[client] Clear the login-required latch after installing the new client

Run() released the latch before setState() swapped in the fresh connect
client, so Status() calls landing in that window still read the previous
run's context state — which holds the NeedsLogin that prompted the login
— and re-latched what had just been cleared. The generation guard does
not catch this: the clear precedes the observation, so the generation
matches and the store counts as current. The state-change goroutine
calls Status() on every recorder tick, and the login path emits several,
so the window is ordinary traffic rather than a rare interleaving.

Clear once the replacement client is installed instead.

.goreleaser.yaml

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

client/android/client.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"os"
 	"slices"
+	"strings"
 	"sync"
 	"time"
 
@@ -75,6 +76,24 @@ type Client struct {
 	connectClient *internal.ConnectClient
 	config        *profilemanager.Config
 	cacheDir      string
+
+	stateChangeMu    sync.Mutex
+	stateChangeSubID string
+	eventSub         *peer.EventSubscription
+	// Closed to stop the watch goroutines from delivering buffered items to a
+	// listener that has been removed or replaced. See stopStateChangeWatchLocked.
+	stateChangeDone chan struct{}
+
+	// Latched "the server wants an interactive login": survives the engine
+	// restarts that replace the run loop's context state. See Client.Status.
+	// Guarded by loginRequiredMu together with loginCleared, which counts
+	// clears so a stale observation cannot re-latch over one.
+	loginRequiredMu sync.Mutex
+	loginRequired   bool
+	loginCleared    uint64
+
+	extendMu     sync.Mutex
+	extendCancel context.CancelFunc
 }
 
 func (c *Client) setState(cfg *profilemanager.Config, cacheDir string, cc *internal.ConnectClient) {
@@ -148,11 +167,16 @@ func (c *Client) Run(platformFiles PlatformFiles, urlOpener URLOpener, isAndroid
 	if err != nil {
 		return err
 	}
-
 	// 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)
+	// 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()
+	// still reads the previous run's context state, which holds the NeedsLogin
+	// that prompted this login, and would re-latch what was just cleared.
+	c.clearLoginRequired()
 	return connectClient.RunOnAndroid(c.tunAdapter, c.iFaceDiscover, c.networkChangeListener, slices.Clone(dns.items), dnsReadyListener, stateFile, cacheDir)
 }
 
@@ -299,6 +323,13 @@ func (c *Client) SetInfoLogLevel() {
 // PeersList return with the list of the PeerInfos
 func (c *Client) PeersList() *PeerInfoArray {
 
+	// The recorder only caches transfer counters and handshake times; nothing
+	// refreshes them on its own, so without this they read as zero. The desktop
+	// daemon does the same before serving a full peer status.
+	if err := c.recorder.RefreshWireGuardStats(); err != nil {
+		log.Debugf("failed to refresh WireGuard stats: %v", err)
+	}
+
 	fullStatus := c.recorder.GetFullStatus()
 
 	peerInfos := make([]PeerInfo, len(fullStatus.Peers))
@@ -309,6 +340,20 @@ func (c *Client) PeersList() *PeerInfoArray {
 			FQDN:       p.FQDN,
 			ConnStatus: int(p.ConnStatus),
 			Routes:     PeerRoutes{routes: maps.Keys(p.GetRoutes())},
+
+			PubKey:                     p.PubKey,
+			Latency:                    formatDuration(p.Latency),
+			LatencyMs:                  p.Latency.Milliseconds(),
+			BytesRx:                    p.BytesRx,
+			BytesTx:                    p.BytesTx,
+			ConnStatusUpdate:           formatTime(p.ConnStatusUpdate),
+			Relayed:                    p.Relayed,
+			RosenpassEnabled:           p.RosenpassEnabled,
+			LastWireguardHandshake:     formatTime(p.LastWireguardHandshake),
+			LocalIceCandidateType:      p.LocalIceCandidateType,
+			RemoteIceCandidateType:     p.RemoteIceCandidateType,
+			LocalIceCandidateEndpoint:  p.LocalIceCandidateEndpoint,
+			RemoteIceCandidateEndpoint: p.RemoteIceCandidateEndpoint,
 		}
 		peerInfos[n] = pi
 	}
@@ -439,10 +484,6 @@ func (c *Client) RemoveConnectionListener() {
 	c.recorder.RemoveConnectionListener()
 }
 
-func (c *Client) toggleRoute(command routeCommand) error {
-	return command.toggleRoute()
-}
-
 func (c *Client) getRouteManager() (routemanager.Manager, error) {
 	client := c.getConnectClient()
 	if client == nil {
@@ -462,22 +503,22 @@ func (c *Client) getRouteManager() (routemanager.Manager, error) {
 	return manager, nil
 }
 
-func (c *Client) SelectRoute(route string) error {
+func (c *Client) SelectRoute(id string) error {
 	manager, err := c.getRouteManager()
 	if err != nil {
 		return err
 	}
 
-	return c.toggleRoute(selectRouteCommand{route: route, manager: manager})
+	return manager.SelectRoutes([]route.NetID{route.NetID(id)}, true)
 }
 
-func (c *Client) DeselectRoute(route string) error {
+func (c *Client) DeselectRoute(id string) error {
 	manager, err := c.getRouteManager()
 	if err != nil {
 		return err
 	}
 
-	return c.toggleRoute(deselectRouteCommand{route: route, manager: manager})
+	return manager.DeselectRoutes([]route.NetID{route.NetID(id)})
 }
 
 // getNetworkDomainsFromRoute extracts domains from a route and enriches each domain
@@ -512,3 +553,28 @@ func exportEnvList(list *EnvList) {
 		}
 	}
 }
+
+// formatDuration renders a duration for display, trimming the fractional part
+// to two digits so latencies read as "12.34ms" rather than "12.345678ms".
+func formatDuration(d time.Duration) string {
+	ds := d.String()
+	dotIndex := strings.Index(ds, ".")
+	if dotIndex == -1 {
+		return ds
+	}
+
+	endIndex := min(dotIndex+3, len(ds))
+
+	// Skip the remaining digits so only the unit suffix is appended back.
+	unitStart := endIndex
+	for unitStart < len(ds) && ds[unitStart] >= '0' && ds[unitStart] <= '9' {
+		unitStart++
+	}
+	return ds[:endIndex] + ds[unitStart:]
+}
+
+// formatTime renders a timestamp in UTC using a fixed layout. The zero time is
+// passed through as-is so the UI can recognise it and show "never" instead.
+func formatTime(t time.Time) string {
+	return t.UTC().Format("2006-01-02 15:04:05")
+}

client/android/peer_notifier.go

@@ -12,12 +12,30 @@ const (
 )
 
 // PeerInfo describe information about the peers. It designed for the UI usage
+//
+// The fields below ConnStatus back the peer detail screen. Durations and times
+// are pre-formatted into strings so the UI does not have to know Go's layouts;
+// Latency is additionally exposed as LatencyMs for colour coding.
 type PeerInfo struct {
 	IP         string
 	IPv6       string
 	FQDN       string
 	ConnStatus int
 	Routes     PeerRoutes
+
+	PubKey                     string
+	Latency                    string
+	LatencyMs                  int64
+	BytesRx                    int64
+	BytesTx                    int64
+	ConnStatusUpdate           string
+	Relayed                    bool
+	RosenpassEnabled           bool
+	LastWireguardHandshake     string
+	LocalIceCandidateType      string
+	RemoteIceCandidateType     string
+	LocalIceCandidateEndpoint  string
+	RemoteIceCandidateEndpoint string
 }
 
 func (p *PeerInfo) GetPeerRoutes() *PeerRoutes {

client/android/profile_manager.go

@@ -189,6 +189,19 @@ func (pm *ProfileManager) LogoutProfile(id string) error {
 	return nil
 }
 
+// RenameProfile changes a profile's display name. The profile ID, and therefore
+// its on-disk filename, is left untouched: only the "name" field of the config
+// is rewritten. This works for the default profile too, whose config lives in
+// netbird.cfg rather than under profiles/.
+func (pm *ProfileManager) RenameProfile(id string, newName string) error {
+	if err := pm.serviceMgr.RenameProfile(profilemanager.ID(id), androidUsername, newName); err != nil {
+		return fmt.Errorf("failed to rename profile: %w", err)
+	}
+
+	log.Infof("renamed profile %s to: %s", id, newName)
+	return nil
+}
+
 // RemoveProfile deletes a profile
 func (pm *ProfileManager) RemoveProfile(id string) error {
 	// Use ServiceManager (removes profile from profiles/ directory)

client/android/route_command.go

@@ -1,70 +0,0 @@
-//go:build android
-
-package android
-
-import (
-	"fmt"
-
-	log "github.com/sirupsen/logrus"
-	"golang.org/x/exp/maps"
-
-	"github.com/netbirdio/netbird/client/internal/routemanager"
-	"github.com/netbirdio/netbird/route"
-)
-
-func executeRouteToggle(id string, manager routemanager.Manager,
-	operationName string,
-	routeOperation func(routes []route.NetID, allRoutes []route.NetID) error) error {
-	netID := route.NetID(id)
-	routes := []route.NetID{netID}
-
-	routesMap := manager.GetClientRoutesWithNetID()
-	routes = route.ExpandV6ExitPairs(routes, routesMap)
-
-	log.Debugf("%s with ids: %v", operationName, routes)
-
-	if err := routeOperation(routes, maps.Keys(routesMap)); err != nil {
-		log.Debugf("error when %s: %s", operationName, err)
-		return fmt.Errorf("error %s: %w", operationName, err)
-	}
-
-	manager.TriggerSelection(manager.GetClientRoutes())
-
-	return nil
-}
-
-type routeCommand interface {
-	toggleRoute() error
-}
-
-type selectRouteCommand struct {
-	route   string
-	manager routemanager.Manager
-}
-
-func (s selectRouteCommand) toggleRoute() error {
-	routeSelector := s.manager.GetRouteSelector()
-	if routeSelector == nil {
-		return fmt.Errorf("no route selector available")
-	}
-
-	routeOperation := func(routes []route.NetID, allRoutes []route.NetID) error {
-		return routeSelector.SelectRoutes(routes, true, allRoutes)
-	}
-
-	return executeRouteToggle(s.route, s.manager, "selecting route", routeOperation)
-}
-
-type deselectRouteCommand struct {
-	route   string
-	manager routemanager.Manager
-}
-
-func (d deselectRouteCommand) toggleRoute() error {
-	routeSelector := d.manager.GetRouteSelector()
-	if routeSelector == nil {
-		return fmt.Errorf("no route selector available")
-	}
-
-	return executeRouteToggle(d.route, d.manager, "deselecting route", routeSelector.DeselectRoutes)
-}

client/android/session.go

@@ -0,0 +1,309 @@
+//go:build android
+
+package android
+
+import (
+	"context"
+	"fmt"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal"
+	"github.com/netbirdio/netbird/client/internal/auth"
+	"github.com/netbirdio/netbird/client/internal/auth/sessionwatch"
+	"github.com/netbirdio/netbird/client/internal/peer"
+	cProto "github.com/netbirdio/netbird/client/proto"
+)
+
+// StateChangeListener receives client state notifications.
+//
+// OnStateChanged is a payload-free wake-up whenever the state snapshot
+// changed: connection state, the run-loop status label (e.g. NeedsLogin) or
+// the session deadline. It mirrors the daemon's SubscribeStatus stream
+// trigger — on each signal the consumer pulls the fresh values via
+// Status() / SessionExpiresAtUnix().
+//
+// OnSessionExpiring forwards the engine's session-expiry warnings, fired at
+// sessionwatch.WarningLead before the deadline and again at FinalWarningLead
+// (finalWarning true). The second one is suppressed when the user dismissed
+// the first via DismissSessionWarning. The daemon turns the same events into
+// its tray notification.
+type StateChangeListener interface {
+	OnStateChanged()
+	OnSessionExpiring(expiresAtUnix int64, leadMinutes int64, finalWarning bool)
+}
+
+// Status returns the connect run-loop's status label — the same value the
+// desktop daemon serves in StatusResponse.Status. "NeedsLogin" means the
+// management server rejected the peer and an interactive login is required.
+//
+// The label is latched: the run loop keeps its status in a per-run context
+// state, which a restart replaces with a fresh Idle one, so an engine restart
+// (network change, always-on) would otherwise erase the fact that the peer
+// still needs to log in. Only a successful interactive login or extend clears
+// it — see clearLoginRequired.
+func (c *Client) Status() string {
+	latched, generation := c.loginRequiredState()
+	if latched {
+		return string(internal.StatusNeedsLogin)
+	}
+	cc := c.getConnectClient()
+	if cc == nil {
+		return string(internal.StatusIdle)
+	}
+	status := cc.Status()
+	if status == internal.StatusNeedsLogin {
+		c.latchLoginRequired(generation)
+	}
+	return string(status)
+}
+
+func (c *Client) loginRequiredState() (bool, uint64) {
+	c.loginRequiredMu.Lock()
+	defer c.loginRequiredMu.Unlock()
+	return c.loginRequired, c.loginCleared
+}
+
+// latchLoginRequired records a NeedsLogin observation, unless a clear landed
+// while the caller was reading the run loop's status: cc.Status() is read
+// outside the lock, so a login or extend completing in that window would
+// otherwise be undone by this stale observation, stranding the UI on
+// "login required" over a healthy session.
+func (c *Client) latchLoginRequired(observedGeneration uint64) {
+	c.loginRequiredMu.Lock()
+	defer c.loginRequiredMu.Unlock()
+	if c.loginCleared != observedGeneration {
+		return
+	}
+	c.loginRequired = true
+}
+
+// clearLoginRequired releases the latch after a successful interactive login
+// or session extend, and invalidates any observation already in flight.
+func (c *Client) clearLoginRequired() {
+	c.loginRequiredMu.Lock()
+	defer c.loginRequiredMu.Unlock()
+	c.loginRequired = false
+	c.loginCleared++
+}
+
+// SessionExpiresAtUnix returns the SSO session deadline as unix seconds, or 0
+// when no deadline is known (not SSO-registered, expiry disabled, or the
+// engine has not received one yet). A past value means the session expired.
+// Mirror of StatusResponse.sessionExpiresAt on the desktop daemon.
+func (c *Client) SessionExpiresAtUnix() int64 {
+	deadline := c.recorder.GetSessionExpiresAt()
+	if deadline.IsZero() {
+		return 0
+	}
+	return deadline.Unix()
+}
+
+// SetStateChangeListener registers the state notification listener.
+// Replaces any previously registered listener; remove it with
+// RemoveStateChangeListener.
+func (c *Client) SetStateChangeListener(listener StateChangeListener) {
+	c.stateChangeMu.Lock()
+	defer c.stateChangeMu.Unlock()
+	c.stopStateChangeWatchLocked()
+	if listener == nil {
+		return
+	}
+
+	// Both subscriptions are buffered (one pending tick, ten pending events),
+	// so unsubscribing is not enough to stop callbacks: the loops would drain
+	// what is already queued and deliver it to a listener the caller has
+	// already removed or replaced. Gate every callback on this registration's
+	// own signal, which is closed before unsubscribing.
+	done := make(chan struct{})
+	c.stateChangeDone = done
+
+	id, ch := c.recorder.SubscribeToStateChanges()
+	c.stateChangeSubID = id
+	// The channel is closed by UnsubscribeFromStateChanges, which ends the
+	// goroutine. Ticks are coalesced (buffer of one), so a burst of changes
+	// wakes the listener once.
+	go func() {
+		for range ch {
+			select {
+			case <-done:
+				return
+			default:
+			}
+			listener.OnStateChanged()
+		}
+	}()
+
+	c.eventSub = c.recorder.SubscribeToEvents()
+	go watchSessionWarnings(c.eventSub, listener, done)
+}
+
+// RemoveStateChangeListener unregisters the state notification listener.
+func (c *Client) RemoveStateChangeListener() {
+	c.stateChangeMu.Lock()
+	defer c.stateChangeMu.Unlock()
+	c.stopStateChangeWatchLocked()
+}
+
+// DismissSessionWarning records the user's "Dismiss" on the first expiry
+// warning and suppresses the final one for the current deadline. A refreshed
+// deadline re-arms both. No-op while the engine is not running.
+func (c *Client) DismissSessionWarning() {
+	cc := c.getConnectClient()
+	if cc == nil {
+		return
+	}
+	engine := cc.Engine()
+	if engine == nil {
+		return
+	}
+	engine.DismissSessionWarning()
+}
+
+// ExtendAuthSession runs the interactive SSO flow to obtain a fresh JWT and
+// asks the management server to extend the session deadline. The tunnel is
+// untouched: no resync, no reconnect. Async; the result arrives on the
+// listener. Mirror of the daemon's RequestExtendAuthSession /
+// WaitExtendAuthSession RPC pair, with URLOpener playing the "UI opens the
+// browser" role.
+//
+// Only one flow may be in flight: the PKCE step binds a fixed loopback port,
+// so a second concurrent flow would fail on that bind. Call
+// CancelExtendAuthSession when the user abandons the browser.
+func (c *Client) ExtendAuthSession(urlOpener URLOpener, isAndroidTV bool, resultListener ErrListener) {
+	ctx, err := c.beginExtend()
+	if err != nil {
+		resultListener.OnError(err)
+		return
+	}
+
+	go func() {
+		defer c.endExtend()
+		if err := c.extendAuthSession(ctx, urlOpener, isAndroidTV); err != nil {
+			resultListener.OnError(err)
+			return
+		}
+		resultListener.OnSuccess()
+	}()
+}
+
+// CancelExtendAuthSession aborts an in-flight ExtendAuthSession. The tunnel is
+// left alone — unlike the login flow, which cancels the whole client context
+// by stopping the engine. Without this the abandoned PKCE wait keeps its
+// loopback port for the full flow timeout and blocks every later attempt.
+// No-op when no flow is running.
+func (c *Client) CancelExtendAuthSession() {
+	c.extendMu.Lock()
+	defer c.extendMu.Unlock()
+	if c.extendCancel != nil {
+		c.extendCancel()
+	}
+}
+
+func (c *Client) stopStateChangeWatchLocked() {
+	// Signal first, unsubscribe second: closing the channels only stops new
+	// items, and the loops would still hand whatever is buffered to a listener
+	// that is no longer registered.
+	if c.stateChangeDone != nil {
+		close(c.stateChangeDone)
+		c.stateChangeDone = nil
+	}
+	if c.stateChangeSubID != "" {
+		c.recorder.UnsubscribeFromStateChanges(c.stateChangeSubID)
+		c.stateChangeSubID = ""
+	}
+	if c.eventSub != nil {
+		// Closes the channel, which ends watchSessionWarnings.
+		c.recorder.UnsubscribeFromEvents(c.eventSub)
+		c.eventSub = nil
+	}
+}
+
+// watchSessionWarnings forwards the engine's session-expiry warnings to the
+// listener. The event stream also carries unrelated traffic — network-map
+// updates on every sync, DNS and route errors — so everything but an
+// AUTHENTICATION event carrying the session-warning marker is dropped. Exits
+// when the subscription is closed by UnsubscribeFromEvents, or earlier when
+// done is closed — the stream buffers up to ten events, and a deregistered
+// listener must not receive the ones already queued.
+func watchSessionWarnings(sub *peer.EventSubscription, listener StateChangeListener, done <-chan struct{}) {
+	for ev := range sub.Events() {
+		select {
+		case <-done:
+			return
+		default:
+		}
+		if ev.GetCategory() != cProto.SystemEvent_AUTHENTICATION {
+			continue
+		}
+		meta := ev.GetMetadata()
+		if meta[sessionwatch.MetaSessionWarning] != "true" {
+			// Other AUTHENTICATION events exist (e.g. a deadline rejected as
+			// out of range); they carry no warning marker.
+			continue
+		}
+		deadline, err := sessionwatch.ParseExpiresAt(meta[sessionwatch.MetaSessionExpiresAt])
+		if err != nil {
+			log.Warnf("session warning event with unparsable deadline: %v", err)
+			continue
+		}
+		lead, err := sessionwatch.ParseLeadMinutes(meta[sessionwatch.MetaSessionLeadMinutes])
+		if err != nil {
+			// Informational only — the deadline above is what drives the UI.
+			lead = 0
+		}
+		listener.OnSessionExpiring(deadline.Unix(), int64(lead),
+			meta[sessionwatch.MetaSessionFinal] == "true")
+	}
+}
+
+func (c *Client) beginExtend() (context.Context, error) {
+	c.extendMu.Lock()
+	defer c.extendMu.Unlock()
+	if c.extendCancel != nil {
+		return nil, fmt.Errorf("session extend already in progress")
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	c.extendCancel = cancel
+	return ctx, nil
+}
+
+func (c *Client) endExtend() {
+	c.extendMu.Lock()
+	defer c.extendMu.Unlock()
+	if c.extendCancel != nil {
+		c.extendCancel()
+		c.extendCancel = nil
+	}
+}
+
+func (c *Client) extendAuthSession(ctx context.Context, urlOpener URLOpener, isAndroidTV bool) error {
+	cfg, _, cc := c.stateSnapshot()
+	if cfg == nil || cc == nil {
+		return fmt.Errorf("engine is not running")
+	}
+	engine := cc.Engine()
+	if engine == nil {
+		return fmt.Errorf("engine is not initialized")
+	}
+
+	authClient, err := auth.NewAuth(ctx, cfg.PrivateKey, cfg.ManagementURL, cfg)
+	if err != nil {
+		return fmt.Errorf("failed to create auth client: %v", err)
+	}
+	defer authClient.Close()
+
+	a := &Auth{ctx: ctx, config: cfg}
+	tokenInfo, err := a.foregroundGetTokenInfo(authClient, urlOpener, isAndroidTV)
+	if err != nil {
+		return fmt.Errorf("interactive sso login failed: %v", err)
+	}
+
+	if _, err := engine.ExtendAuthSession(ctx, tokenInfo.GetTokenToUse()); err != nil {
+		return err
+	}
+	c.clearLoginRequired()
+
+	go urlOpener.OnLoginSuccess()
+	return nil
+}

client/internal/dns/handler_chain.go

@@ -3,7 +3,6 @@ package dns
 import (
 	"context"
 	"fmt"
-	"maps"
 	"math"
 	"net"
 	"slices"
@@ -32,26 +31,6 @@ type SubdomainMatcher interface {
 	MatchSubdomains() bool
 }
 
-// responseMeta holds the annotations handlers attach to a request to explain the
-// response the chain ends up writing. It survives a deferral, so an answer that
-// did not come from the handler that owns the name still says who stepped aside
-// and why.
-type responseMeta map[resutil.MetaKey]string
-
-// format renders the annotations for the response log line. The order is stable
-// so the same event reads the same way every time; a map's own order is not.
-func (m responseMeta) format() string {
-	if len(m) == 0 {
-		return ""
-	}
-
-	var b strings.Builder
-	for _, k := range slices.Sorted(maps.Keys(m)) {
-		b.WriteString(" " + string(k) + "=" + m[k])
-	}
-	return b.String()
-}
-
 type HandlerEntry struct {
 	Handler         dns.Handler
 	Priority        int
@@ -73,19 +52,8 @@ type ResponseWriterChain struct {
 	origPattern    string
 	requestID      string
 	shouldContinue bool
-	// softNegative suppresses a poisoning negative verdict for this request. A
-	// handler that owns the name but cannot answer this query type sets it
-	// before deferring, and it stays set for every handler that runs after.
-	softNegative bool
-	// clientHasEdns records whether the original query advertised EDNS0, taken
-	// before any handler ran: handlers add EDNS0 to the query they forward
-	// upstream, so the message itself no longer answers the question later.
-	clientHasEdns bool
-	response      *dns.Msg
-	// meta is handed to the next handler when this one defers, so the same map
-	// outlives the writer that created it. Handlers must only set metadata from
-	// within ServeDNS, never from a goroutine that outlives the call.
-	meta responseMeta
+	response       *dns.Msg
+	meta           map[string]string
 }
 
 // RequestID returns the request ID for tracing
@@ -94,64 +62,26 @@ func (w *ResponseWriterChain) RequestID() string {
 }
 
 // SetMeta sets a metadata key-value pair for logging
-func (w *ResponseWriterChain) SetMeta(key resutil.MetaKey, value string) {
+func (w *ResponseWriterChain) SetMeta(key, value string) {
 	if w.meta == nil {
-		w.meta = make(responseMeta)
+		w.meta = make(map[string]string)
 	}
 	w.meta[key] = value
 }
 
-// RequestSoftNegative marks the request so a downstream NXDOMAIN is turned into
-// NODATA before it reaches the client. Set by a handler that defers a query for
-// a name it owns but a type it cannot resolve.
-func (w *ResponseWriterChain) RequestSoftNegative() {
-	w.softNegative = true
-}
-
 func (w *ResponseWriterChain) WriteMsg(m *dns.Msg) error {
 	// Check if this is a continue signal (NXDOMAIN with Zero bit set)
 	if m.Rcode == dns.RcodeNameError && m.MsgHdr.Zero {
 		w.shouldContinue = true
 		return nil
 	}
-	if w.softNegative && m.Rcode == dns.RcodeNameError {
-		m = softenNegative(m, w.clientHasEdns)
-		w.SetMeta(resutil.MetaKeySoftened, "nxdomain->nodata")
-	}
 	w.response = m
 	if m.MsgHdr.Truncated {
-		w.SetMeta(resutil.MetaKeyTruncated, "true")
+		w.SetMeta("truncated", "true")
 	}
 	return w.ResponseWriter.WriteMsg(m)
 }
 
-// softenNegative downgrades an NXDOMAIN to NODATA for a request a handler that
-// owns the name deferred. NXDOMAIN is cached for the name and every type below
-// it, so a resolver that has never heard of a routed name would take out the
-// record types the route does serve; NODATA is cached for this name and type
-// only. The authority section goes with it: the negative TTL of a zone we just
-// overruled does not apply, and RFC 2308 keeps a negative answer that carries no
-// SOA out of caches altogether, so the rewrite cannot outlive the route.
-//
-// The rewrite is ours, not the answering resolver's, so an EDNS0 client is told
-// as much: the reply we hand back travels a path no capture on this host sees,
-// and an empty answer is otherwise indistinguishable from a real one. Returns a
-// copy so the answering handler keeps whatever it may hold on to.
-func softenNegative(m *dns.Msg, clientHasEdns bool) *dns.Msg {
-	out := m.Copy()
-	out.Rcode = dns.RcodeSuccess
-	out.Ns = nil
-
-	if clientHasEdns {
-		resutil.AttachEDE(out, resutil.EDENetbirdSoftenedNegative,
-			"netbird: name is served locally, NXDOMAIN from the fallthrough resolver suppressed")
-	} else {
-		resutil.StripOPT(out)
-	}
-
-	return out
-}
-
 func NewHandlerChain() *HandlerChain {
 	return &HandlerChain{
 		handlers: make([]HandlerEntry, 0),
@@ -293,19 +223,6 @@ func (c *HandlerChain) dispatch(w dns.ResponseWriter, r *dns.Msg, maxPriority in
 	handlers := slices.Clone(c.handlers)
 	c.mu.RUnlock()
 
-	// Carried across deferrals: once a handler that owns the name defers, no
-	// handler after it may answer with a verdict that poisons the name. The
-	// metadata of a handler that stepped aside is carried too, so the response
-	// log line explains an answer that did not come from the handler that owns
-	// the name.
-	var softNegative bool
-	var carried responseMeta
-
-	// Taken before any handler runs: handlers advertise EDNS0 on the query they
-	// forward upstream, so afterwards the message no longer tells us whether the
-	// client did.
-	clientHasEdns := r.IsEdns0() != nil
-
 	// Try handlers in priority order
 	for _, entry := range handlers {
 		if entry.Priority > maxPriority {
@@ -328,16 +245,11 @@ func (c *HandlerChain) dispatch(w dns.ResponseWriter, r *dns.Msg, maxPriority in
 			ResponseWriter: w,
 			origPattern:    entry.OrigPattern,
 			requestID:      requestID,
-			softNegative:   softNegative,
-			clientHasEdns:  clientHasEdns,
-			meta:           carried,
 		}
 		entry.Handler.ServeDNS(chainWriter, r)
 
 		// If handler wants to continue, try next handler
 		if chainWriter.shouldContinue {
-			softNegative = softNegative || chainWriter.softNegative
-			carried = chainWriter.meta
 			if entry.Priority != PriorityMgmtCache {
 				logger.Tracef("handler requested continue for domain=%s", qname)
 			}
@@ -353,59 +265,30 @@ func (c *HandlerChain) dispatch(w dns.ResponseWriter, r *dns.Msg, maxPriority in
 		qname, dns.TypeToString[question.Qtype], dns.ClassToString[question.Qclass])
 	resp := &dns.Msg{}
 	resp.SetRcode(r, dns.RcodeRefused)
-	// A handler that owns the name deferred and nothing below it could answer
-	// (a client with no primary nameserver group). The name exists as far as
-	// this client is concerned, since the route serves its addresses, so REFUSED
-	// would contradict the route: a stub that takes it as "not served here" and
-	// retries another resolver can come back with an NXDOMAIN that takes the
-	// whole name down. Answer "no records of this type" instead, without an SOA,
-	// so RFC 2308 keeps it out of negative caches, and tell an EDNS0 client the
-	// empty answer is ours rather than a resolver's.
-	if softNegative {
-		resp.Rcode = dns.RcodeSuccess
-		if clientHasEdns {
-			resutil.AttachEDE(resp, resutil.EDENetbirdSoftenedNegative,
-				"netbird: name is served locally, no fallthrough resolver for this query type")
-		}
-		// logResponse never runs on this path, so the carried metadata is
-		// appended here or the reason for the deferral is lost in exactly the
-		// case that is hardest to diagnose.
-		logger.Tracef("no handler below the deferring one for domain=%s type=%s, answering NODATA%s",
-			qname, dns.TypeToString[question.Qtype], carried.format())
-	}
 	if err := w.WriteMsg(resp); err != nil {
 		logger.Errorf("failed to write DNS response: %v", err)
 	}
 }
 
 func (c *HandlerChain) logResponse(logger *log.Entry, cw *ResponseWriterChain, qname string, startTime time.Time) {
-	// Runs for every query, and the arguments below are not free: Len() packs
-	// the message to measure it, and formatting the answers and the metadata
-	// allocates. None of it is worth doing when the line is discarded.
-	if !log.IsLevelEnabled(log.TraceLevel) {
-		return
-	}
-
 	if cw.response == nil {
 		return
 	}
 
+	var meta string
+	for k, v := range cw.meta {
+		meta += " " + k + "=" + v
+	}
+
 	logger.Tracef("response: domain=%s rcode=%s answers=%s size=%dB%s took=%s",
 		qname, dns.RcodeToString[cw.response.Rcode], resutil.FormatAnswers(cw.response.Answer),
-		cw.response.Len(), cw.meta.format(), time.Since(startTime))
+		cw.response.Len(), meta, time.Since(startTime))
 }
 
 // ResolveInternal runs an in-process DNS query against the chain, skipping any
 // handler with priority > maxPriority. Used by internal callers (e.g. the mgmt
-// cache refresher) that must bypass themselves to avoid loops.
-//
-// "Nothing answered" is read off RcodeRefused, which a request soft-negatived by
-// a deferring handler never carries: it ends in an empty NOERROR instead, and
-// would look resolved. No caller can reach that today, since every handler that
-// defers sits above the maxPriority any caller passes. Lowering one below it
-// means this check needs the soft-negative case too.
-//
-// Honors ctx cancellation; on ctx.Done the dispatch goroutine is left to drain on its own
+// cache refresher) that must bypass themselves to avoid loops. Honors ctx
+// cancellation; on ctx.Done the dispatch goroutine is left to drain on its own
 // (bounded by the invoked handler's internal timeout).
 func (c *HandlerChain) ResolveInternal(ctx context.Context, r *dns.Msg, maxPriority int) (*dns.Msg, error) {
 	if len(r.Question) == 0 {

client/internal/dns/handler_chain_test.go

@@ -3,19 +3,15 @@ package dns_test
 import (
 	"context"
 	"net"
-	"strings"
 	"testing"
 	"time"
 
 	"github.com/miekg/dns"
-	log "github.com/sirupsen/logrus"
-	logtest "github.com/sirupsen/logrus/hooks/test"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/mock"
 	"github.com/stretchr/testify/require"
 
 	nbdns "github.com/netbirdio/netbird/client/internal/dns"
-	"github.com/netbirdio/netbird/client/internal/dns/resutil"
 	"github.com/netbirdio/netbird/client/internal/dns/test"
 )
 
@@ -1242,276 +1238,6 @@ func TestHandlerChain_ResolveInternal_HonorsContextTimeout(t *testing.T) {
 	assert.Less(t, elapsed, 500*time.Millisecond, "ResolveInternal must return shortly after ctx deadline")
 }
 
-// requestSoftNegative asks the chain to soften a negative verdict produced by
-// the handlers that run after the caller defers, reporting whether the chain
-// supports the signal. Written as a type assertion so these tests compile
-// against a chain that does not support it yet.
-func requestSoftNegative(w dns.ResponseWriter) bool {
-	sn, ok := w.(interface{ RequestSoftNegative() })
-	if ok {
-		sn.RequestSoftNegative()
-	}
-	return ok
-}
-
-// deferringHandler defers to the next handler in the chain, optionally asking
-// for the negative verdict of whatever answers instead to be softened. This is
-// what a DNS route handler does for a record type its routing peer cannot
-// resolve.
-type deferringHandler struct {
-	softNegative bool
-	called       bool
-	supported    bool
-}
-
-func (h *deferringHandler) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
-	h.called = true
-	if h.softNegative {
-		h.supported = requestSoftNegative(w)
-		// The real handler records why it stepped aside, for the response log
-		// line of whichever handler answers instead.
-		resutil.SetMeta(w, resutil.MetaKeyDeferredBy, "test handler")
-	}
-	resp := new(dns.Msg)
-	resp.SetRcode(r, dns.RcodeNameError)
-	resp.MsgHdr.Zero = true
-	_ = w.WriteMsg(resp)
-}
-
-// nxdomainHandler answers an authoritative NXDOMAIN with an SOA in the
-// authority section, the way a public resolver answers for a name that only
-// exists inside the routed network.
-type nxdomainHandler struct{}
-
-func (h *nxdomainHandler) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
-	resp := new(dns.Msg)
-	resp.SetRcode(r, dns.RcodeNameError)
-	resp.Ns = []dns.RR{&dns.SOA{
-		Hdr:     dns.RR_Header{Name: "example.com.", Rrtype: dns.TypeSOA, Class: dns.ClassINET, Ttl: 3600},
-		Ns:      "ns1.example.com.",
-		Mbox:    "hostmaster.example.com.",
-		Minttl:  3600,
-		Expire:  604800,
-		Refresh: 7200,
-		Retry:   3600,
-	}}
-	_ = w.WriteMsg(resp)
-}
-
-// TestHandlerChain_SoftNegative_DowngradesDownstreamNXDOMAIN is the whole point
-// of the soft-negative signal: a route handler may only defer a query to the
-// public chain if the answer cannot poison the routed name. NXDOMAIN is cached
-// for the name and every type under it (RFC 2308, RFC 8020), so it has to be
-// rewritten to NODATA, which is cached per name and type only.
-func TestHandlerChain_SoftNegative_DowngradesDownstreamNXDOMAIN(t *testing.T) {
-	chain := nbdns.NewHandlerChain()
-
-	route := &deferringHandler{softNegative: true}
-	chain.AddHandler("*.example.com.", route, nbdns.PriorityDNSRoute)
-	chain.AddHandler(".", &nxdomainHandler{}, nbdns.PriorityDefault)
-
-	r := new(dns.Msg)
-	r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-
-	mw := &test.MockResponseWriter{}
-	chain.ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-	resp := mw.GetLastResponse()
-	require.NotNil(t, resp, "a response must reach the client")
-	assert.True(t, route.called, "the route handler must run first")
-	require.True(t, route.supported, "the chain writer must accept a soft-negative request")
-	assert.Equal(t, dns.RcodeSuccess, resp.Rcode, "NXDOMAIN must be softened to NODATA")
-	assert.Empty(t, resp.Answer, "a softened negative carries no answer")
-	assert.Empty(t, resp.Ns, "the downstream zone's SOA must not set the negative TTL for a name we overrode")
-}
-
-// responseLineFor returns the chain's response log line for one query. Raising
-// the level to trace also unmutes whatever else is logging in this package, so
-// the line has to be picked by the name it was asked about rather than by being
-// the last one seen.
-func responseLineFor(hook *logtest.Hook, qname string) string {
-	for _, e := range hook.AllEntries() {
-		if strings.HasPrefix(e.Message, "response:") && strings.Contains(e.Message, qname) {
-			return e.Message
-		}
-	}
-	return ""
-}
-
-// TestHandlerChain_SoftNegative_IsVisibleToTheClient covers observability of the
-// rewrite. The reply we hand the application travels over loopback, which the
-// bundle capture does not see, so a softened verdict has to say so on the wire:
-// without it an empty answer is indistinguishable from a real "no such record"
-// in a dig output or a capture taken next to the application.
-func TestHandlerChain_SoftNegative_IsVisibleToTheClient(t *testing.T) {
-	// One hook for the whole test: logtest installs it on the standard logger
-	// and logrus has no way to take it off again, so a hook per subtest would
-	// leave several behind buffering every later line in the package.
-	hook := logtest.NewGlobal()
-	t.Cleanup(hook.Reset)
-
-	newChain := func() *nbdns.HandlerChain {
-		chain := nbdns.NewHandlerChain()
-		chain.AddHandler("*.example.com.", &deferringHandler{softNegative: true}, nbdns.PriorityDNSRoute)
-		chain.AddHandler(".", &nxdomainHandler{}, nbdns.PriorityDefault)
-		return chain
-	}
-
-	t.Run("EDNS0 client gets an extended error", func(t *testing.T) {
-		r := new(dns.Msg)
-		r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-		r.SetEdns0(dns.DefaultMsgSize, false)
-
-		mw := &test.MockResponseWriter{}
-		newChain().ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-		resp := mw.GetLastResponse()
-		require.NotNil(t, resp)
-		require.Equal(t, dns.RcodeSuccess, resp.Rcode)
-
-		ede, ok := resutil.ExtractEDE(resp)
-		require.True(t, ok, "a softened verdict must carry an extended DNS error")
-		assert.Equal(t, resutil.EDENetbirdSoftenedNegative, ede.InfoCode)
-		assert.Contains(t, ede.ExtraText, "netbird", "the text must name us as the source of the rewrite")
-	})
-
-	t.Run("plain client gets no OPT", func(t *testing.T) {
-		r := new(dns.Msg)
-		r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-
-		mw := &test.MockResponseWriter{}
-		newChain().ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-		resp := mw.GetLastResponse()
-		require.NotNil(t, resp)
-		assert.Nil(t, resp.IsEdns0(), "RFC 6891 forbids an OPT toward a client that did not advertise EDNS0")
-	})
-
-	// The response line is what support reads out of a debug bundle, and it now
-	// carries several annotations at once. Built from a map, their order would
-	// differ on every query, so the same event never looks the same twice.
-	t.Run("log fields keep a stable order", func(t *testing.T) {
-		const qname = "_mongodb._tcp.stable.example.com."
-
-		prev := log.GetLevel()
-		log.SetLevel(log.TraceLevel)
-		t.Cleanup(func() { log.SetLevel(prev) })
-
-		lineFor := func() string {
-			hook.Reset()
-
-			r := new(dns.Msg)
-			r.SetQuestion(qname, dns.TypeSRV)
-
-			mw := &test.MockResponseWriter{}
-			newChain().ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-			line := responseLineFor(hook, qname)
-			// The duration differs per query and is not what we compare.
-			line, _, _ = strings.Cut(line, " took=")
-			return line
-		}
-
-		first := lineFor()
-		require.NotEmpty(t, first, "the chain must log the response it wrote")
-		require.Contains(t, first, "deferred_by=", "the line must carry more than one annotation to be worth ordering")
-		require.Contains(t, first, "softened=")
-
-		for range 20 {
-			assert.Equal(t, first, lineFor(), "the same event must produce the same line")
-		}
-	})
-
-	t.Run("logged with the reason it was deferred", func(t *testing.T) {
-		const qname = "_mongodb._tcp.reason.example.com."
-
-		hook.Reset()
-
-		prev := log.GetLevel()
-		log.SetLevel(log.TraceLevel)
-		t.Cleanup(func() { log.SetLevel(prev) })
-
-		r := new(dns.Msg)
-		r.SetQuestion(qname, dns.TypeSRV)
-
-		mw := &test.MockResponseWriter{}
-		newChain().ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-		response := responseLineFor(hook, qname)
-		require.NotEmpty(t, response, "the chain must log the response it wrote")
-		assert.Contains(t, response, "softened=", "the log must show the verdict was rewritten")
-		assert.Contains(t, response, "deferred_by=", "the log must name the handler that deferred")
-	})
-}
-
-// TestHandlerChain_SoftNegative_NoHandlerBelow covers a client with no primary
-// nameserver group: the deferred query reaches the end of the chain unanswered.
-// REFUSED would say the name is not served here while the route serves its
-// addresses, and a stub that acts on that by asking elsewhere can bring back an
-// NXDOMAIN for the whole name. The answer must be an empty, uncacheable NODATA.
-func TestHandlerChain_SoftNegative_NoHandlerBelow(t *testing.T) {
-	chain := nbdns.NewHandlerChain()
-
-	route := &deferringHandler{softNegative: true}
-	chain.AddHandler("*.example.com.", route, nbdns.PriorityDNSRoute)
-
-	r := new(dns.Msg)
-	r.SetQuestion("db.example.com.", dns.TypeHTTPS)
-
-	mw := &test.MockResponseWriter{}
-	chain.ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-	resp := mw.GetLastResponse()
-	require.NotNil(t, resp, "a response must reach the client")
-	assert.Equal(t, dns.RcodeSuccess, resp.Rcode, "an unanswered soft-negative query must be NODATA, not REFUSED")
-	assert.Empty(t, resp.Answer)
-	assert.Empty(t, resp.Ns,
-		"no SOA, so RFC 2308 keeps the empty answer out of negative caches and it cannot outlive the route")
-}
-
-// TestHandlerChain_SoftNegative_KeepsRealAnswers guards the other direction:
-// softening applies to negative verdicts only. A real answer from a downstream
-// handler must reach the client untouched.
-func TestHandlerChain_SoftNegative_KeepsRealAnswers(t *testing.T) {
-	chain := nbdns.NewHandlerChain()
-
-	route := &deferringHandler{softNegative: true}
-	chain.AddHandler("*.example.com.", route, nbdns.PriorityDNSRoute)
-	chain.AddHandler(".", &answeringHandler{name: "public", ip: "203.0.113.10"}, nbdns.PriorityDefault)
-
-	r := new(dns.Msg)
-	r.SetQuestion("db.example.com.", dns.TypeA)
-
-	mw := &test.MockResponseWriter{}
-	chain.ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-	resp := mw.GetLastResponse()
-	require.NotNil(t, resp, "a response must reach the client")
-	assert.Equal(t, dns.RcodeSuccess, resp.Rcode)
-	require.Len(t, resp.Answer, 1, "the downstream answer must pass through")
-}
-
-// TestHandlerChain_NXDOMAINPreservedWithoutSoftNegative makes sure the
-// softening is opt-in: an ordinary chain continuation still yields NXDOMAIN, so
-// genuine non-existence keeps being reported.
-func TestHandlerChain_NXDOMAINPreservedWithoutSoftNegative(t *testing.T) {
-	chain := nbdns.NewHandlerChain()
-
-	route := &deferringHandler{}
-	chain.AddHandler("*.example.com.", route, nbdns.PriorityDNSRoute)
-	chain.AddHandler(".", &nxdomainHandler{}, nbdns.PriorityDefault)
-
-	r := new(dns.Msg)
-	r.SetQuestion("nope.example.com.", dns.TypeA)
-
-	mw := &test.MockResponseWriter{}
-	chain.ServeDNS(&nbdns.ResponseWriterChain{ResponseWriter: mw}, r)
-
-	resp := mw.GetLastResponse()
-	require.NotNil(t, resp, "a response must reach the client")
-	assert.Equal(t, dns.RcodeNameError, resp.Rcode, "without the signal a real NXDOMAIN must survive")
-}
-
 func TestHandlerChain_HasRootHandlerAtOrBelow(t *testing.T) {
 	chain := nbdns.NewHandlerChain()
 	h := &answeringHandler{name: "h", ip: "10.0.0.1"}

client/internal/dns/resutil/resolve.go

@@ -19,34 +19,6 @@ import (
 // uses when a resolved host has no addresses of the requested family.
 const errNoSuitableAddress = "no suitable address found"
 
-// Extended DNS Error info codes NetBird emits so a client can see why an answer
-// looks the way it does without reading this peer's logs. They live in the RFC
-// 8914 Private Use range (49152-65535) and are registered here, in one place,
-// because nothing else guarantees two NetBird components pick distinct codes.
-const (
-	// EDENetbirdUpstreamTimeout: a DNS forwarder's upstream did not answer.
-	EDENetbirdUpstreamTimeout uint16 = 49152
-	// EDENetbirdUpstreamFailure: a DNS forwarder's upstream failed.
-	EDENetbirdUpstreamFailure uint16 = 49153
-	// EDENetbirdSoftenedNegative: the empty answer is ours, not the answering
-	// resolver's. A handler that owns the name deferred this query type, and the
-	// negative verdict that came back was downgraded so it cannot poison the
-	// name.
-	EDENetbirdSoftenedNegative uint16 = 49154
-)
-
-// AttachEDE adds an Extended DNS Error (RFC 8914) option to a message, creating
-// the OPT pseudo-record if it has none. Callers must only use it toward a client
-// that advertised EDNS0: per RFC 6891 an OPT must not appear otherwise.
-func AttachEDE(msg *dns.Msg, code uint16, text string) {
-	opt := msg.IsEdns0()
-	if opt == nil {
-		msg.SetEdns0(dns.DefaultMsgSize, false)
-		opt = msg.IsEdns0()
-	}
-	opt.Option = append(opt.Option, &dns.EDNS0_EDE{InfoCode: code, ExtraText: text})
-}
-
 // GenerateRequestID creates a random 8-character hex string for request tracing.
 func GenerateRequestID() string {
 	bytes := make([]byte, 4)
@@ -106,38 +78,10 @@ type resolver interface {
 	LookupNetIP(ctx context.Context, network, host string) ([]netip.Addr, error)
 }
 
-// MetaKey names an annotation a handler attaches to a request to explain the
-// response the chain ends up writing. The set is closed and rendered onto a
-// single log line, so an unrecognized key is a typo inventing a field rather
-// than a new fact.
-type MetaKey string
-
-const (
-	// MetaKeyProtocol: the transport a query arrived on.
-	MetaKeyProtocol MetaKey = "protocol"
-	// MetaKeyUpstream: the upstream that answered.
-	MetaKeyUpstream MetaKey = "upstream"
-	// MetaKeyUpstreamProtocol: the transport used toward that upstream.
-	MetaKeyUpstreamProtocol MetaKey = "upstream_protocol"
-	// MetaKeyPeer: the routing peer whose forwarder answered.
-	MetaKeyPeer MetaKey = "peer"
-	// MetaKeyEDE: an Extended DNS Error carried by the answer.
-	MetaKeyEDE MetaKey = "ede"
-	// MetaKeyTruncated: the answer did not fit and was truncated.
-	MetaKeyTruncated MetaKey = "truncated"
-	// MetaKeySoftened: a negative verdict was rewritten so it cannot poison a
-	// name served locally.
-	MetaKeySoftened MetaKey = "softened"
-	// MetaKeyDeferredBy: the handler that owned the name and stepped aside.
-	MetaKeyDeferredBy MetaKey = "deferred_by"
-	// MetaKeyDeferredReason: why it stepped aside.
-	MetaKeyDeferredReason MetaKey = "deferred_reason"
-)
-
 // chainedWriter is implemented by ResponseWriters that carry request metadata
 type chainedWriter interface {
 	RequestID() string
-	SetMeta(key MetaKey, value string)
+	SetMeta(key, value string)
 }
 
 // GetRequestID extracts a request ID from the ResponseWriter if available,
@@ -152,30 +96,12 @@ func GetRequestID(w dns.ResponseWriter) string {
 }
 
 // SetMeta sets metadata on the ResponseWriter if it supports it.
-func SetMeta(w dns.ResponseWriter, key MetaKey, value string) {
+func SetMeta(w dns.ResponseWriter, key, value string) {
 	if cw, ok := w.(chainedWriter); ok {
 		cw.SetMeta(key, value)
 	}
 }
 
-// softNegativeRequester is implemented by chain writers that can soften the
-// negative verdict of the handlers a deferring handler falls through to.
-type softNegativeRequester interface {
-	RequestSoftNegative()
-}
-
-// RequestSoftNegative asks the handler chain to downgrade an NXDOMAIN from the
-// handlers that run after the caller defers. A handler that owns a name but
-// cannot answer one query type for it needs this: the resolvers it falls
-// through to cannot prove the name absent, and an NXDOMAIN from them is cached
-// for the name and every type under it (RFC 2308, RFC 8020), taking the
-// addresses the handler does serve down with it.
-func RequestSoftNegative(w dns.ResponseWriter) {
-	if sn, ok := w.(softNegativeRequester); ok {
-		sn.RequestSoftNegative()
-	}
-}
-
 // LookupResult contains the result of an external DNS lookup
 type LookupResult struct {
 	IPs   []netip.Addr
@@ -273,27 +199,11 @@ type RecordResolver interface {
 	LookupAddr(ctx context.Context, addr string) ([]string, error)
 }
 
-// SupportedRecordQtype reports whether LookupRecords can resolve qtype. The
-// set is bounded by the net.Resolver API, which exposes no way to query an
-// arbitrary record type, and going around it with a raw exchange would mean
-// picking nameservers ourselves instead of resolving the way the host does.
-//
-// Both ends read this: a DNS forwarder answers these types for the domains it
-// routes, and a client uses it to tell which types are worth forwarding to a
-// peer at all.
-func SupportedRecordQtype(qtype uint16) bool {
-	switch qtype {
-	case dns.TypeMX, dns.TypeTXT, dns.TypeNS, dns.TypeSRV, dns.TypeCNAME, dns.TypePTR:
-		return true
-	default:
-		return false
-	}
-}
-
 // LookupRecords resolves a non-address DNS record type through the host
-// resolver and returns the resource records and the DNS rcode. Types outside
-// SupportedRecordQtype yield NODATA so that a routed name is never poisoned
-// with NXDOMAIN for a type we cannot look up.
+// resolver and returns the resource records and the DNS rcode. Types the host
+// resolver cannot answer (anything not covered by the net.Resolver Lookup*
+// methods) yield NODATA so that a routed name is never poisoned with NXDOMAIN
+// for an unsupported type.
 func LookupRecords(ctx context.Context, r RecordResolver, name string, qtype uint16, ttl uint32) ([]dns.RR, int) {
 	fqdn := dns.Fqdn(name)
 

client/internal/dns/upstream.go

@@ -295,7 +295,7 @@ func (u *upstreamResolverBase) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	if addr := w.RemoteAddr(); addr != nil {
 		network := addr.Network()
 		ctx = contextWithDNSProtocol(ctx, network)
-		resutil.SetMeta(w, resutil.MetaKeyProtocol, network)
+		resutil.SetMeta(w, "protocol", network)
 	}
 
 	ok, failures := u.tryUpstreamServers(ctx, w, r, logger)
@@ -331,7 +331,7 @@ func (u *upstreamResolverBase) tryOnlyRace(ctx context.Context, w dns.ResponseWr
 		return false, res.failures
 	}
 	if res.ede != "" {
-		resutil.SetMeta(w, resutil.MetaKeyEDE, res.ede)
+		resutil.SetMeta(w, "ede", res.ede)
 	}
 	u.writeSuccessResponse(w, res.msg, res.upstream, r.Question[0].Name, res.protocol, logger)
 	return true, res.failures
@@ -361,7 +361,7 @@ func (u *upstreamResolverBase) raceAll(ctx context.Context, w dns.ResponseWriter
 			failures = append(failures, res.failures...)
 			if res.msg != nil {
 				if res.ede != "" {
-					resutil.SetMeta(w, resutil.MetaKeyEDE, res.ede)
+					resutil.SetMeta(w, "ede", res.ede)
 				}
 				u.writeSuccessResponse(w, res.msg, res.upstream, r.Question[0].Name, res.protocol, logger)
 				return true, failures
@@ -550,9 +550,9 @@ func (u *upstreamResolverBase) debugUpstreamTimeout(upstream netip.AddrPort) str
 }
 
 func (u *upstreamResolverBase) writeSuccessResponse(w dns.ResponseWriter, rm *dns.Msg, upstream netip.AddrPort, domain string, proto string, logger *log.Entry) {
-	resutil.SetMeta(w, resutil.MetaKeyUpstream, upstream.String())
+	resutil.SetMeta(w, "upstream", upstream.String())
 	if proto != "" {
-		resutil.SetMeta(w, resutil.MetaKeyUpstreamProtocol, proto)
+		resutil.SetMeta(w, "upstream_protocol", proto)
 	}
 
 	// Clear Zero bit from external responses to prevent upstream servers from

client/internal/dnsfwd/forwarder.go

@@ -26,6 +26,15 @@ import (
 const errResolveFailed = "failed to resolve query for domain=%s: %v"
 const upstreamTimeout = 15 * time.Second
 
+// EDE info codes the forwarder emits on upstream failures so the querying
+// client can see the reason without inspecting this peer's logs. They live in
+// the RFC 8914 Private Use range (49152-65535); the Go resolver never exposes a
+// real upstream EDE here, so these cannot collide with a genuine code.
+const (
+	edeNetbirdUpstreamTimeout uint16 = 49152
+	edeNetbirdUpstreamFailure uint16 = 49153
+)
+
 type resolver interface {
 	LookupNetIP(ctx context.Context, network, host string) ([]netip.Addr, error)
 	LookupMX(ctx context.Context, name string) ([]*net.MX, error)
@@ -207,9 +216,6 @@ func (f *DNSForwarder) handleDNSQuery(logger *log.Entry, w dns.ResponseWriter, q
 		qname, dns.TypeToString[question.Qtype], dns.ClassToString[question.Qclass])
 
 	resp := query.SetReply(query)
-	// Every answer here comes from a recursive lookup on this peer. SetReply
-	// leaves RA unset, which reads to a client as a server that cannot recurse.
-	resp.RecursionAvailable = true
 
 	mostSpecificResId, matchingEntries := f.getMatchingEntries(strings.TrimSuffix(qname, "."))
 	if mostSpecificResId == "" {
@@ -223,22 +229,20 @@ func (f *DNSForwarder) handleDNSQuery(logger *log.Entry, w dns.ResponseWriter, q
 
 	reqHasEdns := query.IsEdns0() != nil
 
-	switch {
-	case question.Qtype == dns.TypeA || question.Qtype == dns.TypeAAAA:
+	switch question.Qtype {
+	case dns.TypeA, dns.TypeAAAA:
 		f.handleAddressQuery(ctx, logger, w, resp, mostSpecificResId, matchingEntries, reqHasEdns, startTime)
-	case resutil.SupportedRecordQtype(question.Qtype):
+	case dns.TypeMX, dns.TypeTXT, dns.TypeNS, dns.TypeSRV, dns.TypeCNAME, dns.TypePTR:
 		f.handleRecordQuery(ctx, logger, w, resp, startTime)
 	default:
 		// The domain is routed here, so any other type is answered NODATA
 		// (NOERROR, empty answer) rather than falling back to a resolver that
 		// would poison the name with NXDOMAIN. The Extended DNS Error lets a
 		// client tell this capability-driven NODATA apart from an
-		// authoritative one; a current client knows the type is unsupported and
-		// resolves it through its own handler chain instead of asking at all.
-		// The OPT pseudo-record must not appear unless the query advertised
-		// EDNS0.
+		// authoritative one. The OPT pseudo-record must not appear unless the
+		// query advertised EDNS0.
 		if reqHasEdns {
-			resutil.AttachEDE(resp, dns.ExtendedErrorCodeNotSupported, "netbird forwarder: unsupported query type")
+			attachEDE(resp, dns.ExtendedErrorCodeNotSupported, "netbird forwarder: unsupported query type")
 		}
 		f.writeResponse(logger, w, resp, qname, startTime)
 	}
@@ -437,7 +441,7 @@ func (f *DNSForwarder) handleDNSError(
 	}
 
 	if reqHasEdns {
-		resutil.AttachEDE(resp, edeCodeFor(dnsErr), edeText(dnsErr))
+		attachEDE(resp, edeCodeFor(dnsErr), edeText(dnsErr))
 	}
 
 	f.writeResponse(logger, w, resp, domain, startTime)
@@ -484,9 +488,9 @@ func (f *DNSForwarder) getMatchingEntries(domain string) (route.ResID, []*Forwar
 // edeCodeFor maps an upstream lookup error to the NetBird EDE info code.
 func edeCodeFor(dnsErr *net.DNSError) uint16 {
 	if dnsErr != nil && dnsErr.IsTimeout {
-		return resutil.EDENetbirdUpstreamTimeout
+		return edeNetbirdUpstreamTimeout
 	}
-	return resutil.EDENetbirdUpstreamFailure
+	return edeNetbirdUpstreamFailure
 }
 
 // edeText builds the EDE extra-text describing the class of upstream failure.
@@ -499,3 +503,14 @@ func edeText(dnsErr *net.DNSError) string {
 	}
 	return "netbird forwarder: upstream failure"
 }
+
+// attachEDE adds an Extended DNS Error (RFC 8914) option to the response,
+// creating the OPT pseudo-record if the response does not already carry one.
+func attachEDE(resp *dns.Msg, code uint16, text string) {
+	opt := resp.IsEdns0()
+	if opt == nil {
+		resp.SetEdns0(dns.DefaultMsgSize, false)
+		opt = resp.IsEdns0()
+	}
+	opt.Option = append(opt.Option, &dns.EDNS0_EDE{InfoCode: code, ExtraText: text})
+}

client/internal/dnsfwd/forwarder_test.go

@@ -649,39 +649,6 @@ func TestDNSForwarder_ResponseCodes(t *testing.T) {
 	}
 }
 
-// TestDNSForwarder_RecursionAvailable covers the RA bit. Every answer this
-// forwarder produces comes from a recursive lookup on the routing peer, but
-// dns.Msg.SetReply leaves RA unset, so clients report "recursion not available"
-// and some stub resolvers treat the server as unable to serve the query.
-func TestDNSForwarder_RecursionAvailable(t *testing.T) {
-	t.Run("record answer", func(t *testing.T) {
-		mockResolver := &MockResolver{}
-		forwarder := newRecordTestForwarder(t, mockResolver, "example.com")
-
-		mockResolver.On("LookupMX", mock.Anything, "example.com.").
-			Return([]*net.MX{{Host: "mail.example.com.", Pref: 10}}, nil).Once()
-
-		resp := runRecordQuery(t, forwarder, "example.com", dns.TypeMX)
-		assert.True(t, resp.RecursionAvailable, "the forwarder resolves recursively")
-	})
-
-	t.Run("unsupported type NODATA", func(t *testing.T) {
-		forwarder := newRecordTestForwarder(t, &MockResolver{}, "example.com")
-
-		resp := runRecordQuery(t, forwarder, "example.com", dns.TypeCAA)
-		require.Equal(t, dns.RcodeSuccess, resp.Rcode)
-		assert.True(t, resp.RecursionAvailable, "RA describes the server, not the query type")
-	})
-
-	t.Run("unauthorized domain", func(t *testing.T) {
-		forwarder := newRecordTestForwarder(t, &MockResolver{}, "example.com")
-
-		resp := runRecordQuery(t, forwarder, "other.com", dns.TypeMX)
-		require.Equal(t, dns.RcodeRefused, resp.Rcode)
-		assert.True(t, resp.RecursionAvailable, "RA describes the server, not the verdict")
-	})
-}
-
 func hasEDE(m *dns.Msg, code uint16) bool {
 	opt := m.IsEdns0()
 	if opt == nil {
@@ -892,7 +859,7 @@ func TestDNSForwarder_UpstreamFailureEDE(t *testing.T) {
 			lookupErr:   &net.DNSError{Err: "i/o timeout", Server: "10.0.0.53:53", IsTimeout: true},
 			reqEdns:     true,
 			wantEDE:     true,
-			wantCode:    resutil.EDENetbirdUpstreamTimeout,
+			wantCode:    edeNetbirdUpstreamTimeout,
 			wantTextHas: "netbird forwarder: upstream timeout",
 		},
 		{
@@ -900,7 +867,7 @@ func TestDNSForwarder_UpstreamFailureEDE(t *testing.T) {
 			lookupErr:   &net.DNSError{Err: "server misbehaving", Server: "10.0.0.53:53"},
 			reqEdns:     true,
 			wantEDE:     true,
-			wantCode:    resutil.EDENetbirdUpstreamFailure,
+			wantCode:    edeNetbirdUpstreamFailure,
 			wantTextHas: "netbird forwarder: upstream failure",
 		},
 		{

client/internal/routemanager/dnsinterceptor/handler.go

@@ -22,6 +22,7 @@ import (
 	nbdns "github.com/netbirdio/netbird/client/internal/dns"
 	"github.com/netbirdio/netbird/client/internal/dns/resutil"
 	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
 	"github.com/netbirdio/netbird/client/internal/routemanager/common"
 	"github.com/netbirdio/netbird/client/internal/routemanager/fakeip"
 	iface "github.com/netbirdio/netbird/client/internal/routemanager/iface"
@@ -39,61 +40,6 @@ type internalDNATer interface {
 	AddInternalDNATMapping(netip.Addr, netip.Addr) error
 }
 
-// peerAllowedIPs reports the tunnel address a peer is reachable on.
-type peerAllowedIPs interface {
-	AllowedIP(pubKey string) (netip.Addr, bool)
-}
-
-// disposition says where a query for a given record type has to be answered.
-type disposition int
-
-const (
-	// dispositionPeer: the routing peer owns the answer. Address records are
-	// what a DNS route exists for, and their answer programs the routes,
-	// allowed IPs and firewall sets, so they are never resolved anywhere else.
-	dispositionPeer disposition = iota
-	// dispositionPeerFirst: the peer's forwarder resolves the type, and the
-	// records may only exist inside the routed network, so it has to be asked.
-	// A peer running a client that predates that support cannot answer, which
-	// only its reply reveals.
-	dispositionPeerFirst
-	// dispositionChain: no forwarder resolves the type, so asking would cost a
-	// tunnel round trip for a reply that says nothing. Public records for the
-	// name are still better than none, so the query goes to the rest of the
-	// chain.
-	dispositionChain
-)
-
-func dispositionFor(qtype uint16) disposition {
-	switch {
-	case qtype == dns.TypeA || qtype == dns.TypeAAAA:
-		return dispositionPeer
-	case resutil.SupportedRecordQtype(qtype):
-		return dispositionPeerFirst
-	default:
-		return dispositionChain
-	}
-}
-
-// peerCannotAnswer reports whether a forwarder reply means the peer is unable to
-// resolve this query type, as opposed to having resolved it and found nothing.
-// A forwarder older than the one that learned to resolve non-address types
-// answers NOTIMP to all of them, before it even looks at the domain; FORMERR
-// covers one that cannot parse the EDNS0 we add to the query.
-//
-// REFUSED is deliberately not here. It says the peer does not hold the name,
-// which happens while the client and the peer disagree about the route set, and
-// falling through then would hand the name of an internal-only domain to a
-// public resolver. No forwarder version reports a missing record type that way.
-func peerCannotAnswer(rcode int) bool {
-	switch rcode {
-	case dns.RcodeNotImplemented, dns.RcodeFormatError:
-		return true
-	default:
-		return false
-	}
-}
-
 type DnsInterceptor struct {
 	mu                   sync.RWMutex
 	route                *route.Route
@@ -104,7 +50,7 @@ type DnsInterceptor struct {
 	currentPeerKey       string
 	interceptedDomains   domainMap
 	wgInterface          iface.WGIface
-	peerStore            peerAllowedIPs
+	peerStore            *peerstore.Store
 	firewall             firewall.Manager
 	fakeIPManager        *fakeip.Manager
 	forwarderPort        *atomic.Uint32
@@ -280,14 +226,11 @@ func (d *DnsInterceptor) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 		return
 	}
 
-	qtype := r.Question[0].Qtype
-	dispose := dispositionFor(qtype)
-
-	if dispose == dispositionChain {
-		d.deferToChain(w, r, logger, "unsupported-qtype")
-		return
-	}
-
+	// All query types for an intercepted domain are forwarded to the peer's
+	// DNS forwarder, which owns the name. Falling through to the system
+	// resolver would let it answer NXDOMAIN for a name it isn't authoritative
+	// for, poisoning the whole name (including the A/AAAA records the route
+	// does serve). The forwarder answers NODATA for types it cannot resolve.
 	d.mu.RLock()
 	peerKey := d.currentPeerKey
 	d.mu.RUnlock()
@@ -303,33 +246,35 @@ func (d *DnsInterceptor) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 		return
 	}
 
-	query, hadEdns := peerQuery(r)
+	if r.Extra == nil {
+		r.MsgHdr.AuthenticatedData = true
+	}
+
+	// Advertise EDNS0 to the forwarder so it may return an Extended DNS Error
+	// describing why a lookup failed. The OPT is stripped from the reply when
+	// the original client did not request EDNS0.
+	hadEdns := r.IsEdns0() != nil
+	if !hadEdns {
+		r.SetEdns0(dns.DefaultMsgSize, false)
+	}
 
 	upstream := net.JoinHostPort(upstreamIP.String(), strconv.FormatUint(uint64(d.forwarderPort.Load()), 10))
 	ctx, cancel := context.WithTimeout(context.Background(), dnsTimeout)
 	defer cancel()
 
-	reply := d.queryUpstreamDNS(ctx, w, query, upstream, upstreamIP, peerKey, logger)
+	reply := d.queryUpstreamDNS(ctx, w, r, upstream, upstreamIP, peerKey, logger)
 	if reply == nil {
-		// queryUpstreamDNS already logged the failure and answered the client
-		return
-	}
-
-	// The peer owns the name but its forwarder cannot resolve this type. No
-	// capability is announced anywhere, so the round trip above is the probe.
-	if dispose == dispositionPeerFirst && peerCannotAnswer(reply.Rcode) {
-		d.deferToChain(w, r, logger, "peer-rcode-"+dns.RcodeToString[reply.Rcode])
 		return
 	}
 
 	if ede, ok := resutil.ExtractEDE(reply); ok {
-		resutil.SetMeta(w, resutil.MetaKeyEDE, fmt.Sprintf("%d %s", ede.InfoCode, ede.ExtraText))
+		resutil.SetMeta(w, "ede", fmt.Sprintf("%d %s", ede.InfoCode, ede.ExtraText))
 	}
 	if !hadEdns {
 		resutil.StripOPT(reply)
 	}
 
-	resutil.SetMeta(w, resutil.MetaKeyPeer, peerKey)
+	resutil.SetMeta(w, "peer", peerKey)
 
 	reply.Id = r.Id
 	if err := d.writeMsg(w, reply, logger); err != nil {
@@ -337,30 +282,6 @@ func (d *DnsInterceptor) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	}
 }
 
-// deferToChain hands the query to the next handler in the chain and asks the
-// chain to soften the negative verdict of whatever answers instead. The
-// resolvers below cannot prove a name inside the routed network absent, so their
-// NXDOMAIN must not reach the client: it would be cached for the name and every
-// type under it, taking the addresses this route does serve with it.
-func (d *DnsInterceptor) deferToChain(w dns.ResponseWriter, r *dns.Msg, logger *log.Entry, reason string) {
-	logger.Tracef("continuing to next handler for domain=%s type=%s reason=%s",
-		r.Question[0].Name, dns.TypeToString[r.Question[0].Qtype], reason)
-
-	resutil.RequestSoftNegative(w)
-	// Carried to the chain's response log line: without it the answer looks like
-	// it came from the fallthrough resolver on its own.
-	resutil.SetMeta(w, resutil.MetaKeyDeferredBy, "dns-route")
-	resutil.SetMeta(w, resutil.MetaKeyDeferredReason, reason)
-
-	resp := new(dns.Msg)
-	resp.SetRcode(r, dns.RcodeNameError)
-	// Set Zero bit to signal handler chain to continue
-	resp.MsgHdr.Zero = true
-	if err := w.WriteMsg(resp); err != nil {
-		logger.Errorf("failed writing DNS continue response: %v", err)
-	}
-}
-
 func (d *DnsInterceptor) writeDNSError(w dns.ResponseWriter, r *dns.Msg, logger *log.Entry, reason string) {
 	logger.Warnf("failed to query upstream for domain=%s: %s", r.Question[0].Name, reason)
 
@@ -700,26 +621,3 @@ func (d *DnsInterceptor) debugPeerTimeout(peerIP netip.Addr, peerKey string) str
 
 	return fmt.Sprintf(" (peer %s)", nbdns.FormatPeerStatus(&peerState))
 }
-
-// peerQuery builds the query sent to a peer's DNS forwarder and reports whether
-// the client itself advertised EDNS0. EDNS0 is added so the forwarder can return
-// an Extended DNS Error describing an upstream failure, and the OPT is stripped
-// from the reply again when the client did not ask for it. The client's message
-// is left untouched: it may still be handed to the next handler in the chain,
-// and neither the OPT nor the AD bit is ours to put on the wire on its behalf.
-func peerQuery(r *dns.Msg) (query *dns.Msg, hadEdns bool) {
-	query = r.Copy()
-	hadEdns = query.IsEdns0() != nil
-
-	// AD tells the forwarder we understand authenticated data. Only set when the
-	// client sent no additional section of its own, so we never overrule what it
-	// asked for.
-	if len(query.Extra) == 0 {
-		query.MsgHdr.AuthenticatedData = true
-	}
-	if !hadEdns {
-		query.SetEdns0(dns.DefaultMsgSize, false)
-	}
-
-	return query, hadEdns
-}

client/internal/routemanager/dnsinterceptor/handler_test.go

@@ -1,399 +0,0 @@
-package dnsinterceptor
-
-import (
-	"net"
-	"net/netip"
-	"sync"
-	"sync/atomic"
-	"testing"
-	"time"
-
-	"github.com/miekg/dns"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-	"golang.zx2c4.com/wireguard/tun/netstack"
-
-	"github.com/netbirdio/netbird/client/iface/device"
-	"github.com/netbirdio/netbird/client/iface/wgaddr"
-	"github.com/netbirdio/netbird/client/internal/dns/resutil"
-	"github.com/netbirdio/netbird/client/internal/dns/test"
-	"github.com/netbirdio/netbird/client/internal/peer"
-	"github.com/netbirdio/netbird/route"
-)
-
-// softNegativeWriter records what the handler told the chain: whether to soften
-// a negative verdict from the handlers it defers to, and the metadata that ends
-// up on the chain's response log line.
-type softNegativeWriter struct {
-	test.MockResponseWriter
-	softNegative bool
-	meta         map[resutil.MetaKey]string
-}
-
-func (w *softNegativeWriter) RequestSoftNegative() { w.softNegative = true }
-
-func (w *softNegativeWriter) RequestID() string { return "test" }
-
-func (w *softNegativeWriter) SetMeta(key resutil.MetaKey, value string) {
-	if w.meta == nil {
-		w.meta = make(map[resutil.MetaKey]string)
-	}
-	w.meta[key] = value
-}
-
-// TestServeDNS_QtypeTheForwarderCannotResolve covers the record types no DNS
-// forwarder can answer, because the host resolver exposes no API for them.
-// Asking the peer only burns a tunnel round trip, so the query goes straight to
-// the rest of the chain. No peer key is configured, proving no round trip is
-// attempted.
-func TestServeDNS_QtypeTheForwarderCannotResolve(t *testing.T) {
-	qtypes := []uint16{
-		dns.TypeHTTPS,
-		dns.TypeSVCB,
-		dns.TypeCAA,
-		dns.TypeNAPTR,
-		dns.TypeTLSA,
-		dns.TypeSOA,
-	}
-
-	for _, qtype := range qtypes {
-		t.Run(dns.TypeToString[qtype], func(t *testing.T) {
-			d := &DnsInterceptor{}
-			w := &softNegativeWriter{}
-
-			r := new(dns.Msg)
-			r.SetQuestion("db.example.com.", qtype)
-
-			d.ServeDNS(w, r)
-
-			resp := w.GetLastResponse()
-			require.NotNil(t, resp, "a response must be written")
-			assert.Equal(t, dns.RcodeNameError, resp.Rcode, "chain continuation is signalled as NXDOMAIN")
-			assert.True(t, resp.MsgHdr.Zero, "Zero bit must be set so the chain continues")
-			assert.True(t, w.softNegative,
-				"a downstream NXDOMAIN must be softened, or the fallthrough poisons the routed name")
-			assert.Equal(t, "dns-route", w.meta[resutil.MetaKeyDeferredBy],
-				"the chain's response log line must name us as the handler that stepped aside")
-			assert.Equal(t, "unsupported-qtype", w.meta[resutil.MetaKeyDeferredReason],
-				"the reason must survive to the log line, or the fallthrough is invisible")
-		})
-	}
-}
-
-// TestServeDNS_AddressQtypeNeverFallsThrough locks in the opposite case: A and
-// AAAA are what the route exists for, and the peer is authoritative for them.
-// A missing peer is an error the client must see, never a fallthrough to a
-// resolver that knows nothing about the routed name.
-func TestServeDNS_AddressQtypeNeverFallsThrough(t *testing.T) {
-	for _, qtype := range []uint16{dns.TypeA, dns.TypeAAAA} {
-		t.Run(dns.TypeToString[qtype], func(t *testing.T) {
-			d := &DnsInterceptor{}
-			w := &softNegativeWriter{}
-
-			r := new(dns.Msg)
-			r.SetQuestion("db.example.com.", qtype)
-
-			d.ServeDNS(w, r)
-
-			resp := w.GetLastResponse()
-			require.NotNil(t, resp, "a response must be written")
-			assert.Equal(t, dns.RcodeServerFailure, resp.Rcode, "an unusable route must fail, not fall through")
-			assert.False(t, resp.MsgHdr.Zero, "the chain must not continue for address queries")
-			assert.False(t, w.softNegative, "no fallthrough means nothing to soften")
-		})
-	}
-}
-
-// TestDispositionFor pins the query-type policy, and ties the types we ask a
-// peer about to the types a forwarder can actually resolve, so the two ends
-// cannot drift apart.
-func TestDispositionFor(t *testing.T) {
-	tests := []struct {
-		qtype uint16
-		want  disposition
-	}{
-		{dns.TypeA, dispositionPeer},
-		{dns.TypeAAAA, dispositionPeer},
-		{dns.TypeMX, dispositionPeerFirst},
-		{dns.TypeTXT, dispositionPeerFirst},
-		{dns.TypeNS, dispositionPeerFirst},
-		{dns.TypeSRV, dispositionPeerFirst},
-		{dns.TypeCNAME, dispositionPeerFirst},
-		{dns.TypePTR, dispositionPeerFirst},
-		{dns.TypeHTTPS, dispositionChain},
-		{dns.TypeSVCB, dispositionChain},
-		{dns.TypeCAA, dispositionChain},
-		{dns.TypeNAPTR, dispositionChain},
-		{dns.TypeTLSA, dispositionChain},
-		{dns.TypeDS, dispositionChain},
-		{dns.TypeDNSKEY, dispositionChain},
-		{dns.TypeSOA, dispositionChain},
-		{dns.TypeANY, dispositionChain},
-	}
-
-	for _, tt := range tests {
-		t.Run(dns.TypeToString[tt.qtype], func(t *testing.T) {
-			assert.Equal(t, tt.want, dispositionFor(tt.qtype), "disposition for %s", dns.TypeToString[tt.qtype])
-
-			// Address types are resolved by the forwarder's address path, not by
-			// its record path, so they are deliberately outside
-			// SupportedRecordQtype and their disposition does not depend on it.
-			if tt.want == dispositionPeer {
-				assert.False(t, resutil.SupportedRecordQtype(tt.qtype),
-					"address types take the forwarder's address path")
-				return
-			}
-
-			assert.Equal(t, tt.want == dispositionPeerFirst, resutil.SupportedRecordQtype(tt.qtype),
-				"only types a forwarder resolves may be sent to a peer")
-		})
-	}
-}
-
-// TestPeerCannotAnswer separates a peer that cannot resolve a type from one that
-// resolved it and found nothing. Only the former may be retried elsewhere:
-// treating NODATA or NXDOMAIN as a capability failure would send the client to a
-// public resolver behind the peer's back and prefer a public record over the
-// private one the route exists to reach.
-func TestPeerCannotAnswer(t *testing.T) {
-	cannot := []int{dns.RcodeNotImplemented, dns.RcodeFormatError}
-	for _, rcode := range cannot {
-		assert.True(t, peerCannotAnswer(rcode), "%s means the peer cannot answer", dns.RcodeToString[rcode])
-	}
-
-	// REFUSED belongs here, not above: it says the peer does not hold the name,
-	// so retrying elsewhere would leak the name of an internal-only domain to a
-	// public resolver. No forwarder version reports a missing record type this
-	// way, so nothing is lost by keeping it out.
-	can := []int{
-		dns.RcodeSuccess,
-		dns.RcodeNameError,
-		dns.RcodeServerFailure,
-		dns.RcodeNotAuth,
-		dns.RcodeRefused,
-	}
-	for _, rcode := range can {
-		assert.False(t, peerCannotAnswer(rcode), "%s is the peer's answer, not a capability failure", dns.RcodeToString[rcode])
-	}
-}
-
-// fakeWGIface is the minimum an interceptor needs to reach a peer's DNS
-// forwarder over a plain socket. A nil netstack keeps the exchange on the host
-// stack so the test can point it at a loopback listener.
-type fakeWGIface struct{}
-
-func (fakeWGIface) AddAllowedIP(string, netip.Prefix) error    { return nil }
-func (fakeWGIface) RemoveAllowedIP(string, netip.Prefix) error { return nil }
-func (fakeWGIface) Name() string                               { return "wt0" }
-func (fakeWGIface) Address() wgaddr.Address                    { return wgaddr.Address{} }
-func (fakeWGIface) ToInterface() *net.Interface                { return nil }
-func (fakeWGIface) IsUserspaceBind() bool                      { return false }
-func (fakeWGIface) GetFilter() device.PacketFilter             { return nil }
-func (fakeWGIface) GetDevice() *device.FilteredDevice          { return nil }
-func (fakeWGIface) GetNet() *netstack.Net                      { return nil }
-
-// fakePeerIPs resolves every peer key to the loopback address the test's
-// forwarder stub listens on.
-type fakePeerIPs struct {
-	addr netip.Addr
-}
-
-func (p fakePeerIPs) AllowedIP(string) (netip.Addr, bool) { return p.addr, p.addr.IsValid() }
-
-// forwarderStub stands in for the DNS forwarder of a routing peer, recording
-// what it was asked and answering with a canned reply.
-type forwarderStub struct {
-	mu       sync.Mutex
-	queries  []*dns.Msg
-	reply    func(q *dns.Msg) *dns.Msg
-	port     uint16
-	shutdown func()
-}
-
-func (s *forwarderStub) received() []*dns.Msg {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	return append([]*dns.Msg(nil), s.queries...)
-}
-
-func newForwarderStub(t *testing.T, reply func(q *dns.Msg) *dns.Msg) *forwarderStub {
-	t.Helper()
-
-	conn, err := net.ListenUDP("udp", net.UDPAddrFromAddrPort(netip.MustParseAddrPort("127.0.0.1:0")))
-	require.NoError(t, err, "listen for the forwarder stub")
-
-	stub := &forwarderStub{reply: reply}
-	stub.port = uint16(conn.LocalAddr().(*net.UDPAddr).Port)
-
-	mux := dns.NewServeMux()
-	mux.HandleFunc(".", func(w dns.ResponseWriter, q *dns.Msg) {
-		stub.mu.Lock()
-		stub.queries = append(stub.queries, q.Copy())
-		stub.mu.Unlock()
-		_ = w.WriteMsg(stub.reply(q))
-	})
-
-	srv := &dns.Server{PacketConn: conn, Handler: mux}
-	started := make(chan struct{})
-	srv.NotifyStartedFunc = func() { close(started) }
-	go func() {
-		_ = srv.ActivateAndServe()
-	}()
-	select {
-	case <-started:
-	case <-time.After(5 * time.Second):
-		t.Fatal("forwarder stub did not start")
-	}
-
-	stub.shutdown = func() { _ = srv.Shutdown() }
-	t.Cleanup(stub.shutdown)
-
-	return stub
-}
-
-func newTestInterceptor(t *testing.T, stub *forwarderStub) *DnsInterceptor {
-	t.Helper()
-
-	port := new(atomic.Uint32)
-	port.Store(uint32(stub.port))
-
-	return &DnsInterceptor{
-		route:              &route.Route{Domains: nil},
-		statusRecorder:     peer.NewRecorder("https://mgm"),
-		currentPeerKey:     "peer-key",
-		interceptedDomains: make(domainMap),
-		wgInterface:        fakeWGIface{},
-		peerStore:          fakePeerIPs{addr: netip.MustParseAddr("127.0.0.1")},
-		forwarderPort:      port,
-	}
-}
-
-// TestServeDNS_PeerCannotResolveQtype is the reported regression: a routing peer
-// running a client older than the one that taught the forwarder about non-address
-// record types answers NOTIMP for every SRV query, and the interceptor used to
-// hand that straight to the application, breaking SRV-based service discovery.
-// The client cannot know the peer's capability up front, so it must try the peer
-// and fall back to the chain when the peer says it cannot answer.
-func TestServeDNS_PeerCannotResolveQtype(t *testing.T) {
-	for _, rcode := range []int{dns.RcodeNotImplemented, dns.RcodeFormatError} {
-		t.Run(dns.RcodeToString[rcode], func(t *testing.T) {
-			stub := newForwarderStub(t, func(q *dns.Msg) *dns.Msg {
-				resp := new(dns.Msg)
-				resp.SetRcode(q, rcode)
-				return resp
-			})
-			d := newTestInterceptor(t, stub)
-			w := &softNegativeWriter{}
-
-			r := new(dns.Msg)
-			r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-
-			d.ServeDNS(w, r)
-
-			require.Len(t, stub.received(), 1, "the peer must be asked before giving up on it")
-
-			resp := w.GetLastResponse()
-			require.NotNil(t, resp, "a response must be written")
-			assert.Equal(t, dns.RcodeNameError, resp.Rcode, "chain continuation is signalled as NXDOMAIN")
-			assert.True(t, resp.MsgHdr.Zero, "Zero bit must be set so the chain continues")
-			assert.True(t, w.softNegative, "the fallthrough must not be allowed to poison the routed name")
-			assert.Equal(t, "dns-route", w.meta[resutil.MetaKeyDeferredBy])
-			assert.Equal(t, "peer-rcode-"+dns.RcodeToString[rcode], w.meta[resutil.MetaKeyDeferredReason],
-				"the peer's verdict must be visible in the log, it is the only trace of the probe")
-		})
-	}
-}
-
-// TestServeDNS_PeerVerdictPassesThrough covers the replies that are the peer's
-// answer rather than a statement about its capability. The peer is authoritative
-// for a name routed to it, so its verdict reaches the client as it is: retrying
-// these elsewhere would send the name of an internal-only domain to a public
-// resolver and prefer whatever that resolver says over the route.
-func TestServeDNS_PeerVerdictPassesThrough(t *testing.T) {
-	// REFUSED means the peer does not consider the name routed to it, which
-	// happens while the client and the peer disagree about the route set. It is
-	// never how a peer reports a record type it cannot resolve: a forwarder too
-	// old for non-address types answers NOTIMP before it looks at the domain.
-	for _, rcode := range []int{dns.RcodeRefused, dns.RcodeNameError, dns.RcodeSuccess} {
-		t.Run(dns.RcodeToString[rcode], func(t *testing.T) {
-			stub := newForwarderStub(t, func(q *dns.Msg) *dns.Msg {
-				resp := new(dns.Msg)
-				resp.SetRcode(q, rcode)
-				return resp
-			})
-			d := newTestInterceptor(t, stub)
-			w := &softNegativeWriter{}
-
-			r := new(dns.Msg)
-			r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-
-			d.ServeDNS(w, r)
-
-			require.Len(t, stub.received(), 1, "the peer must be asked exactly once")
-
-			resp := w.GetLastResponse()
-			require.NotNil(t, resp, "a response must be written")
-			assert.Equal(t, rcode, resp.Rcode, "the peer's verdict must reach the client unchanged")
-			assert.False(t, resp.MsgHdr.Zero, "the chain must not continue past an answered query")
-			assert.False(t, w.softNegative, "nothing was deferred, so nothing may be softened")
-		})
-	}
-}
-
-// TestServeDNS_PeerResolvesQtype guards the common case: a peer that can answer
-// the type stays authoritative, and its records reach the client unchanged.
-func TestServeDNS_PeerResolvesQtype(t *testing.T) {
-	stub := newForwarderStub(t, func(q *dns.Msg) *dns.Msg {
-		resp := new(dns.Msg)
-		resp.SetReply(q)
-		resp.Answer = []dns.RR{&dns.SRV{
-			Hdr:    dns.RR_Header{Name: q.Question[0].Name, Rrtype: dns.TypeSRV, Class: dns.ClassINET, Ttl: 60},
-			Target: "shard-00.db.example.com.",
-			Port:   27017,
-		}}
-		return resp
-	})
-	d := newTestInterceptor(t, stub)
-	w := &softNegativeWriter{}
-
-	r := new(dns.Msg)
-	r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-
-	d.ServeDNS(w, r)
-
-	resp := w.GetLastResponse()
-	require.NotNil(t, resp, "a response must be written")
-	assert.Equal(t, dns.RcodeSuccess, resp.Rcode)
-	require.Len(t, resp.Answer, 1, "the peer's records must pass through")
-	assert.False(t, resp.MsgHdr.Zero, "an answered query must not continue the chain")
-	assert.False(t, w.softNegative)
-}
-
-// TestServeDNS_DeferredQueryIsPristine covers what the fallthrough hands to the
-// next handler. The interceptor advertises EDNS0 to the peer so the forwarder can
-// return an Extended DNS Error, and sets AD, but neither belongs to the client's
-// query: on the deferred path they would travel to the public resolver and the
-// reply could come back carrying an OPT the client never advertised, which
-// RFC 6891 forbids us from passing on.
-func TestServeDNS_DeferredQueryIsPristine(t *testing.T) {
-	stub := newForwarderStub(t, func(q *dns.Msg) *dns.Msg {
-		resp := new(dns.Msg)
-		resp.SetRcode(q, dns.RcodeNotImplemented)
-		return resp
-	})
-	d := newTestInterceptor(t, stub)
-	w := &softNegativeWriter{}
-
-	r := new(dns.Msg)
-	r.SetQuestion("_mongodb._tcp.db.example.com.", dns.TypeSRV)
-
-	d.ServeDNS(w, r)
-
-	sent := stub.received()
-	require.Len(t, sent, 1)
-	assert.NotNil(t, sent[0].IsEdns0(), "the peer must be asked with EDNS0 so it can return an EDE")
-
-	assert.Nil(t, r.IsEdns0(), "the deferred query must not carry an OPT the client never sent")
-	assert.False(t, r.AuthenticatedData, "the deferred query must not carry an AD bit the client never set")
-	assert.Empty(t, r.Extra, "the deferred query must reach the next handler as the client sent it")
-}

client/internal/routemanager/dnsinterceptor/handler_windows_test.go

@@ -1,7 +0,0 @@
-//go:build windows
-
-package dnsinterceptor
-
-// GetInterfaceGUIDString completes iface.WGIface on Windows, which requires the
-// interface GUID for DNS registration.
-func (fakeWGIface) GetInterfaceGUIDString() (string, error) { return "", nil }

client/internal/routemanager/manager.go

@@ -52,6 +52,10 @@ type Manager interface {
 	UpdateRoutes(updateSerial uint64, serverRoutes map[route.ID]*route.Route, clientRoutes route.HAMap, useNewDNSRoute bool) error
 	ClassifyRoutes(newRoutes []*route.Route) (map[route.ID]*route.Route, route.HAMap)
 	TriggerSelection(route.HAMap)
+	SelectRoutes(ids []route.NetID, appendRoute bool) error
+	DeselectRoutes(ids []route.NetID) error
+	SelectAllRoutes()
+	DeselectAllRoutes()
 	GetRouteSelector() *routeselector.RouteSelector
 	GetClientRoutes() route.HAMap
 	GetSelectedClientRoutes() route.HAMap
@@ -800,7 +804,7 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 	var info exitNodeInfo
 
 	for haID, routes := range clientRoutes {
-		if !m.isExitNodeRoute(routes) {
+		if !isExitNodeRoutes(routes) {
 			continue
 		}
 
@@ -820,13 +824,6 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 	return info
 }
 
-func (m *DefaultManager) isExitNodeRoute(routes []*route.Route) bool {
-	if len(routes) == 0 {
-		return false
-	}
-	return route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network)
-}
-
 func (m *DefaultManager) categorizeUserSelection(netID route.NetID, info *exitNodeInfo) {
 	if m.routeSelector.IsSelected(netID) {
 		info.userSelected = append(info.userSelected, netID)

client/internal/routemanager/mock.go

@@ -16,6 +16,8 @@ type MockManager struct {
 	ClassifyRoutesFunc           func(routes []*route.Route) (map[route.ID]*route.Route, route.HAMap)
 	UpdateRoutesFunc             func(updateSerial uint64, serverRoutes map[route.ID]*route.Route, clientRoutes route.HAMap, useNewDNSRoute bool) error
 	TriggerSelectionFunc         func(haMap route.HAMap)
+	SelectRoutesFunc             func(ids []route.NetID, appendRoute bool) error
+	DeselectRoutesFunc           func(ids []route.NetID) error
 	GetRouteSelectorFunc         func() *routeselector.RouteSelector
 	GetClientRoutesFunc          func() route.HAMap
 	GetSelectedClientRoutesFunc  func() route.HAMap
@@ -55,6 +57,30 @@ func (m *MockManager) TriggerSelection(networks route.HAMap) {
 	}
 }
 
+// SelectRoutes mock implementation of SelectRoutes from Manager interface
+func (m *MockManager) SelectRoutes(ids []route.NetID, appendRoute bool) error {
+	if m.SelectRoutesFunc != nil {
+		return m.SelectRoutesFunc(ids, appendRoute)
+	}
+	return nil
+}
+
+// DeselectRoutes mock implementation of DeselectRoutes from Manager interface
+func (m *MockManager) DeselectRoutes(ids []route.NetID) error {
+	if m.DeselectRoutesFunc != nil {
+		return m.DeselectRoutesFunc(ids)
+	}
+	return nil
+}
+
+// SelectAllRoutes mock implementation of SelectAllRoutes from Manager interface
+func (m *MockManager) SelectAllRoutes() {
+}
+
+// DeselectAllRoutes mock implementation of DeselectAllRoutes from Manager interface
+func (m *MockManager) DeselectAllRoutes() {
+}
+
 // GetRouteSelector mock implementation of GetRouteSelector from Manager interface
 func (m *MockManager) GetRouteSelector() *routeselector.RouteSelector {
 	if m.GetRouteSelectorFunc != nil {

client/internal/routemanager/selection.go

@@ -0,0 +1,138 @@
+package routemanager
+
+import (
+	"fmt"
+	"slices"
+
+	"github.com/hashicorp/go-multierror"
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/exp/maps"
+
+	nberrors "github.com/netbirdio/netbird/client/errors"
+	"github.com/netbirdio/netbird/route"
+)
+
+// SelectRoutes selects the routes with the given network IDs and applies the
+// new selection. V4/v6 exit-node pairs are expanded automatically. Exit nodes
+// are mutually exclusive: if the selection activates an exit node, every other
+// available exit node is deselected so two can't be active at once. With
+// appendRoute=false the previous selection is replaced instead of extended.
+func (m *DefaultManager) SelectRoutes(ids []route.NetID, appendRoute bool) error {
+	if err := m.selectRoutes(ids, appendRoute); err != nil {
+		return err
+	}
+	m.TriggerSelection(m.GetClientRoutes())
+	return nil
+}
+
+// DeselectRoutes removes the routes with the given network IDs from the
+// selection and applies the change. V4/v6 exit-node pairs are expanded
+// automatically.
+func (m *DefaultManager) DeselectRoutes(ids []route.NetID) error {
+	if err := m.deselectRoutes(ids); err != nil {
+		return err
+	}
+	m.TriggerSelection(m.GetClientRoutes())
+	return nil
+}
+
+func (m *DefaultManager) deselectRoutes(ids []route.NetID) error {
+	routesMap := m.GetClientRoutesWithNetID()
+	routes := route.ExpandV6ExitPairs(slices.Clone(ids), routesMap)
+
+	log.Debugf("deselecting routes with ids: %v", routes)
+
+	if err := m.routeSelector.DeselectRoutes(routes, maps.Keys(routesMap)); err != nil {
+		return fmt.Errorf("deselect routes: %w", err)
+	}
+
+	return nil
+}
+
+// SelectAllRoutes selects every available route and applies the selection.
+// Exit nodes stay mutually exclusive: at most one remains active.
+func (m *DefaultManager) SelectAllRoutes() {
+	m.selectAllRoutes()
+	m.TriggerSelection(m.GetClientRoutes())
+}
+
+func (m *DefaultManager) selectAllRoutes() {
+	m.routeSelector.SelectAllRoutes()
+
+	// Select-all wipes every explicit selection, so exit nodes fall back to
+	// management's auto-apply flags — which may mark several at once.
+	// Reconcile immediately so at most one exit node stays active instead of
+	// waiting for the next network map to enforce it.
+	m.mux.Lock()
+	defer m.mux.Unlock()
+	m.updateRouteSelectorFromManagement(m.clientRoutes)
+}
+
+// DeselectAllRoutes deselects every route and applies the change.
+func (m *DefaultManager) DeselectAllRoutes() {
+	m.routeSelector.DeselectAllRoutes()
+	m.TriggerSelection(m.GetClientRoutes())
+}
+
+func (m *DefaultManager) selectRoutes(ids []route.NetID, appendRoute bool) error {
+	routesMap := m.GetClientRoutesWithNetID()
+	routes := route.ExpandV6ExitPairs(slices.Clone(ids), routesMap)
+	allIDs := maps.Keys(routesMap)
+
+	log.Debugf("selecting routes with ids: %v", routes)
+
+	// A partial failure (e.g. an unknown ID in the request) still selects the
+	// valid routes, so exclusivity below must run regardless of the error.
+	var merr *multierror.Error
+	if err := m.routeSelector.SelectRoutes(routes, appendRoute, allIDs); err != nil {
+		merr = multierror.Append(merr, fmt.Errorf("select routes: %w", err))
+	}
+
+	// Exit nodes are mutually exclusive: if this selection activates an
+	// exit node, deselect every other available exit node so two can't be
+	// selected at once. Non-exit route selections are left untouched.
+	if requestActivatesExitNode(routes, routesMap) {
+		if others := otherExitNodeIDs(routesMap, routes); len(others) > 0 {
+			if err := m.routeSelector.DeselectRoutes(others, allIDs); err != nil {
+				merr = multierror.Append(merr, fmt.Errorf("deselect sibling exit nodes: %w", err))
+			}
+		}
+	}
+
+	return nberrors.FormatErrorOrNil(merr)
+}
+
+func isExitNodeRoutes(routes []*route.Route) bool {
+	return len(routes) > 0 && (route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network))
+}
+
+// requestActivatesExitNode reports whether any requested NetID maps to an exit
+// node (default route) in the current route table.
+func requestActivatesExitNode(requested []route.NetID, routesMap map[route.NetID][]*route.Route) bool {
+	for _, id := range requested {
+		if isExitNodeRoutes(routesMap[id]) {
+			return true
+		}
+	}
+	return false
+}
+
+// otherExitNodeIDs returns every available exit-node NetID that is not in the
+// requested set — the siblings to deselect so a single exit node stays active.
+func otherExitNodeIDs(routesMap map[route.NetID][]*route.Route, requested []route.NetID) []route.NetID {
+	keep := make(map[route.NetID]struct{}, len(requested))
+	for _, id := range requested {
+		keep[id] = struct{}{}
+	}
+	var others []route.NetID
+	for id, routes := range routesMap {
+		if !isExitNodeRoutes(routes) {
+			continue
+		}
+		if _, ok := keep[id]; ok {
+			continue
+		}
+		others = append(others, id)
+	}
+	return others
+}

client/internal/routemanager/selection_test.go

@@ -0,0 +1,129 @@
+package routemanager
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/routeselector"
+	"github.com/netbirdio/netbird/route"
+)
+
+func v6ExitRoute(netID, peer string) *route.Route {
+	return &route.Route{
+		NetID:   route.NetID(netID),
+		Network: netip.MustParsePrefix("::/0"),
+		Peer:    peer,
+	}
+}
+
+func newSelectionTestManager() *DefaultManager {
+	return &DefaultManager{
+		routeSelector: routeselector.NewRouteSelector(),
+		clientRoutes: route.HAMap{
+			"exitA|0.0.0.0/0":    {exitRoute("exitA", "p1", true)},
+			"exitA-v6|::/0":      {v6ExitRoute("exitA-v6", "p1")},
+			"exitB|0.0.0.0/0":    {exitRoute("exitB", "p2", true)},
+			"lan|192.168.1.0/24": {{NetID: "lan", Network: netip.MustParsePrefix("192.168.1.0/24"), Peer: "p3"}},
+		},
+	}
+}
+
+func TestSelectRoutes_ExitNodeExclusivity(t *testing.T) {
+	m := newSelectionTestManager()
+
+	// Selecting an exit node selects its v6 pair and deselects the sibling.
+	require.NoError(t, m.selectRoutes([]route.NetID{"exitA"}, true))
+	assert.True(t, m.routeSelector.IsSelected("exitA"), "exitA should be selected")
+	assert.True(t, m.routeSelector.IsSelected("exitA-v6"), "the v6 pair follows its v4 base")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "the sibling exit node must be deselected")
+
+	// Switching to the sibling deselects the previous exit node and its v6 pair.
+	require.NoError(t, m.selectRoutes([]route.NetID{"exitB"}, true))
+	assert.True(t, m.routeSelector.IsSelected("exitB"), "exitB should now be selected")
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "the previous exit node must be deselected")
+	assert.False(t, m.routeSelector.IsSelected("exitA-v6"), "the previous exit node's v6 pair must be deselected")
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit route selection is untouched")
+
+	// Selecting a non-exit route leaves the active exit node alone.
+	require.NoError(t, m.selectRoutes([]route.NetID{"lan"}, true))
+	assert.True(t, m.routeSelector.IsSelected("exitB"), "selecting a non-exit route keeps the exit node")
+
+	// Deselecting the active exit node turns every exit node off.
+	require.NoError(t, m.deselectRoutes([]route.NetID{"exitB"}))
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "exitB should be deselected")
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "exitA stays deselected")
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit route selection is untouched")
+}
+
+func TestSelectRoutes_PartialErrorStillEnforcesExclusivity(t *testing.T) {
+	// The unknown ID must be reported, but the valid exit node in the same
+	// request is still selected — so its sibling must still be deselected.
+	// Both orderings are covered: processing must continue past the invalid
+	// ID wherever it sits in the request.
+	requests := map[string][]route.NetID{
+		"invalid id first": {"missing", "exitB"},
+		"invalid id last":  {"exitB", "missing"},
+	}
+
+	for name, ids := range requests {
+		t.Run(name, func(t *testing.T) {
+			m := newSelectionTestManager()
+
+			require.NoError(t, m.selectRoutes([]route.NetID{"exitA"}, true))
+
+			err := m.selectRoutes(ids, true)
+			assert.Error(t, err, "unknown id must be reported")
+			assert.True(t, m.routeSelector.IsSelected("exitB"), "valid exit node from the request is selected")
+			assert.False(t, m.routeSelector.IsSelected("exitA"), "sibling exit node must be deselected despite the error")
+			assert.False(t, m.routeSelector.IsSelected("exitA-v6"), "sibling's v6 pair must be deselected too")
+		})
+	}
+}
+
+func TestSelectAllRoutes_KeepsSingleExitNode(t *testing.T) {
+	// Both exit nodes are marked for auto-apply by management
+	// (SkipAutoApply=false), the state where select-all could turn on two at
+	// once without the immediate reconciliation.
+	m := &DefaultManager{
+		routeSelector: routeselector.NewRouteSelector(),
+		clientRoutes: route.HAMap{
+			"exitA|0.0.0.0/0":    {exitRoute("exitA", "p1", false)},
+			"exitB|0.0.0.0/0":    {exitRoute("exitB", "p2", false)},
+			"lan|192.168.1.0/24": {{NetID: "lan", Network: netip.MustParsePrefix("192.168.1.0/24"), Peer: "p3"}},
+		},
+	}
+
+	require.NoError(t, m.selectRoutes([]route.NetID{"exitB"}, true))
+
+	m.selectAllRoutes()
+
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit routes are all selected")
+	assert.True(t, m.routeSelector.IsSelected("exitA"), "the deterministic management pick stays active")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "select-all must not leave a second exit node active")
+}
+
+func TestSelectRoutes_UnknownRoute(t *testing.T) {
+	m := newSelectionTestManager()
+
+	assert.Error(t, m.selectRoutes([]route.NetID{"missing"}, true), "selecting an unavailable route must fail")
+	assert.Error(t, m.deselectRoutes([]route.NetID{"missing"}), "deselecting an unavailable route must fail")
+}
+
+func TestExitNodeSelectionHelpers(t *testing.T) {
+	routesMap := map[route.NetID][]*route.Route{
+		"exitA": {{Network: netip.MustParsePrefix("0.0.0.0/0")}},
+		"exitB": {{Network: netip.MustParsePrefix("::/0")}},
+		"lan":   {{Network: netip.MustParsePrefix("192.168.0.0/16")}},
+	}
+
+	assert.True(t, requestActivatesExitNode([]route.NetID{"exitA"}, routesMap), "v4 default route is an exit node")
+	assert.True(t, requestActivatesExitNode([]route.NetID{"exitB"}, routesMap), "v6 default route is an exit node")
+	assert.False(t, requestActivatesExitNode([]route.NetID{"lan"}, routesMap), "lan route is not an exit node")
+	assert.False(t, requestActivatesExitNode([]route.NetID{"missing"}, routesMap), "unknown id is not an exit node")
+
+	others := otherExitNodeIDs(routesMap, []route.NetID{"exitB"})
+	assert.ElementsMatch(t, []route.NetID{"exitA"}, others, "only the other exit node is a sibling; the lan route is ignored")
+}

client/ios/NetBirdSDK/client.go

@@ -13,7 +13,6 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
-	"golang.org/x/exp/maps"
 
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/auth"
@@ -637,23 +636,18 @@ func (c *Client) SelectRoute(id string) error {
 	}
 
 	routeManager := engine.GetRouteManager()
-	routeSelector := routeManager.GetRouteSelector()
 	if id == "All" {
 		log.Debugf("select all routes")
-		routeSelector.SelectAllRoutes()
-	} else {
-		log.Debugf("select route with id: %s", id)
-		routes := toNetIDs([]string{id})
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		if err := routeSelector.SelectRoutes(routes, true, maps.Keys(routesMap)); err != nil {
-			log.Debugf("error when selecting routes: %s", err)
-			return fmt.Errorf("select routes: %w", err)
-		}
+		routeManager.SelectAllRoutes()
+		return nil
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
-	return nil
 
+	log.Debugf("select route with id: %s", id)
+	if err := routeManager.SelectRoutes(toNetIDs([]string{id}), true); err != nil {
+		log.Debugf("error when selecting routes: %s", err)
+		return err
+	}
+	return nil
 }
 
 func (c *Client) DeselectRoute(id string) error {
@@ -667,21 +661,17 @@ func (c *Client) DeselectRoute(id string) error {
 	}
 
 	routeManager := engine.GetRouteManager()
-	routeSelector := routeManager.GetRouteSelector()
 	if id == "All" {
 		log.Debugf("deselect all routes")
-		routeSelector.DeselectAllRoutes()
-	} else {
-		log.Debugf("deselect route with id: %s", id)
-		routes := toNetIDs([]string{id})
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		if err := routeSelector.DeselectRoutes(routes, maps.Keys(routesMap)); err != nil {
-			log.Debugf("error when deselecting routes: %s", err)
-			return fmt.Errorf("deselect routes: %w", err)
-		}
+		routeManager.DeselectAllRoutes()
+		return nil
+	}
+
+	log.Debugf("deselect route with id: %s", id)
+	if err := routeManager.DeselectRoutes(toNetIDs([]string{id})); err != nil {
+		log.Debugf("error when deselecting routes: %s", err)
+		return err
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
 	return nil
 }
 

client/ios/NetBirdSDK/version.go

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

client/server/network.go

@@ -8,7 +8,6 @@ import (
 	"sort"
 	"strings"
 
-	"golang.org/x/exp/maps"
 	"google.golang.org/grpc/codes"
 	gstatus "google.golang.org/grpc/status"
 
@@ -161,30 +160,11 @@ func (s *Server) SelectNetworks(_ context.Context, req *proto.SelectNetworksRequ
 		return nil, fmt.Errorf("no route manager")
 	}
 
-	routeSelector := routeManager.GetRouteSelector()
 	if req.GetAll() {
-		routeSelector.SelectAllRoutes()
-	} else {
-		routes := toNetIDs(req.GetNetworkIDs())
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		netIdRoutes := maps.Keys(routesMap)
-		if err := routeSelector.SelectRoutes(routes, req.GetAppend(), netIdRoutes); err != nil {
-			return nil, fmt.Errorf("select routes: %w", err)
-		}
-
-		// Exit nodes are mutually exclusive: if this selection activates an
-		// exit node, deselect every other available exit node so two can't be
-		// selected at once. Non-exit route selections are left untouched.
-		if requestActivatesExitNode(routes, routesMap) {
-			if others := otherExitNodeIDs(routesMap, routes); len(others) > 0 {
-				if err := routeSelector.DeselectRoutes(others, netIdRoutes); err != nil {
-					return nil, fmt.Errorf("deselect sibling exit nodes: %w", err)
-				}
-			}
-		}
+		routeManager.SelectAllRoutes()
+	} else if err := routeManager.SelectRoutes(toNetIDs(req.GetNetworkIDs()), req.GetAppend()); err != nil {
+		return nil, err
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
 
 	s.statusRecorder.PublishEvent(
 		proto.SystemEvent_INFO,
@@ -224,19 +204,11 @@ func (s *Server) DeselectNetworks(_ context.Context, req *proto.SelectNetworksRe
 		return nil, fmt.Errorf("no route manager")
 	}
 
-	routeSelector := routeManager.GetRouteSelector()
 	if req.GetAll() {
-		routeSelector.DeselectAllRoutes()
-	} else {
-		routes := toNetIDs(req.GetNetworkIDs())
-		routesMap := routeManager.GetClientRoutesWithNetID()
-		routes = route.ExpandV6ExitPairs(routes, routesMap)
-		netIdRoutes := maps.Keys(routesMap)
-		if err := routeSelector.DeselectRoutes(routes, netIdRoutes); err != nil {
-			return nil, fmt.Errorf("deselect routes: %w", err)
-		}
+		routeManager.DeselectAllRoutes()
+	} else if err := routeManager.DeselectRoutes(toNetIDs(req.GetNetworkIDs())); err != nil {
+		return nil, err
 	}
-	routeManager.TriggerSelection(routeManager.GetClientRoutes())
 
 	s.statusRecorder.PublishEvent(
 		proto.SystemEvent_INFO,
@@ -261,37 +233,3 @@ func toNetIDs(routes []string) []route.NetID {
 	return netIDs
 }
 
-func isExitNodeRoutes(routes []*route.Route) bool {
-	return len(routes) > 0 && (route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network))
-}
-
-// requestActivatesExitNode reports whether any requested NetID maps to an exit
-// node (default route) in the current route table.
-func requestActivatesExitNode(requested []route.NetID, routesMap map[route.NetID][]*route.Route) bool {
-	for _, id := range requested {
-		if isExitNodeRoutes(routesMap[id]) {
-			return true
-		}
-	}
-	return false
-}
-
-// otherExitNodeIDs returns every available exit-node NetID that is not in the
-// requested set — the siblings to deselect so a single exit node stays active.
-func otherExitNodeIDs(routesMap map[route.NetID][]*route.Route, requested []route.NetID) []route.NetID {
-	keep := make(map[route.NetID]struct{}, len(requested))
-	for _, id := range requested {
-		keep[id] = struct{}{}
-	}
-	var others []route.NetID
-	for id, routes := range routesMap {
-		if !isExitNodeRoutes(routes) {
-			continue
-		}
-		if _, ok := keep[id]; ok {
-			continue
-		}
-		others = append(others, id)
-	}
-	return others
-}

client/server/network_exitnode_test.go

@@ -1,26 +0,0 @@
-package server
-
-import (
-	"net/netip"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-
-	"github.com/netbirdio/netbird/route"
-)
-
-func TestExitNodeSelectionHelpers(t *testing.T) {
-	routesMap := map[route.NetID][]*route.Route{
-		"exitA": {{Network: netip.MustParsePrefix("0.0.0.0/0")}},
-		"exitB": {{Network: netip.MustParsePrefix("::/0")}},
-		"lan":   {{Network: netip.MustParsePrefix("192.168.0.0/16")}},
-	}
-
-	assert.True(t, requestActivatesExitNode([]route.NetID{"exitA"}, routesMap), "v4 default route is an exit node")
-	assert.True(t, requestActivatesExitNode([]route.NetID{"exitB"}, routesMap), "v6 default route is an exit node")
-	assert.False(t, requestActivatesExitNode([]route.NetID{"lan"}, routesMap), "lan route is not an exit node")
-	assert.False(t, requestActivatesExitNode([]route.NetID{"missing"}, routesMap), "unknown id is not an exit node")
-
-	others := otherExitNodeIDs(routesMap, []route.NetID{"exitB"})
-	assert.ElementsMatch(t, []route.NetID{"exitA"}, others, "only the other exit node is a sibling; the lan route is ignored")
-}

client/ui/frontend/src/modules/auto-update/UpdateVersionCard.tsx

@@ -2,6 +2,7 @@ import { type ReactNode } from "react";
 import { useTranslation } from "react-i18next";
 import { Browser } from "@wailsio/runtime";
 import { DownloadIcon, NotepadText } from "lucide-react";
+import { Update as UpdateSvc } from "@bindings/services";
 import { Button } from "@/components/buttons/Button";
 import { useClientVersion } from "@/contexts/ClientVersionContext";
 import { cn } from "@/lib/cn";
@@ -14,6 +15,12 @@ function openUrl(url: string) {
     });
 }
 
+function openInstallerDownload() {
+    UpdateSvc.DownloadURL()
+        .then(openUrl)
+        .catch(() => openUrl(GITHUB_RELEASES));
+}
+
 export function UpdateVersionCard() {
     const { t } = useTranslation();
     const { updateVersion, enforced, triggerUpdate } = useClientVersion();
@@ -37,11 +44,7 @@ export function UpdateVersionCard() {
                         {t("update.card.installNow")}
                     </Button>
                 ) : (
-                    <Button
-                        variant={"primary"}
-                        size={"xs"}
-                        onClick={() => openUrl(GITHUB_RELEASES)}
-                    >
+                    <Button variant={"primary"} size={"xs"} onClick={openInstallerDownload}>
                         <DownloadIcon size={14} />
                         {t("update.card.getInstaller")}
                     </Button>

client/ui/services/update.go

@@ -10,6 +10,7 @@ import (
 
 	"github.com/netbirdio/netbird/client/proto"
 	"github.com/netbirdio/netbird/client/ui/updater"
+	"github.com/netbirdio/netbird/version"
 )
 
 // UpdateResult mirrors TriggerUpdateResponse.
@@ -33,6 +34,12 @@ func (s *Update) GetState() updater.State {
 	return s.holder.Get()
 }
 
+// DownloadURL returns the platform-appropriate installer download link for
+// manual (non-enforced) updates.
+func (s *Update) DownloadURL() string {
+	return version.DownloadUrl()
+}
+
 // Quit exits the app. Scheduled off the calling goroutine so the JS caller's
 // response returns before the runtime tears down.
 func (s *Update) Quit() {

client/ui/tray.go

@@ -32,9 +32,8 @@ const (
 
 	quitDownTimeout = 5 * time.Second
 
-	urlGitHubRepo     = "https://github.com/netbirdio/netbird"
-	urlGitHubReleases = "https://github.com/netbirdio/netbird/releases/latest"
-	urlDocs           = "https://docs.netbird.io"
+	urlGitHubRepo = "https://github.com/netbirdio/netbird"
+	urlDocs       = "https://docs.netbird.io"
 )
 
 // TrayServices bundles the services the tray menu needs, grouped so NewTray

client/ui/tray_update.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/netbirdio/netbird/client/ui/services"
 	"github.com/netbirdio/netbird/client/ui/updater"
+	"github.com/netbirdio/netbird/version"
 )
 
 // trayUpdater owns the tray UI that reacts to auto-update. Composed inside Tray.
@@ -76,15 +77,15 @@ func (u *trayUpdater) applyLanguage() {
 	u.refreshMenuItem(state)
 }
 
-// handleClick opens the GitHub releases page when not Enforced, otherwise shows
-// the progress page and asks the daemon to start the installer.
+// handleClick opens the installer download link when not Enforced, otherwise
+// shows the progress page and asks the daemon to start the installer.
 func (u *trayUpdater) handleClick() {
 	u.mu.Lock()
 	state := u.state
 	u.mu.Unlock()
 
 	if !state.Enforced {
-		_ = u.app.Browser.OpenURL(urlGitHubReleases)
+		_ = u.app.Browser.OpenURL(version.DownloadUrl())
 		return
 	}
 

docs/agent-networks/01-end-to-end-flows.md

@@ -109,7 +109,7 @@ sequenceDiagram
         Chk->>Inj: continue
         Inj->>Inj: inject NetBird identity headers per provider config
         Inj->>Grd: continue
-        Grd->>Grd: enforce model allowlist
+        Grd->>Grd: enforce per-provider allowlist (fail-closed backstop)
         Grd->>Up: forward (over WireGuard)
         Up-->>Resp: response (JSON or SSE stream)
         Resp->>Resp: parse usage tokens, completion
@@ -135,6 +135,21 @@ sequenceDiagram
   (`redact_pii = settings.RedactPii`). Phones, emails, credit cards,
   PII names — see `redact.go` for the full set. See
   [`modules/31-proxy-middleware-builtin.md`](modules/31-proxy-middleware-builtin.md).
+- The model allowlist is enforced in TWO places. `CheckLLMPolicyLimits`
+  is authoritative: it resolves the policy that governs this
+  (provider, caller-groups) and denies (`deny_code = llm_policy.model_blocked`)
+  when no applicable policy permits the model — so an allowlist scoped to
+  one group/provider never leaks to another, and an un-guardrailed policy
+  is genuinely unrestricted. `llm_guardrail` is a per-provider fail-closed
+  backstop: it only carries an allowlist for a provider every authorising
+  policy restricts, and blocks unknown/undetermined models even when
+  management is unreachable. Because that backstop allowlist is the UNION
+  of every restricting policy's models, per-group narrowing lives only in
+  the authoritative check: during a `CheckLLMPolicyLimits` outage
+  `llm_limit_check` fails open, so a caller can reach any model in the
+  provider's union — a group scoped to model A could reach model B if
+  another group restricts the same provider to B. This is the documented
+  fail-open trade-off; a future flag may switch it to fail-closed.
 - SSE streaming requires special handling on the response side; the
   parser must handle partial chunks without buffering the whole
   stream. See [`modules/32-proxy-llm-parsers.md`](modules/32-proxy-llm-parsers.md).

docs/agent-networks/modules/21-management-agentnetwork.md

@@ -122,7 +122,7 @@ At request time the path is independent: the proxy calls `SelectPolicyForRequest
   | on_request | 1 | `llm_router` | `{"providers":[{id, models[], upstream_*, auth_header_*, allowed_group_ids[]}]}` | **true** |
   | on_request | 2 | `llm_limit_check` | `{}` | – |
   | on_request | 3 | `llm_identity_inject` | `{"providers":[{provider_id, header_pair?, json_metadata?, extra_headers?}]}` | **true** |
-  | on_request | 4 | `llm_guardrail` | `{"model_allowlist"?, "prompt_capture":{enabled,redact_pii}}` | – |
+  | on_request | 4 | `llm_guardrail` | `{"provider_allowlists"?: {providerID: []model}, "prompt_capture":{enabled,redact_pii}}` | – |
   | on_response | 5 | `llm_limit_record` | `{}` (runs LAST at runtime) | – |
   | on_response | 6 | `cost_meter` | `{}` | – |
   | on_response | 7 | `llm_response_parser` | `{"capture_completion": <bool>, "redact_pii"?: true}` | – |

docs/agent-networks/modules/31-proxy-middleware-builtin.md

@@ -244,7 +244,7 @@ no mocks. Tests: `TestChain_AllowPath_StampsAttributionAndRecordsCounter`
 | `llm_router` | `{providers: [{id, models, upstream_scheme, upstream_host, upstream_path?, auth_header_name, auth_header_value, allowed_group_ids}]}` |
 | `llm_limit_check` | `{}` — pulls `MgmtClient` from `FactoryContext` |
 | `llm_identity_inject` | `{providers: [{provider_id, header_pair?|json_metadata?, extra_headers?}]}` |
-| `llm_guardrail` | `{model_allowlist: []string, prompt_capture: {enabled, redact_pii}}` |
+| `llm_guardrail` | `{provider_allowlists: {providerID: []string}, prompt_capture: {enabled, redact_pii}}` — allowlist keyed by resolved provider id; a provider absent from the map is unrestricted (fail-closed backstop; authoritative per-policy/group check is management's `CheckLLMPolicyLimits`) |
 | `llm_response_parser` | `{redact_pii?, capture_completion?: *bool}` |
 | `cost_meter` | `{pricing_path?}` (basename inside data-dir; defaults `pricing.yaml`) |
 | `llm_limit_record` | `{}` — same pattern as `llm_limit_check` |

e2e/agentnetwork/guardrail_block_test.go

@@ -0,0 +1,209 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// pathRoutedGuardrailCase is one provider's self-contained scenario: its own
+// provider, its own guardrail whose allowlist holds ONLY that provider's
+// allowed model, and its own policy. Each case runs in isolation (its own
+// proxy + client), so the guardrail the proxy enforces contains exactly this
+// provider's model — never a mixed cross-provider list.
+type pathRoutedGuardrailCase struct {
+	name       string
+	catalogID  string // agent-network catalog provider id
+	wire       string // harness.WireVertex | harness.WireBedrock
+	allowEntry string // the single model id put on the guardrail allowlist
+	allowModel string // model id sent that MUST be served (200)
+	blockModel string // model id sent that MUST be denied (403 model_blocked)
+}
+
+// TestGuardrailBlocksUnselectedModel_PathRouted is the end-to-end regression
+// guard for the customer report that a model-allowlist guardrail attached to a
+// policy has no effect for PATH-ROUTED providers — where the model travels in
+// the URL, not the JSON body: Google Vertex (…/models/{model}:rawPredict) and
+// AWS Bedrock (/model/{id}/invoke).
+//
+// Each provider is tested in isolation with a guardrail allowlisting a single
+// model of its own: the allowed model (in the URL path) is served (200) and an
+// unselected model (in the URL path) is denied 403 by the guardrail
+// (llm_policy.model_blocked) before the upstream. The Vertex case mirrors the
+// customer verbatim — allow Sonnet, and the unselected model is the exact
+// claude-opus-4-6 they reported reaching the model unblocked. The Bedrock case
+// sends a region-prefixed, versioned inference-profile id so URL-path model
+// normalization is exercised too.
+//
+// The provider is catch-all (no models), so the router forwards any model and a
+// 403 can only come from the guardrail, never model_not_routable. Only the
+// upstream LLM is mocked (the vLLM nginx answers any path with 200); management
+// synth/reconcile, the proxy middleware chain (URL-path model extraction,
+// router, guardrail) and the tunnel are all real, and the guardrail denies
+// before the upstream is dialed so the mock cannot influence the block. A
+// static bearer api key is used so the router injects a static Authorization
+// header instead of minting a GCP token — the only reason path-routed providers
+// normally need live credentials — so the test runs with none and is always on.
+func TestGuardrailBlocksUnselectedModel_PathRouted(t *testing.T) {
+	cases := []pathRoutedGuardrailCase{
+		{
+			name:       "vertex",
+			catalogID:  "vertex_ai_api",
+			wire:       harness.WireVertex,
+			allowEntry: "claude-sonnet-4-5",
+			allowModel: "claude-sonnet-4-5",
+			blockModel: "claude-opus-4-6", // the customer-reported model
+		},
+		{
+			name:       "bedrock",
+			catalogID:  "bedrock_api",
+			wire:       harness.WireBedrock,
+			allowEntry: "anthropic.claude-sonnet-4-5", // normalized catalog id
+			allowModel: "us.anthropic.claude-sonnet-4-5-v1:0",
+			blockModel: "us.anthropic.claude-opus-4-8-v1:0",
+		},
+	}
+
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			runPathRoutedGuardrailCase(t, tc)
+		})
+	}
+}
+
+func runPathRoutedGuardrailCase(t *testing.T, tc pathRoutedGuardrailCase) {
+	t.Helper()
+
+	const (
+		vertexProject = "e2e-project"
+		vertexRegion  = "global"
+	)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-" + tc.name})
+	require.NoError(t, err, "create group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-guardrail-" + tc.name + "-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grp.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	// Catch-all provider (no models) so the router forwards any model; a static
+	// bearer key means the router injects a static auth header instead of minting
+	// a GCP token. Bootstraps the cluster if it isn't already.
+	staticKey := "static-e2e-token"
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:             tc.name,
+		ProviderId:       tc.catalogID,
+		UpstreamUrl:      vllm.URL,
+		ApiKey:           &staticKey,
+		Enabled:          ptr(true),
+		BootstrapCluster: ptr(harness.AgentNetworkCluster),
+	})
+	require.NoError(t, err, "create %s provider", tc.name)
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	// Guardrail allowlisting ONLY this provider's allowed model.
+	var gr api.AgentNetworkGuardrailRequest
+	gr.Name = "e2e-guardrail-" + tc.name
+	gr.Checks.ModelAllowlist.Enabled = true
+	gr.Checks.ModelAllowlist.Models = []string{tc.allowEntry}
+	guard, err := srv.CreateGuardrail(ctx, gr)
+	require.NoError(t, err, "create guardrail")
+	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), guard.Id) })
+
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-" + tc.name,
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: []string{prov.Id},
+		GuardrailIds:           &[]string{guard.Id},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-guardrail-"+tc.name+"-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Probe first: the GET resolves the endpoint and its first packet wakes the
+	// lazy proxy peer, so WaitProxyPeer then observes it connected.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+
+	send := func(model string) (int, string) {
+		var code int
+		var body string
+		var cerr error
+		switch tc.wire {
+		case harness.WireVertex:
+			code, body, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, vertexProject, vertexRegion, model, "Reply with exactly: pong", "")
+		case harness.WireBedrock:
+			code, body, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, model, "Reply with exactly: pong", "")
+		default:
+			t.Fatalf("unsupported wire %q", tc.wire)
+		}
+		require.NoError(t, cerr, "request must reach the proxy for %s", tc.name)
+		return code, body
+	}
+
+	// Allowed model (in the URL path) is served. Retry to absorb tunnel/DNS
+	// jitter on the first call over the freshly warmed tunnel.
+	var code int
+	var body string
+	deadline := time.Now().Add(90 * time.Second)
+	for time.Now().Before(deadline) {
+		code, body = send(tc.allowModel)
+		if code == 200 {
+			break
+		}
+		time.Sleep(5 * time.Second)
+	}
+	assert.Equal(t, 200, code,
+		"allowed %s model (URL path) must be served; body: %s\n=== proxy logs ===\n%s", tc.name, body, px.Logs(context.Background()))
+
+	// Unselected model (in the URL path) must be blocked by the guardrail.
+	code, body = send(tc.blockModel)
+	assert.Equal(t, 403, code,
+		"unselected %s model (URL path) must be denied, not served; body: %s\n=== proxy logs ===\n%s", tc.name, body, px.Logs(context.Background()))
+	assert.Contains(t, body, "llm_policy.model_blocked",
+		"%s denial must come from the guardrail allowlist, not routing; body: %s", tc.name, body)
+}

e2e/agentnetwork/guardrail_groupswitch_test.go

@@ -0,0 +1,205 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestGuardrailGroupSwitchTakesEffectAfterTTL proves that moving a peer between
+// groups flips its model-allowlist decision once the proxy's tunnel-peer cache
+// expires. The peer's groups reach the guardrail via ValidateTunnelPeer, which
+// the proxy caches; the switch is invisible until that cache expires. The proxy
+// runs with a short NB_PROXY_TUNNEL_CACHE_TTL so the flip happens in seconds
+// instead of the 5-minute default.
+//
+// Setup: one catch-all provider declaring modelA + modelB; polA (grpA -> allow
+// modelA) and polB (grpB -> allow modelB). The client starts in grpA. modelA is
+// served and modelB denied; after switching the client grpA -> grpB, modelB is
+// served and modelA denied. The cross-group deny comes from management's
+// per-policy/group CheckLLMPolicyLimits (the proxy backstop carries the union).
+func TestGuardrailGroupSwitchTakesEffectAfterTTL(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	const (
+		modelA = "e2e-model-a"
+		modelB = "e2e-model-b"
+		// Short tunnel-cache TTL so a group switch propagates in seconds.
+		// Exercises the NB_PROXY_TUNNEL_CACHE_TTL override.
+		cacheTTL = 3 * time.Second
+	)
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	grpA, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-gswitch-a"})
+	require.NoError(t, err, "create group A")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpA.Id) })
+
+	grpB, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-gswitch-b"})
+	require.NoError(t, err, "create group B")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpB.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-gswitch-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grpA.Id}, // client starts in group A
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	staticKey := "static-e2e-token"
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "gswitch",
+		ProviderId:  "openai_api",
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &staticKey,
+		Enabled:     ptr(true),
+		Models: &[]api.AgentNetworkProviderModel{
+			{Id: modelA, InputPer1k: 0.001, OutputPer1k: 0.001},
+			{Id: modelB, InputPer1k: 0.001, OutputPer1k: 0.001},
+		},
+		BootstrapCluster: ptr(harness.AgentNetworkCluster),
+	})
+	require.NoError(t, err, "create provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	mkGuard := func(name, model string) api.AgentNetworkGuardrail {
+		var gr api.AgentNetworkGuardrailRequest
+		gr.Name = name
+		gr.Checks.ModelAllowlist.Enabled = true
+		gr.Checks.ModelAllowlist.Models = []string{model}
+		g, gerr := srv.CreateGuardrail(ctx, gr)
+		require.NoError(t, gerr, "create guardrail %s", name)
+		t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
+		return g
+	}
+	gA := mkGuard("e2e-gswitch-a", modelA)
+	gB := mkGuard("e2e-gswitch-b", modelB)
+
+	enabled := true
+	polA, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-gswitch-a",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpA.Id},
+		DestinationProviderIds: []string{prov.Id},
+		GuardrailIds:           &[]string{gA.Id},
+	})
+	require.NoError(t, err, "create policy A")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polA.Id) })
+
+	polB, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-gswitch-b",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpB.Id},
+		DestinationProviderIds: []string{prov.Id},
+		GuardrailIds:           &[]string{gB.Id},
+	})
+	require.NoError(t, err, "create policy B")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polB.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-gswitch-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken, map[string]string{
+		"NB_PROXY_TUNNEL_CACHE_TTL": cacheTTL.String(),
+	})
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+
+	send := func(model string) (int, string) {
+		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
+		require.NoError(t, cerr, "request must reach the proxy")
+		return code, body
+	}
+	sendUntil := func(model string, want int, timeout time.Duration) (int, string) {
+		var code int
+		var body string
+		deadline := time.Now().Add(timeout)
+		for time.Now().Before(deadline) {
+			code, body = send(model)
+			if code == want {
+				return code, body
+			}
+			time.Sleep(2 * time.Second)
+		}
+		return code, body
+	}
+
+	// Phase 1 — client is in group A: modelA served, modelB denied.
+	code, body := sendUntil(modelA, 200, 90*time.Second)
+	assert.Equal(t, 200, code,
+		"group-A model must be served while the client is in group A; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	code, body = send(modelB)
+	assert.Equal(t, 403, code,
+		"group-B model must be denied while the client is in group A; body: %s", body)
+	assert.Contains(t, body, "llm_policy.model_blocked")
+
+	// Switch the client peer from group A to group B.
+	peerID := clientPeerInGroup(t, ctx, grpA.Id)
+	_, err = srv.API().Groups.Update(ctx, grpB.Id, api.PutApiGroupsGroupIdJSONRequestBody{
+		Name:  grpB.Name,
+		Peers: &[]string{peerID},
+	})
+	require.NoError(t, err, "add peer to group B")
+	_, err = srv.API().Groups.Update(ctx, grpA.Id, api.PutApiGroupsGroupIdJSONRequestBody{
+		Name:  grpA.Name,
+		Peers: &[]string{},
+	})
+	require.NoError(t, err, "remove peer from group A")
+
+	// Phase 2 — after the short TTL expires the proxy re-validates the peer,
+	// sees group B, and the decision flips. Poll to absorb TTL + re-validation.
+	code, body = sendUntil(modelB, 200, 60*time.Second)
+	assert.Equal(t, 200, code,
+		"after the group switch + TTL, the group-B model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	code, body = send(modelA)
+	assert.Equal(t, 403, code,
+		"after the switch, the old group-A model must be denied; body: %s", body)
+	assert.Contains(t, body, "llm_policy.model_blocked")
+}
+
+// clientPeerInGroup returns the id of the single peer that is a member of the
+// given group — the test client. The proxy peer is never added to test groups.
+func clientPeerInGroup(t *testing.T, ctx context.Context, groupID string) string {
+	t.Helper()
+	peers, err := srv.API().Peers.List(ctx)
+	require.NoError(t, err, "list peers")
+	for _, p := range peers {
+		for _, g := range p.Groups {
+			if g.Id == groupID {
+				return p.Id
+			}
+		}
+	}
+	t.Fatalf("no peer found in group %s", groupID)
+	return ""
+}

e2e/agentnetwork/guardrail_multipolicy_test.go

@@ -0,0 +1,201 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestGuardrailMultiPolicyModelAllowlist: modelSelected served (200), grpOther's
+// modelOther denied for the grpMain client (403 model_blocked, no cross-group
+// leak), and openModel on the un-guardrailed policy's provider served (200).
+func TestGuardrailMultiPolicyModelAllowlist(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	const (
+		modelSelected = "e2e-selected"
+		modelOther    = "e2e-other"
+		openModel     = "e2e-open"
+	)
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	grpMain, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-mp-main"})
+	require.NoError(t, err, "create main group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpMain.Id) })
+
+	grpOther, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-mp-other"})
+	require.NoError(t, err, "create other group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpOther.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-guardrail-mp-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grpMain.Id}, // client joins grpMain only
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	staticKey := "static-e2e-token"
+	models := func(ids ...string) *[]api.AgentNetworkProviderModel {
+		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
+		for _, id := range ids {
+			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
+		}
+		return &out
+	}
+
+	// pRestricted declares the two guardrailed models so routing is deterministic
+	// (model -> provider). Created first, so it carries the bootstrap cluster.
+	pRestricted, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:             "restricted",
+		ProviderId:       "openai_api",
+		UpstreamUrl:      vllm.URL,
+		ApiKey:           &staticKey,
+		Enabled:          ptr(true),
+		Models:           models(modelSelected, modelOther),
+		BootstrapCluster: ptr(harness.AgentNetworkCluster),
+	})
+	require.NoError(t, err, "create restricted provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), pRestricted.Id) })
+
+	pOpen, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "open",
+		ProviderId:  "openai_api",
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &staticKey,
+		Enabled:     ptr(true),
+		Models:      models(openModel),
+	})
+	require.NoError(t, err, "create open provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), pOpen.Id) })
+
+	mkGuardrail := func(name, model string) api.AgentNetworkGuardrail {
+		var gr api.AgentNetworkGuardrailRequest
+		gr.Name = name
+		gr.Checks.ModelAllowlist.Enabled = true
+		gr.Checks.ModelAllowlist.Models = []string{model}
+		g, gerr := srv.CreateGuardrail(ctx, gr)
+		require.NoError(t, gerr, "create guardrail %s", name)
+		t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
+		return g
+	}
+	gMain := mkGuardrail("e2e-guardrail-mp-main", modelSelected)
+	gOther := mkGuardrail("e2e-guardrail-mp-other", modelOther)
+
+	enabled := true
+	// polMain: grpMain restricted to modelSelected on pRestricted.
+	polMain, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-mp-main",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpMain.Id},
+		DestinationProviderIds: []string{pRestricted.Id},
+		GuardrailIds:           &[]string{gMain.Id},
+	})
+	require.NoError(t, err, "create main policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMain.Id) })
+
+	// polOther: grpOther restricted to modelOther on the SAME provider. The
+	// client is not in grpOther, so modelOther must never be usable by it.
+	polOther, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-mp-other",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpOther.Id},
+		DestinationProviderIds: []string{pRestricted.Id},
+		GuardrailIds:           &[]string{gOther.Id},
+	})
+	require.NoError(t, err, "create other policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOther.Id) })
+
+	// polOpen: grpMain on pOpen with NO guardrail — unrestricted.
+	polOpen, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-mp-open",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpMain.Id},
+		DestinationProviderIds: []string{pOpen.Id},
+	})
+	require.NoError(t, err, "create open policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOpen.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-guardrail-mp-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+
+	send := func(model string) (int, string) {
+		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
+		require.NoError(t, cerr, "request must reach the proxy")
+		return code, body
+	}
+	// sendUntil200 absorbs first-call tunnel/DNS jitter on the freshly warmed tunnel.
+	sendUntil200 := func(model string) (int, string) {
+		var code int
+		var body string
+		deadline := time.Now().Add(90 * time.Second)
+		for time.Now().Before(deadline) {
+			code, body = send(model)
+			if code == 200 {
+				break
+			}
+			time.Sleep(5 * time.Second)
+		}
+		return code, body
+	}
+
+	t.Run("selected model allowed for its group", func(t *testing.T) {
+		code, body := sendUntil200(modelSelected)
+		assert.Equal(t, 200, code,
+			"grpMain's allowlisted model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	})
+
+	t.Run("other group's model does not leak", func(t *testing.T) {
+		// modelOther is allowlisted only for grpOther. The grpMain client must be
+		// denied by management's per-policy/group check — not waved through by an
+		// account-wide union. This is the security-critical wrong-ALLOW guard.
+		code, body := send(modelOther)
+		assert.Equal(t, 403, code,
+			"another group's allowlisted model must be denied for this caller; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+		assert.Contains(t, body, "llm_policy.model_blocked",
+			"denial must be a model-allowlist decision; body: %s", body)
+	})
+
+	t.Run("unguarded policy leaves its provider unrestricted", func(t *testing.T) {
+		// polOpen carries no guardrail, so pOpen is unrestricted for grpMain. The
+		// old account-wide union would have blocked openModel (it is on no
+		// allowlist); it must now be served — the false-DENY guard.
+		code, body := sendUntil200(openModel)
+		assert.Equal(t, 200, code,
+			"an un-guardrailed policy's provider must not be blocked by another policy's allowlist; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	})
+}

e2e/agentnetwork/guardrail_pergroup_providers_test.go

@@ -0,0 +1,422 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// pergroupCase describes one provider surface for the per-group allowlist matrix.
+// selectedReq/otherReq are the model identifiers as they travel in the request
+// (URL path for Bedrock/Vertex, body "model" for chat/messages). selectedAllow/
+// otherAllow are the (normalized) forms the guardrail allowlist holds — for
+// Bedrock these differ from the request form so path normalization is exercised.
+type pergroupCase struct {
+	name          string
+	catalogID     string
+	wire          string // "chat", "messages", "vertex", "bedrock"
+	models        *[]api.AgentNetworkProviderModel
+	selectedReq   string
+	selectedAllow string
+	otherReq      string
+	otherAllow    string
+
+	providerID string // filled during setup
+}
+
+// TestGuardrailPerGroupAllowlist_AllProviders proves the per-policy/group model
+// allowlist end to end across every always-on provider surface, including the
+// path-routed ones (Vertex, Bedrock) where the model travels in the URL.
+//
+// For each provider two policies target it: grpMain (the client) is allowed only
+// selectedReq; grpOther (which the client is NOT in) is allowed only otherReq.
+// The client must get selectedReq served (200) and otherReq denied (403,
+// llm_policy.model_blocked) — the cross-group no-leak property. The deny is the
+// authoritative per-policy/group decision from management (the proxy per-provider
+// backstop carries the union of both models), so this also confirms management
+// receives the correct normalized model for path-routed providers.
+func TestGuardrailPerGroupAllowlist_AllProviders(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	const (
+		vertexProject = "e2e-project"
+		vertexRegion  = "global"
+	)
+
+	priced := func(ids ...string) *[]api.AgentNetworkProviderModel {
+		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
+		for _, id := range ids {
+			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
+		}
+		return &out
+	}
+
+	cases := []*pergroupCase{
+		{
+			name: "openai", catalogID: "openai_api", wire: harness.WireChat,
+			models:      priced("oai-model-a", "oai-model-b"),
+			selectedReq: "oai-model-a", selectedAllow: "oai-model-a",
+			otherReq: "oai-model-b", otherAllow: "oai-model-b",
+		},
+		{
+			name: "anthropic", catalogID: "anthropic_api", wire: harness.WireMessages,
+			models:      priced("ant-model-a", "ant-model-b"),
+			selectedReq: "ant-model-a", selectedAllow: "ant-model-a",
+			otherReq: "ant-model-b", otherAllow: "ant-model-b",
+		},
+		{
+			// Vertex catalog ids travel bare in the rawPredict path.
+			name: "vertex", catalogID: "vertex_ai_api", wire: "vertex",
+			selectedReq: "claude-sonnet-4-5", selectedAllow: "claude-sonnet-4-5",
+			otherReq: "claude-opus-4-6", otherAllow: "claude-opus-4-6",
+		},
+		{
+			// Bedrock request ids are region-prefixed/versioned; the parser
+			// normalizes them to the catalog key the allowlist holds.
+			name: "bedrock", catalogID: "bedrock_api", wire: "bedrock",
+			selectedReq: "us.anthropic.claude-sonnet-4-5-v1:0", selectedAllow: "anthropic.claude-sonnet-4-5",
+			otherReq: "us.anthropic.claude-opus-4-8-v1:0", otherAllow: "anthropic.claude-opus-4-8",
+		},
+	}
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	grpMain, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-pergroup-main"})
+	require.NoError(t, err, "create main group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpMain.Id) })
+
+	grpOther, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-pergroup-other"})
+	require.NoError(t, err, "create other group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpOther.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-pergroup-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grpMain.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	staticKey := "static-e2e-token"
+	enabled := true
+
+	for i, c := range cases {
+		req := api.AgentNetworkProviderRequest{
+			Name:        "e2e-pergroup-" + c.name,
+			ProviderId:  c.catalogID,
+			UpstreamUrl: vllm.URL,
+			ApiKey:      &staticKey,
+			Enabled:     ptr(true),
+			Models:      c.models,
+		}
+		if i == 0 {
+			req.BootstrapCluster = ptr(harness.AgentNetworkCluster)
+		}
+		prov, perr := srv.CreateProvider(ctx, req)
+		require.NoError(t, perr, "create provider %s", c.name)
+		c.providerID = prov.Id
+		t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+		gSel := mkAllowGuardrail(t, ctx, "e2e-pergroup-"+c.name+"-sel", c.selectedAllow)
+		gOth := mkAllowGuardrail(t, ctx, "e2e-pergroup-"+c.name+"-oth", c.otherAllow)
+
+		polMain, merr := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+			Name:                   "e2e-pergroup-" + c.name + "-main",
+			Enabled:                &enabled,
+			SourceGroups:           []string{grpMain.Id},
+			DestinationProviderIds: []string{prov.Id},
+			GuardrailIds:           &[]string{gSel.Id},
+		})
+		require.NoError(t, merr, "create main policy %s", c.name)
+		t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMain.Id) })
+
+		polOther, oerr := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+			Name:                   "e2e-pergroup-" + c.name + "-other",
+			Enabled:                &enabled,
+			SourceGroups:           []string{grpOther.Id},
+			DestinationProviderIds: []string{prov.Id},
+			GuardrailIds:           &[]string{gOth.Id},
+		})
+		require.NoError(t, oerr, "create other policy %s", c.name)
+		t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOther.Id) })
+	}
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-pergroup-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+
+	send := func(c *pergroupCase, model string) (int, string) {
+		var code int
+		var body string
+		var cerr error
+		switch c.wire {
+		case "vertex":
+			code, body, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, vertexProject, vertexRegion, model, "Reply with exactly: pong", "")
+		case "bedrock":
+			code, body, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, model, "Reply with exactly: pong", "")
+		default:
+			code, body, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, c.wire, model, "Reply with exactly: pong", "")
+		}
+		require.NoError(t, cerr, "request must reach the proxy for %s", c.name)
+		return code, body
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			// grpMain's own model is served. Retry to absorb tunnel/DNS jitter on
+			// the first call over the freshly warmed tunnel.
+			var code int
+			var body string
+			deadline := time.Now().Add(90 * time.Second)
+			for time.Now().Before(deadline) {
+				code, body = send(c, c.selectedReq)
+				if code == 200 {
+					break
+				}
+				time.Sleep(5 * time.Second)
+			}
+			assert.Equal(t, 200, code,
+				"%s: grpMain's allowlisted model must be served; body: %s\n=== proxy logs ===\n%s", c.name, body, px.Logs(context.Background()))
+
+			// grpOther's model must NOT leak to the grpMain client.
+			code, body = send(c, c.otherReq)
+			assert.Equal(t, 403, code,
+				"%s: another group's allowlisted model must be denied for this caller; body: %s\n=== proxy logs ===\n%s", c.name, body, px.Logs(context.Background()))
+			assert.Contains(t, body, "llm_policy.model_blocked",
+				"%s: denial must be a model-allowlist decision, not routing; body: %s", c.name, body)
+		})
+	}
+}
+
+// TestGuardrailMultiGroupUser proves the per-policy/group decision for a caller
+// that belongs to MULTIPLE groups at once. Two scenarios, one shared stack:
+//
+//   - union across the user's groups: the client is in gUX and gUY, each with
+//     its own policy+guardrail on provider P1 (gUX->union-a, gUY->union-b). The
+//     client may use BOTH models (the union of its groups' allowlists) while a
+//     third, un-allowlisted model is denied.
+//   - an un-guardrailed group lifts the restriction: the client is in gMP and
+//     gMQ on provider P2, where gMP restricts to mix-a but gMQ's policy carries
+//     NO guardrail. Because one applicable policy is unrestricted, the client may
+//     use a model on no allowlist (mix-z) as well as mix-a.
+func TestGuardrailMultiGroupUser(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	const (
+		unionA = "mg-union-a"
+		unionB = "mg-union-b"
+		unionC = "mg-union-c" // allowlisted by neither group
+		mixA   = "mg-mix-a"
+		mixZ   = "mg-mix-z" // on no allowlist; reachable only via the un-guardrailed policy
+	)
+
+	priced := func(ids ...string) *[]api.AgentNetworkProviderModel {
+		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
+		for _, id := range ids {
+			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
+		}
+		return &out
+	}
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	mkGroup := func(name string) *api.Group {
+		g, gerr := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: name})
+		require.NoError(t, gerr, "create group %s", name)
+		t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), g.Id) })
+		return g
+	}
+	gUX := mkGroup("e2e-mg-union-x")
+	gUY := mkGroup("e2e-mg-union-y")
+	gMP := mkGroup("e2e-mg-mix-p")
+	gMQ := mkGroup("e2e-mg-mix-q")
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-mg-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{gUX.Id, gUY.Id, gMP.Id, gMQ.Id}, // client in all four groups
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	staticKey := "static-e2e-token"
+	enabled := true
+
+	// P1 — union scenario: two restricting policies, one per group.
+	p1, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:             "e2e-mg-union",
+		ProviderId:       "openai_api",
+		UpstreamUrl:      vllm.URL,
+		ApiKey:           &staticKey,
+		Enabled:          ptr(true),
+		Models:           priced(unionA, unionB, unionC),
+		BootstrapCluster: ptr(harness.AgentNetworkCluster),
+	})
+	require.NoError(t, err, "create union provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), p1.Id) })
+
+	polUX, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-mg-union-x",
+		Enabled:                &enabled,
+		SourceGroups:           []string{gUX.Id},
+		DestinationProviderIds: []string{p1.Id},
+		GuardrailIds:           &[]string{mkAllowGuardrail(t, ctx, "e2e-mg-union-x", unionA).Id},
+	})
+	require.NoError(t, err, "create union policy X")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polUX.Id) })
+
+	polUY, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-mg-union-y",
+		Enabled:                &enabled,
+		SourceGroups:           []string{gUY.Id},
+		DestinationProviderIds: []string{p1.Id},
+		GuardrailIds:           &[]string{mkAllowGuardrail(t, ctx, "e2e-mg-union-y", unionB).Id},
+	})
+	require.NoError(t, err, "create union policy Y")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polUY.Id) })
+
+	// P2 — mixed scenario: one restricting policy + one un-guardrailed policy.
+	p2, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "e2e-mg-mix",
+		ProviderId:  "openai_api",
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &staticKey,
+		Enabled:     ptr(true),
+		Models:      priced(mixA, mixZ),
+	})
+	require.NoError(t, err, "create mix provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), p2.Id) })
+
+	polMP, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-mg-mix-p",
+		Enabled:                &enabled,
+		SourceGroups:           []string{gMP.Id},
+		DestinationProviderIds: []string{p2.Id},
+		GuardrailIds:           &[]string{mkAllowGuardrail(t, ctx, "e2e-mg-mix-p", mixA).Id},
+	})
+	require.NoError(t, err, "create mix policy P")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMP.Id) })
+
+	polMQ, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-mg-mix-q",
+		Enabled:                &enabled,
+		SourceGroups:           []string{gMQ.Id},
+		DestinationProviderIds: []string{p2.Id}, // NO guardrail -> unrestricted
+	})
+	require.NoError(t, err, "create mix policy Q")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMQ.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings")
+	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-mg-proxy")
+	require.NoError(t, err, "mint proxy token")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+
+	send := func(model string) (int, string) {
+		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
+		require.NoError(t, cerr, "request must reach the proxy")
+		return code, body
+	}
+	sendUntil200 := func(model string) (int, string) {
+		var code int
+		var body string
+		deadline := time.Now().Add(90 * time.Second)
+		for time.Now().Before(deadline) {
+			code, body = send(model)
+			if code == 200 {
+				break
+			}
+			time.Sleep(5 * time.Second)
+		}
+		return code, body
+	}
+
+	t.Run("union across the user's groups", func(t *testing.T) {
+		code, body := sendUntil200(unionA)
+		assert.Equal(t, 200, code, "model allowed by group X must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+
+		code, body = sendUntil200(unionB)
+		assert.Equal(t, 200, code, "model allowed by group Y must also be served (union across the user's groups); body: %s", body)
+
+		code, body = send(unionC)
+		assert.Equal(t, 403, code, "a model on neither group's allowlist must be denied; body: %s", body)
+		assert.Contains(t, body, "llm_policy.model_blocked")
+	})
+
+	t.Run("an un-guardrailed group lifts the restriction", func(t *testing.T) {
+		code, body := sendUntil200(mixA)
+		assert.Equal(t, 200, code, "the restricted group's model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+
+		code, body = sendUntil200(mixZ)
+		assert.Equal(t, 200, code,
+			"a non-allowlisted model must be served because the user is also in a group whose policy has no guardrail; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	})
+}
+
+// mkAllowGuardrail creates a guardrail whose model allowlist is enabled and holds
+// exactly the given model, registering cleanup.
+func mkAllowGuardrail(t *testing.T, ctx context.Context, name, model string) api.AgentNetworkGuardrail {
+	t.Helper()
+	var gr api.AgentNetworkGuardrailRequest
+	gr.Name = name
+	gr.Checks.ModelAllowlist.Enabled = true
+	gr.Checks.ModelAllowlist.Models = []string{model}
+	g, err := srv.CreateGuardrail(ctx, gr)
+	require.NoError(t, err, "create guardrail %s", name)
+	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
+	return g
+}

e2e/harness/proxy.go

@@ -38,7 +38,11 @@ type Proxy struct {
 // network, registered via the given account proxy token and serving the
 // AgentNetworkCluster over a self-signed wildcard cert. It does not wait for
 // peer connectivity — callers poll management for the proxy peer.
-func StartProxy(ctx context.Context, c *Combined, proxyToken string) (*Proxy, error) {
+// StartProxy launches the reverse-proxy container. Optional envOverrides are
+// merged into the container environment after the defaults, so callers can set
+// or override any NB_PROXY_* var (e.g. NB_PROXY_TUNNEL_CACHE_TTL for tests that
+// need a short authorization-cache window).
+func StartProxy(ctx context.Context, c *Combined, proxyToken string, envOverrides ...map[string]string) (*Proxy, error) {
 	root, err := repoRoot()
 	if err != nil {
 		return nil, err
@@ -93,6 +97,12 @@ func StartProxy(ctx context.Context, c *Combined, proxyToken string) (*Proxy, er
 		WaitingFor: wait.ForLog("Initial mapping sync complete").WithStartupTimeout(90 * time.Second),
 	}
 
+	for _, ov := range envOverrides {
+		for k, v := range ov {
+			req.Env[k] = v
+		}
+	}
+
 	ctr, err := testcontainers.GenericContainer(ctx, testcontainers.GenericContainerRequest{
 		ContainerRequest: req,
 		Started:          true,

management/internals/modules/agentnetwork/manager.go

@@ -86,12 +86,17 @@ type Manager interface {
 
 // PolicySelectionInput is the per-request selection envelope. The
 // proxy populates it from CapturedData (account, user, groups) plus
-// the provider llm_router resolved.
+// the provider llm_router resolved and the model it extracted.
 type PolicySelectionInput struct {
 	AccountID  string
 	UserID     string
 	GroupIDs   []string
 	ProviderID string
+	// Model is the already-normalised upstream model id the proxy extracted
+	// (parser strips Bedrock region/version, Vertex @version), so a
+	// case-insensitive compare suffices. Empty = undetermined → not permitted
+	// (fail closed).
+	Model string
 }
 
 // PolicySelectionResult names the policy that "pays" for this request

management/internals/modules/agentnetwork/policyselect.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"math"
 	"sort"
+	"strings"
 	"time"
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
@@ -35,6 +36,10 @@ const (
 	denyCodeAccountTokenCapExceeded = "llm_account.token_cap_exceeded"
 	//nolint:gosec // account deny code label, not a credential
 	denyCodeAccountBudgetCapExceeded = "llm_account.budget_cap_exceeded"
+	// denyCodeModelBlocked is returned when policies govern the request's
+	// (provider, caller-groups) but none permits the model. Matches the proxy
+	// guardrail's code so both layers surface the same label.
+	denyCodeModelBlocked = "llm_policy.model_blocked"
 )
 
 // consumptionCache holds the consumption counters prefetched for one
@@ -159,6 +164,25 @@ func (m *managerImpl) SelectPolicyForRequest(ctx context.Context, in PolicySelec
 	}
 	candidates := filterApplicablePolicies(policies, in)
 
+	// Model-allowlist gate scoped to the matched policies: keep candidates whose
+	// guardrails permit the model (none enabled = unrestricted), deny when
+	// policies apply but none permits it. Skip the load when none has a guardrail.
+	if len(candidates) > 0 && anyPolicyHasGuardrails(candidates) {
+		guardrailsByID, gErr := m.loadGuardrailsByID(ctx, in.AccountID)
+		if gErr != nil {
+			return nil, gErr
+		}
+		permitted := filterModelPermittedPolicies(candidates, guardrailsByID, in.Model)
+		if len(permitted) == 0 {
+			return &PolicySelectionResult{
+				Allow:      false,
+				DenyCode:   denyCodeModelBlocked,
+				DenyReason: modelBlockedReason(in.Model),
+			}, nil
+		}
+		candidates = permitted
+	}
+
 	// Prefetch every consumption counter the ceiling + candidate policies will
 	// read, in a single store round-trip, then score against the cache.
 	cache, err := m.prefetchConsumption(ctx, in, rules, candidates, now)
@@ -250,6 +274,90 @@ func filterApplicablePolicies(policies []*types.Policy, in PolicySelectionInput)
 	return out
 }
 
+// anyPolicyHasGuardrails reports whether any policy references at least one
+// guardrail, so the selector can skip loading guardrails when none do.
+func anyPolicyHasGuardrails(policies []*types.Policy) bool {
+	for _, p := range policies {
+		if p != nil && len(p.GuardrailIDs) > 0 {
+			return true
+		}
+	}
+	return false
+}
+
+// loadGuardrailsByID loads the account's guardrails indexed by ID. Used by the
+// model-allowlist gate to resolve each candidate policy's attached guardrails.
+func (m *managerImpl) loadGuardrailsByID(ctx context.Context, accountID string) (map[string]*types.Guardrail, error) {
+	guardrails, err := m.store.GetAccountAgentNetworkGuardrails(ctx, store.LockingStrengthNone, accountID)
+	if err != nil {
+		return nil, fmt.Errorf("list account guardrails: %w", err)
+	}
+	byID := make(map[string]*types.Guardrail, len(guardrails))
+	for _, g := range guardrails {
+		if g != nil {
+			byID[g.ID] = g
+		}
+	}
+	return byID, nil
+}
+
+// filterModelPermittedPolicies returns the subset of policies whose guardrails
+// permit the model. Order is preserved so downstream scoring is unaffected.
+func filterModelPermittedPolicies(policies []*types.Policy, byID map[string]*types.Guardrail, model string) []*types.Policy {
+	out := make([]*types.Policy, 0, len(policies))
+	for _, p := range policies {
+		if policyPermitsModel(p, byID, model) {
+			out = append(out, p)
+		}
+	}
+	return out
+}
+
+// policyPermitsModel reports whether a policy permits the model. No
+// allowlist-enabled guardrail = unrestricted (permits any, incl. empty);
+// otherwise the model must be in the union of its allowlists, so an
+// empty/undetermined model fails closed.
+func policyPermitsModel(p *types.Policy, byID map[string]*types.Guardrail, model string) bool {
+	if p == nil {
+		return false
+	}
+	wanted := normaliseModelID(model)
+	restricted := false
+	for _, gID := range p.GuardrailIDs {
+		g, ok := byID[gID]
+		if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
+			continue
+		}
+		restricted = true
+		if wanted == "" {
+			continue
+		}
+		for _, allowed := range g.Checks.ModelAllowlist.Models {
+			if normaliseModelID(allowed) == wanted {
+				return true
+			}
+		}
+	}
+	return !restricted
+}
+
+// normaliseModelID lowercases and trims a model identifier so the allowlist
+// compare is case-insensitive and trim-tolerant. Mirrors the proxy guardrail's
+// normaliseModel so both layers agree on what "same model" means.
+func normaliseModelID(model string) string {
+	return strings.ToLower(strings.TrimSpace(model))
+}
+
+// modelBlockedReason builds the human-readable deny reason for a model-allowlist
+// rejection. The model is quoted when known; an undetermined model is reported
+// as such so the access log distinguishes "wrong model" from "no model".
+func modelBlockedReason(model string) string {
+	if normaliseModelID(model) == "" {
+		return "request model could not be determined for the policy allowlist"
+	}
+	return fmt.Sprintf("model %q is not permitted by any applicable policy allowlist", model)
+}
+
 // candidate is the per-policy intermediate the selector ranks. A
 // policy that's been exhausted on any enabled cap never makes it
 // into this slice; the selector's deny envelope carries the latest

management/internals/modules/agentnetwork/policyselect_model_test.go

@@ -0,0 +1,329 @@
+package agentnetwork
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"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/store"
+)
+
+// guardedPolicy builds an enabled, uncapped policy that authorises sourceGroups
+// to reach providerID under the given guardrails. Uncapped keeps the selector's
+// headroom scoring trivial so these tests isolate the model-allowlist gate.
+func guardedPolicy(id, account string, sourceGroups []string, providerID string, guardrailIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		AccountID:              account,
+		Enabled:                true,
+		SourceGroups:           sourceGroups,
+		DestinationProviderIDs: []string{providerID},
+		GuardrailIDs:           guardrailIDs,
+		CreatedAt:              time.Now().UTC(),
+	}
+}
+
+// allowlistGuardrail builds a guardrail whose model allowlist is enabled and
+// carries the given models.
+func allowlistGuardrail(id, account string, models ...string) *types.Guardrail {
+	return &types.Guardrail{
+		ID:        id,
+		AccountID: account,
+		Checks: types.GuardrailChecks{
+			ModelAllowlist: types.GuardrailModelAllowlist{Enabled: true, Models: models},
+		},
+	}
+}
+
+func expectPolicies(mockStore *store.MockStore, account string, policies ...*types.Policy) {
+	mockStore.EXPECT().
+		GetAccountAgentNetworkPolicies(gomock.Any(), gomock.Any(), account).
+		Return(policies, nil)
+}
+
+func expectGuardrails(mockStore *store.MockStore, account string, guardrails ...*types.Guardrail) {
+	mockStore.EXPECT().
+		GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), account).
+		Return(guardrails, nil)
+}
+
+// TestSelectPolicy_ModelBlockedByAllowlist proves the authoritative allowlist
+// decision: a policy authorises the (provider, group) but restricts the model,
+// and the requested model isn't on the list, so the request is denied.
+func TestSelectPolicy_ModelBlockedByAllowlist(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		UserID:     "user-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "a model outside the only applicable policy's allowlist must be denied")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode, "deny code must be model_blocked")
+	assert.NotEmpty(t, res.DenyReason, "deny reason must be populated")
+}
+
+// TestSelectPolicy_ModelAllowedByAllowlist is the allow counterpart: the model
+// is on the applicable policy's allowlist, so selection proceeds normally.
+func TestSelectPolicy_ModelAllowedByAllowlist(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o", "claude-opus-4"))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		UserID:     "user-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a model on the applicable policy's allowlist must be allowed")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_CaseInsensitiveModelMatch proves the compare tolerates case
+// and surrounding whitespace, matching the proxy guardrail's normalisation.
+func TestSelectPolicy_CaseInsensitiveModelMatch(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "  GPT-4o  "))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "case/whitespace variants must match the allowlist entry")
+}
+
+// TestSelectPolicy_UnguardedPolicyIsUnrestricted is the false-deny fix: when two
+// policies authorise the same (provider, group) and one has no guardrail, that
+// policy makes the request unrestricted — not caught by the other's allowlist.
+func TestSelectPolicy_UnguardedPolicyIsUnrestricted(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	restricted := guardedPolicy("pol-restricted", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	open := guardedPolicy("pol-open", "acc-1", []string{"grp-eng"}, "prov-1") // no guardrail
+	expectPolicies(mockStore, "acc-1", restricted, open)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "an un-guardrailed policy for the same (provider, group) must leave the request unrestricted")
+	assert.Equal(t, "pol-open", res.SelectedPolicyID, "the unrestricted policy must be the one that pays")
+}
+
+// TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups is the false-allow fix: a
+// model allowlisted only for grp-b must not be usable by a grp-a caller. The
+// selector considers only policies applicable to the caller's groups.
+func TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	polA := guardedPolicy("pol-a", "acc-1", []string{"grp-a"}, "prov-1", "g-a")
+	polB := guardedPolicy("pol-b", "acc-1", []string{"grp-b"}, "prov-1", "g-b")
+	expectPolicies(mockStore, "acc-1", polA, polB)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-a", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-b", "acc-1", "claude-opus-4"),
+	)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-a"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only allowed for grp-b
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "grp-b's allowlisted model must not leak to a grp-a caller")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
+}
+
+// TestSelectPolicy_UndeterminedModelFailsClosed proves the fail-closed contract
+// mirrors the proxy: with a restricted applicable policy and an empty model
+// (e.g. a path-routed shape the parser couldn't map), the request is denied.
+func TestSelectPolicy_UndeterminedModelFailsClosed(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "", // undetermined
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "an undetermined model must fail closed against a restricted policy")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
+}
+
+// TestSelectPolicy_DisabledAllowlistDoesNotRestrict proves a guardrail whose
+// model allowlist is disabled imposes no model restriction, even though the
+// policy references it.
+func TestSelectPolicy_DisabledAllowlistDoesNotRestrict(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	disabled := &types.Guardrail{
+		ID:        "g-1",
+		AccountID: "acc-1",
+		Checks: types.GuardrailChecks{
+			ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}},
+		},
+	}
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", disabled)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "anything-goes",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a disabled allowlist must not restrict the model")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_UnionAcrossPolicyGuardrails proves a policy with multiple
+// allowlist guardrails permits the union of their models (not just the first).
+func TestSelectPolicy_UnionAcrossPolicyGuardrails(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1", "g-2")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-1", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-2", "acc-1", "claude-opus-4"),
+	)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only in the second guardrail's list
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a model in any of the policy's allowlist guardrails must be permitted")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_GuardrailLookupErrorPropagates proves a store failure while
+// resolving the candidate policies' guardrails surfaces as an error, not a
+// silent allow/deny.
+func TestSelectPolicy_GuardrailLookupErrorPropagates(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	mockStore.EXPECT().
+		GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), "acc-1").
+		Return(nil, errors.New("store unavailable"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.Error(t, err, "a guardrail-lookup failure must surface as an error")
+	assert.Nil(t, res)
+}
+
+// TestSelectPolicy_MissingGuardrailReferenceTreatedAsUnrestricted proves a
+// policy referencing a guardrail ID absent from the account's set (a stale
+// reference) imposes no model restriction — same as no guardrail.
+func TestSelectPolicy_MissingGuardrailReferenceTreatedAsUnrestricted(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-missing")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1")
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "anything-goes",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "an orphaned guardrail reference must not restrict the model")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_PartialCandidatesPermittedAfterModelFilter proves the model
+// gate narrows candidates before cap scoring: the permitting policy is selected
+// even though the blocked one has a larger, more attractive cap.
+func TestSelectPolicy_PartialCandidatesPermittedAfterModelFilter(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	polBig := guardedPolicy("pol-big", "acc-1", []string{"grp-eng"}, "prov-1", "g-restrict")
+	polBig.Limits = types.PolicyLimits{
+		TokenLimit: types.PolicyTokenLimit{Enabled: true, GroupCap: 1_000_000, WindowSeconds: 3600},
+	}
+	polSmall := guardedPolicy("pol-small", "acc-1", []string{"grp-eng"}, "prov-1", "g-permit")
+	polSmall.Limits = types.PolicyLimits{
+		TokenLimit: types.PolicyTokenLimit{Enabled: true, GroupCap: 100, WindowSeconds: 3600},
+	}
+	expectPolicies(mockStore, "acc-1", polBig, polSmall)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-restrict", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-permit", "acc-1", "claude-opus-4"),
+	)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only pol-small's guardrail permits this
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow)
+	assert.Equal(t, "pol-small", res.SelectedPolicyID,
+		"the model filter must exclude pol-big before cap scoring")
+}

management/internals/modules/agentnetwork/synthesizer.go

@@ -235,7 +235,12 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 
 	mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
 	applyAccountCollectionControls(&mergedGuardrails, settings)
-	guardrailJSON, err := marshalGuardrailConfig(mergedGuardrails)
+	// The proxy guardrail is a per-provider fail-closed backstop; the
+	// authoritative per-policy/group decision is management's
+	// SelectPolicyForRequest. A provider lands in this map only when every
+	// authorising policy restricts models.
+	providerAllowlists := buildProviderAllowlists(enabledPolicies, guardrailsByID)
+	guardrailJSON, err := marshalGuardrailConfig(providerAllowlists, mergedGuardrails.PromptCapture)
 	if err != nil {
 		return nil, err
 	}
@@ -780,10 +785,12 @@ func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byt
 
 // guardrailConfig is the JSON shape the proxy-side llm_guardrail
 // middleware expects. Mirrors the proxy registration documented in
-// the management→proxy contract.
+// the management→proxy contract. provider_allowlists is keyed by the
+// resolved provider id llm_router stamps; a provider absent from the map is
+// unrestricted at the proxy layer.
 type guardrailConfig struct {
-	ModelAllowlist []string               `json:"model_allowlist,omitempty"`
-	PromptCapture  guardrailPromptCapture `json:"prompt_capture"`
+	ProviderAllowlists map[string][]string    `json:"provider_allowlists,omitempty"`
+	PromptCapture      guardrailPromptCapture `json:"prompt_capture"`
 }
 
 type guardrailPromptCapture struct {
@@ -828,13 +835,10 @@ func applyAccountCollectionControls(merged *MergedGuardrails, settings *types.Se
 	merged.PromptCapture.RedactPii = settings.RedactPii || merged.PromptCapture.RedactPii
 }
 
-func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
+func marshalGuardrailConfig(providerAllowlists map[string][]string, capture MergedPromptCapture) ([]byte, error) {
 	cfg := guardrailConfig{
-		ModelAllowlist: merged.ModelAllowlist,
-		PromptCapture: guardrailPromptCapture{
-			Enabled:   merged.PromptCapture.Enabled,
-			RedactPii: merged.PromptCapture.RedactPii,
-		},
+		ProviderAllowlists: providerAllowlists,
+		PromptCapture:      guardrailPromptCapture(capture),
 	}
 	out, err := json.Marshal(cfg)
 	if err != nil {
@@ -843,6 +847,74 @@ func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
 	return out, nil
 }
 
+// buildProviderAllowlists returns the proxy's per-provider backstop: a provider
+// is included only when every authorising policy restricts models (their union);
+// if any leaves it unrestricted it is omitted, so management decides per group.
+func buildProviderAllowlists(policies []*types.Policy, byID map[string]*types.Guardrail) map[string][]string {
+	type providerAcc struct {
+		models          map[string]struct{}
+		anyUnrestricted bool
+	}
+	accs := make(map[string]*providerAcc)
+	for _, p := range policies {
+		if p == nil {
+			continue
+		}
+		restricted, models := policyModelAllowlist(p, byID)
+		for _, providerID := range p.DestinationProviderIDs {
+			if providerID == "" {
+				continue
+			}
+			acc, ok := accs[providerID]
+			if !ok {
+				acc = &providerAcc{models: make(map[string]struct{})}
+				accs[providerID] = acc
+			}
+			if !restricted {
+				acc.anyUnrestricted = true
+				continue
+			}
+			for _, m := range models {
+				acc.models[m] = struct{}{}
+			}
+		}
+	}
+	out := make(map[string][]string, len(accs))
+	for providerID, acc := range accs {
+		if acc.anyUnrestricted {
+			continue
+		}
+		models := make([]string, 0, len(acc.models))
+		for m := range acc.models {
+			models = append(models, m)
+		}
+		sort.Strings(models)
+		out[providerID] = models
+	}
+	return out
+}
+
+// policyModelAllowlist reports whether a policy restricts models (has an
+// allowlist-enabled guardrail) and the union of allowed models. Models are
+// verbatim; the proxy factory lowercases/trims them at decode time.
+func policyModelAllowlist(p *types.Policy, byID map[string]*types.Guardrail) (bool, []string) {
+	restricted := false
+	var models []string
+	for _, gID := range p.GuardrailIDs {
+		g, ok := byID[gID]
+		if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
+			continue
+		}
+		restricted = true
+		for _, m := range g.Checks.ModelAllowlist.Models {
+			if m != "" {
+				models = append(models, m)
+			}
+		}
+	}
+	return restricted, models
+}
+
 // buildAccountService composes the per-account gateway Service. The
 // target carries the noop placeholder URL — the router middleware
 // rewrites every request to the matched provider's upstream before the
@@ -986,38 +1058,11 @@ func unionSourceGroups(policies []*types.Policy) []string {
 	return out
 }
 
-// MergedGuardrails is the JSON shape passed to the proxy via the
-// guardrail middleware's config_json. Mirrors the proxy-side
-// expectations and is intentionally distinct from
-// types.GuardrailChecks so we can evolve either side independently.
+// MergedGuardrails is the synthesiser's fold target. Only prompt capture is
+// merged here — the model allowlist is emitted per-provider, and
+// token/budget/retention moved onto Policy.Limits and account Settings.
 type MergedGuardrails struct {
-	ModelAllowlist []string            `json:"model_allowlist,omitempty"`
-	TokenLimits    MergedTokenLimits   `json:"token_limits"`
-	Budget         MergedBudget        `json:"budget"`
-	PromptCapture  MergedPromptCapture `json:"prompt_capture"`
-	Retention      MergedRetention     `json:"retention"`
-}
-
-type MergedTokenLimits struct {
-	Hourly  *MergedTokenWindow `json:"hourly,omitempty"`
-	Daily   *MergedTokenWindow `json:"daily,omitempty"`
-	Monthly *MergedTokenWindow `json:"monthly,omitempty"`
-}
-
-type MergedTokenWindow struct {
-	MaxInputTokens  int `json:"max_input_tokens,omitempty"`
-	MaxOutputTokens int `json:"max_output_tokens,omitempty"`
-}
-
-type MergedBudget struct {
-	Hourly  *MergedBudgetWindow `json:"hourly,omitempty"`
-	Daily   *MergedBudgetWindow `json:"daily,omitempty"`
-	Monthly *MergedBudgetWindow `json:"monthly,omitempty"`
-}
-
-type MergedBudgetWindow struct {
-	SoftCapUSD float64 `json:"soft_cap_usd,omitempty"`
-	HardCapUSD float64 `json:"hard_cap_usd,omitempty"`
+	PromptCapture MergedPromptCapture
 }
 
 type MergedPromptCapture struct {
@@ -1025,64 +1070,31 @@ type MergedPromptCapture struct {
 	RedactPii bool `json:"redact_pii"`
 }
 
-type MergedRetention struct {
-	Enabled bool `json:"enabled"`
-	Days    int  `json:"days"`
-}
-
-// mergeGuardrails computes the effective guardrail spec applied at the
-// proxy, given the referencing policies and the account's guardrail
-// catalogue. Policy enabled-ness is the caller's responsibility — only
-// enabled policies should be passed in.
+// mergeGuardrails folds the referencing policies' guardrails into the
+// prompt-capture decision only. The model allowlist is enforced per-policy/group
+// in management and shipped per-provider; token/budget/retention live off
+// guardrails now.
 //
-// Merge rules:
-//   - Model allowlist:   union of allowlists across policies that enable it.
-//   - Token / Budget:    most-restrictive (min of non-zero caps) per window.
-//   - Prompt capture:    enabled if any policy enables it; redact_pii sticks
-//     if any enabling policy turns it on.
-//   - Retention:         enabled if any enables it; smallest non-zero days wins.
+// Merge rule — prompt capture: enabled if any policy enables it; redact_pii
+// sticks if any enabling policy turns it on.
 func mergeGuardrails(policies []*types.Policy, byID map[string]*types.Guardrail) MergedGuardrails {
 	merged := MergedGuardrails{}
-	allowlist := make(map[string]struct{})
-	allowlistEnabled := false
-
 	for _, policy := range policies {
 		for _, gID := range policy.GuardrailIDs {
 			g, ok := byID[gID]
 			if !ok || g == nil {
 				continue
 			}
-			mergeGuardrail(g, &merged, allowlist, &allowlistEnabled)
+			mergeGuardrail(g, &merged)
 		}
 	}
-
-	if allowlistEnabled {
-		merged.ModelAllowlist = make([]string, 0, len(allowlist))
-		for m := range allowlist {
-			merged.ModelAllowlist = append(merged.ModelAllowlist, m)
-		}
-		sort.Strings(merged.ModelAllowlist)
-	}
 	return merged
 }
 
-// mergeGuardrail folds a single guardrail's enabled checks into the
-// running merge: model-allowlist models join the shared set (and flip
-// allowlistEnabled), and prompt-capture / redact-pii stick once any
-// enabling guardrail turns them on.
-//
-// TokenLimits, Budget, and Retention have moved off guardrails — token
-// and budget caps now live on the Policy itself (Policy.Limits) and
-// retention moves to account-level Settings — so they are not merged here.
-func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails, allowlist map[string]struct{}, allowlistEnabled *bool) {
-	if g.Checks.ModelAllowlist.Enabled {
-		*allowlistEnabled = true
-		for _, m := range g.Checks.ModelAllowlist.Models {
-			if m != "" {
-				allowlist[m] = struct{}{}
-			}
-		}
-	}
+// mergeGuardrail folds a single guardrail's prompt-capture settings into the
+// running merge: enabled / redact-pii stick once any enabling guardrail turns
+// them on.
+func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails) {
 	if g.Checks.PromptCapture.Enabled {
 		merged.PromptCapture.Enabled = true
 		if g.Checks.PromptCapture.RedactPii {

management/internals/modules/agentnetwork/synthesizer_guardrail_realstore_test.go

@@ -81,8 +81,8 @@ func TestSynthesizeServices_RealStore_PromptCaptureAccountIsSoleControl(t *testi
 	require.Len(t, services, 1)
 
 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	assert.Equal(t, []string{"gpt-5.4"}, cfg.ModelAllowlist,
-		"model allowlist is a pure policy guardrail and must always reach the config")
+	assert.Equal(t, map[string][]string{"prov-1": {"gpt-5.4"}}, cfg.ProviderAllowlists,
+		"model allowlist is a pure policy guardrail and must reach the per-provider config")
 	assert.False(t, cfg.PromptCapture.Enabled,
 		"prompt capture must be off when the account toggle is off, even with a capture-enabled guardrail")
 }
@@ -172,7 +172,7 @@ func TestSynthesizeServices_RealStore_NoGuardrail_CaptureOff(t *testing.T) {
 	require.Len(t, services, 1, "exactly one synth service expected")
 
 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	assert.Empty(t, cfg.ModelAllowlist, "no guardrail → no allowlist")
+	assert.Empty(t, cfg.ProviderAllowlists, "no guardrail → provider unrestricted (absent from map)")
 	assert.False(t, cfg.PromptCapture.Enabled, "no guardrail → prompt capture off by default")
 	assert.False(t, cfg.PromptCapture.RedactPii, "no guardrail → redact off by default")
 }

management/internals/modules/agentnetwork/synthesizer_provider_allowlist_test.go

@@ -0,0 +1,95 @@
+package agentnetwork
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+)
+
+// policyForProviders builds an enabled policy authorising the given providers
+// under the given guardrails (both optional). Groups are irrelevant to
+// buildProviderAllowlists, which keys purely on destination provider.
+func policyForProviders(id string, guardrailIDs []string, providerIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		Enabled:                true,
+		DestinationProviderIDs: providerIDs,
+		GuardrailIDs:           guardrailIDs,
+	}
+}
+
+func TestBuildProviderAllowlists(t *testing.T) {
+	byID := map[string]*types.Guardrail{
+		"g-4o":       allowlistGuardrail("g-4o", "acc-1", "gpt-4o"),
+		"g-opus":     allowlistGuardrail("g-opus", "acc-1", "claude-opus-4"),
+		"g-disabled": {ID: "g-disabled", Checks: types.GuardrailChecks{ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}}}},
+	}
+
+	t.Run("all authorising policies restrict yields per-provider union", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", []string{"g-opus"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, map[string][]string{"prov-x": {"claude-opus-4", "gpt-4o"}}, got,
+			"a provider every policy restricts carries the sorted union of their models")
+	})
+
+	t.Run("any un-guardrailed policy leaves the provider unrestricted (omitted)", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", nil, "prov-x"), // no guardrail
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.NotContains(t, got, "prov-x",
+			"a provider reachable by an un-guardrailed policy must be omitted so the proxy treats it as unrestricted")
+	})
+
+	t.Run("a disabled allowlist counts as unrestricted", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-disabled"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.NotContains(t, got, "prov-x",
+			"a policy whose only guardrail has a disabled allowlist is unrestricted")
+	})
+
+	t.Run("providers are isolated from one another", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", []string{"g-opus"}, "prov-y"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-x"], "prov-x keeps only its own model")
+		assert.Equal(t, []string{"claude-opus-4"}, got["prov-y"], "prov-y keeps only its own model")
+	})
+
+	t.Run("one policy authorising two providers restricts both", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x", "prov-y"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-x"])
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-y"])
+	})
+
+	t.Run("union across a single policy's guardrails", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o", "g-opus"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.ElementsMatch(t, []string{"claude-opus-4", "gpt-4o"}, got["prov-x"],
+			"a policy's own multiple allowlist guardrails union together")
+	})
+
+	t.Run("an enabled allowlist with no models denies everything", func(t *testing.T) {
+		empty := map[string]*types.Guardrail{"g-empty": allowlistGuardrail("g-empty", "acc-1")}
+		got := buildProviderAllowlists([]*types.Policy{
+			policyForProviders("p1", []string{"g-empty"}, "prov-x"),
+		}, empty)
+		assert.Equal(t, map[string][]string{"prov-x": {}}, got,
+			"an enabled-but-empty allowlist is restricted with an empty set, not unrestricted")
+	})
+}

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -1031,8 +1031,12 @@ func TestSynthesizeServices_GuardrailMerge_AllowlistUnion_LimitsRestrictive(t *t
 
 	var cfg guardrailConfig
 	require.NoError(t, json.Unmarshal(guardrailJSON, &cfg), "guardrail config must unmarshal cleanly")
-	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ModelAllowlist,
-		"model allowlist union must keep both models")
+	// Both policies restrict the same provider, so the per-provider backstop
+	// carries the union of their models — a coarse gate that management's
+	// per-policy/group check narrows; it only blocks models outside the union
+	// when management is down.
+	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ProviderAllowlists["prov-1"],
+		"per-provider allowlist union must keep both models")
 }
 
 func TestSynthesizeServices_BackfillsMissingSessionKeys(t *testing.T) {

management/internals/shared/grpc/proxy.go

@@ -285,6 +285,7 @@ func (s *ProxyServiceServer) CheckLLMPolicyLimits(ctx context.Context, req *prot
 		UserID:     req.GetUserId(),
 		GroupIDs:   req.GetGroupIds(),
 		ProviderID: req.GetProviderId(),
+		Model:      req.GetModel(),
 	})
 	if err != nil {
 		log.WithContext(ctx).Errorf("select policy for request: %v", err)

management/internals/shared/grpc/proxy_llm_policy_limits_test.go

@@ -0,0 +1,138 @@
+package grpc
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc/codes"
+	grpcstatus "google.golang.org/grpc/status"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
+	"github.com/netbirdio/netbird/shared/management/proto"
+)
+
+// fakeAgentNetworkLimits records the PolicySelectionInput it was invoked with
+// and returns a pre-programmed result, so tests can assert what the handler
+// forwards to the selector.
+type fakeAgentNetworkLimits struct {
+	gotInput agentnetwork.PolicySelectionInput
+	result   *agentnetwork.PolicySelectionResult
+	err      error
+}
+
+func (f *fakeAgentNetworkLimits) SelectPolicyForRequest(_ context.Context, in agentnetwork.PolicySelectionInput) (*agentnetwork.PolicySelectionResult, error) {
+	f.gotInput = in
+	if f.err != nil {
+		return nil, f.err
+	}
+	return f.result, nil
+}
+
+func (f *fakeAgentNetworkLimits) RecordUsage(_ context.Context, _ agentnetwork.RecordUsageInput) error {
+	return nil
+}
+
+// TestCheckLLMPolicyLimits_ForwardsModelToSelector proves the wiring added here:
+// the model the proxy extracted must reach the selector's Model unchanged,
+// alongside the account/user/group/provider fields.
+func TestCheckLLMPolicyLimits_ForwardsModelToSelector(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{Allow: true, SelectedPolicyID: "pol-1"}}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	req := &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		UserId:     "user-1",
+		GroupIds:   []string{"grp-a", "grp-b"},
+		ProviderId: "prov-1",
+		Model:      "claude-opus-4",
+	}
+
+	resp, err := s.CheckLLMPolicyLimits(context.Background(), req)
+	require.NoError(t, err)
+	require.NotNil(t, resp)
+
+	assert.Equal(t, "acc-1", fake.gotInput.AccountID)
+	assert.Equal(t, "user-1", fake.gotInput.UserID)
+	assert.Equal(t, []string{"grp-a", "grp-b"}, fake.gotInput.GroupIDs)
+	assert.Equal(t, "prov-1", fake.gotInput.ProviderID)
+	assert.Equal(t, "claude-opus-4", fake.gotInput.Model,
+		"the request's model must be forwarded to the selector")
+}
+
+// TestCheckLLMPolicyLimits_EmptyModelForwardedAsEmpty proves an undetermined
+// model (empty string) is forwarded as-is; the selector decides how to treat it.
+func TestCheckLLMPolicyLimits_EmptyModelForwardedAsEmpty(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{Allow: true}}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	req := &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+	}
+
+	_, err := s.CheckLLMPolicyLimits(context.Background(), req)
+	require.NoError(t, err)
+	assert.Equal(t, "", fake.gotInput.Model, "an absent model must be forwarded as empty, not fabricated")
+}
+
+// TestCheckLLMPolicyLimits_DenyResponseCarriesModelBlockedCode proves the deny
+// envelope surfaces the model-allowlist deny code + reason through the response.
+func TestCheckLLMPolicyLimits_DenyResponseCarriesModelBlockedCode(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{
+		Allow:      false,
+		DenyCode:   "llm_policy.model_blocked",
+		DenyReason: `model "claude-opus-4" is not permitted by any applicable policy allowlist`,
+	}}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	resp, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	require.NotNil(t, resp)
+	assert.Equal(t, "deny", resp.Decision)
+	assert.Equal(t, "llm_policy.model_blocked", resp.DenyCode)
+	assert.NotEmpty(t, resp.DenyReason)
+	assert.Empty(t, resp.SelectedPolicyId, "a denied request must carry no selected policy")
+}
+
+// TestCheckLLMPolicyLimits_SelectorErrorSurfacesAsInternal proves a selector
+// failure surfaces as an Internal gRPC error rather than a silent allow.
+func TestCheckLLMPolicyLimits_SelectorErrorSurfacesAsInternal(t *testing.T) {
+	fake := &fakeAgentNetworkLimits{err: errors.New("boom")}
+	s := &ProxyServiceServer{}
+	s.SetAgentNetworkLimitsService(fake)
+
+	_, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.Error(t, err)
+	st, ok := grpcstatus.FromError(err)
+	require.True(t, ok)
+	assert.Equal(t, codes.Internal, st.Code(), "selector errors must never fail open on the hot path")
+}
+
+// TestCheckLLMPolicyLimits_UnconfiguredServiceReturnsUnimplemented locks the
+// fallback: with no limits service wired the RPC returns Unimplemented.
+func TestCheckLLMPolicyLimits_UnconfiguredServiceReturnsUnimplemented(t *testing.T) {
+	s := &ProxyServiceServer{}
+
+	_, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
+		AccountId:  "acc-1",
+		ProviderId: "prov-1",
+	})
+	require.Error(t, err)
+	st, ok := grpcstatus.FromError(err)
+	require.True(t, ok)
+	assert.Equal(t, codes.Unimplemented, st.Code())
+}

proxy/internal/auth/tunnel_cache.go

@@ -3,20 +3,30 @@ package auth
 import (
 	"context"
 	"net/netip"
+	"os"
+	"strings"
 	"sync"
 	"time"
 
+	log "github.com/sirupsen/logrus"
 	"golang.org/x/sync/singleflight"
 
 	"github.com/netbirdio/netbird/proxy/internal/types"
 	"github.com/netbirdio/netbird/shared/management/proto"
 )
 
-// tunnelCacheTTL caps how long a positive ValidateTunnelPeer result is
-// reused before re-fetching from management. 5 minutes balances freshness
-// against management load on busy mesh networks.
+// tunnelCacheTTL is the default cap on how long a positive ValidateTunnelPeer
+// result is reused before re-fetching from management. 5 minutes balances
+// freshness against management load on busy mesh networks. Override it with
+// envTunnelCacheTTL when an account needs authorization changes to take effect
+// sooner (at the cost of more ValidateTunnelPeer RPCs).
 const tunnelCacheTTL = 300 * time.Second
 
+// envTunnelCacheTTL overrides tunnelCacheTTL. The value is a Go duration string
+// (e.g. "30s", "2m"); an unset, unparseable, or non-positive value keeps the
+// default.
+const envTunnelCacheTTL = "NB_PROXY_TUNNEL_CACHE_TTL"
+
 // tunnelCachePerAccount caps the number of cached identities per account.
 // Bounded eviction avoids memory growth in pathological cases (huge peer
 // roster, brief request bursts) while staying generous for normal use.
@@ -60,16 +70,35 @@ type accountBucket struct {
 	order []tunnelCacheKey
 }
 
-// newTunnelValidationCache constructs a cache with default TTL and bounds.
+// newTunnelValidationCache constructs a cache with the configured TTL
+// (envTunnelCacheTTL override or default) and default bounds.
 func newTunnelValidationCache() *tunnelValidationCache {
 	return &tunnelValidationCache{
 		entries: make(map[types.AccountID]*accountBucket),
-		ttl:     tunnelCacheTTL,
+		ttl:     tunnelCacheTTLFromEnv(),
 		maxSize: tunnelCachePerAccount,
 		now:     time.Now,
 	}
 }
 
+// tunnelCacheTTLFromEnv returns the tunnel-cache TTL, honoring the
+// envTunnelCacheTTL override. The override must be a positive Go duration
+// string (e.g. "30s", "2m"); anything unset, unparseable, or non-positive
+// falls back to tunnelCacheTTL.
+func tunnelCacheTTLFromEnv() time.Duration {
+	raw := strings.TrimSpace(os.Getenv(envTunnelCacheTTL))
+	if raw == "" {
+		return tunnelCacheTTL
+	}
+	d, err := time.ParseDuration(raw)
+	if err != nil || d <= 0 {
+		log.Warnf("ignoring invalid %s=%q (want a positive Go duration like 30s or 2m); using default %s",
+			envTunnelCacheTTL, raw, tunnelCacheTTL)
+		return tunnelCacheTTL
+	}
+	return d
+}
+
 // get returns a cached response for the key, or nil when missing or
 // expired. Expired entries are evicted lazily on read.
 func (c *tunnelValidationCache) get(key tunnelCacheKey) *proto.ValidateTunnelPeerResponse {

proxy/internal/auth/tunnel_cache_test.go

@@ -169,3 +169,32 @@ func TestTunnelCache_BoundedSizeEvictsOldest(t *testing.T) {
 	assert.NotNil(t, cache.get(keys[1]), "second-newest must remain cached")
 	assert.NotNil(t, cache.get(keys[2]), "newest must remain cached")
 }
+
+func TestTunnelCacheTTLFromEnv(t *testing.T) {
+	t.Run("unset uses default", func(t *testing.T) {
+		t.Setenv(envTunnelCacheTTL, "")
+		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
+	})
+	t.Run("valid duration overrides", func(t *testing.T) {
+		t.Setenv(envTunnelCacheTTL, "45s")
+		assert.Equal(t, 45*time.Second, tunnelCacheTTLFromEnv())
+	})
+	t.Run("whitespace trimmed", func(t *testing.T) {
+		t.Setenv(envTunnelCacheTTL, "  2m ")
+		assert.Equal(t, 2*time.Minute, tunnelCacheTTLFromEnv())
+	})
+	t.Run("unparseable uses default", func(t *testing.T) {
+		t.Setenv(envTunnelCacheTTL, "nonsense")
+		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
+	})
+	t.Run("non-positive uses default", func(t *testing.T) {
+		t.Setenv(envTunnelCacheTTL, "0s")
+		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
+		t.Setenv(envTunnelCacheTTL, "-30s")
+		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
+	})
+	t.Run("constructor honors override", func(t *testing.T) {
+		t.Setenv(envTunnelCacheTTL, "90s")
+		assert.Equal(t, 90*time.Second, newTunnelValidationCache().ttl)
+	})
+}

proxy/internal/llm/pricing/defaults_coverage_test.go

@@ -28,12 +28,12 @@ func TestDefaultTable_FirstPartyModelCoverage(t *testing.T) {
 			"ministral-8b-latest", "mistral-embed",
 		},
 		"anthropic": {
-			"claude-fable-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6",
+			"claude-fable-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6",
 			"claude-opus-4-1", "claude-sonnet-4-6", "claude-sonnet-4-5", "claude-haiku-4-5",
 		},
 		// bedrock keys are the normalized ids the request parser emits.
 		"bedrock": {
-			"anthropic.claude-opus-4-8", "anthropic.claude-opus-4-7", "anthropic.claude-opus-4-6",
+			"anthropic.claude-opus-5", "anthropic.claude-opus-4-8", "anthropic.claude-opus-4-7", "anthropic.claude-opus-4-6",
 			"anthropic.claude-opus-4-1", "anthropic.claude-sonnet-4-6", "anthropic.claude-sonnet-4-5",
 			"anthropic.claude-haiku-4-5", "meta.llama3-3-70b-instruct",
 			"amazon.nova-pro", "amazon.nova-lite", "amazon.nova-micro", "amazon.nova-2-lite",

proxy/internal/llm/pricing/defaults_pricing.yaml

@@ -180,6 +180,11 @@ anthropic:
     output_per_1k: 0.050
     cache_read_per_1k: 0.001
     cache_creation_per_1k: 0.0125
+  claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
   claude-opus-4-8:
     input_per_1k: 0.005
     output_per_1k: 0.025
@@ -236,6 +241,11 @@ bedrock:
   # eu.anthropic.claude-sonnet-4-5-20250929-v1:0 -> anthropic.claude-sonnet-4-5.
   # Anthropic-on-Bedrock keeps the additive cache buckets (read ≈0.1x input,
   # write ≈1.25x input); Nova / Llama report no cache, so cost is input+output.
+  anthropic.claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
   anthropic.claude-opus-4-8:
     input_per_1k: 0.005
     output_per_1k: 0.025

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

@@ -10,11 +10,15 @@ import (
 )
 
 // Config is the JSON-decoded shape accepted by the factory. The
-// runtime path consumes the normalised allowlist; raw config is not
+// runtime path consumes the normalised allowlists; raw config is not
 // retained beyond construction.
 type Config struct {
-	ModelAllowlist []string      `json:"model_allowlist"`
-	PromptCapture  PromptCapture `json:"prompt_capture"`
+	// ProviderAllowlists maps a resolved provider id (KeyLLMResolvedProviderID) to
+	// its model allowlist. A provider present is restricted to those models; one
+	// absent is unrestricted. Kept per-provider so one provider's list can't leak
+	// onto another.
+	ProviderAllowlists map[string][]string `json:"provider_allowlists,omitempty"`
+	PromptCapture      PromptCapture       `json:"prompt_capture"`
 }
 
 // PromptCapture toggles the optional prompt capture + redaction step
@@ -54,21 +58,28 @@ func isEmptyJSON(raw []byte) bool {
 	return false
 }
 
-// normaliseConfig lowercases and trims allowlist entries so the runtime
-// match is case-insensitive. Empty entries are dropped.
+// normaliseConfig lowercases and trims allowlist entries for case-insensitive
+// matching; empty entries drop. A provider whose entries all drop keeps an empty
+// (non-nil) list — "deny every model" — distinct from an absent provider
+// (unrestricted).
 func normaliseConfig(cfg Config) Config {
-	if len(cfg.ModelAllowlist) == 0 {
+	if len(cfg.ProviderAllowlists) == 0 {
+		cfg.ProviderAllowlists = nil
 		return cfg
 	}
-	cleaned := make([]string, 0, len(cfg.ModelAllowlist))
-	for _, entry := range cfg.ModelAllowlist {
-		n := normaliseModel(entry)
-		if n == "" {
-			continue
+	cleaned := make(map[string][]string, len(cfg.ProviderAllowlists))
+	for provider, models := range cfg.ProviderAllowlists {
+		list := make([]string, 0, len(models))
+		for _, entry := range models {
+			n := normaliseModel(entry)
+			if n == "" {
+				continue
+			}
+			list = append(list, n)
 		}
-		cleaned = append(cleaned, n)
+		cleaned[provider] = list
 	}
-	cfg.ModelAllowlist = cleaned
+	cfg.ProviderAllowlists = cleaned
 	return cfg
 }
 

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

@@ -83,8 +83,9 @@ func (m *Middleware) MutationsSupported() bool { return false }
 // prompt capture only affects the metadata emitted alongside an allow.
 func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middleware.Output, error) {
 	model, modelPresent := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
+	providerID, _ := lookupMetadata(in.Metadata, middleware.KeyLLMResolvedProviderID)
 
-	if denial := m.evaluateAllowlist(model, modelPresent); denial != nil {
+	if denial := m.evaluateAllowlist(providerID, model, modelPresent); denial != nil {
 		return denial, nil
 	}
 
@@ -110,20 +111,32 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 // is a no-op.
 func (m *Middleware) Close() error { return nil }
 
-// evaluateAllowlist returns a deny Output when the configured allowlist
-// rejects the model. A nil return means the request should proceed.
-func (m *Middleware) evaluateAllowlist(model string, modelPresent bool) *middleware.Output {
-	if len(m.cfg.ModelAllowlist) == 0 {
+// evaluateAllowlist denies when the resolved provider's allowlist rejects the
+// model; nil means proceed. Scoped to the provider llm_router resolved, so an
+// unrestricted provider (absent from config) is never caught by another's list.
+func (m *Middleware) evaluateAllowlist(providerID, model string, modelPresent bool) *middleware.Output {
+	if len(m.cfg.ProviderAllowlists) == 0 {
 		return nil
 	}
-	// Fail closed: with an allowlist configured, a request whose model the
-	// upstream parser could not extract (absent or empty) must be denied rather
-	// than allowed. This is what enforces the allowlist for URL/path-routed
-	// providers (Bedrock, Vertex, ...) whose model lives outside the JSON body.
+	// Restrictions exist but the resolved provider is unknown, so we can't tell
+	// if this request targets a restricted provider — fail closed. llm_router
+	// normally stamps the provider first, so this is a defensive guard.
+	if providerID == "" {
+		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
+	}
+	allowlist, restricted := m.cfg.ProviderAllowlists[providerID]
+	if !restricted {
+		// This provider has no allowlist (some authorising policy left it
+		// unrestricted); management owns any per-policy/group decision.
+		return nil
+	}
+	// Fail closed: with an allowlist in effect for this provider, a request whose
+	// model the parser couldn't extract (absent/empty) is denied. This enforces
+	// the allowlist for path-routed providers (Bedrock, Vertex) with no body model.
 	if !modelPresent || normaliseModel(model) == "" {
 		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
-	if m.modelInAllowlist(model) {
+	if modelInAllowlist(allowlist, model) {
 		return nil
 	}
 	return denyModel(model, denyCodeModel, denyMessageModel, denyReasonModel)
@@ -151,14 +164,15 @@ func denyModel(model, code, message, reason string) *middleware.Output {
 	}
 }
 
-// modelInAllowlist reports whether the model matches any allowlist
-// entry under the case-insensitive, trim-tolerant comparison rule.
-func (m *Middleware) modelInAllowlist(model string) bool {
+// modelInAllowlist reports whether the model matches any entry in the supplied
+// (already-normalised) allowlist under the case-insensitive, trim-tolerant
+// comparison rule.
+func modelInAllowlist(allowlist []string, model string) bool {
 	normalised := normaliseModel(model)
 	if normalised == "" {
 		return false
 	}
-	for _, allowed := range m.cfg.ModelAllowlist {
+	for _, allowed := range allowlist {
 		if allowed == normalised {
 			return true
 		}

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

@@ -26,6 +26,25 @@ func newInput(meta ...middleware.KV) *middleware.Input {
 	return &middleware.Input{Slot: middleware.SlotOnRequest, Metadata: meta}
 }
 
+const (
+	testProvider  = "prov-1"
+	otherProvider = "prov-2"
+)
+
+// providerCfg builds a Config restricting testProvider to the given models.
+func providerCfg(models ...string) Config {
+	return Config{ProviderAllowlists: map[string][]string{testProvider: models}}
+}
+
+// newInputProvider builds an input that carries a resolved provider id (as
+// llm_router would stamp) plus any extra metadata.
+func newInputProvider(provider string, meta ...middleware.KV) *middleware.Input {
+	all := make([]middleware.KV, 0, len(meta)+1)
+	all = append(all, middleware.KV{Key: middleware.KeyLLMResolvedProviderID, Value: provider})
+	all = append(all, meta...)
+	return &middleware.Input{Slot: middleware.SlotOnRequest, Metadata: all}
+}
+
 func TestMiddlewareIdentity(t *testing.T) {
 	mw := New(Config{})
 	assert.Equal(t, ID, mw.ID(), "middleware ID must be llm_guardrail")
@@ -47,12 +66,12 @@ func TestMiddlewareIdentity(t *testing.T) {
 
 func TestAllowlistEmptyAllowsAnyModel(t *testing.T) {
 	mw := New(Config{})
-	out, err := mw.Invoke(context.Background(), newInput(
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "empty allowlist must allow any model")
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no provider allowlists must allow any model")
 	v, ok := metaValue(t, out.Metadata, middleware.KeyLLMPolicyDecision)
 	require.True(t, ok, "decision metadata must be emitted")
 	assert.Equal(t, "allow", v, "decision must be allow")
@@ -62,8 +81,8 @@ func TestAllowlistEmptyAllowsAnyModel(t *testing.T) {
 }
 
 func TestAllowlistMatchAllows(t *testing.T) {
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o", "claude-opus-4"}})
-	out, err := mw.Invoke(context.Background(), newInput(
+	mw := New(providerCfg("gpt-4o", "claude-opus-4"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
@@ -71,8 +90,8 @@ func TestAllowlistMatchAllows(t *testing.T) {
 }
 
 func TestAllowlistMissDenies(t *testing.T) {
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
-	out, err := mw.Invoke(context.Background(), newInput(
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
 	))
 	require.NoError(t, err)
@@ -91,10 +110,10 @@ func TestAllowlistMissDenies(t *testing.T) {
 }
 
 func TestAllowlistCaseInsensitive(t *testing.T) {
-	mw := New(Config{ModelAllowlist: []string{"  GPT-4o  ", "Claude-OPUS-4"}})
+	mw := New(providerCfg("  GPT-4o  ", "Claude-OPUS-4"))
 	cases := []string{"gpt-4o", "GPT-4O", "  claude-opus-4 "}
 	for _, model := range cases {
-		out, err := mw.Invoke(context.Background(), newInput(
+		out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 			middleware.KV{Key: middleware.KeyLLMModel, Value: model},
 		))
 		require.NoError(t, err)
@@ -103,14 +122,15 @@ func TestAllowlistCaseInsensitive(t *testing.T) {
 }
 
 func TestAllowlistMissingModelKeyDenies(t *testing.T) {
-	// Fail closed: with an allowlist configured, a request whose model the
-	// parser could not extract (URL/path-routed providers such as Bedrock or
-	// Vertex whose shape wasn't recognised) must be denied, not allowed.
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
-	out, err := mw.Invoke(context.Background(), newInput())
+	// Fail closed: with an allowlist in effect for the resolved provider, a
+	// request whose model the parser could not extract (URL/path-routed
+	// providers such as Bedrock or Vertex whose shape wasn't recognised) must be
+	// denied, not allowed.
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when an allowlist is set")
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when the provider is restricted")
 	assert.Equal(t, 403, out.DenyStatus, "deny status must be 403")
 	require.NotNil(t, out.DenyReason, "deny reason must be populated")
 	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
@@ -122,26 +142,101 @@ func TestAllowlistMissingModelKeyDenies(t *testing.T) {
 
 func TestAllowlistEmptyModelValueDenies(t *testing.T) {
 	// A present-but-empty model is as undeterminable as an absent one.
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
-	out, err := mw.Invoke(context.Background(), newInput(
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "   "},
 	))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when an allowlist is set")
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when the provider is restricted")
 	require.NotNil(t, out.DenyReason, "deny reason must be populated")
 	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
 }
 
 func TestAllowlistEmptyListAllowsMissingModel(t *testing.T) {
-	// Without an allowlist there is nothing to enforce, so a missing model is
-	// still allowed — the fail-closed rule only applies when a list is set.
+	// Without any provider allowlists there is nothing to enforce, so a missing
+	// model is still allowed — the fail-closed rule only applies when a
+	// restriction is in effect.
 	mw := New(Config{})
 	out, err := mw.Invoke(context.Background(), newInput())
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no allowlist must allow even without a model")
 }
 
+func TestUnrestrictedProviderAllowsAnyModel(t *testing.T) {
+	// The request resolved to otherProvider, which has no allowlist, so its
+	// traffic must not be caught by testProvider's restriction — the
+	// cross-provider-leak / false-deny guard.
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(otherProvider,
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
+	))
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "an unrestricted provider must not inherit another provider's allowlist")
+}
+
+func TestPerProviderAllowlistsAreIsolated(t *testing.T) {
+	// gpt-4o is allowed only on testProvider; claude-opus-4 only on
+	// otherProvider. A model allowlisted for one provider must not be usable on
+	// the other — the fail-closed layer never unions allowlists across providers.
+	mw := New(Config{ProviderAllowlists: map[string][]string{
+		testProvider:  {"gpt-4o"},
+		otherProvider: {"claude-opus-4"},
+	}})
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
+	))
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "claude-opus-4 is allowed only on otherProvider, not testProvider")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "cross-provider model must be blocked, not model_unknown")
+}
+
+func TestRestrictionsButNoResolvedProviderFailsClosed(t *testing.T) {
+	// Restrictions exist for the account but the resolved provider id is absent,
+	// so the request cannot be scoped to a provider. Fail closed rather than
+	// wave it through.
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInput(
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "missing resolved provider must fail closed when restrictions exist")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
+}
+
+func TestEnabledButEmptyAllowlistDeniesEveryModel(t *testing.T) {
+	// An allowlist-enabled provider with zero models is distinct from an
+	// unrestricted (absent) provider: it must deny every model.
+	mw := New(providerCfg())
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "an enabled-but-empty allowlist must deny every model")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "deny code must be model_blocked, not model_unknown")
+}
+
+func TestFactoryAllEmptyEntriesDenyEveryModel(t *testing.T) {
+	// All the provider's entries are blank; they collapse to a non-nil empty
+	// list (deny everything for that provider), not "no restriction".
+	raw := []byte(`{"provider_allowlists":{"prov-1":["","   "]}}`)
+	mw, err := Factory{}.New(raw)
+	require.NoError(t, err)
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "all-blank allowlist entries must still restrict the provider")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "deny code must be model_blocked")
+}
+
 func TestPromptCaptureDisabledEmitsNoPrompt(t *testing.T) {
 	mw := New(Config{})
 	out, err := mw.Invoke(context.Background(), newInput(
@@ -217,8 +312,8 @@ func TestFactoryAcceptsZeroConfigs(t *testing.T) {
 
 func TestFactoryDecodesValidConfig(t *testing.T) {
 	cfg := Config{
-		ModelAllowlist: []string{"gpt-4o"},
-		PromptCapture:  PromptCapture{Enabled: true, RedactPii: true},
+		ProviderAllowlists: map[string][]string{testProvider: {"gpt-4o"}},
+		PromptCapture:      PromptCapture{Enabled: true, RedactPii: true},
 	}
 	raw, err := json.Marshal(cfg)
 	require.NoError(t, err, "marshalling test config must succeed")
@@ -234,15 +329,15 @@ func TestFactoryRejectsMalformedJSON(t *testing.T) {
 }
 
 func TestFactoryNormalisesAllowlist(t *testing.T) {
-	raw := []byte(`{"model_allowlist":["  GPT-4o  ","",""," Claude-3 "]}`)
+	raw := []byte(`{"provider_allowlists":{"prov-1":["  GPT-4o  ","",""," Claude-3 "]}}`)
 	mw, err := Factory{}.New(raw)
 	require.NoError(t, err)
-	out, err := mw.Invoke(context.Background(), newInput(
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "factory must lowercase + trim allowlist entries")
-	out2, err := mw.Invoke(context.Background(), newInput(
+	out2, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-3"},
 	))
 	require.NoError(t, err)

proxy/internal/middleware/builtin/llm_limit_check/middleware.go

@@ -175,7 +175,7 @@ func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.Ou
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
 			Code:    code,
-			Message: "LLM policy limit exceeded",
+			Message: denyMessageForCode(code),
 		},
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMPolicyDecision, Value: "deny"},
@@ -184,6 +184,21 @@ func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.Ou
 	}
 }
 
+// denyMessageForCode maps a management deny code to a public message.
+// Model-allowlist rejections get a model-specific message matching the
+// local guardrail; everything else keeps the generic quota wording. The
+// message stays generic so it never leaks internal quota detail.
+func denyMessageForCode(code string) string {
+	switch code {
+	case "llm_policy.model_blocked":
+		return "model is not in the policy allowlist"
+	case "llm_policy.model_unknown":
+		return "request model could not be determined for the policy allowlist"
+	default:
+		return "LLM policy limit exceeded"
+	}
+}
+
 // lookupKV returns the value associated with key, or the empty
 // string when absent.
 func lookupKV(kvs []middleware.KV, key string) string {

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

@@ -115,6 +115,46 @@ func TestInvoke_DenyConvertsToProxyDeny(t *testing.T) {
 	assert.NotContains(t, out.DenyReason.Message, "1000", "internal cap numbers must not reach the caller")
 }
 
+// TestInvoke_ModelDenyMessages proves a model-allowlist rejection gets a
+// model-specific public message rather than the generic quota wording, so a
+// blocked or undetermined model reads consistently with the local guardrail.
+func TestInvoke_ModelDenyMessages(t *testing.T) {
+	cases := []struct {
+		name    string
+		code    string
+		message string
+	}{
+		{"blocked", "llm_policy.model_blocked", "model is not in the policy allowlist"},
+		{"unknown", "llm_policy.model_unknown", "request model could not be determined for the policy allowlist"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			mgmt := &fakeMgmt{
+				checkResp: &proto.CheckLLMPolicyLimitsResponse{
+					Decision: "deny",
+					DenyCode: tc.code,
+				},
+			}
+			m := New(mgmt, nil)
+
+			out := runInvoke(t, m, &middleware.Input{
+				AccountID:  "acc-1",
+				UserGroups: []string{"grp-engineers"},
+				Metadata: []middleware.KV{
+					{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-1"},
+					{Key: middleware.KeyLLMModel, Value: "some-model"},
+				},
+			})
+
+			assert.Equal(t, middleware.DecisionDeny, out.Decision)
+			require.NotNil(t, out.DenyReason, "deny envelope must carry a reason payload")
+			assert.Equal(t, tc.code, out.DenyReason.Code, "canonical deny code surfaces to the caller")
+			assert.Equal(t, tc.message, out.DenyReason.Message,
+				"model denials must use a model-specific message, matching the local guardrail")
+		})
+	}
+}
+
 // TestInvoke_NoMgmtClientPassesThrough proves the partial-wiring
 // safety: a middleware constructed without a management client
 // allows every request without attribution. This makes a half-set-up

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

@@ -25,10 +25,17 @@ func runParserGuardrail(t *testing.T, url string, body []byte, allowlist []strin
 	})
 	require.NoError(t, err, "parser must not error")
 
-	guard := llm_guardrail.New(llm_guardrail.Config{ModelAllowlist: allowlist})
+	const providerID = "prov-under-test"
+	guard := llm_guardrail.New(llm_guardrail.Config{
+		ProviderAllowlists: map[string][]string{providerID: allowlist},
+	})
+	// The real chain has llm_router stamp the resolved provider id before the
+	// guardrail runs; the parser doesn't, so add it here so the guardrail can
+	// scope the allowlist to this provider.
+	meta := append([]middleware.KV{{Key: middleware.KeyLLMResolvedProviderID, Value: providerID}}, parsed.Metadata...)
 	out, err := guard.Invoke(context.Background(), &middleware.Input{
 		Slot:     middleware.SlotOnRequest,
-		Metadata: parsed.Metadata,
+		Metadata: meta,
 	})
 	require.NoError(t, err, "guardrail must not error")
 	require.NotNil(t, out, "guardrail must return an output")