[proxy,management] Serve Bedrock model discovery from the control plane (#7250)

[proxy,management] Serve Bedrock model discovery from the control plane

A Bedrock provider could never answer a model-discovery request. The router
sent GET /inference-profiles to the record's upstream, which has to be
bedrock-runtime.<region> for InvokeModel to work, and that host does not
implement the operation. ListInferenceProfiles is a control-plane operation on
bedrock.<region>.amazonaws.com, and one provider record carries one upstream,
so the two hosts genuinely differ.

The route now carries a discovery host, taken from the catalog's declaration
with the region read back out of the configured upstream, and the listing — and
only the listing — goes there. Inference is untouched. A proxied or self-hosted
Bedrock endpoint gets no discovery host at all rather than a guessed one, since
inventing a host would send the operator's credential somewhere they never
configured.

Two things had to follow for the listing to be usable once it arrives. The
response filter only understood OpenAI's {"data":[{"id":…}]}, so a Bedrock
listing fell through it untouched, offering every profile in the account
whatever the policy said. And discoverableModels intersected by exact string,
so a record registering the raw profile id while a guardrail names the catalog
key intersected to nothing — bounding a working provider's listing down to
empty.

Normalisation is the third. The geography in front of a cross-region profile
was matched against a hardcoded list of four, so every profile issued under jp,
au, ca, sa or us-gov carried its prefix into the pricing key, matched no
catalog entry and metered at zero. It is now recognised by either the geography
or the vendor that follows it, so an id has to be new on both axes at once to
slip through — a live eu-central-1 listing returned "global.xai.grok-4.6" days
after the vendor list was first written.

management/internals/modules/agentnetwork/modeldiscovery/discovery.go

@@ -191,7 +191,7 @@ func (c *Client) discoveryURL(entry catalog.Provider, req Request) (string, erro
 			// A provider record carries no region field: the region lives
 			// inside the upstream host the operator already configured, so
 			// read it back out rather than asking them for it twice.
-			region = regionFromUpstream(entry, req.UpstreamURL)
+			region = RegionFromUpstream(entry, req.UpstreamURL)
 		}
 		if region == "" {
 			return "", fmt.Errorf("%w: %s discovery needs a region, and none could be read from the provider upstream",
@@ -207,13 +207,13 @@ func (c *Client) discoveryURL(entry catalog.Provider, req Request) (string, erro
 	return target.String(), nil
 }
 
-// regionFromUpstream recovers the region an operator embedded in the provider
+// RegionFromUpstream recovers the region an operator embedded in the provider
 // upstream, by matching it against the catalog's own host template. Bedrock's
 // template is "bedrock-runtime.<region>.amazonaws.com" and Vertex's is
 // "<region>-aiplatform.googleapis.com", so the region is whatever sits between
 // the fixed halves. Returns empty when the upstream does not match the
 // template, which is the case for a custom or proxied endpoint.
-func regionFromUpstream(entry catalog.Provider, upstreamURL string) string {
+func RegionFromUpstream(entry catalog.Provider, upstreamURL string) string {
 	prefix, suffix, found := strings.Cut(entry.DefaultHost, catalog.RegionPlaceholder)
 	if !found {
 		return ""

management/internals/modules/agentnetwork/modeldiscovery/discovery_test.go

@@ -490,7 +490,43 @@ func TestRegionFromUpstream(t *testing.T) {
 		{"bedrock regionless without scheme", bedrock, "bedrock-runtime.amazonaws.com", ""},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
-			assert.Equal(t, tc.want, regionFromUpstream(tc.entry, tc.upstream))
+			assert.Equal(t, tc.want, RegionFromUpstream(tc.entry, tc.upstream))
 		})
 	}
 }
+
+// bedrockGeoListing carries profiles from geographies the original prefix list
+// did not name. Every one reduces to a catalog key, so every one must arrive
+// priced — an unstripped geography is what made a real account's listing come
+// back almost entirely at zero.
+const bedrockGeoListing = `{"inferenceProfileSummaries":[
+  {"inferenceProfileId":"jp.anthropic.claude-sonnet-5-20260514-v1:0",
+   "inferenceProfileName":"JP Anthropic Claude Sonnet 5","status":"ACTIVE","type":"SYSTEM_DEFINED"},
+  {"inferenceProfileId":"au.anthropic.claude-haiku-4-5-20251001-v1:0",
+   "inferenceProfileName":"AU Anthropic Claude Haiku 4.5","status":"ACTIVE","type":"SYSTEM_DEFINED"},
+  {"inferenceProfileId":"us-gov.anthropic.claude-sonnet-5-20260514-v1:0",
+   "inferenceProfileName":"GovCloud Anthropic Claude Sonnet 5","status":"ACTIVE","type":"SYSTEM_DEFINED"}
+]}`
+
+func TestBedrockProfilesFromAnyGeographyArrivePriced(t *testing.T) {
+	cl, _ := newStubClient(http.StatusOK, bedrockGeoListing)
+
+	models, err := cl.Fetch(context.Background(), Request{
+		CatalogID:   "bedrock_api",
+		UpstreamURL: "https://bedrock-runtime.eu-central-1.amazonaws.com",
+		APIKey:      "aws-token",
+	})
+	require.NoError(t, err)
+	require.Len(t, models, 3)
+
+	for _, m := range models {
+		assert.True(t, m.PricingKnown, "%s must resolve to a catalog rate", m.ID)
+		assert.Greater(t, m.InputPer1k, 0.0, "input rate for %s", m.ID)
+		assert.Greater(t, m.OutputPer1k, 0.0, "output rate for %s", m.ID)
+		assert.Greater(t, m.CacheReadPer1k, 0.0, "cache-read rate for %s", m.ID)
+	}
+
+	// The wire id is preserved whatever the pricing key reduced to: it is the
+	// only form that works at invoke time.
+	assert.Equal(t, "jp.anthropic.claude-sonnet-5-20260514-v1:0", models[0].ID)
+}

management/internals/modules/agentnetwork/synthesizer.go

@@ -10,6 +10,7 @@ import (
 	"strings"
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/modeldiscovery"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
 	"github.com/netbirdio/netbird/management/internals/modules/reverseproxy/sessionkey"
@@ -380,6 +381,9 @@ type routerProviderRoute struct {
 	// proxy dials this provider's upstream. For self-hosted / internal gateways
 	// behind a private or self-signed certificate.
 	SkipTLSVerify bool `json:"skip_tls_verify,omitempty"`
+	// DiscoveryHost, when set, is the host serving this provider's model
+	// listing, for a vendor that does not serve it from the inference host.
+	DiscoveryHost string `json:"discovery_host,omitempty"`
 }
 
 // indexProviderGroups walks the enabled policies and returns, per
@@ -447,6 +451,9 @@ func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]
 		if err != nil {
 			return nil, fmt.Errorf("router config for provider %s: %w", p.ID, err)
 		}
+		// Lookup rather than assume: an unknown provider id yields the zero
+		// entry, which declares no discovery and so contributes nothing.
+		catalogEntry, _ := catalog.Lookup(p.ProviderID)
 		headerName, headerValue, gcpSAKeyB64, err := providerAuthHeader(p)
 		if err != nil {
 			return nil, err
@@ -466,6 +473,7 @@ func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]
 			Bedrock:                 catalog.IsBedrockPathStyle(p.ProviderID),
 			GCPServiceAccountKeyB64: gcpSAKeyB64,
 			SkipTLSVerify:           p.SkipTLSVerification,
+			DiscoveryHost:           discoveryHost(catalogEntry, p.UpstreamURL),
 		})
 	}
 	out, err := json.Marshal(cfg)
@@ -475,6 +483,33 @@ func buildRouterConfigJSON(providers []*types.Provider, groupIndex map[string][]
 	return out, nil
 }
 
+// discoveryHost returns the host serving this provider's model listing when it
+// differs from the inference host, and empty when the two are the same — which
+// is true of every vendor but Bedrock, whose ListInferenceProfiles is a control
+// plane operation on bedrock.<region> while InvokeModel must go to
+// bedrock-runtime.<region>. One provider record therefore needs two hosts.
+//
+// The catalog declares the listing host; the region is recovered from the
+// upstream the operator configured, since a provider record carries no region
+// field. An upstream matching no catalog template yields empty rather than a
+// guess: a proxied or self-hosted Bedrock endpoint may serve both from one
+// place, and inventing a host would send the credential somewhere the operator
+// never configured.
+func discoveryHost(entry catalog.Provider, upstreamURL string) string {
+	if entry.Discovery == nil || entry.Discovery.Host == "" {
+		return ""
+	}
+	host := entry.Discovery.Host
+	if !strings.Contains(host, catalog.RegionPlaceholder) {
+		return host
+	}
+	region := modeldiscovery.RegionFromUpstream(entry, upstreamURL)
+	if region == "" {
+		return ""
+	}
+	return strings.ReplaceAll(host, catalog.RegionPlaceholder, region)
+}
+
 // providerVendor returns the parser surface ("openai", "anthropic", …)
 // the provider speaks, sourced from its catalog entry's ParserID. The
 // router uses it to keep a request the parser tagged with a vendor on a

management/internals/modules/agentnetwork/synthesizer_pricing_test.go

@@ -103,3 +103,37 @@ func TestBuildCostMeterConfig_OrphanAndGatewayProviders(t *testing.T) {
 	assert.NotContains(t, cfg.Pricing.Providers, "prov-litellm", "empty-models gateway needs no per-record entry")
 	assert.NotEmpty(t, cfg.Pricing.Defaults["openai"], "defaults still ship so the gateway's catalog-model traffic is priced")
 }
+
+// TestBuildCostMeterConfig_BedrockGeographyOutsideTheOriginalFour is the
+// accounting half of the geography bug. The docs tell operators to register a
+// Bedrock id exactly as AWS issues it, region prefix included, and the cost
+// meter keys its table by the normalized form. While the geography was matched
+// against a list of four, a profile issued anywhere else kept its prefix,
+// missed the catalog entry it was meant to inherit from, and billed with a
+// zero entry underneath the operator's own rates — so every cache bucket
+// metered free and a model priced only by catalog defaults metered at nothing
+// at all.
+func TestBuildCostMeterConfig_BedrockGeographyOutsideTheOriginalFour(t *testing.T) {
+	for _, geo := range []string{"jp", "au", "ca", "sa", "us-gov"} {
+		t.Run(geo, func(t *testing.T) {
+			bedrock := &types.Provider{
+				ID:         "prov-bedrock",
+				ProviderID: "bedrock_api",
+				Enabled:    true,
+				Models: []types.ProviderModel{
+					{ID: geo + ".anthropic.claude-sonnet-5-20260514-v1:0", InputPer1k: 0.003, OutputPer1k: 0.015},
+				},
+			}
+			raw, err := buildCostMeterConfigJSON([]*types.Provider{bedrock}, map[string][]string{"prov-bedrock": {"grp"}})
+			require.NoError(t, err)
+			cfg := decodeCostMeterConfig(t, raw)
+
+			e, ok := cfg.Pricing.Providers["prov-bedrock"]["anthropic.claude-sonnet-5"]
+			require.True(t, ok, "a %s profile must key by the same normalized id the parser emits", geo)
+			assert.InDelta(t, 0.0003, e.CacheReadPer1k, 1e-9,
+				"cache read must be inherited from the bedrock default entry, not left at zero")
+			assert.InDelta(t, 0.00375, e.CacheCreationPer1k, 1e-9,
+				"cache creation must be inherited from the bedrock default entry, not left at zero")
+		})
+	}
+}

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -10,6 +10,7 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
 	"github.com/netbirdio/netbird/management/server/store"
@@ -1245,3 +1246,57 @@ func TestSynthesizeServices_EmptyAPIKey_FailsClosed(t *testing.T) {
 	require.Error(t, err, "synthesis must refuse a provider with no api key")
 	assert.Contains(t, err.Error(), "no api key", "error must surface the missing credential")
 }
+
+// TestDiscoveryHost pins which providers get a separate listing host. Getting
+// this wrong in either direction is costly: a missing host leaves Bedrock
+// discovery 404ing at AWS, and a host on the wrong provider would send that
+// provider's listing — and its credential — somewhere the operator never
+// configured.
+func TestDiscoveryHost(t *testing.T) {
+	entry := func(id string) catalog.Provider {
+		p, ok := catalog.Lookup(id)
+		require.True(t, ok, "catalog entry %s must exist", id)
+		return p
+	}
+
+	for _, tc := range []struct {
+		name     string
+		entry    catalog.Provider
+		upstream string
+		want     string
+	}{
+		{
+			// ListInferenceProfiles is a control-plane operation; the runtime
+			// host answers <UnknownOperationException/> for it.
+			name:  "bedrock splits the listing off the runtime host",
+			entry: entry("bedrock_api"), upstream: "https://bedrock-runtime.eu-central-1.amazonaws.com",
+			want: "bedrock.eu-central-1.amazonaws.com",
+		},
+		{
+			name:  "bedrock in another region",
+			entry: entry("bedrock_api"), upstream: "https://bedrock-runtime.us-west-2.amazonaws.com",
+			want: "bedrock.us-west-2.amazonaws.com",
+		},
+		{
+			// A proxied Bedrock endpoint may well serve both from one place,
+			// and there is no region to read back out of it.
+			name:  "proxied bedrock upstream yields no discovery host",
+			entry: entry("bedrock_api"), upstream: "https://bedrock.internal.example.com",
+			want: "",
+		},
+		{
+			name:  "openai serves its listing from the same host",
+			entry: entry("openai_api"), upstream: "https://api.openai.com",
+			want: "",
+		},
+		{
+			name:  "vertex serves its listing from the same host",
+			entry: entry("vertex_ai_api"), upstream: "https://us-east5-aiplatform.googleapis.com",
+			want: "",
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.want, discoveryHost(tc.entry, tc.upstream))
+		})
+	}
+}