[management, proxy] Management-owned LLM pricing: file-backed defaults + (#6965)

2026-09-02 13:01:29 +02:00 · 2026-07-31 20:52:56 +02:00
parent feecb993f4
commit 0780a806f2
47 changed files with 3196 additions and 1540 deletions
--- a/proxy/internal/middleware/builtin/builtin.go
+++ b/proxy/internal/middleware/builtin/builtin.go
@@ -36,15 +36,13 @@ var defaultRegistry = middleware.NewRegistry()

 // FactoryContext is the per-process bag that concrete factories may
 // consult during construction. It carries the proxy-lifetime context,
-// the data directory used for static config files (pricing tables,
-// allowlists), the OTel meter, and the proxy logger.
+// the OTel meter, and the proxy logger.
 //
 // Configure must be called once at boot before any chain build calls
 // Resolve. Calling it twice overwrites the prior value; tests may rely
 // on this to reset state.
 type FactoryContext struct {
 	Context    context.Context
-	DataDir    string
 	Meter      metric.Meter
 	Logger     *log.Logger
 	MgmtClient MgmtClient
@@ -58,12 +56,11 @@ var (
 // Configure stores the per-process FactoryContext. Concrete factories
 // reach for it via Context(). mgmt may be nil on tests / standalone
 // builds with no management server; consumers must guard.
-func Configure(ctx context.Context, dataDir string, meter metric.Meter, logger *log.Logger, mgmt MgmtClient) {
+func Configure(ctx context.Context, meter metric.Meter, logger *log.Logger, mgmt MgmtClient) {
 	ctxMu.Lock()
 	defer ctxMu.Unlock()
 	ctxStore = FactoryContext{
 		Context:    ctx,
-		DataDir:    dataDir,
 		Meter:      meter,
 		Logger:     logger,
 		MgmtClient: mgmt,
--- a/proxy/internal/middleware/builtin/cost_calculation_matrix_test.go
+++ b/proxy/internal/middleware/builtin/cost_calculation_matrix_test.go
@@ -12,6 +12,7 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"

+	mgmtpricing "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/cost_meter"
@@ -19,17 +20,20 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_response_parser"
 )

-// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the embedded default pricing
-// table and asserts exact USD amounts hardcoded from the vendors' published prices, including the cache split.
+// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the REAL default pricing
+// table management ships (mgmtpricing.DefaultTable, catalog-derived) and asserts exact USD amounts hardcoded from
+// the vendors' published prices, including the cache split. This is the cross-stack pricing contract test: the
+// management-side Entry JSON must decode into the proxy-side table and produce these exact costs.
 func TestCostCalculation_ProviderMatrix(t *testing.T) {
-	// Empty data dir → embedded defaults, like a proxy with no pricing override.
-	builtin.Configure(context.Background(), t.TempDir(), nil, nil, nil)
+	builtin.Configure(context.Background(), nil, nil, nil)

 	reqMW, err := llm_request_parser.Factory{}.New(nil)
 	require.NoError(t, err, "build llm_request_parser")
 	respMW, err := llm_response_parser.Factory{}.New(nil)
 	require.NoError(t, err, "build llm_response_parser")
-	costMW, err := cost_meter.Factory{}.New(nil)
+	costCfgJSON, err := json.Marshal(map[string]any{"pricing": map[string]any{"defaults": mgmtpricing.DefaultTable()}})
+	require.NoError(t, err, "marshal management default table into cost_meter config")
+	costMW, err := cost_meter.Factory{}.New(costCfgJSON)
 	require.NoError(t, err, "build cost_meter")
 	t.Cleanup(func() { _ = costMW.Close() })

--- a/proxy/internal/middleware/builtin/cost_meter/factory.go
+++ b/proxy/internal/middleware/builtin/cost_meter/factory.go
@@ -2,7 +2,6 @@ package cost_meter

 import (
 	"bytes"
-	"context"
 	"encoding/json"
 	"fmt"

@@ -11,16 +10,27 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 )

-// defaultPricingFilename is the basename probed inside the proxy data
-// directory when no override is configured.
-const defaultPricingFilename = "pricing.yaml"
-
-// Config is the on-wire configuration for the middleware.
+// Config is the on-wire configuration for the middleware, synthesized by
+// management (buildCostMeterConfigJSON). The proxy has no embedded price
+// list: this payload is the only pricing source, and updates arrive as
+// ordinary mapping pushes that rebuild the chain (and with it this
+// middleware instance) — no per-request fetches, no reload loops.
 type Config struct {
-	// PricingPath optionally overrides the basename of the pricing
-	// file probed inside the proxy data directory. When empty the
-	// loader falls back to "pricing.yaml".
-	PricingPath string `json:"pricing_path"`
+	Pricing *PricingConfig `json:"pricing"`
+}
+
+// PricingConfig carries the full pricing table:
+//   - Defaults: parser surface ("openai"/"anthropic"/"bedrock") ->
+//     normalized model id -> rates, matched against llm.provider +
+//     llm.model.
+//   - Providers: provider record id -> normalized model id -> rates,
+//     matched against the llm.resolved_provider_id metadata llm_router
+//     stamps. Entries arrive fully materialized (management folds default
+//     cache rates in at synth time), so lookup order is simply
+//     per-record first, defaults second.
+type PricingConfig struct {
+	Defaults  map[string]map[string]pricing.EntryJSON `json:"defaults"`
+	Providers map[string]map[string]pricing.EntryJSON `json:"providers"`
 }

 // Factory builds cost_meter instances from raw config bytes.
@@ -29,45 +39,45 @@ type Factory struct{}
 // ID returns the registry identifier.
 func (Factory) ID() string { return ID }

-// New constructs a middleware instance. Empty, null, and {} configs
-// are accepted; non-empty rawConfig that fails to unmarshal is
-// rejected so misconfigurations surface at chain build time. The
-// pricing loader is built once per instance and reused across
-// invocations.
+// New constructs a middleware instance. Empty, null, and {} configs are
+// accepted for backward compatibility with a management server that
+// predates config-delivered pricing — the instance then skips every cost
+// computation (unknown_model) and a warning is logged once at build time.
+// Non-empty rawConfig that fails to unmarshal, or a table carrying a
+// non-finite / negative rate, is rejected so misconfigurations surface at
+// chain build time.
 func (Factory) New(rawConfig []byte) (middleware.Middleware, error) {
 	cfg, err := decodeConfig(rawConfig)
 	if err != nil {
 		return nil, err
 	}

-	fctx := builtin.Context()
-	pricingPath := cfg.PricingPath
-	if pricingPath == "" {
-		pricingPath = defaultPricingFilename
+	if cfg.Pricing == nil {
+		if logger := builtin.Context().Logger; logger != nil {
+			logger.Warnf("cost_meter: no pricing table in middleware config; management predates config-delivered pricing — every request will record cost.skipped=unknown_model ($0)")
+		}
+		return newMiddleware(mustEmptyTable(), nil), nil
 	}

-	loader, err := pricing.NewLoader(fctx.DataDir, pricingPath, ID, nil)
+	defaults, err := pricing.NewTable(cfg.Pricing.Defaults)
 	if err != nil {
-		return nil, fmt.Errorf("init pricing loader: %w", err)
+		return nil, fmt.Errorf("cost_meter pricing defaults: %w", err)
 	}
-
-	cancel := startReloader(fctx.Context, loader)
-
-	return newMiddleware(loader, cancel), nil
+	perRecord, err := pricing.NewEntries(cfg.Pricing.Providers)
+	if err != nil {
+		return nil, fmt.Errorf("cost_meter per-provider pricing: %w", err)
+	}
+	return newMiddleware(defaults, perRecord), nil
 }

-// startReloader binds the loader's mtime-poll goroutine to a context
-// derived from the proxy-lifetime context and returns its cancel func so
-// the owning middleware can stop the goroutine on teardown. Returns nil
-// when there's nothing to watch (nil context or defaults-only loader), in
-// which case the middleware's Close is a no-op.
-func startReloader(ctx context.Context, loader *pricing.Loader) context.CancelFunc {
-	if ctx == nil || !loader.WatchesFile() {
-		return nil
+// mustEmptyTable returns a valid empty table. NewTable on a nil map cannot
+// fail; the panic guard documents that invariant.
+func mustEmptyTable() *pricing.Table {
+	t, err := pricing.NewTable(nil)
+	if err != nil {
+		panic(fmt.Sprintf("cost_meter: empty pricing table must build: %v", err))
 	}
-	cctx, cancel := context.WithCancel(ctx)
-	go loader.Reload(cctx)
-	return cancel
+	return t
 }

 // decodeConfig accepts empty, null, and {} configs, returning a
--- a/proxy/internal/middleware/builtin/cost_meter/middleware.go
+++ b/proxy/internal/middleware/builtin/cost_meter/middleware.go
@@ -1,7 +1,9 @@
 // Package cost_meter implements the SlotOnResponse middleware that
 // converts token-usage metadata emitted by llm_response_parser into a
-// per-request USD cost estimate. The middleware uses the shared pricing
-// loader so operator pricing overrides apply to the chain.
+// per-request USD cost estimate. Pricing arrives from management inside
+// the middleware config: a per-provider-record table (the operator's
+// stored prices, matched via llm.resolved_provider_id) consulted first,
+// then the surface-keyed defaults table.
 package cost_meter

 import (
@@ -17,7 +19,9 @@ import (
 const ID = "cost_meter"

 // Version is the implementation version emitted via the spec merge.
-const Version = "1.0.0"
+// 1.1.0: pricing is config-delivered (defaults + per-provider-record
+// entries) instead of proxy-embedded.
+const Version = "1.1.0"

 // Skip reasons emitted under KeyCostSkipped. The set is closed; the
 // dashboard surfaces these verbatim.
@@ -42,19 +46,21 @@ var metadataKeys = []string{
 }

 // Middleware computes a per-response cost estimate from the token
-// counts emitted upstream by llm_response_parser.
+// counts emitted upstream by llm_response_parser. Both tables are
+// immutable — a pricing change arrives as a mapping push that rebuilds
+// the chain with a fresh instance.
 type Middleware struct {
-	loader *pricing.Loader
-	// cancel stops this instance's pricing-reload goroutine. Non-nil only
-	// when the loader watches an override file; Close calls it so a chain
-	// rebuild doesn't leak a poll goroutine per retired instance.
-	cancel context.CancelFunc
+	// defaults is the surface-keyed table (llm.provider x llm.model).
+	defaults *pricing.Table
+	// perRecord is keyed by provider record id (llm.resolved_provider_id)
+	// then normalized model id; entries arrive fully materialized from
+	// management. Consulted before defaults. May be nil.
+	perRecord map[string]map[string]pricing.Entry
 }

-// newMiddleware constructs a Middleware bound to the given pricing loader.
-// cancel may be nil (defaults-only loader with no reloader to stop).
-func newMiddleware(loader *pricing.Loader, cancel context.CancelFunc) *Middleware {
-	return &Middleware{loader: loader, cancel: cancel}
+// newMiddleware constructs a Middleware over the given pricing tables.
+func newMiddleware(defaults *pricing.Table, perRecord map[string]map[string]pricing.Entry) *Middleware {
+	return &Middleware{defaults: defaults, perRecord: perRecord}
 }

 // ID returns the registry identifier.
@@ -79,16 +85,9 @@ func (m *Middleware) MetadataKeys() []string {
 // response.
 func (m *Middleware) MutationsSupported() bool { return false }

-// Close stops this instance's pricing-reload goroutine, if any. Called by
-// the chain when a rebuild retires the instance, so the mtime-poll loop
-// doesn't outlive the chain it belonged to. Safe to call on a nil receiver
-// and on an instance with no reloader.
-func (m *Middleware) Close() error {
-	if m != nil && m.cancel != nil {
-		m.cancel()
-	}
-	return nil
-}
+// Close releases resources owned by the middleware. Stateless — the
+// pricing tables are plain maps owned by this instance.
+func (m *Middleware) Close() error { return nil }

 // Invoke reads provider, model, and token metadata, looks up pricing,
 // and emits either KeyCostUSDTotal or KeyCostSkipped. The decision is
@@ -144,8 +143,7 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 		return out, nil
 	}

-	table := m.loader.Get()
-	costs, ok := table.Costs(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+	costs, ok := m.lookupCosts(in.Metadata, provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
 	if !ok {
 		out.Metadata = skip(skipUnknownModel)
 		return out, nil
@@ -164,6 +162,26 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 	return out, nil
 }

+// lookupCosts resolves the price for this request and computes the cost
+// split. Resolution order:
+//
+//  1. Per-provider-record entry: the operator's stored price for the
+//     provider route that served the request, keyed by the
+//     llm.resolved_provider_id metadata llm_router stamped on the allow
+//     path. Absent metadata (e.g. no router in the chain) skips this tier.
+//  2. Surface defaults: the catalog-derived table keyed by llm.provider.
+//
+// The surface always selects the cache formula — a per-record entry for an
+// 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 {
+			return pricing.EntryCosts(entry, surface, inTokens, outTokens, cachedTokens, cacheCreationTokens), true
+		}
+	}
+	return m.defaults.Costs(surface, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+}
+
 // usd renders a cost as the fixed-precision string every cost.usd_* key
 // carries, so the per-bucket values and the aggregates round identically.
 //
--- a/proxy/internal/middleware/builtin/cost_meter/middleware_test.go
+++ b/proxy/internal/middleware/builtin/cost_meter/middleware_test.go
@@ -3,39 +3,50 @@ package cost_meter
 import (
 	"context"
 	"encoding/json"
-	"os"
-	"path/filepath"
 	"testing"

 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"

+	"github.com/netbirdio/netbird/proxy/internal/llm/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
-	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 )

-const fixturePricing = `openai:
-  gpt-4o:
-    input_per_1k: 0.0025
-    output_per_1k: 0.01
-  gpt-4o-mini:
-    input_per_1k: 0.00015
-    output_per_1k: 0.0006
-anthropic:
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-`
-
-// configureBuiltin points the package-level FactoryContext at a tmp
-// directory containing the test pricing fixture. Returns the path so
-// callers can override files later if needed.
-func configureBuiltin(t *testing.T) string {
+// fixtureConfig mirrors what management's buildCostMeterConfigJSON ships:
+// a surface-keyed defaults table. Rates match the retired YAML fixture so
+// every cost assertion below is byte-identical to the pre-feature values.
+func fixtureConfig(t *testing.T) []byte {
 	t.Helper()
-	dir := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(dir, "pricing.yaml"), []byte(fixturePricing), 0o600), "write pricing fixture")
-	builtin.Configure(context.Background(), dir, nil, nil, nil)
-	return dir
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {
+				"gpt-4o":      {InputPer1K: 0.0025, OutputPer1K: 0.01},
+				"gpt-4o-mini": {InputPer1K: 0.00015, OutputPer1K: 0.0006},
+			},
+			"anthropic": {
+				"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015},
+			},
+		},
+	}})
+	require.NoError(t, err)
+	return raw
+}
+
+// fixtureConfigWithCache adds the cache-rate fields.
+func fixtureConfigWithCache(t *testing.T) []byte {
+	t.Helper()
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {
+				"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01, CachedInputPer1K: 0.00125},
+			},
+			"anthropic": {
+				"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015, CacheReadPer1K: 0.0003, CacheCreationPer1K: 0.00375},
+			},
+		},
+	}})
+	require.NoError(t, err)
+	return raw
 }

 func metaValue(t *testing.T, kvs []middleware.KV, key string) (string, bool) {
@@ -56,8 +67,7 @@ func buildMiddleware(t *testing.T, raw []byte) middleware.Middleware {
 }

 func TestMiddleware_StaticSurface(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	assert.Equal(t, ID, mw.ID(), "ID must match the registered constant")
 	assert.Equal(t, Version, mw.Version(), "Version must match the constant")
@@ -79,8 +89,10 @@ func TestMiddleware_StaticSurface(t *testing.T) {
 	assert.Equal(t, expected, keys, "metadata key allowlist must match the spec")
 }

+// TestFactory_AcceptsEmptyAndJSONConfig: empty/null/{} configs are what an
+// old management (pre config-delivered pricing) sends — they must build a
+// working (all-skip) instance, never fail the chain.
 func TestFactory_AcceptsEmptyAndJSONConfig(t *testing.T) {
-	configureBuiltin(t)
 	cases := [][]byte{nil, {}, []byte("null"), []byte("{}"), []byte("   ")}
 	for _, raw := range cases {
 		mw, err := Factory{}.New(raw)
@@ -90,15 +102,57 @@ func TestFactory_AcceptsEmptyAndJSONConfig(t *testing.T) {
 }

 func TestFactory_RejectsMalformedConfig(t *testing.T) {
-	configureBuiltin(t)
 	mw, err := Factory{}.New([]byte("{not json"))
 	require.Error(t, err, "malformed config must surface at construction")
 	assert.Nil(t, mw, "no instance is returned on error")
 }

-func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+// TestFactory_RejectsInvalidRates: a non-finite or negative rate anywhere
+// in the table fails the chain build (defense-in-depth behind management's
+// API validation) rather than silently mispricing.
+func TestFactory_RejectsInvalidRates(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {"gpt-4o": {InputPer1K: -0.0025, OutputPer1K: 0.01}},
+		},
+	}})
+	require.NoError(t, err)
+	mw, err := Factory{}.New(raw)
+	require.Error(t, err, "negative rate must fail the build")
+	assert.Nil(t, mw)
+
+	raw, err = json.Marshal(Config{Pricing: &PricingConfig{
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-1": {"m": {InputPer1K: 0.01, OutputPer1K: 0.01, CacheReadPer1K: -1}},
+		},
+	}})
+	require.NoError(t, err)
+	_, err = Factory{}.New(raw)
+	require.Error(t, err, "per-record tables validate too")
+}
+
+// TestFactory_NilPricingSkipsEverything is the version-skew contract: a
+// new proxy under an old management ({} config) must build, allow, and
+// skip with unknown_model — degraded but never broken.
+func TestFactory_NilPricingSkipsEverything(t *testing.T) {
+	mw := buildMiddleware(t, []byte("{}"))
+	out, err := mw.Invoke(context.Background(), &middleware.Input{
+		Metadata: []middleware.KV{
+			{Key: middleware.KeyLLMProvider, Value: "openai"},
+			{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+			{Key: middleware.KeyLLMInputTokens, Value: "1000"},
+			{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
+		},
+	})
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "cost_meter always allows")
+	value, ok := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
+	require.True(t, ok, "no pricing table means every request skips")
+	assert.Equal(t, skipUnknownModel, value)
+}
+
+func TestFactory_ConfigDefaultsPriceRequests(t *testing.T) {
+	mw := buildMiddleware(t, fixtureConfig(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -116,34 +170,78 @@ func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {
 	assert.Equal(t, "0.000750000", value, "0.00015 + 0.0006 per 1k tokens, 9-decimal format")
 }

-func TestFactory_PricingPathOverride(t *testing.T) {
-	dir := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(dir, "custom.yaml"), []byte(fixturePricing), 0o600), "write custom pricing")
-	builtin.Configure(context.Background(), dir, nil, nil, nil)
-
-	raw, err := json.Marshal(Config{PricingPath: "custom.yaml"})
+// TestInvoke_PerRecordEntryBeatsDefaults: when llm_router resolved a
+// provider record whose operator pinned a price for the model, that price
+// wins over the surface default.
+func TestInvoke_PerRecordEntryBeatsDefaults(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
+		},
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-azure": {"gpt-4o": {InputPer1K: 0.005, OutputPer1K: 0.02}},
+		},
+	}})
 	require.NoError(t, err)
-
 	mw := buildMiddleware(t, raw)
+
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMProvider, Value: "openai"},
 			{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-			{Key: middleware.KeyLLMInputTokens, Value: "2000"},
+			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-azure"},
+			{Key: middleware.KeyLLMInputTokens, Value: "1000"},
 			{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
 		},
 	})
 	require.NoError(t, err)
-
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok, "cost.usd_total must be emitted with custom pricing path")
-	assert.Equal(t, "0.015000000", value, "2*0.0025 + 1*0.01 = 0.015 with 9-decimal format")
+	require.True(t, ok)
+	assert.Equal(t, "0.025000000", value, "operator's per-record price (0.005+0.02) wins over the default (0.0025+0.01)")
 }

-func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+// TestInvoke_PerRecordMissFallsBackToDefaults: a resolved record with no
+// entry for this model (or no entries at all) falls through to the
+// surface defaults — gateway providers rely on exactly this.
+func TestInvoke_PerRecordMissFallsBackToDefaults(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Defaults: map[string]map[string]pricing.EntryJSON{
+			"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
+		},
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-1": {"some-other-model": {InputPer1K: 1, OutputPer1K: 1}},
+		},
+	}})
+	require.NoError(t, err)
+	mw := buildMiddleware(t, raw)

+	for name, recordID := range map[string]string{
+		"record with other models": "prov-1",
+		"record with no entries":   "prov-gateway",
+	} {
+		t.Run(name, func(t *testing.T) {
+			out, err := mw.Invoke(context.Background(), &middleware.Input{
+				Metadata: []middleware.KV{
+					{Key: middleware.KeyLLMProvider, Value: "openai"},
+					{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+					{Key: middleware.KeyLLMResolvedProviderID, Value: recordID},
+					{Key: middleware.KeyLLMInputTokens, Value: "1000"},
+					{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
+				},
+			})
+			require.NoError(t, err)
+			value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
+			require.True(t, ok, "per-record miss must fall back to the surface default, not skip")
+			assert.Equal(t, "0.012500000", value, "default rates apply")
+		})
+	}
+}
+
+// TestInvoke_NoResolvedProviderIDUsesDefaults: metadata without a
+// resolved provider id (router denied, or a chain without llm_router)
+// prices from the defaults table directly.
+func TestInvoke_NoResolvedProviderIDUsesDefaults(t *testing.T) {
+	mw := buildMiddleware(t, fixtureConfig(t))
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMProvider, Value: "anthropic"},
@@ -153,17 +251,15 @@ func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
 		},
 	})
 	require.NoError(t, err)
-
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok, "cost.usd_total must be emitted")
+	require.True(t, ok)
 	assert.Equal(t, "0.018000000", value, "0.003 + 0.015 = 0.018 with 9-decimal format")
 	_, skipped := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
 	assert.False(t, skipped, "cost.skipped must not be set when cost is computed")
 }

 func TestInvoke_MissingProvider(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -179,8 +275,7 @@ func TestInvoke_MissingProvider(t *testing.T) {
 }

 func TestInvoke_MissingModel(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -196,8 +291,7 @@ func TestInvoke_MissingModel(t *testing.T) {
 }

 func TestInvoke_MissingTokens(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	cases := []struct {
 		name string
@@ -240,8 +334,7 @@ func TestInvoke_MissingTokens(t *testing.T) {
 }

 func TestInvoke_UnparseableTokens(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	cases := []struct {
 		name string
@@ -271,8 +364,7 @@ func TestInvoke_UnparseableTokens(t *testing.T) {
 }

 func TestInvoke_ZeroTokens(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -291,8 +383,7 @@ func TestInvoke_ZeroTokens(t *testing.T) {
 }

 func TestInvoke_UnknownModel(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -309,8 +400,7 @@ func TestInvoke_UnknownModel(t *testing.T) {
 }

 func TestInvoke_NilInput(t *testing.T) {
-	configureBuiltin(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfig(t))

 	out, err := mw.Invoke(context.Background(), nil)
 	require.NoError(t, err)
@@ -319,36 +409,12 @@ func TestInvoke_NilInput(t *testing.T) {
 	assert.Empty(t, out.Metadata, "no metadata must be emitted on nil input")
 }

-const fixturePricingWithCache = `openai:
-  gpt-4o:
-    input_per_1k: 0.0025
-    output_per_1k: 0.01
-    cached_input_per_1k: 0.00125
-anthropic:
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-`
-
-// configureBuiltinWithCacheRates points the package-level
-// FactoryContext at a tmp directory containing pricing entries that
-// include the cache rate fields.
-func configureBuiltinWithCacheRates(t *testing.T) {
-	t.Helper()
-	dir := t.TempDir()
-	require.NoError(t, os.WriteFile(filepath.Join(dir, "pricing.yaml"), []byte(fixturePricingWithCache), 0o600), "write cache-aware pricing fixture")
-	builtin.Configure(context.Background(), dir, nil, nil, nil)
-}
-
 // TestInvoke_OpenAICachedSubsetDiscount proves the OpenAI shape end
 // to end through the middleware: cached_input_tokens is treated as a
 // SUBSET of input_tokens and discounted at the configured rate, not
 // added on top.
 func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -390,8 +456,7 @@ func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
 // shape: cache_read and cache_creation are additive to input_tokens
 // and each carries its own rate.
 func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -429,8 +494,37 @@ func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
 		"output bucket bills 200 tokens at 0.015/1k")
 }

+// TestInvoke_PerRecordEntryUsesSurfaceFormula: a per-record entry for an
+// anthropic-surface request must bill its cache buckets additively — the
+// formula follows llm.provider, not which table the entry came from.
+func TestInvoke_PerRecordEntryUsesSurfaceFormula(t *testing.T) {
+	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
+		Providers: map[string]map[string]pricing.EntryJSON{
+			"prov-ant": {"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015, CacheReadPer1K: 0.0003, CacheCreationPer1K: 0.00375}},
+		},
+	}})
+	require.NoError(t, err)
+	mw := buildMiddleware(t, raw)
+
+	out, err := mw.Invoke(context.Background(), &middleware.Input{
+		Metadata: []middleware.KV{
+			{Key: middleware.KeyLLMProvider, Value: "anthropic"},
+			{Key: middleware.KeyLLMModel, Value: "claude-sonnet-4-5"},
+			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-ant"},
+			{Key: middleware.KeyLLMInputTokens, Value: "256"},
+			{Key: middleware.KeyLLMOutputTokens, Value: "200"},
+			{Key: middleware.KeyLLMCachedInputTokens, Value: "768"},
+			{Key: middleware.KeyLLMCacheCreationTokens, Value: "512"},
+		},
+	})
+	require.NoError(t, err)
+	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
+	require.True(t, ok)
+	assert.Equal(t, "0.005918400", value, "identical math to the defaults-table entry with the same rates")
+}
+
 // assertBucket asserts one per-bucket cost key carries the expected
-// 6-decimal value.
+// 9-decimal value.
 func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
 	t.Helper()
 	got, ok := metaValue(t, md, key)
@@ -439,13 +533,10 @@ func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
 }

 // TestInvoke_CachedTokensAbsentFallsBackToBaseFormula covers the
-// "operator hasn't opted in" path: with no cached metadata keys
-// emitted, the meter must produce exactly the same cost as before
-// the feature landed. Critical so operators with the new binary but
-// no YAML changes see no behavioural drift on OpenAI requests.
+// no-cache-metadata path: with no cached keys emitted, the meter must
+// produce exactly the input+output cost.
 func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -460,7 +551,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 1000 input * 0.0025 + 500 output * 0.01 = 0.0025 + 0.005 = 0.0075
-	assert.Equal(t, "0.007500000", value, "no cached metadata = same cost as before the feature landed")
+	assert.Equal(t, "0.007500000", value, "no cached metadata = plain input+output cost")
 }

 // TestInvoke_UnparseableCachedTokensSkippedSilently proves the
@@ -469,8 +560,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 // regular formula. Cache buckets are a refinement, never a reason to
 // abort cost computation.
 func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
-	configureBuiltinWithCacheRates(t)
-	mw := buildMiddleware(t, nil)
+	mw := buildMiddleware(t, fixtureConfigWithCache(t))

 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -487,22 +577,10 @@ func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
 	assert.Equal(t, "0.007500000", value, "same as the no-cached-metadata path")
 }

-// TestMiddleware_CloseCancelsReloader proves Close stops the per-instance
-// pricing-reload goroutine: a chain rebuild retires the old instance and
-// calls Close, which must invoke the cancel func startReloader handed it so
-// the mtime-poll loop doesn't outlive the chain.
-func TestMiddleware_CloseCancelsReloader(t *testing.T) {
-	ctx, cancel := context.WithCancel(context.Background())
-	m := newMiddleware(nil, cancel)
-
-	require.NoError(t, m.Close(), "Close must not error")
-	require.Error(t, ctx.Err(), "Close must cancel the reloader context so the poll goroutine exits")
-}
-
-// TestMiddleware_CloseNilSafe confirms Close is a no-op (no panic) for an
-// instance with no reloader and for a nil receiver.
+// TestMiddleware_CloseNilSafe confirms Close is a no-op (no panic) even
+// for a nil receiver.
 func TestMiddleware_CloseNilSafe(t *testing.T) {
-	require.NoError(t, newMiddleware(nil, nil).Close(), "no-reloader Close must be a no-op")
+	require.NoError(t, newMiddleware(nil, nil).Close(), "Close must be a no-op")
 	var m *Middleware
 	require.NoError(t, m.Close(), "nil-receiver Close must be safe")
 }
--- a/proxy/internal/middleware/builtin/llm_request_parser/bedrock_test.go
+++ b/proxy/internal/middleware/builtin/llm_request_parser/bedrock_test.go
@@ -6,23 +6,6 @@ import (
 	"github.com/stretchr/testify/require"
 )

-func TestNormalizeBedrockModel(t *testing.T) {
-	cases := map[string]string{
-		"eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
-		"us.anthropic.claude-opus-4-8-20250101-v1:0":   "anthropic.claude-opus-4-8",
-		"apac.anthropic.claude-haiku-4-5-v1:0":         "anthropic.claude-haiku-4-5",
-		"anthropic.claude-sonnet-4-5-20250929-v1:0":    "anthropic.claude-sonnet-4-5",
-		"meta.llama3-3-70b-instruct-v1:0":              "meta.llama3-3-70b-instruct",
-		"amazon.nova-pro-v1:0":                         "amazon.nova-pro",
-		"amazon.nova-2-lite-v1:0":                      "amazon.nova-2-lite",
-		// Inference-profile ARN — model id lives in the last path segment.
-		"arn:aws:bedrock:eu-central-1:123456789012:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
-	}
-	for in, want := range cases {
-		require.Equal(t, want, normalizeBedrockModel(in), "normalize %q", in)
-	}
-}
-
 func TestParseBedrockPath(t *testing.T) {
 	tests := []struct {
 		path   string
--- a/proxy/internal/middleware/builtin/llm_request_parser/middleware.go
+++ b/proxy/internal/middleware/builtin/llm_request_parser/middleware.go
@@ -8,7 +8,6 @@ package llm_request_parser
 import (
 	"context"
 	"net/url"
-	"regexp"
 	"strconv"
 	"strings"
 	"unicode/utf8"
@@ -253,9 +252,7 @@ func parseVertexPath(reqPath string) (vertexRequest, bool) {
 	if c := strings.LastIndex(rest, ":"); c >= 0 {
 		model, action = rest[:c], rest[c+1:]
 	}
-	if at := strings.Index(model, "@"); at >= 0 {
-		model = model[:at]
-	}
+	model = llm.NormalizeVertexModel(model)
 	if model == "" {
 		return vertexRequest{}, false
 	}
@@ -343,14 +340,6 @@ func trimBedrockNamespace(reqPath string) string {
 	return reqPath
 }

-// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
-// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
-var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
-
-// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
-// version/throughput suffix of a Bedrock model id.
-var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
-
 // parseBedrockPath extracts the model and streaming/converse flags from an AWS
 // Bedrock runtime model endpoint:
 //
@@ -375,7 +364,7 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 	if decoded, err := url.PathUnescape(rawModel); err == nil {
 		rawModel = decoded
 	}
-	model := normalizeBedrockModel(rawModel)
+	model := llm.NormalizeBedrockModel(rawModel)
 	if model == "" {
 		return bedrockRequest{}, false
 	}
@@ -389,30 +378,6 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 	}
 }

-// normalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
-// prefix, and the version/throughput suffix from a Bedrock model id so it
-// matches the catalog/pricing key, e.g.
-// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
-// and "arn:aws:bedrock:eu-central-1:123:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0"
-// -> "anthropic.claude-sonnet-4-5".
-func normalizeBedrockModel(modelID string) string {
-	m := modelID
-	// A full ARN (inference-profile / provisioned-throughput / foundation-model)
-	// carries the model id in its last path segment.
-	if strings.HasPrefix(m, "arn:") {
-		if i := strings.LastIndex(m, "/"); i >= 0 {
-			m = m[i+1:]
-		}
-	}
-	for _, p := range bedrockRegionPrefixes {
-		if strings.HasPrefix(m, p) {
-			m = m[len(p):]
-			break
-		}
-	}
-	return bedrockVersionSuffix.ReplaceAllString(m, "")
-}
-
 // invokeBedrock emits the model/provider/session/prompt for an AWS Bedrock
 // request. Bedrock is metered under the dedicated "bedrock" parser, which reads
 // both the InvokeModel and Converse response shapes.