[e2e] Retry the endpoint probe when the tunnel is not up yet

ResolveProxyIP exists to wake the lazy proxy peer, and it retried only
curl exit 6 — DNS. The wake-up attempt that arrives before WireGuard has
brought the tunnel up fails with exit 7 instead, and that returned
immediately:

  no HTTP response from vast-azalea.netbird.local: exit status 7
  (curl: (7) Failed to connect ... after 0 ms)

So the one function whose job is to tolerate a not-yet-ready endpoint
failed on the readiness state it was written for, one second after the
client container reported ready. Retry both exit codes within the same
window; anything else would still be failing when the window closed and
still fails immediately.

Raise the access-log ingest window to 60s for the same reason. The proxy
streams each entry with a 10s send timeout of its own, so 30s left
barely three attempts of headroom before a test that had already got its
200 was failed for a row still in flight.

agent-network/README.md

@@ -40,6 +40,35 @@ You can then use this private endpoint to configure your AI agents, whether that
 Full step-by-step setup:
 **https://docs.netbird.io/agent-network/quickstart**
 
+## Client settings that don't follow the endpoint
+
+Most of an agent's traffic follows the base URL you hand it, but a few
+client-side checks call their vendor directly and never reach the proxy. On a
+network that blocks direct egress they fail even though inference works, so
+they are worth setting once when you roll the endpoint out.
+
+For Claude Code:
+
+- **Fast mode** checks availability against `api.anthropic.com` rather than the
+  configured base URL. Set `CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK=1` when the
+  agent authenticates with `ANTHROPIC_AUTH_TOKEN` alone (the usual shape when
+  the proxy injects the real provider key) or when a TLS-inspecting proxy
+  answers the check itself. Set
+  `CLAUDE_CODE_SKIP_FAST_MODE_NETWORK_ERRORS=1` when the network refuses the
+  connection outright. Fast mode is an Anthropic-API feature, so it is
+  unavailable on a Bedrock- or Vertex-backed endpoint whatever you set.
+- **Model discovery** is off by default. Set
+  `CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1` for the picker to list the
+  models your policies authorise; the proxy filters the response to that set.
+  The client gives discovery a three-second budget and treats any redirect as
+  a failure, so the endpoint must serve `/v1/models` directly.
+- **The WebFetch domain safety check** also calls `api.anthropic.com` directly
+  and is unaffected by the variables above.
+
+Allowing direct egress to `api.anthropic.com` covers the network cases but not
+the credential one, where the check reaches Anthropic and is rejected because
+the agent presents a proxy-issued key.
+
 ## Architecture 
 
 Agent Network is built on two existing NetBird capabilities:

client/internal/profilemanager/state.go

@@ -87,9 +87,10 @@ func (pm *ProfileManager) SetActiveProfileState(state *ProfileState) error {
 
 // RemoveProfileState deletes the per-profile state file (which holds the
 // account email used for the SSO login hint and the UI display). Called after
-// a successful logout so a logged-out profile no longer shows a stale account
-// email. The state file only stores the email, so deleting it is equivalent to
-// clearing it; the next SSO login recreates it. A missing file is not an error.
+// profile removal; logout keeps the file so the next login can pass the email
+// as the login_hint. The state file only stores the email, so deleting it is
+// equivalent to clearing it; the next SSO login recreates it. A missing file
+// is not an error.
 func (pm *ProfileManager) RemoveProfileState(profileName string) error {
 	configDir, err := getConfigDir()
 	if err != nil {

client/internal/routemanager/systemops/routeselection_windows_test.go

@@ -0,0 +1,82 @@
+//go:build windows
+
+package systemops
+
+import (
+	"math"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestSortRouteCandidates(t *testing.T) {
+	tests := []struct {
+		name       string
+		candidates []candidateRoute
+		wantOrder  []uint32
+	}{
+		{
+			name: "longest prefix wins over metrics",
+			candidates: []candidateRoute{
+				{interfaceIndex: 1, prefixLength: 0, routeMetric: 0, interfaceMetric: 5},
+				{interfaceIndex: 2, prefixLength: 24, routeMetric: 100, interfaceMetric: 50},
+			},
+			wantOrder: []uint32{2, 1},
+		},
+		{
+			// Windows ranks equal-length prefixes by route metric + interface metric,
+			// so a higher route metric on a low metric interface can still win.
+			name: "combined metric beats route metric alone",
+			candidates: []candidateRoute{
+				{interfaceIndex: 8, prefixLength: 0, routeMetric: 0, interfaceMetric: 100},
+				{interfaceIndex: 5, prefixLength: 0, routeMetric: 10, interfaceMetric: 5},
+			},
+			wantOrder: []uint32{5, 8},
+		},
+		{
+			name: "lower combined metric wins",
+			candidates: []candidateRoute{
+				{interfaceIndex: 5, prefixLength: 0, routeMetric: 300, interfaceMetric: 5},
+				{interfaceIndex: 8, prefixLength: 0, routeMetric: 0, interfaceMetric: 100},
+			},
+			wantOrder: []uint32{8, 5},
+		},
+		{
+			name: "equal combined metric falls back to route metric",
+			candidates: []candidateRoute{
+				{interfaceIndex: 1, prefixLength: 0, routeMetric: 20, interfaceMetric: 10},
+				{interfaceIndex: 2, prefixLength: 0, routeMetric: 5, interfaceMetric: 25},
+			},
+			wantOrder: []uint32{2, 1},
+		},
+		{
+			// The metrics are uint32 on the Windows side, so the sum must not wrap.
+			name: "combined metric beyond the uint32 range",
+			candidates: []candidateRoute{
+				{interfaceIndex: 1, prefixLength: 0, routeMetric: math.MaxUint32, interfaceMetric: 5},
+				{interfaceIndex: 2, prefixLength: 0, routeMetric: math.MaxUint32 - 10, interfaceMetric: 5},
+			},
+			wantOrder: []uint32{2, 1},
+		},
+		{
+			name: "unknown interface metric ranks on route metric only",
+			candidates: []candidateRoute{
+				{interfaceIndex: 1, prefixLength: 0, routeMetric: 30, interfaceMetric: -1},
+				{interfaceIndex: 2, prefixLength: 0, routeMetric: 5, interfaceMetric: 10},
+			},
+			wantOrder: []uint32{2, 1},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			sortRouteCandidates(tt.candidates)
+
+			got := make([]uint32, 0, len(tt.candidates))
+			for _, c := range tt.candidates {
+				got = append(got, c.interfaceIndex)
+			}
+			assert.Equal(t, tt.wantOrder, got)
+		})
+	}
+}

client/internal/routemanager/systemops/systemops_windows.go

@@ -882,26 +882,40 @@ func getInterfaceMetric(interfaceIndex uint32, family int16) int {
 	return int(ipInterfaceRow.Metric)
 }
 
-// sortRouteCandidates sorts route candidates by priority: prefix length -> route metric -> interface metric
+// sortRouteCandidates sorts route candidates by priority: prefix length -> combined metric -> route metric.
+// Windows prefers the longest matching prefix and, among prefixes of the same length, the lowest metric, see
+// https://learn.microsoft.com/en-us/windows-hardware/customize/desktop/unattend/microsoft-windows-tcpip-interfaces-interface-routes-route-metric
 func sortRouteCandidates(candidates []candidateRoute) {
 	sort.Slice(candidates, func(i, j int) bool {
 		if candidates[i].prefixLength != candidates[j].prefixLength {
 			return candidates[i].prefixLength > candidates[j].prefixLength
 		}
-		if candidates[i].routeMetric != candidates[j].routeMetric {
-			return candidates[i].routeMetric < candidates[j].routeMetric
+		mi, mj := combinedMetric(candidates[i]), combinedMetric(candidates[j])
+		if mi != mj {
+			return mi < mj
 		}
-		return candidates[i].interfaceMetric < candidates[j].interfaceMetric
+		return candidates[i].routeMetric < candidates[j].routeMetric
 	})
 }
 
+// combinedMetric returns the effective metric Windows uses to rank routes with an equal prefix length:
+// the sum of the route metric and the metric of the interface the route is on, see
+// https://learn.microsoft.com/en-us/windows-server/networking/technologies/network-subsystem/net-sub-interface-metric
+// An unknown interface metric contributes nothing.
+func combinedMetric(candidate candidateRoute) uint64 {
+	if candidate.interfaceMetric < 0 {
+		return uint64(candidate.routeMetric)
+	}
+	return uint64(candidate.routeMetric) + uint64(candidate.interfaceMetric)
+}
+
 // GetBestInterface finds the best interface for reaching a destination,
 // excluding the VPN interface to avoid routing loops.
 //
 // Route selection priority:
 // 1. Longest prefix match (most specific route)
-// 2. Lowest route metric
-// 3. Lowest interface metric
+// 2. Lowest combined metric (route metric + interface metric)
+// 3. Lowest route metric.
 func GetBestInterface(dest netip.Addr, vpnIntf string) (*net.Interface, error) {
 	var skipInterfaceIndex int
 	if vpnIntf != "" {
@@ -925,7 +939,6 @@ func GetBestInterface(dest netip.Addr, vpnIntf string) (*net.Interface, error) {
 		return nil, fmt.Errorf("no route to %s", dest)
 	}
 
-	// Sort routes: prefix length -> route metric -> interface metric
 	sortRouteCandidates(candidates)
 
 	for _, candidate := range candidates {

client/internal/routemanager/systemops/v6route_linux_test.go

@@ -5,6 +5,7 @@ package systemops
 import (
 	"errors"
 	"net"
+	"net/netip"
 	"syscall"
 	"testing"
 
@@ -29,6 +30,7 @@ func ensureIPv6DefaultRoute(t *testing.T) {
 	}
 	if err := netlink.RouteAdd(route); err != nil {
 		if errors.Is(err, syscall.EEXIST) {
+			requireUsableIPv6Nexthop(t)
 			return
 		}
 		t.Skipf("install IPv6 fallback default route: %v", err)
@@ -38,4 +40,36 @@ func ensureIPv6DefaultRoute(t *testing.T) {
 			t.Logf("delete IPv6 fallback default route: %v", err)
 		}
 	})
+
+	requireUsableIPv6Nexthop(t)
+}
+
+// requireUsableIPv6Nexthop skips the test unless the resolved IPv6 default
+// nexthop can actually carry a route. Installing the default route succeeding
+// does not imply the kernel accepts it as a nexthop for a concrete prefix.
+func requireUsableIPv6Nexthop(t *testing.T) {
+	t.Helper()
+
+	nexthop, err := GetNextHop(netip.IPv6Unspecified())
+	if err != nil {
+		t.Skipf("resolve IPv6 default nexthop: %v", err)
+	}
+
+	probe := &netlink.Route{
+		Scope:  netlink.SCOPE_UNIVERSE,
+		Table:  syscall.RT_TABLE_MAIN,
+		Family: netlink.FAMILY_V6,
+		Dst:    &net.IPNet{IP: net.ParseIP("100::64"), Mask: net.CIDRMask(128, 128)},
+	}
+	require.NoError(t, addNextHop(nexthop, probe), "build IPv6 probe route")
+
+	switch err := netlink.RouteAdd(probe); {
+	case err == nil:
+		if err := netlink.RouteDel(probe); err != nil && !errors.Is(err, syscall.ESRCH) {
+			t.Logf("delete IPv6 probe route: %v", err)
+		}
+	case errors.Is(err, syscall.EEXIST):
+	default:
+		t.Skipf("IPv6 nexthop %s unusable for route installation: %v", nexthop, err)
+	}
 }

client/ui/authsession/service.go

@@ -6,9 +6,11 @@ import (
 	"context"
 	"time"
 
+	log "github.com/sirupsen/logrus"
 	"google.golang.org/grpc/codes"
 	gstatus "google.golang.org/grpc/status"
 
+	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/proto"
 )
 
@@ -60,9 +62,19 @@ func (s *Session) RequestExtend(ctx context.Context, p ExtendStartParams) (Exten
 
 	// a request from the UI implies a graphical session, which the daemon cannot detect itself
 	req := &proto.RequestExtendAuthSessionRequest{HasGraphicalSession: true}
-	if p.Hint != "" {
-		h := p.Hint
-		req.Hint = &h
+	hint := p.Hint
+	if hint == "" {
+		pm := profilemanager.NewProfileManager()
+		if active, perr := pm.GetActiveProfile(); perr != nil {
+			log.Debugf("failed to get active profile for login hint: %v", perr)
+		} else if state, serr := pm.GetProfileState(active.ID); serr != nil {
+			log.Debugf("failed to get profile state for login hint: %v", serr)
+		} else {
+			hint = state.Email
+		}
+	}
+	if hint != "" {
+		req.Hint = &hint
 	}
 
 	resp, err := cli.RequestExtendAuthSession(ctx, req)

client/ui/services/connection.go

@@ -123,8 +123,16 @@ func (s *Connection) Login(ctx context.Context, p LoginParams) (LoginResult, err
 	if p.PreSharedKey != "" {
 		req.OptionalPreSharedKey = ptrStr(p.PreSharedKey)
 	}
-	if p.Hint != "" {
-		req.Hint = ptrStr(p.Hint)
+	hint := p.Hint
+	if hint == "" && profileID != "" {
+		if state, serr := profilemanager.NewProfileManager().GetProfileState(profilemanager.ID(profileID)); serr == nil {
+			hint = state.Email
+		} else {
+			log.Debugf("failed to get profile state for login hint: %v", serr)
+		}
+	}
+	if hint != "" {
+		req.Hint = ptrStr(hint)
 	}
 
 	resp, err := cli.Login(ctx, req)
@@ -228,16 +236,6 @@ func (s *Connection) Logout(ctx context.Context, p LogoutParams) error {
 		return s.classifyDaemonError(err)
 	}
 
-	// The daemon runs as root and can't reach the user-owned per-profile state
-	// file holding the account email (see Profiles.List), so clear the stale
-	// email here; the next SSO login recreates it.
-	if p.ProfileName != "" {
-		if err := profilemanager.NewProfileManager().RemoveProfileState(p.ProfileName); err != nil {
-			// Non-fatal: the logout itself succeeded.
-			log.Warnf("failed to remove profile state for %s: %v", p.ProfileName, err)
-		}
-	}
-
 	return nil
 }
 
@@ -261,7 +259,7 @@ func (s *Connection) waitSSOLogin(ctx context.Context, p WaitSSOParams) (string,
 
 	// Persist the account email the same way the CLI does after its own
 	// WaitSSOLogin: the daemon returns it but cannot store it, since it runs as
-	// root and the per-profile state file is user-owned (see Logout below).
+	// root and the per-profile state file is user-owned (see Profiles.List).
 	// Without this the profile has no email, so Profiles.List shows no account
 	// and later logins and session extends go out without a login_hint —
 	// leaving the IdP to guess which account was meant.

client/ui/services/profile.go

@@ -162,8 +162,9 @@ func (s *Profiles) Remove(ctx context.Context, p ProfileRef) error {
 	}
 
 	// The daemon deletes what it owns but runs as root, so it leaves the
-	// user-owned state file holding the account email behind (same split as
-	// Connection.Logout). Legacy profiles are keyed by name rather than by a
+	// user-owned state file holding the account email behind. Logout keeps the
+	// email on purpose so later logins can pass it as the login_hint; profile
+	// removal is what deletes it. Legacy profiles are keyed by name rather than by a
 	// generated ID, so a recreated profile of the same name would inherit the
 	// deleted one's email and offer it as the login_hint.
 	//

e2e/agentnetwork/custom_pricing_test.go

@@ -23,9 +23,10 @@ import (
 // model the client asks for. The proxy prices off the REQUEST model, not the
 // upstream response model, so a made-up model id billed at operator rates lets
 // these tests assert exact costs without a real vendor key.
+// Sourced from the harness so the counts can't drift from the mock's config.
 const (
-	vllmPromptTokens     = 11
-	vllmCompletionTokens = 2
+	vllmPromptTokens     = harness.VLLMChatInputTokens
+	vllmCompletionTokens = harness.VLLMChatOutputTokens
 )
 
 // pricedEnv is a connected single-provider agent-network deployment pointed at
@@ -169,23 +170,48 @@ func chatOnce(t *testing.T, ctx context.Context, env pricedEnv, model, sessionID
 	return body
 }
 
-// findAccessLogBySession polls the access-log page for the row carrying sessionID.
-func findAccessLogBySession(t *testing.T, ctx context.Context, sessionID string) api.AgentNetworkAccessLog {
-	t.Helper()
-	var row api.AgentNetworkAccessLog
-	require.Eventually(t, func() bool {
-		logs, lerr := srv.ListAccessLogs(ctx)
-		if lerr != nil {
-			return false
-		}
-		for _, r := range logs.Data {
-			if r.SessionId != nil && *r.SessionId == sessionID {
-				row = r
-				return true
+// accessLogIngestWindow is how long a single request's access-log row is given
+// to appear before the caller gives up on it.
+// accessLogIngestWindow bounds how long a row may take to appear after its
+// request returned. The proxy streams each entry to management with a 10s send
+// timeout of its own, so a request whose send hits one full timeout and is
+// retried has not yet missed anything real — 30s left barely three send
+// attempts of headroom and lost the race on a loaded runner.
+const accessLogIngestWindow = 60 * time.Second
+
+// lookupAccessLogBySession polls the access-log page for the row carrying
+// sessionID and reports whether it arrived within the window. It never fails
+// the test: callers that can recover — by firing a fresh request under a new
+// session — need to see the miss rather than die on it.
+func lookupAccessLogBySession(ctx context.Context, sessionID string, within time.Duration) (api.AgentNetworkAccessLog, bool) {
+	deadline := time.Now().Add(within)
+	for {
+		if logs, lerr := srv.ListAccessLogs(ctx); lerr == nil {
+			for _, r := range logs.Data {
+				if r.SessionId != nil && *r.SessionId == sessionID {
+					return r, true
+				}
 			}
 		}
-		return false
-	}, 30*time.Second, 2*time.Second, "session id %q must be recorded in an access-log row", sessionID)
+		if time.Now().After(deadline) {
+			return api.AgentNetworkAccessLog{}, false
+		}
+		select {
+		case <-ctx.Done():
+			return api.AgentNetworkAccessLog{}, false
+		case <-time.After(2 * time.Second):
+		}
+	}
+}
+
+// findAccessLogBySession polls the access-log page for the row carrying
+// sessionID, failing the test if it never lands. Use it for a request whose row
+// must exist; where a missing row is a recoverable race, use
+// lookupAccessLogBySession and retry.
+func findAccessLogBySession(t *testing.T, ctx context.Context, sessionID string) api.AgentNetworkAccessLog {
+	t.Helper()
+	row, ok := lookupAccessLogBySession(ctx, sessionID, accessLogIngestWindow)
+	require.True(t, ok, "session id %q must be recorded in an access-log row", sessionID)
 	return row
 }
 
@@ -319,6 +345,11 @@ func TestPriceChangeUpdatesRecordedCost(t *testing.T) {
 		outRateA    = 0.020
 		inRateB     = 0.050 // 5x / 4x the original, so a repriced row is unmistakable
 		outRateB    = 0.080
+		// Per-attempt ingest wait, shorter than the default so a request that
+		// produces no row costs one retry rather than most of the budget, and an
+		// overall deadline long enough to hold several attempts.
+		repriceIngestWindow = 20 * time.Second
+		repriceDeadline     = 180 * time.Second
 	)
 
 	env := provisionPricedProvider(t, ctx, "reprice", []api.AgentNetworkProviderModel{
@@ -353,10 +384,15 @@ func TestPriceChangeUpdatesRecordedCost(t *testing.T) {
 	// reading its cost, so an un-ingested row is never mistaken for "still rate A".
 	// The expected new input cost is unmistakably higher than rate A, so a
 	// lingering old-rate row can't satisfy the check.
+	//
+	// Every way an iteration can come up short — the request failing, its row not
+	// landing, or the row still carrying rate A — is a symptom of the same
+	// in-flight rebuild, so each one retries under a fresh session rather than
+	// ending the test. Only the outer deadline is fatal.
 	wantInputB := float64(vllmPromptTokens) / 1000 * inRateB
 	var repriced api.AgentNetworkAccessLog
 	var lastSession string
-	deadline := time.Now().Add(90 * time.Second)
+	deadline := time.Now().Add(repriceDeadline)
 	for time.Now().Before(deadline) {
 		lastSession = fmt.Sprintf("e2e-session-reprice-b-%d", time.Now().UnixNano())
 		code, _, cerr := env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireChat, customModel, "Reply with exactly: pong", lastSession)
@@ -364,7 +400,15 @@ func TestPriceChangeUpdatesRecordedCost(t *testing.T) {
 			time.Sleep(5 * time.Second)
 			continue
 		}
-		row := findAccessLogBySession(t, ctx, lastSession)
+		row, ok := lookupAccessLogBySession(ctx, lastSession, repriceIngestWindow)
+		if !ok {
+			// No row for this request. The provider update rebuilds the proxy's
+			// middleware chain, and a request served mid-rebuild can complete
+			// without a resolved provider — 200 to the caller, nothing to
+			// attribute, so no row is ever written for it. Fire another one.
+			t.Logf("no access-log row for session %q within %s; retrying under a fresh session", lastSession, repriceIngestWindow)
+			continue
+		}
 		if inDelta(row.InputCostUsd, wantInputB, 1e-6) {
 			repriced = row
 			break
@@ -630,3 +674,47 @@ func inDelta(a, b, tol float64) bool {
 	}
 	return d <= tol
 }
+
+// TestCustomDatedModelKeepsItsOwnPrice covers the review fix that anchored the
+// release-date fallback to Claude ids. Pricing looks every model up through
+// that helper, so while it matched a bare trailing date any operator id ending
+// in eight digits inherited the rate of its undated sibling — a silent
+// mis-bill on models NetBird knows nothing about.
+func TestCustomDatedModelKeepsItsOwnPrice(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	const (
+		baseModel  = "internal-llm"
+		datedModel = "internal-llm-20250101"
+		baseIn     = 0.010
+		baseOut    = 0.020
+		// An order of magnitude apart, so a row billed at the wrong entry is
+		// unmistakable rather than a rounding argument.
+		datedIn  = 0.100
+		datedOut = 0.200
+	)
+
+	env := provisionPricedProvider(t, ctx, "customdated", []api.AgentNetworkProviderModel{
+		{Id: baseModel, InputPer1k: baseIn, OutputPer1k: baseOut},
+		{Id: datedModel, InputPer1k: datedIn, OutputPer1k: datedOut},
+	})
+
+	t.Run("the undated id bills at its own rate", func(t *testing.T) {
+		session := fmt.Sprintf("e2e-session-customdated-base-%d", time.Now().UnixNano())
+		chatOnce(t, ctx, env, baseModel, session)
+		assertOpenAICostAtRates(t, findAccessLogBySession(t, ctx, session), baseIn, baseOut)
+	})
+
+	t.Run("the dated id keeps its own rate", func(t *testing.T) {
+		session := fmt.Sprintf("e2e-session-customdated-dated-%d", time.Now().UnixNano())
+		chatOnce(t, ctx, env, datedModel, session)
+		row := findAccessLogBySession(t, ctx, session)
+		assertOpenAICostAtRates(t, row, datedIn, datedOut)
+
+		// Spelled out because it is the regression: inheriting the sibling's
+		// rate would bill this request at a tenth of its price.
+		assert.Greater(t, row.InputCostUsd, float64(vllmPromptTokens)/1000*baseIn*2,
+			"a custom dated id must not inherit the undated entry's rate")
+	})
+}

e2e/agentnetwork/discovery_live_test.go

@@ -0,0 +1,400 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"sort"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	sharedllm "github.com/netbirdio/netbird/shared/llm"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// TestLiveModelDiscovery drives model discovery against the REAL vendor
+// endpoints — OpenAI, Anthropic, Bedrock and Vertex — rather than the mock.
+//
+// The mock upstream proves the filter's mechanics: it advertises ids we chose,
+// so a listing narrowing to the ones we authorised is arithmetic we already
+// controlled both sides of. What it cannot prove is that the filter survives
+// contact with a real catalogue — ids we never enumerated, dated builds whose
+// suffix the vendor picks, surfaces that answer a listing request with
+// something other than a listing. That is what this covers, and it is the part
+// a QA engineer would otherwise have to walk through by hand.
+//
+// One proxy serves every case. Each provider gets its own group, policy and
+// client, because a model-less request matches exactly ONE route
+// (matchModelless): with two providers authorised for the same caller, the
+// listing would go to whichever won the tiebreak and the other would go
+// untested. Group-scoping the caller makes each provider the only candidate
+// for its own client.
+func TestLiveModelDiscovery(t *testing.T) {
+	cases := liveDiscoveryCases()
+	if len(cases) == 0 {
+		t.Skip("no provider keys set; source ~/.llm-keys to run live model discovery")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Minute)
+	defer cancel()
+
+	t.Logf("[discovery] live matrix: %s", strings.Join(caseNames(cases), ", "))
+
+	// Provision every provider, group and policy before the proxy starts: the
+	// proxy takes a configuration snapshot at connect time and does not
+	// reconcile provider changes made afterwards.
+	keys := make(map[string]string, len(cases))
+	for i := range cases {
+		keys[cases[i].name] = provisionLiveDiscovery(t, ctx, &cases[i])
+	}
+
+	endpoint, firstIP, firstClient, px := connectClient(t, ctx, "disc-live", keys[cases[0].name])
+	clients := map[string]*harness.Client{cases[0].name: firstClient}
+	ips := map[string]string{cases[0].name: firstIP}
+	for _, tc := range cases[1:] {
+		cl := joinClient(t, ctx, px, endpoint, keys[tc.name])
+		ip, err := cl.ResolveProxyIP(ctx, endpoint)
+		require.NoError(t, err, "resolve endpoint from the %s client", tc.name)
+		clients[tc.name] = cl
+		ips[tc.name] = ip
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			runLiveDiscoveryCase(t, ctx, tc, clients[tc.name], endpoint, ips[tc.name])
+		})
+	}
+}
+
+// discoveryOutcome is what a discovery request must produce end to end. The
+// three are genuinely different contracts, not degrees of success: only the
+// first puts a bounded listing in front of the caller.
+type discoveryOutcome int
+
+const (
+	// outcomeFiltered: the proxy routes the request and bounds the response to
+	// what the caller may use.
+	outcomeFiltered discoveryOutcome = iota
+	// outcomeDenied: no provider of this shape can serve the surface, so the
+	// proxy refuses rather than rewriting the request onto an upstream that
+	// would 404 it. The caller gets a NetBird error, not a vendor one.
+	outcomeDenied
+	// outcomeUpstreamNoListing: the proxy routes the request to the configured
+	// upstream, and the vendor does not implement the endpoint there. Proxy
+	// side correct, product side a dead end — see the Bedrock case.
+	outcomeUpstreamNoListing
+)
+
+// liveDiscoveryCase is one provider's discovery surface and what the proxy
+// must make of it.
+type liveDiscoveryCase struct {
+	name      string
+	catalogID string
+	upstream  string
+	apiKey    string
+
+	// path is the discovery endpoint the client calls. Not every surface uses
+	// /v1/models: Bedrock lists inference profiles instead.
+	path string
+	// headers the vendor requires on a bare GET (Anthropic versions its API
+	// through a header, and rejects a request without one).
+	headers []string
+
+	// models the provider record enumerates. Empty models a gateway record,
+	// which enumerates nothing and claims everything.
+	models []string
+	// allowlist, when non-empty, is a guardrail narrowing the policy below the
+	// provider's own enumeration — the second of the two bounds discovery
+	// applies, and the only one a provider record alone cannot demonstrate.
+	allowlist []string
+
+	// outcome is what this surface must produce end to end.
+	outcome discoveryOutcome
+
+	// permitted is every id allowed to survive filtering, in the form the
+	// provider record registers it. A surviving id counts as permitted when it
+	// matches one of these outright or after Anthropic date-normalisation.
+	permitted []string
+	// wantHidden are ids the upstream is known to advertise and the bound must
+	// remove. Only set where we enumerate the model ourselves, so the
+	// expectation cannot rot when a vendor changes its catalogue.
+	wantHidden []string
+}
+
+// liveDiscoveryCases builds the matrix from whichever provider credentials are
+// present, mirroring availableProviders' env-var gating so a partial key set
+// still yields partial coverage.
+func liveDiscoveryCases() []liveDiscoveryCase {
+	var cases []liveDiscoveryCase
+
+	// OpenAI enumerates TWO real models and the policy permits one. That is
+	// the only case here where both bounds are observable at once: the
+	// upstream advertises dozens of ids, the provider record cuts them to two,
+	// and the guardrail cuts those to one.
+	if k := os.Getenv("OPENAI_TOKEN"); k != "" {
+		cases = append(cases, liveDiscoveryCase{
+			name: "openai", catalogID: "openai_api", upstream: "https://api.openai.com", apiKey: k,
+			path:       "/v1/models",
+			models:     []string{"gpt-4o-mini", "gpt-4o"},
+			allowlist:  []string{"gpt-4o-mini"},
+			outcome:    outcomeFiltered,
+			permitted:  []string{"gpt-4o-mini"},
+			wantHidden: []string{"gpt-4o"},
+		})
+	}
+
+	// Anthropic is the surface Claude Code actually calls. Its listing returns
+	// DATED build ids (claude-haiku-4-5-20251001) while the provider record
+	// registers the undated id, so this is the case that proves the filter's
+	// date-normalisation against ids the vendor chose rather than ids we wrote.
+	if k := os.Getenv("ANTHROPIC_TOKEN"); k != "" {
+		cases = append(cases, liveDiscoveryCase{
+			name: "anthropic", catalogID: "anthropic_api", upstream: "https://api.anthropic.com", apiKey: k,
+			path:      "/v1/models",
+			headers:   []string{"anthropic-version: 2023-06-01"},
+			models:    []string{"claude-haiku-4-5"},
+			outcome:   outcomeFiltered,
+			permitted: []string{"claude-haiku-4-5"},
+		})
+	}
+
+	// Bedrock lists inference profiles, not models: matchModelless routes
+	// /inference-profiles to a Bedrock route and refuses /v1/models for one.
+	//
+	// The request reaches AWS and AWS refuses it — bedrock-runtime answers
+	// <UnknownOperationException/>, because ListInferenceProfiles is a CONTROL
+	// PLANE operation served by bedrock.<region>.amazonaws.com, not the runtime
+	// host. A provider record carries one upstream and it has to be the runtime
+	// host for InvokeModel to work, so no Bedrock record can serve a listing as
+	// the model stands today.
+	//
+	// The mock upstream hides this entirely: it answers /inference-profiles on
+	// the same listener as everything else, so the routing test passes there
+	// while the real endpoint 404s. That is the whole reason this file exists,
+	// so the case is kept, asserting what actually happens.
+	if k := os.Getenv("AWS_BEARER_TOKEN_BEDROCK"); k != "" {
+		region := os.Getenv("AWS_REGION")
+		if region == "" {
+			region = "eu-central-1"
+		}
+		model := os.Getenv("AWS_BEDROCK_MODEL")
+		if model == "" {
+			model = "global.anthropic.claude-sonnet-4-6"
+		}
+		cases = append(cases, liveDiscoveryCase{
+			name: "bedrock", catalogID: "bedrock_api",
+			upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k,
+			path:    "/inference-profiles",
+			models:  []string{sharedllm.NormalizeAnthropicModel(strings.TrimPrefix(model, "global."))},
+			outcome: outcomeUpstreamNoListing,
+		})
+	}
+
+	// Vertex carries the model in the rawPredict path and serves no listing
+	// endpoint at all, so the proxy must refuse discovery rather than rewrite
+	// it onto an upstream that would 404.
+	if sa := os.Getenv("GOOGLE_VERTEX_SA_BASE64"); sa != "" {
+		if project := os.Getenv("GOOGLE_VERTEX_PROJECT"); project != "" {
+			region := os.Getenv("GOOGLE_VERTEX_REGION")
+			if region == "" {
+				region = "global"
+			}
+			host := "aiplatform.googleapis.com"
+			if region != "global" {
+				host = region + "-aiplatform.googleapis.com"
+			}
+			cases = append(cases, liveDiscoveryCase{
+				name: "vertex", catalogID: "vertex_ai_api", upstream: "https://" + host,
+				apiKey:  "keyfile::" + sa,
+				path:    "/v1/models",
+				outcome: outcomeDenied,
+			})
+		}
+	}
+
+	return cases
+}
+
+// provisionLiveDiscovery creates the group, provider, optional guardrail and
+// policy for one case, and returns the setup key a client joins that group
+// with. Scoping each provider to its own group is what keeps it the only
+// candidate for its own client's model-less request.
+func provisionLiveDiscovery(t *testing.T, ctx context.Context, tc *liveDiscoveryCase) string {
+	t.Helper()
+
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-disc-live-" + tc.name})
+	require.NoError(t, err, "create group for %s", tc.name)
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-disc-live-" + tc.name,
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grp.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key for %s", tc.name)
+	require.NotEmpty(t, sk.Key, "setup key plaintext for %s", tc.name)
+
+	req := api.AgentNetworkProviderRequest{
+		Name:        "e2e-disc-live-" + tc.name,
+		ProviderId:  tc.catalogID,
+		UpstreamUrl: tc.upstream,
+		ApiKey:      &tc.apiKey,
+		Enabled:     ptr(true),
+	}
+	if len(tc.models) > 0 {
+		models := make([]api.AgentNetworkProviderModel, 0, len(tc.models))
+		for _, id := range tc.models {
+			models = append(models, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.002})
+		}
+		req.Models = &models
+	}
+	prov, err := srv.CreateProvider(ctx, req)
+	require.NoError(t, err, "create provider %s", tc.name)
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	polReq := api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-disc-live-" + tc.name,
+		Enabled:                ptr(true),
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: []string{prov.Id},
+	}
+	if len(tc.allowlist) > 0 {
+		var gr api.AgentNetworkGuardrailRequest
+		gr.Name = "e2e-disc-live-" + tc.name
+		gr.Checks.ModelAllowlist.Enabled = true
+		gr.Checks.ModelAllowlist.Models = tc.allowlist
+		g, gerr := srv.CreateGuardrail(ctx, gr)
+		require.NoError(t, gerr, "create guardrail for %s", tc.name)
+		t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
+		polReq.GuardrailIds = &[]string{g.Id}
+	}
+	pol, err := srv.CreatePolicy(ctx, polReq)
+	require.NoError(t, err, "create policy for %s", tc.name)
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	return sk.Key
+}
+
+// runLiveDiscoveryCase issues the discovery request and reports everything the
+// vendor said before asserting on any of it. The log is the point on the first
+// run: a live catalogue is the one input we do not control, so a failure has to
+// arrive with the response that caused it rather than just a count.
+func runLiveDiscoveryCase(t *testing.T, ctx context.Context, tc liveDiscoveryCase, cl *harness.Client, endpoint, proxyIP string) {
+	t.Helper()
+
+	// A single request is enough for the two non-listing outcomes, and retrying
+	// them would burn the retry window waiting for a status that is never
+	// coming.
+	if tc.outcome != outcomeFiltered {
+		code, body, err := cl.Get(ctx, endpoint, proxyIP, tc.path, tc.headers)
+		require.NoError(t, err, "request must reach the proxy")
+		t.Logf("[discovery] %s GET %s -> %d; body: %s", tc.name, tc.path, code, truncate(body, 2000))
+		assert.NotEqual(t, 200, code,
+			"%s serves no bounded listing, so a 200 here would mean the caller was handed a picker nothing narrows; body: %s",
+			tc.name, truncate(body, 2000))
+
+		// Which side refused is the whole distinction between these two
+		// outcomes, and a NetBird error is the thing that tells them apart: the
+		// middleware chain stamps its own name on anything it generates.
+		if tc.outcome == outcomeDenied {
+			assert.True(t, isProxyError(body),
+				"%s serves no listing endpoint at all, so the proxy must refuse the request itself rather than forward it to an upstream that would answer for us; body: %s",
+				tc.name, truncate(body, 2000))
+			return
+		}
+		assert.False(t, isProxyError(body),
+			"%s discovery must be routed to the configured upstream and refused by the vendor, not blocked by the proxy; body: %s",
+			tc.name, truncate(body, 2000))
+		return
+	}
+
+	code, body := callUntil(t, func() (int, string, error) {
+		return cl.Get(ctx, endpoint, proxyIP, tc.path, tc.headers)
+	}, 200)
+	t.Logf("[discovery] %s GET %s -> %d; body: %s", tc.name, tc.path, code, truncate(body, 4000))
+	require.Equal(t, 200, code, "%s discovery must be served; body: %s", tc.name, truncate(body, 2000))
+
+	ids, ok := listingIDs(body)
+	require.Truef(t, ok,
+		"%s answered discovery with something other than a {\"data\":[{\"id\":…}]} listing, which the filter forwards untouched — the caller would get an unbounded picker; body: %s",
+		tc.name, truncate(body, 2000))
+	sort.Strings(ids)
+	t.Logf("[discovery] %s: %d ids after filtering: %s", tc.name, len(ids), strings.Join(ids, ", "))
+
+	require.NotEmpty(t, ids, "%s filtered the listing down to nothing; the caller would see an empty picker", tc.name)
+
+	permitted := make(map[string]struct{}, len(tc.permitted)*2)
+	for _, id := range tc.permitted {
+		permitted[id] = struct{}{}
+		permitted[sharedllm.NormalizeAnthropicModel(id)] = struct{}{}
+	}
+	for _, id := range ids {
+		_, direct := permitted[id]
+		_, normalised := permitted[sharedllm.NormalizeAnthropicModel(id)]
+		assert.Truef(t, direct || normalised,
+			"%s offered %q, which no policy on this route permits — every entry the picker shows must be a request the guardrail would allow", tc.name, id)
+	}
+	for _, hidden := range tc.wantHidden {
+		assert.NotContainsf(t, ids, hidden,
+			"%s offered %q, which the provider enumerates but the policy does not permit", tc.name, hidden)
+	}
+}
+
+// isProxyError reports whether a response body was generated by the middleware
+// chain rather than forwarded from a vendor. Every chain-generated error names
+// the middleware that raised it, which no upstream's error body does — so this
+// separates "the proxy refused" from "the proxy routed it and the vendor
+// refused", the two failures that otherwise look alike from the client side.
+func isProxyError(body string) bool {
+	return strings.Contains(body, `"middleware":`)
+}
+
+// listingIDs pulls the model ids out of a listing response. ok is false when
+// the body is not the {"data":[{"id":…}]} shape the filter recognises.
+func listingIDs(body string) ([]string, bool) {
+	var doc struct {
+		Data []struct {
+			ID string `json:"id"`
+		} `json:"data"`
+	}
+	if err := json.Unmarshal([]byte(body), &doc); err != nil {
+		return nil, false
+	}
+	if doc.Data == nil {
+		return nil, false
+	}
+	ids := make([]string, 0, len(doc.Data))
+	for _, entry := range doc.Data {
+		ids = append(ids, entry.ID)
+	}
+	return ids, true
+}
+
+func caseNames(cases []liveDiscoveryCase) []string {
+	names := make([]string, 0, len(cases))
+	for _, c := range cases {
+		names = append(names, c.name)
+	}
+	return names
+}
+
+// truncate bounds a logged response body. A live catalogue can run to tens of
+// kilobytes, and the useful part is the front.
+func truncate(s string, limit int) string {
+	if len(s) <= limit {
+		return s
+	}
+	return s[:limit] + "… (" + strconv.Itoa(len(s)-limit) + " more bytes)"
+}

e2e/agentnetwork/discovery_multipolicy_test.go

@@ -0,0 +1,168 @@
+//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"
+)
+
+// TestDiscoveryBoundToCallersPolicies covers a model listing on a provider two
+// teams reach under different allowlists.
+//
+// Bounding the listing by the provider's enumerated models alone is not enough
+// once more than one policy is in play: the caller would be offered every model
+// any team may use, and each one outside their own policy is a request the
+// guardrail refuses a moment later — the empty-or-wrong picker this endpoint
+// exists to avoid, just moved one level up.
+//
+// The client joins the main group only. Both models are enumerated by the same
+// provider and both are advertised by the upstream, so a listing that leaked
+// the other team's model would visibly contain it.
+func TestDiscoveryBoundToCallersPolicies(t *testing.T) {
+	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()) })
+
+	grpMain, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-disc-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-disc-mp-other"})
+	require.NoError(t, err, "create other group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpOther.Id) })
+
+	ephemeral := false
+	mkKey := func(name, groupID string) string {
+		sk, kerr := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+			Name:       name,
+			Type:       "reusable",
+			ExpiresIn:  86400,
+			UsageLimit: 0,
+			AutoGroups: []string{groupID},
+			Ephemeral:  &ephemeral,
+		})
+		require.NoError(t, kerr, "mint setup key %s", name)
+		require.NotEmpty(t, sk.Key, "setup key plaintext")
+		return sk.Key
+	}
+	// One client per group. The second is what makes the first assertion mean
+	// something: without a client that DOES see the other team's model, its
+	// absence from the main client's listing could equally be a policy that
+	// never propagated.
+	keyMain := mkKey("e2e-disc-mp-main-client", grpMain.Id)
+	keyOther := mkKey("e2e-disc-mp-other-client", grpOther.Id)
+
+	// One provider enumerating both models the upstream advertises, so the
+	// listing is narrowed by policy rather than by what the provider serves.
+	staticKey := "static-e2e-token"
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "e2e-disc-mp",
+		ProviderId:  "openai_api",
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &staticKey,
+		Enabled:     ptr(true),
+		Models: &[]api.AgentNetworkProviderModel{
+			{Id: harness.VLLMModel, InputPer1k: 0.001, OutputPer1k: 0.001},
+			{Id: harness.VLLMUnlistedModel, InputPer1k: 0.001, OutputPer1k: 0.001},
+		},
+	})
+	require.NoError(t, err, "create provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.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-disc-mp-main", harness.VLLMModel)
+	gOther := mkGuardrail("e2e-disc-mp-other", harness.VLLMUnlistedModel)
+
+	enabled := true
+	polMain, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-disc-mp-main",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpMain.Id},
+		DestinationProviderIds: []string{prov.Id},
+		GuardrailIds:           &[]string{gMain.Id},
+	})
+	require.NoError(t, err, "create main policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMain.Id) })
+
+	// The other team's policy, on the same provider, permitting the model the
+	// client must never be offered.
+	polOther, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-disc-mp-other",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpOther.Id},
+		DestinationProviderIds: []string{prov.Id},
+		GuardrailIds:           &[]string{gOther.Id},
+	})
+	require.NoError(t, err, "create other policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOther.Id) })
+
+	endpoint, proxyIP, clMain, px := connectClient(t, ctx, "disc-mp", keyMain)
+	clOther := joinClient(t, ctx, px, endpoint, keyOther)
+
+	listing := func(t *testing.T, cl *harness.Client, ip string) string {
+		t.Helper()
+		code, body := callUntil(t, func() (int, string, error) {
+			return cl.Get(ctx, endpoint, ip, "/v1/models?limit=1000", nil)
+		}, 200)
+		require.Equal(t, 200, code, "discovery must be served; body: %s", body)
+		return body
+	}
+
+	otherIP, err := clOther.ResolveProxyIP(ctx, endpoint)
+	require.NoError(t, err, "resolve endpoint from the other client")
+
+	// The other team's client first: seeing its own model proves polOther is
+	// live, so the main client's listing is narrowed by policy scoping rather
+	// than by the other policy having failed to apply at all.
+	otherBody := listing(t, clOther, otherIP)
+	assert.Contains(t, otherBody, harness.VLLMUnlistedModel,
+		"the other group's policy must be in force, or this test proves nothing")
+	assert.NotContains(t, otherBody, harness.VLLMModel,
+		"and it must not be offered the main group's model either — isolation runs both ways")
+
+	mainBody := listing(t, clMain, proxyIP)
+	assert.Contains(t, mainBody, harness.VLLMModel,
+		"the model the caller's own policy permits must reach the picker")
+	assert.NotContains(t, mainBody, harness.VLLMUnlistedModel,
+		"a model only another group's policy permits must not be offered to this caller")
+}
+
+// joinClient starts a second tunnel client against an already-running proxy, so
+// a test can drive the same endpoint as two different group memberships without
+// paying for a second proxy.
+func joinClient(t *testing.T, ctx context.Context, px *harness.Proxy, endpoint, setupKey string) *harness.Client {
+	t.Helper()
+
+	cl, err := harness.StartClient(ctx, srv, setupKey)
+	require.NoError(t, err, "start second client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "second client must connect to management")
+	if _, err := cl.ResolveProxyIP(ctx, endpoint); err != nil {
+		t.Fatalf("second client could not resolve the endpoint: %v", err)
+	}
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("second client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+	return cl
+}

e2e/agentnetwork/gateway_protocol_test.go

@@ -0,0 +1,455 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"fmt"
+	"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"
+)
+
+// Models each catalog surface is registered with in the matrix below. They
+// differ per provider so the router's choice is unambiguous: a request that
+// lands on the wrong provider record fails the surface assertion instead of
+// passing by coincidence.
+const (
+	matrixAnthropicModel = "claude-sonnet-5"
+	matrixBedrockModel   = "anthropic.claude-sonnet-5"
+	// matrixBedrockPathModel is what a Bedrock SDK client puts in the URL: a
+	// cross-region inference profile with a release date and version suffix.
+	// The proxy must normalise it back to matrixBedrockModel to route and price.
+	matrixBedrockPathModel = "us.anthropic.claude-sonnet-5-20250101-v1:0"
+	// matrixVertexModel differs from the Anthropic record's model on purpose:
+	// a shared id would leave two routes claiming it and make which one serves
+	// /v1/messages depend on declaration order.
+	matrixVertexModel   = "claude-haiku-4-5"
+	matrixVertexProject = "e2e-project"
+	matrixVertexRegion  = "us-east5"
+)
+
+// gatewayEnv is a connected client plus a set of provider records, all pointed
+// at one mock upstream, so several wire shapes can be driven over a single
+// tunnel.
+type gatewayEnv struct {
+	endpoint string
+	proxyIP  string
+	client   *harness.Client
+	proxy    *harness.Proxy
+	vllm     *harness.VLLM
+	// providerIDs maps the catalog id to the created provider record id.
+	providerIDs map[string]string
+}
+
+// provisionGatewayMatrix brings up one mock upstream and one provider record
+// per catalog surface, all authorised for the same group by a single policy.
+// Sharing one proxy and client keeps the wire-shape cases to one tunnel setup;
+// each case still creates its own session id so its access-log row is findable.
+func provisionGatewayMatrix(t *testing.T, ctx context.Context) gatewayEnv {
+	t.Helper()
+
+	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-gw-matrix"})
+	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-gw-matrix-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")
+
+	// The mock ignores auth, so a dummy credential satisfies each catalog
+	// entry's auth template. Vertex is the exception: its api_key is a GCP
+	// service-account keyfile the proxy mints an OAuth token from, and a dummy
+	// one cannot mint. That is deliberate — the Vertex case below asserts on
+	// routing, which happens before the token mint.
+	dummyKey := "sk-gw-e2e"
+	dummyKeyfile := "keyfile::" + "e2e-not-a-real-service-account-key"
+
+	specs := []struct {
+		name      string
+		catalogID string
+		apiKey    string
+		models    []api.AgentNetworkProviderModel
+	}{
+		{
+			name: "openai", catalogID: "openai_api", apiKey: dummyKey,
+			models: []api.AgentNetworkProviderModel{{Id: harness.VLLMModel, InputPer1k: 0.001, OutputPer1k: 0.002}},
+		},
+		{
+			name: "anthropic", catalogID: "anthropic_api", apiKey: dummyKey,
+			models: []api.AgentNetworkProviderModel{{Id: matrixAnthropicModel, InputPer1k: 0.003, OutputPer1k: 0.015}},
+		},
+		{
+			name: "bedrock", catalogID: "bedrock_api", apiKey: dummyKey,
+			models: []api.AgentNetworkProviderModel{{Id: matrixBedrockModel, InputPer1k: 0.003, OutputPer1k: 0.015}},
+		},
+		{
+			name: "vertex", catalogID: "vertex_ai_api", apiKey: dummyKeyfile,
+			models: []api.AgentNetworkProviderModel{{Id: matrixVertexModel, InputPer1k: 0.001, OutputPer1k: 0.005}},
+		},
+	}
+
+	providerIDs := make(map[string]string, len(specs))
+	ids := make([]string, 0, len(specs))
+	for _, spec := range specs {
+		key := spec.apiKey
+		models := spec.models
+		prov, perr := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+			Name:        "e2e-gw-" + spec.name,
+			ProviderId:  spec.catalogID,
+			UpstreamUrl: vllm.URL,
+			ApiKey:      &key,
+			Enabled:     ptr(true),
+			Models:      &models,
+		})
+		require.NoError(t, perr, "create %s provider", spec.name)
+		id := prov.Id
+		t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), id) })
+		providerIDs[spec.catalogID] = id
+		ids = append(ids, id)
+	}
+
+	// Uncapped token limit: never blocks the handful of tokens driven here, but
+	// switches on usage metering so consumption and cost land in the row.
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-gw-matrix",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: ids,
+		Limits: &api.AgentNetworkPolicyLimits{
+			TokenLimit: api.AgentNetworkPolicyTokenLimit{
+				Enabled:       true,
+				GroupCap:      10_000_000,
+				UserCap:       10_000_000,
+				WindowSeconds: 60,
+			},
+		},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	endpoint, proxyIP, cl, px := connectClient(t, ctx, "gw-matrix", sk.Key)
+	return gatewayEnv{
+		endpoint:    endpoint,
+		proxyIP:     proxyIP,
+		client:      cl,
+		proxy:       px,
+		vllm:        vllm,
+		providerIDs: providerIDs,
+	}
+}
+
+// connectClient starts a proxy and a tunnel client for the shared account and
+// waits until the client can reach the proxy peer, returning the endpoint and
+// the proxy's tunnel IP to pin requests to.
+func connectClient(t *testing.T, ctx context.Context, name, setupKey string) (string, string, *harness.Client, *harness.Proxy) {
+	t.Helper()
+
+	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-"+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, setupKey)
+	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")
+	// The probe 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()))
+	}
+	return settings.Endpoint, proxyIP, cl, px
+}
+
+// callUntil retries an HTTP call through the tunnel until it returns one of the
+// wanted statuses or the deadline passes, absorbing the DNS and tunnel jitter
+// the first call through a fresh tunnel can hit. The last status and body are
+// returned either way so the caller can assert with real detail.
+func callUntil(t *testing.T, call func() (int, string, error), want ...int) (int, string) {
+	t.Helper()
+	wanted := make(map[int]struct{}, len(want))
+	for _, w := range want {
+		wanted[w] = struct{}{}
+	}
+
+	var code int
+	var body string
+	deadline := time.Now().Add(90 * time.Second)
+	for time.Now().Before(deadline) {
+		c, b, err := call()
+		if err == nil {
+			code, body = c, b
+			if _, ok := wanted[code]; ok {
+				return code, body
+			}
+		}
+		time.Sleep(5 * time.Second)
+	}
+	return code, body
+}
+
+// TestGatewayProtocolProviderMatrix drives one request per wire shape over a
+// single tunnel, with a provider record per catalog surface behind it. It is
+// the regression net for the routing and parser-selection changes: each case
+// asserts the surface the request was metered under and the token counts that
+// surface's own usage block carries, so a request parsed by the wrong provider's
+// parser meters zero and fails rather than passing on a coincidence.
+func TestGatewayProtocolProviderMatrix(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	env := provisionGatewayMatrix(t, ctx)
+	diag := func() string {
+		return fmt.Sprintf("\n=== upstream logs ===\n%s\n=== proxy logs ===\n%s",
+			env.vllm.Logs(context.Background()), env.proxy.Logs(context.Background()))
+	}
+
+	t.Run("openai chat completions", func(t *testing.T) {
+		session := "e2e-gw-openai"
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireChat, harness.VLLMModel, "ping", session)
+		}, 200)
+		require.Equal(t, 200, code, "openai chat must succeed; body: %s%s", body, diag())
+		require.Contains(t, body, "chat.completion", "body must be an OpenAI completion; got: %s", body)
+
+		row := findAccessLogBySession(t, ctx, session)
+		require.NotNil(t, row.Provider)
+		assert.Equal(t, "openai", *row.Provider, "the OpenAI chat path must meter under the openai surface")
+		assert.Equal(t, int64(harness.VLLMChatInputTokens), row.InputTokens, "OpenAI usage block must be read")
+		assert.Equal(t, int64(harness.VLLMChatOutputTokens), row.OutputTokens)
+	})
+
+	t.Run("anthropic messages", func(t *testing.T) {
+		session := "e2e-gw-anthropic"
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireMessages, matrixAnthropicModel, "ping", session)
+		}, 200)
+		require.Equal(t, 200, code, "anthropic messages must succeed; body: %s%s", body, diag())
+
+		row := findAccessLogBySession(t, ctx, session)
+		require.NotNil(t, row.Provider)
+		assert.Equal(t, "anthropic", *row.Provider, "the /v1/messages path must meter under the anthropic surface")
+		// These counts only appear if the Anthropic parser read the response:
+		// its usage fields are named differently from the OpenAI block.
+		assert.Equal(t, int64(harness.VLLMMessagesInputTokens), row.InputTokens,
+			"Anthropic input_tokens must be read; zero here means the wrong parser ran")
+		assert.Equal(t, int64(harness.VLLMMessagesOutputTokens), row.OutputTokens)
+		assert.Positive(t, row.CachedInputTokens, "the Anthropic cache-read bucket must be recorded")
+		assert.Positive(t, row.CostUsd, "a metered request must carry a cost")
+		require.NotNil(t, row.ResolvedProviderId)
+		assert.Equal(t, env.providerIDs["anthropic_api"], *row.ResolvedProviderId,
+			"a vendor-tagged request must not cross to another provider's record")
+	})
+
+	t.Run("bedrock invoke normalises the path model", func(t *testing.T) {
+		session := "e2e-gw-bedrock"
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Bedrock(ctx, env.endpoint, env.proxyIP, matrixBedrockPathModel, "ping", session)
+		}, 200)
+		require.Equal(t, 200, code, "bedrock invoke must succeed; body: %s%s", body, diag())
+
+		row := findAccessLogBySession(t, ctx, session)
+		require.NotNil(t, row.Provider)
+		assert.Equal(t, "bedrock", *row.Provider, "a native Bedrock path must meter under the bedrock surface")
+		require.NotNil(t, row.Model)
+		assert.Equal(t, matrixBedrockModel, *row.Model,
+			"the inference-profile prefix, release date and version suffix must be normalised away")
+		assert.Equal(t, int64(harness.VLLMMessagesInputTokens), row.InputTokens)
+	})
+
+	t.Run("anthropic token counting", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.PostJSON(ctx, env.endpoint, env.proxyIP, "/v1/messages/count_tokens",
+				fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":"ping"}]}`, matrixAnthropicModel),
+				[]string{"anthropic-version: 2023-06-01"})
+		}, 200)
+		assert.Equal(t, 200, code, "token counting must route rather than deny; body: %s%s", body, diag())
+	})
+
+	t.Run("bedrock token counting", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.PostJSON(ctx, env.endpoint, env.proxyIP,
+				"/model/"+matrixBedrockPathModel+"/count-tokens",
+				`{"input":{"converse":{"messages":[{"role":"user","content":[{"text":"ping"}]}]}}}`, nil)
+		}, 200)
+		assert.Equal(t, 200, code,
+			"the Bedrock count-tokens action must route; denying it pushes counting onto the billable inference path; body: %s%s",
+			body, diag())
+	})
+
+	t.Run("vertex token counting reaches its provider", func(t *testing.T) {
+		// The dummy service-account key cannot mint an OAuth token, so the
+		// request stops at the upstream credential. Both outcomes render as
+		// 403, so the deny code is what distinguishes them: upstream_auth_failed
+		// means the path resolved to the Vertex route and only the credential
+		// failed, while model_not_routable would mean the method segment was
+		// swallowed into the model id and no route ever claimed it.
+		path := fmt.Sprintf("/v1/projects/%s/locations/%s/publishers/anthropic/models/%s/count-tokens:rawPredict",
+			matrixVertexProject, matrixVertexRegion, matrixVertexModel)
+		_, body := callUntil(t, func() (int, string, error) {
+			return env.client.PostJSON(ctx, env.endpoint, env.proxyIP, path,
+				`{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"ping"}]}`, nil)
+		}, 403)
+		assert.NotContains(t, body, "model_not_routable",
+			"the count-tokens method segment must not be parsed as part of the model id; body: %s%s", body, diag())
+		assert.Contains(t, body, "llm_policy.upstream_auth_failed",
+			"the request must reach the Vertex route and fail only at the credential; body: %s%s", body, diag())
+	})
+
+	t.Run("connection warming probe", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Get(ctx, env.endpoint, env.proxyIP, "/api/hello", nil)
+		}, 200)
+		assert.NotEqual(t, 403, code,
+			"the warm-up probe carries no model and must not be refused as unroutable; body: %s%s", body, diag())
+	})
+
+	t.Run("unknown model denies in the caller's error shape", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireMessages,
+				"claude-not-a-real-model-9", "ping", "e2e-gw-unknown")
+		}, 403)
+		require.Equal(t, 403, code, "a model no provider claims must still be refused; body: %s%s", body, diag())
+
+		// The NetBird fields stay where they were for existing consumers.
+		assert.Contains(t, body, "llm_policy.model_not_routable", "the deny code must be preserved")
+		// And the vendor's own envelope rides alongside, so the client can show
+		// the reason instead of an unexplained API error.
+		assert.Contains(t, body, `"type":"error"`, "an Anthropic caller must get the Anthropic error envelope")
+		assert.Contains(t, body, "permission_error", "403 must map to the vendor's permission error type")
+	})
+}
+
+// TestModelDiscoveryWithModelAllowlist covers gateway model discovery on an
+// account that restricts models, which is the configuration that broke: the
+// listing carries no model, and the per-model allowlist fails closed on an
+// undetermined one, so discovery denied for exactly the accounts using the
+// feature. It also asserts the allowlist still refuses a model outside it, so
+// the exemption cannot be read as a way around the gate.
+func TestModelDiscoveryWithModelAllowlist(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*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-gw-discovery"})
+	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-gw-discovery-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")
+
+	// One provider enumerating a single model, while the upstream's own listing
+	// advertises two. The proxy must serve the shorter list.
+	dummyKey := "sk-discovery-e2e"
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "e2e-gw-discovery",
+		ProviderId:  "openai_api",
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &dummyKey,
+		Enabled:     ptr(true),
+		Models: &[]api.AgentNetworkProviderModel{
+			{Id: harness.VLLMModel, InputPer1k: 0.001, OutputPer1k: 0.002},
+		},
+	})
+	require.NoError(t, err, "create provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	// The model allowlist is what makes this a regression test: without a
+	// guardrail enabled, discovery was never gated in the first place.
+	var gr api.AgentNetworkGuardrailRequest
+	gr.Name = "e2e-gw-discovery-allowlist"
+	gr.Checks.ModelAllowlist.Enabled = true
+	gr.Checks.ModelAllowlist.Models = []string{harness.VLLMModel}
+	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-gw-discovery",
+		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) })
+
+	endpoint, proxyIP, cl, px := connectClient(t, ctx, "gw-discovery", sk.Key)
+	diag := func() string {
+		return fmt.Sprintf("\n=== upstream logs ===\n%s\n=== proxy logs ===\n%s",
+			vllm.Logs(context.Background()), px.Logs(context.Background()))
+	}
+
+	t.Run("listing is served and bounded by policy", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return cl.Get(ctx, endpoint, proxyIP, "/v1/models?limit=1000", nil)
+		}, 200)
+		require.Equal(t, 200, code,
+			"discovery must not be refused because the request carries no model; body: %s%s", body, diag())
+
+		assert.Contains(t, body, harness.VLLMModel, "the authorised model must reach the picker")
+		assert.NotContains(t, body, harness.VLLMUnlistedModel,
+			"a model the policy does not authorise must not be offered; body: %s", body)
+	})
+
+	t.Run("allowlist still refuses a model outside it", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return cl.Chat(ctx, endpoint, proxyIP, harness.WireChat,
+				harness.VLLMUnlistedModel, "ping", "e2e-gw-discovery-blocked")
+		}, 403)
+		require.Equal(t, 403, code,
+			"exempting model-less endpoints must not exempt inference; body: %s%s", body, diag())
+		assert.True(t,
+			strings.Contains(body, "llm_policy.model_blocked") || strings.Contains(body, "llm_policy.model_not_routable"),
+			"the refusal must name a model policy code; body: %s", body)
+	})
+
+	t.Run("allowlisted model still routes", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return cl.Chat(ctx, endpoint, proxyIP, harness.WireChat,
+				harness.VLLMModel, "ping", "e2e-gw-discovery-allowed")
+		}, 200)
+		require.Equal(t, 200, code, "the allowlisted model must still be served; body: %s%s", body, diag())
+	})
+}

e2e/agentnetwork/gateway_review_test.go

@@ -0,0 +1,242 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"fmt"
+	"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"
+)
+
+// The cases in this file cover behaviour that arrived from code review, after
+// the gateway-protocol end-to-end tests were written. Each had unit coverage
+// only; none needed a new harness capability, which is why they belong here
+// rather than on a manual checklist.
+
+// TestNonInferenceEndpointsAreAuthorised covers the two review findings on the
+// endpoints that carry no body: the per-model lookup must be authorised
+// against the same allowlist that bounds the listing beside it, and only a read
+// method may claim the non-inference exemption that skips the token pre-flight.
+func TestNonInferenceEndpointsAreAuthorised(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	env := provisionDiscoveryProvider(t, ctx)
+
+	t.Run("lookup of an authorised model succeeds", func(t *testing.T) {
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Get(ctx, env.endpoint, env.proxyIP, "/v1/models/"+harness.VLLMModel, nil)
+		}, 200)
+		assert.Equal(t, 200, code, "an allowlisted model must remain reachable; body: %s", body)
+	})
+
+	t.Run("lookup of an unauthorised model is refused", func(t *testing.T) {
+		code, body, err := env.client.Get(ctx, env.endpoint, env.proxyIP, "/v1/models/"+harness.VLLMUnlistedModel, nil)
+		require.NoError(t, err, "request must reach the proxy")
+		assert.Equal(t, 403, code,
+			"a model the policy does not authorise must not be confirmed by the detail lookup; body: %s", body)
+	})
+
+	// A write must not claim the exemption that lets the listing skip the token
+	// pre-flight. The body names no model on purpose: that is what a request
+	// probing for the exemption looks like, and it is the case the method gate
+	// exists to refuse. (A POST that does name a model is a different thing —
+	// it routes and meters as the inference request it is.)
+	for _, path := range []string{"/v1/models", "/v1/models/" + harness.VLLMModel, "/api/hello"} {
+		t.Run("write to "+path+" is refused", func(t *testing.T) {
+			code, body, err := env.client.PostJSON(ctx, env.endpoint, env.proxyIP, path,
+				`{"messages":[{"role":"user","content":"hi"}]}`, nil)
+			require.NoError(t, err, "request must reach the proxy")
+			assert.NotEqual(t, 200, code,
+				"a write to a non-inference path must not be served unmetered; body: %s", body)
+		})
+	}
+
+	// A request carrying the sub-agent attribution headers must still be served
+	// and metered normally. Asserting the ids themselves is not possible yet:
+	// the parser lifts them onto the request's metadata, but nothing persists
+	// them, so they have no queryable surface to check against.
+	t.Run("sub-agent headers do not disturb the request", func(t *testing.T) {
+		sessionID := fmt.Sprintf("e2e-session-agentid-%d", time.Now().UnixNano())
+		code, body, err := env.client.PostJSON(ctx, env.endpoint, env.proxyIP, "/v1/chat/completions",
+			fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":"Reply with exactly: pong"}]}`, harness.VLLMModel),
+			[]string{
+				"x-session-id: " + sessionID,
+				"x-claude-code-agent-id: agent-child-7",
+				"x-claude-code-parent-agent-id: agent-root-1",
+			})
+		require.NoError(t, err, "request must reach the proxy")
+		require.Equal(t, 200, code, "the request must succeed; body: %s", body)
+
+		row := findAccessLogBySession(t, ctx, sessionID)
+		assert.Positive(t, row.InputTokens, "the request must still be metered normally")
+	})
+}
+
+// TestDatedModelIdRouting covers both halves of the dated-id rule that review
+// tightened: a dated id still reaches an undated registration, but a route
+// pinned to one dated build must never serve a different one.
+func TestDatedModelIdRouting(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	const (
+		undated = "claude-sonnet-9"
+		datedA  = "claude-sonnet-9-20250101"
+		datedB  = "claude-sonnet-9-20250202"
+	)
+
+	t.Run("a dated id reaches its undated registration", func(t *testing.T) {
+		env := provisionModelProvider(t, ctx, "dated-undated", "anthropic_api", undated)
+
+		sessionID := fmt.Sprintf("e2e-session-dated-%d", time.Now().UnixNano())
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireMessages, datedA, "Reply with exactly: pong", sessionID)
+		}, 200)
+		require.Equal(t, 200, code, "a pinned release of a registered family must route; body: %s", body)
+
+		row := findAccessLogBySession(t, ctx, sessionID)
+		assert.Positive(t, row.InputTokens, "the dated request must price at the registered rate, not zero")
+	})
+
+	t.Run("a route pinned to one dated build refuses another", func(t *testing.T) {
+		env := provisionModelProvider(t, ctx, "dated-pinned", "anthropic_api", datedA)
+
+		code, body := callUntil(t, func() (int, string, error) {
+			return env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireMessages, datedA, "Reply with exactly: pong", "")
+		}, 200)
+		require.Equal(t, 200, code, "the exact dated id must still route; body: %s", body)
+
+		code, body, err := env.client.Chat(ctx, env.endpoint, env.proxyIP, harness.WireMessages, datedB, "Reply with exactly: pong", "")
+		require.NoError(t, err, "request must reach the proxy")
+		assert.Equal(t, 403, code,
+			"a provider pinned to one dated build must not serve another; body: %s", body)
+	})
+}
+
+// TestBedrockInferenceProfilesReachTheUpstream covers the startup lookup a
+// Bedrock client makes. The proxy forwards it to the configured upstream rather
+// than denying it, so what comes back is the upstream's answer — never a
+// NetBird policy rejection.
+func TestBedrockInferenceProfilesReachTheUpstream(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	env := provisionModelProvider(t, ctx, "infprofiles", "bedrock_api", "anthropic.claude-sonnet-5")
+
+	code, body := callUntil(t, func() (int, string, error) {
+		return env.client.Get(ctx, env.endpoint, env.proxyIP, "/inference-profiles", nil)
+	}, 200)
+
+	assert.Equal(t, 200, code, "the lookup must reach the upstream; body: %s", body)
+	assert.NotContains(t, body, "llm_policy.",
+		"the proxy must not answer a control-plane lookup with a policy denial")
+	assert.Contains(t, body, "inferenceProfileSummaries",
+		"the upstream's own answer must come back untouched")
+}
+
+// provisionDiscoveryProvider brings up one mock-backed provider enumerating a
+// single model, with an allowlist guardrail in effect, plus a connected client.
+func provisionDiscoveryProvider(t *testing.T, ctx context.Context) pricedEnv {
+	t.Helper()
+	env := provisionModelProvider(t, ctx, "noninference", "openai_api", harness.VLLMModel)
+
+	var gr api.AgentNetworkGuardrailRequest
+	gr.Name = "e2e-noninference-allowlist-" + fmt.Sprint(time.Now().UnixNano())
+	gr.Checks.ModelAllowlist.Enabled = true
+	gr.Checks.ModelAllowlist.Models = []string{harness.VLLMModel}
+	guard, err := srv.CreateGuardrail(ctx, gr)
+	require.NoError(t, err, "create guardrail")
+	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), guard.Id) })
+
+	enabled := true
+	_, err = srv.UpdatePolicy(ctx, env.policyID, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-noninference",
+		Enabled:                &enabled,
+		SourceGroups:           []string{env.groupID},
+		DestinationProviderIds: []string{env.providerID},
+		GuardrailIds:           &[]string{guard.Id},
+	})
+	require.NoError(t, err, "attach guardrail to policy")
+	return env
+}
+
+// provisionModelProvider brings up the mock, one provider under the given
+// catalog id enumerating exactly one model, an authorising policy, and a
+// connected proxy + client.
+func provisionModelProvider(t *testing.T, ctx context.Context, name, catalogID, model string) pricedEnv {
+	t.Helper()
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	suffix := strings.ToLower(name)
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-gwr-" + suffix})
+	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-gwr-" + suffix + "-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")
+
+	dummyKey := "sk-gwr-e2e"
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "e2e-gwr-" + suffix,
+		ProviderId:  catalogID,
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &dummyKey,
+		Enabled:     ptr(true),
+		Models: &[]api.AgentNetworkProviderModel{
+			{Id: model, InputPer1k: 0.001, OutputPer1k: 0.002},
+		},
+	})
+	require.NoError(t, err, "create provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-gwr-" + suffix,
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: []string{prov.Id},
+		Limits: &api.AgentNetworkPolicyLimits{
+			TokenLimit: api.AgentNetworkPolicyTokenLimit{
+				Enabled:       true,
+				GroupCap:      10_000_000,
+				UserCap:       10_000_000,
+				WindowSeconds: 60,
+			},
+		},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	endpoint, proxyIP, cl, px := connectClient(t, ctx, "gwr-"+suffix, sk.Key)
+	return pricedEnv{
+		providerID: prov.Id,
+		groupID:    grp.Id,
+		policyID:   pol.Id,
+		upstream:   vllm.URL,
+		endpoint:   endpoint,
+		proxyIP:    proxyIP,
+		client:     cl,
+		proxy:      px,
+	}
+}

e2e/agentnetwork/streaming_test.go

@@ -0,0 +1,199 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"fmt"
+	"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"
+)
+
+// streamedModel is priced high enough that a mis-metered request is obvious in
+// the recorded cost, and named so it cannot collide with another test's route.
+const streamedModel = "e2e-streamed-model"
+
+const (
+	streamInRate  = 0.010
+	streamOutRate = 0.020
+	// The cache-read bucket is priced separately from input, so a run that
+	// folded the two together fails the per-bucket assertions below.
+	streamCacheReadRate = 0.001
+)
+
+// TestStreamingResponseMetersInputTokens is the end-to-end guard for the
+// metering bug this endpoint's gateway-protocol work fixed.
+//
+// On a streamed answer the input-token count exists only in the opening
+// message_start event; every later frame reports output. A response read with
+// the wrong vendor's parser — the shape a gateway record produces when it names
+// one API surface and serves another — never looks at that event, so input
+// metered as zero and the bulk of the bill silently vanished. Nothing in the
+// suite sent stream: true before this test, so the whole branch went unrun.
+//
+// The provider points at the mock's streaming listener, which answers every
+// request as SSE with token counts that differ from the buffered surface. That
+// difference is the point: passing these assertions is only possible if the
+// stream accumulator ran.
+func TestStreamingResponseMetersInputTokens(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	env := provisionStreamingProvider(t, ctx, "anthropic_api")
+
+	sessionID := fmt.Sprintf("e2e-session-stream-%d", time.Now().UnixNano())
+	code, body := chatStreamUntil(t, ctx, env, harness.WireMessages, streamedModel, sessionID)
+	require.Equal(t, 200, code, "streamed chat must succeed; body: %s", body)
+	assert.Contains(t, body, "message_start",
+		"the client must receive the event stream itself, not a buffered rewrite of it")
+
+	row := findAccessLogBySession(t, ctx, sessionID)
+
+	assert.Equal(t, harness.VLLMStreamInputTokens, int(row.InputTokens),
+		"input tokens live in message_start; zero here is the bug this test exists for")
+	assert.Equal(t, harness.VLLMStreamOutputTokens, int(row.OutputTokens),
+		"output tokens ride message_delta and supersede the message_start seed")
+	assert.Equal(t, harness.VLLMStreamCacheReadTokens, int(row.CachedInputTokens),
+		"the Anthropic cache bucket rides message_start too, and only its own parser reads it")
+
+	// The Anthropic surface bills cache reads additively, so the input bucket
+	// prices the full input count rather than a remainder.
+	wantInput := float64(harness.VLLMStreamInputTokens) / 1000 * streamInRate
+	wantOutput := float64(harness.VLLMStreamOutputTokens) / 1000 * streamOutRate
+	assert.InDelta(t, wantInput, row.InputCostUsd, 1e-6, "input cost must price the streamed input tokens")
+	assert.InDelta(t, wantOutput, row.OutputCostUsd, 1e-6, "output cost must price the streamed output tokens")
+	assert.Greater(t, row.CostUsd, 0.0, "a streamed request must never record as free")
+}
+
+// TestStreamingOnGatewayTypedProvider drives the same streamed Anthropic call
+// through a provider record whose catalog id names the OpenAI surface — the
+// exact misconfiguration that hid the bug, since gateway records commonly pin
+// one parser while the upstream serves another shape entirely.
+//
+// The router must choose the parser from the request path rather than the
+// record's provider id, or the Anthropic usage block goes unread and input
+// meters at zero all over again.
+func TestStreamingOnGatewayTypedProvider(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	env := provisionStreamingProvider(t, ctx, "openai_api")
+
+	sessionID := fmt.Sprintf("e2e-session-stream-gw-%d", time.Now().UnixNano())
+	code, body := chatStreamUntil(t, ctx, env, harness.WireMessages, streamedModel, sessionID)
+	require.Equal(t, 200, code, "streamed chat through a gateway record must succeed; body: %s", body)
+
+	row := findAccessLogBySession(t, ctx, sessionID)
+
+	assert.Equal(t, harness.VLLMStreamInputTokens, int(row.InputTokens),
+		"a record typed openai_api must still read the Anthropic usage block it is actually serving")
+	assert.Equal(t, harness.VLLMStreamOutputTokens, int(row.OutputTokens),
+		"output tokens must survive the surface mismatch too")
+	assert.InDelta(t, float64(harness.VLLMStreamInputTokens)/1000*streamInRate, row.InputCostUsd, 1e-6,
+		"the request must be priced on the surface it spoke, not the one the record names")
+}
+
+// provisionStreamingProvider brings up the mock, one provider pointed at its
+// streaming listener under the given catalog id, a policy authorising it, and a
+// connected proxy + client.
+func provisionStreamingProvider(t *testing.T, ctx context.Context, catalogID string) pricedEnv {
+	t.Helper()
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	name := "stream-" + catalogID
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-" + 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-" + 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")
+
+	dummyKey := "sk-stream-e2e"
+	cacheRead := streamCacheReadRate
+	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        name,
+		ProviderId:  catalogID,
+		UpstreamUrl: vllm.StreamURL,
+		ApiKey:      &dummyKey,
+		Enabled:     ptr(true),
+		Models: &[]api.AgentNetworkProviderModel{{
+			Id:             streamedModel,
+			InputPer1k:     streamInRate,
+			OutputPer1k:    streamOutRate,
+			CacheReadPer1k: &cacheRead,
+		}},
+	})
+	require.NoError(t, err, "create provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
+
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-" + name,
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: []string{prov.Id},
+		Limits: &api.AgentNetworkPolicyLimits{
+			TokenLimit: api.AgentNetworkPolicyTokenLimit{
+				Enabled:       true,
+				GroupCap:      10_000_000,
+				UserCap:       10_000_000,
+				WindowSeconds: 60,
+			},
+		},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	endpoint, proxyIP, cl, px := connectClient(t, ctx, name, sk.Key)
+	return pricedEnv{
+		providerID: prov.Id,
+		groupID:    grp.Id,
+		policyID:   pol.Id,
+		upstream:   vllm.StreamURL,
+		endpoint:   endpoint,
+		proxyIP:    proxyIP,
+		client:     cl,
+		proxy:      px,
+	}
+}
+
+// chatStreamUntil drives one streamed chat, retrying to absorb the tunnel and
+// DNS jitter a first call through a fresh peer can hit.
+func chatStreamUntil(t *testing.T, ctx context.Context, env pricedEnv, kind, model, sessionID string) (int, string) {
+	t.Helper()
+	var code int
+	var body string
+	deadline := time.Now().Add(90 * time.Second)
+	for time.Now().Before(deadline) {
+		c, b, cerr := env.client.ChatStream(ctx, env.endpoint, env.proxyIP, kind, model, "Reply with exactly: pong", sessionID)
+		if cerr == nil {
+			code, body = c, b
+			if code == 200 {
+				break
+			}
+		}
+		time.Sleep(5 * time.Second)
+	}
+	if code != 200 {
+		t.Logf("=== proxy logs ===\n%s", env.proxy.Logs(context.Background()))
+	}
+	return code, body
+}

e2e/harness/client.go

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"net/http"
 	"os/exec"
 	"strconv"
 	"strings"
@@ -199,12 +200,18 @@ func (cl *Client) pollStatus(ctx context.Context, timeout time.Duration, want st
 const (
 	// curlExitCouldNotResolve is curl's exit code for a DNS resolution failure, distinct from connection-level failures.
 	curlExitCouldNotResolve = 6
-	// dnsProbeRetryWindow bounds DNS-failure retries: the synthesized zone lands a beat after management connects, so early NXDOMAIN is propagation; a zone still absent after this window is a real failure.
-	dnsProbeRetryWindow   = 30 * time.Second
-	dnsProbeRetryInterval = 2 * time.Second
+	// curlExitCouldNotConnect is curl's exit code for a connection that never
+	// established. The probe exists to WAKE the lazy proxy peer, so the first
+	// attempt legitimately arrives before WireGuard has brought the tunnel up
+	// and fails here — which is propagation, exactly like an early NXDOMAIN,
+	// and belongs inside the retry window rather than failing the test outright.
+	curlExitCouldNotConnect = 7
+	// endpointProbeRetryWindow bounds retries of the transient failures above: the synthesized zone and the tunnel both land a beat after management connects. Still failing after this window is a real failure.
+	endpointProbeRetryWindow   = 30 * time.Second
+	endpointProbeRetryInterval = 2 * time.Second
 )
 
-// ResolveProxyIP GETs https://<endpoint>/ from the client's netns: any HTTP status proves DNS + tunnel and wakes the lazy proxy peer; only DNS failures retry, within dnsProbeRetryWindow. Returns the connected IP for --resolve pinning.
+// ResolveProxyIP GETs https://<endpoint>/ from the client's netns: any HTTP status proves DNS + tunnel and wakes the lazy proxy peer; DNS and connect failures retry, within endpointProbeRetryWindow. Returns the connected IP for --resolve pinning.
 func (cl *Client) ResolveProxyIP(ctx context.Context, endpoint string) (string, error) {
 	args := []string{
 		"run", "--rm",
@@ -215,7 +222,7 @@ func (cl *Client) ResolveProxyIP(ctx context.Context, endpoint string) (string,
 		"-w", "%{remote_ip}",
 		"https://" + endpoint + "/",
 	}
-	deadline := time.Now().Add(dnsProbeRetryWindow)
+	deadline := time.Now().Add(endpointProbeRetryWindow)
 	for {
 		cmd := exec.CommandContext(ctx, "docker", args...)
 		var stdout, stderr strings.Builder
@@ -231,21 +238,29 @@ func (cl *Client) ResolveProxyIP(ctx context.Context, endpoint string) (string,
 		}
 
 		var exitErr *exec.ExitError
-		if !errors.As(err, &exitErr) || exitErr.ExitCode() != curlExitCouldNotResolve {
+		if !errors.As(err, &exitErr) || !isTransientProbeExit(exitErr.ExitCode()) {
 			return "", fmt.Errorf("no HTTP response from %s: %w (%s)", endpoint, err, strings.TrimSpace(stderr.String()))
 		}
-		dnsErr := fmt.Errorf("DNS resolution failed for %s: %s", endpoint, strings.TrimSpace(stderr.String()))
-		if time.Until(deadline) < dnsProbeRetryInterval {
-			return "", dnsErr
+		probeErr := fmt.Errorf("endpoint %s not reachable yet: %s", endpoint, strings.TrimSpace(stderr.String()))
+		if time.Until(deadline) < endpointProbeRetryInterval {
+			return "", probeErr
 		}
 		select {
 		case <-ctx.Done():
-			return "", fmt.Errorf("%w (%w)", dnsErr, ctx.Err())
-		case <-time.After(dnsProbeRetryInterval):
+			return "", fmt.Errorf("%w (%w)", probeErr, ctx.Err())
+		case <-time.After(endpointProbeRetryInterval):
 		}
 	}
 }
 
+// isTransientProbeExit reports whether a curl exit code describes a state the
+// endpoint is expected to pass THROUGH on its way up, rather than a settled
+// failure. Anything else — TLS refusal, a protocol error, a bad argument —
+// would still be failing after the retry window, so it fails immediately.
+func isTransientProbeExit(code int) bool {
+	return code == curlExitCouldNotResolve || code == curlExitCouldNotConnect
+}
+
 // Wire shapes for Chat.
 const (
 	// WireChat is the OpenAI-compatible /v1/chat/completions shape.
@@ -292,6 +307,27 @@ func (cl *Client) ChatPrefixed(ctx context.Context, endpoint, proxyIP, pathPrefi
 	return cl.post(ctx, endpoint, proxyIP, pathPrefix+path, body, withSessionID(headers, sessionID))
 }
 
+// ChatStream is Chat with "stream": true in the request body, so the proxy's
+// request parser marks the call as streaming and its response parser takes the
+// SSE accumulator rather than the buffered-body path. Pair it with a provider
+// pointed at VLLM.StreamURL, which answers every request as an event stream.
+func (cl *Client) ChatStream(ctx context.Context, endpoint, proxyIP, kind, model, prompt, sessionID string) (int, string, error) {
+	var path, body string
+	var headers []string
+	switch kind {
+	case WireMessages:
+		path = "/v1/messages"
+		headers = []string{"anthropic-version: 2023-06-01"}
+		body = fmt.Sprintf(`{"model":%q,"max_tokens":2048,"stream":true,"messages":[{"role":"user","content":%q}]}`, model, prompt)
+	default:
+		path = "/v1/chat/completions"
+		// include_usage is what makes a real OpenAI stream emit its final usage
+		// frame; without it the last chunk carries no tokens at all.
+		body = fmt.Sprintf(`{"model":%q,"stream":true,"stream_options":{"include_usage":true},"messages":[{"role":"user","content":%q}]}`, model, prompt)
+	}
+	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(headers, sessionID))
+}
+
 // Vertex issues an Anthropic-on-Vertex rawPredict POST over the tunnel. Unlike
 // Chat, the model is carried in the request path (project/region/model), so the
 // proxy routes by path and mints the service-account OAuth token; the body uses
@@ -322,10 +358,29 @@ func withSessionID(headers []string, sessionID string) []string {
 	return append(headers, "x-session-id: "+sessionID)
 }
 
-// post runs curl in a throwaway container sharing the client's network
-// namespace so the request traverses the WireGuard tunnel, pinning the endpoint
-// to the proxy IP. It returns the HTTP status and response body.
+// Get issues a GET to the agent-network endpoint over the client's tunnel.
+// Model discovery and the connection-warming probe are read-only endpoints
+// that carry no body, so they can't go through the chat helpers.
+func (cl *Client) Get(ctx context.Context, endpoint, proxyIP, path string, extraHeaders []string) (int, string, error) {
+	return cl.do(ctx, http.MethodGet, endpoint, proxyIP, path, "", extraHeaders)
+}
+
+// PostJSON issues an arbitrary JSON POST over the client's tunnel, for wire
+// shapes the typed helpers don't cover (token counting, say).
+func (cl *Client) PostJSON(ctx context.Context, endpoint, proxyIP, path, body string, extraHeaders []string) (int, string, error) {
+	return cl.do(ctx, http.MethodPost, endpoint, proxyIP, path, body, extraHeaders)
+}
+
+// post issues a JSON POST. Retained as the shorthand the chat helpers use.
 func (cl *Client) post(ctx context.Context, endpoint, proxyIP, path, body string, extraHeaders []string) (int, string, error) {
+	return cl.do(ctx, http.MethodPost, endpoint, proxyIP, path, body, extraHeaders)
+}
+
+// do runs curl in a throwaway container sharing the client's network
+// namespace so the request traverses the WireGuard tunnel, pinning the endpoint
+// to the proxy IP. It returns the HTTP status and response body. An empty body
+// sends no payload, which is what a GET needs.
+func (cl *Client) do(ctx context.Context, method, endpoint, proxyIP, path, body string, extraHeaders []string) (int, string, error) {
 	url := "https://" + endpoint + path
 	args := []string{
 		"run", "--rm",
@@ -334,13 +389,15 @@ func (cl *Client) post(ctx context.Context, endpoint, proxyIP, path, body string
 		"-sk", "--connect-timeout", "5", "--max-time", "90",
 		"--resolve", endpoint + ":443:" + proxyIP,
 		"-o", "/dev/stderr", "-w", "%{http_code}",
-		"-X", "POST", url,
+		"-X", method, url,
 		"-H", "Content-Type: application/json",
 	}
 	for _, h := range extraHeaders {
 		args = append(args, "-H", h)
 	}
-	args = append(args, "--data", body)
+	if body != "" {
+		args = append(args, "--data", body)
+	}
 	cmd := exec.CommandContext(ctx, "docker", args...)
 	// -w writes the status code to stdout; -o /dev/stderr writes the body to
 	// stderr so we can capture both separately.

e2e/harness/vllm.go

@@ -18,18 +18,63 @@ const (
 	vllmImage = "nginx:alpine"
 	vllmAlias = "vllm"
 	vllmPort  = "8000/tcp"
+	// vllmStreamPort serves the same wire shapes as an SSE stream. See the
+	// nginx config for why streaming lives on its own listener.
+	vllmStreamPort = "8001/tcp"
 	// VLLMModel is the served model id the mock advertises and echoes back. It
 	// matches a real small model commonly served by vLLM so the provider's
 	// enumerated model and the client's request line up.
 	VLLMModel = "Qwen/Qwen2.5-0.5B-Instruct"
+	// VLLMUnlistedModel is a second id the mock's model listing advertises but
+	// no test provider enumerates, so a filtered listing is observably shorter
+	// than the upstream's own.
+	VLLMUnlistedModel = "Qwen/Qwen2.5-7B-Instruct"
+)
+
+// Token counts the mock reports per wire shape. Tests assert on these rather
+// than on "> 0" so a response parsed with the wrong provider's parser (which
+// would read a different field, or none) fails loudly instead of passing on
+// a coincidental non-zero.
+const (
+	// VLLMChatInputTokens / VLLMChatOutputTokens ride the OpenAI usage block.
+	VLLMChatInputTokens  = 11
+	VLLMChatOutputTokens = 2
+	// VLLMMessagesInputTokens / VLLMMessagesOutputTokens ride the Anthropic
+	// usage block, whose field names the OpenAI parser cannot read.
+	VLLMMessagesInputTokens  = 17
+	VLLMMessagesOutputTokens = 3
+)
+
+// Token counts the streaming surface reports. They differ from the
+// non-streaming ones on purpose: a test that asserts these numbers proves the
+// SSE accumulator ran, rather than a buffered JSON body having been parsed.
+//
+// Input and cache-read arrive on message_start; output arrives on
+// message_delta and supersedes the seed value message_start carries. Any
+// parser that cannot read message_start reports zero input tokens — which is
+// exactly the bug these counts exist to catch.
+const (
+	VLLMStreamInputTokens     = 29
+	VLLMStreamOutputTokens    = 5
+	VLLMStreamCacheReadTokens = 7
 )
 
 // vllmNginxConf emulates a vLLM OpenAI-compatible server over plain HTTP (vLLM's
-// default: no TLS, port 8000). It answers /v1/models with a one-model list and
-// any chat/completions path with a canned OpenAI-shaped chat completion carrying
-// a non-zero usage block, so the proxy's OpenAI parser records real token
-// consumption. Running actual vLLM in CI is infeasible (GPU + multi-GB model
+// default: no TLS, port 8000), and additionally answers the wire shapes the
+// other catalog surfaces speak so one mock can stand in for every provider the
+// proxy routes to. Running actual vLLM in CI is infeasible (GPU + multi-GB model
 // download), so this stands in for the wire contract the proxy depends on.
+//
+// Each shape answers with its own vendor's usage block, so a response parsed
+// under the wrong surface meters zero rather than passing by accident:
+//
+//   - /v1/chat/completions (and any unmatched path): OpenAI chat completion.
+//   - /v1/messages: Anthropic Messages, snake_case usage plus a cache bucket.
+//   - /model/{id}/invoke: Bedrock InvokeModel, which carries the Anthropic body.
+//   - the token-counting endpoints: a count, with no usage block at all.
+//
+// The model listing advertises two models so a policy that authorises one
+// produces an observably shorter list than the upstream's own.
 const vllmNginxConf = `pid /tmp/nginx.pid;
 events {}
 http {
@@ -37,13 +82,75 @@ http {
     listen 8000;
     location = /v1/models {
       default_type application/json;
-      return 200 '{"object":"list","data":[{"id":"Qwen/Qwen2.5-0.5B-Instruct","object":"model","owned_by":"vllm"}]}';
+      return 200 '{"object":"list","data":[{"id":"Qwen/Qwen2.5-0.5B-Instruct","object":"model","owned_by":"vllm"},{"id":"Qwen/Qwen2.5-7B-Instruct","object":"model","owned_by":"vllm"}]}';
+    }
+    location = /v1/messages {
+      default_type application/json;
+      return 200 '{"id":"msg_e2e","type":"message","role":"assistant","model":"claude-sonnet-5","content":[{"type":"text","text":"pong"}],"stop_reason":"end_turn","usage":{"input_tokens":17,"output_tokens":3,"cache_read_input_tokens":5}}';
+    }
+    location = /v1/messages/count_tokens {
+      default_type application/json;
+      return 200 '{"input_tokens":7}';
+    }
+    location ~ ^/model/.+/invoke$ {
+      default_type application/json;
+      return 200 '{"id":"msg_e2e_bedrock","type":"message","role":"assistant","content":[{"type":"text","text":"pong"}],"stop_reason":"end_turn","usage":{"input_tokens":17,"output_tokens":3,"cache_read_input_tokens":5}}';
+    }
+    location ~ ^/model/.+/count-tokens$ {
+      default_type application/json;
+      return 200 '{"inputTokens":9}';
+    }
+    location = /api/hello {
+      return 200;
+    }
+    location = /inference-profiles {
+      default_type application/json;
+      return 200 '{"inferenceProfileSummaries":[{"inferenceProfileId":"us.anthropic.claude-sonnet-5","status":"ACTIVE"}]}';
     }
     location / {
       default_type application/json;
       return 200 '{"id":"chatcmpl-e2e-vllm","object":"chat.completion","created":1700000000,"model":"Qwen/Qwen2.5-0.5B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"pong"},"finish_reason":"stop"}],"usage":{"prompt_tokens":11,"completion_tokens":2,"total_tokens":13}}';
     }
   }
+
+  # The streaming surface, on its own port so the response content type is a
+  # property of the listener rather than of a per-request branch: nginx sets
+  # Content-Type from default_type, which cannot be varied inside an "if", and
+  # a second Content-Type via add_header would leave the proxy reading the
+  # wrong one. A provider record pointed at this port streams every answer.
+  #
+  # Input and cache-read tokens ride message_start, output rides message_delta
+  # — the split that makes a stream different from a buffered body, and the
+  # reason a parser that ignores message_start meters input as zero.
+  server {
+    listen 8001;
+    location = /v1/messages {
+      default_type text/event-stream;
+      return 200 'event: message_start
+data: {"type":"message_start","message":{"id":"msg_e2e_stream","type":"message","role":"assistant","model":"claude-sonnet-5","content":[],"usage":{"input_tokens":29,"output_tokens":1,"cache_read_input_tokens":7}}}
+
+event: content_block_delta
+data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"pong"}}
+
+event: message_delta
+data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":5}}
+
+event: message_stop
+data: {"type":"message_stop"}
+
+';
+    }
+    location / {
+      default_type text/event-stream;
+      return 200 'data: {"choices":[{"delta":{"content":"pong"}}]}
+
+data: {"choices":[],"usage":{"prompt_tokens":29,"completion_tokens":5,"total_tokens":34}}
+
+data: [DONE]
+
+';
+    }
+  }
 }
 `
 
@@ -55,6 +162,10 @@ type VLLM struct {
 	workDir   string
 	// URL is the upstream URL the vllm provider points at (http://<alias>:8000).
 	URL string
+	// StreamURL is the same mock's streaming listener. A provider pointed here
+	// answers every request as SSE, so the proxy's streaming accumulator runs
+	// instead of its buffered-body parser.
+	StreamURL string
 }
 
 // StartVLLM runs the mock vLLM server on the shared network over plain HTTP.
@@ -73,14 +184,17 @@ func StartVLLM(ctx context.Context, c *Combined) (*VLLM, error) {
 
 	req := testcontainers.ContainerRequest{
 		Image:          vllmImage,
-		ExposedPorts:   []string{vllmPort},
+		ExposedPorts:   []string{vllmPort, vllmStreamPort},
 		Networks:       []string{c.network.Name},
 		NetworkAliases: map[string][]string{c.network.Name: {vllmAlias}},
 		Cmd:            []string{"nginx", "-c", "/conf/nginx.conf", "-g", "daemon off;"},
 		HostConfigModifier: func(hc *container.HostConfig) {
 			hc.Binds = append(hc.Binds, workDir+":/conf:ro")
 		},
-		WaitingFor: wait.ForListeningPort(vllmPort).WithStartupTimeout(60 * time.Second),
+		WaitingFor: wait.ForAll(
+			wait.ForListeningPort(vllmPort),
+			wait.ForListeningPort(vllmStreamPort),
+		).WithStartupTimeout(60 * time.Second),
 	}
 
 	ctr, err := testcontainers.GenericContainer(ctx, testcontainers.GenericContainerRequest{
@@ -92,7 +206,12 @@ func StartVLLM(ctx context.Context, c *Combined) (*VLLM, error) {
 		return nil, fmt.Errorf("start vllm container: %w", err)
 	}
 
-	return &VLLM{container: ctr, workDir: workDir, URL: "http://" + vllmAlias + ":8000"}, nil
+	return &VLLM{
+		container: ctr,
+		workDir:   workDir,
+		URL:       "http://" + vllmAlias + ":8000",
+		StreamURL: "http://" + vllmAlias + ":8001",
+	}, nil
 }
 
 // Logs returns the vLLM container logs, for diagnostics on failure.

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

@@ -296,6 +296,8 @@ var providers = []Provider{
 		// account to be on >= 30-day data retention or all requests
 		// 400.
 		Models: []Model{
+			{ID: "claude-opus-5", Label: "Claude Opus 5", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-sonnet-5", Label: "Claude Sonnet 5", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
 			{ID: "claude-fable-5", Label: "Claude Fable 5", InputPer1k: 0.010, OutputPer1k: 0.050, CacheReadPer1k: 0.001, CacheCreationPer1k: 0.0125, ContextWindow: 1000000},
 			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
 			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
@@ -355,6 +357,8 @@ var providers = []Provider{
 		// Llama 3.3 70B entry kept unchanged — LiteLLM tracks only
 		// per-region Llama 3 entries; standalone 3.3 not yet listed.
 		Models: []Model{
+			{ID: "anthropic.claude-opus-5", Label: "Claude Opus 5 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "anthropic.claude-sonnet-5", Label: "Claude Sonnet 5 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
 			{ID: "anthropic.claude-opus-4-8", Label: "Claude Opus 4.8 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
 			{ID: "anthropic.claude-opus-4-7", Label: "Claude Opus 4.7 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
 			{ID: "anthropic.claude-opus-4-6", Label: "Claude Opus 4.6 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
@@ -406,6 +410,8 @@ var providers = []Provider{
 		// exists — the router denies unmeterable publishers rather than forward
 		// them uncounted.
 		Models: []Model{
+			{ID: "claude-opus-5", Label: "Claude Opus 5 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
+			{ID: "claude-sonnet-5", Label: "Claude Sonnet 5 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
 			{ID: "claude-fable-5", Label: "Claude Fable 5 (Vertex)", InputPer1k: 0.010, OutputPer1k: 0.050, CacheReadPer1k: 0.001, CacheCreationPer1k: 0.0125, ContextWindow: 1000000},
 			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
 			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},

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

@@ -0,0 +1,36 @@
+package catalog
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestClaudeLineupSelectable pins the models Claude Code resolves to by
+// default. A model absent from the lineup can't be ticked on a provider
+// record, so llm_router denies it as not-routable and the operator has no
+// way to authorise the client's own default.
+func TestClaudeLineupSelectable(t *testing.T) {
+	for providerID, wanted := range map[string][]string{
+		"anthropic_api": {"claude-opus-5", "claude-sonnet-5", "claude-haiku-4-5"},
+		"bedrock_api":   {"anthropic.claude-opus-5", "anthropic.claude-sonnet-5", "anthropic.claude-haiku-4-5"},
+		"vertex_ai_api": {"claude-opus-5", "claude-sonnet-5", "claude-haiku-4-5"},
+	} {
+		provider, ok := Lookup(providerID)
+		require.True(t, ok, "catalog must define %s", providerID)
+
+		selectable := make(map[string]Model, len(provider.Models))
+		for _, m := range provider.Models {
+			selectable[m.ID] = m
+		}
+		for _, id := range wanted {
+			model, found := selectable[id]
+			require.True(t, found, "%s must offer %s", providerID, id)
+			assert.NotEmpty(t, model.Label, "%s/%s needs a label for the picker", providerID, id)
+			assert.Positive(t, model.InputPer1k, "%s/%s needs an input rate", providerID, id)
+			assert.Positive(t, model.OutputPer1k, "%s/%s needs an output rate", providerID, id)
+			assert.Positive(t, model.ContextWindow, "%s/%s needs a context window", providerID, id)
+		}
+	}
+}

management/internals/modules/agentnetwork/pricing/defaults.go

@@ -47,17 +47,11 @@ var supplementalDefaults = map[string]map[string]Entry{
 		"gpt-5-nano": {InputPer1k: 0.00005, OutputPer1k: 0.0004, CachedInputPer1k: 0.000005},
 	},
 	"anthropic": {
-		// claude-opus-5 is not yet in the catalog lineup but gateway /
-		// grandfathered traffic uses it; priced so it isn't skipped.
-		"claude-opus-5": {InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625},
 		// "kimi-k3[1m]" is the 1M-context alias some Claude Code guides
 		// configure against Moonshot's Anthropic-compatible endpoint;
 		// priced identically to kimi-k3 so those requests aren't skipped.
 		"kimi-k3[1m]": {InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003},
 	},
-	"bedrock": {
-		"anthropic.claude-opus-5": {InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625},
-	},
 }
 
 var (

management/internals/modules/agentnetwork/pricing/defaults_llm_pricing.example.yaml

@@ -82,6 +82,11 @@ anthropic:
     output_per_1k: 0.015
     cache_read_per_1k: 0.0003
     cache_creation_per_1k: 0.00375
+  claude-sonnet-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
   kimi-k3:
     input_per_1k: 0.003
     output_per_1k: 0.015
@@ -145,6 +150,11 @@ bedrock:
     output_per_1k: 0.015
     cache_read_per_1k: 0.0003
     cache_creation_per_1k: 0.00375
+  anthropic.claude-sonnet-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
   meta.llama3-3-70b-instruct:
     input_per_1k: 0.00072
     output_per_1k: 0.00072

management/internals/modules/agentnetwork/pricing/defaults_test.go

@@ -116,11 +116,13 @@ func TestDefaultTable_PinnedRates(t *testing.T) {
 	assert.InDelta(t, 0.010, fable.InputPer1k, 1e-9, "claude-fable-5 input")
 	assert.InDelta(t, 0.0125, fable.CacheCreationPer1k, 1e-9, "claude-fable-5 cache creation")
 
-	// Supplementals present on their surfaces.
+	// Every id below must stay priced whichever source provides it: the
+	// catalog lineup for the current Claude 5 family, supplementalDefaults
+	// for the ids the dashboard deliberately doesn't offer.
 	for surface, ids := range map[string][]string{
 		"openai":    {"gpt-5", "gpt-5-mini", "gpt-5-nano"},
-		"anthropic": {"claude-opus-5", "kimi-k3[1m]", "kimi-k3"},
-		"bedrock":   {"anthropic.claude-opus-5"},
+		"anthropic": {"claude-opus-5", "claude-sonnet-5", "kimi-k3[1m]", "kimi-k3"},
+		"bedrock":   {"anthropic.claude-opus-5", "anthropic.claude-sonnet-5"},
 	} {
 		for _, id := range ids {
 			_, ok := table[surface][id]

management/internals/modules/agentnetwork/synthesizer.go

@@ -211,7 +211,19 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 
 	groupIndex := indexProviderGroups(enabledPolicies)
 
-	routerCfgJSON, err := buildRouterConfigJSON(enabledProviders, groupIndex)
+	// The proxy guardrail is a per-provider fail-closed backstop; the
+	// authoritative per-policy/group decision is management's
+	// SelectPolicyForRequest. A provider lands in that map only when every
+	// authorising policy restricts models.
+	providerAllowlists := buildProviderAllowlists(enabledPolicies, guardrailsByID)
+
+	// Discovery gets the finer view: per policy rather than flattened per
+	// provider, so a listing can be bounded to what the calling groups may
+	// actually use instead of the union across everyone who reaches the
+	// provider.
+	modelPolicies := buildModelPolicies(enabledPolicies, guardrailsByID)
+
+	routerCfgJSON, err := buildRouterConfigJSON(enabledProviders, groupIndex, modelPolicies)
 	if err != nil {
 		return nil, err
 	}
@@ -228,11 +240,6 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 
 	mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
 	applyAccountCollectionControls(&mergedGuardrails, settings)
-	// 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
@@ -351,6 +358,11 @@ type routerProviderRoute struct {
 	AuthHeaderName  string   `json:"auth_header_name"`
 	AuthHeaderValue string   `json:"auth_header_value"`
 	AllowedGroupIDs []string `json:"allowed_group_ids,omitempty"`
+	// ModelPolicies is one entry per enabled policy authorising this provider,
+	// carrying that policy's source groups and the models it permits. The
+	// router bounds a model listing with it, so a provider two groups reach
+	// under different allowlists offers each only its own.
+	ModelPolicies []routerModelPolicy `json:"model_policies,omitempty"`
 	// Vertex marks a Google Vertex AI provider, whose requests carry the
 	// model in the URL path. The router selects it by path, bypassing the
 	// model/vendor table.
@@ -422,7 +434,7 @@ func indexProviderGroups(policies []*types.Policy) map[string][]string {
 // path-prefix tiebreak. Providers no enabled policy authorises
 // (orphans) are intentionally OMITTED so the router never observes a
 // route with an empty ACL.
-func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]string) ([]byte, error) {
+func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]string, modelPolicies map[string][]routerModelPolicy) ([]byte, error) {
 	cfg := routerConfig{Providers: make([]routerProviderRoute, 0, len(providers))}
 	for _, p := range providers {
 		groups, hasPolicy := groupIndex[p.ID]
@@ -449,6 +461,7 @@ func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]
 			AuthHeaderName:          headerName,
 			AuthHeaderValue:         headerValue,
 			AllowedGroupIDs:         groups,
+			ModelPolicies:           modelPolicies[p.ID],
 			Vertex:                  catalog.IsVertexPathStyle(p.ProviderID),
 			Bedrock:                 catalog.IsBedrockPathStyle(p.ProviderID),
 			GCPServiceAccountKeyB64: gcpSAKeyB64,
@@ -1098,3 +1111,46 @@ func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails) {
 		}
 	}
 }
+
+// routerModelPolicy mirrors the router's ModelPolicyRule: one authorising
+// policy's source groups plus the models it permits. Models is nil for a
+// policy that sets no model allowlist, which lifts the restriction for the
+// groups it binds — so nil and empty must survive the round trip distinctly.
+type routerModelPolicy struct {
+	GroupIDs []string `json:"group_ids"`
+	Models   []string `json:"models"`
+}
+
+// buildModelPolicies indexes, per provider, one rule for each enabled policy
+// authorising it: the policy's source groups and the models its guardrail
+// permits.
+//
+// This is deliberately finer than buildProviderAllowlists, which flattens the
+// same inputs into one list per provider for the proxy's fail-closed guardrail.
+// A flattened list cannot answer "what may THIS caller see", so a provider two
+// teams reach under different allowlists would offer each team the other's
+// models — a picker full of entries the next request refuses. Keeping the
+// source groups alongside the models lets the router answer it at request time,
+// where it knows the caller's groups.
+func buildModelPolicies(policies []*types.Policy, byID map[string]*types.Guardrail) map[string][]routerModelPolicy {
+	out := make(map[string][]routerModelPolicy)
+	for _, p := range policies {
+		if p == nil || len(p.SourceGroups) == 0 {
+			continue
+		}
+		restricted, models := policyModelAllowlist(p, byID)
+		rule := routerModelPolicy{GroupIDs: append([]string(nil), p.SourceGroups...)}
+		if restricted {
+			// Never nil when restricted: an allowlist permitting nothing must
+			// stay distinguishable from no allowlist at all.
+			rule.Models = append([]string{}, models...)
+		}
+		for _, providerID := range p.DestinationProviderIDs {
+			if providerID == "" {
+				continue
+			}
+			out[providerID] = append(out[providerID], rule)
+		}
+	}
+	return out
+}

management/internals/modules/agentnetwork/synthesizer_provider_allowlist_test.go

@@ -4,6 +4,7 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 )
@@ -93,3 +94,75 @@ func TestBuildProviderAllowlists(t *testing.T) {
 			"an enabled-but-empty allowlist is restricted with an empty set, not unrestricted")
 	})
 }
+
+// policyForGroups builds an enabled policy binding the given source groups to
+// the given providers under an optional guardrail.
+func policyForGroups(id string, groups []string, guardrailIDs []string, providerIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		Enabled:                true,
+		SourceGroups:           groups,
+		DestinationProviderIDs: providerIDs,
+		GuardrailIDs:           guardrailIDs,
+	}
+}
+
+// TestBuildModelPolicies covers the finer index discovery needs. Where
+// buildProviderAllowlists flattens every authorising policy into one list per
+// provider — enough for a fail-closed backstop, but blind to who is asking —
+// this keeps each policy's source groups beside its models so the router can
+// bound a listing to the calling groups.
+func TestBuildModelPolicies(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("each policy keeps its own groups and models", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForGroups("p1", []string{"grp-eng"}, []string{"g-4o"}, "prov-x"),
+			policyForGroups("p2", []string{"grp-sales"}, []string{"g-opus"}, "prov-x"),
+		}
+		got := buildModelPolicies(policies, byID)
+		assert.Equal(t, []routerModelPolicy{
+			{GroupIDs: []string{"grp-eng"}, Models: []string{"gpt-4o"}},
+			{GroupIDs: []string{"grp-sales"}, Models: []string{"claude-opus-4"}},
+		}, got["prov-x"],
+			"the two policies must stay separable so neither group is offered the other's models")
+	})
+
+	t.Run("an unrestricted policy carries nil models", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForGroups("p1", []string{"grp-eng"}, []string{"g-4o"}, "prov-x"),
+			policyForGroups("p2", []string{"grp-admin"}, nil, "prov-x"),
+		}
+		got := buildModelPolicies(policies, byID)
+		assert.Nil(t, got["prov-x"][1].Models,
+			"no allowlist must reach the router as nil, which lifts the restriction for its groups")
+	})
+
+	t.Run("a disabled allowlist is not a restriction", func(t *testing.T) {
+		policies := []*types.Policy{policyForGroups("p1", []string{"grp-eng"}, []string{"g-disabled"}, "prov-x")}
+		got := buildModelPolicies(policies, byID)
+		assert.Nil(t, got["prov-x"][0].Models,
+			"a guardrail with the allowlist check off restricts nothing")
+	})
+
+	t.Run("an enabled allowlist with no models permits nothing", func(t *testing.T) {
+		byIDEmpty := map[string]*types.Guardrail{
+			"g-empty": {ID: "g-empty", Checks: types.GuardrailChecks{ModelAllowlist: types.GuardrailModelAllowlist{Enabled: true}}},
+		}
+		policies := []*types.Policy{policyForGroups("p1", []string{"grp-eng"}, []string{"g-empty"}, "prov-x")}
+		got := buildModelPolicies(policies, byIDEmpty)
+		require.NotNil(t, got["prov-x"][0].Models,
+			"an empty allowlist must not arrive as nil — that would read as unrestricted")
+		assert.Empty(t, got["prov-x"][0].Models)
+	})
+
+	t.Run("a policy binding no groups is skipped", func(t *testing.T) {
+		policies := []*types.Policy{policyForGroups("p1", nil, []string{"g-4o"}, "prov-x")}
+		assert.Empty(t, buildModelPolicies(policies, byID),
+			"a policy with no source groups authorises nobody, so it bounds nobody's listing")
+	})
+}

proxy/internal/llm/model.go

@@ -13,6 +13,14 @@ func NormalizeBedrockModel(modelID string) string {
 	return sharedllm.NormalizeBedrockModel(modelID)
 }
 
+// NormalizeAnthropicModel strips the trailing "-YYYYMMDD" release-date suffix
+// from an Anthropic model id so a dated id a client pins matches the undated
+// one the operator registered. Thin delegate to shared/llm for the same
+// contract reason as the two below.
+func NormalizeAnthropicModel(modelID string) string {
+	return sharedllm.NormalizeAnthropicModel(modelID)
+}
+
 // NormalizeVertexModel strips the "@version" suffix from a Vertex AI model id
 // so it matches the catalog/pricing key. Thin delegate to shared/llm, kept
 // beside NormalizeBedrockModel for the same contract reason.

proxy/internal/llm/pricing/pricing.go

@@ -10,6 +10,8 @@ package pricing
 import (
 	"fmt"
 	"math"
+
+	sharedllm "github.com/netbirdio/netbird/shared/llm"
 )
 
 // Entry is a single model's input and output pricing, expressed in USD per
@@ -92,7 +94,10 @@ func NewTable(raw map[string]map[string]EntryJSON) (*Table, error) {
 	return &Table{entries: entries}, nil
 }
 
-// Lookup returns the entry for the given provider surface and model.
+// Lookup returns the entry for the given provider surface and model. A
+// dated Anthropic id falls back to its undated form, so a client pinning
+// "claude-sonnet-4-5-20250929" bills at the registered "claude-sonnet-4-5"
+// rate instead of recording no cost at all.
 func (t *Table) Lookup(provider, model string) (Entry, bool) {
 	if t == nil {
 		return Entry{}, false
@@ -101,7 +106,14 @@ func (t *Table) Lookup(provider, model string) (Entry, bool) {
 	if !ok {
 		return Entry{}, false
 	}
-	e, ok := byModel[model]
+	if e, found := byModel[model]; found {
+		return e, true
+	}
+	undated := sharedllm.NormalizeAnthropicModel(model)
+	if undated == model {
+		return Entry{}, false
+	}
+	e, ok := byModel[undated]
 	return e, ok
 }
 

proxy/internal/llm/pricing/pricing_test.go

@@ -175,3 +175,22 @@ func TestNewTable_NilAndEmpty(t *testing.T) {
 	require.NoError(t, err)
 	assert.Empty(t, entries, "nil in, empty (never-matching) map out for the per-record map")
 }
+
+// TestLookup_DatedAnthropicIDFallsBackToUndated covers a client pinning a
+// release date on a model priced under its undated id. Without the
+// fallback the request records no cost at all.
+func TestLookup_DatedAnthropicIDFallsBackToUndated(t *testing.T) {
+	table, err := NewTable(map[string]map[string]EntryJSON{
+		"anthropic": {
+			"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015},
+		},
+	})
+	require.NoError(t, err, "table must build from a valid defaults map")
+
+	entry, ok := table.Lookup("anthropic", "claude-sonnet-4-5-20250929")
+	require.True(t, ok, "a dated id must resolve to the undated entry")
+	assert.InDelta(t, 0.003, entry.InputPer1K, 1e-9, "dated id must bill at the registered rate")
+
+	_, ok = table.Lookup("anthropic", "claude-sonnet-9-9-20250929")
+	assert.False(t, ok, "an unknown family must stay unpriced")
+}

proxy/internal/middleware/builtin/cost_meter/middleware.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"strconv"
 
+	"github.com/netbirdio/netbird/proxy/internal/llm"
 	"github.com/netbirdio/netbird/proxy/internal/llm/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 )
@@ -175,13 +176,28 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 // Anthropic route still bills its cache buckets additively.
 func (m *Middleware) lookupCosts(md []middleware.KV, surface, model string, inTokens, outTokens, cachedTokens, cacheCreationTokens int64) (pricing.Costs, bool) {
 	if recordID := lookupKV(md, middleware.KeyLLMResolvedProviderID); recordID != "" {
-		if entry, ok := m.perRecord[recordID][model]; ok {
+		if entry, ok := perRecordEntry(m.perRecord[recordID], model); ok {
 			return pricing.EntryCosts(entry, surface, inTokens, outTokens, cachedTokens, cacheCreationTokens), true
 		}
 	}
 	return m.defaults.Costs(surface, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
 }
 
+// perRecordEntry resolves the operator's stored price for a model on one
+// provider record, falling back to the undated form of a dated Anthropic id
+// so a client that pins a release date still bills at the registered rate.
+func perRecordEntry(byModel map[string]pricing.Entry, model string) (pricing.Entry, bool) {
+	if entry, ok := byModel[model]; ok {
+		return entry, true
+	}
+	undated := llm.NormalizeAnthropicModel(model)
+	if undated == model {
+		return pricing.Entry{}, false
+	}
+	entry, ok := byModel[undated]
+	return entry, ok
+}
+
 // usd renders a cost as the fixed-precision string every cost.usd_* key
 // carries, so the per-bucket values and the aggregates round identically.
 //

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

@@ -84,8 +84,10 @@ func (m *Middleware) MutationsSupported() bool { return false }
 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)
+	surface, _ := lookupMetadata(in.Metadata, middleware.KeyLLMProvider)
+	nonInference, _ := lookupMetadata(in.Metadata, middleware.KeyLLMNonInference)
 
-	if denial := m.evaluateAllowlist(providerID, model, modelPresent); denial != nil {
+	if denial := m.evaluateAllowlist(providerID, surface, model, modelPresent, nonInference == "true"); denial != nil {
 		return denial, nil
 	}
 
@@ -114,7 +116,7 @@ func (m *Middleware) Close() error { return nil }
 // 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 {
+func (m *Middleware) evaluateAllowlist(providerID, surface, model string, modelPresent, nonInference bool) *middleware.Output {
 	if len(m.cfg.ProviderAllowlists) == 0 {
 		return nil
 	}
@@ -122,7 +124,7 @@ func (m *Middleware) evaluateAllowlist(providerID, model string, modelPresent bo
 	// 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)
+		return denyModel(surface, "", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
 	allowlist, restricted := m.cfg.ProviderAllowlists[providerID]
 	if !restricted {
@@ -133,18 +135,29 @@ func (m *Middleware) evaluateAllowlist(providerID, model string, modelPresent bo
 	// 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.
+	//
+	// The exception is a non-inference endpoint the router already authorised.
+	// The model listing and the connection-warming probe name no model
+	// anywhere — not in a body, not in the path — so failing closed here
+	// rejected model discovery for exactly the accounts that configured an
+	// allowlist, which is the outage this endpoint is meant to avoid. The
+	// per-model lookup does name one (the router stamps it from the path), so
+	// it still falls through to the allowlist check below.
 	if !modelPresent || normaliseModel(model) == "" {
-		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
+		if nonInference {
+			return nil
+		}
+		return denyModel(surface, "", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
 	if modelInAllowlist(allowlist, model) {
 		return nil
 	}
-	return denyModel(model, denyCodeModel, denyMessageModel, denyReasonModel)
+	return denyModel(surface, 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 {
+func denyModel(surface, model, code, message, reason string) *middleware.Output {
 	details := map[string]string{}
 	if model != "" {
 		details["model"] = model
@@ -156,6 +169,7 @@ func denyModel(model, code, message, reason string) *middleware.Output {
 			Code:    code,
 			Message: message,
 			Details: details,
+			Surface: surface,
 		},
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMPolicyDecision, Value: "deny"},

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

@@ -343,3 +343,52 @@ func TestFactoryNormalisesAllowlist(t *testing.T) {
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out2.Decision, "trimmed entry must still match")
 }
+
+// TestAllowlistSkipsNonInferenceWithoutModel covers the reported regression:
+// GET /v1/models carries no model anywhere, so the fail-closed rule above
+// denied model discovery for exactly the accounts that configured a provider
+// allowlist — the clients that read a 403 here render an empty model picker.
+// The router authorises those endpoints by path before the guardrail sees
+// them, so an absent model there is expected rather than undeterminable.
+func TestAllowlistSkipsNonInferenceWithoutModel(t *testing.T) {
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+		middleware.KV{Key: middleware.KeyLLMNonInference, Value: "true"},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision,
+		"model discovery must not be refused because it names no model")
+}
+
+// TestAllowlistStillAppliesToNonInferenceWithModel pins that the exemption is
+// scoped to requests that genuinely name nothing. The per-model lookup
+// (GET /v1/models/{id}) is non-inference too, but the router stamps the model
+// from its path, so the allowlist must still decide it — otherwise the
+// exemption becomes a way to confirm a model the policy blocks.
+func TestAllowlistStillAppliesToNonInferenceWithModel(t *testing.T) {
+	mw := New(providerCfg("gpt-4o"))
+
+	t.Run("model in the allowlist", func(t *testing.T) {
+		out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+			middleware.KV{Key: middleware.KeyLLMNonInference, Value: "true"},
+			middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+		))
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision,
+			"an allowlisted model must stay reachable")
+	})
+
+	t.Run("model outside the allowlist", func(t *testing.T) {
+		out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+			middleware.KV{Key: middleware.KeyLLMNonInference, Value: "true"},
+			middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-5"},
+		))
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionDeny, out.Decision,
+			"non-inference must not become a way past the allowlist")
+		require.NotNil(t, out.DenyReason)
+		assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code,
+			"a named but blocked model is blocked, not unknown")
+	})
+}

proxy/internal/middleware/builtin/llm_identity_inject/middleware.go

@@ -217,6 +217,32 @@ func applyHeaderPair(rule *HeaderPairRule, in *middleware.Input) *middleware.Mut
 	return mutations
 }
 
+// bodyInjectableSurfaces are the request-body dialects that accept the
+// OpenAI-standard identity fields this middleware writes. A surface
+// outside this set gets header-only stamping: "user" and "metadata.tags"
+// are not part of the Anthropic Messages schema, which rejects unknown
+// top-level fields and permits only "user_id" under metadata, so writing
+// them into an Anthropic-shaped body turns a working request into a 400.
+// Claude Code speaks that shape through gateway records pinned to the
+// OpenAI parser, so the check keys on the detected surface rather than
+// on the provider record.
+var bodyInjectableSurfaces = map[string]struct{}{
+	"openai": {},
+	// An empty surface means no parser claimed the path (a custom gateway
+	// base). Those upstreams are OpenAI-compatible by convention, so keep
+	// the long-standing behaviour rather than silently dropping identity.
+	"": {},
+}
+
+// bodyAcceptsOpenAIIdentity reports whether the request body may carry the
+// OpenAI-standard identity fields, read from the surface llm_request_parser
+// resolved from the request path.
+func bodyAcceptsOpenAIIdentity(in *middleware.Input) bool {
+	surface, _ := lookupMetadata(in.Metadata, middleware.KeyLLMProvider)
+	_, ok := bodyInjectableSurfaces[surface]
+	return ok
+}
+
 // injectIntoBody parses the request body and writes the supplied
 // identity dimensions into it. Tags land at metadata.tags (creating
 // the metadata object when absent); the user identity lands at the
@@ -225,6 +251,8 @@ func applyHeaderPair(rule *HeaderPairRule, in *middleware.Input) *middleware.Mut
 // was written. Returns ok=false (no mutation) when:
 //
 //   - both inputs are empty (nothing to write);
+//   - the body speaks a dialect without these fields (see
+//     bodyInjectableSurfaces);
 //   - the body is empty or truncated (we don't have the full document
 //     to safely round-trip);
 //   - the body isn't a JSON object (skip silently — this middleware
@@ -245,6 +273,9 @@ func injectIntoBody(in *middleware.Input, tags []string, userID string) ([]byte,
 	if in == nil || len(in.Body) == 0 || in.BodyTruncated {
 		return nil, false
 	}
+	if !bodyAcceptsOpenAIIdentity(in) {
+		return nil, false
+	}
 	var doc map[string]any
 	if err := json.Unmarshal(in.Body, &doc); err != nil {
 		return nil, false

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

@@ -704,3 +704,57 @@ func TestInject_ExtraHeaders_EmptyValueSkipped(t *testing.T) {
 			"empty extra value must not be stamped")
 	}
 }
+
+// TestInject_AnthropicBodyIsNotRewritten pins the shape gate. Claude Code
+// reaches a LiteLLM record on /v1/messages, where "user" is not a
+// permitted top-level field and metadata accepts only "user_id", so
+// writing the OpenAI-standard fields would turn a working request into a
+// 400 naming a field the client never sent. Header stamping still runs, so
+// spend tracking and per-end-user budgets keep working.
+func TestInject_AnthropicBodyIsNotRewritten(t *testing.T) {
+	rule := liteLLMRuleWithBody()
+	rule.HeaderPair.EndUserIDInBody = true
+	mw := New(Config{Providers: []ProviderInjection{rule}})
+
+	in := newInput(litellmProvider, "alice", []string{"grp-eng"})
+	in.UserEmail = "alice@example.com"
+	in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "anthropic"})
+	in.Body = []byte(`{"model":"claude-sonnet-5","messages":[]}`)
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out.Mutations)
+	assert.Empty(t, out.Mutations.BodyReplace,
+		"an Anthropic-shaped body must reach the upstream unmodified")
+
+	var endUser string
+	for _, kv := range out.Mutations.HeadersAdd {
+		if kv.Key == "x-litellm-end-user-id" {
+			endUser = kv.Value
+		}
+	}
+	assert.Equal(t, "alice@example.com", endUser,
+		"header stamping must still carry identity when body inject is skipped")
+}
+
+// TestInject_OpenAIBodyStillRewritten guards the gate against
+// over-reaching: the OpenAI surface must keep its body-level identity,
+// which is the only path LiteLLM's tag-budget check reads.
+func TestInject_OpenAIBodyStillRewritten(t *testing.T) {
+	mw := New(Config{Providers: []ProviderInjection{liteLLMRuleWithBody()}})
+
+	in := newInput(litellmProvider, "alice", []string{"grp-eng"})
+	in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "openai"})
+	in.Body = []byte(`{"model":"gpt-4o-mini","messages":[]}`)
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out.Mutations)
+	require.NotEmpty(t, out.Mutations.BodyReplace, "the OpenAI surface still gets body tags")
+
+	var doc map[string]any
+	require.NoError(t, json.Unmarshal(out.Mutations.BodyReplace, &doc))
+	meta, ok := doc["metadata"].(map[string]any)
+	require.True(t, ok, "metadata must be an object")
+	assert.NotEmpty(t, meta["tags"], "metadata.tags must still be written")
+}

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

@@ -84,6 +84,15 @@ func (m *Middleware) Invoke(ctx context.Context, in *middleware.Input) (*middlew
 		return allowNoAttribution(), nil
 	}
 
+	// Model-listing and other non-inference endpoints carry no model, and
+	// management's per-model allowlist fails closed on an empty one. The
+	// router has already authorised the route against the caller's groups
+	// and the request consumes no tokens, so gating it on a model that
+	// cannot exist would only break gateway model discovery.
+	if lookupKV(in.Metadata, middleware.KeyLLMNonInference) == "true" {
+		return allowNoAttribution(), nil
+	}
+
 	providerID := lookupKV(in.Metadata, middleware.KeyLLMResolvedProviderID)
 	if providerID == "" {
 		// llm_router didn't emit a resolved provider id — usually
@@ -117,7 +126,7 @@ func (m *Middleware) Invoke(ctx context.Context, in *middleware.Input) (*middlew
 	}
 
 	if resp.GetDecision() == "deny" {
-		return denyFromManagement(resp), nil
+		return denyFromManagement(resp, lookupKV(in.Metadata, middleware.KeyLLMProvider)), nil
 	}
 	return allowFromManagement(resp), nil
 }
@@ -161,7 +170,7 @@ func allowFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.O
 // envelope. The deny code surfaces verbatim through the framework's
 // fixed JSON template; arbitrary middleware bytes can't reach the
 // wire.
-func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.Output {
+func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse, surface string) *middleware.Output {
 	code := resp.GetDenyCode()
 	if code == "" {
 		code = "llm_policy.cap_exceeded"
@@ -176,6 +185,7 @@ func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.Ou
 		DenyReason: &middleware.DenyReason{
 			Code:    code,
 			Message: denyMessageForCode(code),
+			Surface: surface,
 		},
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMPolicyDecision, Value: "deny"},

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

@@ -224,3 +224,35 @@ func TestMetadataKeys_Allowlist(t *testing.T) {
 	}
 	assert.ElementsMatch(t, want, keys)
 }
+
+// TestInvoke_NonInferenceSkipsPreflight covers gateway model discovery:
+// GET /v1/models carries no model, and management's per-model allowlist
+// fails closed on an empty one, so a pre-flight would deny discovery for
+// exactly the accounts that use the model allowlist. The router marks the
+// request non-inference after authorising the route, and the gate must
+// then allow without calling management at all.
+func TestInvoke_NonInferenceSkipsPreflight(t *testing.T) {
+	mgmt := &fakeMgmt{
+		checkResp: &proto.CheckLLMPolicyLimitsResponse{
+			Decision: "deny",
+			DenyCode: "llm_policy.model_blocked",
+		},
+	}
+	m := New(mgmt, nil)
+
+	out := runInvoke(t, m, &middleware.Input{
+		AccountID:  "acc-1",
+		UserID:     "user-bob",
+		UserGroups: []string{"grp-engineers"},
+		Metadata: []middleware.KV{
+			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-1"},
+			{Key: middleware.KeyLLMNonInference, Value: "true"},
+		},
+	})
+
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "model-less endpoints must not be gated on a model")
+	assert.Nil(t, mgmt.checkReq, "no pre-flight may be sent for a non-inference request")
+
+	assert.Empty(t, lookupKV(out.Metadata, middleware.KeyLLMSelectedPolicyID),
+		"no policy is attributed when nothing was metered")
+}

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

@@ -1,9 +1,13 @@
 package llm_request_parser
 
 import (
+	"context"
 	"testing"
 
+	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
 )
 
 func TestParseBedrockPath(t *testing.T) {
@@ -36,3 +40,25 @@ func TestParseBedrockPath(t *testing.T) {
 		}
 	}
 }
+
+// TestInvoke_BedrockCountTokens covers the dedicated token-counting
+// endpoint. Denying it does not break the client, it just pushes context
+// counting back onto the inference endpoint, which is billable.
+func TestInvoke_BedrockCountTokens(t *testing.T) {
+	mw := newMiddleware(t)
+
+	out, err := mw.Invoke(context.Background(), &middleware.Input{
+		URL:  "/model/us.anthropic.claude-sonnet-4-5-20250929-v1:0/count-tokens",
+		Body: []byte(`{"input":{"converse":{"messages":[]}}}`),
+	})
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision)
+
+	model, ok := metaValue(t, out.Metadata, middleware.KeyLLMModel)
+	require.True(t, ok, "count-tokens carries a model in the path and must emit it")
+	assert.Equal(t, "anthropic.claude-sonnet-4-5", model, "model must be normalized like any other action")
+
+	stream, _ := metaValue(t, out.Metadata, middleware.KeyLLMStream)
+	assert.Equal(t, "false", stream, "count-tokens never streams")
+}

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

@@ -61,6 +61,8 @@ func (middlewareImpl) MetadataKeys() []string {
 		middleware.KeyLLMRequestPromptRaw,
 		middleware.KeyLLMCaptureTruncated,
 		middleware.KeyLLMSessionID,
+		middleware.KeyLLMAgentID,
+		middleware.KeyLLMParentAgentID,
 	}
 }
 
@@ -72,9 +74,9 @@ func (middlewareImpl) Close() error { return nil }
 
 // Invoke detects the LLM provider, parses request facts, and emits
 // metadata. Always returns DecisionAllow; never errors. Provider
-// selection prefers the configured providerID (synthesiser-stamped on
-// agent-network targets) so requests routed to a custom upstream URL
-// still resolve. Falls back to URL sniffing when no providerID is set.
+// selection prefers the request path, falling back to the configured
+// providerID (synthesiser-stamped on agent-network targets) so requests
+// routed to a custom upstream URL still resolve.
 func (m middlewareImpl) Invoke(_ context.Context, in *middleware.Input) (*middleware.Output, error) {
 	out := &middleware.Output{Decision: middleware.DecisionAllow}
 	if in == nil {
@@ -92,9 +94,14 @@ func (m middlewareImpl) Invoke(_ context.Context, in *middleware.Input) (*middle
 		return m.invokeBedrock(in, br), nil
 	}
 
-	parser, ok := llm.ParserByName(m.providerID)
+	// A path that names an API surface wins over the configured providerID:
+	// a gateway record pinned to "openai" still serves Claude Code on
+	// /v1/messages, and reading that body with the OpenAI parser loses the
+	// Anthropic usage block and prices the request on the wrong surface.
+	// providerID stays the fallback for upstreams whose path says nothing.
+	parser, ok := llm.DetectParser(extractPath(in.URL))
 	if !ok {
-		parser, ok = llm.DetectParser(extractPath(in.URL))
+		parser, ok = llm.ParserByName(m.providerID)
 	}
 	if !ok {
 		return out, nil
@@ -116,9 +123,9 @@ func (m middlewareImpl) Invoke(_ context.Context, in *middleware.Input) (*middle
 	}
 	appendSessionID := func(md []middleware.KV) []middleware.KV {
 		if sessionID != "" {
-			return append(md, middleware.KV{Key: middleware.KeyLLMSessionID, Value: sessionID})
+			md = append(md, middleware.KV{Key: middleware.KeyLLMSessionID, Value: sessionID})
 		}
-		return md
+		return appendAgentIDs(md, in.Headers)
 	}
 
 	facts, err := parser.ParseRequest(in.Body)
@@ -160,6 +167,41 @@ func (m middlewareImpl) Invoke(_ context.Context, in *middleware.Input) (*middle
 	return out, nil
 }
 
+// agentIDHeader and parentAgentIDHeader carry sub-agent attribution: a
+// coding agent that spawns helpers stamps the spawned agent's id, plus the
+// spawning agent's when that helper is itself nested. Both are opaque
+// identifiers rather than content, so they're emitted regardless of the
+// prompt-collection toggle, the same way the session id is.
+const (
+	agentIDHeader       = "x-claude-code-agent-id"
+	parentAgentIDHeader = "x-claude-code-parent-agent-id"
+)
+
+// appendAgentIDs stamps the sub-agent attribution headers onto the metadata
+// bag, skipping either one the request doesn't carry.
+func appendAgentIDs(md []middleware.KV, headers []middleware.KV) []middleware.KV {
+	for _, pair := range []struct{ key, header string }{
+		{middleware.KeyLLMAgentID, agentIDHeader},
+		{middleware.KeyLLMParentAgentID, parentAgentIDHeader},
+	} {
+		if v := headerValue(headers, pair.header); v != "" {
+			md = append(md, middleware.KV{Key: pair.key, Value: v})
+		}
+	}
+	return md
+}
+
+// headerValue returns the first non-empty value for the named header.
+// Headers arrive in canonical form, so the match is case-insensitive.
+func headerValue(headers []middleware.KV, want string) string {
+	for _, kv := range headers {
+		if strings.EqualFold(kv.Key, want) && kv.Value != "" {
+			return kv.Value
+		}
+	}
+	return ""
+}
+
 // sessionIDHeaders are request header names that may carry a client
 // session identifier, checked in order, case-insensitively. Matching is
 // against Go's canonical header form, so use the hyphenated names the
@@ -173,10 +215,8 @@ var sessionIDHeaders = []string{"x-claude-code-session-id", "session-id", "x-ses
 // canonical form, so the match is case-insensitive.
 func sessionIDFromHeaders(headers []middleware.KV) string {
 	for _, want := range sessionIDHeaders {
-		for _, kv := range headers {
-			if strings.EqualFold(kv.Key, want) && kv.Value != "" {
-				return kv.Value
-			}
+		if v := headerValue(headers, want); v != "" {
+			return v
 		}
 	}
 	return ""
@@ -252,6 +292,12 @@ func parseVertexPath(reqPath string) (vertexRequest, bool) {
 	if c := strings.LastIndex(rest, ":"); c >= 0 {
 		model, action = rest[:c], rest[c+1:]
 	}
+	// Token counting hangs off the model as its own path segment
+	// (".../models/{model}/count-tokens:rawPredict"), so anything past the
+	// first "/" belongs to the method rather than the model id.
+	if slash := strings.Index(model, "/"); slash >= 0 {
+		model = model[:slash]
+	}
 	model = llm.NormalizeVertexModel(model)
 	if model == "" {
 		return vertexRequest{}, false
@@ -298,6 +344,7 @@ func (m middlewareImpl) invokeVertex(in *middleware.Input, vx vertexRequest) *mi
 	if sessionID != "" {
 		md = append(md, middleware.KV{Key: middleware.KeyLLMSessionID, Value: sessionID})
 	}
+	md = appendAgentIDs(md, in.Headers)
 
 	promptTruncated := false
 	if parser != nil && m.capturePrompt {
@@ -345,7 +392,9 @@ func trimBedrockNamespace(reqPath string) string {
 //
 //	/model/{modelId}/{action}
 //
-// action ∈ {invoke, invoke-with-response-stream, converse, converse-stream}.
+// action ∈ {invoke, invoke-with-response-stream, converse, converse-stream,
+// count-tokens}. Token counting carries a model and no usage, so it routes
+// like any other action and meters to zero.
 // The modelId may be URL-encoded and may carry a cross-region inference-profile
 // prefix and a version suffix; normalizeBedrockModel strips both so the model
 // matches catalog pricing.
@@ -369,7 +418,7 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 		return bedrockRequest{}, false
 	}
 	switch action {
-	case "invoke", "converse":
+	case "invoke", "converse", "count-tokens":
 		return bedrockRequest{model: model}, true
 	case "invoke-with-response-stream", "converse-stream":
 		return bedrockRequest{model: model, stream: true}, true
@@ -397,6 +446,7 @@ func (m middlewareImpl) invokeBedrock(in *middleware.Input, br bedrockRequest) *
 	if sessionID != "" {
 		md = append(md, middleware.KV{Key: middleware.KeyLLMSessionID, Value: sessionID})
 	}
+	md = appendAgentIDs(md, in.Headers)
 
 	promptTruncated := false
 	if parser != nil && m.capturePrompt {

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

@@ -45,6 +45,8 @@ func TestMiddleware_StaticSurface(t *testing.T) {
 		middleware.KeyLLMRequestPromptRaw,
 		middleware.KeyLLMCaptureTruncated,
 		middleware.KeyLLMSessionID,
+		middleware.KeyLLMAgentID,
+		middleware.KeyLLMParentAgentID,
 	}
 	assert.Equal(t, expected, keys, "metadata key allowlist must match the spec")
 }
@@ -230,6 +232,31 @@ func TestInvoke_ProviderIDConfigBypassesURLSniff(t *testing.T) {
 	assert.Equal(t, "gpt-4o-mini", model)
 }
 
+func TestInvoke_PathSurfaceBeatsProviderIDConfig(t *testing.T) {
+	// Gateway records (LiteLLM, Portkey, OpenRouter) pin provider_id
+	// "openai", but the same record serves Claude Code on /v1/messages.
+	// Parsing that body as OpenAI reads no usage off the Anthropic
+	// response and prices the request on a surface where no claude-*
+	// model exists, so the path has to win.
+	mw, err := Factory{}.New([]byte(`{"provider_id":"openai"}`))
+	require.NoError(t, err, "factory must accept provider_id config")
+
+	out, err := mw.Invoke(context.Background(), &middleware.Input{
+		URL:  "/v1/messages",
+		Body: []byte(`{"model":"claude-sonnet-5","stream":true,"messages":[{"role":"user","content":"Hi"}]}`),
+	})
+	require.NoError(t, err)
+	require.NotNil(t, out)
+
+	provider, ok := metaValue(t, out.Metadata, middleware.KeyLLMProvider)
+	require.True(t, ok, "provider must be emitted")
+	assert.Equal(t, "anthropic", provider, "the /v1/messages path selects the Anthropic surface")
+
+	model, ok := metaValue(t, out.Metadata, middleware.KeyLLMModel)
+	require.True(t, ok, "model must be extracted")
+	assert.Equal(t, "claude-sonnet-5", model)
+}
+
 func TestInvoke_UnknownProviderIDFallsBackToURL(t *testing.T) {
 	mw, err := Factory{}.New([]byte(`{"provider_id":"not-a-real-parser"}`))
 	require.NoError(t, err, "factory must accept any provider_id string")
@@ -416,3 +443,81 @@ func TestInvoke_NilInputAllows(t *testing.T) {
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "nil input still allows")
 	assert.Empty(t, out.Metadata, "nil input emits no metadata")
 }
+
+// TestParseVertexPath_CountTokensKeepsModel covers Vertex token counting,
+// where the method hangs off the model as its own path segment. Splitting
+// only on the final colon swallowed "/count-tokens" into the model id, so
+// the router saw a model no route could claim.
+func TestParseVertexPath_CountTokensKeepsModel(t *testing.T) {
+	cases := map[string]struct {
+		model  string
+		stream bool
+	}{
+		"/v1/projects/p/locations/global/publishers/anthropic/models/claude-sonnet-5:rawPredict":                       {model: "claude-sonnet-5"},
+		"/v1/projects/p/locations/global/publishers/anthropic/models/claude-sonnet-5:streamRawPredict":                 {model: "claude-sonnet-5", stream: true},
+		"/v1/projects/p/locations/global/publishers/anthropic/models/claude-sonnet-5/count-tokens:rawPredict":          {model: "claude-sonnet-5"},
+		"/v1/projects/p/locations/global/publishers/anthropic/models/claude-sonnet-5@20250929/count-tokens:rawPredict": {model: "claude-sonnet-5"},
+	}
+	for path, want := range cases {
+		vx, ok := parseVertexPath(path)
+		require.True(t, ok, "must parse %q", path)
+		assert.Equal(t, want.model, vx.model, "model for %q", path)
+		assert.Equal(t, want.stream, vx.stream, "stream flag for %q", path)
+		assert.Equal(t, "anthropic", vx.publisher, "publisher for %q", path)
+	}
+}
+
+// TestInvoke_EmitsAgentIDs covers sub-agent attribution: several agents run
+// in parallel inside one session, and without their ids every request in
+// the session attributes to the session alone.
+func TestInvoke_EmitsAgentIDs(t *testing.T) {
+	mw := newMiddleware(t)
+
+	t.Run("spawned agent", func(t *testing.T) {
+		out, err := mw.Invoke(context.Background(), &middleware.Input{
+			URL:  "/v1/messages",
+			Body: []byte(`{"model":"claude-sonnet-5","messages":[]}`),
+			Headers: []middleware.KV{
+				{Key: "X-Claude-Code-Session-Id", Value: "sess-1"},
+				{Key: "X-Claude-Code-Agent-Id", Value: "agent-7"},
+			},
+		})
+		require.NoError(t, err)
+
+		agent, ok := metaValue(t, out.Metadata, middleware.KeyLLMAgentID)
+		require.True(t, ok, "the spawned agent's id must be emitted")
+		assert.Equal(t, "agent-7", agent)
+
+		_, ok = metaValue(t, out.Metadata, middleware.KeyLLMParentAgentID)
+		assert.False(t, ok, "a top-level agent has no parent to emit")
+	})
+
+	t.Run("nested agent", func(t *testing.T) {
+		out, err := mw.Invoke(context.Background(), &middleware.Input{
+			URL:  "/v1/messages",
+			Body: []byte(`{"model":"claude-sonnet-5","messages":[]}`),
+			Headers: []middleware.KV{
+				{Key: "X-Claude-Code-Agent-Id", Value: "agent-9"},
+				{Key: "X-Claude-Code-Parent-Agent-Id", Value: "agent-7"},
+			},
+		})
+		require.NoError(t, err)
+
+		agent, _ := metaValue(t, out.Metadata, middleware.KeyLLMAgentID)
+		assert.Equal(t, "agent-9", agent)
+		parent, ok := metaValue(t, out.Metadata, middleware.KeyLLMParentAgentID)
+		require.True(t, ok, "a nested agent must carry the spawning agent's id")
+		assert.Equal(t, "agent-7", parent)
+	})
+
+	t.Run("absent on a plain request", func(t *testing.T) {
+		out, err := mw.Invoke(context.Background(), &middleware.Input{
+			URL:  "/v1/messages",
+			Body: []byte(`{"model":"claude-sonnet-5","messages":[]}`),
+		})
+		require.NoError(t, err)
+
+		_, ok := metaValue(t, out.Metadata, middleware.KeyLLMAgentID)
+		assert.False(t, ok, "no key is emitted when the client sends no agent id")
+	})
+}

proxy/internal/middleware/builtin/llm_router/bedrock_route_test.go

@@ -1,9 +1,13 @@
 package llm_router
 
 import (
+	"context"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
 )
 
 // TestRouteClaimsModel_BedrockNormalizesCandidate guards the fix for the native
@@ -28,3 +32,86 @@ func TestRouteClaimsModel_BedrockNormalizesCandidate(t *testing.T) {
 	assert.False(t, routeClaimsModel(openai, "us.gpt-4o"),
 		"non-Bedrock routes must not strip a us. prefix")
 }
+
+// TestRouter_BedrockCountTokensRoutes pins that the token-counting action
+// reaches the Bedrock route instead of denying as not-routable.
+func TestRouter_BedrockCountTokensRoutes(t *testing.T) {
+	mw := New(Config{Providers: []ProviderRoute{{
+		ID:              "bedrock-prod",
+		Bedrock:         true,
+		Models:          []string{"anthropic.claude-sonnet-4-5"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "bedrock-runtime.eu-central-1.amazonaws.com",
+	}}})
+
+	in := newInputWithModelAndURL("anthropic.claude-sonnet-4-5",
+		"/model/anthropic.claude-sonnet-4-5-20250929-v1:0/count-tokens")
+	in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "bedrock"})
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "count-tokens must route, not deny")
+	require.NotNil(t, out.Mutations)
+	require.NotNil(t, out.Mutations.RewriteUpstream)
+	assert.Equal(t, "bedrock-runtime.eu-central-1.amazonaws.com", out.Mutations.RewriteUpstream.Host)
+}
+
+// TestRouter_BedrockInferenceProfilesRoutes covers the startup lookups a
+// client makes to resolve a configured inference profile. They carry no
+// model, so before they were recognised they denied and wrote a policy
+// rejection into the access log on every session start.
+func TestRouter_BedrockInferenceProfilesRoutes(t *testing.T) {
+	bedrock := ProviderRoute{
+		ID:              "bedrock-prod",
+		Bedrock:         true,
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "bedrock-runtime.eu-central-1.amazonaws.com",
+	}
+	openai := ProviderRoute{
+		ID:              "openai-prod",
+		Models:          []string{"gpt-4o"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "api.openai.com",
+	}
+	mw := New(Config{Providers: []ProviderRoute{openai, bedrock}})
+
+	for _, path := range []string{
+		"/inference-profiles?type=SYSTEM_DEFINED",
+		"/inference-profiles/us.anthropic.claude-sonnet-5",
+	} {
+		out, err := mw.Invoke(context.Background(), newModellessInput(path))
+		require.NoError(t, err)
+		require.NotNil(t, out)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision, "%s must route", path)
+		require.NotNil(t, out.Mutations)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Equal(t, "bedrock-runtime.eu-central-1.amazonaws.com", out.Mutations.RewriteUpstream.Host,
+			"%s must reach the Bedrock provider, not the first authorised one", path)
+
+		nonInference, _ := metaValue(t, out.Metadata, middleware.KeyLLMNonInference)
+		assert.Equal(t, "true", nonInference, "%s carries no model to gate on", path)
+	}
+}
+
+// TestRouter_BedrockNamespacedInferenceProfilesStripsPrefix pins that the
+// optional gateway namespace is removed before the request goes upstream.
+func TestRouter_BedrockNamespacedInferenceProfilesStripsPrefix(t *testing.T) {
+	mw := New(Config{Providers: []ProviderRoute{{
+		ID:              "bedrock-prod",
+		Bedrock:         true,
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "bedrock-runtime.eu-central-1.amazonaws.com",
+	}}})
+
+	out, err := mw.Invoke(context.Background(), newModellessInput("/bedrock/inference-profiles"))
+	require.NoError(t, err)
+	require.NotNil(t, out.Mutations)
+	require.NotNil(t, out.Mutations.RewriteUpstream)
+	assert.Equal(t, "/bedrock", out.Mutations.RewriteUpstream.StripPathPrefix,
+		"the namespace prefix must not reach the real Bedrock endpoint")
+}

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

@@ -44,6 +44,12 @@ type ProviderRoute struct {
 	AuthHeaderName  string   `json:"auth_header_name"`
 	AuthHeaderValue string   `json:"auth_header_value"`
 	AllowedGroupIDs []string `json:"allowed_group_ids"`
+	// ModelPolicies carries, per authorising policy, the source groups it
+	// binds and the models it permits. The router uses it to bound a model
+	// listing to what THIS caller may use: a provider reachable by two groups
+	// under different allowlists must not offer either group the other's
+	// models. Empty means no policy restricts models on this route.
+	ModelPolicies []ModelPolicyRule `json:"model_policies,omitempty"`
 	// Vertex marks a Google Vertex AI provider. Vertex requests carry the
 	// model in the URL path, so the router selects this route by path
 	// (isVertexPath) and bypasses the model/vendor table entirely.
@@ -65,6 +71,18 @@ type ProviderRoute struct {
 	SkipTLSVerify bool `json:"skip_tls_verify,omitempty"`
 }
 
+// ModelPolicyRule is one authorising policy's contribution to what a caller
+// may use on a route: the source groups it binds, and the models it permits.
+//
+// Models is nil when the policy sets no model allowlist — an unrestricted
+// policy, which lifts the restriction for the groups it binds. That is why
+// nil and empty must stay distinct: an empty list is a guardrail that permits
+// nothing, and collapsing the two would let a listing fail open.
+type ModelPolicyRule struct {
+	GroupIDs []string `json:"group_ids"`
+	Models   []string `json:"models"`
+}
+
 // Config is the on-wire configuration accepted by the factory. An
 // empty Providers slice yields a router that denies every request as
 // not-routable; the synthesiser is responsible for stamping the

proxy/internal/middleware/builtin/llm_router/middleware.go

@@ -109,6 +109,10 @@ func (m *Middleware) MetadataKeys() []string {
 		middleware.KeyLLMAuthorisingGroups,
 		middleware.KeyLLMPolicyDecision,
 		middleware.KeyLLMPolicyReason,
+		middleware.KeyLLMNonInference,
+		// Emitted only for the per-model lookup, whose model lives in the path
+		// rather than a body the parser could read.
+		middleware.KeyLLMModel,
 	}
 }
 
@@ -137,29 +141,26 @@ const (
 // known to a provider that no policy authorises for the caller deny
 // with no_authorised_provider.
 func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middleware.Output, error) {
+	reqPath := requestPath(in.URL)
+	// The caller's API dialect, used to mirror a denial in the vendor's own
+	// error shape so the client can explain it to the user.
+	surface, _ := lookupMetadata(in.Metadata, middleware.KeyLLMProvider)
+	model, _ := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
+
 	// Vertex AI carries the model in the URL path, not the body, and is
 	// selected by path rather than by the model/vendor table. Route it before
 	// the model lookup so a model the parser extracted from the path can't be
 	// claimed by a same-vendor direct provider (e.g. claude-* on api.anthropic.com).
-	reqPath := requestPath(in.URL)
 	if isVertexPath(reqPath) {
-		model, _ := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
 		// The request parser emits no llm.provider for a Vertex publisher it
 		// can't parse (e.g. google/gemini). Forwarding such a request would
 		// bypass token/budget metering, so deny it rather than serve it
 		// unmetered.
-		if vendor, _ := lookupMetadata(in.Metadata, middleware.KeyLLMProvider); vendor == "" {
-			return denyUnmeterable(), nil
+		if surface == "" {
+			return denyUnmeterable(surface), nil
 		}
 		route, outcome := m.matchVertex(reqPath, model, in.UserGroups)
-		switch outcome {
-		case matchOutcomeFound:
-			return m.allowWithRoute(route, in.UserGroups), nil
-		case matchOutcomeUnauthorised:
-			return denyNoAuthorisedRoute(model), nil
-		default:
-			return denyUnknownModel(model), nil
-		}
+		return m.decide(route, outcome, surface, model, in.UserGroups, nil), nil
 	}
 
 	// Bedrock likewise carries the model in the URL path (/model/{id}/{action}),
@@ -167,52 +168,211 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 	// before the model lookup; when the prefix is present, strip it from the
 	// forwarded path so the real Bedrock endpoint receives its native path.
 	if isBedrockPath(reqPath) {
-		model, _ := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
 		native, hadPrefix := splitBedrockNamespace(reqPath)
 		route, outcome := m.matchBedrock(native, model, in.UserGroups)
-		switch outcome {
-		case matchOutcomeFound:
-			out := m.allowWithRoute(route, in.UserGroups)
-			if hadPrefix && out.Mutations != nil && out.Mutations.RewriteUpstream != nil {
-				out.Mutations.RewriteUpstream.StripPathPrefix = bedrockNamespacePrefix
+		return m.decide(route, outcome, surface, model, in.UserGroups, func(out *middleware.Output) {
+			if hadPrefix {
+				stripBedrockNamespace(out)
 			}
-			return out, nil
-		case matchOutcomeUnauthorised:
-			return denyNoAuthorisedRoute(model), nil
-		default:
-			return denyUnknownModel(model), nil
-		}
+		}), nil
 	}
 
-	model, ok := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
-	if !ok || model == "" {
-		// Non-inference endpoints (model listing) carry no model but still
-		// need rewriting from the synth placeholder to a real upstream;
-		// clients such as Codex call GET /v1/models at startup to enumerate
-		// availability and read a 403 as "model unavailable".
-		route, outcome := m.matchModelless(requestPath(in.URL), in.UserGroups)
-		switch outcome {
-		case matchOutcomeFound:
-			return m.allowWithRoute(route, in.UserGroups), nil
-		case matchOutcomeUnauthorised:
-			// A recognised model-less endpoint exists but no provider
-			// authorises the caller — deny as an authorisation failure
-			// rather than masking it as a missing model.
-			return denyNoAuthorisedRoute(model), nil
-		default:
-			return denyMissingModel(), nil
-		}
+	// GET /v1/models/{id} carries no body, so no model reaches the router in
+	// metadata — but the path names one, and answering it confirms a model
+	// exists and is reachable. Authorise it against the model table like any
+	// other per-model request, then mark it non-inference so it still skips
+	// the token pre-flight it would otherwise charge nothing against.
+	if detail, isDetail := modelDetailID(reqPath); isDetail && isNonInferenceMethod(in.Method) {
+		route, outcome := m.matchRoute(detail, surface, reqPath, in.UserGroups)
+		return m.decide(route, outcome, surface, detail, in.UserGroups, func(out *middleware.Output) {
+			markNonInference(out)
+			// The parser reads models from JSON bodies only, and this request
+			// has none, so stamp the one the path names. Without it the
+			// guardrail's own allowlist — a separate, possibly narrower list
+			// than the route's — never sees a model to check.
+			out.Metadata = append(out.Metadata, middleware.KV{Key: middleware.KeyLLMModel, Value: detail})
+		}), nil
 	}
 
-	vendor, _ := lookupMetadata(in.Metadata, middleware.KeyLLMProvider)
-	route, outcome := m.matchRoute(model, vendor, requestPath(in.URL), in.UserGroups)
+	if model == "" {
+		return m.routeModelless(reqPath, surface, in.Method, in.UserGroups), nil
+	}
+
+	route, outcome := m.matchRoute(model, surface, reqPath, in.UserGroups)
+	return m.decide(route, outcome, surface, model, in.UserGroups, nil), nil
+}
+
+// decide turns a per-model match result into the middleware's decision. Every
+// surface that routes by model shares the same two denial arms — a model no
+// route claims is not routable, one that some route claims but none authorises
+// for this caller is an authorisation failure — so they live here once.
+// decorate, when non-nil, adjusts the allow with whatever that surface needs.
+func (m *Middleware) decide(
+	route ProviderRoute,
+	outcome matchOutcome,
+	surface, model string,
+	userGroups []string,
+	decorate func(*middleware.Output),
+) *middleware.Output {
 	switch outcome {
 	case matchOutcomeFound:
-		return m.allowWithRoute(route, in.UserGroups), nil
+		out := m.allowWithRoute(route, surface, userGroups)
+		if decorate != nil {
+			decorate(out)
+		}
+		return out
 	case matchOutcomeUnauthorised:
-		return denyNoAuthorisedRoute(model), nil
+		return denyNoAuthorisedRoute(surface, model)
 	default:
-		return denyUnknownModel(model), nil
+		return denyUnknownModel(surface, model)
+	}
+}
+
+// routeModelless serves the endpoints that name no model at all: the model
+// listing, the connection-warming probe, and the Bedrock inference-profile
+// lookup. They still need rewriting from the synth placeholder to a real
+// upstream — clients such as Codex call GET /v1/models at startup to enumerate
+// availability and read a 403 as "model unavailable".
+func (m *Middleware) routeModelless(reqPath, surface, method string, userGroups []string) *middleware.Output {
+	route, outcome := m.matchModelless(reqPath, method, userGroups)
+	switch outcome {
+	case matchOutcomeFound:
+		out := m.allowWithRoute(route, surface, userGroups)
+		markNonInference(out)
+		if _, hadPrefix := splitBedrockNamespace(reqPath); hadPrefix {
+			stripBedrockNamespace(out)
+		}
+		// What the caller may actually use bounds what the picker may offer:
+		// every entry outside it is a request the chain will deny a moment
+		// later.
+		if reqPath == modelListingPath && out.Mutations != nil && out.Mutations.RewriteUpstream != nil {
+			if models, bounded := discoverableModels(route, userGroups); bounded {
+				out.Mutations.RewriteUpstream.DiscoveryModels = models
+			}
+		}
+		return out
+	case matchOutcomeUnauthorised:
+		// A recognised model-less endpoint exists but no provider authorises
+		// the caller — deny as an authorisation failure rather than masking it
+		// as a missing model.
+		return denyNoAuthorisedRoute(surface, "")
+	default:
+		return denyMissingModel(surface)
+	}
+}
+
+// isNonInferenceMethod reports whether a request method is one the
+// non-inference endpoints actually use: the listing and the per-model lookup
+// are GET, the connection-warming probe is HEAD or GET. The method is the only
+// thing separating "GET /v1/models/{id}" from a POST to the same path carrying
+// an inference body, and the non-inference mark exempts a request from the
+// token pre-flight — so anything else falls through to normal per-model
+// routing, which denies when the request names no model.
+func isNonInferenceMethod(method string) bool {
+	return method == http.MethodGet || method == http.MethodHead
+}
+
+// discoverableModels returns the model ids a caller in userGroups may actually
+// use on this route, and whether the listing should be bounded to them at all.
+//
+// Two things narrow a listing, and both must apply or the picker offers models
+// the very next request refuses:
+//
+//   - the provider's own enumerated models, when it lists any (a gateway record
+//     enumerates nothing and claims everything);
+//   - the model allowlists of the policies that authorise THIS caller. A
+//     provider reachable by two groups under different allowlists must not
+//     offer either group the other's models, which is why the rules carry their
+//     source groups rather than arriving pre-flattened.
+//
+// A policy that sets no allowlist lifts the restriction for the groups it
+// binds, so a caller holding one unrestricted policy sees the provider's full
+// list. bounded is false when nothing narrows the listing — an unrestricted
+// caller on a route that enumerates nothing — in which case the upstream's own
+// answer passes through untouched.
+func discoverableModels(route ProviderRoute, userGroups []string) ([]string, bool) {
+	permitted, restricted := policyPermittedModels(route, userGroups)
+
+	switch {
+	case !restricted && len(route.Models) == 0:
+		return nil, false
+	case !restricted:
+		return append([]string(nil), route.Models...), true
+	case len(route.Models) == 0:
+		// A gateway record enumerates nothing, so the allowlist is the whole
+		// bound — previously such a record offered the upstream's entire
+		// catalogue however narrow the policy was.
+		return sortedModels(permitted), true
+	}
+
+	// Both bound: only what the provider serves and the policy permits.
+	intersection := make(map[string]struct{}, len(route.Models))
+	for _, m := range route.Models {
+		if _, ok := permitted[m]; ok {
+			intersection[m] = struct{}{}
+		}
+	}
+	return sortedModels(intersection), true
+}
+
+// policyPermittedModels folds the rules whose groups intersect the caller's
+// into the set of models they permit. restricted is false when the caller
+// holds at least one authorising policy that sets no allowlist, or when no
+// rule binds them at all.
+func policyPermittedModels(route ProviderRoute, userGroups []string) (map[string]struct{}, bool) {
+	permitted := make(map[string]struct{})
+	restricted := false
+	for _, rule := range route.ModelPolicies {
+		if !groupsIntersect(rule.GroupIDs, userGroups) {
+			continue
+		}
+		if rule.Models == nil {
+			// An unrestricted policy the caller holds lifts the restriction
+			// entirely, whatever the others say.
+			return nil, false
+		}
+		restricted = true
+		for _, m := range rule.Models {
+			permitted[m] = struct{}{}
+		}
+	}
+	return permitted, restricted
+}
+
+// groupsIntersect reports whether the two group-id sets share a member.
+func groupsIntersect(a, b []string) bool {
+	for _, x := range a {
+		for _, y := range b {
+			if x == y {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// sortedModels flattens a model set into a stable slice so the bound the proxy
+// applies — and any test asserting on it — does not depend on map order.
+func sortedModels(set map[string]struct{}) []string {
+	out := make([]string, 0, len(set))
+	for m := range set {
+		out = append(out, m)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// markNonInference tags an allow as a request that spends no tokens, so the
+// limit check skips the management pre-flight it would charge nothing against.
+func markNonInference(out *middleware.Output) {
+	out.Metadata = append(out.Metadata, middleware.KV{Key: middleware.KeyLLMNonInference, Value: "true"})
+}
+
+// stripBedrockNamespace tells the rewrite to drop the optional "/bedrock"
+// gateway namespace so the upstream receives its native Bedrock path.
+func stripBedrockNamespace(out *middleware.Output) {
+	if out.Mutations != nil && out.Mutations.RewriteUpstream != nil {
+		out.Mutations.RewriteUpstream.StripPathPrefix = bedrockNamespacePrefix
 	}
 }
 
@@ -300,12 +460,60 @@ func (m *Middleware) matchRoute(model, vendor, reqPath string, userGroups []stri
 	return best, matchOutcomeFound
 }
 
-// isModelLessPath reports whether reqPath is a known OpenAI-shaped
-// non-inference endpoint that legitimately carries no model in its
-// request (the model-listing endpoints). These must route to an upstream
-// rather than deny, so model enumeration works end to end.
+// connectionWarmPath is the probe Anthropic clients send before their first
+// inference request to open the upstream connection early. Forwarding it
+// warms the connection the request will actually use; denying it only fills
+// the access log with rejections at every session start.
+const connectionWarmPath = "/api/hello"
+
+// modelListingPath is the endpoint clients read at startup to populate
+// their model picker. Its response is a list the proxy can bound; the
+// per-model "/v1/models/{id}" lookup returns a single object and is left
+// alone.
+const modelListingPath = "/v1/models"
+
+// isModelLessPath reports whether reqPath is a known non-inference endpoint
+// that legitimately carries no model at all: the model listing and the
+// connection-warming probe. These must route to an upstream rather than
+// deny, so model enumeration works end to end. The per-model
+// "/v1/models/{id}" lookup is deliberately excluded — it names a model, so
+// it is authorised against the model table instead (see modelDetailID).
 func isModelLessPath(reqPath string) bool {
-	return reqPath == "/v1/models" || strings.HasPrefix(reqPath, "/v1/models/")
+	return reqPath == modelListingPath || reqPath == connectionWarmPath
+}
+
+// modelDetailID returns the model id named by a "/v1/models/{id}" lookup.
+// reqPath comes from url.URL.Path, which is already percent-decoded, so an
+// id carrying a "/" (a self-hosted "Qwen/Qwen2.5-0.5B-Instruct" sent as
+// "Qwen%2FQwen2.5-...") arrives whole and everything after the prefix is the
+// id, separators included.
+func modelDetailID(reqPath string) (string, bool) {
+	if !strings.HasPrefix(reqPath, modelListingPath+"/") {
+		return "", false
+	}
+	id := strings.TrimPrefix(reqPath, modelListingPath+"/")
+	if id == "" {
+		return "", false
+	}
+	return id, true
+}
+
+// isBedrockModelLessPath reports whether reqPath is a Bedrock
+// inference-profile lookup, optionally behind the "/bedrock" gateway
+// namespace. Clients read these at startup to resolve a configured profile
+// to its underlying model. They carry no model of their own, so they route
+// by path to a Bedrock provider rather than through the model table.
+//
+// On native AWS these live on the control plane ("bedrock.<region>") while a
+// provider's upstream is normally the runtime host ("bedrock-runtime.<region>"),
+// so forwarding yields a 404 there. That is deliberate: a client has one base
+// URL, so pointing it straight at the runtime host 404s identically, and
+// forwarding keeps the proxy transparent instead of inventing a policy denial
+// the client would never otherwise see. Operators whose Bedrock upstream is a
+// gateway that does serve the lookup get a working answer.
+func isBedrockModelLessPath(reqPath string) bool {
+	native, _ := splitBedrockNamespace(reqPath)
+	return native == "/inference-profiles" || strings.HasPrefix(native, "/inference-profiles/")
 }
 
 // isVertexPath reports whether reqPath is a Google Vertex AI publisher
@@ -332,20 +540,33 @@ func splitBedrockNamespace(reqPath string) (string, bool) {
 	return reqPath, false
 }
 
+// bedrockActions are the runtime actions that follow the model id in a
+// Bedrock path. count-tokens is here so a client can price its context
+// against the dedicated endpoint; denying it pushes that work back onto
+// the inference endpoint, which bills for it.
+var bedrockActions = []string{
+	"/invoke",
+	"/invoke-with-response-stream",
+	"/converse",
+	"/converse-stream",
+	"/count-tokens",
+}
+
 // isBedrockPath reports whether reqPath is an AWS Bedrock runtime model
-// endpoint: /model/{modelId}/{action} where action is invoke,
-// invoke-with-response-stream, converse, or converse-stream — optionally behind
-// a "/bedrock" gateway-namespace prefix. The model lives in the path, so these
-// requests are routed by path to the Bedrock provider.
+// endpoint: /model/{modelId}/{action} — optionally behind a "/bedrock"
+// gateway-namespace prefix. The model lives in the path, so these requests
+// are routed by path to the Bedrock provider.
 func isBedrockPath(reqPath string) bool {
 	native, _ := splitBedrockNamespace(reqPath)
 	if !strings.HasPrefix(native, "/model/") {
 		return false
 	}
-	return strings.HasSuffix(native, "/invoke") ||
-		strings.HasSuffix(native, "/invoke-with-response-stream") ||
-		strings.HasSuffix(native, "/converse") ||
-		strings.HasSuffix(native, "/converse-stream")
+	for _, action := range bedrockActions {
+		if strings.HasSuffix(native, action) {
+			return true
+		}
+	}
+	return false
 }
 
 // matchVertex selects the Vertex provider authorised for the caller's groups
@@ -425,19 +646,26 @@ func (m *Middleware) matchPathRoute(reqPath, model string, userGroups []string,
 // declaration order), matchOutcomeUnauthorised when no provider authorises
 // the caller, or matchOutcomeUnknownModel when the path isn't a recognised
 // model-less endpoint.
-func (m *Middleware) matchModelless(reqPath string, userGroups []string) (ProviderRoute, matchOutcome) {
-	if !isModelLessPath(reqPath) {
+func (m *Middleware) matchModelless(reqPath, method string, userGroups []string) (ProviderRoute, matchOutcome) {
+	if !isNonInferenceMethod(method) {
 		return ProviderRoute{}, matchOutcomeUnknownModel
 	}
-	var candidates []ProviderRoute
-	for _, route := range m.cfg.Providers {
+	var eligible func(ProviderRoute) bool
+	switch {
+	case isBedrockModelLessPath(reqPath):
+		eligible = func(r ProviderRoute) bool { return r.Bedrock }
+	case isModelLessPath(reqPath):
 		// Vertex/Bedrock are path-routed and don't serve OpenAI-style
 		// model-listing endpoints; including them here could rewrite a
 		// GET /v1/models to an upstream that 404s it.
-		if route.Vertex || route.Bedrock {
-			continue
-		}
-		if routeAuthorisesGroups(route, userGroups) {
+		eligible = func(r ProviderRoute) bool { return !r.Vertex && !r.Bedrock }
+	default:
+		return ProviderRoute{}, matchOutcomeUnknownModel
+	}
+
+	var candidates []ProviderRoute
+	for _, route := range m.cfg.Providers {
+		if eligible(route) && routeAuthorisesGroups(route, userGroups) {
 			candidates = append(candidates, route)
 		}
 	}
@@ -564,6 +792,16 @@ func routeClaimsModel(route ProviderRoute, model string) bool {
 		if route.Bedrock && llm.NormalizeBedrockModel(candidate) == model {
 			return true
 		}
+		// A client may pin a dated Anthropic id ("claude-sonnet-4-5-20250929")
+		// where the operator registered the undated one. Only an undated
+		// registration absorbs a dated request: normalising both sides would
+		// let a route pinned to one dated release claim a different one, so an
+		// operator who deliberately pinned a build would silently serve
+		// another — and with several such routes, ordering would decide which.
+		if candidate == llm.NormalizeAnthropicModel(candidate) &&
+			candidate == llm.NormalizeAnthropicModel(model) {
+			return true
+		}
 	}
 	return false
 }
@@ -612,7 +850,7 @@ func requestPath(raw string) string {
 // provider id so identity-stamping middlewares (llm_identity_inject)
 // tag the request with ONLY the groups that authorised this specific
 // route — not every group the peer happens to be in.
-func (m *Middleware) allowWithRoute(route ProviderRoute, userGroups []string) *middleware.Output {
+func (m *Middleware) allowWithRoute(route ProviderRoute, surface string, userGroups []string) *middleware.Output {
 	rewrite := &middleware.UpstreamRewrite{
 		Scheme: route.UpstreamScheme,
 		Host:   route.UpstreamHost,
@@ -634,7 +872,7 @@ func (m *Middleware) allowWithRoute(route ProviderRoute, userGroups []string) *m
 		// request time (cached + auto-refreshed) instead of a static value.
 		bearer, err := m.gcpBearer(route.GCPServiceAccountKeyB64)
 		if err != nil {
-			return denyUpstreamAuth()
+			return denyUpstreamAuth(surface)
 		}
 		authValue = bearer
 	}
@@ -704,11 +942,12 @@ func (m *Middleware) gcpTokenSource(saKeyB64 string) (oauth2.TokenSource, error)
 // denyUpstreamAuth is returned when the router cannot obtain the upstream
 // credential (e.g. a malformed service-account key or an unreachable token
 // endpoint). It surfaces as a 502 — an upstream problem, not a policy denial.
-func denyUpstreamAuth() *middleware.Output {
+func denyUpstreamAuth(surface string) *middleware.Output {
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 502,
 		DenyReason: &middleware.DenyReason{
+			Surface: surface,
 			Code:    denyCodeUpstreamAuth,
 			Message: "could not obtain upstream credential",
 		},
@@ -722,11 +961,12 @@ func denyUpstreamAuth() *middleware.Output {
 // denyUnmeterable returns the deny envelope for a path-routed request whose
 // publisher has no parser surface, so its usage can't be metered. Serving it
 // would bypass token/budget caps, so it is rejected with a 403.
-func denyUnmeterable() *middleware.Output {
+func denyUnmeterable(surface string) *middleware.Output {
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
+			Surface: surface,
 			Code:    denyCodeUnmeterable,
 			Message: "request publisher is not supported for metering",
 		},
@@ -739,11 +979,12 @@ func denyUnmeterable() *middleware.Output {
 
 // denyMissingModel returns the deny envelope for a request whose
 // envelope has no llm.model metadata.
-func denyMissingModel() *middleware.Output {
+func denyMissingModel(surface string) *middleware.Output {
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
+			Surface: surface,
 			Code:    denyCodeNotRoutable,
 			Message: "missing llm.model on request envelope",
 		},
@@ -756,11 +997,12 @@ func denyMissingModel() *middleware.Output {
 
 // denyUnknownModel returns the deny envelope for a model that no
 // configured provider claims.
-func denyUnknownModel(model string) *middleware.Output {
+func denyUnknownModel(surface, model string) *middleware.Output {
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
+			Surface: surface,
 			Code:    denyCodeNotRoutable,
 			Message: fmt.Sprintf("no provider configured for model %s", model),
 			Details: map[string]string{"model": model},
@@ -775,11 +1017,12 @@ func denyUnknownModel(model string) *middleware.Output {
 // denyNoAuthorisedRoute returns the deny envelope for a model that one
 // or more providers claim, but where no policy authorises the caller's
 // groups for any of those providers.
-func denyNoAuthorisedRoute(model string) *middleware.Output {
+func denyNoAuthorisedRoute(surface, model string) *middleware.Output {
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
+			Surface: surface,
 			Code:    denyCodeNoAuthorisedRoute,
 			Message: fmt.Sprintf("no policy authorises model %s for the caller's groups", model),
 			Details: map[string]string{"model": model},

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

@@ -2,6 +2,7 @@ package llm_router
 
 import (
 	"context"
+	"net/http"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -60,6 +61,8 @@ func TestMiddlewareIdentity(t *testing.T) {
 		[]string{
 			middleware.KeyLLMResolvedProviderID,
 			middleware.KeyLLMAuthorisingGroups,
+			middleware.KeyLLMNonInference,
+			middleware.KeyLLMModel,
 			middleware.KeyLLMPolicyDecision,
 			middleware.KeyLLMPolicyReason,
 		},
@@ -171,8 +174,12 @@ func TestRouter_MissingModel(t *testing.T) {
 // from which a model could be parsed). UserGroups matches defaultTestGroup.
 func newModellessInput(reqURL string) *middleware.Input {
 	return &middleware.Input{
-		Slot:       middleware.SlotOnRequest,
-		URL:        reqURL,
+		Slot: middleware.SlotOnRequest,
+		URL:  reqURL,
+		// The non-inference endpoints are read requests; the method is what
+		// separates them from an inference body posted to the same path, so
+		// state it rather than leaning on the zero value.
+		Method:     http.MethodGet,
 		UserGroups: []string{defaultTestGroup},
 	}
 }
@@ -197,6 +204,12 @@ func TestRouter_ModelLessPath_RoutesToAuthorisedProvider(t *testing.T) {
 
 	provider, _ := metaValue(t, out.Metadata, middleware.KeyLLMResolvedProviderID)
 	assert.Equal(t, "openai-prod", provider, "resolved provider must be the authorised route")
+
+	// The limits gate reads this to tell "no model applies here" from
+	// "the model could not be determined", which fails closed.
+	nonInference, ok := metaValue(t, out.Metadata, middleware.KeyLLMNonInference)
+	require.True(t, ok, "model-less allow must mark the request non-inference")
+	assert.Equal(t, "true", nonInference)
 }
 
 func TestRouter_ModelLessPath_MultiProviderDeclarationOrder(t *testing.T) {
@@ -873,3 +886,403 @@ func TestRouter_EmptyModelsClaimsAnyModel(t *testing.T) {
 	resolved, _ := metaValue(t, out.Metadata, middleware.KeyLLMResolvedProviderID)
 	assert.Equal(t, "litellm", resolved)
 }
+
+// TestRouter_DatedAnthropicModelRoutes covers a client pinning a release
+// date on a model the operator registered undated. Exact matches still win,
+// so an operator who registers both dated releases keeps them distinct.
+func TestRouter_DatedAnthropicModelRoutes(t *testing.T) {
+	mw := New(Config{Providers: []ProviderRoute{{
+		ID:              "anthropic-prod",
+		Vendor:          "anthropic",
+		Models:          []string{"claude-sonnet-4-5"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "api.anthropic.com",
+	}}})
+
+	in := newInputWithModelAndURL("claude-sonnet-4-5-20250929", "/v1/messages")
+	in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "anthropic"})
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "a dated id must route to the undated registration")
+	require.NotNil(t, out.Mutations)
+	require.NotNil(t, out.Mutations.RewriteUpstream)
+	assert.Equal(t, "api.anthropic.com", out.Mutations.RewriteUpstream.Host)
+}
+
+// TestRouter_ConnectionWarmProbeRoutes covers the HEAD /api/hello probe an
+// Anthropic client sends before its first request. Forwarding it warms the
+// connection that request will use; denying it only wrote a rejection into
+// the access log at every session start.
+func TestRouter_ConnectionWarmProbeRoutes(t *testing.T) {
+	mw := New(Config{Providers: []ProviderRoute{{
+		ID:              "anthropic-prod",
+		Vendor:          "anthropic",
+		Models:          []string{"claude-sonnet-5"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "api.anthropic.com",
+	}}})
+
+	in := newModellessInput("/api/hello")
+	in.Method = http.MethodHead
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "the warm-up probe must reach the upstream")
+	require.NotNil(t, out.Mutations)
+	require.NotNil(t, out.Mutations.RewriteUpstream)
+	assert.Equal(t, "api.anthropic.com", out.Mutations.RewriteUpstream.Host)
+
+	nonInference, _ := metaValue(t, out.Metadata, middleware.KeyLLMNonInference)
+	assert.Equal(t, "true", nonInference, "the probe carries no model to gate on")
+}
+
+// TestRouter_ModelListingCarriesAuthorisedModels pins the list the proxy
+// bounds the discovery response with. A catch-all route enumerates nothing,
+// so it must not bound the upstream's list at all.
+func TestRouter_ModelListingCarriesAuthorisedModels(t *testing.T) {
+	enumerated := ProviderRoute{
+		ID:              "anthropic-prod",
+		Models:          []string{"claude-sonnet-5", "claude-haiku-4-5"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "api.anthropic.com",
+	}
+
+	t.Run("enumerated route bounds the listing", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{enumerated}})
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models"))
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Equal(t, []string{"claude-sonnet-5", "claude-haiku-4-5"},
+			out.Mutations.RewriteUpstream.DiscoveryModels,
+			"the picker must be bounded by what the route authorises")
+	})
+
+	t.Run("catch-all route leaves the listing alone", func(t *testing.T) {
+		catchAll := enumerated
+		catchAll.Models = nil
+		mw := New(Config{Providers: []ProviderRoute{catchAll}})
+
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models"))
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Empty(t, out.Mutations.RewriteUpstream.DiscoveryModels,
+			"a route that claims every model cannot bound the upstream's list")
+	})
+
+	t.Run("per-model lookup is not a listing", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{enumerated}})
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models/claude-sonnet-5"))
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Empty(t, out.Mutations.RewriteUpstream.DiscoveryModels,
+			"the single-object lookup has no data array to filter")
+	})
+}
+
+// TestRouter_ModelDetailHonoursAllowlist pins that GET /v1/models/{id} is
+// authorised against the model table. It carries no body model, so treating
+// it as a model-less endpoint would let a caller confirm a model the route
+// does not list — the listing itself is bounded to the allowlist, so the
+// detail lookup must be too.
+func TestRouter_ModelDetailHonoursAllowlist(t *testing.T) {
+	enumerated := ProviderRoute{
+		ID:              "anthropic-prod",
+		Models:          []string{"claude-sonnet-5"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "api.anthropic.com",
+	}
+
+	t.Run("allowlisted model routes and skips metering", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{enumerated}})
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models/claude-sonnet-5"))
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision)
+		require.NotNil(t, out.Mutations)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Equal(t, "api.anthropic.com", out.Mutations.RewriteUpstream.Host)
+
+		nonInference, _ := metaValue(t, out.Metadata, middleware.KeyLLMNonInference)
+		assert.Equal(t, "true", nonInference, "a detail lookup spends no tokens")
+	})
+
+	t.Run("model outside the allowlist denies", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{enumerated}})
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models/claude-opus-5"))
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionDeny, out.Decision,
+			"a model no route lists must not be confirmed by the detail lookup")
+	})
+
+	t.Run("dated id matches its undated registration", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{enumerated}})
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models/claude-sonnet-5-20250929"))
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision,
+			"a pinned release of an allowlisted family stays reachable")
+	})
+
+	t.Run("catch-all route still answers every lookup", func(t *testing.T) {
+		catchAll := enumerated
+		catchAll.Models = nil
+		mw := New(Config{Providers: []ProviderRoute{catchAll}})
+
+		out, err := mw.Invoke(context.Background(), newModellessInput("/v1/models/anything-at-all"))
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision,
+			"a gateway that enumerates nothing cannot refuse a lookup")
+	})
+}
+
+// TestRouter_NonInferenceRequiresReadMethod pins that the non-inference mark —
+// which exempts a request from the token pre-flight — is reachable only by the
+// read methods these endpoints actually use. A POST to the same path could
+// carry an inference body, so it must not buy the exemption; it falls through
+// to normal per-model routing instead, which denies when no model is named.
+func TestRouter_NonInferenceRequiresReadMethod(t *testing.T) {
+	route := ProviderRoute{
+		ID:              "gateway",
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "gateway.example.com",
+	}
+
+	for _, path := range []string{"/v1/models", "/v1/models/claude-sonnet-5", "/api/hello"} {
+		t.Run("POST "+path, func(t *testing.T) {
+			mw := New(Config{Providers: []ProviderRoute{route}})
+
+			in := newModellessInput(path)
+			in.Method = http.MethodPost
+
+			out, err := mw.Invoke(context.Background(), in)
+			require.NoError(t, err)
+			assert.Equal(t, middleware.DecisionDeny, out.Decision,
+				"a write to a non-inference path must not route unmetered")
+
+			nonInference, _ := metaValue(t, out.Metadata, middleware.KeyLLMNonInference)
+			assert.NotEqual(t, "true", nonInference,
+				"only a read method may skip the token pre-flight")
+		})
+	}
+
+	t.Run("HEAD keeps the warm probe working", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{route}})
+
+		in := newModellessInput(connectionWarmPath)
+		in.Method = http.MethodHead
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision,
+			"the HEAD warm probe must still reach the upstream")
+
+		nonInference, _ := metaValue(t, out.Metadata, middleware.KeyLLMNonInference)
+		assert.Equal(t, "true", nonInference,
+			"the HEAD warm probe carries no model to meter")
+	})
+}
+
+// TestRouter_PinnedDatedModelStaysDistinct pins that a route registered
+// against one dated Anthropic release does not claim another. Normalising
+// both sides of the comparison made every dated build of a family
+// interchangeable, so an operator who deliberately pinned a build would have
+// served a different one — and with several such routes, declaration or path
+// order would have decided which.
+func TestRouter_PinnedDatedModelStaysDistinct(t *testing.T) {
+	pinned := ProviderRoute{
+		ID:              "anthropic-pinned",
+		Vendor:          "anthropic",
+		Models:          []string{"claude-sonnet-4-5-20250101"},
+		AllowedGroupIDs: []string{defaultTestGroup},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "pinned.example.com",
+	}
+
+	t.Run("a different dated release is not claimed", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{pinned}})
+		in := newInputWithModelAndURL("claude-sonnet-4-5-20250202", "/v1/messages")
+		in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "anthropic"})
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionDeny, out.Decision,
+			"a route pinned to one dated build must not serve another")
+	})
+
+	t.Run("its own dated release still routes", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{pinned}})
+		in := newInputWithModelAndURL("claude-sonnet-4-5-20250101", "/v1/messages")
+		in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "anthropic"})
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		assert.Equal(t, middleware.DecisionAllow, out.Decision, "the exact match must still route")
+	})
+
+	t.Run("two pinned builds each route to their own provider", func(t *testing.T) {
+		other := pinned
+		other.ID = "anthropic-pinned-newer"
+		other.Models = []string{"claude-sonnet-4-5-20250202"}
+		other.UpstreamHost = "newer.example.com"
+		mw := New(Config{Providers: []ProviderRoute{pinned, other}})
+
+		in := newInputWithModelAndURL("claude-sonnet-4-5-20250202", "/v1/messages")
+		in.Metadata = append(in.Metadata, middleware.KV{Key: middleware.KeyLLMProvider, Value: "anthropic"})
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		require.Equal(t, middleware.DecisionAllow, out.Decision)
+		require.NotNil(t, out.Mutations)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Equal(t, "newer.example.com", out.Mutations.RewriteUpstream.Host,
+			"declaration order must not decide between two deliberately pinned builds")
+	})
+}
+
+// TestRouter_DiscoveryBoundToCallersPolicies pins that a model listing is
+// bounded by the policies that authorise the caller, not by the union across
+// everyone who can reach the provider. Two teams sharing one provider record
+// under different allowlists is the case that makes the difference visible: a
+// flattened per-provider list would offer each team the other's models, and
+// every one of those entries is a request the guardrail then refuses.
+func TestRouter_DiscoveryBoundToCallersPolicies(t *testing.T) {
+	const (
+		eng   = "grp-eng"
+		sales = "grp-sales"
+	)
+	route := ProviderRoute{
+		ID:              "shared-gateway",
+		Models:          []string{"claude-sonnet-5", "claude-haiku-4-5", "gpt-4o"},
+		AllowedGroupIDs: []string{eng, sales},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "gateway.example.com",
+		ModelPolicies: []ModelPolicyRule{
+			{GroupIDs: []string{eng}, Models: []string{"claude-sonnet-5"}},
+			{GroupIDs: []string{sales}, Models: []string{"gpt-4o"}},
+		},
+	}
+
+	listingFor := func(t *testing.T, group string) []string {
+		t.Helper()
+		mw := New(Config{Providers: []ProviderRoute{route}})
+		in := newModellessInput(modelListingPath)
+		in.UserGroups = []string{group}
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		require.Equal(t, middleware.DecisionAllow, out.Decision)
+		require.NotNil(t, out.Mutations)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		return out.Mutations.RewriteUpstream.DiscoveryModels
+	}
+
+	t.Run("each group sees only its own policy's models", func(t *testing.T) {
+		assert.Equal(t, []string{"claude-sonnet-5"}, listingFor(t, eng),
+			"engineering must not be offered the model only sales may use")
+		assert.Equal(t, []string{"gpt-4o"}, listingFor(t, sales),
+			"sales must not be offered the model only engineering may use")
+	})
+
+	t.Run("a model no policy allows is offered to nobody", func(t *testing.T) {
+		for _, group := range []string{eng, sales} {
+			assert.NotContains(t, listingFor(t, group), "claude-haiku-4-5",
+				"the provider serves it, but no policy permits it")
+		}
+	})
+}
+
+// TestRouter_DiscoveryUnrestrictedPolicy covers the lifting rule: a caller
+// holding one policy without a model allowlist sees everything the provider
+// enumerates, whatever the other policies say.
+func TestRouter_DiscoveryUnrestrictedPolicy(t *testing.T) {
+	const (
+		eng   = "grp-eng"
+		admin = "grp-admin"
+	)
+	route := ProviderRoute{
+		ID:              "shared-gateway",
+		Models:          []string{"claude-sonnet-5", "gpt-4o"},
+		AllowedGroupIDs: []string{eng, admin},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "gateway.example.com",
+		ModelPolicies: []ModelPolicyRule{
+			{GroupIDs: []string{eng}, Models: []string{"claude-sonnet-5"}},
+			// nil Models: a policy that sets no allowlist at all.
+			{GroupIDs: []string{admin}},
+		},
+	}
+	mw := New(Config{Providers: []ProviderRoute{route}})
+
+	in := newModellessInput(modelListingPath)
+	in.UserGroups = []string{eng, admin}
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out.Mutations.RewriteUpstream)
+	assert.ElementsMatch(t, []string{"claude-sonnet-5", "gpt-4o"},
+		out.Mutations.RewriteUpstream.DiscoveryModels,
+		"an unrestricted policy the caller holds lifts the restriction")
+}
+
+// TestRouter_DiscoveryOnGatewayRecord covers a record that enumerates no
+// models. It previously offered the upstream's whole catalogue however narrow
+// the policy was, because there was nothing to intersect against; the policy
+// allowlist is now the bound on its own.
+func TestRouter_DiscoveryOnGatewayRecord(t *testing.T) {
+	const eng = "grp-eng"
+	base := ProviderRoute{
+		ID:              "litellm",
+		AllowedGroupIDs: []string{eng},
+		UpstreamScheme:  "https",
+		UpstreamHost:    "litellm.internal",
+	}
+
+	t.Run("a policy allowlist bounds it", func(t *testing.T) {
+		route := base
+		route.ModelPolicies = []ModelPolicyRule{{GroupIDs: []string{eng}, Models: []string{"gpt-4o"}}}
+		mw := New(Config{Providers: []ProviderRoute{route}})
+
+		in := newModellessInput(modelListingPath)
+		in.UserGroups = []string{eng}
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		assert.Equal(t, []string{"gpt-4o"}, out.Mutations.RewriteUpstream.DiscoveryModels,
+			"a catch-all record must still be bounded by what policy permits")
+	})
+
+	t.Run("an allowlist permitting nothing offers nothing", func(t *testing.T) {
+		route := base
+		route.ModelPolicies = []ModelPolicyRule{{GroupIDs: []string{eng}, Models: []string{}}}
+		mw := New(Config{Providers: []ProviderRoute{route}})
+
+		in := newModellessInput(modelListingPath)
+		in.UserGroups = []string{eng}
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Empty(t, out.Mutations.RewriteUpstream.DiscoveryModels,
+			"an empty allowlist permits nothing, and must not be read as unrestricted")
+	})
+
+	t.Run("no policy restriction leaves the listing alone", func(t *testing.T) {
+		mw := New(Config{Providers: []ProviderRoute{base}})
+
+		in := newModellessInput(modelListingPath)
+		in.UserGroups = []string{eng}
+
+		out, err := mw.Invoke(context.Background(), in)
+		require.NoError(t, err)
+		require.NotNil(t, out.Mutations.RewriteUpstream)
+		assert.Nil(t, out.Mutations.RewriteUpstream.DiscoveryModels,
+			"nothing narrows the listing, so the upstream's own answer passes through")
+	})
+}

proxy/internal/middleware/decision.go

@@ -11,11 +11,78 @@ var codeRegex = regexp.MustCompile(`^[a-z][a-z0-9._-]{0,63}$`)
 // denyResponse is the on-wire shape rendered by RenderDenyResponse.
 // Keeping this as a typed struct ensures we never leak
 // middleware-supplied bytes outside known fields.
+//
+// Type and Error mirror the denial in the vendor's own error shape when
+// the request reached a known LLM surface. LLM clients only parse their
+// provider's envelope, so without the mirror a budget stop reaches the
+// user as an unexplained API error. The NetBird fields stay where they
+// were, so the body is a superset and existing consumers are unaffected.
 type denyResponse struct {
 	Code       string            `json:"code"`
 	Message    string            `json:"message,omitempty"`
 	Details    map[string]string `json:"details,omitempty"`
 	Middleware string            `json:"middleware,omitempty"`
+	Type       string            `json:"type,omitempty"`
+	Error      *providerError    `json:"error,omitempty"`
+}
+
+// providerError is the nested error object both vendor envelopes carry.
+type providerError struct {
+	Type    string `json:"type"`
+	Message string `json:"message,omitempty"`
+	Code    string `json:"code,omitempty"`
+}
+
+// Vendor error types keyed by HTTP status, per each provider's published
+// error reference.
+const (
+	anthropicErrInvalidRequest = "invalid_request_error"
+	anthropicErrPermission     = "permission_error"
+	anthropicErrRateLimit      = "rate_limit_error"
+	anthropicErrAPI            = "api_error"
+	openAIErrInvalidRequest    = "invalid_request_error"
+	openAIErrRateLimit         = "rate_limit_error"
+)
+
+// providerEnvelope returns the vendor-shaped mirror for a denial on the
+// given surface, or nil when the surface has no envelope we can speak.
+// message is the already-redacted public message.
+func providerEnvelope(surface, code, message string, status int) (string, *providerError) {
+	switch surface {
+	case "anthropic":
+		return "error", &providerError{
+			Type:    anthropicErrorType(status),
+			Message: message,
+		}
+	case "openai":
+		return "", &providerError{
+			Type:    openAIErrorType(status),
+			Message: message,
+			Code:    code,
+		}
+	default:
+		return "", nil
+	}
+}
+
+func anthropicErrorType(status int) string {
+	switch status {
+	case http.StatusForbidden:
+		return anthropicErrPermission
+	case http.StatusTooManyRequests:
+		return anthropicErrRateLimit
+	case http.StatusBadRequest:
+		return anthropicErrInvalidRequest
+	default:
+		return anthropicErrAPI
+	}
+}
+
+func openAIErrorType(status int) string {
+	if status == http.StatusTooManyRequests {
+		return openAIErrRateLimit
+	}
+	return openAIErrInvalidRequest
 }
 
 // RenderDenyResponse writes a structured JSON deny body. Status is
@@ -36,6 +103,7 @@ func RenderDenyResponse(w http.ResponseWriter, middlewareID string, reason *Deny
 		Message:    truncate(Scan(reason.Message), 256),
 		Middleware: truncate(Scan(middlewareID), 64),
 	}
+	resp.Type, resp.Error = providerEnvelope(reason.Surface, resp.Code, resp.Message, status)
 	if n := len(reason.Details); n > 0 {
 		resp.Details = make(map[string]string, min(n, 8))
 		for k, v := range reason.Details {

proxy/internal/middleware/decision_test.go

@@ -0,0 +1,92 @@
+package middleware
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// decodeDeny renders a denial and returns the parsed body plus the status.
+func decodeDeny(t *testing.T, reason *DenyReason, status int) (map[string]any, int) {
+	t.Helper()
+	rec := httptest.NewRecorder()
+	RenderDenyResponse(rec, "llm_limit_check", reason, status)
+
+	var body map[string]any
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &body), "deny body must be valid JSON")
+	return body, rec.Code
+}
+
+// TestRenderDeny_AnthropicSurfaceMirrorsVendorShape covers a budget stop
+// reaching Claude Code. The client only parses the Anthropic envelope, so
+// without the mirror the user sees an unexplained API error instead of the
+// reason their request was refused.
+func TestRenderDeny_AnthropicSurfaceMirrorsVendorShape(t *testing.T) {
+	body, status := decodeDeny(t, &DenyReason{
+		Code:    "llm_policy.budget_cap_exceeded",
+		Message: "LLM policy limit exceeded",
+		Surface: "anthropic",
+	}, http.StatusForbidden)
+
+	assert.Equal(t, http.StatusForbidden, status)
+	assert.Equal(t, "error", body["type"], "Anthropic errors carry type=error at the top level")
+
+	errObj, ok := body["error"].(map[string]any)
+	require.True(t, ok, "error must be an object")
+	assert.Equal(t, "permission_error", errObj["type"], "403 maps to permission_error")
+	assert.Equal(t, "LLM policy limit exceeded", errObj["message"])
+
+	// The NetBird fields stay put so existing consumers keep working.
+	assert.Equal(t, "llm_policy.budget_cap_exceeded", body["code"])
+	assert.Equal(t, "LLM policy limit exceeded", body["message"])
+	assert.Equal(t, "llm_limit_check", body["middleware"])
+}
+
+// TestRenderDeny_OpenAISurfaceMirrorsVendorShape pins the OpenAI envelope,
+// which nests the code and carries no top-level type.
+func TestRenderDeny_OpenAISurfaceMirrorsVendorShape(t *testing.T) {
+	body, _ := decodeDeny(t, &DenyReason{
+		Code:    "llm_policy.model_blocked",
+		Message: "model is not in the policy allowlist",
+		Surface: "openai",
+	}, http.StatusForbidden)
+
+	assert.NotContains(t, body, "type", "OpenAI errors have no top-level type")
+
+	errObj, ok := body["error"].(map[string]any)
+	require.True(t, ok, "error must be an object")
+	assert.Equal(t, "invalid_request_error", errObj["type"])
+	assert.Equal(t, "llm_policy.model_blocked", errObj["code"], "the NetBird code rides in the vendor code field")
+	assert.Equal(t, "model is not in the policy allowlist", errObj["message"])
+}
+
+// TestRenderDeny_RateLimitStatusMapsToVendorRateLimit pins the mapping a
+// client's backoff keys on.
+func TestRenderDeny_RateLimitStatusMapsToVendorRateLimit(t *testing.T) {
+	body, status := decodeDeny(t, &DenyReason{
+		Code:    "llm_policy.token_cap_exceeded",
+		Message: "LLM policy limit exceeded",
+		Surface: "anthropic",
+	}, http.StatusTooManyRequests)
+
+	assert.Equal(t, http.StatusTooManyRequests, status, "429 must survive the status clamp")
+	errObj := body["error"].(map[string]any)
+	assert.Equal(t, "rate_limit_error", errObj["type"])
+}
+
+// TestRenderDeny_NoSurfaceKeepsLegacyShape guards non-LLM middlewares and
+// denials raised before a surface is known.
+func TestRenderDeny_NoSurfaceKeepsLegacyShape(t *testing.T) {
+	body, _ := decodeDeny(t, &DenyReason{
+		Code:    "llm_policy.model_not_routable",
+		Message: "no provider configured for model x",
+	}, http.StatusForbidden)
+
+	assert.NotContains(t, body, "type", "no surface means no vendor mirror")
+	assert.NotContains(t, body, "error", "no surface means no vendor mirror")
+	assert.Equal(t, "llm_policy.model_not_routable", body["code"])
+}

proxy/internal/middleware/keys.go

@@ -22,6 +22,15 @@ const (
 	// body. Empty for clients that don't send one.
 	KeyLLMSessionID = "llm.session_id"
 
+	// Sub-agent attribution (emitted by llm_request_parser from the
+	// client's request headers). A coding agent that spawns helpers
+	// stamps the spawned agent's id, and the spawning agent's id when
+	// the helper is itself nested, so cost within one session can be
+	// split across the agents that ran in parallel. These identify an
+	// agent, not a person or a device: never treat them as a user id.
+	KeyLLMAgentID       = "llm.agent_id"
+	KeyLLMParentAgentID = "llm.parent_agent_id"
+
 	// LLM response-side metadata (emitted by llm_response_parser).
 	//nolint:gosec // metadata key name, not a credential
 	KeyLLMInputTokens = "llm.input_tokens"
@@ -66,6 +75,14 @@ const (
 	// downstream gateways' spend logs.
 	KeyLLMAuthorisingGroups = "llm.authorising_groups"
 
+	// LLM non-inference marker (emitted by llm_router on the allow path
+	// for endpoints that legitimately carry no model, such as model
+	// listing). The router still authorises these against the caller's
+	// groups; the marker only tells the limits gate that a per-model
+	// allowlist has nothing to evaluate, so an empty model must not be
+	// read as an undetermined one. Never derived from client input.
+	KeyLLMNonInference = "llm.non_inference"
+
 	// LLM policy attribution (emitted by llm_limit_check on the allow
 	// path). Names the policy that paid for this request and the
 	// dimension counters the post-flight llm_limit_record middleware

proxy/internal/middleware/types.go

@@ -179,6 +179,12 @@ type DenyReason struct {
 	Code    string
 	Message string
 	Details map[string]string
+	// Surface names the LLM API dialect the caller speaks (the
+	// llm.provider value), so the rendered body can mirror the denial in
+	// that vendor's error shape alongside the NetBird fields. Empty for
+	// non-LLM middlewares and for denials raised before a surface was
+	// resolved; the body then carries the NetBird fields alone.
+	Surface string
 }
 
 // Output is the value each middleware returns to the dispatcher. The
@@ -247,6 +253,12 @@ type UpstreamRewrite struct {
 	// without verifying its TLS certificate. Set by llm_router from the
 	// provider's skip_tls_verification for self-hosted / internal gateways.
 	SkipTLSVerify bool
+	// DiscoveryModels, when non-empty, is the set of model ids the resolved
+	// route authorises, and the proxy drops everything else from the
+	// model-listing response. Empty leaves the upstream's list untouched,
+	// which is what a route that claims every model wants. Set by
+	// llm_router on a model-listing request only.
+	DiscoveryModels []string
 }
 
 // AuthHeader is a single name/value pair the proxy injects on the

proxy/internal/proxy/discovery_filter.go

@@ -0,0 +1,191 @@
+package proxy
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+
+	sharedllm "github.com/netbirdio/netbird/shared/llm"
+)
+
+// maxDiscoveryBodyBytes bounds the model-listing response the filter will
+// buffer. A listing is a few kilobytes of ids; anything larger is not a
+// listing we recognise, and buffering it to rewrite would cost more than
+// the filtering is worth.
+const maxDiscoveryBodyBytes = 1 << 20
+
+// modelDiscoveryFilter returns a ModifyResponse hook that drops models the
+// caller's policy does not authorise from a model-listing response, then
+// delegates to next (which may be nil).
+//
+// Clients populate their model picker from this endpoint, so an unfiltered
+// list offers models the very next request denies. The filter is
+// best-effort: a response it cannot safely rewrite passes through
+// untouched rather than reaching the client corrupted.
+func modelDiscoveryFilter(allowed []string, next func(*http.Response) error) func(*http.Response) error {
+	permitted := make(map[string]struct{}, len(allowed)*2)
+	for _, id := range allowed {
+		permitted[id] = struct{}{}
+		permitted[sharedllm.NormalizeAnthropicModel(id)] = struct{}{}
+	}
+
+	return func(resp *http.Response) error {
+		if err := filterModelListing(resp, permitted); err != nil {
+			return err
+		}
+		if next == nil {
+			return nil
+		}
+		return next(resp)
+	}
+}
+
+// filterModelListing rewrites the response body in place, keeping only the
+// entries whose id the policy authorises. Responses that are not a plain
+// JSON listing are left alone.
+func filterModelListing(resp *http.Response, permitted map[string]struct{}) error {
+	if !isPlainJSONListing(resp) {
+		return nil
+	}
+
+	// One byte past the cap, so an oversized body is detectable without
+	// buffering all of it.
+	body, err := io.ReadAll(io.LimitReader(resp.Body, maxDiscoveryBodyBytes+1))
+	if err != nil {
+		_ = resp.Body.Close()
+		return err
+	}
+	if len(body) > maxDiscoveryBodyBytes {
+		// Too large to filter. Put the bytes already read back in front of the
+		// unread remainder and forward the response exactly as the upstream
+		// sent it, headers included. Buffering what was read and closing here
+		// would truncate the body at the cap and hand the client a short,
+		// invalid listing — worse than not filtering at all.
+		resp.Body = spliceBody(body, resp.Body)
+		return nil
+	}
+	if err := resp.Body.Close(); err != nil {
+		return err
+	}
+
+	filtered, ok := filterListingBody(body, permitted)
+	if !ok {
+		restoreBody(resp, body)
+		return nil
+	}
+	restoreBody(resp, filtered)
+	return nil
+}
+
+// isPlainJSONListing reports whether the response is a JSON body the filter
+// can parse. A content-encoded body is skipped: the transport only
+// transparently decompresses what it negotiated itself, and the client
+// negotiates its own encoding on this request.
+func isPlainJSONListing(resp *http.Response) bool {
+	if resp == nil || resp.Body == nil {
+		return false
+	}
+	if resp.StatusCode != http.StatusOK {
+		return false
+	}
+	if enc := resp.Header.Get("Content-Encoding"); enc != "" && !strings.EqualFold(enc, "identity") {
+		return false
+	}
+	return strings.Contains(strings.ToLower(resp.Header.Get("Content-Type")), "application/json")
+}
+
+// filterListingBody returns the listing with unauthorised entries removed.
+// ok is false when the body is not a listing shape, in which case the
+// caller must forward the original bytes.
+func filterListingBody(body []byte, permitted map[string]struct{}) ([]byte, bool) {
+	var doc map[string]json.RawMessage
+	if err := json.Unmarshal(body, &doc); err != nil {
+		return nil, false
+	}
+	raw, present := doc["data"]
+	if !present {
+		return nil, false
+	}
+	var entries []map[string]json.RawMessage
+	if err := json.Unmarshal(raw, &entries); err != nil {
+		return nil, false
+	}
+
+	kept := make([]map[string]json.RawMessage, 0, len(entries))
+	for _, entry := range entries {
+		if entryPermitted(entry, permitted) {
+			kept = append(kept, entry)
+		}
+	}
+
+	encoded, err := json.Marshal(kept)
+	if err != nil {
+		return nil, false
+	}
+	doc["data"] = encoded
+	out, err := json.Marshal(doc)
+	if err != nil {
+		return nil, false
+	}
+	return out, true
+}
+
+// entryPermitted reports whether a listing entry names a model the policy
+// authorises, trying every form the same model is written in.
+func entryPermitted(entry map[string]json.RawMessage, permitted map[string]struct{}) bool {
+	raw, ok := entry["id"]
+	if !ok {
+		return false
+	}
+	var id string
+	if err := json.Unmarshal(raw, &id); err != nil {
+		return false
+	}
+	for _, candidate := range modelIDForms(id) {
+		if _, ok := permitted[candidate]; ok {
+			return true
+		}
+	}
+	return false
+}
+
+// modelIDForms returns the forms a single model id may be written in: the id
+// itself, its undated form, and the same two with a gateway's provider
+// prefix removed ("vertex_ai/claude-sonnet-5"). The bare id is tried first,
+// because a self-hosted id can legitimately contain a slash of its own
+// ("Qwen/Qwen2.5-0.5B-Instruct") and must not be cut down to its tail.
+func modelIDForms(id string) []string {
+	if id == "" {
+		return nil
+	}
+	forms := []string{id, sharedllm.NormalizeAnthropicModel(id)}
+	if slash := strings.LastIndex(id, "/"); slash >= 0 {
+		tail := id[slash+1:]
+		forms = append(forms, tail, sharedllm.NormalizeAnthropicModel(tail))
+	}
+	return forms
+}
+
+// restoreBody puts body back on the response and fixes the length headers
+// so the client reads exactly what is there.
+// spliceBody returns a ReadCloser that yields prefix followed by whatever is
+// left in rest, closing rest when closed. It lets the filter put back bytes it
+// consumed while deciding, without owning the rest of the stream.
+func spliceBody(prefix []byte, rest io.ReadCloser) io.ReadCloser {
+	return struct {
+		io.Reader
+		io.Closer
+	}{
+		Reader: io.MultiReader(bytes.NewReader(prefix), rest),
+		Closer: rest,
+	}
+}
+
+func restoreBody(resp *http.Response, body []byte) {
+	resp.Body = io.NopCloser(bytes.NewReader(body))
+	resp.ContentLength = int64(len(body))
+	resp.Header.Set("Content-Length", strconv.Itoa(len(body)))
+}

proxy/internal/proxy/discovery_filter_test.go

@@ -0,0 +1,217 @@
+package proxy
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// jsonListingResponse builds a 200 model-listing response with the given
+// body, as an upstream would return it.
+func jsonListingResponse(body string) *http.Response {
+	resp := &http.Response{
+		StatusCode:    http.StatusOK,
+		Header:        http.Header{},
+		Body:          io.NopCloser(strings.NewReader(body)),
+		ContentLength: int64(len(body)),
+	}
+	resp.Header.Set("Content-Type", "application/json")
+	return resp
+}
+
+// listedIDs runs the filter and returns the ids left in the response.
+func listedIDs(t *testing.T, allowed []string, body string) []string {
+	t.Helper()
+	resp := jsonListingResponse(body)                            //nolint:bodyclose // in-memory body, replaced by the filter
+	require.NoError(t, modelDiscoveryFilter(allowed, nil)(resp)) //nolint:bodyclose // in-memory body, replaced by the filter
+
+	raw, err := io.ReadAll(resp.Body)
+	require.NoError(t, err)
+
+	var doc struct {
+		Data []struct {
+			ID string `json:"id"`
+		} `json:"data"`
+	}
+	require.NoError(t, json.Unmarshal(raw, &doc), "filtered body must stay valid JSON")
+
+	ids := make([]string, 0, len(doc.Data))
+	for _, entry := range doc.Data {
+		ids = append(ids, entry.ID)
+	}
+	return ids
+}
+
+// TestModelDiscoveryFilter_KeepsOnlyAuthorisedModels covers the picker a
+// developer sees: an unfiltered upstream list offers every model the shared
+// key can reach, and each one the policy excludes is a request the chain
+// denies a moment later.
+func TestModelDiscoveryFilter_KeepsOnlyAuthorisedModels(t *testing.T) {
+	ids := listedIDs(t, []string{"claude-sonnet-5", "claude-haiku-4-5"}, `{
+		"data": [
+			{"id": "claude-opus-5", "display_name": "Claude Opus 5"},
+			{"id": "claude-sonnet-5", "display_name": "Claude Sonnet 5"},
+			{"id": "claude-haiku-4-5"}
+		],
+		"has_more": false
+	}`)
+
+	assert.Equal(t, []string{"claude-sonnet-5", "claude-haiku-4-5"}, ids,
+		"only the models the route authorises may reach the picker")
+}
+
+// TestModelDiscoveryFilter_MatchesDatedAndPrefixedIDs pins the two id forms
+// a gateway returns for a model the operator registered plainly.
+func TestModelDiscoveryFilter_MatchesDatedAndPrefixedIDs(t *testing.T) {
+	ids := listedIDs(t, []string{"claude-sonnet-4-5", "anthropic.claude-opus-5"}, `{
+		"data": [
+			{"id": "claude-sonnet-4-5-20250929"},
+			{"id": "bedrock/anthropic.claude-opus-5"},
+			{"id": "gpt-4o"}
+		]
+	}`)
+
+	assert.Equal(t, []string{"claude-sonnet-4-5-20250929", "bedrock/anthropic.claude-opus-5"}, ids,
+		"a dated or provider-prefixed id must match its registered form")
+}
+
+// TestModelDiscoveryFilter_PreservesEnvelopeFields guards the rest of the
+// document: clients read paging fields alongside data.
+func TestModelDiscoveryFilter_PreservesEnvelopeFields(t *testing.T) {
+	resp := jsonListingResponse(`{"data":[{"id":"claude-sonnet-5"}],"has_more":true,"first_id":"x"}`) //nolint:bodyclose // in-memory body, replaced by the filter
+	require.NoError(t, modelDiscoveryFilter([]string{"claude-sonnet-5"}, nil)(resp))                  //nolint:bodyclose // in-memory body, replaced by the filter
+
+	raw, err := io.ReadAll(resp.Body)
+	require.NoError(t, err)
+
+	var doc map[string]any
+	require.NoError(t, json.Unmarshal(raw, &doc))
+	assert.Equal(t, true, doc["has_more"], "paging fields must survive the rewrite")
+	assert.Equal(t, "x", doc["first_id"])
+	assert.Equal(t, strconv.Itoa(len(raw)), resp.Header.Get("Content-Length"),
+		"Content-Length must match the rewritten body")
+}
+
+// TestModelDiscoveryFilter_PassesThroughUnfilterable covers the responses
+// the filter must not touch: a compressed body it cannot parse, a non-JSON
+// body, an error status, and a document with no data array.
+func TestModelDiscoveryFilter_PassesThroughUnfilterable(t *testing.T) {
+	cases := map[string]func() *http.Response{
+		"compressed": func() *http.Response {
+			resp := jsonListingResponse(`{"data":[{"id":"gpt-4o"}]}`)
+			resp.Header.Set("Content-Encoding", "gzip")
+			return resp
+		},
+		"not json": func() *http.Response {
+			resp := jsonListingResponse(`{"data":[{"id":"gpt-4o"}]}`)
+			resp.Header.Set("Content-Type", "text/html")
+			return resp
+		},
+		"error status": func() *http.Response {
+			resp := jsonListingResponse(`{"data":[{"id":"gpt-4o"}]}`)
+			resp.StatusCode = http.StatusInternalServerError
+			return resp
+		},
+		"no data array": func() *http.Response {
+			return jsonListingResponse(`{"object":"list"}`)
+		},
+	}
+
+	for name, build := range cases {
+		t.Run(name, func(t *testing.T) {
+			resp := build() //nolint:bodyclose // in-memory body, replaced by the filter
+			original, err := io.ReadAll(resp.Body)
+			require.NoError(t, err)
+			resp.Body = io.NopCloser(bytes.NewReader(original))
+
+			require.NoError(t, modelDiscoveryFilter([]string{"claude-sonnet-5"}, nil)(resp)) //nolint:bodyclose // in-memory body, replaced by the filter
+
+			got, err := io.ReadAll(resp.Body)
+			require.NoError(t, err)
+			assert.Equal(t, string(original), string(got), "an unfilterable response must reach the client unchanged")
+		})
+	}
+}
+
+// TestModelDiscoveryFilter_RunsNextHook pins that an existing
+// ModifyResponse hook still runs after filtering.
+func TestModelDiscoveryFilter_RunsNextHook(t *testing.T) {
+	called := false
+	next := func(*http.Response) error {
+		called = true
+		return nil
+	}
+
+	resp := jsonListingResponse(`{"data":[{"id":"claude-sonnet-5"}]}`)                //nolint:bodyclose // in-memory body, replaced by the filter
+	require.NoError(t, modelDiscoveryFilter([]string{"claude-sonnet-5"}, next)(resp)) //nolint:bodyclose // in-memory body, replaced by the filter
+	assert.True(t, called, "the chained hook must still run")
+}
+
+// TestModelDiscoveryFilter_KeepsSlashBearingIDs covers self-hosted backends
+// whose model ids carry a slash of their own. Treating the slash as a
+// gateway prefix and keeping only the tail dropped every such model from
+// the picker even though the policy named it exactly.
+func TestModelDiscoveryFilter_KeepsSlashBearingIDs(t *testing.T) {
+	ids := listedIDs(t, []string{"Qwen/Qwen2.5-0.5B-Instruct"}, `{
+		"object": "list",
+		"data": [
+			{"id": "Qwen/Qwen2.5-0.5B-Instruct"},
+			{"id": "Qwen/Qwen2.5-7B-Instruct"}
+		]
+	}`)
+
+	assert.Equal(t, []string{"Qwen/Qwen2.5-0.5B-Instruct"}, ids,
+		"a slash inside the model id is part of the id, not a provider prefix")
+}
+
+// TestModelDiscoveryFilter_ForwardsOversizedBodyIntact covers a listing past
+// the buffering cap. The filter reads one byte beyond the cap to detect the
+// size; forwarding only what it read would hand the client a body truncated
+// at exactly 1 MiB — valid-looking, short, and unparseable as JSON. The bytes
+// already read must be spliced back in front of the unread remainder so the
+// response reaches the client exactly as the upstream sent it.
+func TestModelDiscoveryFilter_ForwardsOversizedBodyIntact(t *testing.T) {
+	// A well-formed listing whose single entry pads the body past the cap.
+	padding := strings.Repeat("x", maxDiscoveryBodyBytes)
+	body := `{"object":"list","data":[{"id":"gpt-4o","note":"` + padding + `"}]}`
+	require.Greater(t, len(body), maxDiscoveryBodyBytes+1,
+		"the fixture must exceed the cap by more than the one-byte probe")
+
+	resp := jsonListingResponse(body)                        //nolint:bodyclose // in-memory body
+	require.NoError(t, modelDiscoveryFilter(nil, nil)(resp)) //nolint:bodyclose // in-memory body
+
+	got, err := io.ReadAll(resp.Body)
+	require.NoError(t, err)
+	assert.Equal(t, len(body), len(got),
+		"an oversized listing must reach the client whole, not truncated at the cap")
+	assert.Equal(t, body, string(got), "the forwarded bytes must be the upstream's own")
+
+	var doc map[string]json.RawMessage
+	assert.NoError(t, json.Unmarshal(got, &doc),
+		"the forwarded body must still parse as JSON")
+}
+
+// TestModelDiscoveryFilter_OversizedBodyKeepsUpstreamHeaders pins that the
+// oversized path leaves the response metadata alone. Rewriting Content-Length
+// to the truncated prefix is what made the corruption invisible to the client
+// until it tried to parse.
+func TestModelDiscoveryFilter_OversizedBodyKeepsUpstreamHeaders(t *testing.T) {
+	padding := strings.Repeat("x", maxDiscoveryBodyBytes)
+	body := `{"object":"list","data":[{"id":"gpt-4o","note":"` + padding + `"}]}`
+
+	resp := jsonListingResponse(body) //nolint:bodyclose // in-memory body
+	resp.Header.Set("Content-Length", strconv.Itoa(len(body)))
+	require.NoError(t, modelDiscoveryFilter(nil, nil)(resp)) //nolint:bodyclose // in-memory body
+
+	assert.Equal(t, int64(len(body)), resp.ContentLength,
+		"ContentLength must keep describing the body the client receives")
+	assert.Equal(t, strconv.Itoa(len(body)), resp.Header.Get("Content-Length"),
+		"the Content-Length header must not be rewritten to the truncated prefix")
+}

proxy/internal/proxy/reverseproxy.go

@@ -363,6 +363,9 @@ func (p *ReverseProxy) forwardUpstream(respWriter http.ResponseWriter, r *http.R
 	if result.rewriteRedirects {
 		rp.ModifyResponse = p.rewriteLocationFunc(effectiveURL, rewriteMatchedPath, r) //nolint:bodyclose
 	}
+	if upstreamRewrite != nil && len(upstreamRewrite.DiscoveryModels) > 0 {
+		rp.ModifyResponse = modelDiscoveryFilter(upstreamRewrite.DiscoveryModels, rp.ModifyResponse) //nolint:bodyclose // the hook replaces the body and closes the original
+	}
 	rp.ServeHTTP(respWriter, r.WithContext(ctx))
 }
 

shared/llm/model.go

@@ -46,6 +46,27 @@ func NormalizeBedrockModel(modelID string) string {
 	return bedrockVersionSuffix.ReplaceAllString(m, "")
 }
 
+// anthropicDatedModel matches a Claude model id carrying the trailing
+// "-YYYYMMDD" release-date suffix Anthropic appends to a pinned release,
+// capturing the id without it. The "claude" anchor is load-bearing: pricing
+// looks every model up through this helper regardless of surface, and an
+// operator may register a custom id with any shape at all, so an unanchored
+// "-\d{8}$" would let "internal-llm-20250101" silently inherit the rate
+// registered for "internal-llm". The anchor also covers the vendor-prefixed
+// forms ("anthropic.claude-...", "us.anthropic.claude-...").
+var anthropicDatedModel = regexp.MustCompile(`(?i)^(.*claude.*)-\d{8}$`)
+
+// NormalizeAnthropicModel strips the trailing release-date suffix from a
+// Claude model id, e.g. "claude-sonnet-4-5-20250929" -> "claude-sonnet-4-5",
+// so a dated id a client pins matches the undated one the operator
+// registered. Ids that are not Claude-family are returned untouched.
+// Callers try the verbatim id first and fall back to this, so two dated
+// releases of the same family stay distinct wherever both are registered
+// explicitly.
+func NormalizeAnthropicModel(modelID string) string {
+	return anthropicDatedModel.ReplaceAllString(modelID, "$1")
+}
+
 // NormalizeVertexModel strips the "@version" suffix from a Vertex AI model id
 // (e.g. "claude-sonnet-4-5@20250929" -> "claude-sonnet-4-5") so it matches
 // the catalog/pricing key. Vertex publisher models are priced under their

shared/llm/model_test.go

@@ -34,3 +34,29 @@ func TestNormalizeVertexModel(t *testing.T) {
 		require.Equal(t, want, NormalizeVertexModel(in), "normalize %q", in)
 	}
 }
+
+func TestNormalizeAnthropicModel(t *testing.T) {
+	cases := map[string]string{
+		"claude-sonnet-4-5-20250929":            "claude-sonnet-4-5",
+		"claude-3-5-haiku-20241022":             "claude-3-5-haiku",
+		"claude-sonnet-5":                       "claude-sonnet-5",
+		"claude-opus-4-8":                       "claude-opus-4-8",
+		"anthropic.claude-haiku-4-5":            "anthropic.claude-haiku-4-5",
+		"anthropic.claude-sonnet-4-5-20250929":  "anthropic.claude-sonnet-4-5",
+		"us.anthropic.claude-opus-4-8-20250101": "us.anthropic.claude-opus-4-8",
+		// Non-Claude ids must survive untouched even when they end in eight
+		// consecutive digits: an operator can register a custom model under
+		// any id, and pricing looks every one of them up through this helper.
+		"gpt-4o":                  "gpt-4o",
+		"gpt-4o-2024-08-06":       "gpt-4o-2024-08-06",
+		"gpt-4o-20240806":         "gpt-4o-20240806",
+		"internal-llm-20250101":   "internal-llm-20250101",
+		"deepseek-r1-20250120":    "deepseek-r1-20250120",
+		"Qwen/Qwen2.5-20250101":   "Qwen/Qwen2.5-20250101",
+		"gemini-2-5-pro-20250101": "gemini-2-5-pro-20250101",
+		"":                        "",
+	}
+	for in, want := range cases {
+		require.Equal(t, want, NormalizeAnthropicModel(in), "normalize %q", in)
+	}
+}