fix(usage): bill Codex Fast from requested tier

apps/manager-server/internal/repository/usageevent/repository.go

@@ -107,9 +107,12 @@ func (r *repository) InsertBatch(ctx context.Context, events []model.UsageEvent)
 		if event.RequestServiceTier == "" {
 			event.RequestServiceTier = event.ServiceTier
 		}
-		if event.ResponseServiceTier != "" {
-			event.ServiceTier = event.ResponseServiceTier
-		}
+		event.ServiceTier = usage.EffectiveServiceTier(usage.CacheInputContext{
+			ExecutorType:     event.ExecutorType,
+			Provider:         event.Provider,
+			ProviderSnapshot: event.AuthProviderSnapshot,
+			AuthType:         event.AuthType,
+		}, event.RequestServiceTier, event.ServiceTier, event.ResponseServiceTier)
 		failed := 0
 		if event.Failed {
 			failed = 1

apps/manager-server/internal/repository/usageevent/stream_test.go

@@ -66,6 +66,44 @@ func TestWriteCompatibleUsageMatchesBuildPayload(t *testing.T) {
 	}
 }
 
+func TestInsertBatchSelectsServiceTierByProviderSemantics(t *testing.T) {
+	db, err := sqliterepo.Open(filepath.Join(t.TempDir(), "usage.sqlite"))
+	if err != nil {
+		t.Fatalf("open database: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+	repo := New(db)
+
+	codex := streamTestEvent("codex-tier", 100, "POST /v1/responses", "gpt-5.4")
+	codex.ExecutorType = "codex"
+	codex.RequestServiceTier = "priority"
+	codex.ResponseServiceTier = "default"
+	codex.ServiceTier = "priority"
+	nonCodex := streamTestEvent("openai-tier", 200, "POST /v1/responses", "gpt-5.4")
+	nonCodex.Provider = "openai-compatible"
+	nonCodex.RequestServiceTier = "priority"
+	nonCodex.ResponseServiceTier = "default"
+	nonCodex.ServiceTier = "priority"
+
+	if _, err := repo.InsertBatch(context.Background(), []usage.Event{codex, nonCodex}); err != nil {
+		t.Fatalf("insert events: %v", err)
+	}
+	recent, err := repo.ListRecent(context.Background(), 2)
+	if err != nil {
+		t.Fatalf("list recent: %v", err)
+	}
+	byHash := make(map[string]usage.Event, len(recent))
+	for _, event := range recent {
+		byHash[event.EventHash] = event
+	}
+	if event := byHash["codex-tier"]; event.ServiceTier != "priority" || event.RequestServiceTier != "priority" || event.ResponseServiceTier != "default" {
+		t.Fatalf("codex tiers = %q/%q/%q", event.ServiceTier, event.RequestServiceTier, event.ResponseServiceTier)
+	}
+	if event := byHash["openai-tier"]; event.ServiceTier != "default" || event.RequestServiceTier != "priority" || event.ResponseServiceTier != "default" {
+		t.Fatalf("non-Codex tiers = %q/%q/%q", event.ServiceTier, event.RequestServiceTier, event.ResponseServiceTier)
+	}
+}
+
 func TestWriteExportJSONLUsesRecentLimitAndAscendingKeysetOrder(t *testing.T) {
 	db, err := sqliterepo.Open(filepath.Join(t.TempDir(), "usage.sqlite"))
 	if err != nil {

apps/manager-server/internal/usage/event.go

@@ -173,11 +173,40 @@ type CacheInputContext struct {
 	ExecutorType     string
 	Provider         string
 	ProviderSnapshot string
+	AuthType         string
 	ResolvedModel    string
 	RequestedModel   string
 	DisplayModel     string
 }
 
+// EffectiveServiceTier returns the tier used for billing and aggregation.
+// Codex reports default/auto even when Fast Mode was requested, so Codex uses
+// the request tier. Other providers retain response-tier precedence.
+func EffectiveServiceTier(context CacheInputContext, requestTier, legacyTier, responseTier string) string {
+	identity := strings.ToLower(strings.Join([]string{
+		context.ExecutorType,
+		context.Provider,
+		context.ProviderSnapshot,
+		context.AuthType,
+	}, " "))
+	if strings.Contains(identity, "codex") {
+		if requestTier != "" {
+			return requestTier
+		}
+		if legacyTier != "" {
+			return legacyTier
+		}
+		return responseTier
+	}
+	if responseTier != "" {
+		return responseTier
+	}
+	if legacyTier != "" {
+		return legacyTier
+	}
+	return requestTier
+}
+
 type RawCacheAccountingHints struct {
 	ExplicitMode     string
 	ExplicitTotal    int64
@@ -501,22 +530,22 @@ func NormalizeRaw(raw []byte) (Event, error) {
 	}
 	provider := readString(record, "provider", "type", "auth_type", "authType")
 	executorType := readString(record, "executor_type", "executorType")
+	authType := readString(record, "auth_type", "authType")
 	requestServiceTier := readString(record, "request_service_tier", "requestServiceTier", "service_tier", "serviceTier")
 	responseServiceTier := readString(record, "response_service_tier", "responseServiceTier")
-	serviceTier := responseServiceTier
-	if serviceTier == "" {
-		serviceTier = requestServiceTier
-	}
 	authProviderSnapshot := readString(record, "auth_provider_snapshot", "authProviderSnapshot")
-	cacheAccounting := NormalizeCacheAccounting(CacheInputContext{
+	usageContext := CacheInputContext{
 		ExplicitMode:     cacheInputModeFromRecord(record),
 		ExecutorType:     executorType,
 		Provider:         provider,
 		ProviderSnapshot: authProviderSnapshot,
+		AuthType:         authType,
 		ResolvedModel:    resolvedModel,
 		RequestedModel:   requestedModel,
 		DisplayModel:     model,
-	}, inputTokens, cachedTokens, cacheTokens, cacheReadTokens, cacheCreationTokens)
+	}
+	serviceTier := EffectiveServiceTier(usageContext, requestServiceTier, "", responseServiceTier)
+	cacheAccounting := NormalizeCacheAccounting(usageContext, inputTokens, cachedTokens, cacheTokens, cacheReadTokens, cacheCreationTokens)
 	if totalTokens <= 0 {
 		totalTokens = cacheAccounting.TotalInputTokens + maxInt64(outputTokens, 0) + maxInt64(reasoningTokens, 0)
 	}
@@ -533,7 +562,7 @@ func NormalizeRaw(raw []byte) (Event, error) {
 		Endpoint:                      endpoint,
 		Method:                        method,
 		Path:                          path,
-		AuthType:                      readString(record, "auth_type", "authType"),
+		AuthType:                      authType,
 		AuthIndex:                     authIndex,
 		Source:                        source,
 		SourceHash:                    hashString(sourceRaw),

apps/manager-server/internal/usage/import.go

@@ -484,6 +484,12 @@ func eventFromExportedRecord(record map[string]any) (Event, bool, error) {
 		RawJSON:                       rawJSON,
 		CreatedAtMS:                   readInt(record, "created_at_ms", "createdAtMs"),
 	}
+	event.ServiceTier = EffectiveServiceTier(CacheInputContext{
+		ExecutorType:     event.ExecutorType,
+		Provider:         event.Provider,
+		ProviderSnapshot: event.AuthProviderSnapshot,
+		AuthType:         event.AuthType,
+	}, event.RequestServiceTier, event.ServiceTier, event.ResponseServiceTier)
 	if event.Endpoint == "" {
 		event.Endpoint = "-"
 	}
@@ -669,6 +675,12 @@ func eventFromLegacyDetail(
 		RawJSON:                       rawJSON,
 		CreatedAtMS:                   now,
 	}
+	event.ServiceTier = EffectiveServiceTier(CacheInputContext{
+		ExecutorType:     event.ExecutorType,
+		Provider:         event.Provider,
+		ProviderSnapshot: event.AuthProviderSnapshot,
+		AuthType:         event.AuthType,
+	}, event.RequestServiceTier, event.ServiceTier, event.ResponseServiceTier)
 	if event.Model == "" {
 		event.Model = "-"
 	}

apps/manager-server/internal/usage/import_test.go

@@ -518,21 +518,41 @@ func TestNormalizeRawHandlesCurrentCPAGPT56QueuePayloadWithoutCacheMode(t *testi
 		event.NormalizedCacheReadTokens != 30 || event.NormalizedCacheCreationTokens != 17 {
 		t.Fatalf("normalized cache accounting = %#v", event)
 	}
-	if event.RequestServiceTier != "priority" || event.ResponseServiceTier != "default" || event.ServiceTier != "default" {
+	if event.RequestServiceTier != "priority" || event.ResponseServiceTier != "default" || event.ServiceTier != "priority" {
 		t.Fatalf("service tiers = %q/%q/%q", event.RequestServiceTier, event.ResponseServiceTier, event.ServiceTier)
 	}
 }
 
-func TestNormalizeRawPrefersResponseServiceTier(t *testing.T) {
+func TestNormalizeRawPrefersRequestServiceTierForCodex(t *testing.T) {
 	payload := `{
 	  "timestamp": "2026-07-10T00:00:00Z",
+	  "executor_type": "codex",
 	  "model": "gpt-5.6-sol",
 	  "service_tier": "priority",
 	  "request_service_tier": "priority",
 	  "response_service_tier": "default",
 	  "tokens": {"input_tokens": 1, "total_tokens": 1}
 	}`
 
+	event, err := NormalizeRaw([]byte(payload))
+	if err != nil {
+		t.Fatalf("normalize raw: %v", err)
+	}
+	if event.RequestServiceTier != "priority" || event.ResponseServiceTier != "default" || event.ServiceTier != "priority" {
+		t.Fatalf("service tiers = %q/%q/%q", event.RequestServiceTier, event.ResponseServiceTier, event.ServiceTier)
+	}
+}
+
+func TestNormalizeRawPrefersResponseServiceTierForNonCodex(t *testing.T) {
+	payload := `{
+	  "timestamp": "2026-07-10T00:00:00Z",
+	  "provider": "openai-compatible",
+	  "model": "gpt-5.4",
+	  "request_service_tier": "priority",
+	  "response_service_tier": "default",
+	  "tokens": {"input_tokens": 1, "total_tokens": 1}
+	}`
+
 	event, err := NormalizeRaw([]byte(payload))
 	if err != nil {
 		t.Fatalf("normalize raw: %v", err)

apps/web/src/features/monitoring/components/RealtimeEventsPanel.test.tsx

@@ -49,6 +49,8 @@ const t = ((key: string, options?: Record<string, unknown>) => {
     'monitoring.result_failed': 'Failed',
     'monitoring.result_success': 'Success',
     'monitoring.service_tier_short': 'Tier',
+    'monitoring.request_service_tier_short': 'Requested tier',
+    'monitoring.response_service_tier_short': 'Reported tier',
     'monitoring.this_call_cost': 'Cost',
     'monitoring.this_call_usage': 'Usage',
     'monitoring.ttft_short': 'TTFT',
@@ -185,6 +187,8 @@ describe('RealtimeEventsPanel', () => {
         executorType: 'codex',
         reasoningEffort: 'medium',
         serviceTier: 'priority',
+        requestServiceTier: 'priority',
+        responseServiceTier: 'default',
         cacheReadTokens: 4,
         cacheCreationTokens: 1,
         failStatusCode: 429,
@@ -198,7 +202,8 @@ describe('RealtimeEventsPanel', () => {
     expect(markup).not.toContain('>Executor: codex<');
     expect(markup).not.toContain('Executor: codex');
     expect(markup).toContain('medium');
-    expect(markup).toContain('Tier: priority');
+    expect(markup).toContain('Requested tier: priority');
+    expect(markup).toContain('Reported tier: default');
     expect(markup).toContain('client-gpt');
     expect(markup).toContain('gpt-5.4');
     expect(markup).not.toContain('Resolved');

apps/web/src/features/monitoring/components/RealtimeEventsPanel.tsx

@@ -813,6 +813,8 @@ export function RealtimeEventsPanel({
                 row.resolvedModel.trim() !== row.model;
               const reasoningEffort = formatOptionalText(row.reasoningEffort);
               const serviceTier = formatOptionalText(row.serviceTier);
+              const requestServiceTier = formatOptionalText(row.requestServiceTier);
+              const responseServiceTier = formatOptionalText(row.responseServiceTier);
               const failureDetails = buildFailureDetails(row, t, locale);
               const failureTooltipId = failureDetails
                 ? `${tooltipIdPrefix}-failure-tooltip-${row.id}`
@@ -861,9 +863,14 @@ export function RealtimeEventsPanel({
                       ) : (
                         <span className={styles.mutedCell}>-</span>
                       )}
-                      {serviceTier !== '-' ? (
+                      {requestServiceTier !== '-' ? (
+                        <small>{`${t('monitoring.request_service_tier_short')}: ${requestServiceTier}`}</small>
+                      ) : serviceTier !== '-' ? (
                         <small>{`${shortLabel(t, 'monitoring.service_tier_short', 'monitoring.service_tier')}: ${serviceTier}`}</small>
                       ) : null}
+                      {responseServiceTier !== '-' ? (
+                        <small>{`${t('monitoring.response_service_tier_short')}: ${responseServiceTier}`}</small>
+                      ) : null}
                     </div>
                   </td>
                   <td>

apps/web/src/features/monitoring/model/eventRows.test.ts

@@ -89,13 +89,18 @@ describe('buildEventRows', () => {
     const [row] = buildRows({
       executor_type: 'codex',
       service_tier: 'priority',
+      request_service_tier: 'priority',
+      response_service_tier: 'default',
       reasoning_effort: 'medium',
     });
 
     expect(row.executorType).toBe('codex');
     expect(row.serviceTier).toBe('priority');
+    expect(row.requestServiceTier).toBe('priority');
+    expect(row.responseServiceTier).toBe('default');
     expect(row.searchText).toContain('codex');
     expect(row.searchText).toContain('priority');
+    expect(row.searchText).toContain('default');
     expect(row.searchText).toContain('medium');
   });
 

apps/web/src/features/monitoring/model/eventRows.ts

@@ -129,7 +129,16 @@ export const buildEventRows = (
       const sourceKey = sourceMeta.identityKey || `source:${sourceLabel}`;
       const taskKey = `${detail.timestamp}|${sourceKey}|${authIndex}`;
       const reasoningEffort = readString(detail.reasoning_effort ?? detail.reasoningEffort);
-      const serviceTier = readString(detail.service_tier ?? detail.serviceTier);
+      const requestServiceTier = readString(
+        detail.request_service_tier ?? detail.requestServiceTier
+      );
+      const responseServiceTier = readString(
+        detail.response_service_tier ?? detail.responseServiceTier
+      );
+      const serviceTier =
+        requestServiceTier ||
+        readString(detail.service_tier ?? detail.serviceTier) ||
+        responseServiceTier;
       const executorType = readString(detail.executor_type ?? detail.executorType);
       const failStatusCodeRaw = detail.fail_status_code ?? detail.failStatusCode;
       const failStatusCode =
@@ -217,6 +226,8 @@ export const buildEventRows = (
         totalCost,
         reasoningEffort,
         serviceTier,
+        requestServiceTier,
+        responseServiceTier,
         executorType,
         failStatusCode: normalizedFailStatusCode,
         failSummary,
@@ -247,6 +258,8 @@ export const buildEventRows = (
           projectId,
           reasoningEffort,
           serviceTier,
+          requestServiceTier,
+          responseServiceTier,
           executorType,
           normalizedFailStatusCode,
           failSummary,

apps/web/src/features/monitoring/model/types.ts

@@ -183,6 +183,8 @@ export type MonitoringEventRow = {
   totalCost: number;
   reasoningEffort?: string;
   serviceTier?: string;
+  requestServiceTier?: string;
+  responseServiceTier?: string;
   executorType?: string;
   failStatusCode?: number | null;
   failSummary?: string;

apps/web/src/i18n/locales/en.json

@@ -1773,6 +1773,8 @@
     "reasoning_effort_short": "Effort",
     "service_tier": "Service Tier",
     "service_tier_short": "Tier",
+    "request_service_tier_short": "Requested tier",
+    "response_service_tier_short": "Reported tier",
     "executor_type": "Executor Type",
     "executor_type_short": "Executor",
     "fail_status_code": "Failure Status Code",

apps/web/src/i18n/locales/ru.json

@@ -1773,6 +1773,8 @@
     "reasoning_effort_short": "Уровень",
     "service_tier": "Уровень сервиса",
     "service_tier_short": "Сервис",
+    "request_service_tier_short": "Запрошенный уровень",
+    "response_service_tier_short": "Уровень ответа",
     "executor_type": "Тип исполнителя",
     "executor_type_short": "Исполнитель",
     "fail_status_code": "Код ошибки",

apps/web/src/i18n/locales/zh-CN.json

@@ -1773,6 +1773,8 @@
     "reasoning_effort_short": "强度",
     "service_tier": "服务等级",
     "service_tier_short": "等级",
+    "request_service_tier_short": "请求等级",
+    "response_service_tier_short": "响应等级",
     "executor_type": "执行器类型",
     "executor_type_short": "执行器",
     "fail_status_code": "失败状态码",

apps/web/src/i18n/locales/zh-TW.json

@@ -1773,6 +1773,8 @@
     "reasoning_effort_short": "強度",
     "service_tier": "服務等級",
     "service_tier_short": "等級",
+    "request_service_tier_short": "請求等級",
+    "response_service_tier_short": "回應等級",
     "executor_type": "執行器類型",
     "executor_type_short": "執行器",
     "fail_status_code": "失敗狀態碼",