[client] make AllowedIPs swap self-healing on WireGuard errors

Address CodeRabbit review on #6799.

Decrement now decides whether to reprogram based on whether the currently
installed peer still holds references (e.peers[e.active] > 0) instead of
whether the released peer was the active one. If a prior swap's remove/add
failed, e.active was left pointing at a peer with zero references and every
later Decrement returned early, permanently stranding the prefix and leaking
the entry. The active peer is now detached only when e.active != "", and a
failed hand-off is retried on the next Decrement/Increment.

Also clear currentPeerKey unconditionally in the static handler's
RemoveAllowedIPs (matching dynamic/dnsinterceptor) and assert Increment
errors in the tests. Added self-heal tests for the failed remove/add paths.

client/internal/routemanager/dnsinterceptor/handler.go

@@ -95,7 +95,7 @@ func (d *DnsInterceptor) RemoveRoute() error {
 
 			// AllowedIPs should use real IPs
 			if d.currentPeerKey != "" {
-				if _, err := d.allowedIPsRefcounter.Decrement(prefix); err != nil {
+				if _, err := d.allowedIPsRefcounter.Decrement(prefix, d.currentPeerKey); err != nil {
 					merr = multierror.Append(merr, fmt.Errorf("remove allowed IP %s: %v", prefix, err))
 				}
 			}
@@ -172,7 +172,7 @@ func (d *DnsInterceptor) removeAllowedIP(realPrefix netip.Prefix) error {
 	}
 
 	// AllowedIPs use real IPs
-	if _, err := d.allowedIPsRefcounter.Decrement(realPrefix); err != nil {
+	if _, err := d.allowedIPsRefcounter.Decrement(realPrefix, d.currentPeerKey); err != nil {
 		return fmt.Errorf("remove allowed IP %s: %v", realPrefix, err)
 	}
 
@@ -205,7 +205,7 @@ func (d *DnsInterceptor) RemoveAllowedIPs() error {
 	for _, prefixes := range d.interceptedDomains {
 		for _, prefix := range prefixes {
 			// AllowedIPs use real IPs
-			if _, err := d.allowedIPsRefcounter.Decrement(prefix); err != nil {
+			if _, err := d.allowedIPsRefcounter.Decrement(prefix, d.currentPeerKey); err != nil {
 				merr = multierror.Append(merr, fmt.Errorf("remove allowed IP %s: %v", prefix, err))
 			}
 		}

client/internal/routemanager/dynamic/route.go

@@ -135,7 +135,7 @@ func (r *Route) RemoveAllowedIPs() error {
 	var merr *multierror.Error
 	for _, domainPrefixes := range r.dynamicDomains {
 		for _, prefix := range domainPrefixes {
-			if _, err := r.allowedIPsRefcounter.Decrement(prefix); err != nil {
+			if _, err := r.allowedIPsRefcounter.Decrement(prefix, r.currentPeerKey); err != nil {
 				merr = multierror.Append(merr, fmt.Errorf("remove allowed IP %s: %w", prefix, err))
 			}
 		}
@@ -320,7 +320,7 @@ func (r *Route) removeRoutes(prefixes []netip.Prefix) ([]netip.Prefix, error) {
 			merr = multierror.Append(merr, fmt.Errorf("remove dynamic route for IP %s: %w", prefix, err))
 		}
 		if r.currentPeerKey != "" {
-			if _, err := r.allowedIPsRefcounter.Decrement(prefix); err != nil {
+			if _, err := r.allowedIPsRefcounter.Decrement(prefix, r.currentPeerKey); err != nil {
 				merr = multierror.Append(merr, fmt.Errorf("remove allowed IP %s: %w", prefix, err))
 			}
 		}

client/internal/routemanager/manager.go

@@ -215,7 +215,7 @@ func (m *DefaultManager) setupRefCounters(useNoop bool) {
 		)
 	}
 
-	m.allowedIPsRefCounter = refcounter.New(
+	m.allowedIPsRefCounter = refcounter.NewAllowedIPs(
 		func(prefix netip.Prefix, peerKey string) (string, error) {
 			// save peerKey to use it in the remove function
 			return peerKey, m.wgInterface.AddAllowedIP(peerKey, prefix)

client/internal/routemanager/refcounter/allowedips.go

@@ -0,0 +1,185 @@
+package refcounter
+
+import (
+	"errors"
+	"fmt"
+	"net/netip"
+	"sort"
+	"sync"
+
+	"github.com/hashicorp/go-multierror"
+
+	nberrors "github.com/netbirdio/netbird/client/errors"
+)
+
+// allowedIPsEntry holds the per-peer reference counts for a single prefix and which peer is
+// currently installed in WireGuard. WireGuard allows a prefix on exactly one peer, so at most
+// one peer is active at a time even when several peers reference the prefix.
+type allowedIPsEntry struct {
+	// peers maps a peerKey to the number of references holding the prefix for that peer.
+	peers map[string]int
+	// active is the peerKey currently installed in WireGuard for this prefix ("" if none).
+	active string
+	// total is the sum of all per-peer reference counts (kept in sync with peers).
+	total int
+}
+
+// AllowedIPsRefCounter is a peer-aware reference counter for WireGuard AllowedIPs.
+//
+// The generic Counter keys only by prefix and remembers a single Out value set by the first
+// caller, which it never changes. That is wrong for AllowedIPs: two independent watchers (or
+// multiple resolved domains) can reference the same prefix through different peers, and when the
+// peer currently installed in WireGuard releases its last reference the prefix must be handed over
+// to a surviving peer instead of being left pointing at the released one.
+//
+// It calls add/remove (which program WireGuard) only on the transitions that matter:
+//   - add on the first reference for a prefix, or when swapping the active peer;
+//   - remove on the last reference for a prefix, or on the old peer during a swap.
+type AllowedIPsRefCounter struct {
+	mu      sync.Mutex
+	entries map[netip.Prefix]*allowedIPsEntry
+	add     AddFunc[netip.Prefix, string, string]
+	remove  RemoveFunc[netip.Prefix, string]
+}
+
+// NewAllowedIPs creates a new peer-aware AllowedIPs reference counter.
+// add programs a prefix on a peer in WireGuard and returns the peerKey to store as the active peer.
+// remove unprograms the prefix from the given peer.
+func NewAllowedIPs(add AddFunc[netip.Prefix, string, string], remove RemoveFunc[netip.Prefix, string]) *AllowedIPsRefCounter {
+	return &AllowedIPsRefCounter{
+		entries: map[netip.Prefix]*allowedIPsEntry{},
+		add:     add,
+		remove:  remove,
+	}
+}
+
+// Increment adds a reference to prefix for peerKey. WireGuard is programmed only for the first
+// reference to a prefix; while a different peer is already installed the prefix is left with it
+// (first peer wins, HA at the WireGuard layer is not possible) and only the reference count is kept.
+func (rm *AllowedIPsRefCounter) Increment(prefix netip.Prefix, peerKey string) (Ref[string], error) {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	e, ok := rm.entries[prefix]
+	if !ok {
+		e = &allowedIPsEntry{peers: map[string]int{}}
+		rm.entries[prefix] = e
+	}
+
+	logCallerF("Increasing allowed IP ref count for prefix %v peer %s [peer %d -> %d, total %d -> %d, active %q]",
+		prefix, peerKey, e.peers[peerKey], e.peers[peerKey]+1, e.total, e.total+1, e.active)
+
+	// Program WireGuard only when nothing is installed yet for this prefix.
+	if e.active == "" {
+		out, err := rm.add(prefix, peerKey)
+		if errors.Is(err, ErrIgnore) {
+			if e.total == 0 {
+				delete(rm.entries, prefix)
+			}
+			return Ref[string]{Count: e.total, Out: e.active}, nil
+		}
+		if err != nil {
+			if e.total == 0 {
+				delete(rm.entries, prefix)
+			}
+			return Ref[string]{}, fmt.Errorf("failed to add allowed IP %v for peer %s: %w", prefix, peerKey, err)
+		}
+		e.active = out
+	}
+
+	e.peers[peerKey]++
+	e.total++
+
+	return Ref[string]{Count: e.total, Out: e.active}, nil
+}
+
+// Decrement removes a reference to prefix for peerKey. When the peer currently installed in
+// WireGuard releases its last reference, the prefix is swapped to a surviving peer if one exists,
+// otherwise it is removed from WireGuard.
+func (rm *AllowedIPsRefCounter) Decrement(prefix netip.Prefix, peerKey string) (Ref[string], error) {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	e, ok := rm.entries[prefix]
+	if !ok {
+		logCallerF("No allowed IP reference found for prefix %v", prefix)
+		return Ref[string]{}, nil
+	}
+
+	if e.peers[peerKey] > 0 {
+		logCallerF("Decreasing allowed IP ref count for prefix %v peer %s [peer %d -> %d, total %d -> %d, active %q]",
+			prefix, peerKey, e.peers[peerKey], e.peers[peerKey]-1, e.total, e.total-1, e.active)
+		e.peers[peerKey]--
+		e.total--
+		if e.peers[peerKey] == 0 {
+			delete(e.peers, peerKey)
+		}
+	} else {
+		logCallerF("No allowed IP reference found for prefix %v peer %s", prefix, peerKey)
+	}
+
+	// If the peer currently installed in WireGuard still holds references, nothing to reprogram.
+	// Keying the check on the active peer (not the one just released) makes this self-healing:
+	// a prior swap whose remove/add failed leaves e.active pointing at a peer with no references,
+	// and this retries the hand-off on the next Decrement instead of getting stuck.
+	if e.active != "" && e.peers[e.active] > 0 {
+		return Ref[string]{Count: e.total, Out: e.active}, nil
+	}
+
+	// Detach the stale/gone active peer from WireGuard before reprogramming.
+	if e.active != "" {
+		if err := rm.remove(prefix, e.active); err != nil {
+			return Ref[string]{Count: e.total, Out: e.active}, fmt.Errorf("remove allowed IP %v for peer %s: %w", prefix, e.active, err)
+		}
+		e.active = ""
+	}
+
+	// Hand the prefix over to a surviving peer, or drop the entry when none remain.
+	if survivor, ok := pickSurvivor(e.peers); ok {
+		out, err := rm.add(prefix, survivor)
+		if err != nil {
+			return Ref[string]{Count: e.total, Out: ""}, fmt.Errorf("swap allowed IP %v to peer %s: %w", prefix, survivor, err)
+		}
+		e.active = out
+		return Ref[string]{Count: e.total, Out: e.active}, nil
+	}
+
+	delete(rm.entries, prefix)
+	return Ref[string]{Count: 0, Out: ""}, nil
+}
+
+// Flush removes all prefixes from WireGuard and clears the counter.
+func (rm *AllowedIPsRefCounter) Flush() error {
+	rm.mu.Lock()
+	defer rm.mu.Unlock()
+
+	var merr *multierror.Error
+	for prefix, e := range rm.entries {
+		if e.active == "" {
+			continue
+		}
+		logCallerF("Flushing allowed IP for prefix %v peer %s", prefix, e.active)
+		if err := rm.remove(prefix, e.active); err != nil {
+			merr = multierror.Append(merr, fmt.Errorf("remove allowed IP %v for peer %s: %w", prefix, e.active, err))
+		}
+	}
+
+	clear(rm.entries)
+
+	return nberrors.FormatErrorOrNil(merr)
+}
+
+// pickSurvivor deterministically selects a peer still referencing the prefix. WireGuard cannot do
+// multipath for a single prefix, so any surviving peer is a valid winner; the choice is made stable
+// (lowest peerKey) for predictable behavior and testability.
+func pickSurvivor(peers map[string]int) (string, bool) {
+	if len(peers) == 0 {
+		return "", false
+	}
+	keys := make([]string, 0, len(peers))
+	for k := range peers {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	return keys[0], true
+}

client/internal/routemanager/refcounter/allowedips_test.go

@@ -0,0 +1,241 @@
+package refcounter
+
+import (
+	"errors"
+	"net/netip"
+	"testing"
+)
+
+// fakeWG models WireGuard's cryptokey routing: a prefix can be installed on exactly one peer.
+// failAdd/failRemove make the next add/remove fail once, to exercise the self-healing error paths.
+type fakeWG struct {
+	installed  map[netip.Prefix]string
+	adds       int
+	removes    int
+	failAdd    bool
+	failRemove bool
+}
+
+func newFakeWG() *fakeWG {
+	return &fakeWG{installed: map[netip.Prefix]string{}}
+}
+
+func (f *fakeWG) counter() *AllowedIPsRefCounter {
+	return NewAllowedIPs(
+		func(prefix netip.Prefix, peerKey string) (string, error) {
+			if f.failAdd {
+				f.failAdd = false
+				return "", errors.New("add failed")
+			}
+			f.adds++
+			f.installed[prefix] = peerKey
+			return peerKey, nil
+		},
+		func(prefix netip.Prefix, peerKey string) error {
+			if f.failRemove {
+				f.failRemove = false
+				return errors.New("remove failed")
+			}
+			f.removes++
+			// only clear if this peer is the one installed, mirroring wg semantics
+			if f.installed[prefix] == peerKey {
+				delete(f.installed, prefix)
+			}
+			return nil
+		},
+	)
+}
+
+func mustPrefix(t *testing.T, s string) netip.Prefix {
+	t.Helper()
+	p, err := netip.ParsePrefix(s)
+	if err != nil {
+		t.Fatalf("parse prefix %q: %v", s, err)
+	}
+	return p
+}
+
+func mustIncrement(t *testing.T, c *AllowedIPsRefCounter, p netip.Prefix, peer string) Ref[string] {
+	t.Helper()
+	ref, err := c.Increment(p, peer)
+	if err != nil {
+		t.Fatalf("Increment(%v, %s): %v", p, peer, err)
+	}
+	return ref
+}
+
+func mustDecrement(t *testing.T, c *AllowedIPsRefCounter, p netip.Prefix, peer string) Ref[string] {
+	t.Helper()
+	ref, err := c.Decrement(p, peer)
+	if err != nil {
+		t.Fatalf("Decrement(%v, %s): %v", p, peer, err)
+	}
+	return ref
+}
+
+// TestAllowedIPs_SwapOnActivePeerRemoval reproduces the reported bug: two networks with the same
+// prefix routed by different peers. Removing the network whose peer is installed must hand the
+// prefix over to the surviving peer instead of leaving it on the removed one.
+func TestAllowedIPs_SwapOnActivePeerRemoval(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p := mustPrefix(t, "10.44.8.0/24")
+
+	mustIncrement(t, c, p, "peerA")
+	mustIncrement(t, c, p, "peerB")
+	// First peer wins while both are present.
+	if got := f.installed[p]; got != "peerA" {
+		t.Fatalf("expected peerA installed, got %q", got)
+	}
+
+	// Remove the active peer's network -> must swap to peerB.
+	mustDecrement(t, c, p, "peerA")
+	if got := f.installed[p]; got != "peerB" {
+		t.Fatalf("BUG: prefix stuck on removed peer, want peerB got %q", got)
+	}
+
+	// Remove the last one -> prefix gone.
+	mustDecrement(t, c, p, "peerB")
+	if _, ok := f.installed[p]; ok {
+		t.Fatalf("expected prefix removed, still installed on %q", f.installed[p])
+	}
+}
+
+// TestAllowedIPs_RemoveNonActivePeer removing a non-installed peer must not touch WireGuard.
+func TestAllowedIPs_RemoveNonActivePeer(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p := mustPrefix(t, "10.44.8.0/24")
+
+	mustIncrement(t, c, p, "peerA")
+	mustIncrement(t, c, p, "peerB")
+	removesBefore := f.removes
+
+	mustDecrement(t, c, p, "peerB")
+	if f.installed[p] != "peerA" {
+		t.Fatalf("active peer must stay peerA, got %q", f.installed[p])
+	}
+	if f.removes != removesBefore {
+		t.Fatalf("removing a non-active peer must not call wg remove")
+	}
+}
+
+// TestAllowedIPs_SamePeerMultipleRefs two references via the same peer must keep the prefix until
+// the last reference is released (the reason the per-peer count must be an int, not a set).
+func TestAllowedIPs_SamePeerMultipleRefs(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p := mustPrefix(t, "10.44.8.0/24")
+
+	mustIncrement(t, c, p, "peerA")
+	mustIncrement(t, c, p, "peerA")
+	if f.adds != 1 {
+		t.Fatalf("expected a single wg add for the same peer, got %d", f.adds)
+	}
+
+	mustDecrement(t, c, p, "peerA")
+	if f.installed[p] != "peerA" {
+		t.Fatalf("prefix must stay while a reference remains, got %q", f.installed[p])
+	}
+	if f.removes != 0 {
+		t.Fatalf("no wg remove expected while a reference remains, got %d", f.removes)
+	}
+
+	mustDecrement(t, c, p, "peerA")
+	if _, ok := f.installed[p]; ok {
+		t.Fatalf("prefix must be removed after last reference")
+	}
+}
+
+// TestAllowedIPs_RefCountAndActive checks the Ref returned to callers (used for the HA-disabled log).
+func TestAllowedIPs_RefCountAndActive(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p := mustPrefix(t, "10.44.8.0/24")
+
+	ref := mustIncrement(t, c, p, "peerA")
+	if ref.Count != 1 || ref.Out != "peerA" {
+		t.Fatalf("want {1, peerA}, got {%d, %q}", ref.Count, ref.Out)
+	}
+	ref = mustIncrement(t, c, p, "peerB")
+	if ref.Count != 2 || ref.Out != "peerA" {
+		t.Fatalf("want {2, peerA}, got {%d, %q}", ref.Count, ref.Out)
+	}
+}
+
+// TestAllowedIPs_Flush removes everything installed and clears the counter.
+func TestAllowedIPs_Flush(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p1 := mustPrefix(t, "10.44.8.0/24")
+	p2 := mustPrefix(t, "10.44.9.0/24")
+
+	mustIncrement(t, c, p1, "peerA")
+	mustIncrement(t, c, p2, "peerB")
+
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(f.installed) != 0 {
+		t.Fatalf("expected all prefixes removed, got %v", f.installed)
+	}
+	// After flush, a fresh increment must add again.
+	mustIncrement(t, c, p1, "peerC")
+	if f.installed[p1] != "peerC" {
+		t.Fatalf("counter not reset after flush")
+	}
+}
+
+// TestAllowedIPs_SelfHealAfterSwapAddError ensures a failed add during a swap does not permanently
+// strand the prefix: the next Decrement (or Increment) must retry and install a surviving peer.
+func TestAllowedIPs_SelfHealAfterSwapAddError(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p := mustPrefix(t, "10.44.8.0/24")
+
+	mustIncrement(t, c, p, "peerA")
+	mustIncrement(t, c, p, "peerB")
+	mustIncrement(t, c, p, "peerC")
+
+	// Removing the active peerA triggers a swap to a survivor; make the add fail once.
+	f.failAdd = true
+	if _, err := c.Decrement(p, "peerA"); err == nil {
+		t.Fatalf("expected error from failed swap add")
+	}
+	if _, ok := f.installed[p]; ok {
+		t.Fatalf("nothing should be installed after a failed swap add, got %q", f.installed[p])
+	}
+
+	// A later Decrement of a non-active survivor must retry the hand-off (self-heal), not stay stuck.
+	ref := mustDecrement(t, c, p, "peerC")
+	if got := f.installed[p]; got == "" {
+		t.Fatalf("self-heal failed: prefix left unrouted after add recovered")
+	}
+	if ref.Out == "" {
+		t.Fatalf("expected an active peer after self-heal, got empty")
+	}
+}
+
+// TestAllowedIPs_SelfHealAfterRemoveError ensures a failed remove during a swap is retried instead
+// of leaving e.active stuck on a peer that no longer holds references.
+func TestAllowedIPs_SelfHealAfterRemoveError(t *testing.T) {
+	f := newFakeWG()
+	c := f.counter()
+	p := mustPrefix(t, "10.44.8.0/24")
+
+	mustIncrement(t, c, p, "peerA")
+	mustIncrement(t, c, p, "peerB")
+
+	// Releasing active peerA must detach it (remove) then add peerB; fail the remove once.
+	f.failRemove = true
+	if _, err := c.Decrement(p, "peerA"); err == nil {
+		t.Fatalf("expected error from failed remove")
+	}
+
+	// Next Decrement of the non-active survivor retries: removes stale peerA, installs peerB.
+	mustDecrement(t, c, p, "peerB")
+	// peerB had only one ref, so after retry the prefix is fully released.
+	if _, ok := f.installed[p]; ok {
+		t.Fatalf("expected prefix released after self-heal, still on %q", f.installed[p])
+	}
+}

client/internal/routemanager/refcounter/types.go

@@ -5,5 +5,7 @@ import "net/netip"
 // RouteRefCounter is a Counter for Route, it doesn't take any input on Increment and doesn't use any output on Decrement
 type RouteRefCounter = Counter[netip.Prefix, struct{}, struct{}]
 
-// AllowedIPsRefCounter is a Counter for AllowedIPs, it takes a peer key on Increment and passes it back to Decrement
-type AllowedIPsRefCounter = Counter[netip.Prefix, string, string]
+// AllowedIPsRefCounter tracks WireGuard AllowedIPs per prefix. Unlike the generic Counter it is peer-aware:
+// a prefix can be claimed by several peers at once and WireGuard allows a given prefix on exactly one peer,
+// so the counter records the per-peer reference count and swaps the installed peer when the active one is released.
+// See allowedips.go.

client/internal/routemanager/static/route.go

@@ -15,6 +15,11 @@ type Route struct {
 	route                *route.Route
 	routeRefCounter      *refcounter.RouteRefCounter
 	allowedIPsRefcounter *refcounter.AllowedIPsRefCounter
+	// currentPeerKey is the routing peer this watcher currently has the prefix installed on
+	// (the HA winner elected by the watcher). It can differ from route.Peer and change on
+	// failover, so it is recorded on AddAllowedIPs and used on RemoveAllowedIPs to decrement
+	// the exact peer that was incremented.
+	currentPeerKey string
 }
 
 func NewRoute(params common.HandlerParams) *Route {
@@ -52,12 +57,15 @@ func (r *Route) AddAllowedIPs(peerKey string) error {
 			ref.Out,
 		)
 	}
+	r.currentPeerKey = peerKey
 	return nil
 }
 
 func (r *Route) RemoveAllowedIPs() error {
-	if _, err := r.allowedIPsRefcounter.Decrement(r.route.Network); err != nil {
-		return err
+	var err error
+	if _, decErr := r.allowedIPsRefcounter.Decrement(r.route.Network, r.currentPeerKey); decErr != nil {
+		err = fmt.Errorf("remove allowed IP %s: %w", r.route.Network, decErr)
 	}
-	return nil
+	r.currentPeerKey = ""
+	return err
 }

e2e/agentnetwork/chat_test.go

@@ -91,7 +91,7 @@ func availableProviders() []providerCase {
 		if region == "" {
 			region = "us-east-1"
 		}
-		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: "us.anthropic.claude-haiku-4-5", kind: harness.WireMessages})
+		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: "us.anthropic.claude-haiku-4-5", kind: harness.WireBedrock})
 	}
 	return ps
 }
@@ -224,9 +224,12 @@ func TestProvidersMatrix(t *testing.T) {
 				var c int
 				var b string
 				var cerr error
-				if pc.kind == harness.WireVertex {
+				switch pc.kind {
+				case harness.WireVertex:
 					c, b, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model, "Reply with exactly: pong", sessionID)
-				} else {
+				case harness.WireBedrock:
+					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, "Reply with exactly: pong", sessionID)
+				default:
 					c, b, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
 				}
 				if cerr == nil {

e2e/agentnetwork/guardrail_test.go

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

e2e/harness/agentnetwork.go

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

e2e/harness/client.go

@@ -194,6 +194,11 @@ const (
 	// WireVertex is the Anthropic-on-Vertex rawPredict shape: the client posts
 	// the full Vertex model path and the proxy mints the SA OAuth token.
 	WireVertex = "vertex"
+	// WireBedrock is the native AWS Bedrock InvokeModel shape: the model id
+	// travels in the URL path (/model/{id}/invoke), not the body, so the proxy
+	// routes by path. This is what a Bedrock SDK client sends and the shape the
+	// model-allowlist guardrail must enforce.
+	WireBedrock = "bedrock"
 )
 
 // Chat issues a chat-completion POST to the agent-network endpoint over the
@@ -226,6 +231,17 @@ func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
+// Bedrock issues a native AWS Bedrock InvokeModel POST over the tunnel. The
+// model id is carried in the request path (/model/{id}/invoke), so the proxy
+// routes by path; the body uses the bedrock anthropic_version rather than a
+// model field. A non-empty sessionID is sent as the universal x-session-id
+// header the proxy records.
+func (cl *Client) Bedrock(ctx context.Context, endpoint, proxyIP, model, prompt, sessionID string) (int, string, error) {
+	path := "/model/" + model + "/invoke"
+	body := fmt.Sprintf(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":64,"messages":[{"role":"user","content":%q}]}`, prompt)
+	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
+}
+
 // withSessionID appends the x-session-id header when sessionID is non-empty.
 func withSessionID(headers []string, sessionID string) []string {
 	if sessionID == "" {

management/internals/modules/peers/manager.go

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

management/main.go

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

management/server/agentnetwork_proxypeer_restart_test.go

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

management/server/types/networkmap_components.go

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

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

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

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

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

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

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