[client] pqkem: gate controller re-offer to kick the KEM exactly once

When the controller receives the responder's (KEM-less) offer it replies with
its own KEM offer instead of answering, so the only transaction that brings the
tunnel up is the one that also carries the PSK. Guard that reply with
ShouldSendBootstrapOffer so it fires only when no exchange is in flight: without
it, every responder offer triggered another offer (an offer-per-offer runaway).
The whole behaviour is isolated to the KEM path (config.PQ != nil); non-PQ
connections answer as before.

.github/workflows/pr-title-check.yml

@@ -16,6 +16,8 @@ jobs:
             const allowedTags = [
               'management',
               'client',
+              'android',
+              'ios',
               'signal',
               'proxy',
               'relay',

.goreleaser_ui.yaml

@@ -93,7 +93,9 @@ nfpms:
       - src: client/ui/build/appicon.png
         dst: /usr/share/pixmaps/netbird.png
     dependencies:
-      - netbird
+      - netbird (>= 0.75.0)
+      - libgtk-4-1 (>= 4.14)
+      - libwebkitgtk-6.0-4
 
   - maintainer: Netbird <dev@netbird.io>
     description: Netbird client UI.
@@ -114,7 +116,9 @@ nfpms:
       - src: client/ui/build/appicon.png
         dst: /usr/share/pixmaps/netbird.png
     dependencies:
-      - netbird
+      - netbird >= 0.75.0
+      - (gtk4 >= 4.14 or libgtk-4-1 >= 4.14)
+      - (webkitgtk6.0 or libwebkitgtk-6_0-4)
 
     rpm:
       signature:

client/android/client.go

@@ -57,6 +57,12 @@ type DnsReadyListener interface {
 	dns.ReadyListener
 }
 
+// TunSettings is a snapshot of the settings the TUN device is rebuilt with
+type TunSettings struct {
+	Routes        string
+	SearchDomains string
+}
+
 func init() {
 	formatter.SetLogcatFormatter(log.StandardLogger())
 }
@@ -76,6 +82,8 @@ type Client struct {
 	connectClient *internal.ConnectClient
 	config        *profilemanager.Config
 	cacheDir      string
+	// Identifies the running profile for the SSO login hint; see profile_state.go.
+	cfgPath string
 
 	stateChangeMu    sync.Mutex
 	stateChangeSubID string
@@ -96,11 +104,12 @@ type Client struct {
 	extendCancel context.CancelFunc
 }
 
-func (c *Client) setState(cfg *profilemanager.Config, cacheDir string, cc *internal.ConnectClient) {
+func (c *Client) setState(cfg *profilemanager.Config, cacheDir string, cfgPath string, cc *internal.ConnectClient) {
 	c.stateMu.Lock()
 	defer c.stateMu.Unlock()
 	c.config = cfg
 	c.cacheDir = cacheDir
+	c.cfgPath = cfgPath
 	c.connectClient = cc
 }
 
@@ -110,6 +119,16 @@ func (c *Client) stateSnapshot() (*profilemanager.Config, string, *internal.Conn
 	return c.config, c.cacheDir, c.connectClient
 }
 
+// authSnapshot returns the config together with the path it was loaded from, in
+// one lock: the path identifies the profile whose account email backs the login
+// hint, so reading it separately could pair one profile's config with another's
+// hint when a profile switch lands in between.
+func (c *Client) authSnapshot() (*profilemanager.Config, string, *internal.ConnectClient) {
+	c.stateMu.RLock()
+	defer c.stateMu.RUnlock()
+	return c.config, c.cfgPath, c.connectClient
+}
+
 func (c *Client) getConnectClient() *internal.ConnectClient {
 	c.stateMu.RLock()
 	defer c.stateMu.RUnlock()
@@ -162,7 +181,7 @@ func (c *Client) Run(platformFiles PlatformFiles, urlOpener URLOpener, isAndroid
 	defer c.ctxCancel()
 	c.ctxCancelLock.Unlock()
 
-	auth := NewAuthWithConfig(ctx, cfg)
+	auth := NewAuthWithConfig(ctx, cfg, cfgFile)
 	err = auth.login(urlOpener, isAndroidTV)
 	if err != nil {
 		return err
@@ -170,7 +189,7 @@ func (c *Client) Run(platformFiles PlatformFiles, urlOpener URLOpener, isAndroid
 	// todo do not throw error in case of cancelled context
 	ctx = internal.CtxInitState(ctx)
 	connectClient := internal.NewConnectClient(ctx, cfg, c.recorder)
-	c.setState(cfg, cacheDir, connectClient)
+	c.setState(cfg, cacheDir, cfgFile, connectClient)
 	// This path runs the interactive SSO flow, so reaching here means the peer
 	// is authenticated again — release the latch Status() reports from. Clear
 	// only once the fresh connect client is installed: until then Status()
@@ -211,7 +230,7 @@ func (c *Client) RunWithoutLogin(platformFiles PlatformFiles, dns *DNSList, dnsR
 	// todo do not throw error in case of cancelled context
 	ctx = internal.CtxInitState(ctx)
 	connectClient := internal.NewConnectClient(ctx, cfg, c.recorder)
-	c.setState(cfg, cacheDir, connectClient)
+	c.setState(cfg, cacheDir, cfgFile, connectClient)
 	return connectClient.RunOnAndroid(c.tunAdapter, c.iFaceDiscover, c.networkChangeListener, slices.Clone(dns.items), dnsReadyListener, stateFile, cacheDir)
 }
 
@@ -240,6 +259,24 @@ func (c *Client) RenewTun(fd int) error {
 	return e.RenewTun(fd)
 }
 
+func (c *Client) GetTunSettings() (*TunSettings, error) {
+	cc := c.getConnectClient()
+	if cc == nil {
+		return nil, fmt.Errorf("engine not running")
+	}
+
+	e := cc.Engine()
+	if e == nil {
+		return nil, fmt.Errorf("engine not initialized")
+	}
+
+	routes, searchDomains := e.TunSettings()
+	return &TunSettings{
+		Routes:        strings.Join(routes, ";"),
+		SearchDomains: strings.Join(searchDomains, ";"),
+	}, nil
+}
+
 // DebugBundle generates a debug bundle, uploads it, and returns the upload key.
 // It works both with and without a running engine.
 func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool) (string, error) {

client/android/login.go

@@ -4,6 +4,8 @@ import (
 	"context"
 	"fmt"
 
+	log "github.com/sirupsen/logrus"
+
 	"github.com/netbirdio/netbird/client/internal/auth"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/system"
@@ -36,12 +38,20 @@ type Auth struct {
 }
 
 // NewAuth instantiate Auth struct and validate the management URL
+//
+// The configuration at cfgPath is reused when one is already there, and only created when it is
+// not. Building a fresh in-memory config unconditionally gives the client a new WireGuard key on
+// every call: the peer registers under that key, the key is written out, and any peer registered by
+// an earlier call is orphaned on the server. It also breaks a client that enrols and then runs from
+// the persisted config, because the identity it registered is not the one it runs with — the
+// management stream rejects it with "no peer auth method provided".
 func NewAuth(cfgPath string, mgmURL string) (*Auth, error) {
 	inputCfg := profilemanager.ConfigInput{
+		ConfigPath:    cfgPath,
 		ManagementURL: mgmURL,
 	}
 
-	cfg, err := profilemanager.CreateInMemoryConfig(inputCfg)
+	cfg, err := profilemanager.UpdateOrCreateConfig(inputCfg)
 	if err != nil {
 		return nil, err
 	}
@@ -53,11 +63,14 @@ func NewAuth(cfgPath string, mgmURL string) (*Auth, error) {
 	}, nil
 }
 
-// NewAuthWithConfig instantiate Auth based on existing config
-func NewAuthWithConfig(ctx context.Context, config *profilemanager.Config) *Auth {
+// NewAuthWithConfig instantiate Auth based on existing config. cfgPath is the
+// file the config was loaded from; it identifies the profile whose account email
+// backs the login_hint.
+func NewAuthWithConfig(ctx context.Context, config *profilemanager.Config, cfgPath string) *Auth {
 	return &Auth{
-		ctx:    ctx,
-		config: config,
+		ctx:     ctx,
+		config:  config,
+		cfgPath: cfgPath,
 	}
 }
 
@@ -150,12 +163,14 @@ func (a *Auth) login(urlOpener URLOpener, isAndroidTV bool) error {
 	}
 
 	jwtToken := ""
+	email := ""
 	if needsLogin {
 		tokenInfo, err := a.foregroundGetTokenInfo(authClient, urlOpener, isAndroidTV)
 		if err != nil {
 			return fmt.Errorf("interactive sso login failed: %v", err)
 		}
 		jwtToken = tokenInfo.GetTokenToUse()
+		email = tokenInfo.Email
 	}
 
 	err, _ = authClient.Login(a.ctx, "", jwtToken)
@@ -163,17 +178,42 @@ func (a *Auth) login(urlOpener URLOpener, isAndroidTV bool) error {
 		return fmt.Errorf("login failed: %v", err)
 	}
 
+	// Stored after Login, not before: a rejected token must not leave a hint
+	// pointing at an account that cannot be used.
+	if email != "" && a.cfgPath != "" {
+		if err := writeProfileEmail(a.cfgPath, email); err != nil {
+			log.Warnf("failed to store profile account email: %v", err)
+		}
+	}
+
 	go urlOpener.OnLoginSuccess()
 
 	return nil
 }
 
+// loginHintSetter is implemented by both concrete flows (PKCE and device code)
+// but absent from the OAuthFlow interface, hence the assertion below — the same
+// way internal/auth wires it in authenticateWithPKCEFlow.
+type loginHintSetter interface {
+	SetLoginHint(hint string)
+}
+
 func (a *Auth) foregroundGetTokenInfo(authClient *auth.Auth, urlOpener URLOpener, isAndroidTV bool) (*auth.TokenInfo, error) {
 	oAuthFlow, err := authClient.GetOAuthFlow(a.ctx, isAndroidTV)
 	if err != nil {
 		return nil, fmt.Errorf("failed to get OAuth flow: %v", err)
 	}
 
+	// An empty hint is deliberate, not a fallback: a fresh or logged-out profile
+	// leaves the choice to the IdP, which is how accounts get switched.
+	if a.cfgPath != "" {
+		if hint := readProfileEmail(a.cfgPath); hint != "" {
+			if setter, ok := oAuthFlow.(loginHintSetter); ok {
+				setter.SetLoginHint(hint)
+			}
+		}
+	}
+
 	flowInfo, err := oAuthFlow.RequestAuthInfo(context.TODO())
 	if err != nil {
 		return nil, fmt.Errorf("getting a request OAuth flow info failed: %v", err)

client/android/login_test.go

@@ -0,0 +1,51 @@
+package android
+
+import (
+	"path/filepath"
+	"testing"
+)
+
+// NewAuth must reuse the configuration already at cfgPath rather than building a fresh one.
+//
+// Creating a new in-memory config on every call gives the client a new WireGuard private key each
+// time. The peer registers under that key and the key is written out, so a peer registered by an
+// earlier call is orphaned on the server — a client that enrols twice leaves two entries and owns
+// neither. It also breaks enrol-then-run: RunWithoutLogin reloads the configuration from disk, so
+// the identity that registered is not the identity that runs, and the management stream rejects it
+// with "no peer auth method provided, please use a setup key or interactive SSO login".
+func TestNewAuth_ReusesPersistedIdentity(t *testing.T) {
+	cfgPath := filepath.Join(t.TempDir(), "config.json")
+
+	first, err := NewAuth(cfgPath, "https://api.example.com:443")
+	if err != nil {
+		t.Fatalf("first NewAuth: %v", err)
+	}
+	if first.config.PrivateKey == "" {
+		t.Fatal("first NewAuth produced no private key")
+	}
+
+	second, err := NewAuth(cfgPath, "https://api.example.com:443")
+	if err != nil {
+		t.Fatalf("second NewAuth: %v", err)
+	}
+
+	if second.config.PrivateKey != first.config.PrivateKey {
+		t.Errorf("private key changed between calls: a second enrolment would orphan the peer registered by the first")
+	}
+}
+
+// A missing configuration is still created, so a first enrolment works unchanged.
+func TestNewAuth_CreatesConfigWhenAbsent(t *testing.T) {
+	cfgPath := filepath.Join(t.TempDir(), "config.json")
+
+	auth, err := NewAuth(cfgPath, "https://api.example.com:443")
+	if err != nil {
+		t.Fatalf("NewAuth: %v", err)
+	}
+	if auth.config == nil || auth.config.PrivateKey == "" {
+		t.Fatal("NewAuth did not create a usable configuration")
+	}
+	if auth.cfgPath != cfgPath {
+		t.Errorf("cfgPath = %q, want %q", auth.cfgPath, cfgPath)
+	}
+}

client/android/profile_manager.go

@@ -13,18 +13,17 @@ import (
 )
 
 const (
-	// Android-specific config filename (different from desktop default.json)
-	defaultConfigFilename = "netbird.cfg"
-	// Subdirectory for non-default profiles (must match Java Preferences.java)
-	profilesSubdir = "profiles"
 	// Android uses a single user context per app (non-empty username required by ServiceManager)
 	androidUsername = "android"
 )
 
 // Profile represents a profile for gomobile
 type Profile struct {
-	ID       string
-	Name     string
+	ID   string
+	Name string
+	// Email is the account this profile last logged in with, "" if it never
+	// completed an SSO login or was logged out. See profile_state.go.
+	Email    string
 	IsActive bool
 }
 
@@ -101,6 +100,7 @@ func (pm *ProfileManager) ListProfiles() (*ProfileArray, error) {
 		profiles = append(profiles, &Profile{
 			ID:       p.ID.String(),
 			Name:     p.Name,
+			Email:    pm.profileEmail(p.ID.String()),
 			IsActive: p.IsActive,
 		})
 	}
@@ -123,7 +123,22 @@ func (pm *ProfileManager) GetActiveProfile() (*Profile, error) {
 	if err != nil {
 		return nil, fmt.Errorf("failed to resolve active profile %q: %w", activeState.ID, err)
 	}
-	return &Profile{ID: prof.ID.String(), Name: prof.Name, IsActive: true}, nil
+	return &Profile{
+		ID:       prof.ID.String(),
+		Name:     prof.Name,
+		Email:    pm.profileEmail(prof.ID.String()),
+		IsActive: true,
+	}, nil
+}
+
+// profileEmail returns the account email recorded for a profile. Display-only, so
+// an unresolvable path degrades to "" rather than an error.
+func (pm *ProfileManager) profileEmail(id string) string {
+	configPath, err := pm.getProfileConfigPath(id)
+	if err != nil {
+		return ""
+	}
+	return readProfileEmail(configPath)
 }
 
 // SwitchProfile switches to a different profile
@@ -185,6 +200,11 @@ func (pm *ProfileManager) LogoutProfile(id string) error {
 		return fmt.Errorf("failed to save config: %w", err)
 	}
 
+	// Not fatal: a stale hint costs an account switch, not the logout itself.
+	if err := removeProfileEmail(configPath); err != nil {
+		log.Warnf("failed to clear stored account email for profile %s: %v", id, err)
+	}
+
 	log.Infof("logged out from profile: %s", id)
 	return nil
 }

client/android/profile_state.go

@@ -0,0 +1,108 @@
+package android
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	"github.com/netbirdio/netbird/util"
+)
+
+const (
+	// Android-specific config filename (different from desktop default.json)
+	defaultConfigFilename = "netbird.cfg"
+	// Subdirectory for non-default profiles (must match Java Preferences.java)
+	profilesSubdir = "profiles"
+	// profileAccountSuffix names the file holding the profile's account email.
+	// Deliberately not ".state.json", which desktop uses for the same data:
+	// there the email and the engine's state manager live in different
+	// directories, but on Android both resolve under files/, so sharing the name
+	// would have the two overwrite each other — the state manager rewrites the
+	// whole file from its own keys (see statemanager.Manager.PersistState), and
+	// this package's writer does the same in reverse.
+	profileAccountSuffix = ".account.json"
+)
+
+// profileAccountPathFor derives the account file path from a profile's config
+// path: netbird.cfg -> netbird.account.json, <id>.json -> <id>.account.json.
+//
+// Deriving from the config path rather than resolving the active profile keeps
+// the write on the profile the login actually ran for: Auth.login runs in a
+// goroutine, so the active profile can change under a flow already in flight.
+func profileAccountPathFor(configPath string) (string, error) {
+	if configPath == "" {
+		return "", fmt.Errorf("empty config path")
+	}
+
+	base := filepath.Base(configPath)
+	stem := strings.TrimSuffix(base, filepath.Ext(base))
+	if stem == "" || stem == "." {
+		return "", fmt.Errorf("config path %q has no filename stem", configPath)
+	}
+
+	return filepath.Join(filepath.Dir(configPath), stem+profileAccountSuffix), nil
+}
+
+// readProfileEmail returns the account email stored for the profile whose config
+// lives at configPath. A missing or unreadable file yields "", which leaves the
+// account choice to the IdP.
+func readProfileEmail(configPath string) string {
+	accountPath, err := profileAccountPathFor(configPath)
+	if err != nil {
+		log.Debugf("no profile account path for login hint: %v", err)
+		return ""
+	}
+
+	var state profilemanager.ProfileState
+	if _, err := util.ReadJson(accountPath, &state); err != nil {
+		if !os.IsNotExist(err) {
+			log.Debugf("failed to read profile account for login hint: %v", err)
+		}
+		return ""
+	}
+
+	return state.Email
+}
+
+// writeProfileEmail records the account email for the profile whose config lives
+// at configPath, so later logins can pass it as an OIDC login_hint. An empty
+// email is ignored rather than blanking what is already stored.
+func writeProfileEmail(configPath string, email string) error {
+	if email == "" {
+		return nil
+	}
+
+	accountPath, err := profileAccountPathFor(configPath)
+	if err != nil {
+		return fmt.Errorf("resolve profile account path: %w", err)
+	}
+
+	state := profilemanager.ProfileState{Email: email}
+	if err := util.WriteJsonWithRestrictedPermission(context.Background(), accountPath, state); err != nil {
+		return fmt.Errorf("write profile account: %w", err)
+	}
+
+	return nil
+}
+
+// removeProfileEmail drops the stored account email. Called on logout: while the
+// email is on disk it goes out as a login_hint, which would steer the next login
+// straight back into the account just logged out of. Mirrors the desktop UI's
+// RemoveProfileState call.
+func removeProfileEmail(configPath string) error {
+	accountPath, err := profileAccountPathFor(configPath)
+	if err != nil {
+		return fmt.Errorf("resolve profile account path: %w", err)
+	}
+
+	if err := os.Remove(accountPath); err != nil && !os.IsNotExist(err) {
+		return fmt.Errorf("remove profile account: %w", err)
+	}
+
+	return nil
+}

client/android/profile_state_test.go

@@ -0,0 +1,161 @@
+package android
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestProfileAccountPathFor(t *testing.T) {
+	tests := []struct {
+		name       string
+		configPath string
+		want       string
+		wantErr    bool
+	}{
+		{
+			name:       "default profile",
+			configPath: "/data/data/io.netbird.client/files/netbird.cfg",
+			want:       filepath.FromSlash("/data/data/io.netbird.client/files/netbird.account.json"),
+		},
+		{
+			name:       "id profile",
+			configPath: "/data/data/io.netbird.client/files/profiles/4c5f5c8198c3989cffb5b5394f5a7ae0.json",
+			want:       filepath.FromSlash("/data/data/io.netbird.client/files/profiles/4c5f5c8198c3989cffb5b5394f5a7ae0.account.json"),
+		},
+		{
+			name:       "legacy name-keyed profile is handled the same way",
+			configPath: "/data/data/io.netbird.client/files/profiles/work.json",
+			want:       filepath.FromSlash("/data/data/io.netbird.client/files/profiles/work.account.json"),
+		},
+		{
+			name:       "empty path is rejected",
+			configPath: "",
+			wantErr:    true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := profileAccountPathFor(tt.configPath)
+			if tt.wantErr {
+				if err == nil {
+					t.Fatalf("expected an error, got path %q", got)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+			if got != tt.want {
+				t.Errorf("got %q, want %q", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestProfileAccountPathForDefaultDoesNotCollide(t *testing.T) {
+	root := "/data/data/io.netbird.client/files"
+
+	defaultAccount, err := profileAccountPathFor(filepath.Join(root, defaultConfigFilename))
+	if err != nil {
+		t.Fatalf("default profile: %v", err)
+	}
+
+	idAccount, err := profileAccountPathFor(filepath.Join(root, profilesSubdir, "abc123.json"))
+	if err != nil {
+		t.Fatalf("id profile: %v", err)
+	}
+
+	if defaultAccount == idAccount {
+		t.Fatalf("default and id profile share an account file: %q", defaultAccount)
+	}
+}
+
+// The account file must never land on the engine state file: on Android both
+// resolve under files/, and the state manager rewrites the whole file from its
+// own keys, so sharing a path would have the two overwrite each other. The
+// expected names here mirror ProfileManager.GetStateFilePath.
+func TestProfileAccountPathAvoidsEngineStateFile(t *testing.T) {
+	root := "/data/data/io.netbird.client/files"
+
+	cases := []struct {
+		configPath  string
+		engineState string
+	}{
+		{
+			configPath:  filepath.Join(root, defaultConfigFilename),
+			engineState: filepath.Join(root, "state.json"),
+		},
+		{
+			configPath:  filepath.Join(root, profilesSubdir, "abc123.json"),
+			engineState: filepath.Join(root, profilesSubdir, "abc123.state.json"),
+		},
+	}
+
+	for _, c := range cases {
+		account, err := profileAccountPathFor(c.configPath)
+		if err != nil {
+			t.Fatalf("%s: %v", c.configPath, err)
+		}
+		if account == c.engineState {
+			t.Errorf("account file collides with the engine state file: %q", account)
+		}
+	}
+}
+
+func TestWriteThenReadProfileEmail(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "profiles", "abc123.json")
+	if err := ensureDirFor(t, configPath); err != nil {
+		t.Fatalf("prepare dir: %v", err)
+	}
+
+	if got := readProfileEmail(configPath); got != "" {
+		t.Errorf("expected no email before a login, got %q", got)
+	}
+
+	const email = "user@example.com"
+	if err := writeProfileEmail(configPath, email); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+
+	if got := readProfileEmail(configPath); got != email {
+		t.Errorf("got %q, want %q", got, email)
+	}
+
+	if err := removeProfileEmail(configPath); err != nil {
+		t.Fatalf("remove: %v", err)
+	}
+	if got := readProfileEmail(configPath); got != "" {
+		t.Errorf("expected no email after logout, got %q", got)
+	}
+
+	// Logout may run on a never-logged-in profile, so a second remove must pass.
+	if err := removeProfileEmail(configPath); err != nil {
+		t.Fatalf("second remove should be a no-op: %v", err)
+	}
+}
+
+func TestWriteProfileEmailIgnoresEmpty(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "profiles", "abc123.json")
+	if err := ensureDirFor(t, configPath); err != nil {
+		t.Fatalf("prepare dir: %v", err)
+	}
+
+	const email = "user@example.com"
+	if err := writeProfileEmail(configPath, email); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+	if err := writeProfileEmail(configPath, ""); err != nil {
+		t.Fatalf("write empty: %v", err)
+	}
+
+	if got := readProfileEmail(configPath); got != email {
+		t.Errorf("empty write clobbered the stored email: got %q, want %q", got, email)
+	}
+}
+
+func ensureDirFor(t *testing.T, path string) error {
+	t.Helper()
+	return os.MkdirAll(filepath.Dir(path), 0o700)
+}

client/android/session.go

@@ -278,7 +278,7 @@ func (c *Client) endExtend() {
 }
 
 func (c *Client) extendAuthSession(ctx context.Context, urlOpener URLOpener, isAndroidTV bool) error {
-	cfg, _, cc := c.stateSnapshot()
+	cfg, cfgPath, cc := c.authSnapshot()
 	if cfg == nil || cc == nil {
 		return fmt.Errorf("engine is not running")
 	}
@@ -293,7 +293,10 @@ func (c *Client) extendAuthSession(ctx context.Context, urlOpener URLOpener, isA
 	}
 	defer authClient.Close()
 
-	a := &Auth{ctx: ctx, config: cfg}
+	// Passing the config path makes the flow pick up the login_hint: an extend
+	// renews the session of the account already signed in, so it must not stop to
+	// offer a choice.
+	a := NewAuthWithConfig(ctx, cfg, cfgPath)
 	tokenInfo, err := a.foregroundGetTokenInfo(authClient, urlOpener, isAndroidTV)
 	if err != nil {
 		return fmt.Errorf("interactive sso login failed: %v", err)

client/internal/connect.go

@@ -113,11 +113,14 @@ func (c *ConnectClient) RunOnAndroid(
 	stateFilePath string,
 	cacheDir string,
 ) error {
+	notifier := tunnelnotifier.New(networkChangeListener, nil)
+	defer notifier.Close()
+
 	// in case of non Android os these variables will be nil
 	mobileDependency := MobileDependency{
 		TunAdapter:            tunAdapter,
 		IFaceDiscover:         iFaceDiscover,
-		NetworkChangeListener: networkChangeListener,
+		NetworkChangeListener: notifier,
 		HostDNSAddresses:      dnsAddresses,
 		DnsReadyListener:      dnsReadyListener,
 		StateFilePath:         stateFilePath,

client/internal/dns/notifier.go

@@ -51,7 +51,5 @@ func (n *notifier) notify() {
 		return
 	}
 
-	go func(l listener.NetworkChangeListener) {
-		l.OnNetworkChanged("")
-	}(n.listener)
+	n.listener.OnNetworkChanged("")
 }

client/internal/dns/server.go

@@ -252,7 +252,7 @@ func NewDefaultServerPermanentUpstream(
 	ds.hostsDNSHolder.set(hostsDnsList)
 	ds.permanent = true
 	ds.currentConfig = dnsConfigToHostDNSConfig(config, ds.service.RuntimeIP(), ds.service.RuntimePort())
-	ds.searchDomainNotifier = newNotifier(ds.SearchDomains())
+	ds.searchDomainNotifier = newNotifier(ds.searchDomains())
 	ds.searchDomainNotifier.setListener(listener)
 	setServerDns(ds)
 	return ds
@@ -602,6 +602,12 @@ func (s *DefaultServer) UpdateDNSServer(serial uint64, update nbdns.Config) erro
 }
 
 func (s *DefaultServer) SearchDomains() []string {
+	s.mux.Lock()
+	defer s.mux.Unlock()
+	return s.searchDomains()
+}
+
+func (s *DefaultServer) searchDomains() []string {
 	var searchDomains []string
 
 	for _, dConf := range s.currentConfig.Domains {
@@ -686,7 +692,7 @@ func (s *DefaultServer) applyConfiguration(update nbdns.Config) error {
 	}()
 
 	if s.searchDomainNotifier != nil {
-		s.searchDomainNotifier.onNewSearchDomains(s.SearchDomains())
+		s.searchDomainNotifier.onNewSearchDomains(s.searchDomains())
 	}
 
 	s.updateNSGroupStates(update.NameServerGroups)

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"
@@ -197,6 +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
 
@@ -555,7 +560,11 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 	publicKey := e.config.WgPrivateKey.PublicKey()
 	e.flowManager = netflow.NewManager(e.wgInterface, publicKey[:], e.statusRecorder)
 
-	if e.config.RosenpassEnabled {
+	// Rosenpass and ML-KEM are mutually exclusive. ML-KEM (NB_ENABLE_PQ_MLKEM) takes precedence
+	if e.config.RosenpassEnabled && pqkem.Enabled() {
+		log.Warnf("rosenpass and ML-KEM post-quantum are mutually exclusive; ML-KEM is enabled, so rosenpass is disabled")
+	}
+	if e.config.RosenpassEnabled && !pqkem.Enabled() {
 		log.Infof("rosenpass is enabled")
 		if e.config.RosenpassPermissive {
 			log.Infof("running rosenpass in permissive mode")
@@ -572,12 +581,7 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 	}
 	e.stateManager.Start()
 
-	initialRoutes, dnsConfig, dnsFeatureFlag, err := e.readInitialSettings()
-	if err != nil {
-		return fmt.Errorf("read initial settings: %w", err)
-	}
-
-	dnsServer, err := e.newDnsServer(dnsConfig)
+	dnsServer, err := e.newDnsServer()
 	if err != nil {
 		return fmt.Errorf("create dns server: %w", err)
 	}
@@ -595,10 +599,8 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 		WGInterface:         e.wgInterface,
 		StatusRecorder:      e.statusRecorder,
 		RelayManager:        e.relayManager,
-		InitialRoutes:       initialRoutes,
 		StateManager:        e.stateManager,
 		DNSServer:           dnsServer,
-		DNSFeatureFlag:      dnsFeatureFlag,
 		PeerStore:           e.peerStore,
 		DisableClientRoutes: e.config.DisableClientRoutes,
 		DisableServerRoutes: e.config.DisableServerRoutes,
@@ -651,6 +653,30 @@ 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)
@@ -914,6 +940,10 @@ 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)
@@ -1900,6 +1930,10 @@ 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,
@@ -2083,6 +2117,10 @@ 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 {
@@ -2102,42 +2140,6 @@ func (e *Engine) close() {
 	}
 }
 
-func (e *Engine) readInitialSettings() ([]*route.Route, *nbdns.Config, bool, error) {
-	if runtime.GOOS != "android" {
-		// nolint:nilnil
-		return nil, nil, false, nil
-	}
-
-	info := system.GetInfo(e.ctx)
-	info.SetFlags(
-		e.config.RosenpassEnabled,
-		e.config.RosenpassPermissive,
-		&e.config.ServerSSHAllowed,
-		e.config.DisableClientRoutes,
-		e.config.DisableServerRoutes,
-		e.config.DisableDNS,
-		e.config.DisableFirewall,
-		e.config.BlockLANAccess,
-		e.config.BlockInbound,
-		e.config.DisableIPv6,
-		e.config.SyncMessageVersion,
-		e.config.EnableSSHRoot,
-		e.config.EnableSSHSFTP,
-		e.config.EnableSSHLocalPortForwarding,
-		e.config.EnableSSHRemotePortForwarding,
-		e.config.DisableSSHAuth,
-	)
-
-	netMap, err := e.mgmClient.GetNetworkMap(info)
-	if err != nil {
-		return nil, nil, false, err
-	}
-	routes := toRoutes(netMap.GetRoutes())
-	dnsCfg := toDNSConfig(netMap.GetDNSConfig(), e.wgInterface.Address())
-	dnsFeatureFlag := toDNSFeatureFlag(netMap)
-	return routes, &dnsCfg, dnsFeatureFlag, nil
-}
-
 func (e *Engine) newWgIface() (*iface.WGIface, error) {
 	transportNet, err := e.newStdNet()
 	if err != nil {
@@ -2172,7 +2174,7 @@ func (e *Engine) newWgIface() (*iface.WGIface, error) {
 func (e *Engine) wgInterfaceCreate() (err error) {
 	switch runtime.GOOS {
 	case "android":
-		err = e.wgInterface.CreateOnAndroid(e.routeManager.InitialRouteRange(), e.dnsServer.DnsIP().String(), e.dnsServer.SearchDomains())
+		err = e.wgInterface.CreateOnAndroid(e.routeManager.CurrentRouteRange(), e.dnsServer.DnsIP().String(), e.dnsServer.SearchDomains())
 	case "ios":
 		e.mobileDep.NetworkChangeListener.SetInterfaceIP(e.config.WgAddr.String())
 		if e.config.WgAddr.HasIPv6() {
@@ -2185,7 +2187,7 @@ func (e *Engine) wgInterfaceCreate() (err error) {
 	return err
 }
 
-func (e *Engine) newDnsServer(dnsConfig *nbdns.Config) (dns.Server, error) {
+func (e *Engine) newDnsServer() (dns.Server, error) {
 	// due to tests where we are using a mocked version of the DNS server
 	if e.dnsServer != nil {
 		return e.dnsServer, nil
@@ -2197,7 +2199,7 @@ func (e *Engine) newDnsServer(dnsConfig *nbdns.Config) (dns.Server, error) {
 			e.ctx,
 			e.wgInterface,
 			e.mobileDep.HostDNSAddresses,
-			*dnsConfig,
+			nbdns.Config{},
 			e.mobileDep.NetworkChangeListener,
 			e.statusRecorder,
 			e.config.DisableDNS,
@@ -2895,6 +2897,8 @@ func convertToOfferAnswer(msg *sProto.Message) (*peer.OfferAnswer, error) {
 		Version:         msg.GetBody().GetNetBirdVersion(),
 		RosenpassPubKey: rosenpassPubKey,
 		RosenpassAddr:   rosenpassAddr,
+		MlkemPayload:    msg.GetBody().GetMlkemPayload(),
+		MlkemPort:       int(msg.GetBody().GetMlkemPort()),
 		RelaySrvAddress: msg.GetBody().GetRelayServerAddress(),
 		RelaySrvIP:      relayIP,
 		SessionID:       sessionID,

client/internal/engine_tunsettings.go

@@ -0,0 +1,20 @@
+package internal
+
+func (e *Engine) TunSettings() ([]string, []string) {
+	e.syncMsgMux.Lock()
+	routeManager := e.routeManager
+	dnsServer := e.dnsServer
+	e.syncMsgMux.Unlock()
+
+	var routes []string
+	if routeManager != nil {
+		routes = routeManager.CurrentRouteRange()
+	}
+
+	var searchDomains []string
+	if dnsServer != nil {
+		searchDomains = dnsServer.SearchDomains()
+	}
+
+	return routes, searchDomains
+}

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 +76,39 @@ 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)
+	// ShouldSendBootstrapOffer reports whether, as the controller, we should reply to a
+	// received responder offer with our own KEM offer (true only when no exchange is
+	// already in flight — so we kick the KEM once and ignore further offers).
+	ShouldSendBootstrapOffer(remoteKey string) bool
+	// 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 announced pq UDP port (port 0 means the peer omitted
+	// it and is on the default 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 +126,12 @@ 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 +190,11 @@ type Conn struct {
 	// pendingFirstPacket is the lazyconn-captured handshake init, replayed once the real
 	// transport is up.
 	pendingFirstPacket []byte
+
+	// pqStrictSentinelKey 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.
+	pqStrictSentinelKey *wgtypes.Key
 }
 
 // injectPendingFirstPacket replays the captured handshake through the proxy if present, else
@@ -206,6 +252,14 @@ 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.pqStrictSentinelKey = &k
+		}
+	}
+
 	return conn, nil
 }
 
@@ -670,6 +724,22 @@ 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 +751,10 @@ 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 +797,7 @@ 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 +816,7 @@ 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 +1020,35 @@ 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 +1090,23 @@ 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.pqStrictSentinelKey != 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.pqStrictSentinelKey
+		}
+	}
+
 	if conn.config.RosenpassConfig.PubKey == nil {
 		return conn.config.WgConfig.PreSharedKey
 	}
@@ -1026,6 +1146,21 @@ 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/conn_pq_test.go

@@ -0,0 +1,86 @@
+package peer
+
+import (
+	"net/netip"
+	"testing"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+	"github.com/stretchr/testify/require"
+	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
+)
+
+// fakePQ is a minimal PQHandshaker: only PSK is exercised by presharedKey, the rest
+// are no-op stubs to satisfy the interface.
+type fakePQ struct {
+	psk wgtypes.Key
+	ok  bool
+}
+
+func (f fakePQ) OfferPayload(string) ([]byte, int)          { return nil, 0 }
+func (f fakePQ) ShouldSendBootstrapOffer(string) bool       { return false }
+func (f fakePQ) AnswerPayload(string, []byte) ([]byte, int) { return nil, 0 }
+func (f fakePQ) OnAnswer(string, []byte)                    {}
+func (f fakePQ) PSK(string) (wgtypes.Key, bool)             { return f.psk, f.ok }
+func (f fakePQ) SetRemoteAddr(string, netip.AddrPort)       {}
+func (f fakePQ) OnDataPathRekeyed(string, time.Duration)    {}
+func (f fakePQ) OnDataPathDown(string)                      {}
+
+// TestConn_presharedKey_PQ covers the post-quantum branch of presharedKey across the
+// three states that matter: a derived PSK is programmed, and — before one exists —
+// strict mode blocks with a sentinel while non-strict falls open to the ordinary key.
+func TestConn_presharedKey_PQ(t *testing.T) {
+	derivedPSK, err := wgtypes.GenerateKey()
+	require.NoError(t, err)
+	nbPSK, err := wgtypes.GenerateKey()
+	require.NoError(t, err)
+
+	newConn := func() *Conn {
+		return &Conn{
+			Log: log.WithField("peer", "pq-test"),
+			config: ConnConfig{
+				Key:             "LLHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU=",
+				LocalKey:        "RRHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU=",
+				WgConfig:        WgConfig{PreSharedKey: &nbPSK},
+				RosenpassConfig: RosenpassConfig{},
+			},
+		}
+	}
+
+	t.Run("derived PSK is programmed", func(t *testing.T) {
+		for _, strict := range []bool{false, true} {
+			c := newConn()
+			c.config.PQ = fakePQ{psk: derivedPSK, ok: true}
+			c.config.PQStrict = strict
+			if strict {
+				sentinel, _ := wgtypes.GenerateKey()
+				c.pqStrictSentinelKey = &sentinel
+			}
+			got := c.presharedKey(nil)
+			require.NotNil(t, got)
+			require.Equal(t, derivedPSK, *got, "the derived PQ PSK must win (strict=%v)", strict)
+		}
+	})
+
+	t.Run("non-strict falls open to the ordinary key before a PSK exists", func(t *testing.T) {
+		c := newConn()
+		c.config.PQ = fakePQ{ok: false}
+		c.config.PQStrict = false
+		got := c.presharedKey(nil)
+		require.NotNil(t, got, "non-strict must not block")
+		require.Equal(t, nbPSK, *got, "non-strict falls through to the NetBird PSK, not a sentinel")
+	})
+
+	t.Run("strict blocks with the per-conn sentinel before a PSK exists", func(t *testing.T) {
+		sentinel, err := wgtypes.GenerateKey()
+		require.NoError(t, err)
+		c := newConn()
+		c.config.PQ = fakePQ{ok: false}
+		c.config.PQStrict = true
+		c.pqStrictSentinelKey = &sentinel
+		got := c.presharedKey(nil)
+		require.NotNil(t, got)
+		require.Equal(t, sentinel, *got, "strict must return the blocking sentinel")
+		require.NotEqual(t, nbPSK, *got, "the sentinel must not be the ordinary key")
+	})
+}

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 +136,42 @@ func (h *Handshaker) Listen(ctx context.Context) {
 
 			h.updateRemoteICEState(&remoteOfferAnswer)
 
+			h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
+
+			// Post-quantum: the KEM material rides only the controller's offer, so the two
+			// peers derive a single shared PSK (a bidirectional KEM would yield two
+			// different PSKs and WireGuard would pick misaligned ones). If we are the
+			// controller and receive the responder's (KEM-less) offer, we must NOT let
+			// that KEM-less transaction drive the connection: it would bring WireGuard up
+			// on a pre-PQ key before the KEM completes (the race). Instead we reply with
+			// OUR offer, which carries the KEM, so the only transaction that establishes
+			// the tunnel is the one that also derives the PSK. This also guarantees a
+			// responder-initiated wake still triggers a KEM offer (no stuck responder).
+			// The re-offer reuses our stable ICE session id, so the peer dedups repeats.
+			if h.config.PQ != nil && isController(h.config) {
+				// Reply with our KEM offer exactly once, to kick the exchange. If one is
+				// already in flight, ignore this offer — re-sending on every responder
+				// offer would be an offer-per-offer runaway.
+				if h.config.PQ.ShouldSendBootstrapOffer(h.config.Key) {
+					h.log.Debugf("pqkem: controller received a responder offer, replying with our KEM offer instead of an answer")
+					if err := h.sendOffer(); err != nil {
+						h.log.Errorf("failed to send KEM offer in response to peer offer: %s", err)
+					}
+				} else {
+					h.log.Debugf("pqkem: controller received a responder offer but a KEM exchange is already in flight, ignoring")
+				}
+				continue
+			}
+
+			// Derive+store the KEM PSK (inside sendAnswer's AnswerPayload) BEFORE bringing
+			// up the connection: the relay/ICE workers configure the WG endpoint, which
+			// pulls the PSK for the first handshake. Notifying them first would race the
+			// KEM exchange and hand the first handshake a not-yet-derived key.
+			if err := h.sendAnswer(&remoteOfferAnswer); err != nil {
+				h.log.Errorf("failed to send remote offer confirmation: %s", err)
+				continue
+			}
+
 			if h.relayListener != nil {
 				h.relayListener.Notify(&remoteOfferAnswer)
 			}
@@ -127,11 +179,6 @@ func (h *Handshaker) Listen(ctx context.Context) {
 			if h.iceListener != nil && h.RemoteICESupported() {
 				h.iceListener(&remoteOfferAnswer)
 			}
-
-			if err := h.sendAnswer(); err != nil {
-				h.log.Errorf("failed to send remote offer confirmation: %s", err)
-				continue
-			}
 		case remoteOfferAnswer := <-h.remoteAnswerCh:
 			h.log.Infof("received answer, running version %s, remote WireGuard listen port %d, session id: %s, remote ICE supported: %t", remoteOfferAnswer.Version, remoteOfferAnswer.WgListenPort, remoteOfferAnswer.SessionIDString(), remoteOfferAnswer.hasICECredentials())
 
@@ -142,6 +189,15 @@ func (h *Handshaker) Listen(ctx context.Context) {
 
 			h.updateRemoteICEState(&remoteOfferAnswer)
 
+			h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
+
+			// Feed the KEM answer (derive+store PSK) BEFORE bringing up the connection so
+			// the WG endpoint config pulls the real PSK for the first handshake instead of
+			// racing ahead of the KEM exchange.
+			if h.config.PQ != nil {
+				h.config.PQ.OnAnswer(h.config.Key, remoteOfferAnswer.MlkemPayload)
+			}
+
 			if h.relayListener != nil {
 				h.relayListener.Notify(&remoteOfferAnswer)
 			}
@@ -156,6 +212,18 @@ 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
+	}
+	// remotePort may be 0 (the peer omitted it, meaning the default port); the adapter
+	// resolves 0 to DefaultPort.
+	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 +263,23 @@ 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/capability_test.go

@@ -0,0 +1,65 @@
+package pqkem
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+// TestManager_NonCapablePeerNotOffered: a peer known not to run the KEM (it advertised
+// no PQ port over signalling) is never offered an exchange, and no failure is raised —
+// this is what stops the reoffer storm against non-PQ peers (e.g. Rosenpass peers).
+func TestManager_NonCapablePeerNotOffered(t *testing.T) {
+	wg := newFakeWG()
+	d := NewManager("bbbb", wg, nil) // initiator vs "aaaa"
+	d.Start(&loopback{ep: epB, sw: newSwitch()})
+	defer d.Stop()
+
+	d.MarkNonCapable("aaaa")
+
+	offer, err := d.SignalOffer("aaaa")
+	require.NoError(t, err)
+	require.Nil(t, offer, "a non-capable peer must not be offered a KEM exchange")
+	require.Empty(t, wg.failed, "a non-capable peer must not raise a rekey failure")
+}
+
+// TestManager_MarkNonCapableCancelsInFlight: if we start an exchange with a peer whose
+// capability is not yet known and then learn it does not run the KEM, the in-flight
+// exchange is cancelled and no further offer is produced (no timeout -> no failure).
+func TestManager_MarkNonCapableCancelsInFlight(t *testing.T) {
+	wg := newFakeWG()
+	d := NewManager("bbbb", wg, nil)
+	d.Start(&loopback{ep: epB, sw: newSwitch()})
+	defer d.Stop()
+
+	// Capability unknown -> the bootstrap offer goes out optimistically.
+	offer, err := d.SignalOffer("aaaa")
+	require.NoError(t, err)
+	require.NotNil(t, offer)
+
+	// Now we learn the peer is non-PQ: the exchange must be dropped.
+	d.MarkNonCapable("aaaa")
+
+	next, err := d.SignalOffer("aaaa")
+	require.NoError(t, err)
+	require.Nil(t, next, "after learning non-capability the peer is no longer offered")
+	require.Empty(t, wg.failed, "cancelling an in-flight exchange must not raise a failure")
+}
+
+// TestManager_EstablishedPeerNotDowngraded: a stray zero-port observation must not tear
+// down a peer we already have a working PQ session with.
+func TestManager_EstablishedPeerNotDowngraded(t *testing.T) {
+	dA, dB, _, wgB, _ := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	bootstrap(t, dA, dB)
+	require.NotEqual(t, PSK{}, wgB.psk("aaaa"), "established a PSK")
+
+	dB.MarkNonCapable("aaaa") // stray zero after establishment
+
+	// The peer keeps its derived PSK (MarkNonCapable is a no-op once established).
+	psk, ok := dB.PSK("aaaa")
+	require.True(t, ok, "an established peer must keep its PSK despite a stray zero")
+	require.NotEqual(t, PSK{}, psk)
+}

client/internal/pqkem/concurrency_test.go

@@ -0,0 +1,77 @@
+package pqkem
+
+import (
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+// TestConcurrency_RecoversViaResignalAfterDataPathBreak exercises the A-light recovery:
+// a data-path rotation can no longer converge (OnRekeyFailed), and re-bootstrapping over
+// signalling resyncs both peers on a fresh PSK — even while the data path stays broken,
+// since the signal channel is independent of it.
+func TestConcurrency_RecoversViaResignalAfterDataPathBreak(t *testing.T) {
+	dA, dB, wgA, wgB, lbB := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	// Tighten B's timings and make a single rotation miss raise OnRekeyFailed. Set
+	// before any exchange loop spawns (the loop reads these fields).
+	dB.retryInterval = 5 * time.Millisecond
+	dB.maxRetries = 2
+	dB.maxRekeyFailures = 1
+
+	bootstrap(t, dA, dB)
+	dA.OnDataPathRekeyed("bbbb", 0)
+	dB.OnDataPathRekeyed("aaaa", 0)
+	psk1 := wgB.psk("aaaa")
+	require.NotEqual(t, PSK{}, psk1)
+	require.Equal(t, psk1, wgA.psk("bbbb"), "converged on the same PSK after bootstrap+rotation")
+
+	// Data path breaks: the rotation can no longer converge -> OnRekeyFailed.
+	lbB.drop.Store(true)
+	_, err := dB.startExchange("aaaa", false, ExchangeID{})
+	require.NoError(t, err)
+	require.Eventually(t, func() bool { return failedCount(wgB) >= 1 }, time.Second, 5*time.Millisecond)
+
+	// Recovery: re-bootstrap over signalling with the data path STILL broken. It must
+	// still converge (signal is independent of the data path) on a fresh PSK.
+	bootstrap(t, dA, dB)
+	psk2 := wgB.psk("aaaa")
+	require.NotEqual(t, psk1, psk2, "recovery derived a fresh PSK")
+	require.Equal(t, psk2, wgA.psk("bbbb"), "both sides resync after recovery")
+}
+
+// TestConcurrency_ConcurrentRekeysNoRace hammers both managers with concurrent rotation
+// clocks from many goroutines. Its primary job (with -race) is to prove the single-lock
+// state machine has no data races or deadlocks under contention; a final deterministic
+// bootstrap then asserts there is no split-brain (both sides on the same PSK).
+func TestConcurrency_ConcurrentRekeysNoRace(t *testing.T) {
+	dA, dB, wgA, wgB, _ := pair(t)
+	defer dA.Stop()
+	defer dB.Stop()
+
+	bootstrap(t, dA, dB)
+
+	var wg sync.WaitGroup
+	for g := 0; g < 8; g++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for i := 0; i < 50; i++ {
+				dB.OnDataPathRekeyed("aaaa", 0) // initiator chains a rotation
+				dA.OnDataPathRekeyed("bbbb", 0) // responder side is a no-op, still stresses the lock
+			}
+		}()
+	}
+	wg.Wait()
+
+	// The storm may leave an exchange mid-flight (concurrent cancellation). Force a
+	// clean convergence over signalling, then assert no split-brain.
+	bootstrap(t, dA, dB)
+	a, b := wgA.psk("bbbb"), wgB.psk("aaaa")
+	require.NotEqual(t, PSK{}, b)
+	require.Equal(t, a, b, "both sides converge on the same PSK, no split-brain")
+}

client/internal/pqkem/convergence.go

@@ -0,0 +1,273 @@
+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.capable[remoteID] = true // a real KEM offer proves the peer runs the exchange
+	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.capable[remoteID] = true // a real KEM answer proves the peer runs the exchange
+	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_framing_test.go

@@ -0,0 +1,105 @@
+package pqkem
+
+import (
+	"crypto/mlkem"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+// TestExchange_TamperedCiphertextFailsClosed verifies the core fail-closed
+// property: mutating the ML-KEM ciphertext in the answer does not error (ML-KEM
+// uses implicit rejection — Decapsulate always returns a value) but yields a
+// different shared secret, so the initiator derives a PSK that does NOT match the
+// responder's. A mismatched PSK means WireGuard passes no bytes: tamper => no data.
+func TestExchange_TamperedCiphertextFailsClosed(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	tampered := append([]byte(nil), answer...)
+	tampered[0] ^= 0xff // flip a bit in the ML-KEM ciphertext
+
+	pskA, err := init.Finish(tampered, Binding{LocalID: wgA, RemoteID: wgB})
+	require.NoError(t, err, "implicit rejection: decapsulate still succeeds")
+	require.NotEqual(t, pskB, pskA, "tampered ciphertext must not yield the responder's PSK")
+}
+
+// TestExchange_TamperedX25519ShareDiverges flips a byte in the answer's X25519
+// share: the classical half of the hybrid secret changes, so the derived PSK
+// diverges from the responder's (fail-closed on the ECDH half too).
+func TestExchange_TamperedX25519ShareDiverges(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	tampered := append([]byte(nil), answer...)
+	tampered[mlkem.CiphertextSize768] ^= 0x01 // first byte of the X25519 public key
+
+	pskA, err := init.Finish(tampered, Binding{LocalID: wgA, RemoteID: wgB})
+	// Either the point is rejected (error) or the ECDH differs (different PSK);
+	// in both cases the honest PSK is never reproduced.
+	if err == nil {
+		require.NotEqual(t, pskB, pskA, "tampered X25519 share must not yield the responder's PSK")
+	}
+}
+
+// TestExchange_AllZeroX25519Rejected feeds an all-zero X25519 share (a low-order
+// point) in the answer. The stdlib ECDH must reject it, so Finish errors rather
+// than deriving a PSK from a degenerate secret.
+func TestExchange_AllZeroX25519Rejected(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	bad := append([]byte(nil), answer...)
+	for i := mlkem.CiphertextSize768; i < len(bad); i++ {
+		bad[i] = 0
+	}
+
+	_, err = init.Finish(bad, Binding{LocalID: wgA, RemoteID: wgB})
+	require.Error(t, err, "all-zero X25519 share (low-order point) must be rejected")
+}
+
+// TestExchange_SizeBoundaries locks the exact-length framing checks: one byte
+// short or long on either message is rejected, not silently truncated/padded.
+func TestExchange_SizeBoundaries(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+	offer := init.Offer()
+
+	_, _, err = Respond(offer[:OfferSize-1], Binding{})
+	require.Error(t, err, "offer one byte short")
+	_, _, err = Respond(append(append([]byte(nil), offer...), 0), Binding{})
+	require.Error(t, err, "offer one byte long")
+
+	answer, _, err := Respond(offer, Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	_, err = init.Finish(answer[:AnswerSize-1], Binding{LocalID: wgA, RemoteID: wgB})
+	require.Error(t, err, "answer one byte short")
+	_, err = init.Finish(append(append([]byte(nil), answer...), 0), Binding{LocalID: wgA, RemoteID: wgB})
+	require.Error(t, err, "answer one byte long")
+}
+
+// TestExchange_BindingIsSymmetric confirms the canonicalisation: the two peers
+// pass their identities in opposite (Local, Remote) order yet derive the same PSK,
+// so identity binding does not depend on who is initiator vs responder.
+func TestExchange_BindingIsSymmetric(t *testing.T) {
+	init, err := NewInitiator()
+	require.NoError(t, err)
+
+	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
+	require.NoError(t, err)
+
+	pskA, err := init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
+	require.NoError(t, err)
+
+	require.Equal(t, pskB, pskA, "swapped Local/Remote order must canonicalise to the same PSK")
+}

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 +1,460 @@
+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)
+	capable     map[RemoteID]bool           // peer runs the KEM (advertised a PQ port); false = known non-capable
+	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),
+		capable:          make(map[RemoteID]bool),
+		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.
+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). This is pure routing and says nothing about capability —
+// PQ capability is decided solely from the peer's KEM payload (see processOffer /
+// processAnswer / MarkNonCapable), never from an endpoint or port.
+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()
+}
+
+// MarkNonCapable records that a peer does not run the KEM: it answered our offer with
+// no KEM material over signalling (the capability signal is the peer's payload, not its
+// optional data-path port). Any in-flight exchange is cancelled and further offers are
+// suppressed (see SignalOffer), so a non-PQ peer never drives the rekey-recovery storm.
+// An already-established peer is left untouched — a stray empty answer must not tear
+// down a working PQ session.
+func (m *Manager) MarkNonCapable(remoteID RemoteID) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if m.established[remoteID] {
+		return
+	}
+	if prev, ok := m.capable[remoteID]; ok && !prev {
+		return // already known non-capable, nothing to do
+	}
+	m.capable[remoteID] = false
+	if ex := m.exchanges[remoteID]; ex != nil {
+		if ex.cancel != nil {
+			ex.cancel()
+		}
+		delete(m.exchanges, remoteID)
+	}
+	m.trace("pqkem: peer advertises no PQ service — treating as non-capable, no KEM attempted", "peer", remoteID)
+}
+
+// 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)
+	delete(m.capable, 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.
+//
+// A signal re-negotiation always re-bootstraps (fresh exchange): the remote may have
+// restarted and lost its PSK, so reusing a locally frozen one would desync. The derived
+// PSK still survives idle in the manager (dropped only on account-level peer removal),
+// so a pure lazy wake with no re-negotiation reuses it via the conn's WG-config pull.
+func (m *Manager) SignalOffer(remoteID RemoteID) ([]byte, error) {
+	if !m.IsInitiator(remoteID) {
+		return nil, nil
+	}
+	m.mu.Lock()
+	if capable, ok := m.capable[remoteID]; ok && !capable {
+		m.mu.Unlock()
+		return nil, nil // peer does not run the KEM; do not offer (avoids a failure/reoffer loop)
+	}
+	// Idempotent while a signalling bootstrap is in flight OR already derived a PSK but
+	// not yet chained a rotation (awaitingRekey): return the SAME offer instead of
+	// starting a new exchange. This matters when the controller both offers on its own
+	// guard AND re-offers in response to the responder's offer — without this, the
+	// second call would start a fresh exchange (a different PSK) and desync the peers.
+	if ex := m.exchanges[remoteID]; ex != nil && ex.viaSignal &&
+		(ex.state == stateAwaitingAnswer || ex.state == stateAwaitingRekey) {
+		last := ex.lastSent
+		m.mu.Unlock()
+		return last, nil
+	}
+	m.mu.Unlock()
+	// bootstrap offer acknowledges nothing (zero AckID).
+	return m.startExchange(remoteID, true, ExchangeID{})
+}
+
+// ShouldSendBootstrapOffer reports whether we should emit a fresh KEM offer to kick a
+// bootstrap for this peer. True only if we are the initiator, the peer is not known
+// non-capable, and no exchange is already in flight. The host uses this when it (as the
+// controller) receives the responder's offer: it replies with a KEM offer exactly once
+// to start the exchange, and ignores further responder offers while one is in flight,
+// avoiding an offer-per-offer runaway.
+func (m *Manager) ShouldSendBootstrapOffer(remoteID RemoteID) bool {
+	if !m.IsInitiator(remoteID) {
+		return false
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if capable, ok := m.capable[remoteID]; ok && !capable {
+		return false
+	}
+	return m.exchanges[remoteID] == nil
+}
+
+// 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 clocks the next chained PSK rotation on a fresh data-path rekey
+// (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 the key works). sinceActivity is how long ago the peer last exchanged real
+// user data; past 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 +1,150 @@
+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).
+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
+}
+
+// announcedPort is the PQ data-path port to advertise to peers. It is omitted (0) when
+// the manager is on DefaultPort, since peers assume the default when no port is sent;
+// only a non-default (collision-forced) port is announced explicitly.
+func (p pqHandshaker) announcedPort() int {
+	if port := p.mgr.LocalPort(); port != DefaultPort {
+		return port
+	}
+	return 0
+}
+
+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.announcedPort()
+}
+
+func (p pqHandshaker) ShouldSendBootstrapOffer(remoteKey string) bool {
+	return p.mgr.ShouldSendBootstrapOffer(pqkem.RemoteID(remoteKey))
+}
+
+func (p pqHandshaker) AnswerPayload(remoteKey string, recvOffer []byte) ([]byte, int) {
+	if len(recvOffer) == 0 {
+		// Capability signal (responder side): the KEM offer flows initiator->responder,
+		// so if we are the responder for this peer (it is the KEM initiator by role) an
+		// empty offer means it does not run the KEM. If we are the initiator, an empty
+		// offer is normal — the peer is the responder and puts its material in the
+		// answer — so we must not flag it.
+		if !p.mgr.IsInitiator(pqkem.RemoteID(remoteKey)) {
+			p.mgr.MarkNonCapable(pqkem.RemoteID(remoteKey))
+		}
+		return nil, p.announcedPort()
+	}
+	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.announcedPort()
+}
+
+func (p pqHandshaker) OnAnswer(remoteKey string, recvAnswer []byte) {
+	if len(recvAnswer) == 0 {
+		// Capability signal (initiator side): the KEM answer flows responder->initiator,
+		// so an empty answer to our offer means the peer does not run the KEM — mark it
+		// non-capable to stop offering (no failure/reoffer storm). Only meaningful when
+		// we are the initiator: as the responder we also receive an (empty) answer to
+		// our own non-KEM offer from a perfectly capable peer, which must not be flagged.
+		if p.mgr.IsInitiator(pqkem.RemoteID(remoteKey)) {
+			p.mgr.MarkNonCapable(pqkem.RemoteID(remoteKey))
+		}
+		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 learned from signalling. A
+// zero port means the peer omitted it (it is on DefaultPort), so we resolve it here —
+// DefaultPort lives in this package, not in peer. 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.Addr().IsValid() {
+		return
+	}
+	port := addr.Port()
+	if port == 0 {
+		port = DefaultPort
+	}
+	p.mgr.AddPeer(pqkem.RemoteID(remoteKey), netip.AddrPortFrom(addr.Addr(), port))
+}
+
+// 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_adapter_test.go

@@ -0,0 +1,37 @@
+package internal
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/pqkem"
+)
+
+type pqNoopHandler struct{}
+
+func (pqNoopHandler) OnNewPSKReady(pqkem.RemoteID, pqkem.PSK) error { return nil }
+func (pqNoopHandler) OnRekeyFailed(pqkem.RemoteID) error            { return nil }
+
+// TestPQAdapter_CapabilityRoleAware locks the role-aware capability signal: the KEM
+// payload only flows initiator-offer -> responder-answer, so an empty message in the
+// other direction comes from a perfectly capable peer and must NOT flag it. Only the
+// message that should carry material (the answer we receive as initiator) marks a peer
+// non-capable when empty.
+func TestPQAdapter_CapabilityRoleAware(t *testing.T) {
+	// localID "zzzz" > "aaaa" => this manager is the KEM initiator for peer "aaaa".
+	mgr := pqkem.NewManager("zzzz", pqNoopHandler{}, nil)
+	defer mgr.Stop()
+	h := pqHandshaker{mgr: mgr}
+
+	// An empty OFFER from our peer is normal here: as the initiator's responder it puts
+	// its material in the answer, not the offer. It must not disable our offering.
+	h.AnswerPayload("aaaa", nil)
+	payload, _ := h.OfferPayload("aaaa")
+	require.NotNil(t, payload, "an empty offer from a responder-role peer must not mark it non-capable")
+
+	// An empty ANSWER to our offer means the peer does not run the KEM -> stop offering.
+	h.OnAnswer("aaaa", nil)
+	payload2, _ := h.OfferPayload("aaaa")
+	require.Nil(t, payload2, "an empty answer to our offer marks the peer non-capable, so we stop offering")
+}

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/state.go

@@ -45,12 +45,35 @@ func (pm *ProfileManager) GetProfileState(id ID) (*ProfileState, error) {
 	return &state, nil
 }
 
-func (pm *ProfileManager) SetActiveProfileState(state *ProfileState) error {
+// SetProfileState writes the state file of the profile identified by id. Prefer
+// it over SetActiveProfileState whenever the caller knows which profile the data
+// belongs to: an SSO login spans seconds of user interaction, and the active
+// profile can change during it, which would file the account email under
+// whichever profile happened to be active when the flow returned.
+func (pm *ProfileManager) SetProfileState(id ID, state *ProfileState) error {
 	configDir, err := getConfigDir()
 	if err != nil {
 		return fmt.Errorf("get config directory: %w", err)
 	}
 
+	if id == "" {
+		return fmt.Errorf("empty profile ID")
+	}
+	if id != defaultProfileName && !IsValidProfileFilenameStem(id) {
+		return fmt.Errorf("invalid profile ID: %q", id)
+	}
+
+	stateFile := filepath.Join(configDir, id.String()+".state.json")
+	if err := util.WriteJsonWithRestrictedPermission(context.Background(), stateFile, state); err != nil {
+		return fmt.Errorf("write profile state: %w", err)
+	}
+
+	return nil
+}
+
+// SetActiveProfileState writes the state file of whichever profile is active at
+// call time. Use SetProfileState when the target profile is known.
+func (pm *ProfileManager) SetActiveProfileState(state *ProfileState) error {
 	activeProf, err := pm.GetActiveProfile()
 	if err != nil {
 		if errors.Is(err, ErrNoActiveProfile) {
@@ -59,18 +82,7 @@ func (pm *ProfileManager) SetActiveProfileState(state *ProfileState) error {
 		return fmt.Errorf("get active profile: %w", err)
 	}
 
-	id := activeProf.ID
-	if id != defaultProfileName && !IsValidProfileFilenameStem(id) {
-		return fmt.Errorf("invalid active profile ID: %q", id)
-	}
-
-	stateFile := filepath.Join(configDir, id.String()+".state.json")
-	err = util.WriteJsonWithRestrictedPermission(context.Background(), stateFile, state)
-	if err != nil {
-		return fmt.Errorf("write profile state: %w", err)
-	}
-
-	return nil
+	return pm.SetProfileState(activeProf.ID, state)
 }
 
 // RemoveProfileState deletes the per-profile state file (which holds the

client/internal/routemanager/dnsinterceptor/handler.go

@@ -479,7 +479,7 @@ func (d *DnsInterceptor) removeDNATMappings(realPrefixes []netip.Prefix, logger
 
 // internalDnatFw checks if the firewall supports internal DNAT
 func (d *DnsInterceptor) internalDnatFw() (internalDNATer, bool) {
-	if d.firewall == nil || runtime.GOOS != "android" {
+	if d.firewall == nil || d.fakeIPManager == nil || runtime.GOOS != "android" {
 		return nil, false
 	}
 	fw, ok := d.firewall.(internalDNATer)

client/internal/routemanager/manager.go

@@ -8,14 +8,13 @@ import (
 	"net/netip"
 	"net/url"
 	"runtime"
-	"slices"
+	"sort"
 	"strings"
 	"sync"
 	"sync/atomic"
 	"syscall"
 	"time"
 
-	"github.com/google/uuid"
 	"github.com/hashicorp/go-multierror"
 	log "github.com/sirupsen/logrus"
 	"golang.org/x/exp/maps"
@@ -62,7 +61,7 @@ type Manager interface {
 	GetActiveClientRoutes() route.HAMap
 	GetClientRoutesWithNetID() map[route.NetID][]*route.Route
 	SetRouteChangeListener(listener listener.NetworkChangeListener)
-	InitialRouteRange() []string
+	CurrentRouteRange() []string
 	SetFirewall(firewall.Manager) error
 	SetDNSForwarderPort(port uint16)
 	ReconcilePeerAllowedIPs(peerKey string) error
@@ -76,10 +75,8 @@ type ManagerConfig struct {
 	WGInterface         iface.WGIface
 	StatusRecorder      *peer.Status
 	RelayManager        *relayClient.Manager
-	InitialRoutes       []*route.Route
 	StateManager        *statemanager.Manager
 	DNSServer           dns.Server
-	DNSFeatureFlag      bool
 	PeerStore           *peerstore.Store
 	DisableClientRoutes bool
 	DisableServerRoutes bool
@@ -149,45 +146,12 @@ func NewManager(config ManagerConfig) *DefaultManager {
 	useNoop := netstack.IsEnabled() || config.DisableClientRoutes
 	dm.setupRefCounters(useNoop)
 
-	// don't proceed with client routes if it is disabled
-	if config.DisableClientRoutes {
-		return dm
-	}
-
-	if runtime.GOOS == "android" {
-		dm.setupAndroidRoutes(config)
-	}
 	return dm
 }
-func (m *DefaultManager) setupAndroidRoutes(config ManagerConfig) {
-	cr := m.initialClientRoutes(config.InitialRoutes)
 
-	routesForComparison := slices.Clone(cr)
-
-	if config.DNSFeatureFlag {
-		m.fakeIPManager = fakeip.NewManager()
-
-		v4ID := uuid.NewString()
-		fakeIPRoute := &route.Route{
-			ID:          route.ID(v4ID),
-			Network:     m.fakeIPManager.GetFakeIPBlock(),
-			NetID:       route.NetID(v4ID),
-			Peer:        m.pubKey,
-			NetworkType: route.IPv4Network,
-		}
-		v6ID := uuid.NewString()
-		fakeIPv6Route := &route.Route{
-			ID:          route.ID(v6ID),
-			Network:     m.fakeIPManager.GetFakeIPv6Block(),
-			NetID:       route.NetID(v6ID),
-			Peer:        m.pubKey,
-			NetworkType: route.IPv6Network,
-		}
-		cr = append(cr, fakeIPRoute, fakeIPv6Route)
-		m.notifier.SetFakeIPRoutes([]*route.Route{fakeIPRoute, fakeIPv6Route})
-	}
-
-	m.notifier.SetInitialClientRoutes(cr, routesForComparison)
+func (m *DefaultManager) enableFakeIPRoutes() {
+	m.fakeIPManager = fakeip.NewManager()
+	m.notifier.NotifyRouteChange()
 }
 
 func (m *DefaultManager) setupRefCounters(useNoop bool) {
@@ -464,6 +428,9 @@ func (m *DefaultManager) UpdateRoutes(
 
 	var merr *multierror.Error
 	if !m.disableClientRoutes {
+		if runtime.GOOS == "android" && useNewDNSRoute && m.fakeIPManager == nil {
+			m.enableFakeIPRoutes()
+		}
 
 		// Update route selector based on management server's isSelected status
 		m.updateRouteSelectorFromManagement(clientRoutes)
@@ -500,9 +467,32 @@ func (m *DefaultManager) SetRouteChangeListener(listener listener.NetworkChangeL
 	m.notifier.SetListener(listener)
 }
 
-// InitialRouteRange return the list of initial routes. It used by mobile systems
-func (m *DefaultManager) InitialRouteRange() []string {
-	return m.notifier.GetInitialRouteRanges()
+// CurrentRouteRange returns the current TUN route list. It is used by mobile systems
+func (m *DefaultManager) CurrentRouteRange() []string {
+	m.mux.Lock()
+	defer m.mux.Unlock()
+
+	if m.disableClientRoutes {
+		return nil
+	}
+
+	filtered := m.routeSelector.FilterSelectedExitNodes(m.clientRoutes)
+	var nets []string
+	for _, routes := range filtered {
+		for _, r := range routes {
+			if r.IsDynamic() {
+				continue
+			}
+			nets = append(nets, r.NetString())
+		}
+	}
+
+	if m.fakeIPManager != nil {
+		nets = append(nets, m.fakeIPManager.GetFakeIPBlock().String(), m.fakeIPManager.GetFakeIPv6Block().String())
+	}
+
+	sort.Strings(nets)
+	return nets
 }
 
 // GetRouteSelector returns the route selector
@@ -700,16 +690,6 @@ func (m *DefaultManager) ClassifyRoutes(newRoutes []*route.Route) (map[route.ID]
 	return newServerRoutesMap, newClientRoutesIDMap
 }
 
-func (m *DefaultManager) initialClientRoutes(initialRoutes []*route.Route) []*route.Route {
-	_, crMap := m.ClassifyRoutes(initialRoutes)
-	rs := make([]*route.Route, 0, len(crMap))
-	for _, routes := range crMap {
-		rs = append(rs, routes...)
-	}
-
-	return rs
-}
-
 func isRouteSupported(route *route.Route) bool {
 	if netstack.IsEnabled() || !nbnet.CustomRoutingDisabled() || route.IsDynamic() {
 		return true

client/internal/routemanager/mock.go

@@ -30,8 +30,8 @@ func (m *MockManager) Init() error {
 	return nil
 }
 
-// InitialRouteRange mock implementation of InitialRouteRange from Manager interface
-func (m *MockManager) InitialRouteRange() []string {
+// CurrentRouteRange mock implementation of CurrentRouteRange from Manager interface
+func (m *MockManager) CurrentRouteRange() []string {
 	return nil
 }
 

client/internal/routemanager/notifier/notifier_android.go

@@ -6,7 +6,6 @@ import (
 	"net/netip"
 	"slices"
 	"sort"
-	"strings"
 	"sync"
 
 	"github.com/netbirdio/netbird/client/internal/listener"
@@ -14,12 +13,15 @@ import (
 )
 
 type Notifier struct {
-	initialRoutes []*route.Route
-	currentRoutes []*route.Route
-	fakeIPRoutes  []*route.Route
+	mu sync.Mutex
 
-	listener    listener.NetworkChangeListener
-	listenerMux sync.Mutex
+	// currentRoutes is the last announced route set. It exists only to
+	// suppress noise: without it every network map sync would trigger the
+	// Java side, even when the routes did not change. The actual TUN route
+	// state is owned by the route manager and pulled from there.
+	currentRoutes []*route.Route
+
+	listener listener.NetworkChangeListener
 }
 
 func NewNotifier() *Notifier {
@@ -27,20 +29,15 @@ func NewNotifier() *Notifier {
 }
 
 func (n *Notifier) SetListener(listener listener.NetworkChangeListener) {
-	n.listenerMux.Lock()
-	defer n.listenerMux.Unlock()
+	n.mu.Lock()
+	defer n.mu.Unlock()
 	n.listener = listener
 }
 
-// SetInitialClientRoutes stores the initial route sets for TUN configuration.
-func (n *Notifier) SetInitialClientRoutes(initialRoutes []*route.Route, routesForComparison []*route.Route) {
-	n.initialRoutes = filterStatic(initialRoutes)
-	n.currentRoutes = filterStatic(routesForComparison)
-}
-
-// SetFakeIPRoutes stores the fake IP routes to be included in every TUN rebuild.
-func (n *Notifier) SetFakeIPRoutes(routes []*route.Route) {
-	n.fakeIPRoutes = routes
+func (n *Notifier) NotifyRouteChange() {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	n.notifyLocked()
 }
 
 func (n *Notifier) OnNewRoutes(idMap route.HAMap) {
@@ -54,46 +51,32 @@ func (n *Notifier) OnNewRoutes(idMap route.HAMap) {
 		}
 	}
 
-	if !n.hasRouteDiff(n.currentRoutes, newRoutes) {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	if !hasRouteDiff(n.currentRoutes, newRoutes) {
 		return
 	}
 
 	n.currentRoutes = newRoutes
-	n.notify()
+	n.notifyLocked()
 }
 
 func (n *Notifier) OnNewPrefixes([]netip.Prefix) {
 	// Not used on Android
 }
 
-func (n *Notifier) notify() {
-	n.listenerMux.Lock()
-	defer n.listenerMux.Unlock()
+func (n *Notifier) notifyLocked() {
 	if n.listener == nil {
 		return
 	}
-
-	allRoutes := slices.Clone(n.currentRoutes)
-	allRoutes = append(allRoutes, n.fakeIPRoutes...)
-
-	routeStrings := n.routesToStrings(allRoutes)
-	sort.Strings(routeStrings)
-	go func(l listener.NetworkChangeListener) {
-		l.OnNetworkChanged(strings.Join(routeStrings, ","))
-	}(n.listener)
+	n.listener.OnNetworkChanged("")
 }
 
-func filterStatic(routes []*route.Route) []*route.Route {
-	out := make([]*route.Route, 0, len(routes))
-	for _, r := range routes {
-		if !r.IsDynamic() {
-			out = append(out, r)
-		}
-	}
-	return out
+func (n *Notifier) Close() {
+	// unused
 }
 
-func (n *Notifier) routesToStrings(routes []*route.Route) []string {
+func routesToStrings(routes []*route.Route) []string {
 	nets := make([]string, 0, len(routes))
 	for _, r := range routes {
 		nets = append(nets, r.NetString())
@@ -101,25 +84,10 @@ func (n *Notifier) routesToStrings(routes []*route.Route) []string {
 	return nets
 }
 
-func (n *Notifier) hasRouteDiff(a []*route.Route, b []*route.Route) bool {
-	slices.SortFunc(a, func(x, y *route.Route) int {
-		return strings.Compare(x.NetString(), y.NetString())
-	})
-	slices.SortFunc(b, func(x, y *route.Route) int {
-		return strings.Compare(x.NetString(), y.NetString())
-	})
-
-	return !slices.EqualFunc(a, b, func(x, y *route.Route) bool {
-		return x.NetString() == y.NetString()
-	})
-}
-
-func (n *Notifier) GetInitialRouteRanges() []string {
-	initialStrings := n.routesToStrings(n.initialRoutes)
-	sort.Strings(initialStrings)
-	return initialStrings
-}
-
-func (n *Notifier) Close() {
-	// unused
+func hasRouteDiff(a []*route.Route, b []*route.Route) bool {
+	as := routesToStrings(a)
+	bs := routesToStrings(b)
+	sort.Strings(as)
+	sort.Strings(bs)
+	return !slices.Equal(as, bs)
 }

client/internal/routemanager/notifier/notifier_ios.go

@@ -29,11 +29,7 @@ func (n *Notifier) SetListener(listener listener.NetworkChangeListener) {
 	n.listener = listener
 }
 
-func (n *Notifier) SetInitialClientRoutes([]*route.Route, []*route.Route) {
-	// iOS doesn't care about initial routes
-}
-
-func (n *Notifier) SetFakeIPRoutes([]*route.Route) {
+func (n *Notifier) NotifyRouteChange() {
 	// Not used on iOS
 }
 

client/internal/routemanager/notifier/notifier_other.go

@@ -19,11 +19,7 @@ func (n *Notifier) SetListener(listener listener.NetworkChangeListener) {
 	// Not used on non-mobile platforms
 }
 
-func (n *Notifier) SetInitialClientRoutes([]*route.Route, []*route.Route) {
-	// Not used on non-mobile platforms
-}
-
-func (n *Notifier) SetFakeIPRoutes([]*route.Route) {
+func (n *Notifier) NotifyRouteChange() {
 	// Not used on non-mobile platforms
 }
 
@@ -35,10 +31,6 @@ func (n *Notifier) OnNewPrefixes(prefixes []netip.Prefix) {
 	// Not used on non-mobile platforms
 }
 
-func (n *Notifier) GetInitialRouteRanges() []string {
-	return []string{}
-}
-
 func (n *Notifier) Close() {
 	// unused
 }

client/internal/updater/installer/installer_run_darwin.go

@@ -98,47 +98,44 @@ func (u *Installer) startDaemon(daemonFolder string) error {
 func (u *Installer) startUIAsUser() error {
 	log.Infof("starting netbird-ui: %s", uiBinary)
 
-	// Get the current console user
-	cmd := exec.Command("stat", "-f", "%Su", "/dev/console")
-	output, err := cmd.Output()
+	username, err := consoleUser()
 	if err != nil {
-		return fmt.Errorf("failed to get console user: %w", err)
+		return err
 	}
 
-	username := strings.TrimSpace(string(output))
-	if username == "" || username == "root" {
-		return fmt.Errorf("no active user session found")
-	}
-
-	log.Infof("starting UI for user: %s", username)
-
-	// Get user's UID
 	userInfo, err := user.Lookup(username)
 	if err != nil {
-		return fmt.Errorf("failed to lookup user %s: %w", username, err)
+		return fmt.Errorf("lookup user %s: %w", username, err)
 	}
 
-	// Start the UI process as the console user using launchctl
-	// This ensures the app runs in the user's context with proper GUI access
-	launchCmd := exec.Command("launchctl", "asuser", userInfo.Uid, "open", "-a", uiBinary)
+	log.Infof("starting UI for user: %s (uid %s)", username, userInfo.Uid)
+
+	launchCmd := exec.Command("launchctl", "asuser", userInfo.Uid, "sudo", "-u", username, "-H", "open", "-a", uiBinary)
 	log.Infof("launchCmd: %s", launchCmd.String())
-	// Set the user's home directory for proper macOS app behavior
-	launchCmd.Env = append(os.Environ(), "HOME="+userInfo.HomeDir)
-	log.Infof("set HOME environment variable: %s", userInfo.HomeDir)
 
-	if err := launchCmd.Start(); err != nil {
-		return fmt.Errorf("failed to start UI process: %w", err)
-	}
-
-	// Release the process so it can run independently
-	if err := launchCmd.Process.Release(); err != nil {
-		log.Warnf("failed to release UI process: %v", err)
+	if err := launchCmd.Run(); err != nil {
+		return fmt.Errorf("run UI launch: %w", err)
 	}
 
 	log.Infof("netbird-ui started successfully for user %s", username)
 	return nil
 }
 
+func consoleUser() (string, error) {
+	output, err := exec.Command("stat", "-f", "%Su", "/dev/console").Output()
+	if err != nil {
+		return "", fmt.Errorf("get console user: %w", err)
+	}
+
+	username := strings.TrimSpace(string(output))
+	switch username {
+	case "", "root", "loginwindow", "_mbsetupuser":
+		return "", fmt.Errorf("no active GUI user session, console user: %q", username)
+	}
+
+	return username, nil
+}
+
 func (u *Installer) installPkgFile(ctx context.Context, path string) error {
 	log.Infof("installing pkg file: %s", path)
 

client/server/login_outcome_test.go

@@ -0,0 +1,110 @@
+package server
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal"
+	"github.com/netbirdio/netbird/client/proto"
+)
+
+// A login that never reached Management is not a decision about the peer's
+// credentials, so it must come back as a retryable error rather than an SSO
+// prompt: the user cannot finish a browser login while Management is down, and
+// the CLI's own backoff resolves the outage on its own once the daemon reports
+// the failure. Reproduces `netbird down; netbird up` printing a device-code URL
+// because Management happened to be restarting when the daemon dialed it.
+func TestLogin_ManagementUnreachableIsReturnedInsteadOfDemandingSSO(t *testing.T) {
+	s, _, _, username, _ := setupServerWithProfile(t)
+	s.rootCtx = internal.CtxInitState(context.Background())
+
+	unreachable := errors.New("create connection: dial context: context deadline exceeded")
+	attempts := 0
+	s.isLoginRequiredFn = func(context.Context) (bool, error) {
+		attempts++
+		return false, unreachable
+	}
+
+	resp, err := s.Login(userCtx(), &proto.LoginRequest{Username: &username})
+	require.Error(t, err)
+	require.ErrorIs(t, err, unreachable, "the transport failure was replaced by something else")
+	require.Nil(t, resp, "a failed login must not answer with a login response")
+	require.Equal(t, 1, attempts)
+	require.Nil(t, s.oauthAuthFlow.flow, "the daemon started an SSO flow for a peer whose login was never decided")
+
+	status, err := internal.CtxGetState(s.rootCtx).Status()
+	require.NoError(t, err)
+	require.Equal(t, internal.StatusLoginFailed, status,
+		"a peer that could not reach Management is not waiting on a login")
+}
+
+// The counterpart: Management refusing the peer's credentials is a decision, and
+// the SSO flow still has to start for it. The profile carries an unusable
+// private key so the flow setup fails immediately instead of dialing, which is
+// enough to show the branch was entered — the refusal itself is never what comes
+// back out.
+func TestLogin_AuthRefusalStartsSSOFlow(t *testing.T) {
+	s, _, _, username, cfgPath := setupServerWithProfile(t)
+	s.rootCtx = internal.CtxInitState(context.Background())
+	breakProfilePrivateKey(t, cfgPath)
+
+	s.isLoginRequiredFn = func(context.Context) (bool, error) {
+		return true, nil
+	}
+
+	_, err := s.Login(userCtx(), &proto.LoginRequest{Username: &username})
+	require.Error(t, err)
+
+	status, stateErr := internal.CtxGetState(s.rootCtx).Status()
+	require.NoError(t, stateErr)
+	require.Equal(t, internal.StatusLoginFailed, status,
+		"the SSO flow setup was never reached with the broken key")
+}
+
+func TestLogin_SetupKeyStillRunsWhenPeerNeedsLogin(t *testing.T) {
+	s, _, _, username, _ := setupServerWithProfile(t)
+	s.rootCtx = internal.CtxInitState(context.Background())
+
+	s.isLoginRequiredFn = func(context.Context) (bool, error) {
+		return true, nil
+	}
+
+	var keysTried []string
+	s.loginAttemptFn = func(_ context.Context, setupKey, _ string) (internal.StatusType, error) {
+		keysTried = append(keysTried, setupKey)
+		return "", nil
+	}
+
+	setupKey := "A2C8E32F-AEB2-4B45-8FD3-8A0C1B2D3E4F"
+	resp, err := s.Login(userCtx(), &proto.LoginRequest{Username: &username, SetupKey: setupKey})
+	require.NoError(t, err, "the probe's outcome leaked out as the login result")
+	require.NotNil(t, resp)
+	require.Equal(t, []string{setupKey}, keysTried, "the setup key never reached the login attempt")
+	require.Nil(t, s.oauthAuthFlow.flow, "a setup-key login started an SSO flow")
+
+	status, err := internal.CtxGetState(s.rootCtx).Status()
+	require.NoError(t, err)
+	require.Equal(t, internal.StatusIdle, status)
+}
+
+// breakProfilePrivateKey replaces the profile's private key with an unparseable
+// one, which makes any attempt to build a Management client fail on the spot.
+func breakProfilePrivateKey(t *testing.T, cfgPath string) {
+	t.Helper()
+
+	raw, err := os.ReadFile(cfgPath)
+	require.NoError(t, err)
+
+	var cfg map[string]any
+	require.NoError(t, json.Unmarshal(raw, &cfg))
+	cfg["PrivateKey"] = "not-a-key"
+
+	patched, err := json.Marshal(cfg)
+	require.NoError(t, err)
+	require.NoError(t, os.WriteFile(cfgPath, patched, 0o600))
+}

client/server/server.go

@@ -135,6 +135,13 @@ type Server struct {
 	updateManager *updater.Manager
 
 	jwtCache *jwtCache
+
+	// loginAttemptFn stands in for the Management login round trip. Tests set
+	// it to drive the login outcomes that need a server on the other end;
+	// production leaves it nil, and every login goes through loginAttempt.
+	loginAttemptFn func(ctx context.Context, setupKey, jwtToken string) (internal.StatusType, error)
+
+	isLoginRequiredFn func(ctx context.Context) (bool, error)
 }
 
 type oauthAuthFlow struct {
@@ -370,7 +377,34 @@ func (s *Server) connectionGoroutineRunning() bool {
 	}
 }
 
-// loginAttempt attempts to login using the provided information. it returns a status in case something fails
+// attemptLogin runs a login round trip against Management, or the stand-in a
+// test installed in place of it.
+func (s *Server) attemptLogin(ctx context.Context, setupKey, jwtToken string) (internal.StatusType, error) {
+	if s.loginAttemptFn != nil {
+		return s.loginAttemptFn(ctx, setupKey, jwtToken)
+	}
+	return s.loginAttempt(ctx, setupKey, jwtToken)
+}
+
+func (s *Server) isLoginRequired(ctx context.Context) (bool, error) {
+	if s.isLoginRequiredFn != nil {
+		return s.isLoginRequiredFn(ctx)
+	}
+
+	authClient, err := auth.NewAuth(ctx, s.config.PrivateKey, s.config.ManagementURL, s.config)
+	if err != nil {
+		log.Errorf("failed to create auth client: %v", err)
+		return false, err
+	}
+	defer authClient.Close()
+
+	return authClient.IsLoginRequired(ctx)
+}
+
+// loginAttempt attempts to login using the provided information. It returns
+// StatusNeedsLogin when Management refused the peer's credentials and
+// StatusLoginFailed for every other failure, so callers can tell an
+// authentication decision apart from a login that never got made.
 func (s *Server) loginAttempt(ctx context.Context, setupKey, jwtToken string) (internal.StatusType, error) {
 	authClient, err := auth.NewAuth(ctx, s.config.PrivateKey, s.config.ManagementURL, s.config)
 	if err != nil {
@@ -623,7 +657,19 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 	s.config = config
 	s.mutex.Unlock()
 
-	if _, err := s.loginAttempt(ctx, "", ""); err == nil {
+	// A probe that errors leaves the login undecided: Management unreachable, a
+	// restart mid-request, an internal error. Those are returned for the caller
+	// to retry, because turning them into an SSO prompt asks the user to solve
+	// something that is not theirs to solve, and a browser login cannot succeed
+	// while Management is unreachable anyway. Only Management refusing the
+	// peer's key is a decision, and IsLoginRequired reports that as
+	// needsLogin=true rather than an error.
+	needsLogin, err := s.isLoginRequired(ctx)
+	if err != nil {
+		state.Set(internal.StatusLoginFailed)
+		return nil, err
+	}
+	if !needsLogin {
 		state.Set(internal.StatusIdle)
 		return &proto.LoginResponse{}, nil
 	}
@@ -684,7 +730,7 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 	// which returns NeedsLogin and parks on the browser leg.
 	state.Set(internal.StatusConnecting)
 
-	if loginStatus, err := s.loginAttempt(ctx, msg.SetupKey, ""); err != nil {
+	if loginStatus, err := s.attemptLogin(ctx, msg.SetupKey, ""); err != nil {
 		state.Set(loginStatus)
 		return nil, err
 	}
@@ -839,7 +885,7 @@ func (s *Server) WaitSSOLogin(callerCtx context.Context, msg *proto.WaitSSOLogin
 	s.oauthAuthFlow.expiresAt = time.Now()
 	s.mutex.Unlock()
 
-	if loginStatus, err := s.loginAttempt(ctx, "", tokenInfo.GetTokenToUse()); err != nil {
+	if loginStatus, err := s.attemptLogin(ctx, "", tokenInfo.GetTokenToUse()); err != nil {
 		state.Set(loginStatus)
 		return nil, err
 	}
@@ -1769,6 +1815,9 @@ func (s *Server) RequestExtendAuthSession(
 	if connectClient == nil {
 		return nil, gstatus.Errorf(codes.FailedPrecondition, "client is not running")
 	}
+	if connectClient.Engine() == nil {
+		return nil, gstatus.Errorf(codes.FailedPrecondition, "session can no longer be extended, log in again to reconnect")
+	}
 
 	hint := ""
 	if msg.Hint != nil {

client/ui/build/linux/nfpm/nfpm.yaml

@@ -26,17 +26,17 @@ contents:
 
 # Default dependencies for the GTK4 + WebKitGTK 6.0 stack (Ubuntu 24.04+ / Debian 13+)
 depends:
-  - libgtk-4-1
+  - libgtk-4-1 (>= 4.14)
   - libwebkitgtk-6.0-4
   - xdg-utils
 
 # Distribution-specific overrides for different package formats
 overrides:
-  # RPM packages for Fedora / RHEL / AlmaLinux / Rocky Linux
+  # RPM packages for Fedora / RHEL / AlmaLinux / Rocky Linux / openSUSE
   rpm:
     depends:
-      - gtk4
-      - webkitgtk6.0
+      - (gtk4 >= 4.14 or libgtk-4-1 >= 4.14)
+      - (webkitgtk6.0 or libwebkitgtk-6_0-4)
       - xdg-utils
 
   # Arch Linux packages

client/ui/frontend/src/lib/connection.ts

@@ -43,7 +43,12 @@ function buildSsoCancelPromise(state: SsoState, signal?: AbortSignal): Promise<v
 }
 
 async function runSsoLogin(
-    result: { verificationUri: string; verificationUriComplete: string; userCode: string },
+    result: {
+        verificationUri: string;
+        verificationUriComplete: string;
+        userCode: string;
+        profileId: string;
+    },
     state: SsoState,
     signal?: AbortSignal,
 ): Promise<void> {
@@ -56,7 +61,7 @@ async function runSsoLogin(
     // suspended, so a frontend-driven Up (a promise continuation) would not
     // fire until the user woke the window (e.g. hovering the tray icon).
     const waitPromise = Connection.WaitSSOLoginAndUp(
-        { userCode: result.userCode, hostname: "" },
+        { userCode: result.userCode, hostname: "", profileId: result.profileId },
         { profileName: "", username: "" },
     );
 

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

@@ -11,7 +11,7 @@ import { DialogHeading } from "@/components/dialog/DialogHeading";
 import { SquareIcon } from "@/components/SquareIcon";
 import { Connection, Profiles as ProfilesSvc, Session, WindowManager } from "@bindings/services";
 import { useAutoSizeWindow } from "@/hooks/useAutoSizeWindow";
-import { EVENT_BROWSER_LOGIN_CANCEL } from "@/lib/connection";
+import { EVENT_BROWSER_LOGIN_CANCEL, EVENT_TRIGGER_LOGIN } from "@/lib/connection";
 import { errorDialog, formatErrorMessage } from "@/lib/errors.ts";
 import { formatRemaining } from "@/lib/formatters";
 
@@ -131,6 +131,21 @@ export default function SessionExpirationDialog() {
         }
     }, [busy, t]);
 
+    const authenticate = useCallback(async () => {
+        if (busy) return;
+        setBusy(true);
+        try {
+            await Events.Emit(EVENT_TRIGGER_LOGIN);
+            await WindowManager.CloseSessionExpiration();
+        } catch (e) {
+            setBusy(false);
+            await errorDialog({
+                Title: t("connect.error.loginTitle"),
+                Message: formatErrorMessage(e),
+            });
+        }
+    }, [busy, t]);
+
     const logout = useCallback(async () => {
         if (busy) return;
         setBusy(true);
@@ -185,7 +200,7 @@ export default function SessionExpirationDialog() {
                     variant={"primary"}
                     size={"md"}
                     className={"w-full"}
-                    onClick={stay}
+                    onClick={expired ? authenticate : stay}
                     disabled={busy}
                 >
                     {expired ? t("sessionExpiration.authenticate") : t("sessionExpiration.stay")}

client/ui/main.go

@@ -14,7 +14,6 @@ import (
 	"github.com/sirupsen/logrus"
 	"github.com/wailsapp/wails/v3/pkg/application"
 	"github.com/wailsapp/wails/v3/pkg/events"
-	"github.com/wailsapp/wails/v3/pkg/services/notifications"
 
 	"github.com/netbirdio/netbird/client/ui/authsession"
 	"github.com/netbirdio/netbird/client/ui/i18n"
@@ -63,7 +62,7 @@ type registeredServices struct {
 	profiles        *services.Profiles
 	update          *services.Update
 	daemonFeed      *services.DaemonFeed
-	notifier        *notifications.NotificationService
+	notifier        *Notifier
 	compat          *services.Compat
 	profileSwitcher *services.ProfileSwitcher
 	bundle          *i18n.Bundle
@@ -102,7 +101,7 @@ func main() {
 	updaterHolder := updater.NewHolder(app.Event)
 	update := services.NewUpdate(conn, updaterHolder)
 	daemonFeed := services.NewDaemonFeed(conn, app.Event, updaterHolder, debugLog)
-	notifier := notifications.New()
+	notifier := newNotifier()
 	compat := services.NewCompat(conn)
 	// macOS shows no toast until permission is requested. Run it after
 	// ApplicationStarted so the notifier's Startup has initialised the
@@ -210,7 +209,7 @@ func main() {
 // requestNotificationAuthorization prompts for macOS notification permission.
 // The request blocks until the user responds (up to 3 minutes), so callers run
 // it in a goroutine. No-op on Linux/Windows.
-func requestNotificationAuthorization(notifier *notifications.NotificationService) {
+func requestNotificationAuthorization(notifier *Notifier) {
 	authorized, err := notifier.CheckNotificationAuthorization()
 	if err != nil {
 		logrus.Debugf("check notification authorization: %v", err)

client/ui/notifier.go

@@ -0,0 +1,101 @@
+//go:build !android && !ios && !freebsd && !js
+
+package main
+
+import (
+	"context"
+	"errors"
+	"sync/atomic"
+
+	log "github.com/sirupsen/logrus"
+	"github.com/wailsapp/wails/v3/pkg/application"
+	"github.com/wailsapp/wails/v3/pkg/services/notifications"
+)
+
+var errNotificationsUnavailable = errors.New("notifications unavailable")
+
+// Notifier wraps the Wails notification service so an unavailable backend
+// disables notifications instead of aborting the app. Startup fails for
+// environment reasons (a bare unbundled binary on macOS has no bundle
+// identifier, a headless Linux session has no D-Bus session bus), and Wails
+// treats a service startup error as fatal. After a failed startup every call
+// is a no-op: on macOS, touching UNUserNotificationCenter without a bundle
+// identifier raises an Objective-C exception that recover() cannot catch.
+type Notifier struct {
+	inner     *notifications.NotificationService
+	available atomic.Bool
+}
+
+func newNotifier() *Notifier {
+	return &Notifier{inner: notifications.New()}
+}
+
+// ServiceName implements the Wails service-name hook for startup logs.
+func (n *Notifier) ServiceName() string {
+	return n.inner.ServiceName()
+}
+
+// ServiceStartup starts the platform notifier, downgrading failure to a
+// warning so the app keeps running without notifications.
+func (n *Notifier) ServiceStartup(ctx context.Context, options application.ServiceOptions) error {
+	if err := n.inner.ServiceStartup(ctx, options); err != nil {
+		log.Warnf("notifications disabled: %v", err)
+		return nil
+	}
+	n.available.Store(true)
+	return nil
+}
+
+func (n *Notifier) ServiceShutdown() error {
+	if !n.available.Load() {
+		return nil
+	}
+	return n.inner.ServiceShutdown()
+}
+
+func (n *Notifier) CheckNotificationAuthorization() (bool, error) {
+	if !n.available.Load() {
+		return false, errNotificationsUnavailable
+	}
+	return n.inner.CheckNotificationAuthorization()
+}
+
+func (n *Notifier) RequestNotificationAuthorization() (bool, error) {
+	if !n.available.Load() {
+		return false, errNotificationsUnavailable
+	}
+	return n.inner.RequestNotificationAuthorization()
+}
+
+// SendNotification delivers a notification, silently dropping it when the
+// backend never started (notifications are best-effort everywhere).
+func (n *Notifier) SendNotification(options notifications.NotificationOptions) error {
+	if !n.available.Load() {
+		log.Debugf("notifications disabled, dropping %q", options.ID)
+		return nil
+	}
+	return n.inner.SendNotification(options)
+}
+
+func (n *Notifier) SendNotificationWithActions(options notifications.NotificationOptions) error {
+	if !n.available.Load() {
+		log.Debugf("notifications disabled, dropping %q", options.ID)
+		return nil
+	}
+	return n.inner.SendNotificationWithActions(options)
+}
+
+func (n *Notifier) RegisterNotificationCategory(category notifications.NotificationCategory) error {
+	if !n.available.Load() {
+		return nil
+	}
+	return n.inner.RegisterNotificationCategory(category)
+}
+
+// OnNotificationResponse registers the response callback. Pure Go state, so
+// it is safe (and simply inert) when the backend never started.
+//
+//wails:ignore
+func (n *Notifier) OnNotificationResponse(callback func(result notifications.NotificationResult)) {
+	n.inner.OnNotificationResponse(callback)
+}

client/ui/preferences/store.go

@@ -246,6 +246,7 @@ func (s *Store) ExistedAtLoad() bool {
 func (s *Store) load() error {
 	if _, err := os.Stat(s.path); err != nil {
 		if errors.Is(err, os.ErrNotExist) {
+			log.Infof("no ui preferences file at %s; using defaults", s.path)
 			return nil
 		}
 		return fmt.Errorf("stat preferences: %w", err)

client/ui/services/connection.go

@@ -33,12 +33,21 @@ type LoginResult struct {
 	UserCode                string `json:"userCode"`
 	VerificationURI         string `json:"verificationUri"`
 	VerificationURIComplete string `json:"verificationUriComplete"`
+	// ProfileID is the ID of the profile this login ran against, or "" when the
+	// caller named the profile itself and no ID was resolved. Pass it back in
+	// WaitSSOParams so the account email lands on this profile even if the
+	// active one changes during SSO.
+	ProfileID string `json:"profileId"`
 }
 
 // WaitSSOParams are the inputs to waitSSOLogin.
 type WaitSSOParams struct {
 	UserCode string `json:"userCode"`
 	Hostname string `json:"hostname"`
+	// ProfileID is the profile the login was started for, used to file the
+	// account email against it rather than against whichever profile is active
+	// when the flow returns. Optional: empty falls back to the active profile.
+	ProfileID string `json:"profileId"`
 }
 
 // UpParams selects the profile to bring up.
@@ -77,11 +86,16 @@ func (s *Connection) Login(ctx context.Context, p LoginParams) (LoginResult, err
 	// Fall back to the daemon's active profile and the current OS user.
 	profileName := p.ProfileName
 	username := p.Username
+	// Only set when the daemon told us the ID. A caller-supplied ProfileName is
+	// a handle — a display name or an ID prefix resolve too — and the state file
+	// is named after the ID, so passing a handle on would name the wrong file.
+	profileID := ""
 	if profileName == "" {
 		if active, aerr := cli.GetActiveProfile(ctx, &proto.GetActiveProfileRequest{}); aerr == nil {
 			// Address the active profile by ID (the daemon resolves it as a
 			// handle); names can collide, the ID cannot.
 			profileName = active.GetId()
+			profileID = profileName
 			if username == "" {
 				username = active.GetUsername()
 			}
@@ -122,6 +136,7 @@ func (s *Connection) Login(ctx context.Context, p LoginParams) (LoginResult, err
 		UserCode:                resp.GetUserCode(),
 		VerificationURI:         resp.GetVerificationURI(),
 		VerificationURIComplete: resp.GetVerificationURIComplete(),
+		ProfileID:               profileID,
 	}, nil
 }
 
@@ -242,6 +257,31 @@ func (s *Connection) waitSSOLogin(ctx context.Context, p WaitSSOParams) (string,
 		return "", s.classifyDaemonError(err)
 	}
 	log.Infof("SSO login completed, daemon reported success")
+
+	// Persist the account email the same way the CLI does after its own
+	// WaitSSOLogin: the daemon returns it but cannot store it, since it runs as
+	// root and the per-profile state file is user-owned (see Logout below).
+	// Without this the profile has no email, so Profiles.List shows no account
+	// and later logins and session extends go out without a login_hint —
+	// leaving the IdP to guess which account was meant.
+	if email := resp.GetEmail(); email != "" {
+		state := &profilemanager.ProfileState{Email: email}
+		pm := profilemanager.NewProfileManager()
+
+		// Against the profile the login was started for: SSO spans seconds of
+		// user interaction, and a profile switch in that window would otherwise
+		// file the email under the wrong profile.
+		if p.ProfileID != "" {
+			err = pm.SetProfileState(profilemanager.ID(p.ProfileID), state)
+		} else {
+			err = pm.SetActiveProfileState(state)
+		}
+		if err != nil {
+			// Non-fatal: the login itself succeeded.
+			log.Warnf("failed to store account email: %v", err)
+		}
+	}
+
 	return resp.GetEmail(), nil
 }
 

client/ui/services/profile.go

@@ -6,6 +6,8 @@ import (
 	"context"
 	"os/user"
 
+	log "github.com/sirupsen/logrus"
+
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/proto"
 )
@@ -151,11 +153,31 @@ func (s *Profiles) Remove(ctx context.Context, p ProfileRef) error {
 	if err != nil {
 		return err
 	}
-	_, err = cli.RemoveProfile(ctx, &proto.RemoveProfileRequest{
+	resp, err := cli.RemoveProfile(ctx, &proto.RemoveProfileRequest{
 		ProfileName: p.ProfileName,
 		Username:    p.Username,
 	})
-	return err
+	if err != nil {
+		return err
+	}
+
+	// The daemon deletes what it owns but runs as root, so it leaves the
+	// user-owned state file holding the account email behind (same split as
+	// Connection.Logout). Legacy profiles are keyed by name rather than by a
+	// generated ID, so a recreated profile of the same name would inherit the
+	// deleted one's email and offer it as the login_hint.
+	//
+	// Keyed on the ID the daemon resolved, not on the request handle: that may
+	// have been a display name or an ID prefix, which would name a different
+	// file (or none).
+	if id := resp.GetId(); id != "" {
+		if err := profilemanager.NewProfileManager().RemoveProfileState(id); err != nil {
+			// Non-fatal: the profile itself is gone.
+			log.Warnf("failed to remove profile state for %s: %v", id, err)
+		}
+	}
+
+	return nil
 }
 
 // Rename changes a profile's display name. The on-disk ID is unaffected, so

client/ui/tray.go

@@ -44,7 +44,7 @@ type TrayServices struct {
 	Profiles        *services.Profiles
 	Networks        *services.Networks
 	DaemonFeed      *services.DaemonFeed
-	Notifier        *notifications.NotificationService
+	Notifier        *Notifier
 	Update          *services.Update
 	ProfileSwitcher *services.ProfileSwitcher
 	WindowManager   *services.WindowManager

client/ui/tray_click_linux.go

@@ -4,17 +4,26 @@ package main
 
 // bindTrayClick wires the tray icon's left-click handler on Linux.
 //
-// Both Linux click paths converge on Wails' linuxSystemTray.Activate, which
-// fires the registered clickHandler:
-//   - Real SNI hosts (KDE Plasma, Waybar, GNOME Shell + AppIndicator) invoke
-//     org.kde.StatusNotifierItem.Activate over D-Bus on left-click.
-//   - The in-process StatusNotifierWatcher + XEmbed host used on minimal WMs
-//     (Fluxbox, i3, dwm, OpenBox) maps a Button1 press to that same Activate
-//     call itself (xembed_host_linux.go), so it routes through the same hook.
-// Registering OnClick here therefore covers both paths with one handler — no
-// changes to the watcher or XEmbed C code are needed. Left-click now opens the
-// main window; right-click still opens the menu via Wails' default
-// SecondaryActivate→OpenMenu handler (and the XEmbed GTK popup on minimal WMs).
+// Expected behaviour per tray host:
+//
+//	Host                           Left click              Right click
+//	KDE Plasma, Waybar             main window (Activate)  menu (host-rendered)
+//	GNOME Shell + AppIndicator     menu only               menu only
+//	Minimal WMs via XEmbed host    main window (Activate)  XEmbed GTK popup
+//
+// OnClick fires only on org.kde.StatusNotifierItem.Activate — a real left
+// click. KDE/Waybar send it over D-Bus; the in-process XEmbed host
+// (xembed_host_linux.go) maps a Button1 press to the same Activate call.
+//
+// GNOME Shell + AppIndicator never sends Activate: it renders the dbusmenu
+// on ANY click and only reports the menu opening via dbusmenu
+// Event("opened"). Upstream Wails treated that event as a click, so on GNOME
+// both buttons raised the main window on top of the menu, and on KDE/Waybar
+// a right click raised it over the freshly opened menu. The netbirdio/wails
+// fork (go.mod replace) drops that heuristic: a menu open never fires
+// OnClick. On GNOME the main window is reached via the "Open NetBird" menu
+// entry; left-click-opens-window is not achievable there anyway, since the
+// host always opens the menu itself.
 //
 // We do NOT register OnDoubleClick: Wails' Linux SNI backend never fires it
 // (unlike Windows). And we deliberately skip AttachWindow — it plus Wails3's

client/ui/tray_notify.go

@@ -44,7 +44,7 @@ func safeSendNotification(send sendFn, what string, opts notifications.Notificat
 // notifyIfDaemonOutdated probes the daemon once and fires an OS toast when it
 // is reachable but too old for this UI. A probe error means the daemon isn't
 // reachable (not outdated), so it is left to the normal connection flow.
-func notifyIfDaemonOutdated(compat *services.Compat, notifier *notifications.NotificationService, loc *Localizer) {
+func notifyIfDaemonOutdated(compat *services.Compat, notifier *Notifier, loc *Localizer) {
 	ready, err := compat.DaemonReady(context.Background())
 	if err != nil {
 		log.Debugf("daemon compatibility probe: %v", err)

client/ui/tray_session.go

@@ -27,11 +27,10 @@ const (
 	finalWarningCountdownSeconds = 120
 )
 
-// handleSessionExpired notifies and brings the window forward so the frontend's /login route drives renewal.
+// handleSessionExpired notifies and brings the window forward so the user can reconnect.
 func (t *Tray) handleSessionExpired() {
 	t.notify(t.loc.T("notify.sessionExpired.title"), t.loc.T("notify.sessionExpired.body"), notifyIDSessionExpired)
 	if t.window != nil {
-		t.window.SetURL("/#/login")
 		t.window.Show()
 		t.window.Focus()
 	}
@@ -308,11 +307,7 @@ func (t *Tray) openSessionExtendFlow() {
 	}
 	seconds := int(time.Until(deadline).Seconds())
 	if seconds <= 0 {
-		if t.window != nil {
-			t.window.SetURL("/#/login")
-			t.window.Show()
-			t.window.Focus()
-		}
+		t.app.Event.Emit(services.EventTriggerLogin)
 		return
 	}
 	if t.svc.WindowManager == nil {

client/ui/tray_update.go

@@ -21,7 +21,7 @@ type trayUpdater struct {
 	app          *application.App
 	window       *application.WebviewWindow
 	update       *services.Update
-	notifier     *notifications.NotificationService
+	notifier     *Notifier
 	loc          *Localizer
 	onIconChange func()
 	// onMenuChange drives a full tray relayout: the update row lives in the
@@ -36,7 +36,7 @@ type trayUpdater struct {
 	progressWindowOpen bool
 }
 
-func newTrayUpdater(app *application.App, window *application.WebviewWindow, update *services.Update, notifier *notifications.NotificationService, loc *Localizer, onIconChange func(), onMenuChange func()) *trayUpdater {
+func newTrayUpdater(app *application.App, window *application.WebviewWindow, update *services.Update, notifier *Notifier, loc *Localizer, onIconChange func(), onMenuChange func()) *trayUpdater {
 	u := &trayUpdater{
 		app:          app,
 		window:       window,

e2e/agentnetwork/management_test.go

@@ -160,8 +160,19 @@ func TestSettingsRoundTrip(t *testing.T) {
 	assert.Equal(t, before.Cluster, flipped.Cluster, "cluster must be immutable across updates")
 	assert.Equal(t, before.Subdomain, flipped.Subdomain, "subdomain must be immutable across updates")
 
+	// A cluster different from the pinned one must be rejected; echoing the
+	// pinned one back is valid.
+	_, err = srv.UpdateSettings(ctx, api.AgentNetworkSettingsRequest{
+		Cluster:                ptr("attacker.cluster.invalid"),
+		EnableLogCollection:    before.EnableLogCollection,
+		EnablePromptCollection: before.EnablePromptCollection,
+		RedactPii:              before.RedactPii,
+	})
+	requireClientError(t, err)
+
 	// Restore the original toggles.
 	_, err = srv.UpdateSettings(ctx, api.AgentNetworkSettingsRequest{
+		Cluster:                ptr(before.Cluster),
 		EnableLogCollection:    before.EnableLogCollection,
 		EnablePromptCollection: before.EnablePromptCollection,
 		RedactPii:              before.RedactPii,

e2e/agentnetwork/settings_bootstrap_test.go

@@ -0,0 +1,114 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// harnessStartFresh boots a dedicated combined server with its own fresh
+// account and registers its teardown on t.
+func harnessStartFresh(ctx context.Context, t *testing.T) (*harness.Combined, error) {
+	t.Helper()
+	fresh, err := harness.StartCombined(ctx)
+	if err != nil {
+		return nil, err
+	}
+	t.Cleanup(func() { _ = fresh.Terminate(context.Background()) })
+	if _, err := fresh.Bootstrap(ctx); err != nil {
+		return nil, err
+	}
+	return fresh, nil
+}
+
+// TestSettingsBootstrapViaPut covers the settings-first bootstrap path on an
+// account that has never been bootstrapped: the GET reads as the defaults
+// with an empty cluster/subdomain/endpoint, a PUT without a cluster has
+// nothing to pin and fails, and a PUT carrying a cluster creates the row and
+// pins it immutably. The shared srv cannot provide that starting state (any
+// provider-creating test bootstraps it, and test order is deliberately not
+// relied on), so this boots a dedicated combined server — the image is
+// already built and cached by TestMain's StartCombined, so the extra cost is
+// one container start.
+func TestSettingsBootstrapViaPut(t *testing.T) {
+	ctx := context.Background()
+
+	fresh, err := harnessStartFresh(ctx, t)
+	require.NoError(t, err, "start dedicated combined server")
+
+	// Before agent-network bootstrap the settings read as the defaults, not
+	// as an error and not as a null body.
+	before, err := fresh.GetSettings(ctx)
+	require.NoError(t, err, "get settings on a fresh account must succeed")
+	assert.Empty(t, before.Cluster, "cluster must be empty before bootstrap")
+	assert.Empty(t, before.Subdomain, "subdomain must be empty before bootstrap")
+	assert.Empty(t, before.Endpoint, "endpoint must be empty before bootstrap, not a bare dot")
+	assert.True(t, before.EnableLogCollection, "defaults must show log collection on, matching bootstrap")
+	assert.False(t, before.EnablePromptCollection, "defaults must show prompt collection off")
+
+	// A PUT without a cluster has nothing to pin the account to.
+	_, err = fresh.UpdateSettings(ctx, api.AgentNetworkSettingsRequest{
+		EnableLogCollection: true,
+	})
+	requireClientError(t, err)
+
+	// A PUT carrying a cluster bootstraps the account and applies the
+	// mutable fields from the same request. Every toggle is set away from
+	// its bootstrap default so each assertion can actually fail.
+	const cluster = "e2e.bootstrap.netbird.selfhosted"
+	bootstrapped, err := fresh.UpdateSettings(ctx, api.AgentNetworkSettingsRequest{
+		Cluster:                ptr(cluster),
+		EnableLogCollection:    false,
+		EnablePromptCollection: true,
+		RedactPii:              true,
+	})
+	require.NoError(t, err, "bootstrap settings via PUT must succeed")
+	assert.Equal(t, cluster, bootstrapped.Cluster, "cluster must be pinned from the request")
+	require.NotEmpty(t, bootstrapped.Subdomain, "subdomain must be assigned at bootstrap")
+	assert.Equal(t, bootstrapped.Subdomain+"."+cluster, bootstrapped.Endpoint, "endpoint must combine subdomain and cluster")
+	assert.False(t, bootstrapped.EnableLogCollection, "log collection from the bootstrap request must override the default")
+	assert.True(t, bootstrapped.EnablePromptCollection, "prompt collection from the bootstrap request must apply")
+	assert.True(t, bootstrapped.RedactPii, "redact toggle from the bootstrap request must apply")
+
+	// The row is persisted: an independent read agrees on every field.
+	after, err := fresh.GetSettings(ctx)
+	require.NoError(t, err, "get settings after bootstrap must succeed")
+	assert.Equal(t, bootstrapped.Endpoint, after.Endpoint, "bootstrap must persist across reads")
+	assert.Equal(t, bootstrapped.EnableLogCollection, after.EnableLogCollection, "log collection must persist")
+	assert.Equal(t, bootstrapped.EnablePromptCollection, after.EnablePromptCollection, "prompt collection must persist")
+	assert.Equal(t, bootstrapped.RedactPii, after.RedactPii, "redact toggle must persist")
+
+	// Once bootstrapped, later updates may omit the cluster entirely.
+	persisted, err := fresh.UpdateSettings(ctx, api.AgentNetworkSettingsRequest{
+		EnableLogCollection:    true,
+		EnablePromptCollection: false,
+		RedactPii:              true,
+	})
+	require.NoError(t, err, "post-bootstrap update without cluster must succeed")
+	assert.Equal(t, cluster, persisted.Cluster, "omitted cluster must keep the pinned value")
+	assert.True(t, persisted.EnableLogCollection, "post-bootstrap toggle must apply")
+	assert.False(t, persisted.EnablePromptCollection, "post-bootstrap toggle must apply")
+
+	// The cluster is immutable: a different value is rejected rather than
+	// silently ignored, and the rejected update must not disturb anything.
+	_, err = fresh.UpdateSettings(ctx, api.AgentNetworkSettingsRequest{
+		Cluster:             ptr("other.cluster.invalid"),
+		EnableLogCollection: false,
+	})
+	requireClientError(t, err)
+
+	final, err := fresh.GetSettings(ctx)
+	require.NoError(t, err, "get settings after the rejected cluster change must succeed")
+	assert.Equal(t, persisted.Cluster, final.Cluster, "rejected update must not change the cluster")
+	assert.Equal(t, persisted.Endpoint, final.Endpoint, "rejected update must not change the endpoint")
+	assert.Equal(t, persisted.EnableLogCollection, final.EnableLogCollection, "rejected update must not apply its toggles")
+	assert.Equal(t, persisted.EnablePromptCollection, final.EnablePromptCollection, "rejected update must not apply its toggles")
+	assert.Equal(t, persisted.RedactPii, final.RedactPii, "rejected update must not apply its toggles")
+}

go.mod

@@ -114,7 +114,7 @@ require (
 	github.com/ti-mo/conntrack v0.5.1
 	github.com/ti-mo/netfilter v0.5.2
 	github.com/vmihailenco/msgpack/v5 v5.4.1
-	github.com/wailsapp/wails/v3 v3.0.0-alpha2.117
+	github.com/wailsapp/wails/v3 v3.0.0-beta.3
 	github.com/yusufpapurcu/wmi v1.2.4
 	github.com/zcalusic/sysinfo v1.1.3
 	go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.67.0
@@ -339,3 +339,5 @@ replace github.com/dexidp/dex => github.com/netbirdio/dex v0.244.1-0.20260716205
 replace github.com/dexidp/dex/api/v2 => github.com/netbirdio/dex/api/v2 v2.0.0-20260512110716-8d70ad8647c1
 
 replace github.com/mailru/easyjson => github.com/netbirdio/easyjson v0.9.0
+
+replace github.com/wailsapp/wails/v3 => github.com/netbirdio/wails/v3 v3.0.0-beta.3.0.20260803205919-ad21e92381f4

go.sum

@@ -490,6 +490,8 @@ github.com/netbirdio/service v0.0.0-20240911161631-f62744f42502 h1:3tHlFmhTdX9ax
 github.com/netbirdio/service v0.0.0-20240911161631-f62744f42502/go.mod h1:CIMRFEJVL+0DS1a3Nx06NaMn4Dz63Ng6O7dl0qH0zVM=
 github.com/netbirdio/signal-dispatcher/dispatcher v0.0.0-20250805121659-6b4ac470ca45 h1:ujgviVYmx243Ksy7NdSwrdGPSRNE3pb8kEDSpH0QuAQ=
 github.com/netbirdio/signal-dispatcher/dispatcher v0.0.0-20250805121659-6b4ac470ca45/go.mod h1:5/sjFmLb8O96B5737VCqhHyGRzNFIaN/Bu7ZodXc3qQ=
+github.com/netbirdio/wails/v3 v3.0.0-beta.3.0.20260803205919-ad21e92381f4 h1:UKztc3QjWvzU5DZk+uYaOWN0x62NSe/pkxuPvzqZIy4=
+github.com/netbirdio/wails/v3 v3.0.0-beta.3.0.20260803205919-ad21e92381f4/go.mod h1:BzATbK71VFikMMMCo434wAi0QcaI03P+xeaWgDvQvjw=
 github.com/netbirdio/wireguard-go v0.0.0-20260628102922-2834bebf6c1a h1:3CWK+yTvRKOcC0Q8VCTGy4l60TEb27CQVS7LkMxwjmw=
 github.com/netbirdio/wireguard-go v0.0.0-20260628102922-2834bebf6c1a/go.mod h1:rpwXGsirqLqN2L0JDJQlwOboGHmptD5ZD6T2VmcqhTw=
 github.com/nxadm/tail v1.4.4/go.mod h1:kenIhsEOeOJmVchQTgglprH7qJGnHDVpk1VPCcaMI8A=
@@ -660,8 +662,6 @@ github.com/vmihailenco/msgpack/v5 v5.4.1 h1:cQriyiUvjTwOHg8QZaPihLWeRAAVoCpE00IU
 github.com/vmihailenco/msgpack/v5 v5.4.1/go.mod h1:GaZTsDaehaPpQVyxrf5mtQlH+pc21PIudVV/E3rRQok=
 github.com/vmihailenco/tagparser/v2 v2.0.0 h1:y09buUbR+b5aycVFQs/g70pqKVZNBmxwAhO7/IwNM9g=
 github.com/vmihailenco/tagparser/v2 v2.0.0/go.mod h1:Wri+At7QHww0WTrCBeu4J6bNtoV6mEfg5OIWRZA9qds=
-github.com/wailsapp/wails/v3 v3.0.0-alpha2.117 h1:udyjqPG3AIgkod5QDR/WblCkpV8R86BFPSrsWxSyt5Y=
-github.com/wailsapp/wails/v3 v3.0.0-alpha2.117/go.mod h1:74WH2FScMsgucZvHHvv7eOefDXCm/CjuIxqhhZgPhKg=
 github.com/wlynxg/anet v0.0.5 h1:J3VJGi1gvo0JwZ/P1/Yc/8p63SoW98B5dHkYDmpgvvU=
 github.com/wlynxg/anet v0.0.5/go.mod h1:eay5PRQr7fIVAMbTbchTnO9gG65Hg/uYGdc7mguHxoA=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=

management/internals/modules/agentnetwork/handlers/handlers_test.go

@@ -17,6 +17,7 @@ import (
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
 	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	"github.com/netbirdio/netbird/management/server/account"
 	nbcontext "github.com/netbirdio/netbird/management/server/context"
 	"github.com/netbirdio/netbird/management/server/permissions"
 	"github.com/netbirdio/netbird/management/server/store"
@@ -61,10 +62,23 @@ func newAgentNetworkHandlerFixture(t *testing.T) *agentNetworkHandlerFixture {
 		Return(true, context.Background(), nil).
 		AnyTimes()
 
-	manager := agentnetwork.NewManager(st, perms, nil, nil)
+	// Swallow activity events so the mutation paths (create/update/delete)
+	// are exercisable through the HTTP layer.
+	accounts := account.NewMockManager(ctrl)
+	accounts.EXPECT().
+		StoreEvent(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any()).
+		AnyTimes()
+	accounts.EXPECT().
+		UpdateAccountPeers(gomock.Any(), gomock.Any(), gomock.Any()).
+		AnyTimes()
+
+	manager := agentnetwork.NewManager(st, perms, accounts, nil)
 	h := &handler{manager: manager}
 
 	router := mux.NewRouter()
+	router.HandleFunc("/agent-network/providers", h.createProvider).Methods("POST")
+	router.HandleFunc("/agent-network/providers/{providerId}", h.getProvider).Methods("GET")
+	router.HandleFunc("/agent-network/providers/{providerId}", h.updateProvider).Methods("PUT")
 	h.addPolicyEndpoints(router)
 	h.addConsumptionEndpoints(router)
 	h.addBudgetRuleEndpoints(router)

management/internals/modules/agentnetwork/handlers/providers_handler_test.go

@@ -1,7 +1,9 @@
 package handlers
 
 import (
+	"encoding/json"
 	"math"
+	nethttp "net/http"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -51,3 +53,50 @@ func TestValidate_ModelRates(t *testing.T) {
 		assert.Error(t, validate(base(m), true), "case %q must be rejected", name)
 	}
 }
+
+// TestProviderHandler_UpdateReplacesFullState pins the update contract shared
+// with the other PUT endpoints: the request replaces the provider's mutable
+// state, so optional fields absent from the JSON land as their zero values.
+// The two exceptions are server-side: the api_key (a secret — omitted means
+// "not rotated") and the session keypair, both preserved by the manager. The
+// identity headers stay on the wire as explicit empty strings so a cleared
+// value round-trips.
+func TestProviderHandler_UpdateReplacesFullState(t *testing.T) {
+	f := newAgentNetworkHandlerFixture(t)
+
+	create := `{
+        "provider_id": "openai_api",
+        "name": "openai",
+        "upstream_url": "https://api.openai.com",
+        "api_key": "sk-test",
+        "enabled": true,
+        "metadata_disabled": true,
+        "skip_tls_verification": true,
+        "extra_values": {"x-portkey-config": "pc-prod-3f2a"},
+        "identity_header_user_id": "x-bf-dim-netbird_user_id",
+        "models": [{"id": "gpt-4o", "input_per_1k": 0.0025, "output_per_1k": 0.01}]
+    }`
+	rec := f.do(t, nethttp.MethodPost, "/agent-network/providers", create)
+	require.Equal(t, nethttp.StatusOK, rec.Code, "create must succeed: %s", rec.Body.String())
+
+	var created api.AgentNetworkProvider
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &created))
+
+	// Minimal update: only the required fields, no api_key. Everything
+	// optional must land as its zero value.
+	update := `{"provider_id": "openai_api", "name": "openai-renamed", "upstream_url": "https://api.openai.com", "enabled": true}`
+	rec = f.do(t, nethttp.MethodPut, "/agent-network/providers/"+created.Id, update)
+	require.Equal(t, nethttp.StatusOK, rec.Code, "update without api_key must succeed (key is preserved): %s", rec.Body.String())
+
+	var updated api.AgentNetworkProvider
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &updated))
+	assert.Equal(t, "openai-renamed", updated.Name, "sent field must apply")
+	assert.True(t, updated.Enabled, "sent field must apply")
+	assert.False(t, updated.MetadataDisabled, "omitted metadata_disabled must land as false — PUT replaces the full state")
+	assert.False(t, updated.SkipTlsVerification, "omitted skip_tls_verification must land as false")
+	assert.Nil(t, updated.ExtraValues, "omitted extra_values must be cleared")
+	assert.Equal(t, "", updated.IdentityHeaderUserId, "omitted identity header must be cleared yet stay on the wire")
+	assert.Empty(t, updated.Models, "omitted models must be cleared")
+	assert.Contains(t, rec.Body.String(), `"identity_header_user_id":""`,
+		"cleared identity header must round-trip as an explicit empty string")
+}

management/internals/modules/agentnetwork/handlers/settings_handler.go

@@ -2,7 +2,6 @@ package handlers
 
 import (
 	"encoding/json"
-	"errors"
 	"net/http"
 
 	"github.com/gorilla/mux"
@@ -11,19 +10,20 @@ import (
 	nbcontext "github.com/netbirdio/netbird/management/server/context"
 	"github.com/netbirdio/netbird/shared/management/http/api"
 	"github.com/netbirdio/netbird/shared/management/http/util"
-	"github.com/netbirdio/netbird/shared/management/status"
 )
 
 // addSettingsEndpoints registers the Agent Network settings routes. The
-// settings row is bootstrapped server-side on first provider create; GET reads
-// it and PUT updates the mutable collection toggles (cluster/subdomain stay
-// immutable).
+// settings row is bootstrapped server-side on first provider create or on the
+// first PUT carrying a cluster; GET reads it and PUT applies a partial update
+// of the mutable collection toggles (cluster/subdomain stay immutable).
 func (h *handler) addSettingsEndpoints(router *mux.Router) {
 	router.HandleFunc("/agent-network/settings", h.getSettings).Methods("GET", "OPTIONS")
 	router.HandleFunc("/agent-network/settings", h.updateSettings).Methods("PUT", "OPTIONS")
 }
 
-// updateSettings applies the collection toggles to the account's settings row.
+// updateSettings replaces the mutable settings fields on the account's row.
+// A request carrying a cluster bootstraps the row when the account doesn't
+// have one yet.
 func (h *handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 	userAuth, err := nbcontext.GetUserAuthFromContext(r.Context())
 	if err != nil {
@@ -48,11 +48,9 @@ func (h *handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 	util.WriteJSONObject(r.Context(), w, updated.ToAPIResponse())
 }
 
-// getSettings returns the account's agent-network settings. The settings
-// row is bootstrapped on first provider create, so freshly-onboarded
-// accounts have nothing to read. Rather than 404-ing in that case (which
-// the dashboard would have to special-case), return a JSON null with 200
-// so consumers can branch on the body alone.
+// getSettings returns the account's agent-network settings. Accounts that
+// haven't been bootstrapped yet read as the defaults with an empty cluster,
+// subdomain and endpoint; the manager synthesises that view.
 func (h *handler) getSettings(w http.ResponseWriter, r *http.Request) {
 	userAuth, err := nbcontext.GetUserAuthFromContext(r.Context())
 	if err != nil {
@@ -62,11 +60,6 @@ func (h *handler) getSettings(w http.ResponseWriter, r *http.Request) {
 
 	settings, err := h.manager.GetSettings(r.Context(), userAuth.AccountId, userAuth.UserId)
 	if err != nil {
-		var sErr *status.Error
-		if errors.As(err, &sErr) && sErr.Type() == status.NotFound {
-			util.WriteJSONObject(r.Context(), w, nil)
-			return
-		}
 		util.WriteError(r.Context(), err, w)
 		return
 	}

management/internals/modules/agentnetwork/handlers/settings_handler_test.go

@@ -0,0 +1,137 @@
+package handlers
+
+import (
+	"encoding/json"
+	"net/http"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestSettingsHandler_GetUnbootstrappedReturnsDefaults pins the settings-read
+// convention shared with the account and DNS settings endpoints: settings
+// always read as a JSON object. Before bootstrap that object carries the
+// defaults with an empty cluster/subdomain/endpoint (the "not bootstrapped"
+// signal) and no timestamps — never a 404 and never the legacy null body.
+func TestSettingsHandler_GetUnbootstrappedReturnsDefaults(t *testing.T) {
+	f := newAgentNetworkHandlerFixture(t)
+
+	rec := f.do(t, http.MethodGet, "/agent-network/settings", "")
+	require.Equal(t, http.StatusOK, rec.Code,
+		"unbootstrapped account must read as 200 with defaults: got %d body=%s", rec.Code, rec.Body.String())
+	require.NotEqual(t, "null", trimSpace(rec.Body.String()),
+		"the legacy 200+null shape must not come back")
+
+	var got api.AgentNetworkSettings
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &got))
+	assert.Empty(t, got.Cluster, "cluster must be empty until bootstrapped")
+	assert.Empty(t, got.Subdomain, "subdomain must be empty until bootstrapped")
+	assert.Empty(t, got.Endpoint, "endpoint must be empty until bootstrapped, not a bare dot")
+	assert.True(t, got.EnableLogCollection, "defaults must show log collection on, matching bootstrap")
+	assert.False(t, got.EnablePromptCollection, "defaults must show prompt collection off")
+	assert.False(t, got.RedactPii, "defaults must show redaction off")
+	require.NotNil(t, got.AccessLogRetentionDays)
+	assert.Equal(t, 30, *got.AccessLogRetentionDays, "defaults must show the bootstrap retention")
+	assert.Nil(t, got.CreatedAt, "no timestamps before a row exists")
+	assert.Nil(t, got.UpdatedAt, "no timestamps before a row exists")
+}
+
+// TestSettingsHandler_PutBootstrapsWithCluster covers the settings-first
+// bootstrap path: a PUT carrying a cluster on an unbootstrapped account
+// creates the row (cluster pinned, subdomain assigned) and applies the
+// mutable fields from the same request.
+func TestSettingsHandler_PutBootstrapsWithCluster(t *testing.T) {
+	f := newAgentNetworkHandlerFixture(t)
+
+	rec := f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"cluster": "eu.proxy.netbird.io", "enable_log_collection": true, "enable_prompt_collection": true, "redact_pii": false, "access_log_retention_days": 30}`)
+	require.Equal(t, http.StatusOK, rec.Code, "bootstrap PUT must succeed: %s", rec.Body.String())
+
+	var got api.AgentNetworkSettings
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &got))
+	assert.Equal(t, "eu.proxy.netbird.io", got.Cluster, "cluster must be pinned from the request")
+	assert.NotEmpty(t, got.Subdomain, "subdomain must be assigned at bootstrap")
+	assert.Equal(t, got.Subdomain+".eu.proxy.netbird.io", got.Endpoint, "endpoint must combine subdomain and cluster")
+	assert.True(t, got.EnableLogCollection, "toggle from the bootstrap request must apply")
+	assert.True(t, got.EnablePromptCollection, "toggle from the bootstrap request must apply")
+	require.NotNil(t, got.AccessLogRetentionDays)
+	assert.Equal(t, 30, *got.AccessLogRetentionDays, "retention from the bootstrap request must apply")
+
+	// The row is now readable via GET.
+	rec = f.do(t, http.MethodGet, "/agent-network/settings", "")
+	require.Equal(t, http.StatusOK, rec.Code, "GET after bootstrap must succeed")
+}
+
+// TestSettingsHandler_PutWithoutClusterOnUnbootstrapped pins that a PUT
+// without a cluster cannot conjure a settings row out of nothing — there is
+// no cluster to pin — and surfaces as 404 like the GET.
+func TestSettingsHandler_PutWithoutClusterOnUnbootstrapped(t *testing.T) {
+	f := newAgentNetworkHandlerFixture(t)
+
+	rec := f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"enable_log_collection": false, "enable_prompt_collection": false, "redact_pii": false}`)
+	assert.Equal(t, http.StatusNotFound, rec.Code,
+		"cluster-less PUT on an unbootstrapped account must 404: got %d body=%s", rec.Code, rec.Body.String())
+	assert.Contains(t, rec.Body.String(), "cluster",
+		"the error must point the caller at the bootstrap paths: %s", rec.Body.String())
+}
+
+// TestSettingsHandler_PutReplacesMutableFields pins the update contract shared
+// with the other PUT endpoints: the request replaces every mutable field, so a
+// toggle absent from the JSON lands as its zero value rather than being
+// preserved. Cluster and subdomain survive untouched.
+func TestSettingsHandler_PutReplacesMutableFields(t *testing.T) {
+	f := newAgentNetworkHandlerFixture(t)
+
+	rec := f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"cluster": "eu.proxy.netbird.io", "enable_log_collection": true, "enable_prompt_collection": true, "redact_pii": true, "access_log_retention_days": 14}`)
+	require.Equal(t, http.StatusOK, rec.Code, "bootstrap PUT must succeed: %s", rec.Body.String())
+
+	var before api.AgentNetworkSettings
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &before))
+
+	rec = f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"enable_log_collection": true, "enable_prompt_collection": false, "redact_pii": false}`)
+	require.Equal(t, http.StatusOK, rec.Code, "update PUT must succeed: %s", rec.Body.String())
+
+	var got api.AgentNetworkSettings
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &got))
+	assert.True(t, got.EnableLogCollection, "sent toggle must apply")
+	assert.False(t, got.EnablePromptCollection, "sent toggle must apply")
+	assert.False(t, got.RedactPii, "sent toggle must apply")
+	require.NotNil(t, got.AccessLogRetentionDays)
+	assert.Equal(t, 0, *got.AccessLogRetentionDays,
+		"retention absent from the request must land as the zero value — PUT replaces all mutable fields")
+	assert.Equal(t, before.Cluster, got.Cluster, "cluster must survive updates untouched")
+	assert.Equal(t, before.Subdomain, got.Subdomain, "subdomain must survive updates untouched")
+}
+
+// TestSettingsHandler_PutRejectsClusterChange pins cluster immutability: once
+// assigned, a differing cluster is rejected as a validation error instead of
+// being silently ignored, so callers never observe a value other than the one
+// they sent. Echoing the assigned cluster back stays valid, which lets
+// declarative clients send their full desired state idempotently.
+func TestSettingsHandler_PutRejectsClusterChange(t *testing.T) {
+	f := newAgentNetworkHandlerFixture(t)
+
+	rec := f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"cluster": "eu.proxy.netbird.io", "enable_log_collection": true, "enable_prompt_collection": false, "redact_pii": false}`)
+	require.Equal(t, http.StatusOK, rec.Code, "bootstrap PUT must succeed: %s", rec.Body.String())
+
+	rec = f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"cluster": "us.proxy.netbird.io", "enable_log_collection": true, "enable_prompt_collection": false, "redact_pii": false}`)
+	assert.Equal(t, http.StatusUnprocessableEntity, rec.Code,
+		"cluster change must be rejected as a validation error: got %d body=%s", rec.Code, rec.Body.String())
+
+	rec = f.do(t, http.MethodPut, "/agent-network/settings",
+		`{"cluster": "eu.proxy.netbird.io", "enable_log_collection": true, "enable_prompt_collection": false, "redact_pii": true}`)
+	require.Equal(t, http.StatusOK, rec.Code, "echoing the assigned cluster must stay valid: %s", rec.Body.String())
+
+	var got api.AgentNetworkSettings
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &got))
+	assert.Equal(t, "eu.proxy.netbird.io", got.Cluster, "cluster must be unchanged")
+	assert.True(t, got.RedactPii, "toggle sent alongside the echoed cluster must apply")
+}

management/internals/modules/agentnetwork/manager.go

@@ -157,14 +157,14 @@ func NewManager(
 }
 
 func (m *managerImpl) GetAllProviders(ctx context.Context, accountID, userID string) ([]*types.Provider, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkProviders, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAccountAgentNetworkProviders(ctx, store.LockingStrengthNone, accountID)
 }
 
 func (m *managerImpl) GetProvider(ctx context.Context, accountID, userID, providerID string) (*types.Provider, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkProviders, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAgentNetworkProviderByID(ctx, store.LockingStrengthNone, accountID, providerID)
@@ -175,9 +175,14 @@ func (m *managerImpl) GetProvider(ctx context.Context, accountID, userID, provid
 // been created yet; otherwise it is ignored (the cluster is pinned on
 // Settings and every provider in the account routes through it).
 func (m *managerImpl) CreateProvider(ctx context.Context, userID string, provider *types.Provider, bootstrapCluster string) (*types.Provider, error) {
-	if err := m.requirePermission(ctx, provider.AccountID, userID, operations.Create); err != nil {
+	if err := m.requirePermission(ctx, provider.AccountID, userID, modules.AgentNetworkProviders, operations.Create); err != nil {
 		return nil, err
 	}
+	if strings.TrimSpace(bootstrapCluster) != "" {
+		if err := m.requireSettingsBootstrapPermission(ctx, provider.AccountID, userID); err != nil {
+			return nil, err
+		}
+	}
 
 	// An empty api_key would silently produce a synthesised service
 	// that 401s on every upstream request. Surface the misconfiguration
@@ -202,7 +207,7 @@ func (m *managerImpl) CreateProvider(ctx context.Context, userID string, provide
 	}
 
 	if strings.TrimSpace(bootstrapCluster) != "" {
-		if _, err := m.bootstrapSettingsIfNeeded(ctx, provider.AccountID, bootstrapCluster); err != nil {
+		if _, err := m.bootstrapSettingsIfNeeded(ctx, m.store, provider.AccountID, bootstrapCluster); err != nil {
 			// The provider create has already succeeded; logging the
 			// bootstrap miss matches the plan's PoC behaviour. The synth
 			// path treats a missing settings row as a no-op, and the next
@@ -218,7 +223,7 @@ func (m *managerImpl) CreateProvider(ctx context.Context, userID string, provide
 }
 
 func (m *managerImpl) UpdateProvider(ctx context.Context, userID string, provider *types.Provider) (*types.Provider, error) {
-	if err := m.requirePermission(ctx, provider.AccountID, userID, operations.Update); err != nil {
+	if err := m.requirePermission(ctx, provider.AccountID, userID, modules.AgentNetworkProviders, operations.Update); err != nil {
 		return nil, err
 	}
 
@@ -257,7 +262,7 @@ func (m *managerImpl) UpdateProvider(ctx context.Context, userID string, provide
 }
 
 func (m *managerImpl) DeleteProvider(ctx context.Context, accountID, userID, providerID string) error {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Delete); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkProviders, operations.Delete); err != nil {
 		return err
 	}
 
@@ -306,21 +311,21 @@ func pluralize(n int, singular, plural string) string {
 }
 
 func (m *managerImpl) GetAllPolicies(ctx context.Context, accountID, userID string) ([]*types.Policy, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkPolicies, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAccountAgentNetworkPolicies(ctx, store.LockingStrengthNone, accountID)
 }
 
 func (m *managerImpl) GetPolicy(ctx context.Context, accountID, userID, policyID string) (*types.Policy, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkPolicies, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAgentNetworkPolicyByID(ctx, store.LockingStrengthNone, accountID, policyID)
 }
 
 func (m *managerImpl) CreatePolicy(ctx context.Context, userID string, policy *types.Policy) (*types.Policy, error) {
-	if err := m.requirePermission(ctx, policy.AccountID, userID, operations.Create); err != nil {
+	if err := m.requirePermission(ctx, policy.AccountID, userID, modules.AgentNetworkPolicies, operations.Create); err != nil {
 		return nil, err
 	}
 
@@ -346,7 +351,7 @@ func (m *managerImpl) CreatePolicy(ctx context.Context, userID string, policy *t
 }
 
 func (m *managerImpl) UpdatePolicy(ctx context.Context, userID string, policy *types.Policy) (*types.Policy, error) {
-	if err := m.requirePermission(ctx, policy.AccountID, userID, operations.Update); err != nil {
+	if err := m.requirePermission(ctx, policy.AccountID, userID, modules.AgentNetworkPolicies, operations.Update); err != nil {
 		return nil, err
 	}
 
@@ -373,7 +378,7 @@ func (m *managerImpl) UpdatePolicy(ctx context.Context, userID string, policy *t
 }
 
 func (m *managerImpl) DeletePolicy(ctx context.Context, accountID, userID, policyID string) error {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Delete); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkPolicies, operations.Delete); err != nil {
 		return err
 	}
 
@@ -393,21 +398,21 @@ func (m *managerImpl) DeletePolicy(ctx context.Context, accountID, userID, polic
 }
 
 func (m *managerImpl) GetAllGuardrails(ctx context.Context, accountID, userID string) ([]*types.Guardrail, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkGuardrails, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAccountAgentNetworkGuardrails(ctx, store.LockingStrengthNone, accountID)
 }
 
 func (m *managerImpl) GetGuardrail(ctx context.Context, accountID, userID, guardrailID string) (*types.Guardrail, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkGuardrails, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAgentNetworkGuardrailByID(ctx, store.LockingStrengthNone, accountID, guardrailID)
 }
 
 func (m *managerImpl) CreateGuardrail(ctx context.Context, userID string, guardrail *types.Guardrail) (*types.Guardrail, error) {
-	if err := m.requirePermission(ctx, guardrail.AccountID, userID, operations.Create); err != nil {
+	if err := m.requirePermission(ctx, guardrail.AccountID, userID, modules.AgentNetworkGuardrails, operations.Create); err != nil {
 		return nil, err
 	}
 
@@ -429,7 +434,7 @@ func (m *managerImpl) CreateGuardrail(ctx context.Context, userID string, guardr
 }
 
 func (m *managerImpl) UpdateGuardrail(ctx context.Context, userID string, guardrail *types.Guardrail) (*types.Guardrail, error) {
-	if err := m.requirePermission(ctx, guardrail.AccountID, userID, operations.Update); err != nil {
+	if err := m.requirePermission(ctx, guardrail.AccountID, userID, modules.AgentNetworkGuardrails, operations.Update); err != nil {
 		return nil, err
 	}
 
@@ -452,7 +457,7 @@ func (m *managerImpl) UpdateGuardrail(ctx context.Context, userID string, guardr
 }
 
 func (m *managerImpl) DeleteGuardrail(ctx context.Context, accountID, userID, guardrailID string) error {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Delete); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkGuardrails, operations.Delete); err != nil {
 		return err
 	}
 
@@ -473,7 +478,7 @@ func (m *managerImpl) DeleteGuardrail(ctx context.Context, accountID, userID, gu
 
 // GetAllBudgetRules returns every account-level budget rule for the account.
 func (m *managerImpl) GetAllBudgetRules(ctx context.Context, accountID, userID string) ([]*types.AccountBudgetRule, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkBudgets, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAccountAgentNetworkBudgetRules(ctx, store.LockingStrengthNone, accountID)
@@ -481,7 +486,7 @@ func (m *managerImpl) GetAllBudgetRules(ctx context.Context, accountID, userID s
 
 // GetBudgetRule returns a single account-level budget rule.
 func (m *managerImpl) GetBudgetRule(ctx context.Context, accountID, userID, ruleID string) (*types.AccountBudgetRule, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkBudgets, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.GetAgentNetworkBudgetRuleByID(ctx, store.LockingStrengthNone, accountID, ruleID)
@@ -491,7 +496,7 @@ func (m *managerImpl) GetBudgetRule(ctx context.Context, accountID, userID, rule
 // enforced at request time (CheckLLMPolicyLimits), not baked into the synth
 // proxy config, so no reconcile is needed.
 func (m *managerImpl) CreateBudgetRule(ctx context.Context, userID string, rule *types.AccountBudgetRule) (*types.AccountBudgetRule, error) {
-	if err := m.requirePermission(ctx, rule.AccountID, userID, operations.Create); err != nil {
+	if err := m.requirePermission(ctx, rule.AccountID, userID, modules.AgentNetworkBudgets, operations.Create); err != nil {
 		return nil, err
 	}
 
@@ -513,7 +518,7 @@ func (m *managerImpl) CreateBudgetRule(ctx context.Context, userID string, rule
 
 // UpdateBudgetRule updates an existing account-level budget rule.
 func (m *managerImpl) UpdateBudgetRule(ctx context.Context, userID string, rule *types.AccountBudgetRule) (*types.AccountBudgetRule, error) {
-	if err := m.requirePermission(ctx, rule.AccountID, userID, operations.Update); err != nil {
+	if err := m.requirePermission(ctx, rule.AccountID, userID, modules.AgentNetworkBudgets, operations.Update); err != nil {
 		return nil, err
 	}
 
@@ -536,7 +541,7 @@ func (m *managerImpl) UpdateBudgetRule(ctx context.Context, userID string, rule
 
 // DeleteBudgetRule removes an account-level budget rule.
 func (m *managerImpl) DeleteBudgetRule(ctx context.Context, accountID, userID, ruleID string) error {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Delete); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkBudgets, operations.Delete); err != nil {
 		return err
 	}
 
@@ -554,40 +559,83 @@ func (m *managerImpl) DeleteBudgetRule(ctx context.Context, accountID, userID, r
 	return nil
 }
 
-// UpdateSettings applies the mutable account-level settings — the collection
-// toggles — onto the existing row. Cluster and Subdomain are immutable and are
-// preserved from the persisted row regardless of the input. Because the
-// collection toggles change the synthesised service config (prompt-capture
-// gating, access-log emission), a reconcile is triggered so the proxy and peer
-// network maps converge on the new state.
+// UpdateSettings replaces the mutable account-level settings — the collection
+// toggles and retention — on the account's row. When the account has no
+// settings row yet, a non-empty settings.Cluster bootstraps one (same path as
+// first provider create); without it the update fails with NotFound. On an
+// existing row the cluster and subdomain are immutable: a differing
+// settings.Cluster is rejected rather than silently ignored so callers never
+// observe a value other than what they sent. Because the collection toggles
+// change the synthesised service config (prompt-capture gating, access-log
+// emission), a reconcile is triggered so the proxy and peer network maps
+// converge on the new state.
 func (m *managerImpl) UpdateSettings(ctx context.Context, userID string, settings *types.Settings) (*types.Settings, error) {
-	if err := m.requirePermission(ctx, settings.AccountID, userID, operations.Update); err != nil {
+	if err := m.requirePermission(ctx, settings.AccountID, userID, modules.AgentNetworkSettings, operations.Update); err != nil {
 		return nil, err
 	}
 
-	existing, err := m.store.GetAgentNetworkSettings(ctx, store.LockingStrengthUpdate, settings.AccountID)
+	requestedCluster := strings.TrimSpace(settings.Cluster)
+
+	// The row lock from LockingStrengthUpdate only holds for the duration of
+	// the surrounding transaction, so the read, the cluster-immutability
+	// check, and the save must share one — otherwise concurrent PUTs could
+	// interleave between them.
+	var updated *types.Settings
+	err := m.store.ExecuteInTransaction(ctx, func(tx store.Store) error {
+		existing, err := tx.GetAgentNetworkSettings(ctx, store.LockingStrengthUpdate, settings.AccountID)
+		switch {
+		case err == nil:
+			if requestedCluster != "" && requestedCluster != existing.Cluster {
+				return status.Errorf(status.InvalidArgument, "cluster is immutable once assigned (current: %s)", existing.Cluster)
+			}
+		case isNotFound(err):
+			if requestedCluster == "" {
+				return status.Errorf(status.NotFound, "agent network settings have not been bootstrapped yet; pass cluster to bootstrap them, or create a provider with bootstrap_cluster set")
+			}
+			// Bootstrapping pins the cluster and subdomain — a settings
+			// create on top of the update the caller already passed, matching
+			// the gate on the provider-create bootstrap path.
+			if err := m.requirePermission(ctx, settings.AccountID, userID, modules.AgentNetworkSettings, operations.Create); err != nil {
+				return err
+			}
+			existing, err = m.bootstrapSettingsIfNeeded(ctx, tx, settings.AccountID, requestedCluster)
+			if err != nil {
+				return err
+			}
+		default:
+			return fmt.Errorf("get agent network settings: %w", err)
+		}
+
+		existing.EnableLogCollection = settings.EnableLogCollection
+		existing.EnablePromptCollection = settings.EnablePromptCollection
+		existing.RedactPii = settings.RedactPii
+		existing.AccessLogRetentionDays = settings.AccessLogRetentionDays
+		existing.UpdatedAt = time.Now().UTC()
+
+		if err := tx.SaveAgentNetworkSettings(ctx, existing); err != nil {
+			return fmt.Errorf("save agent network settings: %w", err)
+		}
+		updated = existing
+		return nil
+	})
 	if err != nil {
-		return nil, fmt.Errorf("get agent network settings: %w", err)
-	}
-
-	existing.EnableLogCollection = settings.EnableLogCollection
-	existing.EnablePromptCollection = settings.EnablePromptCollection
-	existing.RedactPii = settings.RedactPii
-	existing.AccessLogRetentionDays = settings.AccessLogRetentionDays
-	existing.UpdatedAt = time.Now().UTC()
-
-	if err := m.store.SaveAgentNetworkSettings(ctx, existing); err != nil {
-		return nil, fmt.Errorf("save agent network settings: %w", err)
+		return nil, err
 	}
 
 	m.accountManager.StoreEvent(ctx, userID, settings.AccountID, settings.AccountID, activity.AgentNetworkSettingsUpdated, map[string]any{
-		"log_collection":    existing.EnableLogCollection,
-		"prompt_collection": existing.EnablePromptCollection,
-		"redact_pii":        existing.RedactPii,
+		"log_collection":    updated.EnableLogCollection,
+		"prompt_collection": updated.EnablePromptCollection,
+		"redact_pii":        updated.RedactPii,
 	})
 	m.reconcile(ctx, settings.AccountID)
 
-	return existing, nil
+	return updated, nil
+}
+
+// isNotFound reports whether err is a status.NotFound error.
+func isNotFound(err error) bool {
+	var sErr *status.Error
+	return errors.As(err, &sErr) && sErr.Type() == status.NotFound
 }
 
 // validateProviderRefs ensures every destination provider id refers to a
@@ -611,14 +659,38 @@ func (m *managerImpl) validateProviderRefs(ctx context.Context, accountID string
 	return nil
 }
 
-// GetSettings returns the agent-network settings row for the account.
-// Returns the underlying status.NotFound when no row has been
-// bootstrapped yet (i.e. the account has no providers).
+// GetSettings returns the agent-network settings row for the account. When no
+// row has been bootstrapped yet, the defaults are returned (without
+// persisting) with cluster and subdomain empty — settings always read as an
+// object, like the account and DNS settings endpoints.
 func (m *managerImpl) GetSettings(ctx context.Context, accountID, userID string) (*types.Settings, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkSettings, operations.Read); err != nil {
 		return nil, err
 	}
-	return m.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
+	settings, err := m.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
+	switch {
+	case err == nil:
+		return settings, nil
+	case isNotFound(err):
+		return types.DefaultSettings(accountID), nil
+	default:
+		return nil, err
+	}
+}
+
+// requireSettingsBootstrapPermission gates the one-time settings bootstrap a
+// first provider create performs. Pinning the account's cluster and subdomain
+// is a settings write, so it needs the settings permission on top of the
+// provider one. No-op once the settings row exists.
+func (m *managerImpl) requireSettingsBootstrapPermission(ctx context.Context, accountID, userID string) error {
+	_, err := m.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
+	if err == nil {
+		return nil
+	}
+	if !isNotFound(err) {
+		return fmt.Errorf("get agent network settings: %w", err)
+	}
+	return m.requirePermission(ctx, accountID, userID, modules.AgentNetworkSettings, operations.Create)
 }
 
 // bootstrapSettingsIfNeeded creates the per-account agent-network
@@ -626,8 +698,9 @@ func (m *managerImpl) GetSettings(ctx context.Context, accountID, userID string)
 // hint the dashboard sends (auto-picked from the active cluster list);
 // the subdomain is picked from the curated wordlist avoiding
 // collisions on the same cluster. Idempotent: if a row already exists
-// it is returned untouched and the hint is ignored.
-func (m *managerImpl) bootstrapSettingsIfNeeded(ctx context.Context, accountID, providerCluster string) (*types.Settings, error) {
+// it is returned untouched and the hint is ignored. st is the store to
+// operate on — pass the transaction store when calling from within one.
+func (m *managerImpl) bootstrapSettingsIfNeeded(ctx context.Context, st store.Store, accountID, providerCluster string) (*types.Settings, error) {
 	if accountID == "" {
 		return nil, fmt.Errorf("bootstrap settings: account id is required")
 	}
@@ -635,16 +708,15 @@ func (m *managerImpl) bootstrapSettingsIfNeeded(ctx context.Context, accountID,
 		return nil, fmt.Errorf("bootstrap settings: provider cluster is required")
 	}
 
-	existing, err := m.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
+	existing, err := st.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
 	if err == nil {
 		return existing, nil
 	}
-	var sErr *status.Error
-	if !errors.As(err, &sErr) || sErr.Type() != status.NotFound {
+	if !isNotFound(err) {
 		return nil, fmt.Errorf("get agent network settings: %w", err)
 	}
 
-	siblings, err := m.store.GetAgentNetworkSettingsByCluster(ctx, store.LockingStrengthNone, providerCluster)
+	siblings, err := st.GetAgentNetworkSettingsByCluster(ctx, store.LockingStrengthNone, providerCluster)
 	if err != nil {
 		return nil, fmt.Errorf("list agent network settings on cluster: %w", err)
 	}
@@ -663,18 +735,12 @@ func (m *managerImpl) bootstrapSettingsIfNeeded(ctx context.Context, accountID,
 	m.labelRngMu.Unlock()
 
 	now := time.Now().UTC()
-	settings := &types.Settings{
-		AccountID: accountID,
-		Cluster:   providerCluster,
-		Subdomain: subdomain,
-		// Logs on by default; usage is collected regardless. Retention bounds
-		// how long full log rows are kept.
-		EnableLogCollection:    true,
-		AccessLogRetentionDays: types.DefaultAccessLogRetentionDays,
-		CreatedAt:              now,
-		UpdatedAt:              now,
-	}
-	if err := m.store.SaveAgentNetworkSettings(ctx, settings); err != nil {
+	settings := types.DefaultSettings(accountID)
+	settings.Cluster = providerCluster
+	settings.Subdomain = subdomain
+	settings.CreatedAt = now
+	settings.UpdatedAt = now
+	if err := st.SaveAgentNetworkSettings(ctx, settings); err != nil {
 		return nil, fmt.Errorf("save agent network settings: %w", err)
 	}
 	return settings, nil
@@ -685,7 +751,7 @@ func (m *managerImpl) bootstrapSettingsIfNeeded(ctx context.Context, accountID,
 // counter view; permission gate is the same Read role that gates
 // every other agent-network surface.
 func (m *managerImpl) ListConsumption(ctx context.Context, accountID, userID string) ([]*types.Consumption, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkUsage, operations.Read); err != nil {
 		return nil, err
 	}
 	return m.store.ListAgentNetworkConsumption(ctx, store.LockingStrengthNone, accountID)
@@ -694,7 +760,7 @@ func (m *managerImpl) ListConsumption(ctx context.Context, accountID, userID str
 // ListAccessLogs returns a paginated, server-side-filtered page of
 // agent-network access logs plus the total count matching the filter.
 func (m *managerImpl) ListAccessLogs(ctx context.Context, accountID, userID string, filter types.AgentNetworkAccessLogFilter) ([]*types.AgentNetworkAccessLog, int64, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkLogs, operations.Read); err != nil {
 		return nil, 0, err
 	}
 	return m.store.GetAgentNetworkAccessLogs(ctx, store.LockingStrengthNone, accountID, filter)
@@ -704,7 +770,7 @@ func (m *managerImpl) ListAccessLogs(ctx context.Context, accountID, userID stri
 // agent-network access logs grouped by session, plus the total number of
 // sessions matching the filter.
 func (m *managerImpl) ListAccessLogSessions(ctx context.Context, accountID, userID string, filter types.AgentNetworkAccessLogFilter) ([]*types.AgentNetworkAccessLogSession, int64, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkLogs, operations.Read); err != nil {
 		return nil, 0, err
 	}
 	return m.store.GetAgentNetworkAccessLogSessions(ctx, store.LockingStrengthNone, accountID, filter)
@@ -713,7 +779,7 @@ func (m *managerImpl) ListAccessLogSessions(ctx context.Context, accountID, user
 // GetUsageOverview returns the filtered usage rows aggregated into time buckets
 // at the requested granularity, oldest-first.
 func (m *managerImpl) GetUsageOverview(ctx context.Context, accountID, userID string, filter types.AgentNetworkAccessLogFilter, granularity types.UsageGranularity) ([]*types.AgentNetworkUsageBucket, error) {
-	if err := m.requirePermission(ctx, accountID, userID, operations.Read); err != nil {
+	if err := m.requirePermission(ctx, accountID, userID, modules.AgentNetworkUsage, operations.Read); err != nil {
 		return nil, err
 	}
 	rows, err := m.store.GetAgentNetworkUsageRows(ctx, store.LockingStrengthNone, accountID, filter)
@@ -787,8 +853,8 @@ func (m *managerImpl) RecordConsumption(ctx context.Context, accountID string, k
 	return m.store.IncrementAgentNetworkConsumption(ctx, accountID, kind, dimID, windowSeconds, windowStart, tokensIn, tokensOut, costUSD)
 }
 
-func (m *managerImpl) requirePermission(ctx context.Context, accountID, userID string, op operations.Operation) error {
-	ok, _, err := m.permissionsManager.ValidateUserPermissions(ctx, accountID, userID, modules.AgentNetwork, op)
+func (m *managerImpl) requirePermission(ctx context.Context, accountID, userID string, module modules.Module, op operations.Operation) error {
+	ok, _, err := m.permissionsManager.ValidateUserPermissions(ctx, accountID, userID, module, op)
 	if err != nil {
 		return status.NewPermissionValidationError(err)
 	}
@@ -877,8 +943,8 @@ func (*mockManager) UpdateBudgetRule(_ context.Context, _ string, r *types.Accou
 
 func (*mockManager) DeleteBudgetRule(_ context.Context, _, _, _ string) error { return nil }
 
-func (*mockManager) GetSettings(_ context.Context, _, _ string) (*types.Settings, error) {
-	return nil, status.Errorf(status.NotFound, "agent network settings not found")
+func (*mockManager) GetSettings(_ context.Context, accountID, _ string) (*types.Settings, error) {
+	return types.DefaultSettings(accountID), nil
 }
 
 func (*mockManager) UpdateSettings(_ context.Context, _ string, s *types.Settings) (*types.Settings, error) {

management/internals/modules/agentnetwork/provider_bootstrap_test.go

@@ -0,0 +1,134 @@
+package agentnetwork
+
+import (
+	"context"
+	"runtime"
+	"testing"
+
+	"github.com/golang/mock/gomock"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	"github.com/netbirdio/netbird/management/server/account"
+	"github.com/netbirdio/netbird/management/server/permissions"
+	"github.com/netbirdio/netbird/management/server/permissions/modules"
+	"github.com/netbirdio/netbird/management/server/permissions/operations"
+	"github.com/netbirdio/netbird/management/server/store"
+	nbtypes "github.com/netbirdio/netbird/management/server/types"
+	"github.com/netbirdio/netbird/shared/management/status"
+)
+
+// bootstrapFixture wires a real sqlite store to a gomock permissions manager
+// so tests can grant the provider permission while denying (or never
+// expecting) the settings one.
+type bootstrapFixture struct {
+	manager Manager
+	store   store.Store
+	perms   *permissions.MockManager
+}
+
+func newBootstrapFixture(t *testing.T) *bootstrapFixture {
+	t.Helper()
+	if runtime.GOOS == "windows" {
+		t.Skip("sqlite store not properly supported on Windows yet")
+	}
+	t.Setenv("NETBIRD_STORE_ENGINE", string(nbtypes.SqliteStoreEngine))
+
+	st, cleanUp, err := store.NewTestStoreFromSQL(context.Background(), "", t.TempDir())
+	require.NoError(t, err, "test store setup must succeed")
+	t.Cleanup(cleanUp)
+
+	ctrl := gomock.NewController(t)
+	perms := permissions.NewMockManager(ctrl)
+
+	accounts := account.NewMockManager(ctrl)
+	accounts.EXPECT().StoreEvent(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any()).AnyTimes()
+	accounts.EXPECT().UpdateAccountPeers(gomock.Any(), gomock.Any(), gomock.Any()).AnyTimes()
+	accounts.EXPECT().BufferUpdateAccountPeers(gomock.Any(), gomock.Any(), gomock.Any()).AnyTimes()
+
+	return &bootstrapFixture{
+		manager: NewManager(st, perms, accounts, nil),
+		store:   st,
+		perms:   perms,
+	}
+}
+
+func (f *bootstrapFixture) expectPermission(accountID, userID string, module modules.Module, op operations.Operation, allowed bool) {
+	f.perms.EXPECT().
+		ValidateUserPermissions(gomock.Any(), accountID, userID, module, op).
+		Return(allowed, context.Background(), nil)
+}
+
+func newBootstrapProvider(accountID string) *types.Provider {
+	p := types.NewProvider(accountID)
+	p.Name = "openai"
+	p.UpstreamURL = "https://api.openai.com"
+	p.APIKey = "sk-test"
+	p.Enabled = true
+	return p
+}
+
+// TestCreateProviderBootstrapRequiresSettingsPermission pins the gate on the
+// one-time settings bootstrap: creating the first provider with a
+// bootstrap_cluster pins the account's cluster and subdomain, which is a
+// settings write and must not ride on the providers permission alone.
+func TestCreateProviderBootstrapRequiresSettingsPermission(t *testing.T) {
+	ctx := context.Background()
+
+	t.Run("denied without settings permission", func(t *testing.T) {
+		f := newBootstrapFixture(t)
+		f.expectPermission("account1", "user1", modules.AgentNetworkProviders, operations.Create, true)
+		f.expectPermission("account1", "user1", modules.AgentNetworkSettings, operations.Create, false)
+
+		_, err := f.manager.CreateProvider(ctx, "user1", newBootstrapProvider("account1"), "cluster1.example.com")
+		require.Error(t, err, "bootstrap without settings permission must fail")
+		var sErr *status.Error
+		require.ErrorAs(t, err, &sErr)
+		assert.Equal(t, status.PermissionDenied, sErr.Type(), "denial should surface as permission denied")
+
+		providers, err := f.store.GetAccountAgentNetworkProviders(ctx, store.LockingStrengthNone, "account1")
+		require.NoError(t, err)
+		assert.Empty(t, providers, "provider must not be persisted when bootstrap is denied")
+		_, err = f.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, "account1")
+		assert.Error(t, err, "settings row must not be created when bootstrap is denied")
+	})
+
+	t.Run("allowed with settings permission", func(t *testing.T) {
+		f := newBootstrapFixture(t)
+		f.expectPermission("account1", "user1", modules.AgentNetworkProviders, operations.Create, true)
+		f.expectPermission("account1", "user1", modules.AgentNetworkSettings, operations.Create, true)
+
+		created, err := f.manager.CreateProvider(ctx, "user1", newBootstrapProvider("account1"), "cluster1.example.com")
+		require.NoError(t, err, "bootstrap with both permissions must succeed")
+		require.NotNil(t, created)
+
+		settings, err := f.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, "account1")
+		require.NoError(t, err, "bootstrap must create the settings row")
+		assert.Equal(t, "cluster1.example.com", settings.Cluster, "settings should pin the bootstrap cluster")
+	})
+
+	t.Run("existing settings need no settings permission", func(t *testing.T) {
+		f := newBootstrapFixture(t)
+		require.NoError(t, f.store.SaveAgentNetworkSettings(ctx, &types.Settings{
+			AccountID: "account1",
+			Cluster:   "cluster1.example.com",
+			Subdomain: "existing",
+		}), "pre-existing settings row setup must succeed")
+
+		// Only the providers permission may be consulted: gomock fails the
+		// test on any unexpected settings-permission call.
+		f.expectPermission("account1", "user1", modules.AgentNetworkProviders, operations.Create, true)
+
+		_, err := f.manager.CreateProvider(ctx, "user1", newBootstrapProvider("account1"), "cluster1.example.com")
+		require.NoError(t, err, "create with existing settings must not require the settings permission")
+	})
+
+	t.Run("no bootstrap cluster needs no settings permission", func(t *testing.T) {
+		f := newBootstrapFixture(t)
+		f.expectPermission("account1", "user1", modules.AgentNetworkProviders, operations.Create, true)
+
+		_, err := f.manager.CreateProvider(ctx, "user1", newBootstrapProvider("account1"), "")
+		require.NoError(t, err, "create without bootstrap must not require the settings permission")
+	})
+}

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

@@ -164,9 +164,7 @@ func (p *Provider) FromAPIRequest(req *api.AgentNetworkProviderRequest) {
 		p.MetadataDisabled = *req.MetadataDisabled
 	}
 	// Identity-header overrides for catalogs flagged Customizable.
-	// nil pointer = "field omitted on the wire" → leave the stored
-	// value untouched (per the openapi description). Empty string is
-	// an explicit clear that disables stamping for this dimension.
+	// Empty or omitted disables stamping for this dimension.
 	if req.IdentityHeaderUserId != nil {
 		p.IdentityHeaderUserID = strings.TrimSpace(*req.IdentityHeaderUserId)
 	}
@@ -192,16 +190,20 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 	created := p.CreatedAt
 	updated := p.UpdatedAt
 	resp := &api.AgentNetworkProvider{
-		Id:                  p.ID,
-		ProviderId:          p.ProviderID,
-		Name:                p.Name,
-		UpstreamUrl:         p.UpstreamURL,
-		Models:              models,
-		Enabled:             p.Enabled,
-		SkipTlsVerification: p.SkipTLSVerification,
-		MetadataDisabled:    p.MetadataDisabled,
-		CreatedAt:           &created,
-		UpdatedAt:           &updated,
+		Id:          p.ID,
+		ProviderId:  p.ProviderID,
+		Name:        p.Name,
+		UpstreamUrl: p.UpstreamURL,
+		Models:      models,
+		// Always present on the wire so an explicitly cleared header
+		// round-trips as "" instead of vanishing from the response.
+		IdentityHeaderUserId: p.IdentityHeaderUserID,
+		IdentityHeaderGroups: p.IdentityHeaderGroups,
+		Enabled:              p.Enabled,
+		SkipTlsVerification:  p.SkipTLSVerification,
+		MetadataDisabled:     p.MetadataDisabled,
+		CreatedAt:            &created,
+		UpdatedAt:            &updated,
 	}
 	if len(p.ExtraValues) > 0 {
 		out := make(map[string]string, len(p.ExtraValues))
@@ -210,14 +212,6 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 		}
 		resp.ExtraValues = &out
 	}
-	if p.IdentityHeaderUserID != "" {
-		v := p.IdentityHeaderUserID
-		resp.IdentityHeaderUserId = &v
-	}
-	if p.IdentityHeaderGroups != "" {
-		v := p.IdentityHeaderGroups
-		resp.IdentityHeaderGroups = &v
-	}
 	return resp
 }
 

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

@@ -77,3 +77,41 @@ func TestProvider_MetadataDisabled_RoundTrip(t *testing.T) {
 	assert.False(t, p.MetadataDisabled, "explicit false must clear metadata_disabled")
 	assert.False(t, p.ToAPIResponse().MetadataDisabled, "response must reflect the cleared value")
 }
+
+// TestProvider_IdentityHeaders_AlwaysOnWire pins that the identity header
+// fields are always present in the API response — an explicitly cleared
+// ("") header must round-trip as "" rather than vanish, so API consumers
+// (e.g. the Terraform provider) never observe a value other than the one
+// they wrote.
+func TestProvider_IdentityHeaders_AlwaysOnWire(t *testing.T) {
+	set := "x-bf-dim-netbird_user_id"
+	empty := ""
+
+	base := func() *api.AgentNetworkProviderRequest {
+		return &api.AgentNetworkProviderRequest{
+			ProviderId:  "custom",
+			Name:        "bifrost",
+			UpstreamUrl: "https://bifrost.internal",
+		}
+	}
+
+	p := NewProvider("acc-1")
+	resp := p.ToAPIResponse()
+	assert.Equal(t, "", resp.IdentityHeaderUserId, "unset header must surface as empty string, not be omitted")
+	assert.Equal(t, "", resp.IdentityHeaderGroups, "unset header must surface as empty string, not be omitted")
+
+	req := base()
+	req.IdentityHeaderUserId = &set
+	p.FromAPIRequest(req)
+	assert.Equal(t, set, p.ToAPIResponse().IdentityHeaderUserId, "configured header must round-trip")
+
+	// Omitting the field preserves it.
+	p.FromAPIRequest(base())
+	assert.Equal(t, set, p.ToAPIResponse().IdentityHeaderUserId, "omitted header must preserve the stored value")
+
+	// An explicit "" clears it AND stays visible on the wire.
+	req = base()
+	req.IdentityHeaderUserId = &empty
+	p.FromAPIRequest(req)
+	assert.Equal(t, "", p.ToAPIResponse().IdentityHeaderUserId, "cleared header must round-trip as empty string")
+}

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

@@ -1,6 +1,7 @@
 package types
 
 import (
+	"strings"
 	"time"
 
 	"github.com/netbirdio/netbird/shared/management/http/api"
@@ -42,18 +43,34 @@ type Settings struct {
 // schema cohesive.
 func (Settings) TableName() string { return "agent_network_settings" }
 
+// DefaultSettings returns the settings an account observes before its row is
+// bootstrapped: log collection on with the default retention, everything else
+// off, and no cluster/subdomain assigned yet. Bootstrap persists exactly these
+// values plus the assigned cluster and subdomain, so the pre-bootstrap read
+// and the freshly bootstrapped row agree.
+func DefaultSettings(accountID string) *Settings {
+	return &Settings{
+		AccountID:              accountID,
+		EnableLogCollection:    true,
+		AccessLogRetentionDays: DefaultAccessLogRetentionDays,
+	}
+}
+
 // Endpoint returns the bare hostname agents reach this account at:
-// `<subdomain>.<cluster>`.
+// `<subdomain>.<cluster>`. Empty until both halves are assigned at bootstrap.
 func (s *Settings) Endpoint() string {
+	if s.Cluster == "" || s.Subdomain == "" {
+		return ""
+	}
 	return s.Subdomain + "." + s.Cluster
 }
 
-// ToAPIResponse renders the settings as the API representation.
+// ToAPIResponse renders the settings as the API representation. The
+// timestamps are omitted while zero — a default (not yet bootstrapped) view
+// has no persisted row to date.
 func (s *Settings) ToAPIResponse() *api.AgentNetworkSettings {
-	created := s.CreatedAt
-	updated := s.UpdatedAt
 	retention := s.AccessLogRetentionDays
-	return &api.AgentNetworkSettings{
+	resp := &api.AgentNetworkSettings{
 		Cluster:                s.Cluster,
 		Subdomain:              s.Subdomain,
 		Endpoint:               s.Endpoint(),
@@ -61,14 +78,27 @@ func (s *Settings) ToAPIResponse() *api.AgentNetworkSettings {
 		EnablePromptCollection: s.EnablePromptCollection,
 		RedactPii:              s.RedactPii,
 		AccessLogRetentionDays: &retention,
-		CreatedAt:              &created,
-		UpdatedAt:              &updated,
 	}
+	if !s.CreatedAt.IsZero() {
+		created := s.CreatedAt
+		resp.CreatedAt = &created
+	}
+	if !s.UpdatedAt.IsZero() {
+		updated := s.UpdatedAt
+		resp.UpdatedAt = &updated
+	}
+	return resp
 }
 
-// FromAPIRequest applies the mutable settings fields from the request. Cluster
-// and Subdomain are immutable and intentionally not touched here.
+// FromAPIRequest applies the request onto the receiver. The mutable
+// collection fields are always replaced with the request values. Cluster
+// participates only in bootstrap and the immutability check (see
+// Manager.UpdateSettings); Subdomain is server-assigned and never taken
+// from a request.
 func (s *Settings) FromAPIRequest(req *api.AgentNetworkSettingsRequest) {
+	if req.Cluster != nil {
+		s.Cluster = strings.TrimSpace(*req.Cluster)
+	}
 	s.EnableLogCollection = req.EnableLogCollection
 	s.EnablePromptCollection = req.EnablePromptCollection
 	s.RedactPii = req.RedactPii

management/internals/server/boot.go

@@ -24,13 +24,13 @@ import (
 
 	"github.com/netbirdio/netbird/encryption"
 	"github.com/netbirdio/netbird/formatter/hook"
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
 	"github.com/netbirdio/netbird/management/internals/modules/reverseproxy/accesslogs"
 	accesslogsmanager "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/accesslogs/manager"
 	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
 	nbgrpc "github.com/netbirdio/netbird/management/internals/shared/grpc"
 	"github.com/netbirdio/netbird/management/server/activity"
 	activitystore "github.com/netbirdio/netbird/management/server/activity/store"
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
 	nbcache "github.com/netbirdio/netbird/management/server/cache"
 	nbContext "github.com/netbirdio/netbird/management/server/context"
 	nbhttp "github.com/netbirdio/netbird/management/server/http"
@@ -184,6 +184,10 @@ func (s *BaseServer) GRPCServer() *grpc.Server {
 			grpc.ChainStreamInterceptor(realip.StreamServerInterceptorOpts(realipOpts...), streamInterceptor, proxyStream),
 		}
 
+		// Append interceptors contributed by registered gRPC extensions. These
+		// run after the built-in chain (ChainUnaryInterceptor is additive).
+		gRPCOpts = appendExtensionInterceptors(gRPCOpts, s.grpcExtensions)
+
 		if s.Config.HttpConfig.LetsEncryptDomain != "" {
 			certManager, err := encryption.CreateCertManager(s.Config.Datadir, s.Config.HttpConfig.LetsEncryptDomain)
 			if err != nil {
@@ -215,6 +219,9 @@ func (s *BaseServer) GRPCServer() *grpc.Server {
 		mgmtProto.RegisterProxyServiceServer(gRPCAPIHandler, s.ReverseProxyGRPCServer())
 		log.Info("ProxyService registered on gRPC server")
 
+		// Register services contributed by external modules via the extension seam.
+		registerExtensions(gRPCAPIHandler, s.grpcExtensions)
+
 		return gRPCAPIHandler
 	})
 }

management/internals/server/grpc_extension.go

@@ -0,0 +1,74 @@
+package server
+
+import (
+	"context"
+
+	"google.golang.org/grpc"
+)
+
+// GRPCExtension bundles an external module's contribution to the management
+// gRPC server: the registration of one or more services onto the shared
+// grpc.Server, any server-wide interceptors those services require, and an
+// optional shutdown hook. It is a generic extension point with no knowledge of
+// any specific service.
+type GRPCExtension struct {
+	// Register is invoked with the shared grpc.Server (as a ServiceRegistrar)
+	// after the built-in services are registered. It may register any number of
+	// services. May be nil.
+	Register func(grpc.ServiceRegistrar)
+	// UnaryInterceptors are appended to the server's unary interceptor chain,
+	// running after the built-in interceptors. May be empty.
+	UnaryInterceptors []grpc.UnaryServerInterceptor
+	// StreamInterceptors are appended to the server's stream interceptor chain,
+	// running after the built-in interceptors. May be empty.
+	StreamInterceptors []grpc.StreamServerInterceptor
+	// Shutdown, if non-nil, is called once during Stop() with the context
+	// governing server shutdown, which carries a deadline. The hook MUST
+	// return promptly and MUST abandon its work once that context is
+	// cancelled or expires: it runs before the rest of Stop()'s cleanup
+	// (store, event store, embedded IdP) and before Stop() itself checks the
+	// context's deadline, so a hook that ignores the context will delay all
+	// of that cleanup and prevent Stop() from returning on time. May be nil.
+	Shutdown func(ctx context.Context)
+}
+
+// RegisterGRPCExtension registers a gRPC extension. Call before the gRPC server
+// is first built (i.e. before Start); registrations after that have no effect.
+func (s *BaseServer) RegisterGRPCExtension(ext GRPCExtension) {
+	s.grpcExtensions = append(s.grpcExtensions, ext)
+}
+
+// appendExtensionInterceptors appends each extension's interceptors to the gRPC
+// server options as additional chained interceptors. grpc.ChainUnaryInterceptor
+// and grpc.ChainStreamInterceptor are additive, so the returned options run the
+// extension interceptors after any interceptors already present in opts.
+func appendExtensionInterceptors(opts []grpc.ServerOption, exts []GRPCExtension) []grpc.ServerOption {
+	for _, ext := range exts {
+		if len(ext.UnaryInterceptors) > 0 {
+			opts = append(opts, grpc.ChainUnaryInterceptor(ext.UnaryInterceptors...))
+		}
+		if len(ext.StreamInterceptors) > 0 {
+			opts = append(opts, grpc.ChainStreamInterceptor(ext.StreamInterceptors...))
+		}
+	}
+	return opts
+}
+
+// registerExtensions registers each extension's services onto reg.
+func registerExtensions(reg grpc.ServiceRegistrar, exts []GRPCExtension) {
+	for _, ext := range exts {
+		if ext.Register != nil {
+			ext.Register(reg)
+		}
+	}
+}
+
+// runExtensionShutdownHooks calls each extension's shutdown hook, if set,
+// passing ctx through so hooks can honor its deadline/cancellation.
+func runExtensionShutdownHooks(ctx context.Context, exts []GRPCExtension) {
+	for _, ext := range exts {
+		if ext.Shutdown != nil {
+			ext.Shutdown(ctx)
+		}
+	}
+}

management/internals/server/grpc_extension_test.go

@@ -0,0 +1,160 @@
+package server
+
+import (
+	"context"
+	"net"
+	"sync/atomic"
+	"testing"
+
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/credentials/insecure"
+	"google.golang.org/grpc/health"
+	healthgrpc "google.golang.org/grpc/health/grpc_health_v1"
+	"google.golang.org/grpc/test/bufconn"
+)
+
+// Test that an extension's interceptors and service registration are actually
+// wired onto a real in-process gRPC server via the helpers, and that shutdown
+// hooks run. This validates the load-bearing assumption that
+// grpc.ChainUnaryInterceptor is additive (extension interceptors run in
+// addition to any base chain).
+func TestGRPCExtensionAppliedToServer(t *testing.T) {
+	var unaryCalls atomic.Int32
+	var streamShutdownCalled atomic.Bool
+
+	ext := GRPCExtension{
+		Register: func(reg grpc.ServiceRegistrar) {
+			healthgrpc.RegisterHealthServer(reg, health.NewServer())
+		},
+		UnaryInterceptors: []grpc.UnaryServerInterceptor{
+			func(ctx context.Context, req any, info *grpc.UnaryServerInfo, handler grpc.UnaryHandler) (any, error) {
+				unaryCalls.Add(1)
+				return handler(ctx, req)
+			},
+		},
+		Shutdown: func(ctx context.Context) { streamShutdownCalled.Store(true) },
+	}
+	exts := []GRPCExtension{ext}
+
+	// Base options mimic GRPCServer(): a pre-existing chain the extension appends to.
+	var baseUnaryCalls atomic.Int32
+	opts := []grpc.ServerOption{
+		grpc.ChainUnaryInterceptor(func(ctx context.Context, req any, info *grpc.UnaryServerInfo, handler grpc.UnaryHandler) (any, error) {
+			baseUnaryCalls.Add(1)
+			return handler(ctx, req)
+		}),
+	}
+	opts = appendExtensionInterceptors(opts, exts)
+
+	srv := grpc.NewServer(opts...)
+	registerExtensions(srv, exts)
+
+	lis := bufconn.Listen(1024 * 1024)
+	go func() { _ = srv.Serve(lis) }()
+	t.Cleanup(srv.Stop)
+
+	conn, err := grpc.NewClient("passthrough:///bufnet",
+		grpc.WithContextDialer(func(ctx context.Context, _ string) (net.Conn, error) { return lis.DialContext(ctx) }),
+		grpc.WithTransportCredentials(insecure.NewCredentials()))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = conn.Close() })
+
+	_, err = healthgrpc.NewHealthClient(conn).Check(context.Background(), &healthgrpc.HealthCheckRequest{})
+	if err != nil {
+		t.Fatalf("health check via extension-registered service failed: %v", err)
+	}
+	if baseUnaryCalls.Load() != 1 {
+		t.Errorf("base interceptor calls = %d, want 1 (base chain must be preserved)", baseUnaryCalls.Load())
+	}
+	if unaryCalls.Load() != 1 {
+		t.Errorf("extension interceptor calls = %d, want 1", unaryCalls.Load())
+	}
+
+	runExtensionShutdownHooks(context.Background(), exts)
+	if !streamShutdownCalled.Load() {
+		t.Error("extension shutdown hook was not called")
+	}
+}
+
+// TestGRPCExtensionShutdownHookReceivesCallerContext asserts that each hook receives
+// a non-nil context and that it is the very same context the caller passed
+// in, so hooks can rely on values/deadlines placed on it by Stop().
+func TestGRPCExtensionShutdownHookReceivesCallerContext(t *testing.T) {
+	type sentinelKey struct{}
+	want := "shutdown-ctx-sentinel"
+	ctx := context.WithValue(context.Background(), sentinelKey{}, want)
+
+	var called bool
+	ext := GRPCExtension{
+		Shutdown: func(hookCtx context.Context) {
+			called = true
+			if hookCtx == nil {
+				t.Fatal("hook received a nil context")
+			}
+			got, _ := hookCtx.Value(sentinelKey{}).(string)
+			if got != want {
+				t.Errorf("hook context sentinel = %q, want %q (not the caller's context)", got, want)
+			}
+		},
+	}
+
+	runExtensionShutdownHooks(ctx, []GRPCExtension{ext})
+	if !called {
+		t.Fatal("shutdown hook was not called")
+	}
+}
+
+// TestGRPCExtensionShutdownHookObservesCancellation documents, by test, that
+// hooks can honor cancellation/deadlines: a hook given an already-cancelled
+// context must see ctx.Err() != nil and a closed Done() channel.
+func TestGRPCExtensionShutdownHookObservesCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	var called bool
+	ext := GRPCExtension{
+		Shutdown: func(hookCtx context.Context) {
+			called = true
+			if hookCtx.Err() == nil {
+				t.Error("hook context Err() = nil, want non-nil for a cancelled context")
+			}
+			select {
+			case <-hookCtx.Done():
+			default:
+				t.Error("hook context Done() channel is not closed for a cancelled context")
+			}
+		},
+	}
+
+	runExtensionShutdownHooks(ctx, []GRPCExtension{ext})
+	if !called {
+		t.Fatal("shutdown hook was not called")
+	}
+}
+
+// TestGRPCExtensionShutdownHookNilSkipped asserts that an extension
+// with a nil Shutdown hook is skipped without panicking, and that hooks for
+// other extensions still run.
+func TestGRPCExtensionShutdownHookNilSkipped(t *testing.T) {
+	var called atomic.Bool
+	exts := []GRPCExtension{
+		{Shutdown: nil},
+		{Shutdown: func(context.Context) { called.Store(true) }},
+	}
+
+	runExtensionShutdownHooks(context.Background(), exts)
+	if !called.Load() {
+		t.Error("shutdown hook for non-nil extension was not called")
+	}
+}
+
+func TestRegisterGRPCExtensionAccumulates(t *testing.T) {
+	s := &BaseServer{}
+	s.RegisterGRPCExtension(GRPCExtension{})
+	s.RegisterGRPCExtension(GRPCExtension{})
+	if len(s.grpcExtensions) != 2 {
+		t.Fatalf("grpcExtensions len = %d, want 2", len(s.grpcExtensions))
+	}
+}

management/internals/server/server.go

@@ -68,6 +68,11 @@ type BaseServer struct {
 
 	proxyAuthClose func()
 
+	// grpcExtensions holds additional gRPC services, interceptors, and shutdown
+	// hooks registered by external modules via RegisterGRPCExtension. Populated
+	// during boot (single-threaded), consumed by GRPCServer() and Stop().
+	grpcExtensions []GRPCExtension
+
 	listener    net.Listener
 	certManager *autocert.Manager
 	update      *version.Update
@@ -257,6 +262,7 @@ func (s *BaseServer) Stop() error {
 		s.proxyAuthClose()
 		s.proxyAuthClose = nil
 	}
+	runExtensionShutdownHooks(ctx, s.grpcExtensions)
 	_ = s.Store().Close(ctx)
 	_ = s.EventStore().Close(ctx)
 	if s.update != nil {

management/internals/shared/grpc/components_encoder.go

@@ -61,6 +61,8 @@ func EncodeNetworkMapEnvelope(in ComponentsEnvelopeInput) *proto.NetworkMapEnvel
 		return &proto.NetworkMapEnvelope{
 			Payload: &proto.NetworkMapEnvelope_Full{
 				Full: &proto.NetworkMapComponentsFull{
+					Serial:     networkSerial(c.Network),
+					Network:    toAccountNetwork(c.Network),
 					PeerConfig: in.PeerConfig,
 					// components.Peers always contains the target peer
 					Peers:            []*proto.PeerCompact{toPeerCompact(c.Peers[c.PeerID])},

management/internals/shared/grpc/components_encoder_test.go

@@ -758,6 +758,9 @@ func TestEncodeNetworkMapEnvelope_NilComponentsGracefulDegrade(t *testing.T) {
 	assert.Equal(t, "netbird.cloud", full.DnsDomain)
 	assert.Len(t, full.Peers, 1)
 	assert.Empty(t, full.Policies)
+	require.NotNil(t, full.Network, "client runs Calculate() over the envelope and dereferences Network unconditionally; a nil here would crash the receiver")
+	assert.Equal(t, "net-empty", full.Network.Identifier)
+	assert.Equal(t, uint64(9), full.Serial)
 }
 
 func TestEncodeNetworkMapEnvelope_AccountSettingsAlwaysEmitted(t *testing.T) {
@@ -776,6 +779,12 @@ func TestEncodeNetworkMapEnvelope_AccountSettingsAlwaysEmitted(t *testing.T) {
 func emptyNetworkMapComponents() *types.NetworkMapComponents {
 	return types.EmptyNetworkMapComponents(
 		&types.NetworkMapComponents{
-			PeerID: "peer-id", Peers: map[string]*types.ComponentPeer{"peer-id": {}}},
+			PeerID: "peer-id", Peers: map[string]*types.ComponentPeer{"peer-id": {}},
+			Network: &types.Network{
+				Identifier: "net-empty",
+				Net:        net.IPNet{IP: net.IP{100, 64, 0, 0}, Mask: net.CIDRMask(10, 32)},
+				Serial:     9,
+			},
+		},
 	)
 }

management/server/agentnetwork_budgetrule_realstack_test.go

@@ -102,11 +102,20 @@ func TestAgentNetwork_UpdateSettings_PreservesImmutableAndTogglesCollection(t *t
 	require.NotEmpty(t, before.Subdomain, "subdomain pinned at bootstrap")
 	assert.False(t, before.EnablePromptCollection, "prompt collection defaults off")
 
-	// Attempt to flip toggles AND smuggle a different cluster/subdomain — the
-	// immutable fields must be ignored.
+	// A cluster different from the one pinned at bootstrap must be rejected
+	// outright — never silently swapped or ignored.
+	_, err = mgr.UpdateSettings(ctx, adminUserID, &agenttypes.Settings{
+		AccountID:           accountID,
+		Cluster:             "attacker.cluster",
+		EnableLogCollection: true,
+	})
+	require.Error(t, err, "UpdateSettings with a mismatched cluster must fail")
+
+	// Flipping the toggles works with the pinned cluster echoed back (and
+	// with it omitted); the subdomain is never taken from the request.
 	updated, err := mgr.UpdateSettings(ctx, adminUserID, &agenttypes.Settings{
 		AccountID:              accountID,
-		Cluster:                "attacker.cluster",
+		Cluster:                clusterAddr,
 		Subdomain:              "evil",
 		EnableLogCollection:    true,
 		EnablePromptCollection: true,

management/server/group.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"slices"
 
+	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	"github.com/rs/xid"
 	log "github.com/sirupsen/logrus"
 
@@ -744,6 +745,10 @@ func validateDeleteGroup(ctx context.Context, transaction store.Store, group *ty
 		return &GroupLinkError{"network router", linkedRouter.ID}
 	}
 
+	if isLinked, linkedPolicy := isGroupLinkedToAgentNetworkPolicy(ctx, transaction, group.AccountID, group.ID); isLinked {
+		return &GroupLinkError{"agent network policy", linkedPolicy.Name}
+	}
+
 	return checkGroupLinkedToSettings(ctx, transaction, group)
 }
 
@@ -875,6 +880,26 @@ func isGroupLinkedToNetworkRouter(ctx context.Context, transaction store.Store,
 	return false, nil
 }
 
+// isGroupLinkedToAgentNetworkPolicy checks if a group is used as a source group by any
+// agent network policy in the account.
+func isGroupLinkedToAgentNetworkPolicy(ctx context.Context, transaction store.Store, accountID string, groupID string) (bool, *agentNetworkTypes.Policy) {
+	policies, err := transaction.GetAccountAgentNetworkPolicies(ctx, store.LockingStrengthNone, accountID)
+	if err != nil {
+		log.WithContext(ctx).Errorf("error retrieving agent network policies while checking group linkage: %v", err)
+		return false, nil
+	}
+
+	for _, policy := range policies {
+		if policy == nil {
+			continue
+		}
+		if slices.Contains(policy.SourceGroups, groupID) {
+			return true, policy
+		}
+	}
+	return false, nil
+}
+
 // areGroupChangesAffectPeers checks if any changes to the specified groups will affect peers.
 // It fetches each collection once and checks all groupIDs against them in memory.
 func areGroupChangesAffectPeers(ctx context.Context, transaction store.Store, accountID string, groupIDs []string) (bool, error) {

management/server/group_test.go

@@ -18,6 +18,7 @@ import (
 	"golang.org/x/exp/maps"
 
 	nbdns "github.com/netbirdio/netbird/dns"
+	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	"github.com/netbirdio/netbird/management/server/groups"
 	"github.com/netbirdio/netbird/management/server/networks"
 	"github.com/netbirdio/netbird/management/server/networks/resources"
@@ -125,6 +126,11 @@ func TestDefaultAccountManager_DeleteGroup(t *testing.T) {
 			"grp-for-integration",
 			"only service users with admin power can delete integration group",
 		},
+		{
+			"agent network policy",
+			"grp-for-agent-network-policy",
+			"agent network policy",
+		},
 	}
 
 	for _, testCase := range testCases {
@@ -218,6 +224,11 @@ func TestDefaultAccountManager_DeleteGroups(t *testing.T) {
 			groupIDs:        []string{"grp-for-integration"},
 			expectedReasons: []string{"only service users with admin power can delete integration group"},
 		},
+		{
+			name:            "agent network policy",
+			groupIDs:        []string{"grp-for-agent-network-policy"},
+			expectedReasons: []string{"agent network policy"},
+		},
 		{
 			name:            "successfully delete multiple groups",
 			groupIDs:        []string{"group-1", "group-2"},
@@ -406,6 +417,14 @@ func initTestGroupAccount(am *DefaultAccountManager) (*DefaultAccountManager, *t
 		Peers:     make([]string, 0),
 	}
 
+	groupForAgentNetworkPolicy := &types.Group{
+		ID:        "grp-for-agent-network-policy",
+		AccountID: "account-id",
+		Name:      "Group for agent network policies",
+		Issued:    types.GroupIssuedAPI,
+		Peers:     make([]string, 0),
+	}
+
 	routeResource := &route.Route{
 		ID:     "example route",
 		Groups: []string{groupForRoute.ID},
@@ -461,6 +480,18 @@ func initTestGroupAccount(am *DefaultAccountManager) (*DefaultAccountManager, *t
 	_ = am.CreateGroup(context.Background(), accountID, groupAdminUserID, groupForSetupKeys)
 	_ = am.CreateGroup(context.Background(), accountID, groupAdminUserID, groupForUsers)
 	_ = am.CreateGroup(context.Background(), accountID, groupAdminUserID, groupForIntegration)
+	_ = am.CreateGroup(context.Background(), accountID, groupAdminUserID, groupForAgentNetworkPolicy)
+
+	agentNetworkPolicy := &agentNetworkTypes.Policy{
+		ID:           "example agent network policy",
+		AccountID:    accountID,
+		Name:         "Example agent network policy",
+		Enabled:      true,
+		SourceGroups: []string{groupForAgentNetworkPolicy.ID},
+	}
+	if err := am.Store.SaveAgentNetworkPolicy(context.Background(), agentNetworkPolicy); err != nil {
+		return nil, nil, err
+	}
 
 	acc, err := am.Store.GetAccount(context.Background(), account.Id)
 	if err != nil {

management/server/permissions/manager.go

@@ -82,6 +82,9 @@ func (m *managerImpl) ValidateUserPermissions(
 	return m.ValidateRoleModuleAccess(ctx, accountID, role, module, operation), ctxEnriched, nil
 }
 
+// ValidateRoleModuleAccess resolves an operation against the role's explicit
+// grant for the module, then the grant for its parent module when the module
+// is a dotted submodule, and finally the role's AutoAllowNew default.
 func (m *managerImpl) ValidateRoleModuleAccess(
 	ctx context.Context,
 	accountID string,
@@ -89,7 +92,7 @@ func (m *managerImpl) ValidateRoleModuleAccess(
 	module modules.Module,
 	operation operations.Operation,
 ) bool {
-	if permissions, ok := role.Permissions[module]; ok {
+	if permissions, ok := lookupModulePermissions(role, module); ok {
 		if allowed, exists := permissions[operation]; exists {
 			return allowed
 		}
@@ -100,6 +103,21 @@ func (m *managerImpl) ValidateRoleModuleAccess(
 	return role.AutoAllowNew[operation]
 }
 
+// lookupModulePermissions returns the role's explicit permission set for the
+// module, falling back to the parent module's set for dotted submodules. The
+// second return reports whether any explicit set was found.
+func lookupModulePermissions(role roles.RolePermissions, module modules.Module) (map[operations.Operation]bool, bool) {
+	if permissions, ok := role.Permissions[module]; ok {
+		return permissions, true
+	}
+	if parent, hasParent := module.Parent(); hasParent {
+		if permissions, ok := role.Permissions[parent]; ok {
+			return permissions, true
+		}
+	}
+	return nil, false
+}
+
 func (m *managerImpl) ValidateAccountAccess(ctx context.Context, accountID string, user *types.User, allowOwnerAndAdmin bool) (context.Context, error) {
 	if user.AccountID != accountID {
 		return ctx, status.NewUserNotPartOfAccountError()
@@ -119,7 +137,7 @@ func (m *managerImpl) GetPermissionsByRole(ctx context.Context, role types.UserR
 	permissions := roles.Permissions{}
 
 	for k := range modules.All {
-		if rolePermissions, ok := roleMap.Permissions[k]; ok {
+		if rolePermissions, ok := lookupModulePermissions(roleMap, k); ok {
 			permissions[k] = rolePermissions
 			continue
 		}

management/server/permissions/manager_test.go

@@ -0,0 +1,139 @@
+package permissions
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/management/server/permissions/modules"
+	"github.com/netbirdio/netbird/management/server/permissions/operations"
+	"github.com/netbirdio/netbird/management/server/permissions/roles"
+	"github.com/netbirdio/netbird/management/server/types"
+)
+
+func TestValidateRoleModuleAccessSubmoduleCascade(t *testing.T) {
+	manager := NewManager(nil)
+	ctx := context.Background()
+
+	fullAccess := map[operations.Operation]bool{
+		operations.Read:   true,
+		operations.Create: true,
+		operations.Update: true,
+		operations.Delete: true,
+	}
+	readOnly := map[operations.Operation]bool{
+		operations.Read:   true,
+		operations.Create: false,
+		operations.Update: false,
+		operations.Delete: false,
+	}
+	denyAll := map[operations.Operation]bool{
+		operations.Read:   false,
+		operations.Create: false,
+		operations.Update: false,
+		operations.Delete: false,
+	}
+
+	t.Run("parent grant covers submodules", func(t *testing.T) {
+		role := roles.RolePermissions{
+			AutoAllowNew: denyAll,
+			Permissions:  roles.Permissions{modules.AgentNetwork: fullAccess},
+		}
+		assert.True(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkProviders, operations.Create),
+			"parent full grant should allow create on a submodule")
+		assert.True(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkLogs, operations.Read),
+			"parent full grant should allow read on a submodule")
+	})
+
+	t.Run("submodule grant does not leak to parent or siblings", func(t *testing.T) {
+		role := roles.RolePermissions{
+			AutoAllowNew: denyAll,
+			Permissions:  roles.Permissions{modules.AgentNetworkUsage: readOnly},
+		}
+		assert.True(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkUsage, operations.Read),
+			"explicit submodule read should be allowed")
+		assert.False(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkUsage, operations.Create),
+			"read-only submodule grant should not allow create")
+		assert.False(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetwork, operations.Read),
+			"submodule grant should not grant the parent module")
+		assert.False(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkProviders, operations.Read),
+			"submodule grant should not grant a sibling submodule")
+	})
+
+	t.Run("explicit submodule entry wins over parent grant", func(t *testing.T) {
+		role := roles.RolePermissions{
+			AutoAllowNew: denyAll,
+			Permissions: roles.Permissions{
+				modules.AgentNetwork:     fullAccess,
+				modules.AgentNetworkLogs: denyAll,
+			},
+		}
+		assert.False(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkLogs, operations.Read),
+			"explicit submodule deny should override the parent grant")
+		assert.True(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkUsage, operations.Read),
+			"sibling submodules should still resolve through the parent grant")
+	})
+
+	t.Run("auto allow applies when neither submodule nor parent is granted", func(t *testing.T) {
+		role := roles.RolePermissions{
+			AutoAllowNew: readOnly,
+		}
+		assert.True(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkProviders, operations.Read),
+			"auto-allow read should apply to submodules")
+		assert.False(t, manager.ValidateRoleModuleAccess(ctx, "account", role, modules.AgentNetworkProviders, operations.Delete),
+			"auto-allow should not grant unlisted operations")
+	})
+}
+
+// TestExistingRolesKeepAgentNetworkBehaviorOnSubmodules pins the behavior the
+// submodule split must not change: every built-in role resolves the new
+// submodules exactly as it resolved the agent_network module before.
+func TestExistingRolesKeepAgentNetworkBehaviorOnSubmodules(t *testing.T) {
+	manager := NewManager(nil)
+	ctx := context.Background()
+
+	submodules := []modules.Module{
+		modules.AgentNetworkProviders,
+		modules.AgentNetworkPolicies,
+		modules.AgentNetworkGuardrails,
+		modules.AgentNetworkBudgets,
+		modules.AgentNetworkUsage,
+		modules.AgentNetworkLogs,
+		modules.AgentNetworkSettings,
+	}
+	allOperations := []operations.Operation{operations.Read, operations.Create, operations.Update, operations.Delete}
+
+	for _, role := range []types.UserRole{types.UserRoleOwner, types.UserRoleAdmin, types.UserRoleAuditor, types.UserRoleNetworkAdmin, types.UserRoleUser} {
+		rolePermissions, ok := roles.RolesMap[role]
+		require.True(t, ok, "role %s must exist in RolesMap", role)
+
+		for _, sub := range submodules {
+			for _, op := range allOperations {
+				expected := manager.ValidateRoleModuleAccess(ctx, "account", rolePermissions, modules.AgentNetwork, op)
+				actual := manager.ValidateRoleModuleAccess(ctx, "account", rolePermissions, sub, op)
+				assert.Equal(t, expected, actual, "role %s: %s on %s should match the agent_network module", role, op, sub)
+			}
+		}
+	}
+}
+
+func TestGetPermissionsByRoleIncludesSubmodules(t *testing.T) {
+	manager := NewManager(nil)
+	ctx := context.Background()
+
+	permissions, err := manager.GetPermissionsByRole(ctx, types.UserRoleAuditor)
+	require.NoError(t, err, "auditor role must resolve")
+
+	usage, ok := permissions[modules.AgentNetworkUsage]
+	require.True(t, ok, "permissions map should contain the usage submodule")
+	assert.True(t, usage[operations.Read], "auditor should read the usage submodule")
+	assert.False(t, usage[operations.Update], "auditor should not update the usage submodule")
+
+	adminPermissions, err := manager.GetPermissionsByRole(ctx, types.UserRoleAdmin)
+	require.NoError(t, err, "admin role must resolve")
+	providers, ok := adminPermissions[modules.AgentNetworkProviders]
+	require.True(t, ok, "permissions map should contain the providers submodule")
+	assert.True(t, providers[operations.Delete], "admin should delete on the providers submodule")
+}

management/server/permissions/modules/module.go

@@ -1,5 +1,7 @@
 package modules
 
+import "strings"
+
 type Module string
 
 const (
@@ -20,6 +22,17 @@ const (
 	IdentityProviders Module = "identity_providers"
 	Services          Module = "services"
 	AgentNetwork      Module = "agent_network"
+
+	// Agent Network submodules. A role may grant one of these directly
+	// or grant the AgentNetwork parent, which covers all of them (see
+	// permissions.Manager cascade resolution).
+	AgentNetworkProviders  Module = "agent_network.providers"
+	AgentNetworkPolicies   Module = "agent_network.policies"
+	AgentNetworkGuardrails Module = "agent_network.guardrails"
+	AgentNetworkBudgets    Module = "agent_network.budgets"
+	AgentNetworkUsage      Module = "agent_network.usage"
+	AgentNetworkLogs       Module = "agent_network.logs"
+	AgentNetworkSettings   Module = "agent_network.settings"
 )
 
 var All = map[Module]struct{}{
@@ -40,4 +53,21 @@ var All = map[Module]struct{}{
 	IdentityProviders: {},
 	Services:          {},
 	AgentNetwork:      {},
+
+	AgentNetworkProviders:  {},
+	AgentNetworkPolicies:   {},
+	AgentNetworkGuardrails: {},
+	AgentNetworkBudgets:    {},
+	AgentNetworkUsage:      {},
+	AgentNetworkLogs:       {},
+	AgentNetworkSettings:   {},
+}
+
+// Parent returns the module owning a dotted submodule name and true, or the
+// module itself and false when it has no parent.
+func (m Module) Parent() (Module, bool) {
+	if i := strings.IndexByte(string(m), '.'); i > 0 {
+		return Module(string(m)[:i]), true
+	}
+	return m, false
 }

management/server/types/account.go

@@ -1707,14 +1707,34 @@ func (a *Account) injectPrivateServicePolicies(svc *service.Service, proxyPeers
 	if len(proxyPeers) == 0 {
 		return
 	}
+	// A service's AccessGroups can name groups that no longer exist — persisted
+	// services and the agent-network synthesiser both carry the ids verbatim from
+	// their own state. An unresolvable source authorises nothing, so drop it here
+	// rather than let the network-map assembly resolve it to a nil group.
+	sources := a.existingGroupIDs(svc.AccessGroups)
+	if len(sources) == 0 {
+		return
+	}
 	for _, proxyPeer := range proxyPeers {
-		a.Policies = append(a.Policies, a.createPrivateServicePolicy(svc, proxyPeer))
+		a.Policies = append(a.Policies, a.createPrivateServicePolicy(svc, proxyPeer, sources))
 	}
 }
 
-func (a *Account) createPrivateServicePolicy(svc *service.Service, proxyPeer *nbpeer.Peer) *Policy {
+// existingGroupIDs returns the subset of groupIDs that resolve to a group in the account,
+// preserving the input order.
+func (a *Account) existingGroupIDs(groupIDs []string) []string {
+	out := make([]string, 0, len(groupIDs))
+	for _, groupID := range groupIDs {
+		if _, ok := a.Groups[groupID]; ok {
+			out = append(out, groupID)
+		}
+	}
+	return out
+}
+
+func (a *Account) createPrivateServicePolicy(svc *service.Service, proxyPeer *nbpeer.Peer, accessGroups []string) *Policy {
 	policyID := fmt.Sprintf("private-access-%s-%s", svc.ID, proxyPeer.ID)
-	sources := append([]string(nil), svc.AccessGroups...)
+	sources := append([]string(nil), accessGroups...)
 	return &Policy{
 		ID:      policyID,
 		Name:    fmt.Sprintf("Private Access to %s", svc.Name),

management/server/types/proxy_access_token.go

@@ -68,7 +68,7 @@ type ProxyAccessTokenGenerated struct {
 // CreateNewProxyAccessToken generates a new proxy access token.
 // Returns the token with hashed value stored and plain token for one-time display.
 func CreateNewProxyAccessToken(name string, expiresIn time.Duration, accountID *string, createdBy string) (*ProxyAccessTokenGenerated, error) {
-	hashedToken, plainToken, err := generateProxyToken()
+	hashedToken, plainToken, err := GenerateProxyToken()
 	if err != nil {
 		return nil, err
 	}
@@ -94,7 +94,10 @@ func CreateNewProxyAccessToken(name string, expiresIn time.Duration, accountID *
 	}, nil
 }
 
-func generateProxyToken() (HashedProxyToken, PlainProxyToken, error) {
+// GenerateProxyToken generates a new random proxy token, returning its SHA-256
+// hash (for storage) and the one-time plaintext. Exported so external modules
+// can mint tokens in the canonical proxy-token format.
+func GenerateProxyToken() (HashedProxyToken, PlainProxyToken, error) {
 	secret, err := b.Random(ProxyTokenSecretLength)
 	if err != nil {
 		return "", "", err

management/server/types/proxy_access_token_test.go

@@ -1,6 +1,7 @@
 package types
 
 import (
+	"strings"
 	"testing"
 	"time"
 
@@ -123,6 +124,22 @@ func TestCreateNewProxyAccessToken(t *testing.T) {
 	})
 }
 
+func TestGenerateProxyToken(t *testing.T) {
+	hashed, plain, err := GenerateProxyToken()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := plain.Validate(); err != nil {
+		t.Errorf("generated token failed Validate(): %v", err)
+	}
+	if plain.Hash() != hashed {
+		t.Error("returned hashed token does not match Hash(plain)")
+	}
+	if !strings.HasPrefix(string(plain), ProxyTokenPrefix) {
+		t.Errorf("token %q missing prefix %q", plain, ProxyTokenPrefix)
+	}
+}
+
 func TestProxyAccessToken_IsExpired(t *testing.T) {
 	past := time.Now().Add(-1 * time.Hour)
 	future := time.Now().Add(1 * time.Hour)

release_files/darwin_pkg/postinstall

@@ -30,7 +30,23 @@ mkdir -p /usr/local/bin/
     $AGENT service install || true
     $AGENT service start || true
 
-    open $APP
+    console_user=$(stat -f%Su /dev/console 2>/dev/null)
+    case "$console_user" in
+        ""|root|loginwindow|_mbsetupuser)
+            echo "No active GUI user session (console user: '${console_user:-none}'); skipping UI launch."
+            ;;
+        *)
+            uid=$(id -u "$console_user" 2>/dev/null)
+            if [ -z "$uid" ]; then
+                echo "Could not resolve uid for console user '$console_user'; skipping UI launch."
+            else
+                echo "Launching NetBird UI as console user $console_user (uid $uid)."
+                if ! launchctl asuser "$uid" sudo -u "$console_user" -H open "$APP"; then
+                    echo "Failed to launch NetBird UI; if autostart is enabled it will start at next login."
+                fi
+            fi
+            ;;
+    esac
 
     echo "Finished Netbird installation successfully"
     exit 0 # all good

shared/management/client/client.go

@@ -22,7 +22,6 @@ type Client interface {
 	ExtendAuthSession(sysInfo *system.Info, jwtToken string) (*proto.ExtendAuthSessionResponse, error)
 	GetDeviceAuthorizationFlow() (*proto.DeviceAuthorizationFlow, error)
 	GetPKCEAuthorizationFlow() (*proto.PKCEAuthorizationFlow, error)
-	GetNetworkMap(sysInfo *system.Info) (*proto.NetworkMap, error)
 	GetServerURL() string
 	// IsHealthy returns the current connection status without blocking.
 	// Used by the engine to monitor connectivity in the background.

shared/management/client/grpc.go

@@ -436,49 +436,6 @@ func (c *GrpcClient) handleSyncStream(ctx context.Context, serverPubKey wgtypes.
 	return nil
 }
 
-// GetNetworkMap return with the network map
-func (c *GrpcClient) GetNetworkMap(sysInfo *system.Info) (*proto.NetworkMap, error) {
-	serverPubKey, err := c.getServerPublicKey()
-	if err != nil {
-		log.Debugf("failed getting Management Service public key: %s", err)
-		return nil, err
-	}
-
-	ctx, cancelStream := context.WithCancel(c.ctx)
-	defer cancelStream()
-	stream, err := c.connectToSyncStream(ctx, *serverPubKey, sysInfo)
-	if err != nil {
-		log.Debugf("failed to open Management Service stream: %s", err)
-		return nil, err
-	}
-	defer func() {
-		_ = stream.CloseSend()
-	}()
-
-	update, err := stream.Recv()
-	if err == io.EOF {
-		log.Debugf("Management stream has been closed by server: %s", err)
-		return nil, err
-	}
-	if err != nil {
-		log.Debugf("disconnected from Management Service sync stream: %v", err)
-		return nil, err
-	}
-
-	decryptedResp := &proto.SyncResponse{}
-	err = encryption.DecryptMessage(*serverPubKey, c.key, update.Body, decryptedResp)
-	if err != nil {
-		log.Errorf("failed decrypting update message from Management Service: %s", err)
-		return nil, err
-	}
-
-	if decryptedResp.GetNetworkMap() == nil {
-		return nil, fmt.Errorf("invalid msg, required network map")
-	}
-
-	return decryptedResp.GetNetworkMap(), nil
-}
-
 func (c *GrpcClient) connectToSyncStream(ctx context.Context, serverPubKey wgtypes.Key, sysInfo *system.Info) (proto.ManagementService_SyncClient, error) {
 	req := &proto.SyncRequest{Meta: infoToMetaData(sysInfo)}
 

shared/management/client/mock.go

@@ -94,11 +94,6 @@ func (m *MockClient) HealthCheck() error {
 	return m.HealthCheckFunc()
 }
 
-// GetNetworkMap mock implementation of GetNetworkMap from Client interface.
-func (m *MockClient) GetNetworkMap(_ *system.Info) (*proto.NetworkMap, error) {
-	return nil, nil
-}
-
 // GetServerURL mock implementation of GetServerURL from mgm.Client interface
 func (m *MockClient) GetServerURL() string {
 	if m.GetServerURLFunc == nil {

shared/management/client/rest/agentnetwork.go

@@ -0,0 +1,381 @@
+package rest
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// AgentNetworkAPI APIs for the Agent Network (AI/LLM gateway), do not use directly
+// see more: https://docs.netbird.io/api/resources/agent-network
+type AgentNetworkAPI struct {
+	c *Client
+}
+
+// ListCatalogProviders lists the catalog of supported upstream AI providers
+// (openai_api, anthropic_api, bedrock_api, ...) with their default models and
+// pricing, used to prefill provider create forms.
+func (a *AgentNetworkAPI) ListCatalogProviders(ctx context.Context) ([]api.AgentNetworkCatalogProvider, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/catalog/providers", nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[[]api.AgentNetworkCatalogProvider](resp)
+	return ret, err
+}
+
+// ListProviders lists all Agent Network providers
+func (a *AgentNetworkAPI) ListProviders(ctx context.Context) ([]api.AgentNetworkProvider, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/providers", nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[[]api.AgentNetworkProvider](resp)
+	return ret, err
+}
+
+// GetProvider gets Agent Network provider info
+func (a *AgentNetworkAPI) GetProvider(ctx context.Context, providerID string) (*api.AgentNetworkProvider, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/providers/"+providerID, nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkProvider](resp)
+	return &ret, err
+}
+
+// CreateProvider creates a new Agent Network provider. Set
+// request.BootstrapCluster on the account's first provider to bootstrap the
+// per-account gateway endpoint (alternatively bootstrap via UpdateSettings
+// with a cluster).
+func (a *AgentNetworkAPI) CreateProvider(ctx context.Context, request api.PostApiAgentNetworkProvidersJSONRequestBody) (*api.AgentNetworkProvider, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "POST", "/api/agent-network/providers", bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkProvider](resp)
+	return &ret, err
+}
+
+// UpdateProvider updates an Agent Network provider. The request replaces the
+// provider's mutable state; only an omitted api_key keeps the stored key
+// (secrets are never required to round-trip).
+func (a *AgentNetworkAPI) UpdateProvider(ctx context.Context, providerID string, request api.PutApiAgentNetworkProvidersProviderIdJSONRequestBody) (*api.AgentNetworkProvider, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "PUT", "/api/agent-network/providers/"+providerID, bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkProvider](resp)
+	return &ret, err
+}
+
+// DeleteProvider deletes an Agent Network provider. Fails while any policy
+// still references the provider — detach it first.
+func (a *AgentNetworkAPI) DeleteProvider(ctx context.Context, providerID string) error {
+	resp, err := a.c.NewRequest(ctx, "DELETE", "/api/agent-network/providers/"+providerID, nil, nil)
+	if err != nil {
+		return err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+
+	return nil
+}
+
+// ListPolicies lists all Agent Network policies
+func (a *AgentNetworkAPI) ListPolicies(ctx context.Context) ([]api.AgentNetworkPolicy, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/policies", nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[[]api.AgentNetworkPolicy](resp)
+	return ret, err
+}
+
+// GetPolicy gets Agent Network policy info
+func (a *AgentNetworkAPI) GetPolicy(ctx context.Context, policyID string) (*api.AgentNetworkPolicy, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/policies/"+policyID, nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkPolicy](resp)
+	return &ret, err
+}
+
+// CreatePolicy creates a new Agent Network policy
+func (a *AgentNetworkAPI) CreatePolicy(ctx context.Context, request api.PostApiAgentNetworkPoliciesJSONRequestBody) (*api.AgentNetworkPolicy, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "POST", "/api/agent-network/policies", bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkPolicy](resp)
+	return &ret, err
+}
+
+// UpdatePolicy updates an Agent Network policy
+func (a *AgentNetworkAPI) UpdatePolicy(ctx context.Context, policyID string, request api.PutApiAgentNetworkPoliciesPolicyIdJSONRequestBody) (*api.AgentNetworkPolicy, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "PUT", "/api/agent-network/policies/"+policyID, bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkPolicy](resp)
+	return &ret, err
+}
+
+// DeletePolicy deletes an Agent Network policy
+func (a *AgentNetworkAPI) DeletePolicy(ctx context.Context, policyID string) error {
+	resp, err := a.c.NewRequest(ctx, "DELETE", "/api/agent-network/policies/"+policyID, nil, nil)
+	if err != nil {
+		return err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+
+	return nil
+}
+
+// ListGuardrails lists all Agent Network guardrails
+func (a *AgentNetworkAPI) ListGuardrails(ctx context.Context) ([]api.AgentNetworkGuardrail, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/guardrails", nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[[]api.AgentNetworkGuardrail](resp)
+	return ret, err
+}
+
+// GetGuardrail gets Agent Network guardrail info
+func (a *AgentNetworkAPI) GetGuardrail(ctx context.Context, guardrailID string) (*api.AgentNetworkGuardrail, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/guardrails/"+guardrailID, nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkGuardrail](resp)
+	return &ret, err
+}
+
+// CreateGuardrail creates a new Agent Network guardrail
+func (a *AgentNetworkAPI) CreateGuardrail(ctx context.Context, request api.PostApiAgentNetworkGuardrailsJSONRequestBody) (*api.AgentNetworkGuardrail, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "POST", "/api/agent-network/guardrails", bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkGuardrail](resp)
+	return &ret, err
+}
+
+// UpdateGuardrail updates an Agent Network guardrail
+func (a *AgentNetworkAPI) UpdateGuardrail(ctx context.Context, guardrailID string, request api.PutApiAgentNetworkGuardrailsGuardrailIdJSONRequestBody) (*api.AgentNetworkGuardrail, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "PUT", "/api/agent-network/guardrails/"+guardrailID, bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkGuardrail](resp)
+	return &ret, err
+}
+
+// DeleteGuardrail deletes an Agent Network guardrail
+func (a *AgentNetworkAPI) DeleteGuardrail(ctx context.Context, guardrailID string) error {
+	resp, err := a.c.NewRequest(ctx, "DELETE", "/api/agent-network/guardrails/"+guardrailID, nil, nil)
+	if err != nil {
+		return err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+
+	return nil
+}
+
+// ListBudgetRules lists all account-level Agent Network budget rules
+func (a *AgentNetworkAPI) ListBudgetRules(ctx context.Context) ([]api.AgentNetworkBudgetRule, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/budget-rules", nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[[]api.AgentNetworkBudgetRule](resp)
+	return ret, err
+}
+
+// GetBudgetRule gets Agent Network budget rule info
+func (a *AgentNetworkAPI) GetBudgetRule(ctx context.Context, ruleID string) (*api.AgentNetworkBudgetRule, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/budget-rules/"+ruleID, nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkBudgetRule](resp)
+	return &ret, err
+}
+
+// CreateBudgetRule creates a new Agent Network budget rule
+func (a *AgentNetworkAPI) CreateBudgetRule(ctx context.Context, request api.PostApiAgentNetworkBudgetRulesJSONRequestBody) (*api.AgentNetworkBudgetRule, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "POST", "/api/agent-network/budget-rules", bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkBudgetRule](resp)
+	return &ret, err
+}
+
+// UpdateBudgetRule updates an Agent Network budget rule
+func (a *AgentNetworkAPI) UpdateBudgetRule(ctx context.Context, ruleID string, request api.PutApiAgentNetworkBudgetRulesRuleIdJSONRequestBody) (*api.AgentNetworkBudgetRule, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "PUT", "/api/agent-network/budget-rules/"+ruleID, bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkBudgetRule](resp)
+	return &ret, err
+}
+
+// DeleteBudgetRule deletes an Agent Network budget rule
+func (a *AgentNetworkAPI) DeleteBudgetRule(ctx context.Context, ruleID string) error {
+	resp, err := a.c.NewRequest(ctx, "DELETE", "/api/agent-network/budget-rules/"+ruleID, nil, nil)
+	if err != nil {
+		return err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+
+	return nil
+}
+
+// GetSettings gets the account's Agent Network gateway settings (cluster,
+// subdomain, endpoint, collection toggles). An account that has not been
+// bootstrapped yet — via UpdateSettings with a cluster, or by creating the
+// first provider with bootstrap_cluster set — reads as the defaults with an
+// empty Cluster, Subdomain and Endpoint. Management servers prior to that
+// contract answered 200 with a JSON null body instead; that legacy shape is
+// translated to an APIError matchable via IsNotFound rather than fabricating
+// defaults the server never stated.
+func (a *AgentNetworkAPI) GetSettings(ctx context.Context) (*api.AgentNetworkSettings, error) {
+	resp, err := a.c.NewRequest(ctx, "GET", "/api/agent-network/settings", nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, err
+	}
+	if trimmed := bytes.TrimSpace(body); len(trimmed) == 0 || bytes.Equal(trimmed, []byte("null")) {
+		return nil, &APIError{StatusCode: http.StatusNotFound, Message: "agent network settings not found"}
+	}
+	var ret api.AgentNetworkSettings
+	if err := json.Unmarshal(body, &ret); err != nil {
+		return nil, err
+	}
+	return &ret, nil
+}
+
+// UpdateSettings updates the account's Agent Network settings; the request
+// replaces every mutable field (collection toggles and retention). Setting
+// request.Cluster bootstraps the settings row when the account does not have
+// one yet; on a bootstrapped account it must match the assigned cluster (or
+// be nil) and any other value is rejected — the cluster is immutable.
+func (a *AgentNetworkAPI) UpdateSettings(ctx context.Context, request api.PutApiAgentNetworkSettingsJSONRequestBody) (*api.AgentNetworkSettings, error) {
+	requestBytes, err := json.Marshal(request)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := a.c.NewRequest(ctx, "PUT", "/api/agent-network/settings", bytes.NewReader(requestBytes), nil)
+	if err != nil {
+		return nil, err
+	}
+	if resp.Body != nil {
+		defer resp.Body.Close()
+	}
+	ret, err := parseResponse[api.AgentNetworkSettings](resp)
+	return &ret, err
+}

shared/management/client/rest/agentnetwork_test.go

@@ -0,0 +1,497 @@
+//go:build integration
+
+package rest_test
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/shared/management/client/rest"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+	"github.com/netbirdio/netbird/shared/management/http/util"
+)
+
+var (
+	testAgentNetworkProvider = api.AgentNetworkProvider{
+		Id:          "ainp_test",
+		ProviderId:  "openai_api",
+		Name:        "OpenAI",
+		UpstreamUrl: "https://api.openai.com",
+		Models:      []api.AgentNetworkProviderModel{},
+		Enabled:     true,
+	}
+
+	testAgentNetworkPolicy = api.AgentNetworkPolicy{
+		Id:                     "ainpol_test",
+		Name:                   "Engineering → OpenAI",
+		Enabled:                true,
+		SourceGroups:           []string{"grp-eng"},
+		DestinationProviderIds: []string{"ainp_test"},
+	}
+
+	testAgentNetworkGuardrail = api.AgentNetworkGuardrail{
+		Id:   "aingr_test",
+		Name: "No secrets",
+	}
+
+	testAgentNetworkBudgetRule = api.AgentNetworkBudgetRule{
+		Id:      "ainbud_test",
+		Name:    "Org monthly ceiling",
+		Enabled: true,
+	}
+
+	testAgentNetworkSettings = api.AgentNetworkSettings{
+		Cluster:                "eu.proxy.netbird.io",
+		Subdomain:              "violet",
+		Endpoint:               "violet.eu.proxy.netbird.io",
+		EnableLogCollection:    true,
+		AccessLogRetentionDays: ptr(30),
+	}
+)
+
+func TestAgentNetwork_ListCatalogProviders_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/catalog/providers", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal([]api.AgentNetworkCatalogProvider{{Id: "openai_api", Name: "OpenAI"}})
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.ListCatalogProviders(context.Background())
+		require.NoError(t, err)
+		assert.Len(t, ret, 1)
+		assert.Equal(t, "openai_api", ret[0].Id)
+	})
+}
+
+func TestAgentNetwork_ListProviders_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/providers", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal([]api.AgentNetworkProvider{testAgentNetworkProvider})
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.ListProviders(context.Background())
+		require.NoError(t, err)
+		assert.Len(t, ret, 1)
+		assert.Equal(t, testAgentNetworkProvider, ret[0])
+	})
+}
+
+func TestAgentNetwork_GetProvider_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/providers/ainp_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "GET", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkProvider)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetProvider(context.Background(), "ainp_test")
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkProvider, *ret)
+	})
+}
+
+func TestAgentNetwork_GetProvider_Err(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/providers/ainp_test", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(util.ErrorResponse{Message: "not found", Code: 404})
+			w.WriteHeader(404)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		_, err := c.AgentNetwork.GetProvider(context.Background(), "ainp_test")
+		require.Error(t, err)
+		assert.True(t, rest.IsNotFound(err), "a 404 must be matchable via IsNotFound")
+	})
+}
+
+func TestAgentNetwork_CreateProvider_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/providers", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "POST", r.Method)
+			reqBytes, err := io.ReadAll(r.Body)
+			require.NoError(t, err)
+			var req api.PostApiAgentNetworkProvidersJSONRequestBody
+			require.NoError(t, json.Unmarshal(reqBytes, &req))
+			assert.Equal(t, "OpenAI", req.Name)
+			require.NotNil(t, req.BootstrapCluster)
+			assert.Equal(t, "eu.proxy.netbird.io", *req.BootstrapCluster)
+			retBytes, _ := json.Marshal(testAgentNetworkProvider)
+			_, err = w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.CreateProvider(context.Background(), api.PostApiAgentNetworkProvidersJSONRequestBody{
+			ProviderId:       "openai_api",
+			Name:             "OpenAI",
+			UpstreamUrl:      "https://api.openai.com",
+			ApiKey:           ptr("sk-test"),
+			BootstrapCluster: ptr("eu.proxy.netbird.io"),
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkProvider, *ret)
+	})
+}
+
+func TestAgentNetwork_UpdateProvider_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/providers/ainp_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "PUT", r.Method)
+			reqBytes, err := io.ReadAll(r.Body)
+			require.NoError(t, err)
+			// Omitted optional fields must be absent from the wire (not
+			// zero-valued) so the server-side merge preserves them.
+			assert.NotContains(t, string(reqBytes), "api_key")
+			assert.NotContains(t, string(reqBytes), "models")
+			retBytes, _ := json.Marshal(testAgentNetworkProvider)
+			_, err = w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.UpdateProvider(context.Background(), "ainp_test", api.PutApiAgentNetworkProvidersProviderIdJSONRequestBody{
+			ProviderId:  "openai_api",
+			Name:        "OpenAI",
+			UpstreamUrl: "https://api.openai.com",
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkProvider, *ret)
+	})
+}
+
+func TestAgentNetwork_DeleteProvider_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/providers/ainp_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "DELETE", r.Method)
+			_, err := w.Write([]byte("{}"))
+			require.NoError(t, err)
+		})
+		err := c.AgentNetwork.DeleteProvider(context.Background(), "ainp_test")
+		require.NoError(t, err)
+	})
+}
+
+func TestAgentNetwork_ListPolicies_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/policies", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal([]api.AgentNetworkPolicy{testAgentNetworkPolicy})
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.ListPolicies(context.Background())
+		require.NoError(t, err)
+		assert.Len(t, ret, 1)
+		assert.Equal(t, testAgentNetworkPolicy, ret[0])
+	})
+}
+
+func TestAgentNetwork_GetPolicy_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/policies/ainpol_test", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(testAgentNetworkPolicy)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetPolicy(context.Background(), "ainpol_test")
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkPolicy, *ret)
+	})
+}
+
+func TestAgentNetwork_CreatePolicy_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/policies", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "POST", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkPolicy)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.CreatePolicy(context.Background(), api.PostApiAgentNetworkPoliciesJSONRequestBody{
+			Name:                   "Engineering → OpenAI",
+			SourceGroups:           []string{"grp-eng"},
+			DestinationProviderIds: []string{"ainp_test"},
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkPolicy, *ret)
+	})
+}
+
+func TestAgentNetwork_UpdatePolicy_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/policies/ainpol_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "PUT", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkPolicy)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.UpdatePolicy(context.Background(), "ainpol_test", api.PutApiAgentNetworkPoliciesPolicyIdJSONRequestBody{
+			Name:                   "Engineering → OpenAI",
+			SourceGroups:           []string{"grp-eng"},
+			DestinationProviderIds: []string{"ainp_test"},
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkPolicy, *ret)
+	})
+}
+
+func TestAgentNetwork_DeletePolicy_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/policies/ainpol_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "DELETE", r.Method)
+			_, err := w.Write([]byte("{}"))
+			require.NoError(t, err)
+		})
+		err := c.AgentNetwork.DeletePolicy(context.Background(), "ainpol_test")
+		require.NoError(t, err)
+	})
+}
+
+func TestAgentNetwork_ListGuardrails_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/guardrails", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal([]api.AgentNetworkGuardrail{testAgentNetworkGuardrail})
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.ListGuardrails(context.Background())
+		require.NoError(t, err)
+		assert.Len(t, ret, 1)
+		assert.Equal(t, testAgentNetworkGuardrail, ret[0])
+	})
+}
+
+func TestAgentNetwork_GetGuardrail_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/guardrails/aingr_test", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(testAgentNetworkGuardrail)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetGuardrail(context.Background(), "aingr_test")
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkGuardrail, *ret)
+	})
+}
+
+func TestAgentNetwork_CreateGuardrail_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/guardrails", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "POST", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkGuardrail)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.CreateGuardrail(context.Background(), api.PostApiAgentNetworkGuardrailsJSONRequestBody{
+			Name: "No secrets",
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkGuardrail, *ret)
+	})
+}
+
+func TestAgentNetwork_UpdateGuardrail_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/guardrails/aingr_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "PUT", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkGuardrail)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.UpdateGuardrail(context.Background(), "aingr_test", api.PutApiAgentNetworkGuardrailsGuardrailIdJSONRequestBody{
+			Name: "No secrets",
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkGuardrail, *ret)
+	})
+}
+
+func TestAgentNetwork_DeleteGuardrail_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/guardrails/aingr_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "DELETE", r.Method)
+			_, err := w.Write([]byte("{}"))
+			require.NoError(t, err)
+		})
+		err := c.AgentNetwork.DeleteGuardrail(context.Background(), "aingr_test")
+		require.NoError(t, err)
+	})
+}
+
+func TestAgentNetwork_ListBudgetRules_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/budget-rules", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal([]api.AgentNetworkBudgetRule{testAgentNetworkBudgetRule})
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.ListBudgetRules(context.Background())
+		require.NoError(t, err)
+		assert.Len(t, ret, 1)
+		assert.Equal(t, testAgentNetworkBudgetRule, ret[0])
+	})
+}
+
+func TestAgentNetwork_GetBudgetRule_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/budget-rules/ainbud_test", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(testAgentNetworkBudgetRule)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetBudgetRule(context.Background(), "ainbud_test")
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkBudgetRule, *ret)
+	})
+}
+
+func TestAgentNetwork_CreateBudgetRule_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/budget-rules", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "POST", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkBudgetRule)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.CreateBudgetRule(context.Background(), api.PostApiAgentNetworkBudgetRulesJSONRequestBody{
+			Name: "Org monthly ceiling",
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkBudgetRule, *ret)
+	})
+}
+
+func TestAgentNetwork_UpdateBudgetRule_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/budget-rules/ainbud_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "PUT", r.Method)
+			retBytes, _ := json.Marshal(testAgentNetworkBudgetRule)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.UpdateBudgetRule(context.Background(), "ainbud_test", api.PutApiAgentNetworkBudgetRulesRuleIdJSONRequestBody{
+			Name: "Org monthly ceiling",
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkBudgetRule, *ret)
+	})
+}
+
+func TestAgentNetwork_DeleteBudgetRule_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/budget-rules/ainbud_test", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "DELETE", r.Method)
+			_, err := w.Write([]byte("{}"))
+			require.NoError(t, err)
+		})
+		err := c.AgentNetwork.DeleteBudgetRule(context.Background(), "ainbud_test")
+		require.NoError(t, err)
+	})
+}
+
+func TestAgentNetwork_GetSettings_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/settings", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(testAgentNetworkSettings)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetSettings(context.Background())
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkSettings, *ret)
+	})
+}
+
+// TestAgentNetwork_GetSettings_UnbootstrappedDefaults pins the settings-read
+// contract: an unbootstrapped account answers 200 with the defaults and empty
+// cluster/subdomain/endpoint, which the client passes through untouched.
+func TestAgentNetwork_GetSettings_UnbootstrappedDefaults(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/settings", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(api.AgentNetworkSettings{
+				EnableLogCollection:    true,
+				AccessLogRetentionDays: ptr(30),
+			})
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetSettings(context.Background())
+		require.NoError(t, err)
+		assert.Empty(t, ret.Endpoint, "empty endpoint is the not-bootstrapped signal")
+		assert.True(t, ret.EnableLogCollection, "defaults must pass through")
+	})
+}
+
+func TestAgentNetwork_GetSettings_Err(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/settings", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(util.ErrorResponse{Message: "no", Code: 403})
+			w.WriteHeader(403)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		_, err := c.AgentNetwork.GetSettings(context.Background())
+		require.Error(t, err)
+		assert.Equal(t, "no", err.Error())
+	})
+}
+
+// TestAgentNetwork_GetSettings_LegacyNullBody pins the compatibility shim for
+// management servers that answered 200 with a JSON null body before the
+// defaults contract: the client translates that shape into an IsNotFound
+// error instead of returning a bogus zero-valued settings object or
+// fabricating defaults the server never stated.
+func TestAgentNetwork_GetSettings_LegacyNullBody(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/settings", func(w http.ResponseWriter, r *http.Request) {
+			_, err := w.Write([]byte("null"))
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.GetSettings(context.Background())
+		require.Error(t, err)
+		assert.Nil(t, ret)
+		assert.True(t, rest.IsNotFound(err), "the legacy 200+null shape must surface as IsNotFound")
+	})
+}
+
+func TestAgentNetwork_UpdateSettings_200(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/settings", func(w http.ResponseWriter, r *http.Request) {
+			assert.Equal(t, "PUT", r.Method)
+			reqBytes, err := io.ReadAll(r.Body)
+			require.NoError(t, err)
+			var req api.PutApiAgentNetworkSettingsJSONRequestBody
+			require.NoError(t, json.Unmarshal(reqBytes, &req))
+			require.NotNil(t, req.Cluster, "bootstrap cluster must be on the wire")
+			assert.Equal(t, "eu.proxy.netbird.io", *req.Cluster)
+			assert.True(t, req.EnableLogCollection)
+			retBytes, _ := json.Marshal(testAgentNetworkSettings)
+			_, err = w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		ret, err := c.AgentNetwork.UpdateSettings(context.Background(), api.PutApiAgentNetworkSettingsJSONRequestBody{
+			Cluster:             ptr("eu.proxy.netbird.io"),
+			EnableLogCollection: true,
+		})
+		require.NoError(t, err)
+		assert.Equal(t, testAgentNetworkSettings, *ret)
+	})
+}
+
+func TestAgentNetwork_UpdateSettings_Err(t *testing.T) {
+	withMockClient(func(c *rest.Client, mux *http.ServeMux) {
+		mux.HandleFunc("/api/agent-network/settings", func(w http.ResponseWriter, r *http.Request) {
+			retBytes, _ := json.Marshal(util.ErrorResponse{Message: "cluster is immutable once assigned (current: eu.proxy.netbird.io)", Code: 422})
+			w.WriteHeader(422)
+			_, err := w.Write(retBytes)
+			require.NoError(t, err)
+		})
+		_, err := c.AgentNetwork.UpdateSettings(context.Background(), api.PutApiAgentNetworkSettingsJSONRequestBody{
+			Cluster: ptr("us.proxy.netbird.io"),
+		})
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "immutable")
+	})
+}

shared/management/client/rest/client.go

@@ -147,6 +147,10 @@ type Client struct {
 	// ReverseProxyTokens account-scoped proxy access tokens used to register
 	// self-hosted (bring-your-own-proxy) `netbird proxy` instances.
 	ReverseProxyTokens *ReverseProxyTokensAPI
+
+	// AgentNetwork NetBird Agent Network (AI/LLM gateway) APIs: catalog,
+	// providers, policies, guardrails, budget rules and account settings.
+	AgentNetwork *AgentNetworkAPI
 }
 
 // New initialize new Client instance using PAT token
@@ -209,6 +213,7 @@ func (c *Client) initialize() {
 	c.ReverseProxyClusters = &ReverseProxyClustersAPI{c}
 	c.ReverseProxyDomains = &ReverseProxyDomainsAPI{c}
 	c.ReverseProxyTokens = &ReverseProxyTokensAPI{c}
+	c.AgentNetwork = &AgentNetworkAPI{c}
 }
 
 // NewRequest creates and executes new management API request

shared/management/http/api/openapi.yml

@@ -5149,12 +5149,12 @@ components:
         identity_header_user_id:
           type: string
           description: |
-            Wire header name the proxy stamps with the caller's display identity (user email or peer name) when the catalog entry's HeaderPair is `customizable`. Empty disables stamping for this dimension. Ignored when the catalog entry has a fixed HeaderPair (e.g. LiteLLM, Portkey). Used today by Bifrost: typical values are `x-bf-lh-netbird_user_id` (always-on log metadata) or `x-bf-dim-netbird_user_id` (Prometheus / OTEL — requires the label to be pre-declared in the gateway's `client.prometheus_labels` config).
+            Wire header name the proxy stamps with the caller's display identity (user email or peer name) when the catalog entry's HeaderPair is `customizable`. Always present in responses; empty disables stamping for this dimension. Ignored when the catalog entry has a fixed HeaderPair (e.g. LiteLLM, Portkey). Used today by Bifrost: typical values are `x-bf-lh-netbird_user_id` (always-on log metadata) or `x-bf-dim-netbird_user_id` (Prometheus / OTEL — requires the label to be pre-declared in the gateway's `client.prometheus_labels` config).
           example: "x-bf-dim-netbird_user_id"
         identity_header_groups:
           type: string
           description: |
-            Wire header name the proxy stamps with the caller's NetBird groups as a comma-separated list (sorted) when the catalog entry's HeaderPair is `customizable`. Empty disables stamping for this dimension. Same per-catalog semantics as `identity_header_user_id`.
+            Wire header name the proxy stamps with the caller's NetBird groups as a comma-separated list (sorted) when the catalog entry's HeaderPair is `customizable`. Always present in responses; empty disables stamping for this dimension. Same per-catalog semantics as `identity_header_user_id`.
           example: "x-bf-dim-netbird_groups"
         enabled:
           type: boolean
@@ -5186,6 +5186,8 @@ components:
         - name
         - upstream_url
         - models
+        - identity_header_user_id
+        - identity_header_groups
         - enabled
         - skip_tls_verification
         - metadata_disabled
@@ -5222,7 +5224,7 @@ components:
         extra_values:
           type: object
           description: |
-            Operator-typed values for catalog-declared extra headers (see AgentNetworkProvider.extra_values). When present on a request, the whole map replaces the stored values. Empty strings drop the corresponding key.
+            Operator-typed values for catalog-declared extra headers (see AgentNetworkProvider.extra_values). The request's map replaces the stored values; empty strings drop the corresponding key.
           additionalProperties:
             type: string
           example:
@@ -5230,12 +5232,12 @@ components:
         identity_header_user_id:
           type: string
           description: |
-            Wire header name for the caller's display identity. See AgentNetworkProvider.identity_header_user_id. When omitted on a request, the stored value is left unchanged; pass an empty string explicitly to clear it (which disables stamping for this dimension).
+            Wire header name for the caller's display identity. See AgentNetworkProvider.identity_header_user_id. Empty or omitted disables stamping for this dimension.
           example: "x-bf-dim-netbird_user_id"
         identity_header_groups:
           type: string
           description: |
-            Wire header name for the caller's groups CSV. See AgentNetworkProvider.identity_header_groups. Same omit / empty semantics as `identity_header_user_id`.
+            Wire header name for the caller's groups CSV. See AgentNetworkProvider.identity_header_groups. Same semantics as `identity_header_user_id`.
           example: "x-bf-dim-netbird_groups"
         enabled:
           type: boolean
@@ -5243,11 +5245,11 @@ components:
           example: true
         skip_tls_verification:
           type: boolean
-          description: Skip upstream TLS certificate verification when the proxy dials this provider's URL. For self-hosted / internal gateways behind a private or self-signed certificate. Defaults to false. When omitted on update, the stored value is left unchanged.
+          description: Skip upstream TLS certificate verification when the proxy dials this provider's URL. For self-hosted / internal gateways behind a private or self-signed certificate. Defaults to false.
           example: false
         metadata_disabled:
           type: boolean
-          description: Disable identity metadata injection (the caller's user + authorizing group) for this provider. Defaults to false (metadata is injected). When omitted on update, the stored value is left unchanged.
+          description: Disable identity metadata injection (the caller's user + authorizing group) for this provider. Defaults to false (metadata is injected).
           example: false
       required:
         - provider_id
@@ -6191,19 +6193,19 @@ components:
         - cache_cost_usd
     AgentNetworkSettings:
       type: object
-      description: Per-account Agent Network gateway settings. One row per account; cluster and subdomain are auto-assigned on first provider create and immutable thereafter.
+      description: Per-account Agent Network gateway settings. One row per account; cluster and subdomain are assigned at bootstrap and immutable thereafter. Before bootstrap the account reads as the default values with empty cluster, subdomain and endpoint.
       properties:
         cluster:
           type: string
-          description: Address of the NetBird proxy cluster fronting this account's agent-network endpoint.
+          description: Address of the NetBird proxy cluster fronting this account's agent-network endpoint. Empty until the account is bootstrapped.
           example: "eu.proxy.netbird.io"
         subdomain:
           type: string
-          description: Auto-generated DNS-safe label that prefixes the cluster to form the agent-network endpoint.
+          description: Auto-generated DNS-safe label that prefixes the cluster to form the agent-network endpoint. Empty until the account is bootstrapped.
           example: "violet"
         endpoint:
           type: string
-          description: Bare hostname agents call for this account, computed as `<subdomain>.<cluster>`.
+          description: Bare hostname agents call for this account, computed as `<subdomain>.<cluster>`. Empty until the account is bootstrapped.
           example: "violet.eu.proxy.netbird.io"
         enable_log_collection:
           type: boolean
@@ -6224,13 +6226,13 @@ components:
         created_at:
           type: string
           format: date-time
-          description: Timestamp when the settings row was created.
+          description: Timestamp when the settings row was created. Absent until the account is bootstrapped.
           readOnly: true
           example: "2026-04-26T10:30:00Z"
         updated_at:
           type: string
           format: date-time
-          description: Timestamp when the settings row was last updated.
+          description: Timestamp when the settings row was last updated. Absent until the account is bootstrapped.
           readOnly: true
           example: "2026-04-26T10:30:00Z"
       required:
@@ -6240,12 +6242,14 @@ components:
         - enable_log_collection
         - enable_prompt_collection
         - redact_pii
-        - created_at
-        - updated_at
     AgentNetworkSettingsRequest:
       type: object
-      description: Mutable account-level Agent Network settings. Cluster and subdomain are immutable and not accepted here.
+      description: Account-level Agent Network settings update. The request replaces every mutable field. `cluster` additionally bootstraps the per-account settings row when the account does not have one yet; the subdomain is always server-assigned.
       properties:
+        cluster:
+          type: string
+          description: Address of the NetBird proxy cluster fronting this account's agent-network endpoint. When the account has no settings row yet, providing it bootstraps the row (assigning the subdomain that forms the agent endpoint). The cluster is immutable once assigned — later updates must omit it or send the assigned value; any other value is rejected.
+          example: "eu.proxy.netbird.io"
         enable_log_collection:
           type: boolean
           description: Whether per-request access-log entries are collected for this account's agent-network traffic.
@@ -13690,7 +13694,7 @@ paths:
   /api/agent-network/settings:
     get:
       summary: Retrieve Agent Network settings
-      description: Returns the per-account Agent Network gateway settings (cluster, subdomain, endpoint). Returns 404 when no provider has been created yet — settings are lazily bootstrapped on first provider create.
+      description: Returns the per-account Agent Network gateway settings (cluster, subdomain, endpoint). Before the account is bootstrapped — on first provider create (`bootstrap_cluster`) or via PUT with `cluster` — the response carries the default values with empty cluster, subdomain and endpoint.
       tags: [ Agent Network ]
       security:
         - BearerAuth: [ ]
@@ -13706,13 +13710,11 @@ paths:
           "$ref": "#/components/responses/requires_authentication"
         '403':
           "$ref": "#/components/responses/forbidden"
-        '404':
-          "$ref": "#/components/responses/not_found"
         '500':
           "$ref": "#/components/responses/internal_error"
     put:
       summary: Update Agent Network settings
-      description: Updates the mutable account-level Agent Network settings (collection toggles). Cluster and subdomain are immutable and ignored if sent. Returns 404 when settings have not been bootstrapped (no provider created yet).
+      description: Updates the account-level Agent Network settings; the request replaces every mutable field (collection toggles and retention). When the account has no settings row yet, providing `cluster` bootstraps it (assigning the subdomain that forms the agent endpoint); without `cluster` the request returns 404. Sending a `cluster` different from the assigned one is rejected (the cluster is immutable once assigned). The subdomain is always server-assigned and immutable.
       tags: [ Agent Network ]
       security:
         - BearerAuth: [ ]
@@ -13738,6 +13740,8 @@ paths:
           "$ref": "#/components/responses/forbidden"
         '404':
           "$ref": "#/components/responses/not_found"
+        '422':
+          "$ref": "#/components/responses/validation_failed"
         '500':
           "$ref": "#/components/responses/internal_error"
   /api/agent-network/budget-rules:

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

@@ -2275,11 +2275,11 @@ type AgentNetworkProvider struct {
 	// Id Provider ID
 	Id string `json:"id"`
 
-	// IdentityHeaderGroups Wire header name the proxy stamps with the caller's NetBird groups as a comma-separated list (sorted) when the catalog entry's HeaderPair is `customizable`. Empty disables stamping for this dimension. Same per-catalog semantics as `identity_header_user_id`.
-	IdentityHeaderGroups *string `json:"identity_header_groups,omitempty"`
+	// IdentityHeaderGroups Wire header name the proxy stamps with the caller's NetBird groups as a comma-separated list (sorted) when the catalog entry's HeaderPair is `customizable`. Always present in responses; empty disables stamping for this dimension. Same per-catalog semantics as `identity_header_user_id`.
+	IdentityHeaderGroups string `json:"identity_header_groups"`
 
-	// IdentityHeaderUserId Wire header name the proxy stamps with the caller's display identity (user email or peer name) when the catalog entry's HeaderPair is `customizable`. Empty disables stamping for this dimension. Ignored when the catalog entry has a fixed HeaderPair (e.g. LiteLLM, Portkey). Used today by Bifrost: typical values are `x-bf-lh-netbird_user_id` (always-on log metadata) or `x-bf-dim-netbird_user_id` (Prometheus / OTEL — requires the label to be pre-declared in the gateway's `client.prometheus_labels` config).
-	IdentityHeaderUserId *string `json:"identity_header_user_id,omitempty"`
+	// IdentityHeaderUserId Wire header name the proxy stamps with the caller's display identity (user email or peer name) when the catalog entry's HeaderPair is `customizable`. Always present in responses; empty disables stamping for this dimension. Ignored when the catalog entry has a fixed HeaderPair (e.g. LiteLLM, Portkey). Used today by Bifrost: typical values are `x-bf-lh-netbird_user_id` (always-on log metadata) or `x-bf-dim-netbird_user_id` (Prometheus / OTEL — requires the label to be pre-declared in the gateway's `client.prometheus_labels` config).
+	IdentityHeaderUserId string `json:"identity_header_user_id"`
 
 	// MetadataDisabled Whether identity metadata injection is disabled for this provider. When enabled (the default), the proxy stamps the caller's user and authorizing group onto upstream requests as provider-specific metadata (e.g. AWS Bedrock's X-Amzn-Bedrock-Request-Metadata header). Set true to suppress it.
 	MetadataDisabled bool `json:"metadata_disabled"`
@@ -2335,16 +2335,16 @@ type AgentNetworkProviderRequest struct {
 	// Enabled Whether the provider is enabled. Defaults to true on create.
 	Enabled *bool `json:"enabled,omitempty"`
 
-	// ExtraValues Operator-typed values for catalog-declared extra headers (see AgentNetworkProvider.extra_values). When present on a request, the whole map replaces the stored values. Empty strings drop the corresponding key.
+	// ExtraValues Operator-typed values for catalog-declared extra headers (see AgentNetworkProvider.extra_values). The request's map replaces the stored values; empty strings drop the corresponding key.
 	ExtraValues *map[string]string `json:"extra_values,omitempty"`
 
-	// IdentityHeaderGroups Wire header name for the caller's groups CSV. See AgentNetworkProvider.identity_header_groups. Same omit / empty semantics as `identity_header_user_id`.
+	// IdentityHeaderGroups Wire header name for the caller's groups CSV. See AgentNetworkProvider.identity_header_groups. Same semantics as `identity_header_user_id`.
 	IdentityHeaderGroups *string `json:"identity_header_groups,omitempty"`
 
-	// IdentityHeaderUserId Wire header name for the caller's display identity. See AgentNetworkProvider.identity_header_user_id. When omitted on a request, the stored value is left unchanged; pass an empty string explicitly to clear it (which disables stamping for this dimension).
+	// IdentityHeaderUserId Wire header name for the caller's display identity. See AgentNetworkProvider.identity_header_user_id. Empty or omitted disables stamping for this dimension.
 	IdentityHeaderUserId *string `json:"identity_header_user_id,omitempty"`
 
-	// MetadataDisabled Disable identity metadata injection (the caller's user + authorizing group) for this provider. Defaults to false (metadata is injected). When omitted on update, the stored value is left unchanged.
+	// MetadataDisabled Disable identity metadata injection (the caller's user + authorizing group) for this provider. Defaults to false (metadata is injected).
 	MetadataDisabled *bool `json:"metadata_disabled,omitempty"`
 
 	// Models Models exposed through this endpoint, with the operator's per-1k input/output prices. Empty means all catalog models are allowed at catalog prices.
@@ -2356,22 +2356,22 @@ type AgentNetworkProviderRequest struct {
 	// ProviderId Catalog identifier for the upstream AI provider (e.g. openai_api, anthropic_api, azure_openai_api, bedrock_api, vertex_ai_api, mistral_api, custom).
 	ProviderId string `json:"provider_id"`
 
-	// SkipTlsVerification Skip upstream TLS certificate verification when the proxy dials this provider's URL. For self-hosted / internal gateways behind a private or self-signed certificate. Defaults to false. When omitted on update, the stored value is left unchanged.
+	// SkipTlsVerification Skip upstream TLS certificate verification when the proxy dials this provider's URL. For self-hosted / internal gateways behind a private or self-signed certificate. Defaults to false.
 	SkipTlsVerification *bool `json:"skip_tls_verification,omitempty"`
 
 	// UpstreamUrl Full upstream URL (with scheme) that NetBird forwards traffic to.
 	UpstreamUrl string `json:"upstream_url"`
 }
 
-// AgentNetworkSettings Per-account Agent Network gateway settings. One row per account; cluster and subdomain are auto-assigned on first provider create and immutable thereafter.
+// AgentNetworkSettings Per-account Agent Network gateway settings. One row per account; cluster and subdomain are assigned at bootstrap and immutable thereafter. Before bootstrap the account reads as the default values with empty cluster, subdomain and endpoint.
 type AgentNetworkSettings struct {
 	// AccessLogRetentionDays Days to retain full access-log rows; older rows are swept. 0 or less means keep indefinitely. Usage records are retained independently.
 	AccessLogRetentionDays *int `json:"access_log_retention_days,omitempty"`
 
-	// Cluster Address of the NetBird proxy cluster fronting this account's agent-network endpoint.
+	// Cluster Address of the NetBird proxy cluster fronting this account's agent-network endpoint. Empty until the account is bootstrapped.
 	Cluster string `json:"cluster"`
 
-	// CreatedAt Timestamp when the settings row was created.
+	// CreatedAt Timestamp when the settings row was created. Absent until the account is bootstrapped.
 	CreatedAt *time.Time `json:"created_at,omitempty"`
 
 	// EnableLogCollection Whether per-request access-log entries are collected for this account's agent-network traffic.
@@ -2380,24 +2380,27 @@ type AgentNetworkSettings struct {
 	// EnablePromptCollection Master switch for request/response prompt capture. Capture runs only when this is on AND a policy guardrail also enables it.
 	EnablePromptCollection bool `json:"enable_prompt_collection"`
 
-	// Endpoint Bare hostname agents call for this account, computed as `<subdomain>.<cluster>`.
+	// Endpoint Bare hostname agents call for this account, computed as `<subdomain>.<cluster>`. Empty until the account is bootstrapped.
 	Endpoint string `json:"endpoint"`
 
 	// RedactPii Whether captured prompts have PII redacted. Effective redaction is the OR of this and any policy guardrail's redact setting.
 	RedactPii bool `json:"redact_pii"`
 
-	// Subdomain Auto-generated DNS-safe label that prefixes the cluster to form the agent-network endpoint.
+	// Subdomain Auto-generated DNS-safe label that prefixes the cluster to form the agent-network endpoint. Empty until the account is bootstrapped.
 	Subdomain string `json:"subdomain"`
 
-	// UpdatedAt Timestamp when the settings row was last updated.
+	// UpdatedAt Timestamp when the settings row was last updated. Absent until the account is bootstrapped.
 	UpdatedAt *time.Time `json:"updated_at,omitempty"`
 }
 
-// AgentNetworkSettingsRequest Mutable account-level Agent Network settings. Cluster and subdomain are immutable and not accepted here.
+// AgentNetworkSettingsRequest Account-level Agent Network settings update. The request replaces every mutable field. `cluster` additionally bootstraps the per-account settings row when the account does not have one yet; the subdomain is always server-assigned.
 type AgentNetworkSettingsRequest struct {
 	// AccessLogRetentionDays Days to retain full access-log rows; older rows are swept. 0 or less means keep indefinitely.
 	AccessLogRetentionDays *int `json:"access_log_retention_days,omitempty"`
 
+	// Cluster Address of the NetBird proxy cluster fronting this account's agent-network endpoint. When the account has no settings row yet, providing it bootstraps the row (assigning the subdomain that forms the agent endpoint). The cluster is immutable once assigned — later updates must omit it or send the assigned value; any other value is rejected.
+	Cluster *string `json:"cluster,omitempty"`
+
 	// EnableLogCollection Whether per-request access-log entries are collected for this account's agent-network traffic.
 	EnableLogCollection bool `json:"enable_log_collection"`
 

shared/management/networkmap/decode.go

@@ -228,15 +228,17 @@ func DecodeEnvelope(env *proto.NetworkMapEnvelope) (*types.NetworkMapComponents,
 	return c, nil
 }
 
+// decodeAccountNetwork never returns nil — Calculate() dereferences
+// c.Network unconditionally, and servers that predate the fix omit the field
+// entirely from the empty-components envelope.
 func decodeAccountNetwork(an *proto.AccountNetwork) *types.Network {
+	n := &types.Network{}
 	if an == nil {
-		return nil
-	}
-	n := &types.Network{
-		Identifier: an.Identifier,
-		Dns:        an.Dns,
-		Serial:     an.Serial,
+		return n
 	}
+	n.Identifier = an.Identifier
+	n.Dns = an.Dns
+	n.Serial = an.Serial
 	if an.NetCidr != "" {
 		if _, ipnet, err := net.ParseCIDR(an.NetCidr); err == nil && ipnet != nil {
 			n.Net = *ipnet

shared/management/networkmap/envelope_test.go

@@ -221,6 +221,66 @@ func TestEnvelopeRoundTrip_AllGroupShortCircuitParity(t *testing.T) {
 		"client-side Calculate must connect the same remote peers as the server")
 }
 
+// TestEnvelopeToNetworkMap_EmptyComponents covers the graceful-degrade path
+// the server takes for a peer that is missing from the account or absent from
+// the validated-peers map. The legacy server short-circuited before
+// Calculate() and shipped a NetworkMap carrying only the account Network; the
+// components path runs Calculate() on the client instead, so the envelope must
+// carry Network or the client panics dereferencing a nil *types.Network.
+func TestEnvelopeToNetworkMap_EmptyComponents(t *testing.T) {
+	localPeerKey := randomWgKey(t)
+	c := types.EmptyNetworkMapComponents(&types.NetworkMapComponents{
+		PeerID: "peer-A",
+		Network: &types.Network{
+			Identifier: "net-empty",
+			Net:        net.IPNet{IP: net.IP{100, 64, 0, 0}, Mask: net.CIDRMask(10, 32)},
+			Serial:     7,
+		},
+		Peers: map[string]*types.ComponentPeer{
+			"peer-A": {ID: "peer-A", Key: localPeerKey, IP: netip.AddrFrom4([4]byte{100, 64, 0, 1})},
+		},
+	})
+
+	envelope := mgmtgrpc.EncodeNetworkMapEnvelope(mgmtgrpc.ComponentsEnvelopeInput{
+		Components: c,
+		DNSDomain:  "netbird.cloud",
+	})
+	require.NotNil(t, envelope.GetFull().Network, "empty envelope must carry the account Network")
+
+	wire, err := goproto.Marshal(envelope)
+	require.NoError(t, err, "marshal envelope")
+	var decoded proto.NetworkMapEnvelope
+	require.NoError(t, goproto.Unmarshal(wire, &decoded), "unmarshal envelope")
+
+	result, err := nbnetworkmap.EnvelopeToNetworkMap(context.Background(), &decoded, localPeerKey, "netbird.cloud")
+	require.NoError(t, err, "EnvelopeToNetworkMap must degrade gracefully on empty components")
+	require.Equal(t, uint64(7), result.NetworkMap.Serial)
+	require.Empty(t, result.NetworkMap.RemotePeers, "unvalidated peer connects to nobody")
+}
+
+// TestEnvelopeToNetworkMap_MissingNetwork simulates a server that omits
+// AccountNetwork from the envelope. Clients must degrade rather than panic, so
+// they survive talking to a management server that predates the encoder fix.
+func TestEnvelopeToNetworkMap_MissingNetwork(t *testing.T) {
+	c, localPeerKey := buildSmokeComponents(t)
+
+	envelope := mgmtgrpc.EncodeNetworkMapEnvelope(mgmtgrpc.ComponentsEnvelopeInput{
+		Components: c,
+		DNSDomain:  "netbird.cloud",
+	})
+	envelope.GetFull().Network = nil
+
+	wire, err := goproto.Marshal(envelope)
+	require.NoError(t, err, "marshal envelope")
+	var decoded proto.NetworkMapEnvelope
+	require.NoError(t, goproto.Unmarshal(wire, &decoded), "unmarshal envelope")
+
+	result, err := nbnetworkmap.EnvelopeToNetworkMap(context.Background(), &decoded, localPeerKey, "netbird.cloud")
+	require.NoError(t, err, "a missing AccountNetwork must not panic the client")
+	require.NotNil(t, result.Components.Network)
+	require.NotEmpty(t, result.NetworkMap.RemotePeers, "the rest of the snapshot stays usable")
+}
+
 // buildSmokeComponents returns a minimal NetworkMapComponents (2 peers, 1
 // group, 1 allow policy) plus the receiving peer's WG public key. Sufficient
 // to validate the encode → marshal → decode → Calculate pipeline produces

shared/signal/client/client.go

@@ -52,6 +52,11 @@ type CredentialPayload struct {
 	Credential      *Credential
 	RosenpassPubKey []byte
 	RosenpassAddr   string
+	// MlkemPayload is the opaque post-quantum KEM handshake message riding this
+	// OFFER/ANSWER (see Body.mlkemPayload). Nil when not running the PQ exchange.
+	MlkemPayload []byte
+	// MlkemPort is the sender's ML-KEM PQ service UDP port (0 when not running).
+	MlkemPort       int
 	RelaySrvAddress string
 	RelaySrvIP      netip.Addr
 	SessionID       []byte
@@ -89,6 +94,13 @@ func MarshalCredential(myKey wgtypes.Key, remoteKey string, p CredentialPayload)
 	if p.RelaySrvIP.IsValid() {
 		body.RelayServerIP = p.RelaySrvIP.Unmap().AsSlice()
 	}
+	if len(p.MlkemPayload) > 0 {
+		body.MlkemPayload = p.MlkemPayload
+	}
+	if p.MlkemPort > 0 {
+		port := uint32(p.MlkemPort)
+		body.MlkemPort = &port
+	}
 	return &proto.Message{
 		Key:       myKey.PublicKey().String(),
 		RemoteKey: remoteKey,

shared/signal/proto/signalexchange.pb.go

@@ -239,6 +239,16 @@ type Body struct {
 	// fallback dial target when DNS resolution of relayServerAddress fails.
 	// SNI/TLS verification still uses relayServerAddress.
 	RelayServerIP []byte `protobuf:"bytes,11,opt,name=relayServerIP,proto3,oneof" json:"relayServerIP,omitempty"`
+	// mlkemPayload carries a post-quantum X25519MLKEM768 handshake message that
+	// seeds the WireGuard PSK, riding this Body's OFFER/ANSWER: on an OFFER it is
+	// the KEM offer, on an ANSWER the KEM answer. It is opaque to signal — the
+	// pqkem library frames and parses it. Absent when the sender does not run the
+	// ML-KEM PQ exchange; unknown to older clients, which ignore it.
+	MlkemPayload []byte `protobuf:"bytes,12,opt,name=mlkemPayload,proto3,oneof" json:"mlkemPayload,omitempty"`
+	// mlkemPort is the UDP port of the sender's ML-KEM PQ service, bound on its
+	// WireGuard overlay IP. Peers send subsequent rekey messages there over the
+	// data path. Zero/absent when the ML-KEM PQ exchange is not running.
+	MlkemPort *uint32 `protobuf:"varint,13,opt,name=mlkemPort,proto3,oneof" json:"mlkemPort,omitempty"`
 }
 
 func (x *Body) Reset() {
@@ -343,6 +353,20 @@ func (x *Body) GetRelayServerIP() []byte {
 	return nil
 }
 
+func (x *Body) GetMlkemPayload() []byte {
+	if x != nil {
+		return x.MlkemPayload
+	}
+	return nil
+}
+
+func (x *Body) GetMlkemPort() uint32 {
+	if x != nil && x.MlkemPort != nil {
+		return *x.MlkemPort
+	}
+	return 0
+}
+
 // Mode indicates a connection mode
 type Mode struct {
 	state         protoimpl.MessageState
@@ -466,7 +490,7 @@ var file_signalexchange_proto_rawDesc = []byte{
 	0x52, 0x09, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x4b, 0x65, 0x79, 0x12, 0x28, 0x0a, 0x04, 0x62,
 	0x6f, 0x64, 0x79, 0x18, 0x04, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x14, 0x2e, 0x73, 0x69, 0x67, 0x6e,
 	0x61, 0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x42, 0x6f, 0x64, 0x79, 0x52,
-	0x04, 0x62, 0x6f, 0x64, 0x79, 0x22, 0xd2, 0x04, 0x0a, 0x04, 0x42, 0x6f, 0x64, 0x79, 0x12, 0x2d,
+	0x04, 0x62, 0x6f, 0x64, 0x79, 0x22, 0xbd, 0x05, 0x0a, 0x04, 0x42, 0x6f, 0x64, 0x79, 0x12, 0x2d,
 	0x0a, 0x04, 0x74, 0x79, 0x70, 0x65, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x19, 0x2e, 0x73,
 	0x69, 0x67, 0x6e, 0x61, 0x6c, 0x65, 0x78, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x2e, 0x42, 0x6f,
 	0x64, 0x79, 0x2e, 0x54, 0x79, 0x70, 0x65, 0x52, 0x04, 0x74, 0x79, 0x70, 0x65, 0x12, 0x18, 0x0a,
@@ -494,39 +518,46 @@ var file_signalexchange_proto_rawDesc = []byte{
 	0x52, 0x09, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x49, 0x64, 0x88, 0x01, 0x01, 0x12, 0x29,
 	0x0a, 0x0d, 0x72, 0x65, 0x6c, 0x61, 0x79, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x49, 0x50, 0x18,
 	0x0b, 0x20, 0x01, 0x28, 0x0c, 0x48, 0x02, 0x52, 0x0d, 0x72, 0x65, 0x6c, 0x61, 0x79, 0x53, 0x65,
-	0x72, 0x76, 0x65, 0x72, 0x49, 0x50, 0x88, 0x01, 0x01, 0x22, 0x52, 0x0a, 0x04, 0x54, 0x79, 0x70,
-	0x65, 0x12, 0x09, 0x0a, 0x05, 0x4f, 0x46, 0x46, 0x45, 0x52, 0x10, 0x00, 0x12, 0x0a, 0x0a, 0x06,
-	0x41, 0x4e, 0x53, 0x57, 0x45, 0x52, 0x10, 0x01, 0x12, 0x0d, 0x0a, 0x09, 0x43, 0x41, 0x4e, 0x44,
-	0x49, 0x44, 0x41, 0x54, 0x45, 0x10, 0x02, 0x12, 0x08, 0x0a, 0x04, 0x4d, 0x4f, 0x44, 0x45, 0x10,
-	0x04, 0x12, 0x0b, 0x0a, 0x07, 0x47, 0x4f, 0x5f, 0x49, 0x44, 0x4c, 0x45, 0x10, 0x05, 0x12, 0x0d,
-	0x0a, 0x09, 0x48, 0x45, 0x41, 0x52, 0x54, 0x42, 0x45, 0x41, 0x54, 0x10, 0x06, 0x42, 0x15, 0x0a,
-	0x13, 0x5f, 0x72, 0x65, 0x6c, 0x61, 0x79, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x41, 0x64, 0x64,
-	0x72, 0x65, 0x73, 0x73, 0x42, 0x0c, 0x0a, 0x0a, 0x5f, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e,
-	0x49, 0x64, 0x42, 0x10, 0x0a, 0x0e, 0x5f, 0x72, 0x65, 0x6c, 0x61, 0x79, 0x53, 0x65, 0x72, 0x76,
-	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,
+	0x72, 0x76, 0x65, 0x72, 0x49, 0x50, 0x88, 0x01, 0x01, 0x12, 0x27, 0x0a, 0x0c, 0x6d, 0x6c, 0x6b,
+	0x65, 0x6d, 0x50, 0x61, 0x79, 0x6c, 0x6f, 0x61, 0x64, 0x18, 0x0c, 0x20, 0x01, 0x28, 0x0c, 0x48,
+	0x03, 0x52, 0x0c, 0x6d, 0x6c, 0x6b, 0x65, 0x6d, 0x50, 0x61, 0x79, 0x6c, 0x6f, 0x61, 0x64, 0x88,
+	0x01, 0x01, 0x12, 0x21, 0x0a, 0x09, 0x6d, 0x6c, 0x6b, 0x65, 0x6d, 0x50, 0x6f, 0x72, 0x74, 0x18,
+	0x0d, 0x20, 0x01, 0x28, 0x0d, 0x48, 0x04, 0x52, 0x09, 0x6d, 0x6c, 0x6b, 0x65, 0x6d, 0x50, 0x6f,
+	0x72, 0x74, 0x88, 0x01, 0x01, 0x22, 0x52, 0x0a, 0x04, 0x54, 0x79, 0x70, 0x65, 0x12, 0x09, 0x0a,
+	0x05, 0x4f, 0x46, 0x46, 0x45, 0x52, 0x10, 0x00, 0x12, 0x0a, 0x0a, 0x06, 0x41, 0x4e, 0x53, 0x57,
+	0x45, 0x52, 0x10, 0x01, 0x12, 0x0d, 0x0a, 0x09, 0x43, 0x41, 0x4e, 0x44, 0x49, 0x44, 0x41, 0x54,
+	0x45, 0x10, 0x02, 0x12, 0x08, 0x0a, 0x04, 0x4d, 0x4f, 0x44, 0x45, 0x10, 0x04, 0x12, 0x0b, 0x0a,
+	0x07, 0x47, 0x4f, 0x5f, 0x49, 0x44, 0x4c, 0x45, 0x10, 0x05, 0x12, 0x0d, 0x0a, 0x09, 0x48, 0x45,
+	0x41, 0x52, 0x54, 0x42, 0x45, 0x41, 0x54, 0x10, 0x06, 0x42, 0x15, 0x0a, 0x13, 0x5f, 0x72, 0x65,
+	0x6c, 0x61, 0x79, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x41, 0x64, 0x64, 0x72, 0x65, 0x73, 0x73,
+	0x42, 0x0c, 0x0a, 0x0a, 0x5f, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x49, 0x64, 0x42, 0x10,
+	0x0a, 0x0e, 0x5f, 0x72, 0x65, 0x6c, 0x61, 0x79, 0x53, 0x65, 0x72, 0x76, 0x65, 0x72, 0x49, 0x50,
+	0x42, 0x0f, 0x0a, 0x0d, 0x5f, 0x6d, 0x6c, 0x6b, 0x65, 0x6d, 0x50, 0x61, 0x79, 0x6c, 0x6f, 0x61,
+	0x64, 0x42, 0x0c, 0x0a, 0x0a, 0x5f, 0x6d, 0x6c, 0x6b, 0x65, 0x6d, 0x50, 0x6f, 0x72, 0x74, 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, 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,
+	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 (