}
+		entries := make([]*proto.NetworkRouterEntry, 0, len(routers))
+		for peerID, r := range routers {
+			if r == nil {
+				continue
+			}
+			entry := &proto.NetworkRouterEntry{
+				Id:           r.PublicID,
+				PeerGroupIds: e.groupPublicXids(r.PeerGroups),
+				Masquerade:   r.Masquerade,
+				Metric:       int32(r.Metric),
+				Enabled:      r.Enabled,
+			}
+			if idx, ok := e.peerOrder[peerID]; ok {
+				entry.PeerIndexSet = true
+				entry.PeerIndex = idx
+			}
+			entries = append(entries, entry)
+		}
+		out[id] = &proto.NetworkRouterList{Entries: entries}
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeResourcePoliciesMap(rpm map[string][]*types.Policy) map[string]*proto.PolicyIds {
+	if len(rpm) == 0 {
+		return nil
+	}
+	// resourceXIDToPublicID is local to one encode — built from components.NetworkResources
+	// (small slice). Network resources without seq id are dropped, matching how
+	// other components-without-seq are silently filtered.
+	resourceXIDToPublicID := make(map[string]string, len(e.components.NetworkResources))
+	for _, r := range e.components.NetworkResources {
+		if r != nil {
+			resourceXIDToPublicID[r.ID] = r.PublicID
+		}
+	}
+	out := make(map[string]*proto.PolicyIds, len(rpm))
+	for resourceXID, policies := range rpm {
+		resId, ok := resourceXIDToPublicID[resourceXID]
+		if !ok {
+			continue
+		}
+		ids := make([]string, 0, len(policies))
+		for _, pol := range policies {
+			ids = append(ids, pol.PublicID)
+		}
+		if len(ids) == 0 {
+			continue
+		}
+		out[resId] = &proto.PolicyIds{Ids: ids}
+	}
+	return out
+}
+
+func (e *componentEncoder) encodeGroupIDToUserIDs(m map[string][]string) map[string]*proto.UserIDList {
+	if len(m) == 0 {
+		return nil
+	}
+	out := make(map[string]*proto.UserIDList, len(m))
+	for groupID, userIDs := range m {
+		id, ok := e.groupPublicXid(groupID)
+		if !ok || len(userIDs) == 0 {
+			continue
+		}
+		out[id] = &proto.UserIDList{UserIds: userIDs}
+	}
+	return out
+}
+
+func stringSetToSlice(s map[string]struct{}) []string {
+	if len(s) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(s))
+	for k := range s {
+		out = append(out, k)
+	}
+	return out
+}
+
+func (e *componentEncoder) encodePostureFailedPeers(m map[string]map[string]struct{}) map[string]*proto.PeerIndexSet {
+	if len(m) == 0 {
+		return nil
+	}
+	out := make(map[string]*proto.PeerIndexSet, len(m))
+	for checkXID, failedPeerIDs := range m {
+		id, ok := e.components.PostureCheckXIDToPublicID[checkXID]
+		if !ok {
+			continue
+		}
+		idxs := make([]uint32, 0, len(failedPeerIDs))
+		for peerID := range failedPeerIDs {
+			if idx, ok := e.peerOrder[peerID]; ok {
+				idxs = append(idxs, idx)
+			}
+		}
+		if len(idxs) == 0 {
+			continue
+		}
+		out[id] = &proto.PeerIndexSet{PeerIndexes: idxs}
+	}
+	return out
+}
+
+// toAccountSettingsCompact always returns a non-nil message — the client
+// dereferences it unconditionally during Calculate(), so a nil here would
+// crash the receiver. A missing types.AccountSettingsInfo on the server
+// (which shouldn't happen in production but the encoder is exported)
+// degrades to login_expiration_enabled = false, which makes
+// LoginExpired() return false for every peer.
+func toAccountSettingsCompact(s *types.AccountSettingsInfo) *proto.AccountSettingsCompact {
+	if s == nil {
+		return &proto.AccountSettingsCompact{}
+	}
+	return &proto.AccountSettingsCompact{
+		PeerLoginExpirationEnabled: s.PeerLoginExpirationEnabled,
+		PeerLoginExpirationNs:      int64(s.PeerLoginExpiration),
+	}
+}
+
+func toAccountNetwork(n *types.Network) *proto.AccountNetwork {
+	if n == nil {
+		return nil
+	}
+	out := &proto.AccountNetwork{
+		Identifier: n.Identifier,
+		NetCidr:    n.Net.String(),
+		Dns:        n.Dns,
+		Serial:     n.CurrentSerial(),
+	}
+	if len(n.NetV6.IP) > 0 {
+		out.NetV6Cidr = n.NetV6.String()
+	}
+	return out
+}
+
+func toPeerCompact(p *types.ComponentPeer) *proto.PeerCompact {
+	pc := &proto.PeerCompact{
+		WgPubKey:               decodeWgKey(p.Key),
+		SshPubKey:              []byte(p.SSHKey),
+		DnsLabel:               p.DNSLabel,
+		AgentVersion:           p.AgentVersion,
+		AddedWithSsoLogin:      p.AddedWithSSOLogin,
+		LoginExpirationEnabled: p.LoginExpirationEnabled,
+		SshEnabled:             p.SSHEnabled,
+		SupportsIpv6:           p.SupportsIPv6,
+		SupportsSourcePrefixes: p.SupportsSourcePrefixes,
+		ServerSshAllowed:       p.ServerSSHAllowed,
+	}
+	if !p.LastLogin.IsZero() {
+		pc.LastLoginUnixNano = p.LastLogin.UnixNano()
+	}
+	switch {
+	case !p.IP.IsValid():
+		// leave Ip nil
+	case p.IP.Is4() || p.IP.Is4In6():
+		ip := p.IP.Unmap().As4()
+		pc.Ip = ip[:]
+	default:
+		ip := p.IP.As16()
+		pc.Ip = ip[:]
+	}
+	if p.IPv6.IsValid() {
+		ip := p.IPv6.As16()
+		pc.Ipv6 = ip[:]
+	}
+	return pc
+}
+
+// decodeWgKey returns the raw 32 bytes of a base64-encoded WireGuard public
+// key, or nil for an empty / malformed key.
+func decodeWgKey(s string) []byte {
+	if s == "" {
+		return nil
+	}
+	out := make([]byte, wgKeyRawLen)
+	n, err := base64.StdEncoding.Decode(out, []byte(s))
+	if err != nil || n != wgKeyRawLen {
+		return nil
+	}
+	return out
+}
+
+func portsToUint32(ports []string) []uint32 {
+	if len(ports) == 0 {
+		return nil
+	}
+	out := make([]uint32, 0, len(ports))
+	for _, p := range ports {
+		v, err := strconv.ParseUint(p, 10, 16)
+		if err != nil {
+			continue
+		}
+		out = append(out, uint32(v))
+	}
+	return out
+}
+
+func portRangesToProto(ranges []types.RulePortRange) []*proto.PortInfo_Range {
+	if len(ranges) == 0 {
+		return nil
+	}
+	out := make([]*proto.PortInfo_Range, 0, len(ranges))
+	for _, r := range ranges {
+		out = append(out, &proto.PortInfo_Range{
+			Start: uint32(r.Start),
+			End:   uint32(r.End),
+		})
+	}
+	return out
+}

management/internals/shared/grpc/components_encoder_test.go

@@ -0,0 +1,781 @@
+package grpc
+
+import (
+	"bytes"
+	"cmp"
+	"net"
+	"net/netip"
+	"slices"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	goproto "google.golang.org/protobuf/proto"
+
+	nbdns "github.com/netbirdio/netbird/dns"
+	"github.com/netbirdio/netbird/management/server/types"
+	nbroute "github.com/netbirdio/netbird/route"
+	"github.com/netbirdio/netbird/shared/management/proto"
+)
+
+const testWgKeyA = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopq="
+const testWgKeyB = "BBCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopq="
+const testWgKeyC = "CBCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopq="
+
+// canonicalize rewrites a NetworkMapComponentsFull in place into a canonical
+// form: peers reordered by wg_pub_key, with the rest of the message rewritten
+// to reference the new peer indexes. Groups, policies, and router indexes are
+// also sorted. After canonicalize, two envelopes built from the same logical
+// input compare byte-equal via proto.Equal.
+//
+// This lives on the test side — the encoder itself emits in map-iteration
+// order. Test-side normalization is the contract for "two encodes are
+// equivalent".
+func canonicalize(full *proto.NetworkMapComponentsFull) {
+	if full == nil {
+		return
+	}
+
+	type peerEntry struct {
+		peer   *proto.PeerCompact
+		oldIdx uint32
+	}
+	entries := make([]peerEntry, len(full.Peers))
+	for i, p := range full.Peers {
+		entries[i] = peerEntry{peer: p, oldIdx: uint32(i)}
+	}
+	// DnsLabel is unique per peer; it tiebreaks on equal WgPubKey (e.g. both
+	// nil from malformed keys, or both empty for placeholders).
+	slices.SortFunc(entries, func(a, b peerEntry) int {
+		if c := bytes.Compare(a.peer.WgPubKey, b.peer.WgPubKey); c != 0 {
+			return c
+		}
+		return cmp.Compare(a.peer.DnsLabel, b.peer.DnsLabel)
+	})
+
+	remap := make(map[uint32]uint32, len(entries))
+	newPeers := make([]*proto.PeerCompact, len(entries))
+	for newIdx, e := range entries {
+		remap[e.oldIdx] = uint32(newIdx)
+		newPeers[newIdx] = e.peer
+	}
+	full.Peers = newPeers
+
+	full.RouterPeerIndexes = remapAndSort(full.RouterPeerIndexes, remap)
+	for _, g := range full.Groups {
+		g.PeerIndexes = remapAndSort(g.PeerIndexes, remap)
+	}
+	slices.SortFunc(full.Groups, func(a, b *proto.GroupCompact) int { return cmp.Compare(a.Id, b.Id) })
+
+	for _, r := range full.Routes {
+		if r.PeerIndexSet {
+			if newIdx, ok := remap[r.PeerIndex]; ok {
+				r.PeerIndex = newIdx
+			}
+		}
+		slices.Sort(r.GroupIds)
+		slices.Sort(r.AccessControlGroupIds)
+		slices.Sort(r.PeerGroupIds)
+	}
+	slices.SortFunc(full.Routes, func(a, b *proto.RouteRaw) int { return cmp.Compare(a.Id, b.Id) })
+
+	for _, list := range full.RoutersMap {
+		for _, entry := range list.Entries {
+			if entry.PeerIndexSet {
+				if newIdx, ok := remap[entry.PeerIndex]; ok {
+					entry.PeerIndex = newIdx
+				}
+			}
+			slices.Sort(entry.PeerGroupIds)
+		}
+		slices.SortFunc(list.Entries, func(a, b *proto.NetworkRouterEntry) int { return cmp.Compare(a.Id, b.Id) })
+	}
+
+	for _, set := range full.PostureFailedPeers {
+		set.PeerIndexes = remapAndSort(set.PeerIndexes, remap)
+	}
+
+	for _, p := range full.Policies {
+		slices.Sort(p.SourceGroupIds)
+		slices.Sort(p.DestinationGroupIds)
+	}
+	// Sort policies by (Id, source_group_ids, destination_group_ids) so that
+	// multiple PolicyCompact entries sharing the same Id (one per rule, when
+	// a Policy has multiple rules) still get a deterministic order. After
+	// sorting we remap indexes in ResourcePoliciesMap.
+	policyOldOrder := make(map[*proto.PolicyCompact]uint32, len(full.Policies))
+	for i, p := range full.Policies {
+		policyOldOrder[p] = uint32(i)
+	}
+	slices.SortFunc(full.Policies, func(a, b *proto.PolicyCompact) int {
+		if c := cmp.Compare(a.Id, b.Id); c != 0 {
+			return c
+		}
+		if c := slices.Compare(a.SourceGroupIds, b.SourceGroupIds); c != 0 {
+			return c
+		}
+		return slices.Compare(a.DestinationGroupIds, b.DestinationGroupIds)
+	})
+	policyRemap := make(map[uint32]uint32, len(full.Policies))
+	for newIdx, p := range full.Policies {
+		policyRemap[policyOldOrder[p]] = uint32(newIdx)
+	}
+	for _, idxs := range full.ResourcePoliciesMap {
+		slices.Sort(idxs.Ids)
+	}
+	for _, list := range full.GroupIdToUserIds {
+		slices.Sort(list.UserIds)
+	}
+	slices.Sort(full.AllowedUserIds)
+}
+
+func remapAndSort(idxs []uint32, remap map[uint32]uint32) []uint32 {
+	out := make([]uint32, 0, len(idxs))
+	for _, i := range idxs {
+		if newIdx, ok := remap[i]; ok {
+			out = append(out, newIdx)
+		}
+	}
+	slices.Sort(out)
+	return out
+}
+
+// envelopesEquivalent decodes both envelopes, canonicalizes them, and reports
+// whether they're proto.Equal. Use instead of byte-comparing marshaled output:
+// the encoder is intentionally non-deterministic.
+func envelopesEquivalent(a, b *proto.NetworkMapEnvelope) bool {
+	canonicalize(a.GetFull())
+	canonicalize(b.GetFull())
+	return goproto.Equal(a, b)
+}
+
+func newTestComponents() *types.NetworkMapComponents {
+	peerA := &types.ComponentPeer{
+		ID:           "peer-a",
+		Key:          testWgKeyA,
+		IP:           netip.AddrFrom4([4]byte{100, 64, 0, 1}),
+		DNSLabel:     "peera",
+		SSHKey:       "ssh-a",
+		AgentVersion: "0.40.0",
+	}
+	peerB := &types.ComponentPeer{
+		ID:           "peer-b",
+		Key:          testWgKeyB,
+		IP:           netip.AddrFrom4([4]byte{100, 64, 0, 2}),
+		IPv6:         netip.AddrFrom16([16]byte{0xfd, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2}),
+		DNSLabel:     "peerb",
+		AgentVersion: "0.25.0",
+	}
+	peerC := &types.ComponentPeer{
+		ID:           "peer-c",
+		Key:          testWgKeyC,
+		IP:           netip.AddrFrom4([4]byte{100, 64, 0, 3}),
+		DNSLabel:     "peerc",
+		AgentVersion: "0.40.0",
+	}
+
+	return &types.NetworkMapComponents{
+		PeerID: "peer-a",
+		Network: &types.Network{
+			Identifier: "net-test",
+			Net:        net.IPNet{IP: net.IP{100, 64, 0, 0}, Mask: net.CIDRMask(10, 32)},
+			Serial:     7,
+		},
+		AccountSettings: &types.AccountSettingsInfo{
+			PeerLoginExpirationEnabled: true,
+			PeerLoginExpiration:        2 * time.Hour,
+		},
+		Peers: map[string]*types.ComponentPeer{
+			"peer-a": peerA,
+			"peer-b": peerB,
+			"peer-c": peerC,
+		},
+		Groups: map[string]*types.ComponentGroup{
+			"group-src": {ID: "group-src", PublicID: "1", Name: "Src", Peers: []string{"peer-a"}},
+			"group-dst": {ID: "group-dst", PublicID: "2", Name: "Dst", Peers: []string{"peer-b", "peer-c"}},
+		},
+		Policies: []*types.Policy{
+			{
+				ID:       "pol-1",
+				PublicID: "10",
+				Enabled:  true,
+				Rules: []*types.PolicyRule{{
+					ID: "rule-1", Enabled: true, Action: types.PolicyTrafficActionAccept,
+					Protocol: types.PolicyRuleProtocolTCP, Bidirectional: true,
+					Ports:        []string{"22", "80"},
+					PortRanges:   []types.RulePortRange{{Start: 8000, End: 8100}},
+					Sources:      []string{"group-src"},
+					Destinations: []string{"group-dst"},
+				}},
+			},
+		},
+		RouterPeers: map[string]*types.ComponentPeer{"peer-c": peerC},
+	}
+}
+
+func TestEncodeNetworkMapEnvelope_Basic(t *testing.T) {
+	c := newTestComponents()
+	env := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{
+		Components: c,
+		DNSDomain:  "netbird.cloud",
+	})
+
+	require.NotNil(t, env)
+	full := env.GetFull()
+	require.NotNil(t, full, "envelope must contain Full payload")
+
+	assert.EqualValues(t, 7, full.Serial)
+	assert.Equal(t, "netbird.cloud", full.DnsDomain)
+
+	require.NotNil(t, full.Network)
+	assert.Equal(t, "net-test", full.Network.Identifier)
+	assert.Equal(t, "100.64.0.0/10", full.Network.NetCidr)
+
+	require.NotNil(t, full.AccountSettings)
+	assert.True(t, full.AccountSettings.PeerLoginExpirationEnabled)
+	assert.EqualValues(t, (2 * time.Hour).Nanoseconds(), full.AccountSettings.PeerLoginExpirationNs)
+
+	require.Len(t, full.Peers, 3)
+	byLabel := map[string]*proto.PeerCompact{}
+	for _, p := range full.Peers {
+		assert.Len(t, p.WgPubKey, 32, "wg key must be raw 32 bytes")
+		assert.Len(t, p.Ip, 4, "ipv4 must be raw 4 bytes")
+		byLabel[p.DnsLabel] = p
+	}
+	assert.Len(t, byLabel["peerb"].Ipv6, 16, "peer-b has ipv6 → 16 bytes")
+}
+
+func TestEncodeNetworkMapEnvelope_RepeatEncodesEquivalent(t *testing.T) {
+	c := newTestComponents()
+
+	expected := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+
+	// Hammer it 100 times — Go map iteration is randomized per call, so each
+	// run produces different wire bytes, but the canonicalized form must
+	// match.
+	for i := 0; i < 100; i++ {
+		got := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+		require.True(t, envelopesEquivalent(expected, got),
+			"encode #%d must be semantically equivalent to first encode", i)
+	}
+}
+
+func TestEncodeNetworkMapEnvelope_ConcurrentEncodesEquivalent(t *testing.T) {
+	c := newTestComponents()
+
+	expected := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+
+	const goroutines = 50
+	var wg sync.WaitGroup
+	wg.Add(goroutines)
+	results := make([]*proto.NetworkMapEnvelope, goroutines)
+	for i := 0; i < goroutines; i++ {
+		i := i
+		go func() {
+			defer wg.Done()
+			results[i] = EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+		}()
+	}
+	wg.Wait()
+
+	for i, got := range results {
+		require.NotNil(t, got, "goroutine %d returned nil", i)
+		require.True(t, envelopesEquivalent(expected, got),
+			"goroutine %d produced inequivalent envelope", i)
+	}
+}
+
+func TestEncodeNetworkMapEnvelope_GroupsByAccountPublicId(t *testing.T) {
+	c := newTestComponents()
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.Groups, 2)
+
+	groupByID := map[string]*proto.GroupCompact{}
+	for _, g := range full.Groups {
+		groupByID[g.Id] = g
+	}
+	require.Contains(t, groupByID, "1")
+	require.Contains(t, groupByID, "2")
+	assert.Len(t, groupByID["1"].PeerIndexes, 1)
+	assert.Len(t, groupByID["2"].PeerIndexes, 2)
+}
+
+func TestEncodeNetworkMapEnvelope_PolicyExpansion(t *testing.T) {
+	c := newTestComponents()
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.Policies, 1)
+	pc := full.Policies[0]
+	assert.EqualValues(t, "10", pc.Id)
+	assert.Equal(t, proto.RuleAction_ACCEPT, pc.Action)
+	assert.Equal(t, proto.RuleProtocol_TCP, pc.Protocol)
+	assert.True(t, pc.Bidirectional)
+	assert.Equal(t, []uint32{22, 80}, pc.Ports)
+	require.Len(t, pc.PortRanges, 1)
+	assert.EqualValues(t, 8000, pc.PortRanges[0].Start)
+	assert.EqualValues(t, 8100, pc.PortRanges[0].End)
+	assert.Equal(t, []string{"1"}, pc.SourceGroupIds)
+	assert.Equal(t, []string{"2"}, pc.DestinationGroupIds)
+}
+
+func TestEncodeNetworkMapEnvelope_RouterIndexes(t *testing.T) {
+	c := newTestComponents()
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.RouterPeerIndexes, 1)
+	idx := full.RouterPeerIndexes[0]
+	require.Less(t, int(idx), len(full.Peers))
+	assert.Equal(t, "peerc", full.Peers[idx].DnsLabel)
+}
+
+func TestEncodeNetworkMapEnvelope_DisabledPolicySkipped(t *testing.T) {
+	c := newTestComponents()
+	c.Policies[0].Enabled = false
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	assert.Empty(t, full.Policies)
+}
+
+func TestEncodeNetworkMapEnvelope_TwoPeersSameMalformedKey(t *testing.T) {
+	// Both peers have nil WgPubKey after decode; canonicalize must still
+	// produce a stable order using DnsLabel as a tiebreaker, so 100 encodes
+	// canonicalize identically.
+	c := newTestComponents()
+	c.Peers["peer-a"].Key = "garbage-a-!!!"
+	c.Peers["peer-b"].Key = "garbage-b-!!!"
+
+	expected := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+	for i := 0; i < 100; i++ {
+		got := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+		require.True(t, envelopesEquivalent(expected, got),
+			"encode #%d with two same-key peers must canonicalize equivalently", i)
+	}
+}
+
+func TestEncodeNetworkMapEnvelope_MalformedWgKey(t *testing.T) {
+	c := newTestComponents()
+	c.Peers["peer-a"].Key = "not-base64-!!!"
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.Peers, 3)
+
+	var byLabel = map[string]*proto.PeerCompact{}
+	for _, p := range full.Peers {
+		byLabel[p.DnsLabel] = p
+	}
+	assert.Nil(t, byLabel["peera"].WgPubKey, "peer with malformed key encodes nil WgPubKey")
+	assert.Len(t, byLabel["peerb"].WgPubKey, 32, "other peers retain their key")
+}
+
+func TestEncodeNetworkMapEnvelope_IPv6OnlyPeer(t *testing.T) {
+	c := newTestComponents()
+	v6Only := &types.ComponentPeer{
+		ID:           "peer-v6",
+		Key:          testWgKeyA,
+		IPv6:         netip.AddrFrom16([16]byte{0xfd, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9}),
+		DNSLabel:     "peerv6",
+		AgentVersion: "0.40.0",
+	}
+	c.Peers["peer-v6"] = v6Only
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	var found *proto.PeerCompact
+	for _, p := range full.Peers {
+		if p.DnsLabel == "peerv6" {
+			found = p
+		}
+	}
+	require.NotNil(t, found, "ipv6-only peer must be present")
+	assert.Empty(t, found.Ip, "no IPv4 address → empty Ip")
+	assert.Len(t, found.Ipv6, 16)
+}
+
+func TestEncodeNetworkMapEnvelope_PeerWithoutIP(t *testing.T) {
+	c := newTestComponents()
+	c.Peers["peer-noip"] = &types.ComponentPeer{
+		ID:           "peer-noip",
+		Key:          testWgKeyA,
+		DNSLabel:     "peernoip",
+		AgentVersion: "0.40.0",
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	var found *proto.PeerCompact
+	for _, p := range full.Peers {
+		if p.DnsLabel == "peernoip" {
+			found = p
+		}
+	}
+	require.NotNil(t, found)
+	assert.Empty(t, found.Ip)
+	assert.Empty(t, found.Ipv6)
+}
+
+func TestEncodeNetworkMapEnvelope_EmptyInput(t *testing.T) {
+	c := &types.NetworkMapComponents{
+		Network: &types.Network{Identifier: "x", Net: net.IPNet{IP: net.IP{100, 64, 0, 0}, Mask: net.CIDRMask(10, 32)}},
+	}
+
+	env := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c})
+
+	full := env.GetFull()
+	require.NotNil(t, full)
+	assert.Empty(t, full.Peers)
+	assert.Empty(t, full.Groups)
+	assert.Empty(t, full.Policies)
+	assert.Empty(t, full.RouterPeerIndexes)
+	require.NotNil(t, full.AccountSettings, "AccountSettingsCompact must always be emitted (client dereferences it unconditionally)")
+}
+
+func TestEncodeNetworkMapEnvelope_PeerLoginExpirationFields(t *testing.T) {
+	c := newTestComponents()
+	now := time.Date(2024, 1, 2, 3, 4, 5, 0, time.UTC)
+	c.Peers["peer-a"].AddedWithSSOLogin = true
+	c.Peers["peer-a"].LoginExpirationEnabled = true
+	c.Peers["peer-a"].LastLogin = now
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	var pa *proto.PeerCompact
+	for _, p := range full.Peers {
+		if p.DnsLabel == "peera" {
+			pa = p
+		}
+	}
+	require.NotNil(t, pa)
+	assert.True(t, pa.AddedWithSsoLogin)
+	assert.True(t, pa.LoginExpirationEnabled)
+	assert.Equal(t, now.UnixNano(), pa.LastLoginUnixNano)
+
+	// peer-b has no UserID and no LastLogin → all fields zero-value.
+	var pb *proto.PeerCompact
+	for _, p := range full.Peers {
+		if p.DnsLabel == "peerb" {
+			pb = p
+		}
+	}
+	require.NotNil(t, pb)
+	assert.False(t, pb.AddedWithSsoLogin)
+	assert.False(t, pb.LoginExpirationEnabled)
+	assert.Zero(t, pb.LastLoginUnixNano)
+}
+
+func TestEncodeNetworkMapEnvelope_RoutesRoundTrip(t *testing.T) {
+	c := newTestComponents()
+	c.Routes = []*nbroute.Route{
+		{
+			ID:                  "route-peer",
+			PublicID:            "100",
+			NetID:               "net-A",
+			Description:         "via peer-c",
+			Network:             netip.MustParsePrefix("10.0.0.0/16"),
+			Peer:                "peer-c", // peer ID, not WG key
+			Groups:              []string{"group-src"},
+			AccessControlGroups: []string{"group-dst"},
+			Enabled:             true,
+		},
+		{
+			ID:         "route-peergroup",
+			PublicID:   "101",
+			NetID:      "net-B",
+			Network:    netip.MustParsePrefix("10.1.0.0/16"),
+			PeerGroups: []string{"group-src", "group-dst"},
+			Enabled:    true,
+		},
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.Routes, 2)
+	byNetID := map[string]*proto.RouteRaw{}
+	for _, r := range full.Routes {
+		byNetID[r.NetId] = r
+	}
+
+	r1 := byNetID["net-A"]
+	require.NotNil(t, r1)
+	assert.True(t, r1.PeerIndexSet, "route with peer must set peer_index_set")
+	require.Less(t, int(r1.PeerIndex), len(full.Peers))
+	assert.Equal(t, "peerc", full.Peers[r1.PeerIndex].DnsLabel)
+	assert.Equal(t, []string{"1"}, r1.GroupIds, "group-src has AccountSeqID 1")
+	assert.Equal(t, []string{"2"}, r1.AccessControlGroupIds, "group-dst has AccountSeqID 2")
+	assert.Empty(t, r1.PeerGroupIds)
+
+	r2 := byNetID["net-B"]
+	require.NotNil(t, r2)
+	assert.False(t, r2.PeerIndexSet, "route with peer_groups must NOT set peer_index_set")
+	assert.ElementsMatch(t, []string{"1", "2"}, r2.PeerGroupIds)
+}
+
+func TestEncodeNetworkMapEnvelope_RouteWithMissingPeerLeavesIndexUnset(t *testing.T) {
+	c := newTestComponents()
+	c.Routes = []*nbroute.Route{{
+		ID:       "route-x",
+		PublicID: "100",
+		Peer:     "peer-not-in-components",
+		Network:  netip.MustParsePrefix("10.0.0.0/16"),
+		Enabled:  true,
+	}}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.Routes, 1)
+	assert.False(t, full.Routes[0].PeerIndexSet,
+		"missing peer reference must not pretend to point at peer index 0")
+}
+
+func TestEncodeNetworkMapEnvelope_ResourceOnlyPolicyShippedAndIndexed(t *testing.T) {
+	c := newTestComponents()
+	// Policy that exists ONLY in ResourcePoliciesMap, not in c.Policies. This
+	// is the I1 case — without unionPolicies the encoder would silently
+	// drop it from the wire.
+	resourceOnlyPolicy := &types.Policy{
+		ID: "pol-resource", PublicID: "99", Enabled: true,
+		Rules: []*types.PolicyRule{{
+			ID: "rule-r", Enabled: true, Action: types.PolicyTrafficActionAccept,
+			Protocol:     types.PolicyRuleProtocolTCP,
+			Sources:      []string{"group-src"},
+			Destinations: []string{"group-dst"},
+		}},
+	}
+	c.ResourcePoliciesMap = map[string][]*types.Policy{
+		"resource-x": {c.Policies[0], resourceOnlyPolicy}, // shared + resource-only
+	}
+	// Resource must appear in components.NetworkResources with a seq id —
+	// encoder uses that to translate the xid map key to uint32.
+	c.NetworkResources = []*types.ComponentResource{
+		{ID: "resource-x", PublicID: "77", Name: "res-x", Enabled: true},
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.Policies, 2, "encoded policies must include both peer-traffic and resource-only")
+
+	policyByID := map[string]*proto.PolicyCompact{}
+	policyIds := make([]string, 0)
+	for _, p := range full.Policies {
+		policyByID[p.Id] = p
+		policyIds = append(policyIds, p.Id)
+	}
+	require.Contains(t, policyByID, "10", "original peer-traffic policy id 10")
+	require.Contains(t, policyByID, "99", "resource-only policy id 99")
+
+	require.Contains(t, full.ResourcePoliciesMap, "77")
+	ids := full.ResourcePoliciesMap["77"].Ids
+	require.Len(t, ids, 2)
+	assert.ElementsMatch(t, policyIds, ids,
+		"resource policies map must reference both wire policy indexes")
+}
+
+func TestEncodeNetworkMapEnvelope_NameServerGroups(t *testing.T) {
+	c := newTestComponents()
+	c.NameServerGroups = []*nbdns.NameServerGroup{{
+		ID: "nsg-1", PublicID: "50", Name: "Main", Description: "primary",
+		NameServers: []nbdns.NameServer{{
+			IP: netip.MustParseAddr("8.8.8.8"), NSType: nbdns.UDPNameServerType, Port: 53,
+		}},
+		Groups:  []string{"group-src", "group-not-persisted"},
+		Primary: true, Enabled: true,
+		Domains: []string{"corp.example"},
+	}}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.NameserverGroups, 1)
+	nsg := full.NameserverGroups[0]
+	assert.EqualValues(t, "50", nsg.Id)
+	assert.True(t, nsg.Primary)
+	require.Len(t, nsg.Nameservers, 1)
+	assert.Equal(t, "8.8.8.8", nsg.Nameservers[0].IP)
+	assert.Equal(t, []string{"1"}, nsg.GroupIds)
+}
+
+func TestEncodeNetworkMapEnvelope_PostureFailedPeers(t *testing.T) {
+	c := newTestComponents()
+	c.PostureCheckXIDToPublicID = map[string]string{"check-1": "33"}
+	c.PostureFailedPeers = map[string]map[string]struct{}{
+		"check-1": {
+			"peer-a":              {},
+			"peer-b":              {},
+			"peer-not-in-account": {},
+		},
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Contains(t, full.PostureFailedPeers, "33")
+	idxs := full.PostureFailedPeers["33"].PeerIndexes
+	assert.Len(t, idxs, 2, "missing peer is silently dropped (filterPostureFailedPeers guarantees presence in real data)")
+}
+
+func TestEncodeNetworkMapEnvelope_RoutersMap(t *testing.T) {
+	c := newTestComponents()
+	c.NetworkXIDToPublicID = map[string]string{"net-1": "5"}
+	c.RoutersMap = map[string]map[string]*types.ComponentRouter{
+		"net-1": {
+			"peer-c": {
+				PublicID: "200",
+				Peer:     "peer-c", Masquerade: true, Metric: 10, Enabled: true,
+			},
+		},
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Contains(t, full.RoutersMap, "5")
+	entries := full.RoutersMap["5"].Entries
+	require.Len(t, entries, 1)
+	e := entries[0]
+	assert.EqualValues(t, "200", e.Id)
+	assert.True(t, e.PeerIndexSet)
+	require.Less(t, int(e.PeerIndex), len(full.Peers))
+	assert.Equal(t, "peerc", full.Peers[e.PeerIndex].DnsLabel)
+	assert.True(t, e.Masquerade)
+	assert.EqualValues(t, 10, e.Metric)
+	assert.True(t, e.Enabled)
+}
+
+func TestEncodeNetworkMapEnvelope_RouterPeerNotInComponentsPeers(t *testing.T) {
+	// Router peer in c.RouterPeers but NOT in c.Peers (validation may have
+	// filtered it). indexRouterPeers runs before encodeRoutersMap, so the
+	// peer_index reference must still resolve.
+	c := newTestComponents()
+	delete(c.Peers, "peer-c")
+	routerPeer := &types.ComponentPeer{
+		ID: "peer-c", Key: testWgKeyC, IP: netip.AddrFrom4([4]byte{100, 64, 0, 3}),
+		DNSLabel: "peerc", AgentVersion: "0.40.0",
+	}
+	c.RouterPeers = map[string]*types.ComponentPeer{"peer-c": routerPeer}
+	c.NetworkXIDToPublicID = map[string]string{"net-1": "5"}
+	c.RoutersMap = map[string]map[string]*types.ComponentRouter{
+		"net-1": {"peer-c": {PublicID: "1", Peer: "peer-c", Enabled: true}},
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Contains(t, full.RoutersMap, "5")
+	require.Len(t, full.RoutersMap["5"].Entries, 1)
+	e := full.RoutersMap["5"].Entries[0]
+	assert.True(t, e.PeerIndexSet, "router peer must be indexed even when not in c.Peers")
+}
+
+func TestEncodeNetworkMapEnvelope_GroupIDToUserIDs(t *testing.T) {
+	c := newTestComponents()
+	c.GroupIDToUserIDs = map[string][]string{
+		"group-src":     {"user-1", "user-2"},
+		"group-missing": {"user-4"}, // group not in components → drop
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.Len(t, full.GroupIdToUserIds, 1, "only present groups survive")
+	require.Contains(t, full.GroupIdToUserIds, "1")
+	assert.ElementsMatch(t, []string{"user-1", "user-2"}, full.GroupIdToUserIds["1"].UserIds)
+}
+
+func TestToProxyPatch_EmptyInputReturnsNil(t *testing.T) {
+	assert.Nil(t, toProxyPatch(nil, "netbird.cloud", false, false))
+	assert.Nil(t, toProxyPatch(&types.NetworkMap{}, "netbird.cloud", false, false),
+		"empty NetworkMap (no peers, rules, routes etc) → nil patch so proto3 omits the field")
+}
+
+func TestToProxyPatch_PopulatesAllFields(t *testing.T) {
+	nm := &types.NetworkMap{
+		Peers: []*types.ComponentPeer{{
+			ID: "ext-peer", Key: testWgKeyA, IP: netip.AddrFrom4([4]byte{100, 64, 0, 9}),
+			DNSLabel: "extpeer", AgentVersion: "0.40.0",
+		}},
+		FirewallRules: []*types.FirewallRule{{
+			PeerIP: "100.64.0.9", Action: "accept", Direction: 0, Protocol: "tcp",
+		}},
+	}
+
+	patch := toProxyPatch(nm, "netbird.cloud", false, false)
+
+	require.NotNil(t, patch)
+	assert.Len(t, patch.Peers, 1)
+	assert.Len(t, patch.FirewallRules, 1)
+}
+
+// TestEncodeNetworkMapEnvelope_ProxyPatchPropagated covers the ProxyPatch
+// pass-through in both encoder branches (normal path + nil-Components
+// graceful-degrade). Guards against a regression that drops `ProxyPatch:`
+// from one of the envelope struct literals.
+func TestEncodeNetworkMapEnvelope_ProxyPatchPropagated(t *testing.T) {
+	patch := &proto.ProxyPatch{
+		ForwardingRules: []*proto.ForwardingRule{{
+			Protocol:          proto.RuleProtocol_TCP,
+			DestinationPort:   &proto.PortInfo{PortSelection: &proto.PortInfo_Port{Port: 80}},
+			TranslatedAddress: net.IPv4(10, 0, 0, 1).To4(),
+			TranslatedPort:    &proto.PortInfo{PortSelection: &proto.PortInfo_Port{Port: 8080}},
+		}},
+	}
+
+	t.Run("normal_path", func(t *testing.T) {
+		c := newTestComponents()
+		full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{
+			Components: c,
+			ProxyPatch: patch,
+		}).GetFull()
+
+		require.NotNil(t, full.ProxyPatch, "ProxyPatch must propagate through the normal encode path")
+		assert.Len(t, full.ProxyPatch.ForwardingRules, 1)
+	})
+
+	t.Run("empty_components_graceful_degrade", func(t *testing.T) {
+		full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{
+			Components: emptyNetworkMapComponents(),
+			ProxyPatch: patch,
+		}).GetFull()
+
+		require.NotNil(t, full.ProxyPatch, "ProxyPatch must propagate through the nil-Components branch too")
+		assert.Len(t, full.ProxyPatch.ForwardingRules, 1)
+	})
+}
+
+func TestEncodeNetworkMapEnvelope_NilComponentsGracefulDegrade(t *testing.T) {
+	// nil Components → minimal envelope, no crash. Matches the legacy
+	// behaviour for missing/unvalidated peers.
+	env := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{
+		Components: emptyNetworkMapComponents(),
+		DNSDomain:  "netbird.cloud",
+	})
+
+	require.NotNil(t, env)
+	full := env.GetFull()
+	require.NotNil(t, full)
+	require.NotNil(t, full.AccountSettings, "AccountSettings must always be non-nil")
+	assert.Equal(t, "netbird.cloud", full.DnsDomain)
+	assert.Len(t, full.Peers, 1)
+	assert.Empty(t, full.Policies)
+}
+
+func TestEncodeNetworkMapEnvelope_AccountSettingsAlwaysEmitted(t *testing.T) {
+	c := &types.NetworkMapComponents{
+		Network: &types.Network{Identifier: "x", Net: net.IPNet{IP: net.IP{100, 64, 0, 0}, Mask: net.CIDRMask(10, 32)}},
+		// AccountSettings deliberately nil
+	}
+
+	full := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{Components: c}).GetFull()
+
+	require.NotNil(t, full.AccountSettings, "client dereferences AccountSettings unconditionally during Calculate(); a nil here would crash the receiver")
+	assert.False(t, full.AccountSettings.PeerLoginExpirationEnabled)
+	assert.Zero(t, full.AccountSettings.PeerLoginExpirationNs)
+}
+
+func emptyNetworkMapComponents() *types.NetworkMapComponents {
+	return types.EmptyNetworkMapComponents(
+		&types.NetworkMapComponents{
+			PeerID: "peer-id", Peers: map[string]*types.ComponentPeer{"peer-id": {}}},
+	)
+}

management/internals/shared/grpc/components_envelope_response.go

@@ -0,0 +1,200 @@
+package grpc
+
+import (
+	"context"
+
+	integrationsConfig "github.com/netbirdio/management-integrations/integrations/config"
+
+	"github.com/netbirdio/netbird/client/ssh/auth"
+	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
+	nbpeer "github.com/netbirdio/netbird/management/server/peer"
+	"github.com/netbirdio/netbird/management/server/posture"
+	"github.com/netbirdio/netbird/management/server/types"
+	sharedgrpc "github.com/netbirdio/netbird/shared/management/grpc"
+	"github.com/netbirdio/netbird/shared/management/networkmap"
+	"github.com/netbirdio/netbird/shared/management/proto"
+)
+
+// ToComponentSyncResponse builds a SyncResponse carrying the compact
+// NetworkMapEnvelope for capability-aware peers. The legacy proto.NetworkMap
+// field is intentionally left empty — capable peers ignore it and the
+// envelope alone is the authoritative wire shape.
+//
+// PeerConfig is computed once server-side using the receiving peer's own
+// account-level network metadata. EnableSSH inside PeerConfig is left at
+// peer.SSHEnabled (the peer's local setting); account-policy-driven SSH is
+// computed by the client from the envelope's GroupIDToUserIDs / AllowedUserIDs
+// inside Calculate(), so the SshConfig.SshEnabled bit may flip true on the
+// client even though the server-side PeerConfig reports false.
+func ToComponentSyncResponse(
+	ctx context.Context,
+	config *nbconfig.Config,
+	httpConfig *nbconfig.HttpServerConfig,
+	deviceFlowConfig *nbconfig.DeviceAuthorizationFlow,
+	peer *nbpeer.Peer,
+	turnCredentials *Token,
+	relayCredentials *Token,
+	components *types.NetworkMapComponents,
+	proxyPatch *types.NetworkMap,
+	dnsName string,
+	checks []*posture.Checks,
+	settings *types.Settings,
+	extraSettings *types.ExtraSettings,
+	peerGroups []string,
+	dnsFwdPort int64,
+) *proto.SyncResponse {
+	//
+	// 'component' parameter is expected to never be nil
+	// 'peer' parameter is expected to never be nil
+	//
+	// TODO (dmitri) consider using invariants?
+	//
+	enableSSH := computeSSHEnabledForPeer(components, peer)
+	peerConfig := toPeerConfig(peer, components.Network, dnsName, settings, httpConfig, deviceFlowConfig, enableSSH)
+
+	includeIPv6 := peer.SupportsIPv6() && peer.IPv6.IsValid()
+	useSourcePrefixes := peer.SupportsSourcePrefixes()
+
+	userIDClaim := auth.DefaultUserIDClaim
+	if httpConfig != nil && httpConfig.AuthUserIDClaim != "" {
+		userIDClaim = httpConfig.AuthUserIDClaim
+	}
+
+	envelope := EncodeNetworkMapEnvelope(ComponentsEnvelopeInput{
+		Components:       components,
+		PeerConfig:       peerConfig,
+		DNSDomain:        dnsName,
+		DNSForwarderPort: dnsFwdPort,
+		UserIDClaim:      userIDClaim,
+		ProxyPatch:       toProxyPatch(proxyPatch, dnsName, includeIPv6, useSourcePrefixes),
+	})
+
+	resp := &proto.SyncResponse{
+		PeerConfig:         peerConfig,
+		NetworkMapEnvelope: envelope,
+		Checks:             toProtocolChecks(ctx, checks),
+		Version:            int32(sharedgrpc.ComponentNetworkMap),
+	}
+
+	nbConfig := toNetbirdConfig(config, turnCredentials, relayCredentials, extraSettings, settings)
+	resp.NetbirdConfig = integrationsConfig.ExtendNetBirdConfig(peer.ID, peerGroups, nbConfig, extraSettings)
+
+	// settings == nil → field stays nil → "no info in this snapshot", client
+	// preserves the deadline it already had. settings non-nil → emit either a
+	// valid deadline or the explicit-zero "disabled" sentinel via
+	// encodeSessionExpiresAt.
+	if settings != nil {
+		resp.SessionExpiresAt = encodeSessionExpiresAt(
+			peer.SessionExpiresAt(settings.PeerLoginExpirationEnabled, settings.PeerLoginExpiration),
+		)
+	}
+
+	return resp
+}
+
+// toProxyPatch converts a proxy-injected *types.NetworkMap into the wire
+// patch the components envelope ships alongside. Returns nil when there are
+// no fragments to merge — proto3 omits a nil message field, so the receiver
+// sees no patch and skips the merge step entirely.
+//
+// We reuse the legacy proto-conversion helpers (toProtocolRoutes,
+// toProtocolFirewallRules, toProtocolRoutesFirewallRules,
+// appendRemotePeerConfig, ForwardingRule.ToProto) because the proxy
+// delivers fragments pre-expanded — there's no raw component shape to
+// derive them from. Components purity isn't violated: proxy data isn't
+// policy-graph-derived, it's externally injected post-Calculate, so the
+// client merges it on top of its locally-computed NetworkMap.
+func toProxyPatch(nm *types.NetworkMap, dnsName string, includeIPv6, useSourcePrefixes bool) *proto.ProxyPatch {
+	if nm == nil {
+		return nil
+	}
+	if len(nm.Peers) == 0 && len(nm.OfflinePeers) == 0 && len(nm.FirewallRules) == 0 &&
+		len(nm.Routes) == 0 && len(nm.RoutesFirewallRules) == 0 && len(nm.ForwardingRules) == 0 {
+		return nil
+	}
+
+	patch := &proto.ProxyPatch{
+		Peers:              networkmap.AppendRemotePeerConfig(nil, nm.Peers, dnsName, includeIPv6),
+		OfflinePeers:       networkmap.AppendRemotePeerConfig(nil, nm.OfflinePeers, dnsName, includeIPv6),
+		FirewallRules:      networkmap.ToProtocolFirewallRules(nm.FirewallRules, includeIPv6, useSourcePrefixes),
+		Routes:             networkmap.ToProtocolRoutes(nm.Routes),
+		RouteFirewallRules: networkmap.ToProtocolRoutesFirewallRules(nm.RoutesFirewallRules),
+	}
+	if len(nm.ForwardingRules) > 0 {
+		patch.ForwardingRules = make([]*proto.ForwardingRule, 0, len(nm.ForwardingRules))
+		for _, r := range nm.ForwardingRules {
+			patch.ForwardingRules = append(patch.ForwardingRules, r.ToProto())
+		}
+	}
+	return patch
+}
+
+// computeSSHEnabledForPeer mirrors the SSH-server-activation bit that
+// Calculate() folds into NetworkMap.EnableSSH. Components-format peers
+// receive a freshly-computed PeerConfig.SshConfig.SshEnabled at sync time;
+// without this helper the field would be incorrectly false for any peer
+// that's the destination of an SSH-enabling policy without having
+// peer.SSHEnabled set locally.
+//
+// Mirrors the two activation paths Calculate() uses:
+//  1. Explicit: rule.Protocol == NetbirdSSH and peer is in the rule's
+//     destinations.
+//  2. Legacy implicit: rule covers TCP/22 or TCP/22022 (or ALL), peer is in
+//     destinations, AND the peer has SSHEnabled set locally — this is the
+//     "allow-all/TCP-22 implies SSH activation for SSH-capable peers" path.
+//
+// The full SSH AuthorizedUsers map is still produced by the client when it
+// runs Calculate() over the envelope.
+func computeSSHEnabledForPeer(c *types.NetworkMapComponents, peer *nbpeer.Peer) bool {
+	if c == nil || peer == nil {
+		return false
+	}
+	// Mirror Calculate's `getAllPeersFromGroups` invariant: target peer must
+	// exist in c.Peers, otherwise no rule applies to it.
+	if _, ok := c.Peers[peer.ID]; !ok {
+		return false
+	}
+	for _, policy := range c.Policies {
+		if policy == nil || !policy.Enabled {
+			continue
+		}
+		for _, rule := range policy.Rules {
+			if ruleEnablesSSHForPeer(c, rule, peer) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// ruleEnablesSSHForPeer returns true when rule is active, targets peer, and
+// either explicitly authorises SSH or covers the legacy TCP/22 path while the
+// peer itself has SSH enabled locally.
+func ruleEnablesSSHForPeer(c *types.NetworkMapComponents, rule *types.PolicyRule, peer *nbpeer.Peer) bool {
+	if rule == nil || !rule.Enabled {
+		return false
+	}
+	if !peerInDestinations(c, rule, peer.ID) {
+		return false
+	}
+	if rule.Protocol == types.PolicyRuleProtocolNetbirdSSH {
+		return true
+	}
+	return peer.SSHEnabled && types.PolicyRuleImpliesLegacySSH(rule)
+}
+
+// peerInDestinations reports whether peerID is in any of rule.Destinations'
+// groups (or matches DestinationResource if it's a peer-typed resource —
+// for non-peer types Calculate falls through to group lookup, so we mirror
+// that exactly to avoid silent divergence).
+func peerInDestinations(c *types.NetworkMapComponents, rule *types.PolicyRule, peerID string) bool {
+	if rule.DestinationResource.Type == types.ResourceTypePeer && rule.DestinationResource.ID != "" {
+		return rule.DestinationResource.ID == peerID
+	}
+	for _, groupID := range rule.Destinations {
+		if c.IsPeerInGroup(peerID, groupID) {
+			return true
+		}
+	}
+	return false
+}

management/internals/shared/grpc/components_envelope_response_test.go

@@ -0,0 +1,184 @@
+package grpc
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	nbpeer "github.com/netbirdio/netbird/management/server/peer"
+	"github.com/netbirdio/netbird/management/server/types"
+)
+
+// TestComputeSSHEnabledForPeer covers both Calculate-mirroring branches:
+// explicit NetbirdSSH protocol, and the legacy implicit case where a
+// TCP/22 (or 22022 / ALL / port-range-covering-22) rule activates SSH when
+// the destination peer has SSHEnabled=true locally.
+func TestComputeSSHEnabledForPeer(t *testing.T) {
+	const targetPeerID = "target"
+	const targetGroupID = "g_dst"
+
+	mkComponents := func(rule *types.PolicyRule, sshEnabled bool) (*types.NetworkMapComponents, *nbpeer.Peer) {
+		peer := &nbpeer.Peer{ID: targetPeerID, SSHEnabled: sshEnabled}
+		group := &types.ComponentGroup{ID: targetGroupID, Name: "dst", Peers: []string{targetPeerID}}
+		return &types.NetworkMapComponents{
+			Peers:  map[string]*types.ComponentPeer{targetPeerID: peer.ToComponent()},
+			Groups: map[string]*types.ComponentGroup{targetGroupID: group},
+			Policies: []*types.Policy{{
+				ID:      "p",
+				Enabled: true,
+				Rules:   []*types.PolicyRule{rule},
+			}},
+		}, peer
+	}
+
+	cases := []struct {
+		name        string
+		peerSSH     bool
+		rule        types.PolicyRule
+		wantEnabled bool
+	}{
+		{
+			name:    "explicit-netbird-ssh-activates-regardless-of-peer-ssh",
+			peerSSH: false,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolNetbirdSSH,
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: true,
+		},
+		{
+			name:    "implicit-tcp-22-with-peer-ssh",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolTCP, Ports: []string{"22"},
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: true,
+		},
+		{
+			name:    "implicit-tcp-22-without-peer-ssh-disabled",
+			peerSSH: false,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolTCP, Ports: []string{"22"},
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: false,
+		},
+		{
+			name:    "implicit-tcp-22022-with-peer-ssh",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolTCP, Ports: []string{"22022"},
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: true,
+		},
+		{
+			name:    "implicit-all-protocol-with-peer-ssh",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolALL,
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: true,
+		},
+		{
+			name:    "implicit-port-range-covers-22",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled:      true,
+				Protocol:     types.PolicyRuleProtocolTCP,
+				PortRanges:   []types.RulePortRange{{Start: 20, End: 30}},
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: true,
+		},
+		{
+			name:    "tcp-80-no-ssh",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolTCP, Ports: []string{"80"},
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: false,
+		},
+		{
+			name:    "disabled-rule-skipped",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled: false, Protocol: types.PolicyRuleProtocolNetbirdSSH,
+				Destinations: []string{targetGroupID},
+			},
+			wantEnabled: false,
+		},
+		{
+			name:    "peer-not-in-destinations",
+			peerSSH: true,
+			rule: types.PolicyRule{
+				Enabled: true, Protocol: types.PolicyRuleProtocolNetbirdSSH,
+				Destinations: []string{"g_other"}, // target not in this group
+			},
+			wantEnabled: false,
+		},
+		{
+			name:    "peer-typed-destination-resource-matches",
+			peerSSH: false,
+			rule: types.PolicyRule{
+				Enabled:             true,
+				Protocol:            types.PolicyRuleProtocolNetbirdSSH,
+				DestinationResource: types.Resource{ID: targetPeerID, Type: types.ResourceTypePeer},
+			},
+			wantEnabled: true,
+		},
+		{
+			name:    "non-peer-destination-resource-falls-through-to-groups",
+			peerSSH: false,
+			rule: types.PolicyRule{
+				Enabled:             true,
+				Protocol:            types.PolicyRuleProtocolNetbirdSSH,
+				DestinationResource: types.Resource{ID: targetPeerID, Type: "host"}, // wrong type
+				Destinations:        []string{targetGroupID},                        // saved by group fallback
+			},
+			wantEnabled: true,
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			c, peer := mkComponents(&tc.rule, tc.peerSSH)
+			got := computeSSHEnabledForPeer(c, peer)
+			assert.Equal(t, tc.wantEnabled, got)
+		})
+	}
+}
+
+// TestComputeSSHEnabledForPeer_TargetMissingFromComponents covers the
+// belt-and-suspenders presence guard mirroring Calculate's
+// getAllPeersFromGroups invariant.
+func TestComputeSSHEnabledForPeer_TargetMissingFromComponents(t *testing.T) {
+	peer := &nbpeer.Peer{ID: "missing", SSHEnabled: true}
+	c := &types.NetworkMapComponents{
+		Peers: map[string]*types.ComponentPeer{}, // target peer NOT present
+		Groups: map[string]*types.ComponentGroup{
+			"g": {ID: "g", Peers: []string{"missing"}},
+		},
+		Policies: []*types.Policy{{
+			ID: "p", Enabled: true,
+			Rules: []*types.PolicyRule{{
+				Enabled: true, Protocol: types.PolicyRuleProtocolNetbirdSSH,
+				Destinations: []string{"g"},
+			}},
+		}},
+	}
+	assert.False(t, computeSSHEnabledForPeer(c, peer),
+		"missing target peer must short-circuit to false, not consult policies")
+}
+
+// TestComputeSSHEnabledForPeer_NilInputs guards the cheap nil-checks at
+// function entry — Calculate doesn't accept nil either, but the helper is
+// exported indirectly via ToComponentSyncResponse and may receive nil
+// components on graceful-degrade paths.
+func TestComputeSSHEnabledForPeer_NilInputs(t *testing.T) {
+	assert.False(t, computeSSHEnabledForPeer(nil, &nbpeer.Peer{ID: "x"}))
+	assert.False(t, computeSSHEnabledForPeer(&types.NetworkMapComponents{}, nil))
+}

management/internals/shared/grpc/conversion.go

@@ -10,24 +10,20 @@ import (
 
 	"github.com/hashicorp/go-version"
 	nbversion "github.com/netbirdio/netbird/version"
-	log "github.com/sirupsen/logrus"
-	goproto "google.golang.org/protobuf/proto"
 	"google.golang.org/protobuf/types/known/timestamppb"
 
 	integrationsConfig "github.com/netbirdio/management-integrations/integrations/config"
 
 	"github.com/netbirdio/netbird/client/ssh/auth"
 
-	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map/controller/cache"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
 	"github.com/netbirdio/netbird/management/server/posture"
 	"github.com/netbirdio/netbird/management/server/types"
-	nbroute "github.com/netbirdio/netbird/route"
+	"github.com/netbirdio/netbird/shared/management/networkmap"
 	"github.com/netbirdio/netbird/shared/management/proto"
 	"github.com/netbirdio/netbird/shared/netiputil"
-	"github.com/netbirdio/netbird/shared/sshauth"
 )
 
 const (
@@ -169,8 +165,8 @@ func ToSyncResponse(ctx context.Context, config *nbconfig.Config, httpConfig *nb
 		PeerConfig: toPeerConfig(peer, networkMap.Network, dnsName, settings, httpConfig, deviceFlowConfig, networkMap.EnableSSH),
 		NetworkMap: &proto.NetworkMap{
 			Serial:     networkMap.Network.CurrentSerial(),
-			Routes:     toProtocolRoutes(networkMap.Routes),
-			DNSConfig:  toProtocolDNSConfig(networkMap.DNSConfig, dnsCache, dnsFwdPort),
+			Routes:     networkmap.ToProtocolRoutes(networkMap.Routes),
+			DNSConfig:  networkmap.ToProtocolDNSConfig(networkMap.DNSConfig, dnsCache, dnsFwdPort),
 			PeerConfig: toPeerConfig(peer, networkMap.Network, dnsName, settings, httpConfig, deviceFlowConfig, networkMap.EnableSSH),
 		},
 		Checks: toProtocolChecks(ctx, checks),
@@ -183,7 +179,7 @@ func ToSyncResponse(ctx context.Context, config *nbconfig.Config, httpConfig *nb
 	response.NetworkMap.PeerConfig = response.PeerConfig
 
 	remotePeers := make([]*proto.RemotePeerConfig, 0, len(networkMap.Peers)+len(networkMap.OfflinePeers))
-	remotePeers = appendRemotePeerConfig(remotePeers, networkMap.Peers, dnsName, includeIPv6)
+	remotePeers = networkmap.AppendRemotePeerConfig(remotePeers, networkMap.Peers, dnsName, includeIPv6)
 
 	if !shouldSkipSendingDeprecatedRemotePeers(peer.Meta.WtVersion) {
 		response.RemotePeers = remotePeers
@@ -193,13 +189,13 @@ func ToSyncResponse(ctx context.Context, config *nbconfig.Config, httpConfig *nb
 	response.RemotePeersIsEmpty = len(remotePeers) == 0
 	response.NetworkMap.RemotePeersIsEmpty = response.RemotePeersIsEmpty
 
-	response.NetworkMap.OfflinePeers = appendRemotePeerConfig(nil, networkMap.OfflinePeers, dnsName, includeIPv6)
+	response.NetworkMap.OfflinePeers = networkmap.AppendRemotePeerConfig(nil, networkMap.OfflinePeers, dnsName, includeIPv6)
 
-	firewallRules := toProtocolFirewallRules(networkMap.FirewallRules, includeIPv6, useSourcePrefixes)
+	firewallRules := networkmap.ToProtocolFirewallRules(networkMap.FirewallRules, includeIPv6, useSourcePrefixes)
 	response.NetworkMap.FirewallRules = firewallRules
 	response.NetworkMap.FirewallRulesIsEmpty = len(firewallRules) == 0
 
-	routesFirewallRules := toProtocolRoutesFirewallRules(networkMap.RoutesFirewallRules)
+	routesFirewallRules := networkmap.ToProtocolRoutesFirewallRules(networkMap.RoutesFirewallRules)
 	response.NetworkMap.RoutesFirewallRules = routesFirewallRules
 	response.NetworkMap.RoutesFirewallRulesIsEmpty = len(routesFirewallRules) == 0
 
@@ -212,7 +208,7 @@ func ToSyncResponse(ctx context.Context, config *nbconfig.Config, httpConfig *nb
 	}
 
 	if networkMap.AuthorizedUsers != nil {
-		hashedUsers, machineUsers := buildAuthorizedUsersProto(ctx, networkMap.AuthorizedUsers)
+		hashedUsers, machineUsers := networkmap.BuildAuthorizedUsersProto(ctx, networkMap.AuthorizedUsers)
 		userIDClaim := auth.DefaultUserIDClaim
 		if httpConfig != nil && httpConfig.AuthUserIDClaim != "" {
 			userIDClaim = httpConfig.AuthUserIDClaim
@@ -252,33 +248,6 @@ func encodeSessionExpiresAt(deadline time.Time) *timestamppb.Timestamp {
 	return timestamppb.New(deadline)
 }
 
-func buildAuthorizedUsersProto(ctx context.Context, authorizedUsers map[string]map[string]struct{}) ([][]byte, map[string]*proto.MachineUserIndexes) {
-	userIDToIndex := make(map[string]uint32)
-	var hashedUsers [][]byte
-	machineUsers := make(map[string]*proto.MachineUserIndexes, len(authorizedUsers))
-
-	for machineUser, users := range authorizedUsers {
-		indexes := make([]uint32, 0, len(users))
-		for userID := range users {
-			idx, exists := userIDToIndex[userID]
-			if !exists {
-				hash, err := sshauth.HashUserID(userID)
-				if err != nil {
-					log.WithContext(ctx).Errorf("failed to hash user id %s: %v", userID, err)
-					continue
-				}
-				idx = uint32(len(hashedUsers))
-				userIDToIndex[userID] = idx
-				hashedUsers = append(hashedUsers, hash[:])
-			}
-			indexes = append(indexes, idx)
-		}
-		machineUsers[machineUser] = &proto.MachineUserIndexes{Indexes: indexes}
-	}
-
-	return hashedUsers, machineUsers
-}
-
 func shouldSkipSendingDeprecatedRemotePeers(peerVersion string) bool {
 	if nbversion.IsDevelopmentVersion(peerVersion) {
 		return true
@@ -292,51 +261,6 @@ func shouldSkipSendingDeprecatedRemotePeers(peerVersion string) bool {
 	return precomputedDeprecatedRemotePeersConstraint.Check(peerNBVersion)
 }
 
-func appendRemotePeerConfig(dst []*proto.RemotePeerConfig, peers []*nbpeer.Peer, dnsName string, includeIPv6 bool) []*proto.RemotePeerConfig {
-	for _, rPeer := range peers {
-		allowedIPs := []string{rPeer.IP.String() + "/32"}
-		if includeIPv6 && rPeer.IPv6.IsValid() {
-			allowedIPs = append(allowedIPs, rPeer.IPv6.String()+"/128")
-		}
-		dst = append(dst, &proto.RemotePeerConfig{
-			WgPubKey:     rPeer.Key,
-			AllowedIps:   allowedIPs,
-			SshConfig:    &proto.SSHConfig{SshPubKey: []byte(rPeer.SSHKey)},
-			Fqdn:         rPeer.FQDN(dnsName),
-			AgentVersion: rPeer.Meta.WtVersion,
-		})
-	}
-	return dst
-}
-
-// toProtocolDNSConfig converts nbdns.Config to proto.DNSConfig using the cache
-func toProtocolDNSConfig(update nbdns.Config, cache *cache.DNSConfigCache, forwardPort int64) *proto.DNSConfig {
-	protoUpdate := &proto.DNSConfig{
-		ServiceEnable:    update.ServiceEnable,
-		CustomZones:      make([]*proto.CustomZone, 0, len(update.CustomZones)),
-		NameServerGroups: make([]*proto.NameServerGroup, 0, len(update.NameServerGroups)),
-		ForwarderPort:    forwardPort,
-	}
-
-	for _, zone := range update.CustomZones {
-		protoZone := convertToProtoCustomZone(zone)
-		protoUpdate.CustomZones = append(protoUpdate.CustomZones, protoZone)
-	}
-
-	for _, nsGroup := range update.NameServerGroups {
-		cacheKey := nsGroup.ID
-		if cachedGroup, exists := cache.GetNameServerGroup(cacheKey); exists {
-			protoUpdate.NameServerGroups = append(protoUpdate.NameServerGroups, cachedGroup)
-		} else {
-			protoGroup := convertToProtoNameServerGroup(nsGroup)
-			cache.SetNameServerGroup(cacheKey, protoGroup)
-			protoUpdate.NameServerGroups = append(protoUpdate.NameServerGroups, protoGroup)
-		}
-	}
-
-	return protoUpdate
-}
-
 func ToResponseProto(configProto nbconfig.Protocol) proto.HostConfig_Protocol {
 	switch configProto {
 	case nbconfig.UDP:
@@ -354,203 +278,6 @@ func ToResponseProto(configProto nbconfig.Protocol) proto.HostConfig_Protocol {
 	}
 }
 
-func toProtocolRoutes(routes []*nbroute.Route) []*proto.Route {
-	protoRoutes := make([]*proto.Route, 0, len(routes))
-	for _, r := range routes {
-		protoRoutes = append(protoRoutes, toProtocolRoute(r))
-	}
-	return protoRoutes
-}
-
-func toProtocolRoute(route *nbroute.Route) *proto.Route {
-	return &proto.Route{
-		ID:            string(route.ID),
-		NetID:         string(route.NetID),
-		Network:       route.Network.String(),
-		Domains:       route.Domains.ToPunycodeList(),
-		NetworkType:   int64(route.NetworkType),
-		Peer:          route.Peer,
-		Metric:        int64(route.Metric),
-		Masquerade:    route.Masquerade,
-		KeepRoute:     route.KeepRoute,
-		SkipAutoApply: route.SkipAutoApply,
-	}
-}
-
-// toProtocolFirewallRules converts the firewall rules to the protocol firewall rules.
-// When useSourcePrefixes is true, the compact SourcePrefixes field is populated
-// alongside the deprecated PeerIP for forward compatibility.
-// Wildcard rules ("0.0.0.0") are expanded into separate v4 and v6 SourcePrefixes
-// when includeIPv6 is true.
-func toProtocolFirewallRules(rules []*types.FirewallRule, includeIPv6, useSourcePrefixes bool) []*proto.FirewallRule {
-	result := make([]*proto.FirewallRule, 0, len(rules))
-	for i := range rules {
-		rule := rules[i]
-
-		fwRule := &proto.FirewallRule{
-			PolicyID:  []byte(rule.PolicyID),
-			PeerIP:    rule.PeerIP, //nolint:staticcheck // populated for backward compatibility
-			Direction: getProtoDirection(rule.Direction),
-			Action:    getProtoAction(rule.Action),
-			Protocol:  getProtoProtocol(rule.Protocol),
-			Port:      rule.Port,
-		}
-
-		if useSourcePrefixes && rule.PeerIP != "" {
-			result = append(result, populateSourcePrefixes(fwRule, rule, includeIPv6)...)
-		}
-
-		if shouldUsePortRange(fwRule) {
-			fwRule.PortInfo = rule.PortRange.ToProto()
-		}
-
-		result = append(result, fwRule)
-	}
-	return result
-}
-
-// populateSourcePrefixes sets SourcePrefixes on fwRule and returns any
-// additional rules needed (e.g. a v6 wildcard clone when the peer IP is unspecified).
-func populateSourcePrefixes(fwRule *proto.FirewallRule, rule *types.FirewallRule, includeIPv6 bool) []*proto.FirewallRule {
-	addr, err := netip.ParseAddr(rule.PeerIP)
-	if err != nil {
-		return nil
-	}
-
-	if !addr.IsUnspecified() {
-		fwRule.SourcePrefixes = [][]byte{netiputil.EncodeAddr(addr.Unmap())}
-		return nil
-	}
-
-	// IPv4Unspecified/0 is always valid, error is impossible.
-	v4Wildcard, _ := netiputil.EncodePrefix(netip.PrefixFrom(netip.IPv4Unspecified(), 0))
-	fwRule.SourcePrefixes = [][]byte{v4Wildcard}
-
-	if !includeIPv6 {
-		return nil
-	}
-
-	v6Rule := goproto.Clone(fwRule).(*proto.FirewallRule)
-	v6Rule.PeerIP = "::" //nolint:staticcheck // populated for backward compatibility
-	// IPv6Unspecified/0 is always valid, error is impossible.
-	v6Wildcard, _ := netiputil.EncodePrefix(netip.PrefixFrom(netip.IPv6Unspecified(), 0))
-	v6Rule.SourcePrefixes = [][]byte{v6Wildcard}
-	if shouldUsePortRange(v6Rule) {
-		v6Rule.PortInfo = rule.PortRange.ToProto()
-	}
-	return []*proto.FirewallRule{v6Rule}
-}
-
-// getProtoDirection converts the direction to proto.RuleDirection.
-func getProtoDirection(direction int) proto.RuleDirection {
-	if direction == types.FirewallRuleDirectionOUT {
-		return proto.RuleDirection_OUT
-	}
-	return proto.RuleDirection_IN
-}
-
-func toProtocolRoutesFirewallRules(rules []*types.RouteFirewallRule) []*proto.RouteFirewallRule {
-	result := make([]*proto.RouteFirewallRule, len(rules))
-	for i := range rules {
-		rule := rules[i]
-		result[i] = &proto.RouteFirewallRule{
-			SourceRanges: rule.SourceRanges,
-			Action:       getProtoAction(rule.Action),
-			Destination:  rule.Destination,
-			Protocol:     getProtoProtocol(rule.Protocol),
-			PortInfo:     getProtoPortInfo(rule),
-			IsDynamic:    rule.IsDynamic,
-			Domains:      rule.Domains.ToPunycodeList(),
-			PolicyID:     []byte(rule.PolicyID),
-			RouteID:      string(rule.RouteID),
-		}
-	}
-
-	return result
-}
-
-// getProtoAction converts the action to proto.RuleAction.
-func getProtoAction(action string) proto.RuleAction {
-	if action == string(types.PolicyTrafficActionDrop) {
-		return proto.RuleAction_DROP
-	}
-	return proto.RuleAction_ACCEPT
-}
-
-// getProtoProtocol converts the protocol to proto.RuleProtocol.
-func getProtoProtocol(protocol string) proto.RuleProtocol {
-	switch types.PolicyRuleProtocolType(protocol) {
-	case types.PolicyRuleProtocolALL:
-		return proto.RuleProtocol_ALL
-	case types.PolicyRuleProtocolTCP:
-		return proto.RuleProtocol_TCP
-	case types.PolicyRuleProtocolUDP:
-		return proto.RuleProtocol_UDP
-	case types.PolicyRuleProtocolICMP:
-		return proto.RuleProtocol_ICMP
-	default:
-		return proto.RuleProtocol_UNKNOWN
-	}
-}
-
-// getProtoPortInfo converts the port info to proto.PortInfo.
-func getProtoPortInfo(rule *types.RouteFirewallRule) *proto.PortInfo {
-	var portInfo proto.PortInfo
-	if rule.Port != 0 {
-		portInfo.PortSelection = &proto.PortInfo_Port{Port: uint32(rule.Port)}
-	} else if portRange := rule.PortRange; portRange.Start != 0 && portRange.End != 0 {
-		portInfo.PortSelection = &proto.PortInfo_Range_{
-			Range: &proto.PortInfo_Range{
-				Start: uint32(portRange.Start),
-				End:   uint32(portRange.End),
-			},
-		}
-	}
-	return &portInfo
-}
-
-func shouldUsePortRange(rule *proto.FirewallRule) bool {
-	return rule.Port == "" && (rule.Protocol == proto.RuleProtocol_UDP || rule.Protocol == proto.RuleProtocol_TCP)
-}
-
-// Helper function to convert nbdns.CustomZone to proto.CustomZone
-func convertToProtoCustomZone(zone nbdns.CustomZone) *proto.CustomZone {
-	protoZone := &proto.CustomZone{
-		Domain:               zone.Domain,
-		Records:              make([]*proto.SimpleRecord, 0, len(zone.Records)),
-		SearchDomainDisabled: zone.SearchDomainDisabled,
-		NonAuthoritative:     zone.NonAuthoritative,
-	}
-	for _, record := range zone.Records {
-		protoZone.Records = append(protoZone.Records, &proto.SimpleRecord{
-			Name:  record.Name,
-			Type:  int64(record.Type),
-			Class: record.Class,
-			TTL:   int64(record.TTL),
-			RData: record.RData,
-		})
-	}
-	return protoZone
-}
-
-// Helper function to convert nbdns.NameServerGroup to proto.NameServerGroup
-func convertToProtoNameServerGroup(nsGroup *nbdns.NameServerGroup) *proto.NameServerGroup {
-	protoGroup := &proto.NameServerGroup{
-		Primary:              nsGroup.Primary,
-		Domains:              nsGroup.Domains,
-		SearchDomainsEnabled: nsGroup.SearchDomainsEnabled,
-		NameServers:          make([]*proto.NameServer, 0, len(nsGroup.NameServers)),
-	}
-	for _, ns := range nsGroup.NameServers {
-		protoGroup.NameServers = append(protoGroup.NameServers, &proto.NameServer{
-			IP:     ns.IP.String(),
-			Port:   int64(ns.Port),
-			NSType: int64(ns.NSType),
-		})
-	}
-	return protoGroup
-}
-
 // buildJWTConfig constructs JWT configuration for SSH servers from management server config
 func buildJWTConfig(config *nbconfig.HttpServerConfig, deviceFlowConfig *nbconfig.DeviceAuthorizationFlow) *proto.JWTConfig {
 	if config == nil || config.AuthAudience == "" {

management/internals/shared/grpc/conversion_test.go

@@ -15,6 +15,7 @@ import (
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map/controller/cache"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
 	"github.com/netbirdio/netbird/management/server/types"
+	"github.com/netbirdio/netbird/shared/management/networkmap"
 )
 
 func TestToProtocolDNSConfigWithCache(t *testing.T) {
@@ -64,13 +65,13 @@ func TestToProtocolDNSConfigWithCache(t *testing.T) {
 	}
 
 	// First run with config1
-	result1 := toProtocolDNSConfig(config1, &cache, int64(network_map.DnsForwarderPort))
+	result1 := networkmap.ToProtocolDNSConfig(config1, &cache, int64(network_map.DnsForwarderPort))
 
 	// Second run with config2
-	result2 := toProtocolDNSConfig(config2, &cache, int64(network_map.DnsForwarderPort))
+	result2 := networkmap.ToProtocolDNSConfig(config2, &cache, int64(network_map.DnsForwarderPort))
 
 	// Third run with config1 again
-	result3 := toProtocolDNSConfig(config1, &cache, int64(network_map.DnsForwarderPort))
+	result3 := networkmap.ToProtocolDNSConfig(config1, &cache, int64(network_map.DnsForwarderPort))
 
 	// Verify that result1 and result3 are identical
 	if !reflect.DeepEqual(result1, result3) {
@@ -102,15 +103,14 @@ func BenchmarkToProtocolDNSConfig(b *testing.B) {
 
 			b.ResetTimer()
 			for i := 0; i < b.N; i++ {
-				toProtocolDNSConfig(testData, cache, int64(network_map.DnsForwarderPort))
+				networkmap.ToProtocolDNSConfig(testData, cache, int64(network_map.DnsForwarderPort))
 			}
 		})
 
 		b.Run(fmt.Sprintf("WithoutCache-Size%d", size), func(b *testing.B) {
 			b.ResetTimer()
 			for i := 0; i < b.N; i++ {
-				cache := &cache.DNSConfigCache{}
-				toProtocolDNSConfig(testData, cache, int64(network_map.DnsForwarderPort))
+				networkmap.ToProtocolDNSConfig(testData, nil, int64(network_map.DnsForwarderPort))
 			}
 		})
 	}

management/internals/shared/grpc/proxy.go

@@ -608,11 +608,11 @@ func (s *ProxyServiceServer) disconnectProxy(conn *proxyConnection) {
 	if err := s.proxyController.UnregisterProxyFromCluster(context.Background(), conn.address, conn.proxyID); err != nil {
 		log.Warnf("Failed to unregister proxy %s from cluster: %v", conn.proxyID, err)
 	}
+	conn.cancel()
 	if err := s.proxyManager.Disconnect(context.Background(), conn.proxyID, conn.sessionID); err != nil {
 		log.Warnf("Failed to mark proxy %s as disconnected: %v", conn.proxyID, err)
 	}
 
-	conn.cancel()
 	log.Infof("Proxy %s session %s disconnected", conn.proxyID, conn.sessionID)
 }
 

management/internals/shared/grpc/server.go

@@ -25,6 +25,7 @@ import (
 	"google.golang.org/grpc/status"
 
 	"github.com/netbirdio/netbird/shared/management/client/common"
+	"github.com/netbirdio/netbird/shared/management/grpc"
 
 	"github.com/netbirdio/netbird/management/internals/controllers/network_map"
 	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
@@ -245,6 +246,7 @@ func (s *Server) Sync(req *proto.EncryptedMessage, srv proto.ManagementService_S
 	realIP := getRealIP(ctx)
 	sRealIP := realIP.String()
 	peerMeta := extractPeerMeta(ctx, syncReq.GetMeta())
+
 	userID, err := s.accountManager.GetUserIDByPeerKey(ctx, peerKey.String())
 	if err != nil {
 		s.syncSem.Add(-1)
@@ -683,8 +685,9 @@ func extractPeerMeta(ctx context.Context, meta *proto.PeerSystemMeta) nbpeer.Pee
 			LazyConnectionEnabled: meta.GetFlags().GetLazyConnectionEnabled(),
 			DisableIPv6:           meta.GetFlags().GetDisableIPv6(),
 		},
-		Files:        files,
-		Capabilities: capabilitiesToInt32(meta.GetCapabilities()),
+		Files:              files,
+		Capabilities:       capabilitiesToInt32(meta.GetCapabilities()),
+		SyncMessageVersion: int(meta.GetSyncMessageVersion()),
 	}
 }
 
@@ -1016,7 +1019,43 @@ func (s *Server) sendInitialSync(ctx context.Context, peerKey wgtypes.Key, peer
 		return status.Errorf(codes.Internal, "failed to get peer groups %s", err)
 	}
 
-	plainResp := ToSyncResponse(ctx, s.config, s.config.HttpConfig, s.config.DeviceAuthorizationFlow, peer, turnToken, relayToken, networkMap, s.networkMapController.GetDNSDomain(settings), postureChecks, nil, settings, settings.Extra, peerGroups, dnsFwdPort)
+	dnsName := s.networkMapController.GetDNSDomain(settings)
+
+	var plainResp *proto.SyncResponse
+
+	commonSyncMessageVersion := grpc.HighestCommonSyncMessageVersion(
+		s.perAccountOrGlobalSyncMessageVersions(peer.AccountID),
+		grpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion))
+
+	log.WithContext(ctx).
+		WithFields(log.Fields{
+			"sync_message_version":        commonSyncMessageVersion,
+			"server_sync_message_version": s.perAccountOrGlobalSyncMessageVersions(peer.AccountID),
+			"peer_sync_message_version":   grpc.SyncMessageVersionFromConfig(&peer.Meta.SyncMessageVersion),
+		}).Debug("common highest sync message version")
+
+	if commonSyncMessageVersion == grpc.ComponentNetworkMap {
+		// Capable peer: discard the legacy NetworkMap that SyncAndMarkPeer
+		// computed and recompute the raw components instead. This wastes one
+		// Calculate() call per initial-sync — the component-based wire
+		// format is what the peer actually consumes. The streaming path
+		// (network_map.Controller.UpdateAccountPeers) skips this duplication
+		// because it dispatches by capability before computing.
+		//
+		// TODO: refactor SyncPeer / SyncAndMarkPeer / their mocks + manager
+		// interfaces to return PeerNetworkMapResult so the initial-sync path
+		// stops doing duplicate work. Deferred until the client-side
+		// decoder lands and there's a real deployment of capability=3 peers
+		// worth optimizing for.
+		freshPeer, components, proxyPatch, freshPostureChecks, freshDnsFwdPort, err := s.networkMapController.GetValidatedPeerWithComponents(ctx, false, peer.AccountID, peer)
+		if err != nil {
+			log.WithContext(ctx).Errorf("failed to build components for peer %s on initial sync: %v", peer.ID, err)
+			return status.Errorf(codes.Internal, "failed to build initial sync envelope")
+		}
+		plainResp = ToComponentSyncResponse(ctx, s.config, s.config.HttpConfig, s.config.DeviceAuthorizationFlow, freshPeer, turnToken, relayToken, components, proxyPatch, dnsName, freshPostureChecks, settings, settings.Extra, peerGroups, freshDnsFwdPort)
+	} else {
+		plainResp = ToSyncResponse(ctx, s.config, s.config.HttpConfig, s.config.DeviceAuthorizationFlow, peer, turnToken, relayToken, networkMap, dnsName, postureChecks, nil, settings, settings.Extra, peerGroups, dnsFwdPort)
+	}
 
 	key, err := s.secretsManager.GetWGKey()
 	if err != nil {
@@ -1041,6 +1080,13 @@ func (s *Server) sendInitialSync(ctx context.Context, peerKey wgtypes.Key, peer
 	return nil
 }
 
+func (s *Server) perAccountOrGlobalSyncMessageVersions(accountId string) grpc.SyncMessageVersion {
+	if version, ok := s.config.PerAccountHighestSupportedSyncMessageVersion[accountId]; ok {
+		return grpc.SyncMessageVersionFromConfig(&version)
+	}
+	return grpc.SyncMessageVersionFromConfig(s.config.HighestSupportedSyncMessageVersion)
+}
+
 // GetDeviceAuthorizationFlow returns a device authorization flow information
 // This is used for initiating an Oauth 2 device authorization grant flow
 // which will be used by our clients to Login