Rework signal protocol to minimize number of exchanged initial RP 512KB keys

The Rosenpass static keypair was regenerated on every engine start
(rp.GenerateKeyPair in NewManager), so the local ~512KB public key —
and its fingerprint — changed on each client restart.

Persist the keypair to <StateDir>/rosenpass_key.json with 0600
permissions (same protection tier as the WireGuard private key), and
reload it on start so the public key stays stable across restarts. A
missing, corrupt, or version-incompatible file degrades gracefully to
generating a fresh ephemeral keypair (previous behaviour); an empty
StateDir keeps the ephemeral path for callers without a state dir.

This is the foundation for fingerprint-based RP pubkey caching over
signalling (NET-1407): a stable local key lets remote peers keep their
cached copy valid across our restart.

client/cmd/login.go

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

client/cmd/up.go

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

client/internal/engine.go

@@ -551,7 +551,7 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 		} else {
 			log.Infof("running rosenpass in strict mode")
 		}
-		e.rpManager, err = rosenpass.NewManager(e.config.PreSharedKey, e.config.WgIfaceName, publicKey)
+		e.rpManager, err = rosenpass.NewManager(e.config.PreSharedKey, e.config.WgIfaceName, publicKey, e.config.StateDir)
 		if err != nil {
 			return fmt.Errorf("create rosenpass manager: %w", err)
 		}
@@ -1809,6 +1809,7 @@ func (e *Engine) createPeerConn(pubKey string, allowedIPs []netip.Prefix, agentV
 			PubKey:         e.getRosenpassPubKey(),
 			Addr:           e.getRosenpassAddr(),
 			PermissiveMode: e.config.RosenpassPermissive,
+			KeyResolver:    e.rosenpassKeyResolver(),
 		},
 		ICEConfig: e.createICEConfig(),
 	}
@@ -1879,6 +1880,8 @@ func (e *Engine) receiveSignalEvents() error {
 
 				log.Debugf("receiveMSG: took %s to get lock for peer %s with session id %s", gotLock, msg.Key, offerAnswer.SessionID)
 
+				e.applyRosenpassKeyExchange(msg, offerAnswer)
+
 				if msg.Body.Type == sProto.Body_OFFER {
 					conn.OnRemoteOffer(*offerAnswer)
 				} else {
@@ -2222,6 +2225,34 @@ func (e *Engine) getRosenpassAddr() string {
 	return ""
 }
 
+// rosenpassKeyResolver returns the Rosenpass manager as the offer/answer key
+// resolver, or a true nil interface when Rosenpass is disabled (returning the
+// typed-nil *Manager would make the interface non-nil and panic on use).
+func (e *Engine) rosenpassKeyResolver() peer.RosenpassKeyResolver {
+	if e.rpManager == nil {
+		return nil
+	}
+	return e.rpManager
+}
+
+// applyRosenpassKeyExchange reconciles the fingerprint/cache fields of an incoming
+// offer/answer against the Rosenpass manager's cache: it resolves the remote peer's
+// full public key (from the message or the cache) into the OfferAnswer, and records
+// whether the peer acknowledged holding our key. No-op when Rosenpass is disabled.
+func (e *Engine) applyRosenpassKeyExchange(msg *sProto.Message, oa *peer.OfferAnswer) {
+	if e.rpManager == nil {
+		return
+	}
+	cfg := msg.GetBody().GetRosenpassConfig()
+	if cfg == nil {
+		return
+	}
+
+	remoteWgKey := msg.GetKey()
+	oa.RosenpassPubKey = e.rpManager.ResolveRemotePubKey(remoteWgKey, cfg.GetRosenpassPubKey(), cfg.GetRosenpassPubKeyHash())
+	e.rpManager.SetRemoteAck(remoteWgKey, cfg.GetAcknowledgedRosenpassPubKeyHash())
+}
+
 // RunHealthProbes executes health checks for Signal, Management, Relay, and WireGuard services
 // and updates the status recorder with the latest states.
 //
@@ -2605,13 +2636,14 @@ func (e *Engine) updateForwardRules(rules []*mgmProto.ForwardingRule) ([]firewal
 
 func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, peers []*mgmProto.RemotePeerConfig) map[string]bool {
 	excludedPeers := make(map[string]bool)
+
+	// Ingress forward targets: inbound forwarded traffic is initiated remotely and
+	// cannot wake a lazy connection, so the peer routing the target must stay
+	// permanently connected. AllowedIPs are already parsed on the peer conn, so
+	// reuse those typed prefixes instead of re-parsing the network map strings.
 	for _, r := range rules {
-		ip := r.TranslatedAddress
 		for _, p := range peers {
-			for _, allowedIP := range p.GetAllowedIps() {
-				if allowedIP != ip.String() {
-					continue
-				}
+			if e.peerRoutesAddr(p, r.TranslatedAddress) {
 				log.Infof("exclude forwarder peer from lazy connection: %s", p.GetWgPubKey())
 				excludedPeers[p.GetWgPubKey()] = true
 			}
@@ -2621,6 +2653,27 @@ func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, peers
 	return excludedPeers
 }
 
+// peerRoutesAddr reports whether the peer is a router for addr, matched against
+// the peer's already-parsed AllowedIPs from the store (the same typed value the
+// lazy manager consumes) rather than re-parsing the network map strings.
+func (e *Engine) peerRoutesAddr(p *mgmProto.RemotePeerConfig, addr netip.Addr) bool {
+	prefixes, ok := e.peerStore.AllowedIPs(p.GetWgPubKey())
+	if !ok {
+		return false
+	}
+	return prefixesContain(prefixes, addr)
+}
+
+// prefixesContain reports whether addr falls within any of the prefixes.
+func prefixesContain(prefixes []netip.Prefix, addr netip.Addr) bool {
+	for _, prefix := range prefixes {
+		if prefix.Contains(addr) {
+			return true
+		}
+	}
+	return false
+}
+
 // isChecksEqual checks if two slices of checks are equal.
 func isChecksEqual(checks1, checks2 []*mgmProto.Checks) bool {
 	normalize := func(checks []*mgmProto.Checks) []string {

client/internal/engine_lazy_exclude_test.go

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

client/internal/peer/conn.go

@@ -65,6 +65,20 @@ type WgConfig struct {
 	PreSharedKey *wgtypes.Key
 }
 
+// RosenpassKeyResolver lets the handshaker fill the fingerprint/cache fields of an
+// offer/answer without depending on the Rosenpass manager directly. Implemented by
+// rosenpass.Manager and wired in by the engine.
+type RosenpassKeyResolver interface {
+	// LocalPubKeyHash is the SHA256 of our own Rosenpass public key.
+	LocalPubKeyHash() []byte
+	// RemotePubKeyAck is the SHA256 of the remote peer's cached key (nil if we do
+	// not hold it), sent back as an acknowledgement.
+	RemotePubKeyAck(remoteWgKey string) []byte
+	// RemoteHasLocalKey reports whether the peer already holds our key, so the full
+	// key may be omitted.
+	RemoteHasLocalKey(remoteWgKey string) bool
+}
+
 type RosenpassConfig struct {
 	// RosenpassPubKey is this peer's Rosenpass public key
 	PubKey []byte
@@ -72,6 +86,10 @@ type RosenpassConfig struct {
 	Addr string
 
 	PermissiveMode bool
+
+	// KeyResolver drives fingerprint-based key caching over signalling. Nil when
+	// Rosenpass is disabled, which makes the handshaker always send the full key.
+	KeyResolver RosenpassKeyResolver
 }
 
 // ConnConfig is a peer Connection configuration
@@ -203,7 +221,6 @@ func NewConn(config ConnConfig, services ServiceDependencies) (*Conn, error) {
 		statusICE:          worker.NewAtomicStatus(),
 		dumpState:          dumpState,
 		endpointUpdater:    NewEndpointUpdater(connLog, config.WgConfig, isController(config)),
-		wgWatcher:          NewWGWatcher(connLog, config.WgConfig.WgInterface, config.Key, dumpState),
 		metricsRecorder:    services.MetricsRecorder,
 	}
 
@@ -671,11 +688,12 @@ func (conn *Conn) onGuardEvent() {
 	}
 }
 
-func (conn *Conn) onWGDisconnected() {
+func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
-	if conn.ctx.Err() != nil {
+	// watcherCtx guards against a stale watcher tearing down a connection that already superseded it.
+	if conn.ctx.Err() != nil || watcherCtx.Err() != nil {
 		return
 	}
 
@@ -833,25 +851,39 @@ func (conn *Conn) isConnectedOnAllWay() (status guard.ConnStatus) {
 	})
 }
 
+// enableWgWatcherIfNeeded starts a fresh watcher instance per connection attempt, so its
+// lifecycle stays bound to conn.mu and enable/disable can't race an old goroutine's shutdown.
+// Caller must hold conn.mu.
 func (conn *Conn) enableWgWatcherIfNeeded(enabledTime time.Time) {
-	if !conn.wgWatcher.PrepareInitialHandshake() {
+	if conn.wgWatcher != nil {
 		return
 	}
 
+	watcher := NewWGWatcher(conn.Log, conn.config.WgConfig.WgInterface, conn.config.Key, conn.dumpState)
+	watcher.PrepareInitialHandshake()
+
 	wgWatcherCtx, wgWatcherCancel := context.WithCancel(conn.ctx)
+	conn.wgWatcher = watcher
 	conn.wgWatcherCancel = wgWatcherCancel
+
 	conn.wgWatcherWg.Add(1)
 	go func() {
 		defer conn.wgWatcherWg.Done()
-		conn.wgWatcher.EnableWgWatcher(wgWatcherCtx, enabledTime, conn.onWGDisconnected, conn.onWGHandshakeSuccess, conn.onWGCheckSuccess)
+		onDisconnected := func() { conn.onWGDisconnected(wgWatcherCtx) }
+		watcher.EnableWgWatcher(wgWatcherCtx, enabledTime, onDisconnected, conn.onWGHandshakeSuccess, conn.onWGCheckSuccess)
 	}()
 }
 
+// disableWgWatcherIfNeeded cancels and drops the watcher once no transport is active. It never
+// waits for the goroutine: the timeout path reentrantly calls back here under conn.mu, so
+// blocking would deadlock. Caller must hold conn.mu.
 func (conn *Conn) disableWgWatcherIfNeeded() {
-	if conn.currentConnPriority == conntype.None && conn.wgWatcherCancel != nil {
-		conn.wgWatcherCancel()
-		conn.wgWatcherCancel = nil
+	if conn.currentConnPriority != conntype.None || conn.wgWatcher == nil {
+		return
 	}
+	conn.wgWatcherCancel()
+	conn.wgWatcher = nil
+	conn.wgWatcherCancel = nil
 }
 
 func (conn *Conn) newProxy(remoteConn net.Conn) (wgproxy.Proxy, error) {
@@ -874,7 +906,9 @@ func (conn *Conn) resetEndpoint() {
 		return
 	}
 	conn.Log.Infof("reset wg endpoint")
-	conn.wgWatcher.Reset()
+	if conn.wgWatcher != nil {
+		conn.wgWatcher.Reset()
+	}
 	if err := conn.endpointUpdater.RemoveEndpointAddress(); err != nil {
 		conn.Log.Warnf("failed to remove endpoint address before update: %v", err)
 	}

client/internal/peer/conn_test.go

@@ -339,20 +339,20 @@ func TestConn_onWGDisconnected_EscalatesToRosenpassReset(t *testing.T) {
 	conn := newWGTimeoutTestConn(true, &disconnected)
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Empty(t, disconnected, "escalation must not fire below the threshold")
 
-	conn.onWGDisconnected()
+	conn.onWGDisconnected(conn.ctx)
 	assert.Equal(t, []string{conn.config.WgConfig.RemoteKey}, disconnected,
 		"reaching the threshold must report the peer disconnected once")
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Len(t, disconnected, 1, "escalation must restart counting after firing")
 
-	conn.onWGDisconnected()
+	conn.onWGDisconnected(conn.ctx)
 	assert.Len(t, disconnected, 2, "continued timeouts must escalate again")
 }
 
@@ -364,12 +364,12 @@ func TestConn_onWGDisconnected_CheckSuccessResetsEscalation(t *testing.T) {
 	conn := newWGTimeoutTestConn(true, &disconnected)
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	conn.onWGCheckSuccess()
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Empty(t, disconnected, "handshake success must reset the timeout count")
 }
@@ -382,7 +382,7 @@ func TestConn_onWGDisconnected_NoEscalationWithoutRosenpass(t *testing.T) {
 	conn := newWGTimeoutTestConn(false, &disconnected)
 
 	for i := 0; i < wgTimeoutEscalationThreshold*3; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Empty(t, disconnected, "escalation must be limited to rosenpass connections")
 }

client/internal/peer/handshaker.go

@@ -33,8 +33,13 @@ type OfferAnswer struct {
 	// Version of NetBird Agent
 	Version string
 	// RosenpassPubKey is the Rosenpass public key of the remote peer when receiving this message
-	// This value is the local Rosenpass server public key when sending the message
+	// This value is the local Rosenpass server public key when sending the message.
+	// May be empty on send when the remote peer has acknowledged already holding it (see RosenpassPubKeyAck).
 	RosenpassPubKey []byte
+	// RosenpassPubKeyHash is the SHA256 of the sender's own RosenpassPubKey. Always set when Rosenpass is enabled.
+	RosenpassPubKeyHash []byte
+	// RosenpassPubKeyAck is the SHA256 of the remote peer's key the sender holds cached; empty means "send it in full".
+	RosenpassPubKeyAck []byte
 	// RosenpassAddr is the Rosenpass server address (IP:port) of the remote peer when receiving this message
 	// This value is the local Rosenpass server address when sending the message
 	RosenpassAddr string
@@ -209,11 +214,11 @@ func (h *Handshaker) sendAnswer() error {
 
 func (h *Handshaker) buildOfferAnswer() OfferAnswer {
 	answer := OfferAnswer{
-		WgListenPort:    h.config.LocalWgPort,
-		Version:         version.NetbirdVersion(),
-		RosenpassPubKey: h.config.RosenpassConfig.PubKey,
-		RosenpassAddr:   h.config.RosenpassConfig.Addr,
+		WgListenPort:  h.config.LocalWgPort,
+		Version:       version.NetbirdVersion(),
+		RosenpassAddr: h.config.RosenpassConfig.Addr,
 	}
+	h.setRosenpassPubKey(&answer)
 
 	if h.ice != nil && h.RemoteICESupported() {
 		uFrag, pwd := h.ice.GetLocalUserCredentials()
@@ -230,6 +235,30 @@ func (h *Handshaker) buildOfferAnswer() OfferAnswer {
 	return answer
 }
 
+// setRosenpassPubKey fills the Rosenpass key fields of an outgoing offer/answer.
+// With a resolver wired it advertises our key hash and the ack for the remote key
+// we hold, and includes the full public key only when the peer has not yet
+// acknowledged holding it. Without a resolver (Rosenpass disabled, or an older
+// code path) it always sends the full key, preserving the previous behaviour.
+func (h *Handshaker) setRosenpassPubKey(answer *OfferAnswer) {
+	localKey := h.config.RosenpassConfig.PubKey
+	if len(localKey) == 0 {
+		return
+	}
+
+	resolver := h.config.RosenpassConfig.KeyResolver
+	if resolver == nil {
+		answer.RosenpassPubKey = localKey
+		return
+	}
+
+	answer.RosenpassPubKeyHash = resolver.LocalPubKeyHash()
+	answer.RosenpassPubKeyAck = resolver.RemotePubKeyAck(h.config.Key)
+	if !resolver.RemoteHasLocalKey(h.config.Key) {
+		answer.RosenpassPubKey = localKey
+	}
+}
+
 func (h *Handshaker) updateRemoteICEState(offer *OfferAnswer) {
 	hasICE := offer.hasICECredentials()
 	prev := h.remoteICESupported.Swap(hasICE)

client/internal/peer/handshaker_rosenpass_test.go

@@ -0,0 +1,65 @@
+package peer
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+type fakeRPResolver struct {
+	localHash []byte
+	ack       []byte
+	hasLocal  bool
+}
+
+func (f fakeRPResolver) LocalPubKeyHash() []byte              { return f.localHash }
+func (f fakeRPResolver) RemotePubKeyAck(string) []byte        { return f.ack }
+func (f fakeRPResolver) RemoteHasLocalKey(remote string) bool { return f.hasLocal }
+
+func TestSetRosenpassPubKey_NoResolverAlwaysSendsFullKey(t *testing.T) {
+	localKey := []byte{1, 2, 3}
+	h := &Handshaker{config: ConnConfig{RosenpassConfig: RosenpassConfig{PubKey: localKey}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Equal(t, localKey, a.RosenpassPubKey)
+	require.Nil(t, a.RosenpassPubKeyHash)
+	require.Nil(t, a.RosenpassPubKeyAck)
+}
+
+func TestSetRosenpassPubKey_ResolverIncludesFullKeyUntilAcked(t *testing.T) {
+	localKey := []byte{1, 2, 3}
+	res := fakeRPResolver{localHash: []byte{9}, ack: []byte{8}, hasLocal: false}
+	h := &Handshaker{config: ConnConfig{Key: "peerA", RosenpassConfig: RosenpassConfig{PubKey: localKey, KeyResolver: res}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Equal(t, localKey, a.RosenpassPubKey, "full key must be sent until the peer acks it")
+	require.Equal(t, []byte{9}, a.RosenpassPubKeyHash)
+	require.Equal(t, []byte{8}, a.RosenpassPubKeyAck)
+}
+
+func TestSetRosenpassPubKey_ResolverOmitsFullKeyOnceAcked(t *testing.T) {
+	localKey := []byte{1, 2, 3}
+	res := fakeRPResolver{localHash: []byte{9}, ack: []byte{8}, hasLocal: true}
+	h := &Handshaker{config: ConnConfig{Key: "peerA", RosenpassConfig: RosenpassConfig{PubKey: localKey, KeyResolver: res}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Nil(t, a.RosenpassPubKey, "full key must be omitted once the peer holds it")
+	require.Equal(t, []byte{9}, a.RosenpassPubKeyHash)
+	require.Equal(t, []byte{8}, a.RosenpassPubKeyAck)
+}
+
+func TestSetRosenpassPubKey_DisabledSetsNothing(t *testing.T) {
+	h := &Handshaker{config: ConnConfig{RosenpassConfig: RosenpassConfig{}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Nil(t, a.RosenpassPubKey)
+	require.Nil(t, a.RosenpassPubKeyHash)
+}

client/internal/peer/signaler.go

@@ -61,11 +61,13 @@ func (s *Signaler) signalOfferAnswer(offerAnswer OfferAnswer, remoteKey string,
 			UFrag: offerAnswer.IceCredentials.UFrag,
 			Pwd:   offerAnswer.IceCredentials.Pwd,
 		},
-		RosenpassPubKey: offerAnswer.RosenpassPubKey,
-		RosenpassAddr:   offerAnswer.RosenpassAddr,
-		RelaySrvAddress: offerAnswer.RelaySrvAddress,
-		RelaySrvIP:      offerAnswer.RelaySrvIP,
-		SessionID:       sessionIDBytes,
+		RosenpassPubKey:     offerAnswer.RosenpassPubKey,
+		RosenpassPubKeyHash: offerAnswer.RosenpassPubKeyHash,
+		RosenpassPubKeyAck:  offerAnswer.RosenpassPubKeyAck,
+		RosenpassAddr:       offerAnswer.RosenpassAddr,
+		RelaySrvAddress:     offerAnswer.RelaySrvAddress,
+		RelaySrvIP:          offerAnswer.RelaySrvIP,
+		SessionID:           sessionIDBytes,
 	})
 	if err != nil {
 		return err

client/internal/peer/wg_watcher.go

@@ -3,7 +3,6 @@ package peer
 import (
 	"context"
 	"fmt"
-	"sync"
 	"time"
 
 	log "github.com/sirupsen/logrus"
@@ -24,14 +23,14 @@ type WGInterfaceStater interface {
 	GetStats() (map[string]configurer.WGStats, error)
 }
 
+// WGWatcher is single-shot: one instance per connection attempt, run once, then discarded.
+// Lifecycle is owned by Conn under conn.mu, so it keeps no "enabled" state to go stale.
 type WGWatcher struct {
 	log           *log.Entry
 	wgIfaceStater WGInterfaceStater
 	peerKey       string
 	stateDump     *stateDump
 
-	enabled   bool
-	muEnabled sync.Mutex
 	// initialHandshake is not thread-safe; never call PrepareInitialHandshake and EnableWgWatcher concurrently.
 	initialHandshake time.Time
 
@@ -48,25 +47,14 @@ func NewWGWatcher(log *log.Entry, wgIfaceStater WGInterfaceStater, peerKey strin
 	}
 }
 
-// PrepareInitialHandshake reserves the watcher and reads the peer's current WireGuard
-// handshake time. It must be called before the peer is (re)configured on the WireGuard
-// interface, so the captured baseline reflects the state prior to this connection attempt
-// instead of racing with that configuration. Returns ok=false if the watcher is already
-// running, in which case EnableWgWatcher must not be called.
-func (w *WGWatcher) PrepareInitialHandshake() (ok bool) {
-	w.muEnabled.Lock()
-	if w.enabled {
-		w.muEnabled.Unlock()
-		return false
-	}
-
+// PrepareInitialHandshake reads the peer's current WireGuard handshake time. It must be
+// called before the peer is (re)configured on the WireGuard interface, so the captured
+// baseline reflects the state prior to this connection attempt instead of racing with
+// that configuration.
+func (w *WGWatcher) PrepareInitialHandshake() {
 	w.log.Debugf("enable WireGuard watcher")
-	w.enabled = true
-	w.muEnabled.Unlock()
-
 	handshake, _ := w.wgState()
 	w.initialHandshake = handshake
-	return true
 }
 
 // EnableWgWatcher runs the WireGuard watcher loop using the handshake baseline captured by
@@ -76,10 +64,6 @@ func (w *WGWatcher) PrepareInitialHandshake() (ok bool) {
 // handshake, including the first.
 func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time), onCheckSuccessFn func()) {
 	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, onCheckSuccessFn, enabledTime, w.initialHandshake)
-
-	w.muEnabled.Lock()
-	w.enabled = false
-	w.muEnabled.Unlock()
 }
 
 // Reset signals the watcher that the WireGuard peer has been reset and a new
@@ -105,6 +89,7 @@ func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn
 		case <-timer.C:
 			handshake, ok := w.handshakeCheck(lastHandshake)
 			if !ok {
+				// early ctx cancel check return
 				if ctx.Err() != nil {
 					return
 				}
@@ -153,9 +138,9 @@ func (w *WGWatcher) handshakeCheck(lastHandshake time.Time) (*time.Time, bool) {
 
 	w.log.Tracef("previous handshake, handshake: %v, %v", lastHandshake, handshake)
 
-	// the current know handshake did not change
+	// the current known handshake did not change
 	if handshake.Equal(lastHandshake) {
-		w.log.Warnf("WireGuard handshake timed out: %v", handshake)
+		w.log.Warnf("WireGuard handshake not updated: %v", handshake)
 		return nil, false
 	}
 

client/internal/peer/wg_watcher_test.go

@@ -7,7 +7,6 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
-	"github.com/stretchr/testify/require"
 
 	"github.com/netbirdio/netbird/client/iface/configurer"
 )
@@ -62,7 +61,7 @@ func TestWGWatcher_CheckSuccessCallback(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	require.True(t, watcher.PrepareInitialHandshake())
+	watcher.PrepareInitialHandshake()
 
 	firstHandshake := make(chan struct{}, 1)
 	checkSuccess := make(chan struct{}, 1)
@@ -101,8 +100,7 @@ func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
+	watcher.PrepareInitialHandshake()
 
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
@@ -132,8 +130,7 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	watcher := NewWGWatcher(mlog, mocWgIface, "", newStateDump("peer", mlog, &Status{}))
 
 	ctx, cancel := context.WithCancel(context.Background())
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
+	watcher.PrepareInitialHandshake()
 
 	wg := &sync.WaitGroup{}
 	wg.Add(1)
@@ -149,8 +146,7 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	ctx, cancel = context.WithCancel(context.Background())
 	defer cancel()
 
-	ok = watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should be re-enabled after the previous run stopped")
+	watcher.PrepareInitialHandshake()
 
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {

client/internal/rosenpass/cache_test.go

@@ -0,0 +1,62 @@
+package rosenpass
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func newCacheTestManager(spk []byte) *Manager {
+	return &Manager{
+		spk:               spk,
+		remotePubKeys:     make(map[string][]byte),
+		remoteHasLocalKey: make(map[string]bool),
+	}
+}
+
+func TestResolveRemotePubKey(t *testing.T) {
+	m := newCacheTestManager([]byte{0x01, 0x02})
+	full := bytes.Repeat([]byte{0xAB}, 64)
+
+	// a received full key is cached and returned
+	require.Equal(t, full, m.ResolveRemotePubKey("peerA", full, nil))
+
+	// a later hash-only message resolves from the cache
+	require.Equal(t, full, m.ResolveRemotePubKey("peerA", nil, rawRosenpassKeyHash(full)))
+
+	// hash mismatch is a cache miss
+	require.Nil(t, m.ResolveRemotePubKey("peerA", nil, bytes.Repeat([]byte{0x01}, 32)))
+
+	// no key and no hash (remote without Rosenpass) resolves to nil
+	require.Nil(t, m.ResolveRemotePubKey("peerB", nil, nil))
+}
+
+func TestRemotePubKeyAck(t *testing.T) {
+	m := newCacheTestManager([]byte{0x01})
+
+	// unknown peer -> no ack (signals "send me the full key")
+	require.Nil(t, m.RemotePubKeyAck("peerA"))
+
+	full := bytes.Repeat([]byte{0x09}, 48)
+	m.ResolveRemotePubKey("peerA", full, nil)
+	require.Equal(t, rawRosenpassKeyHash(full), m.RemotePubKeyAck("peerA"))
+}
+
+func TestSetRemoteAckAndRemoteHasLocalKey(t *testing.T) {
+	m := newCacheTestManager(bytes.Repeat([]byte{0x07}, 100))
+
+	require.False(t, m.RemoteHasLocalKey("peerA"))
+
+	// an ack matching our own key hash marks the peer as holding our key
+	m.SetRemoteAck("peerA", m.LocalPubKeyHash())
+	require.True(t, m.RemoteHasLocalKey("peerA"))
+
+	// empty ack clears it
+	m.SetRemoteAck("peerA", nil)
+	require.False(t, m.RemoteHasLocalKey("peerA"))
+
+	// a non-matching ack does not count
+	m.SetRemoteAck("peerA", bytes.Repeat([]byte{0x01}, 32))
+	require.False(t, m.RemoteHasLocalKey("peerA"))
+}

client/internal/rosenpass/manager.go

@@ -8,6 +8,7 @@ import (
 	"log/slog"
 	"net"
 	"os"
+	"path/filepath"
 	"strconv"
 	"strings"
 	"sync"
@@ -28,6 +29,11 @@ func hashRosenpassKey(key []byte) string {
 	return hex.EncodeToString(hasher.Sum(nil))
 }
 
+func rawRosenpassKeyHash(key []byte) []byte {
+	sum := sha256.Sum256(key)
+	return sum[:]
+}
+
 // rpServer is the subset of rp.Server used by Manager. Defined as an interface
 // so tests can substitute a mock without spinning up a real UDP server.
 type rpServer interface {
@@ -50,12 +56,29 @@ type Manager struct {
 	lock         sync.Mutex
 	port         int
 	wgIface      PresharedKeySetter
+
+	// remotePubKeys caches remote peers' full Rosenpass public keys keyed by their
+	// WireGuard public key, so a peer that already sent us its (large) key over
+	// signalling need only send its hash on subsequent offers/answers. RAM only —
+	// never persisted (1000 peers x ~512KB would be ~512MB on disk).
+	remotePubKeys map[string][]byte
+	// remoteHasLocalKey tracks, per remote WireGuard key, whether that peer has
+	// acknowledged holding our current Rosenpass public key, letting us omit it.
+	remoteHasLocalKey map[string]bool
 }
 
 // NewManager creates a new Rosenpass manager. localWgKey is the local
-// WireGuard public key, used to derive the per-peer rendezvous key.
-func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string, localWgKey wgtypes.Key) (*Manager, error) {
-	public, secret, err := rp.GenerateKeyPair()
+// WireGuard public key, used to derive the per-peer rendezvous key. When stateDir
+// is non-empty the static keypair is persisted under it and reused across
+// restarts, keeping the public key (and the fingerprint peers cache) stable;
+// an empty stateDir keeps the previous behaviour of an ephemeral per-run keypair.
+func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string, localWgKey wgtypes.Key, stateDir string) (*Manager, error) {
+	var keyPath string
+	if stateDir != "" {
+		keyPath = filepath.Join(stateDir, keypairFileName)
+	}
+
+	public, secret, err := loadOrGenerateKeypair(keyPath)
 	if err != nil {
 		return nil, err
 	}
@@ -76,8 +99,10 @@ func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string, localWgKey wgtype
 		// nil receiver in addPeer -> m.rpWgHandler.AddPeer. generateConfig will
 		// replace it with a fresh handler on each Run() to clear stale peer
 		// state from previous engine sessions.
-		rpWgHandler: NewNetbirdHandler((*[32]byte)(preSharedKey), localWgKey),
-		lock:        sync.Mutex{},
+		rpWgHandler:       NewNetbirdHandler((*[32]byte)(preSharedKey), localWgKey),
+		lock:              sync.Mutex{},
+		remotePubKeys:     make(map[string][]byte),
+		remoteHasLocalKey: make(map[string]bool),
 	}, nil
 }
 
@@ -90,6 +115,68 @@ func (m *Manager) GetAddress() *net.UDPAddr {
 	return &net.UDPAddr{Port: m.port}
 }
 
+// LocalPubKeyHash returns the raw SHA256 of the local Rosenpass public key. It is
+// advertised on every offer/answer so the remote peer can tell (via its cache)
+// whether it already holds our full key.
+func (m *Manager) LocalPubKeyHash() []byte {
+	return rawRosenpassKeyHash(m.spk)
+}
+
+// RemotePubKeyAck returns the SHA256 of the remote peer's cached public key, used
+// as the acknowledgement we send back. Nil means we do not hold the peer's key,
+// which signals the peer to include its full key next time.
+func (m *Manager) RemotePubKeyAck(remoteWgKey string) []byte {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	key, ok := m.remotePubKeys[remoteWgKey]
+	if !ok {
+		return nil
+	}
+	return rawRosenpassKeyHash(key)
+}
+
+// RemoteHasLocalKey reports whether the remote peer acknowledged holding our
+// current public key, so we may omit the full key from the next offer/answer.
+func (m *Manager) RemoteHasLocalKey(remoteWgKey string) bool {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	return m.remoteHasLocalKey[remoteWgKey]
+}
+
+// ResolveRemotePubKey reconciles the Rosenpass key material from a received
+// offer/answer: it caches a received full key, or — when only a hash was sent —
+// returns the cached key matching that hash. It returns nil when the remote peer
+// does not use Rosenpass (no key, no hash) or on a cache miss (hash sent but not
+// held); a miss self-heals because our resulting empty ack makes the peer resend
+// its full key.
+func (m *Manager) ResolveRemotePubKey(remoteWgKey string, full, hash []byte) []byte {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	if len(full) > 0 {
+		m.remotePubKeys[remoteWgKey] = full
+		return full
+	}
+	if len(hash) == 0 {
+		return nil
+	}
+	if cached, ok := m.remotePubKeys[remoteWgKey]; ok && bytes.Equal(rawRosenpassKeyHash(cached), hash) {
+		return cached
+	}
+	return nil
+}
+
+// SetRemoteAck records whether the remote peer's acknowledgement matches our
+// current public key hash, i.e. whether it already holds our key.
+func (m *Manager) SetRemoteAck(remoteWgKey string, ack []byte) {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	m.remoteHasLocalKey[remoteWgKey] = len(ack) > 0 && bytes.Equal(ack, rawRosenpassKeyHash(m.spk))
+}
+
 // addPeer adds a new peer to the Rosenpass server
 func (m *Manager) addPeer(rosenpassPubKey []byte, rosenpassAddr string, wireGuardIP string, wireGuardPubKey string) error {
 	// Defense in depth against issue #4341 (Android crash): if Run() has not

client/internal/rosenpass/manager_test.go

@@ -255,7 +255,7 @@ func TestAddPeer_NilServer_ReturnsErrorNoCrash(t *testing.T) {
 // issue #4341 cannot occur in the window between NewManager and Run().
 func TestNewManager_PreInitializesHandler(t *testing.T) {
 	psk := wgtypes.Key{}
-	m, err := NewManager(&psk, "wt0", wgtypes.Key{0x01})
+	m, err := NewManager(&psk, "wt0", wgtypes.Key{0x01}, "")
 	require.NoError(t, err)
 	require.NotNil(t, m.rpWgHandler, "rpWgHandler must be initialized in NewManager")
 }

client/internal/rosenpass/persistence.go

@@ -0,0 +1,92 @@
+package rosenpass
+
+import (
+	"context"
+	"fmt"
+	"os"
+
+	rp "cunicu.li/go-rosenpass"
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/util"
+)
+
+const (
+	// keypairFileName is the file, relative to the state directory, that holds
+	// the persisted local Rosenpass static keypair.
+	keypairFileName = "rosenpass_key.json"
+
+	// rpStaticPublicKeySize is the byte length of a Rosenpass (Classic McEliece)
+	// static public key as produced by the pinned go-rosenpass version. Used as a
+	// version-compatibility guard: a persisted key of any other size is treated as
+	// stale and regenerated instead of being fed to go-rosenpass (which would fail).
+	rpStaticPublicKeySize = 524160
+
+	// keypairFormatVersion is bumped whenever the on-disk representation changes so
+	// old files are discarded and regenerated rather than misparsed.
+	keypairFormatVersion = 1
+)
+
+// persistedKeypair is the on-disk representation of the local Rosenpass static
+// keypair. Keys are stored raw (base64 via JSON) with the same restricted 0600
+// permission as the WireGuard private key and other client secrets.
+type persistedKeypair struct {
+	Version   int    `json:"version"`
+	PublicKey []byte `json:"public_key"`
+	SecretKey []byte `json:"secret_key"`
+}
+
+// loadOrGenerateKeypair returns a Rosenpass static keypair. When keyPath is set
+// and holds a valid persisted keypair it is reused, so the local public key —
+// and therefore the fingerprint advertised to remote peers over signalling —
+// stays stable across restarts. Otherwise a fresh keypair is generated and, when
+// keyPath is set, persisted for subsequent runs. A missing or corrupt file is not
+// fatal: it degrades to generating an ephemeral keypair, matching the pre-persistence
+// behaviour.
+func loadOrGenerateKeypair(keyPath string) (public []byte, secret []byte, err error) {
+	if keyPath != "" {
+		public, secret, err = loadKeypair(keyPath)
+		switch {
+		case err == nil:
+			return public, secret, nil
+		case os.IsNotExist(err):
+			// first run for this state dir; fall through to generate
+		default:
+			log.Warnf("failed to load persisted rosenpass keypair, generating a new one: %v", err)
+		}
+	}
+
+	pub, sec, err := rp.GenerateKeyPair()
+	if err != nil {
+		return nil, nil, fmt.Errorf("generate rosenpass key pair: %w", err)
+	}
+
+	if keyPath != "" {
+		if err := saveKeypair(keyPath, pub, sec); err != nil {
+			log.Warnf("failed to persist rosenpass keypair, key will be regenerated on next restart: %v", err)
+		}
+	}
+
+	return pub, sec, nil
+}
+
+func loadKeypair(keyPath string) ([]byte, []byte, error) {
+	var kp persistedKeypair
+	if _, err := util.ReadJson(keyPath, &kp); err != nil {
+		return nil, nil, err
+	}
+
+	if kp.Version != keypairFormatVersion || len(kp.PublicKey) != rpStaticPublicKeySize || len(kp.SecretKey) == 0 {
+		return nil, nil, fmt.Errorf("persisted rosenpass keypair is incompatible (version %d, public %d bytes, secret %d bytes)", kp.Version, len(kp.PublicKey), len(kp.SecretKey))
+	}
+
+	return kp.PublicKey, kp.SecretKey, nil
+}
+
+func saveKeypair(keyPath string, public, secret []byte) error {
+	return util.WriteJsonWithRestrictedPermission(context.Background(), keyPath, persistedKeypair{
+		Version:   keypairFormatVersion,
+		PublicKey: public,
+		SecretKey: secret,
+	})
+}

client/internal/rosenpass/persistence_test.go

@@ -0,0 +1,66 @@
+package rosenpass
+
+import (
+	"bytes"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestLoadOrGenerateKeypair_EphemeralWhenNoPath(t *testing.T) {
+	pub, sec, err := loadOrGenerateKeypair("")
+	require.NoError(t, err)
+	require.Len(t, pub, rpStaticPublicKeySize)
+	require.NotEmpty(t, sec)
+}
+
+func TestLoadOrGenerateKeypair_PersistsAndReloads(t *testing.T) {
+	keyPath := filepath.Join(t.TempDir(), keypairFileName)
+
+	pub1, sec1, err := loadOrGenerateKeypair(keyPath)
+	require.NoError(t, err)
+
+	info, err := os.Stat(keyPath)
+	require.NoError(t, err, "keypair file must be written")
+	require.Equal(t, os.FileMode(0600), info.Mode().Perm(), "keypair file must be 0600")
+
+	pub2, sec2, err := loadOrGenerateKeypair(keyPath)
+	require.NoError(t, err)
+	require.True(t, bytes.Equal(pub1, pub2), "public key must be stable across reloads")
+	require.True(t, bytes.Equal(sec1, sec2), "secret key must be stable across reloads")
+}
+
+func TestLoadOrGenerateKeypair_RegeneratesOnCorruptFile(t *testing.T) {
+	keyPath := filepath.Join(t.TempDir(), keypairFileName)
+	require.NoError(t, os.WriteFile(keyPath, []byte("not json"), 0600))
+
+	pub, sec, err := loadOrGenerateKeypair(keyPath)
+	require.NoError(t, err)
+	require.Len(t, pub, rpStaticPublicKeySize)
+	require.NotEmpty(t, sec)
+
+	// the corrupt file must have been overwritten with a valid, reloadable keypair
+	pub2, _, err := loadOrGenerateKeypair(keyPath)
+	require.NoError(t, err)
+	require.True(t, bytes.Equal(pub, pub2))
+}
+
+func TestLoadOrGenerateKeypair_RegeneratesOnVersionMismatch(t *testing.T) {
+	keyPath := filepath.Join(t.TempDir(), keypairFileName)
+
+	pub1, _, err := loadOrGenerateKeypair(keyPath)
+	require.NoError(t, err)
+
+	// rewrite with a bumped/unknown format version -> must be discarded
+	bs, err := json.Marshal(persistedKeypair{Version: keypairFormatVersion + 1, PublicKey: pub1, SecretKey: []byte{0x01}})
+	require.NoError(t, err)
+	require.NoError(t, os.WriteFile(keyPath, bs, 0600))
+
+	pub2, sec2, err := loadOrGenerateKeypair(keyPath)
+	require.NoError(t, err)
+	require.Len(t, pub2, rpStaticPublicKeySize)
+	require.NotEmpty(t, sec2)
+}

client/server/server.go

@@ -181,7 +181,7 @@ func (s *Server) Start() error {
 		log.Warnf("failed to redirect stderr: %v", err)
 	}
 
-	if err := restoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
+	if err := RestoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
 		log.Warnf(errRestoreResidualState, err)
 	}
 
@@ -551,7 +551,7 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 	s.actCancel = cancel
 	s.mutex.Unlock()
 
-	if err := restoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
+	if err := RestoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
 		log.Warnf(errRestoreResidualState, err)
 	}
 
@@ -858,7 +858,7 @@ func (s *Server) Up(callerCtx context.Context, msg *proto.UpRequest) (*proto.UpR
 
 		return s.waitForUp(callerCtx)
 	}
-	if err := restoreResidualState(callerCtx, s.profileManager.GetStatePath()); err != nil {
+	if err := RestoreResidualState(callerCtx, s.profileManager.GetStatePath()); err != nil {
 		log.Warnf(errRestoreResidualState, err)
 	}
 

client/server/state.go

@@ -46,7 +46,7 @@ func (s *Server) CleanState(ctx context.Context, req *proto.CleanStateRequest) (
 
 	if req.All {
 		// Reuse existing cleanup logic for all states
-		if err := restoreResidualState(ctx, statePath); err != nil {
+		if err := RestoreResidualState(ctx, statePath); err != nil {
 			return nil, status.Errorf(codes.Internal, "failed to clean all states: %v", err)
 		}
 
@@ -113,9 +113,9 @@ func (s *Server) DeleteState(ctx context.Context, req *proto.DeleteStateRequest)
 	}, nil
 }
 
-// restoreResidualState checks if the client was not shut down in a clean way and restores residual if required.
+// RestoreResidualState checks if the client was not shut down in a clean way and restores residual if required.
 // Otherwise, we might not be able to connect to the management server to retrieve new config.
-func restoreResidualState(ctx context.Context, statePath string) error {
+func RestoreResidualState(ctx context.Context, statePath string) error {
 	if statePath == "" {
 		return nil
 	}

e2e/agentnetwork/chat_test.go

@@ -91,7 +91,7 @@ func availableProviders() []providerCase {
 		if region == "" {
 			region = "us-east-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.WireMessages})
+		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})
 	}
 	return ps
 }
@@ -224,9 +224,12 @@ func TestProvidersMatrix(t *testing.T) {
 				var c int
 				var b string
 				var cerr error
-				if pc.kind == harness.WireVertex {
+				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)
-				} else {
+				case harness.WireBedrock:
+					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, "Reply with exactly: pong", sessionID)
+				default:
 					c, b, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
 				}
 				if cerr == nil {

e2e/agentnetwork/guardrail_test.go

@@ -0,0 +1,168 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// 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).
+func catalogModel(pc providerCase) string {
+	switch pc.kind {
+	case harness.WireBedrock:
+		return strings.TrimPrefix(pc.model, "us.")
+	case harness.WireVertex:
+		return strings.SplitN(pc.model, "@", 2)[0]
+	default:
+		return pc.model
+	}
+}
+
+// disallowedModel returns a valid-shaped model id for the provider that is NOT
+// the configured/allowed one, so the guardrail must reject it before the
+// request ever reaches the upstream.
+func disallowedModel(pc providerCase) string {
+	switch pc.kind {
+	case harness.WireBedrock:
+		return "us.anthropic.claude-opus-4-8"
+	case harness.WireVertex:
+		return "claude-opus-4-8@20250101"
+	default:
+		return "unlisted-model"
+	}
+}
+
+// sendModel drives one request for the given model through the provider's native
+// wire shape and returns the HTTP status.
+func sendModel(ctx context.Context, t *testing.T, cl *harness.Client, endpoint, proxyIP string, pc providerCase, model string) int {
+	t.Helper()
+	var code int
+	var err error
+	switch pc.kind {
+	case harness.WireBedrock:
+		code, _, err = cl.Bedrock(ctx, endpoint, proxyIP, model, "Reply with exactly: pong", "")
+	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", "")
+	}
+	require.NoError(t, err, "request must reach the proxy for %s", pc.name)
+	return code
+}
+
+// TestModelAllowlistEnforced provisions a Model Allowlist guardrail limiting each
+// path-routed provider (Bedrock, Vertex) to its configured model, then drives
+// requests over the tunnel: the allowed model returns 200 while a model outside
+// the allowlist is denied 403 by the guardrail before it reaches the upstream.
+// This is the coverage missing for #6751 — the model for these providers travels
+// in the URL path, and the allowlist must be enforced there.
+func TestModelAllowlistEnforced(t *testing.T) {
+	var providers []providerCase
+	for _, pc := range availableProviders() {
+		if pc.kind == harness.WireBedrock || pc.kind == harness.WireVertex {
+			providers = append(providers, pc)
+		}
+	}
+	if len(providers) == 0 {
+		t.Skip("no path-routed provider keys set (AWS_BEARER_TOKEN_BEDROCK / GOOGLE_VERTEX_*); source ~/.llm-keys")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-allowlist"})
+	require.NoError(t, err, "create group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-allowlist-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grp.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+
+	// Providers with their configured (allowed) models; the first bootstraps the cluster.
+	ids := make([]string, 0, len(providers))
+	allowed := make([]string, 0, len(providers))
+	for i, pc := range providers {
+		req := providerRequest(pc)
+		if i == 0 {
+			req.BootstrapCluster = ptr(harness.AgentNetworkCluster)
+		}
+		prov, perr := srv.CreateProvider(ctx, req)
+		require.NoError(t, perr, "create provider %s", pc.name)
+		id := prov.Id
+		ids = append(ids, id)
+		allowed = append(allowed, catalogModel(pc))
+		t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), id) })
+	}
+
+	// Guardrail allowlisting exactly the configured models.
+	var gr api.AgentNetworkGuardrailRequest
+	gr.Name = "e2e-allowlist"
+	gr.Checks.ModelAllowlist.Enabled = true
+	gr.Checks.ModelAllowlist.Models = allowed
+	guard, err := srv.CreateGuardrail(ctx, gr)
+	require.NoError(t, err, "create guardrail")
+	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), guard.Id) })
+
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-allowlist",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: ids,
+		GuardrailIds:           &[]string{guard.Id},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings for endpoint")
+	require.NotEmpty(t, settings.Endpoint, "agent-network endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-proxy-allowlist")
+	require.NoError(t, err, "mint proxy token via CLI")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
+
+	for _, pc := range providers {
+		pc := pc
+		t.Run(pc.name, func(t *testing.T) {
+			// The admin's allowlisted model is served end to end.
+			assert.Equal(t, 200, sendModel(ctx, t, cl, settings.Endpoint, proxyIP, pc, pc.model),
+				"allowlisted model must be permitted for %s", pc.name)
+			// A model outside the allowlist is rejected by the guardrail (before
+			// the upstream), regardless of whether it is a real catalog model.
+			assert.Equal(t, 403, sendModel(ctx, t, cl, settings.Endpoint, proxyIP, pc, disallowedModel(pc)),
+				"model outside the allowlist must be denied for %s", pc.name)
+		})
+	}
+}

e2e/harness/agentnetwork.go

@@ -107,6 +107,17 @@ func (c *Combined) DeletePolicy(ctx context.Context, id string) error {
 	return anDelete(ctx, c, "/api/agent-network/policies/"+id)
 }
 
+// CreateGuardrail creates an agent-network guardrail (e.g. a model allowlist)
+// that can then be attached to a policy via its GuardrailIds.
+func (c *Combined) CreateGuardrail(ctx context.Context, req api.AgentNetworkGuardrailRequest) (api.AgentNetworkGuardrail, error) {
+	return anRequest[api.AgentNetworkGuardrail](ctx, c, http.MethodPost, "/api/agent-network/guardrails", req)
+}
+
+// DeleteGuardrail removes a guardrail by id.
+func (c *Combined) DeleteGuardrail(ctx context.Context, id string) error {
+	return anDelete(ctx, c, "/api/agent-network/guardrails/"+id)
+}
+
 // GetSettings returns the account's agent-network settings row. It exists only
 // after the first provider create bootstraps it.
 func (c *Combined) GetSettings(ctx context.Context) (api.AgentNetworkSettings, error) {

e2e/harness/client.go

@@ -194,6 +194,11 @@ const (
 	// WireVertex is the Anthropic-on-Vertex rawPredict shape: the client posts
 	// the full Vertex model path and the proxy mints the SA OAuth token.
 	WireVertex = "vertex"
+	// WireBedrock is the native AWS Bedrock InvokeModel shape: the model id
+	// travels in the URL path (/model/{id}/invoke), not the body, so the proxy
+	// routes by path. This is what a Bedrock SDK client sends and the shape the
+	// model-allowlist guardrail must enforce.
+	WireBedrock = "bedrock"
 )
 
 // Chat issues a chat-completion POST to the agent-network endpoint over the
@@ -226,6 +231,17 @@ func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
+// Bedrock issues a native AWS Bedrock InvokeModel POST over the tunnel. The
+// model id is carried in the request path (/model/{id}/invoke), so the proxy
+// routes by path; the body uses the bedrock anthropic_version rather than a
+// model field. A non-empty sessionID is sent as the universal x-session-id
+// 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)
+	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
+}
+
 // withSessionID appends the x-session-id header when sessionID is non-empty.
 func withSessionID(headers []string, sessionID string) []string {
 	if sessionID == "" {

management/internals/modules/peers/manager.go

@@ -226,30 +226,6 @@ func (m *managerImpl) CreateProxyPeer(ctx context.Context, accountID string, pee
 		return nil
 	}
 
-	// Dedupe stale embedded peer records for the same (account, cluster).
-	// The proxy generates a fresh WireGuard keypair on every startup
-	// (proxy/internal/roundtrip/netbird.go), so without this sweep the
-	// prior embedded peer would linger forever — holding its CGNAT IP
-	// allocation, polluting other peers' rosters, and (most visibly)
-	// leaving the synth DNS pointing at the dead address. The
-	// (account, cluster) tuple identifies "the embedded peer for this
-	// proxy instance at this cluster"; any record matching that tuple
-	// with a different pubkey is by definition stale and must go.
-	staleIDs, err := m.findStaleEmbeddedProxyPeers(ctx, accountID, cluster, peerKey)
-	if err != nil {
-		return fmt.Errorf("scan for stale embedded proxy peers: %w", err)
-	}
-	if len(staleIDs) > 0 {
-		// userID="" + checkConnected=false: the deletion is initiated
-		// by management itself on behalf of the freshly-registering
-		// proxy, not by an end user; the stale peer may still be
-		// marked Connected from its prior session, but its session is
-		// dead by definition (its key no longer exists).
-		if err := m.DeletePeers(ctx, accountID, staleIDs, "", false); err != nil {
-			return fmt.Errorf("delete stale embedded proxy peers %v: %w", staleIDs, err)
-		}
-	}
-
 	name := fmt.Sprintf("proxy-%s", xid.New().String())
 	newPeer := &peer.Peer{
 		Ephemeral: true,
@@ -275,29 +251,3 @@ func (m *managerImpl) CreateProxyPeer(ctx context.Context, accountID string, pee
 
 	return nil
 }
-
-// findStaleEmbeddedProxyPeers returns the peer IDs of embedded proxy peer
-// records in accountID that target the same cluster but carry a different
-// WireGuard pubkey than the freshly-registering one. Used by CreateProxyPeer
-// to garbage-collect stale records left behind when the proxy restarts with a
-// regenerated keypair.
-func (m *managerImpl) findStaleEmbeddedProxyPeers(ctx context.Context, accountID, cluster, newKey string) ([]string, error) {
-	account, err := m.store.GetAccount(ctx, accountID)
-	if err != nil {
-		return nil, err
-	}
-	var stale []string
-	for _, p := range account.Peers {
-		if p == nil || !p.ProxyMeta.Embedded {
-			continue
-		}
-		if p.ProxyMeta.Cluster != cluster {
-			continue
-		}
-		if p.Key == newKey {
-			continue
-		}
-		stale = append(stale, p.ID)
-	}
-	return stale, nil
-}

management/main.go

@@ -1,19 +1,24 @@
 package main
 
 import (
-	"log"
 	"net/http"
 	// nolint:gosec
 	_ "net/http/pprof"
 	"os"
 
+	log "github.com/sirupsen/logrus"
+
 	"github.com/netbirdio/netbird/management/cmd"
 )
 
 func main() {
-	go func() {
-		log.Println(http.ListenAndServe("localhost:6060", nil))
-	}()
+	if pprofAddr := os.Getenv("NB_PPROF_ADDR"); pprofAddr != "" {
+		log.Infof("pprof enabled, listening on: %s", pprofAddr)
+		go func() {
+			log.Println(http.ListenAndServe(pprofAddr, nil))
+		}()
+	}
+
 	if err := cmd.Execute(); err != nil {
 		os.Exit(1)
 	}

management/server/agentnetwork_proxypeer_restart_test.go

@@ -1,199 +0,0 @@
-package server
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/management/internals/modules/peers"
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
-	agenttypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
-	nbpeer "github.com/netbirdio/netbird/management/server/peer"
-	"github.com/netbirdio/netbird/management/server/permissions"
-	"github.com/netbirdio/netbird/management/server/store"
-	"github.com/netbirdio/netbird/management/server/types"
-)
-
-// TestAgentNetwork_ProxyRestart_PropagatesNewPeerAndDropsStale is the no-mock
-// regression guard for the bug the user reported: restarting the proxy creates
-// a fresh embedded peer with a NEW WireGuard public key (the proxy generates
-// the keypair on every startup at proxy/internal/roundtrip/netbird.go:312).
-// The PRIOR embedded peer record is never deleted on management, so the
-// account accumulates a stale peer holding a stale CGNAT IP. Other peers
-// in the account either keep routing to the dead IP, or — if synth DNS
-// picks the wrong record — never see the new IP at all.
-//
-// What this test exercises (no mocks):
-//   - real SQLite test store
-//   - real DefaultAccountManager, network-map controller, peer-update channels
-//   - real peers.Manager.CreateProxyPeer path (the very method the proxy
-//     invokes over gRPC on every startup)
-//   - real agentnetwork.Manager + synth chain so the client receives a
-//     concrete DNS record that must point at the LATEST proxy peer.
-//
-// Pre-fix expected behavior (red): two embedded peers exist after the
-// "restart"; the synth DNS record points at the stale one; the client
-// receives an update reflecting the new peer but the old one lingers.
-// Post-fix expected behavior (green): exactly one embedded peer exists
-// after restart (with the new key) AND the client's network map carries
-// the synth DNS pointing at that new peer's CGNAT IP.
-func TestAgentNetwork_ProxyRestart_PropagatesNewPeerAndDropsStale(t *testing.T) {
-	am, updateManager, err := createManager(t)
-	require.NoError(t, err, "createManager must succeed")
-	ctx := context.Background()
-
-	const (
-		accountID   = "an-restart-acct"
-		adminUserID = "an-restart-admin"
-		groupAID    = "an-restart-grp-A"
-		clusterAddr = "eu.proxy.netbird.io"
-		clientKey   = "BhRPtynAAYRDy08+q4HTMsos8fs4plTP4NOSh7C1ry8="
-		// Two different proxy pubkeys — the "before" and "after" of a
-		// proxy-process restart with fresh-keypair generation.
-		proxyKey1 = "Aaaaa1aaaaYRDy08+q4HTMsos8fs4plTP4NOSh7C1ry8="
-		proxyKey2 = "Bbbbb2bbbbYRDy08+q4HTMsos8fs4plTP4NOSh7C1ry8="
-	)
-
-	// --- Account scaffold ---
-	account := newAccountWithId(ctx, accountID, adminUserID, "an-restart.test", "", "", false)
-	require.NoError(t, am.Store.SaveAccount(ctx, account))
-
-	clientPeer := &nbpeer.Peer{
-		Key:      clientKey,
-		Name:     "an-restart-client",
-		DNSLabel: "an-restart-client",
-		Meta:     nbpeer.PeerSystemMeta{Hostname: "an-restart-client", GoOS: "linux", WtVersion: "development"},
-	}
-	addedClient, _, _, _, err := am.AddPeer(ctx, "", "", adminUserID, clientPeer, false)
-	require.NoError(t, err, "AddPeer for client must succeed")
-	require.NoError(t, am.MarkPeerConnected(ctx, clientKey, accountID, time.Now().UnixNano(), &types.NetworkMap{}),
-		"MarkPeerConnected for the client peer must succeed (affected-peer fan-out skips disconnected peers)")
-
-	// Place the client in group A so the synth policy reaches it.
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	account.Groups[groupAID] = &types.Group{ID: groupAID, Name: "groupA", Peers: []string{addedClient.ID}}
-	require.NoError(t, am.Store.SaveAccount(ctx, account), "SaveAccount must persist group A")
-
-	// --- Real peers + agent-network managers ---
-	permMgr := permissions.NewManager(am.Store)
-	peersMgr := peers.NewManager(am.Store, permMgr)
-	peersMgr.SetAccountManager(am)
-	peersMgr.SetNetworkMapController(am.networkMapController)
-	agentMgr := agentnetwork.NewManager(am.Store, permMgr, am, nil)
-
-	// Subscribe BEFORE any state-mutating call so we don't lose the update
-	// that contains the synth DNS record.
-	clientCh := updateManager.CreateChannel(ctx, addedClient.ID)
-	t.Cleanup(func() { updateManager.CloseChannel(ctx, addedClient.ID) })
-	drain(clientCh)
-
-	// --- First proxy startup: register peer key K1, then mark it
-	// connected. In production the proxy follows CreateProxyPeer with the
-	// regular sync stream which lands on MarkPeerConnected; the synth DNS
-	// path filters out peers that aren't Connected (types/account.go:323),
-	// so without this step no DNS record would be emitted.
-	require.NoError(t, peersMgr.CreateProxyPeer(ctx, accountID, proxyKey1, clusterAddr),
-		"first CreateProxyPeer (proxy startup) must succeed")
-
-	peer1ID, err := am.Store.GetPeerIDByKey(ctx, store.LockingStrengthNone, proxyKey1)
-	require.NoError(t, err, "proxy peer for K1 must be persisted after CreateProxyPeer")
-	require.NotEmpty(t, peer1ID)
-
-	require.NoError(t, am.MarkPeerConnected(ctx, proxyKey1, accountID, time.Now().UnixNano(), &types.NetworkMap{}),
-		"MarkPeerConnected for K1 must succeed")
-
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	proxyIP1 := account.Peers[peer1ID].IP.String()
-	require.NotEmpty(t, proxyIP1, "K1 must have an assigned overlay IP")
-
-	// --- Provider + policy. CreateProvider / CreatePolicy trigger the
-	// agentnetwork reconcile which runs UpdateAccountPeers; the resulting
-	// NetworkMap delivered to the client carries the synth DNS record
-	// pointing at K1's IP. ---
-	provider, err := agentMgr.CreateProvider(ctx, adminUserID, &agenttypes.Provider{
-		AccountID:   accountID,
-		ProviderID:  "openai_api",
-		Name:        "openai-test",
-		UpstreamURL: "https://api.openai.com",
-		APIKey:      "sk-test-key",
-		Enabled:     true,
-		Models:      []agenttypes.ProviderModel{{ID: "gpt-5.4"}},
-	}, clusterAddr)
-	require.NoError(t, err, "CreateProvider must succeed")
-
-	_, err = agentMgr.CreatePolicy(ctx, adminUserID, &agenttypes.Policy{
-		AccountID:              accountID,
-		Name:                   "p1",
-		Enabled:                true,
-		SourceGroups:           []string{groupAID},
-		DestinationProviderIDs: []string{provider.ID},
-	})
-	require.NoError(t, err, "CreatePolicy must succeed")
-
-	settings, err := am.Store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
-	require.NoError(t, err)
-	fqdn := settings.Endpoint()
-
-	rdata1 := awaitZoneRData(clientCh, clusterAddr, fqdn, true)
-	require.Equal(t, proxyIP1, rdata1,
-		"client must receive a synth DNS record pointing at K1's overlay IP after the synth path runs")
-	drain(clientCh)
-
-	// --- Proxy restart: NEW keypair K2, same account, same cluster ---
-	require.NoError(t, peersMgr.CreateProxyPeer(ctx, accountID, proxyKey2, clusterAddr),
-		"second CreateProxyPeer (proxy restart with fresh keypair) must succeed")
-
-	peer2ID, err := am.Store.GetPeerIDByKey(ctx, store.LockingStrengthNone, proxyKey2)
-	require.NoError(t, err, "proxy peer for K2 must be persisted after restart")
-	require.NotEmpty(t, peer2ID)
-
-	require.NoError(t, am.MarkPeerConnected(ctx, proxyKey2, accountID, time.Now().UnixNano(), &types.NetworkMap{}),
-		"MarkPeerConnected for K2 must succeed")
-
-	// In production the agent's sync stream pulls a fresh NetworkMap as
-	// part of its normal reconcile cadence; in this isolated test
-	// MarkPeerConnected's affected-peer fan-out can race the channel-side
-	// buffer in a way that swallows the synth-DNS-bearing update before
-	// our await reads it. Trigger an explicit account-wide fan-out so the
-	// assertion below tests what production actually delivers, not the
-	// in-test buffer race.
-	am.UpdateAccountPeers(ctx, accountID, types.UpdateReason{Resource: types.UpdateResourcePeer, Operation: types.UpdateOperationUpdate})
-
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	proxyIP2 := account.Peers[peer2ID].IP.String()
-	require.NotEmpty(t, proxyIP2, "K2 must have an assigned overlay IP")
-	require.NotEqual(t, proxyIP1, proxyIP2, "K2 must get a different overlay IP than K1 (sanity)")
-
-	// CRITICAL ASSERTION 1: K1 must no longer be in the store. The SqlStore
-	// returns ("", nil) for a missing key rather than NotFound, so assert
-	// on the returned ID being empty.
-	staleID, err := am.Store.GetPeerIDByKey(ctx, store.LockingStrengthNone, proxyKey1)
-	require.NoError(t, err, "GetPeerIDByKey for a missing peer must not error")
-	assert.Empty(t, staleID,
-		"stale embedded proxy peer K1 must be removed when a new embedded peer registers for the same (account, cluster); pre-fix this assertion fails because management never cleans up the prior peer record")
-
-	// CRITICAL ASSERTION 2: exactly one embedded proxy peer remains, and it
-	// is K2.
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	embeddedKeys := []string{}
-	for _, p := range account.Peers {
-		if p.ProxyMeta.Embedded {
-			embeddedKeys = append(embeddedKeys, p.Key)
-		}
-	}
-	assert.Equal(t, []string{proxyKey2}, embeddedKeys,
-		"after a proxy restart exactly one embedded proxy peer should remain — the one with the new key K2")
-
-	// CRITICAL ASSERTION 3: the synth DNS record the client receives now
-	// points at K2's IP, not K1's.
-	rdata2 := awaitZoneRData(clientCh, clusterAddr, fqdn, true)
-	assert.Equal(t, proxyIP2, rdata2,
-		"after proxy restart, the client's synth DNS record must point at the NEW embedded peer's IP, not the stale K1 IP")
-}

management/server/types/networkmap_components.go

@@ -7,6 +7,7 @@ import (
 	"slices"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/netbirdio/netbird/client/ssh/auth"
@@ -42,6 +43,14 @@ type NetworkMapComponents struct {
 	PostureFailedPeers map[string]map[string]struct{}
 
 	RouterPeers map[string]*nbpeer.Peer
+
+	routesByPeerOnce sync.Once
+	routesByPeerIdx  map[string][]routeIndexEntry
+}
+
+type routeIndexEntry struct {
+	route    *route.Route
+	viaGroup bool
 }
 
 type AccountSettingsInfo struct {
@@ -530,33 +539,43 @@ func (c *NetworkMapComponents) getRoutingPeerRoutes(peerID string) (enabledRoute
 		disabledRoutes = append(disabledRoutes, r)
 	}
 
-	for _, r := range c.Routes {
-		for _, groupID := range r.PeerGroups {
-			group := c.GetGroupInfo(groupID)
-			if group == nil {
-				continue
-			}
-			for _, id := range group.Peers {
-				if id != peerID {
-					continue
-				}
-
-				newPeerRoute := r.Copy()
-				newPeerRoute.Peer = id
-				newPeerRoute.PeerGroups = nil
-				newPeerRoute.ID = route.ID(string(r.ID) + ":" + id)
-				takeRoute(newPeerRoute)
-				break
-			}
-		}
-		if r.Peer == peerID {
-			takeRoute(r.Copy())
+	for _, entry := range c.routesByPeer()[peerID] {
+		if entry.viaGroup {
+			newPeerRoute := entry.route.Copy()
+			newPeerRoute.PeerGroups = nil
+			newPeerRoute.ID = route.ID(string(entry.route.ID) + ":" + peerID)
+			takeRoute(newPeerRoute)
+			continue
 		}
+		takeRoute(entry.route.Copy())
 	}
 
 	return enabledRoutes, disabledRoutes
 }
 
+func (c *NetworkMapComponents) routesByPeer() map[string][]routeIndexEntry {
+	c.routesByPeerOnce.Do(func() {
+		idx := make(map[string][]routeIndexEntry)
+		for _, r := range c.Routes {
+			for _, groupID := range r.PeerGroups {
+				group := c.GetGroupInfo(groupID)
+				if group == nil {
+					continue
+				}
+				for _, id := range group.Peers {
+					idx[id] = append(idx[id], routeIndexEntry{route: r, viaGroup: true})
+				}
+			}
+			if r.Peer != "" {
+				idx[r.Peer] = append(idx[r.Peer], routeIndexEntry{route: r})
+			}
+		}
+		c.routesByPeerIdx = idx
+	})
+
+	return c.routesByPeerIdx
+}
+
 func (c *NetworkMapComponents) filterRoutesByGroups(routes []*route.Route, groupListMap LookupMap) []*route.Route {
 	var filteredRoutes []*route.Route
 	for _, r := range routes {

proxy/internal/middleware/builtin/llm_guardrail/middleware.go

@@ -25,6 +25,14 @@ const (
 	denyCodeModel    = "llm_policy.model_blocked"
 	denyReasonModel  = "model_blocked"
 	denyMessageModel = "model is not in the policy allowlist"
+	// Deny reason used when an allowlist is configured but the request model
+	// could not be determined. URL/path-routed providers (AWS Bedrock, Google
+	// Vertex, ...) carry the model outside the JSON body, so a request shape the
+	// parser does not recognise reaches the guardrail with no model. Such a
+	// request must be denied (fail closed), never waved through.
+	denyCodeModelUnknown    = "llm_policy.model_unknown"
+	denyReasonModelUnknown  = "model_unknown"
+	denyMessageModelUnknown = "request model could not be determined for the policy allowlist"
 )
 
 // Middleware enforces the model allowlist and optionally captures the
@@ -108,23 +116,37 @@ func (m *Middleware) evaluateAllowlist(model string, modelPresent bool) *middlew
 	if len(m.cfg.ModelAllowlist) == 0 {
 		return nil
 	}
-	if !modelPresent {
-		return nil
+	// Fail closed: with an allowlist configured, a request whose model the
+	// upstream parser could not extract (absent or empty) must be denied rather
+	// than allowed. This is what enforces the allowlist for URL/path-routed
+	// providers (Bedrock, Vertex, ...) whose model lives outside the JSON body.
+	if !modelPresent || normaliseModel(model) == "" {
+		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
 	if m.modelInAllowlist(model) {
 		return nil
 	}
+	return denyModel(model, denyCodeModel, denyMessageModel, denyReasonModel)
+}
+
+// denyModel builds a 403 deny Output for a model-allowlist rejection. model is
+// included in the details only when non-empty.
+func denyModel(model, code, message, reason string) *middleware.Output {
+	details := map[string]string{}
+	if model != "" {
+		details["model"] = model
+	}
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
-			Code:    denyCodeModel,
-			Message: denyMessageModel,
-			Details: map[string]string{"model": model},
+			Code:    code,
+			Message: message,
+			Details: details,
 		},
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMPolicyDecision, Value: "deny"},
-			{Key: middleware.KeyLLMPolicyReason, Value: denyReasonModel},
+			{Key: middleware.KeyLLMPolicyReason, Value: reason},
 		},
 	}
 }

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

@@ -102,13 +102,44 @@ func TestAllowlistCaseInsensitive(t *testing.T) {
 	}
 }
 
-func TestAllowlistMissingModelKeyAllows(t *testing.T) {
+func TestAllowlistMissingModelKeyDenies(t *testing.T) {
+	// Fail closed: with an allowlist configured, a request whose model the
+	// parser could not extract (URL/path-routed providers such as Bedrock or
+	// Vertex whose shape wasn't recognised) must be denied, not allowed.
 	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
 	out, err := mw.Invoke(context.Background(), newInput())
 	require.NoError(t, err)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "missing model key must allow even with non-empty allowlist")
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when an allowlist is set")
+	assert.Equal(t, 403, out.DenyStatus, "deny status must be 403")
+	require.NotNil(t, out.DenyReason, "deny reason must be populated")
+	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
 	dec, _ := metaValue(t, out.Metadata, middleware.KeyLLMPolicyDecision)
-	assert.Equal(t, "allow", dec, "decision must be allow when model key is absent")
+	assert.Equal(t, "deny", dec, "decision must be deny when model key is absent")
+	reason, _ := metaValue(t, out.Metadata, middleware.KeyLLMPolicyReason)
+	assert.Equal(t, "model_unknown", reason, "reason metadata must be model_unknown")
+}
+
+func TestAllowlistEmptyModelValueDenies(t *testing.T) {
+	// A present-but-empty model is as undeterminable as an absent one.
+	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
+	out, err := mw.Invoke(context.Background(), newInput(
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "   "},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when an allowlist is set")
+	require.NotNil(t, out.DenyReason, "deny reason must be populated")
+	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
+}
+
+func TestAllowlistEmptyListAllowsMissingModel(t *testing.T) {
+	// Without an allowlist there is nothing to enforce, so a missing model is
+	// still allowed — the fail-closed rule only applies when a list is set.
+	mw := New(Config{})
+	out, err := mw.Invoke(context.Background(), newInput())
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no allowlist must allow even without a model")
 }
 
 func TestPromptCaptureDisabledEmitsNoPrompt(t *testing.T) {

proxy/internal/middleware/builtin/llm_request_parser/guardrail_allowlist_test.go

@@ -0,0 +1,106 @@
+package llm_request_parser
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_guardrail"
+)
+
+// runParserGuardrail runs the request parser then the model-allowlist guardrail
+// in SlotOnRequest order, threading the parser's metadata into the guardrail the
+// same way the real chain does. It returns the guardrail decision so tests can
+// assert allowlist enforcement for URL/path-routed providers end to end.
+func runParserGuardrail(t *testing.T, url string, body []byte, allowlist []string) *middleware.Output {
+	t.Helper()
+	parser := newMiddleware(t)
+	parsed, err := parser.Invoke(context.Background(), &middleware.Input{
+		Slot: middleware.SlotOnRequest,
+		URL:  url,
+		Body: body,
+	})
+	require.NoError(t, err, "parser must not error")
+
+	guard := llm_guardrail.New(llm_guardrail.Config{ModelAllowlist: allowlist})
+	out, err := guard.Invoke(context.Background(), &middleware.Input{
+		Slot:     middleware.SlotOnRequest,
+		Metadata: parsed.Metadata,
+	})
+	require.NoError(t, err, "guardrail must not error")
+	require.NotNil(t, out, "guardrail must return an output")
+	return out
+}
+
+// TestModelAllowlist_URLRoutedProviders validates that the model allowlist is
+// enforced for providers whose model travels in the URL path (AWS Bedrock,
+// Google Vertex) rather than the JSON body. The "unknown action" case is the
+// regression guard for #6751: a Bedrock request shape the parser cannot map to a
+// model must fail closed under an allowlist instead of bypassing it.
+func TestModelAllowlist_URLRoutedProviders(t *testing.T) {
+	const bedrockBody = `{"anthropic_version":"bedrock-2023-05-31","messages":[{"role":"user","content":"hi"}]}`
+	const vertexBody = `{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"hi"}]}`
+
+	tests := []struct {
+		name      string
+		url       string
+		body      string
+		allowlist []string
+		decision  middleware.Decision
+		denyCode  string
+	}{
+		{
+			name:      "bedrock allowed model passes",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-haiku-4-5-v1:0/invoke",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			name:      "bedrock disallowed model denied",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-opus-4-8-v1:0/invoke",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "bedrock unknown action fails closed",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-opus-4-8-v1:0/some-future-action",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_unknown",
+		},
+		{
+			name:      "vertex disallowed model denied",
+			url:       "/v1/projects/p/locations/global/publishers/anthropic/models/claude-opus-4-8@20250101:rawPredict",
+			body:      vertexBody,
+			allowlist: []string{"claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "vertex allowed model passes",
+			url:       "/v1/projects/p/locations/global/publishers/anthropic/models/claude-haiku-4-5@20250101:rawPredict",
+			body:      vertexBody,
+			allowlist: []string{"claude-haiku-4-5"},
+			decision:  middleware.DecisionAllow,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			out := runParserGuardrail(t, tt.url, []byte(tt.body), tt.allowlist)
+			assert.Equal(t, tt.decision, out.Decision, "unexpected decision for %s", tt.name)
+			if tt.decision == middleware.DecisionDeny {
+				require.NotNil(t, out.DenyReason, "deny reason must be set for %s", tt.name)
+				assert.Equal(t, 403, out.DenyStatus, "deny status must be 403 for %s", tt.name)
+				assert.Equal(t, tt.denyCode, out.DenyReason.Code, "deny code for %s", tt.name)
+			}
+		})
+	}
+}

shared/relay/client/early_msg_buffer.go

@@ -10,7 +10,7 @@ import (
 
 const (
 	earlyMsgTTL      = 5 * time.Second
-	earlyMsgCapacity = 1000
+	earlyMsgCapacity = 10000
 )
 
 // earlyMsgBuffer buffers transport messages that arrive before the corresponding

shared/signal/client/client.go

@@ -51,10 +51,17 @@ type CredentialPayload struct {
 	WgListenPort    int
 	Credential      *Credential
 	RosenpassPubKey []byte
-	RosenpassAddr   string
-	RelaySrvAddress string
-	RelaySrvIP      netip.Addr
-	SessionID       []byte
+	// RosenpassPubKeyHash is the SHA256 of the sender's own RosenpassPubKey (empty
+	// when Rosenpass is disabled). RosenpassPubKey may be omitted when the peer has
+	// already acknowledged this hash. See RosenpassConfig in the proto.
+	RosenpassPubKeyHash []byte
+	// RosenpassPubKeyAck is the SHA256 of the remote peer's key the sender holds
+	// cached; empty means "send me the full key".
+	RosenpassPubKeyAck []byte
+	RosenpassAddr      string
+	RelaySrvAddress    string
+	RelaySrvIP         netip.Addr
+	SessionID          []byte
 }
 
 // UnMarshalCredential parses the credentials from the message and returns a Credential instance
@@ -78,8 +85,10 @@ func MarshalCredential(myKey wgtypes.Key, remoteKey string, p CredentialPayload)
 		WgListenPort:   uint32(p.WgListenPort),
 		NetBirdVersion: version.NetbirdVersion(),
 		RosenpassConfig: &proto.RosenpassConfig{
-			RosenpassPubKey:     p.RosenpassPubKey,
-			RosenpassServerAddr: p.RosenpassAddr,
+			RosenpassPubKey:                 p.RosenpassPubKey,
+			RosenpassServerAddr:             p.RosenpassAddr,
+			RosenpassPubKeyHash:             p.RosenpassPubKeyHash,
+			AcknowledgedRosenpassPubKeyHash: p.RosenpassPubKeyAck,
 		},
 		SessionId: p.SessionID,
 	}

shared/signal/proto/signalexchange.pb.go

@@ -1,7 +1,7 @@
 // Code generated by protoc-gen-go. DO NOT EDIT.
 // versions:
 // 	protoc-gen-go v1.26.0
-// 	protoc        v3.21.12
+// 	protoc        v6.33.1
 // source: signalexchange.proto
 
 package proto
@@ -399,6 +399,17 @@ type RosenpassConfig struct {
 	RosenpassPubKey []byte `protobuf:"bytes,1,opt,name=rosenpassPubKey,proto3" json:"rosenpassPubKey,omitempty"`
 	// rosenpassServerAddr is an IP:port of the rosenpass service
 	RosenpassServerAddr string `protobuf:"bytes,2,opt,name=rosenpassServerAddr,proto3" json:"rosenpassServerAddr,omitempty"`
+	// rosenpassPubKeyHash is the SHA256 of the sender's own rosenpassPubKey. It is
+	// always set when Rosenpass is enabled and lets the receiver detect (via a
+	// per-peer cache) whether it already holds the sender's full public key,
+	// avoiding re-sending the large key on every offer/answer.
+	RosenpassPubKeyHash []byte `protobuf:"bytes,3,opt,name=rosenpassPubKeyHash,proto3" json:"rosenpassPubKeyHash,omitempty"`
+	// acknowledgedRosenpassPubKeyHash is the SHA256 of the remote peer's rosenpassPubKey
+	// that the sender currently holds cached. When it matches the receiver's own key hash
+	// the receiver may omit its full rosenpassPubKey from the message. Empty means the
+	// sender does not have the remote key and needs it sent in full. Absent from peers
+	// that predate this field, which keeps them receiving the full key as before.
+	AcknowledgedRosenpassPubKeyHash []byte `protobuf:"bytes,4,opt,name=acknowledgedRosenpassPubKeyHash,proto3" json:"acknowledgedRosenpassPubKeyHash,omitempty"`
 }
 
 func (x *RosenpassConfig) Reset() {
@@ -447,6 +458,20 @@ func (x *RosenpassConfig) GetRosenpassServerAddr() string {
 	return ""
 }
 
+func (x *RosenpassConfig) GetRosenpassPubKeyHash() []byte {
+	if x != nil {
+		return x.RosenpassPubKeyHash
+	}
+	return nil
+}
+
+func (x *RosenpassConfig) GetAcknowledgedRosenpassPubKeyHash() []byte {
+	if x != nil {
+		return x.AcknowledgedRosenpassPubKeyHash
+	}
+	return nil
+}
+
 var File_signalexchange_proto protoreflect.FileDescriptor
 
 var file_signalexchange_proto_rawDesc = []byte{
@@ -506,27 +531,35 @@ var file_signalexchange_proto_rawDesc = []byte{
 	0x65, 0x72, 0x49, 0x50, 0x4a, 0x04, 0x08, 0x09, 0x10, 0x0a, 0x22, 0x2e, 0x0a, 0x04, 0x4d, 0x6f,
 	0x64, 0x65, 0x12, 0x1b, 0x0a, 0x06, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x18, 0x01, 0x20, 0x01,
 	0x28, 0x08, 0x48, 0x00, 0x52, 0x06, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x88, 0x01, 0x01, 0x42,
-	0x09, 0x0a, 0x07, 0x5f, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x22, 0x6d, 0x0a, 0x0f, 0x52, 0x6f,
-	0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x12, 0x28, 0x0a,
-	0x0f, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x50, 0x75, 0x62, 0x4b, 0x65, 0x79,
-	0x18, 0x01, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x0f, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73,
-	0x73, 0x50, 0x75, 0x62, 0x4b, 0x65, 0x79, 0x12, 0x30, 0x0a, 0x13, 0x72, 0x6f, 0x73, 0x65, 0x6e,
-	0x70, 0x61, 0x73, 0x73, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x41, 0x64, 0x64, 0x72, 0x18, 0x02,
-	0x20, 0x01, 0x28, 0x09, 0x52, 0x13, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x53,
-	0x65, 0x72, 0x76, 0x65, 0x72, 0x41, 0x64, 0x64, 0x72, 0x32, 0xb9, 0x01, 0x0a, 0x0e, 0x53, 0x69,
-	0x67, 0x6e, 0x61, 0x6c, 0x45, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x12, 0x4c, 0x0a, 0x04,
-	0x53, 0x65, 0x6e, 0x64, 0x12, 0x20, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63,
-	0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63, 0x72, 0x79, 0x70, 0x74, 0x65, 0x64, 0x4d,
-	0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x1a, 0x20, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x65,
-	0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63, 0x72, 0x79, 0x70, 0x74, 0x65,
-	0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x00, 0x12, 0x59, 0x0a, 0x0d, 0x43, 0x6f,
-	0x6e, 0x6e, 0x65, 0x63, 0x74, 0x53, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x12, 0x20, 0x2e, 0x73, 0x69,
-	0x67, 0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63,
-	0x72, 0x79, 0x70, 0x74, 0x65, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x1a, 0x20, 0x2e,
-	0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45,
-	0x6e, 0x63, 0x72, 0x79, 0x70, 0x74, 0x65, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22,
-	0x00, 0x28, 0x01, 0x30, 0x01, 0x42, 0x08, 0x5a, 0x06, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62,
-	0x06, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x33,
+	0x09, 0x0a, 0x07, 0x5f, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x22, 0xe9, 0x01, 0x0a, 0x0f, 0x52,
+	0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x12, 0x28,
+	0x0a, 0x0f, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x50, 0x75, 0x62, 0x4b, 0x65,
+	0x79, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x0f, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61,
+	0x73, 0x73, 0x50, 0x75, 0x62, 0x4b, 0x65, 0x79, 0x12, 0x30, 0x0a, 0x13, 0x72, 0x6f, 0x73, 0x65,
+	0x6e, 0x70, 0x61, 0x73, 0x73, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x41, 0x64, 0x64, 0x72, 0x18,
+	0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x13, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73,
+	0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x41, 0x64, 0x64, 0x72, 0x12, 0x30, 0x0a, 0x13, 0x72, 0x6f,
+	0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x50, 0x75, 0x62, 0x4b, 0x65, 0x79, 0x48, 0x61, 0x73,
+	0x68, 0x18, 0x03, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x13, 0x72, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61,
+	0x73, 0x73, 0x50, 0x75, 0x62, 0x4b, 0x65, 0x79, 0x48, 0x61, 0x73, 0x68, 0x12, 0x48, 0x0a, 0x1f,
+	0x61, 0x63, 0x6b, 0x6e, 0x6f, 0x77, 0x6c, 0x65, 0x64, 0x67, 0x65, 0x64, 0x52, 0x6f, 0x73, 0x65,
+	0x6e, 0x70, 0x61, 0x73, 0x73, 0x50, 0x75, 0x62, 0x4b, 0x65, 0x79, 0x48, 0x61, 0x73, 0x68, 0x18,
+	0x04, 0x20, 0x01, 0x28, 0x0c, 0x52, 0x1f, 0x61, 0x63, 0x6b, 0x6e, 0x6f, 0x77, 0x6c, 0x65, 0x64,
+	0x67, 0x65, 0x64, 0x52, 0x6f, 0x73, 0x65, 0x6e, 0x70, 0x61, 0x73, 0x73, 0x50, 0x75, 0x62, 0x4b,
+	0x65, 0x79, 0x48, 0x61, 0x73, 0x68, 0x32, 0xb9, 0x01, 0x0a, 0x0e, 0x53, 0x69, 0x67, 0x6e, 0x61,
+	0x6c, 0x45, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x12, 0x4c, 0x0a, 0x04, 0x53, 0x65, 0x6e,
+	0x64, 0x12, 0x20, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e,
+	0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63, 0x72, 0x79, 0x70, 0x74, 0x65, 0x64, 0x4d, 0x65, 0x73, 0x73,
+	0x61, 0x67, 0x65, 0x1a, 0x20, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63, 0x68,
+	0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63, 0x72, 0x79, 0x70, 0x74, 0x65, 0x64, 0x4d, 0x65,
+	0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x00, 0x12, 0x59, 0x0a, 0x0d, 0x43, 0x6f, 0x6e, 0x6e, 0x65,
+	0x63, 0x74, 0x53, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x12, 0x20, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61,
+	0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63, 0x72, 0x79, 0x70,
+	0x74, 0x65, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x1a, 0x20, 0x2e, 0x73, 0x69, 0x67,
+	0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x45, 0x6e, 0x63, 0x72,
+	0x79, 0x70, 0x74, 0x65, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x00, 0x28, 0x01,
+	0x30, 0x01, 0x42, 0x08, 0x5a, 0x06, 0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x62, 0x06, 0x70, 0x72,
+	0x6f, 0x74, 0x6f, 0x33,
 }
 
 var (

shared/signal/proto/signalexchange.proto

@@ -86,4 +86,15 @@ message RosenpassConfig {
   bytes rosenpassPubKey = 1;
   // rosenpassServerAddr is an IP:port of the rosenpass service
   string rosenpassServerAddr = 2;
+  // rosenpassPubKeyHash is the SHA256 of the sender's own rosenpassPubKey. It is
+  // always set when Rosenpass is enabled and lets the receiver detect (via a
+  // per-peer cache) whether it already holds the sender's full public key,
+  // avoiding re-sending the large key on every offer/answer.
+  bytes rosenpassPubKeyHash = 3;
+  // acknowledgedRosenpassPubKeyHash is the SHA256 of the remote peer's rosenpassPubKey
+  // that the sender currently holds cached. When it matches the receiver's own key hash
+  // the receiver may omit its full rosenpassPubKey from the message. Empty means the
+  // sender does not have the remote key and needs it sent in full. Absent from peers
+  // that predate this field, which keeps them receiving the full key as before.
+  bytes acknowledgedRosenpassPubKeyHash = 4;
 }