fix(usage): WS passthrough effort 提取补入映射后模型候选

Fixes #3924

backend/internal/service/openai_fast_policy_ws_test.go

@@ -1015,7 +1015,7 @@ func TestPassthroughUsageMeta_TracksReasoningEffortAcrossTurns(t *testing.T) {
 	firstOut, firstBlocked, firstErr := svc.applyOpenAIFastPolicyToWSResponseCreate(context.Background(), account, capturedSessionModel, firstFrame)
 	require.NoError(t, firstErr)
 	require.Nil(t, firstBlocked)
-	meta.initFromFirstFrame(firstOut)
+	meta.initFromFirstFrame(firstOut, capturedSessionModel)
 	require.NotNil(t, meta.reasoningEffort.Load())
 	require.Equal(t, "medium", *meta.reasoningEffort.Load())
 
@@ -1032,7 +1032,7 @@ func TestPassthroughUsageMeta_TracksReasoningEffortAcrossTurns(t *testing.T) {
 		out, blocked, policyErr := svc.applyOpenAIFastPolicyToWSResponseCreate(context.Background(), account, model, payload)
 		if policyErr == nil && blocked == nil &&
 			strings.TrimSpace(gjson.GetBytes(payload, "type").String()) == "response.create" {
-			meta.updateFromResponseCreate(out, requestModelForThisFrame)
+			meta.updateFromResponseCreate(out, model, requestModelForThisFrame)
 		}
 		return out, blocked, policyErr
 	}

backend/internal/service/openai_ws_v2_passthrough_adapter.go

@@ -142,12 +142,12 @@ func newOpenAIWSPassthroughUsageMeta(initialRequestModel string, firstFrame []by
 	return meta
 }
 
-func (m *openAIWSPassthroughUsageMeta) initFromFirstFrame(policyOutput []byte) {
+func (m *openAIWSPassthroughUsageMeta) initFromFirstFrame(policyOutput []byte, mappedModel string) {
 	if m == nil {
 		return
 	}
 	m.serviceTier.Store(extractOpenAIServiceTierFromBody(policyOutput))
-	m.reasoningEffort.Store(extractOpenAIReasoningEffortFromBody(policyOutput, m.sessionRequestModel))
+	m.reasoningEffort.Store(extractOpenAIReasoningEffortFromBody(policyOutput, mappedModel, m.sessionRequestModel))
 }
 
 func (m *openAIWSPassthroughUsageMeta) updateSessionRequestModel(payload []byte) {
@@ -169,12 +169,12 @@ func (m *openAIWSPassthroughUsageMeta) requestModelForFrame(payload []byte) stri
 	return m.sessionRequestModel
 }
 
-func (m *openAIWSPassthroughUsageMeta) updateFromResponseCreate(policyOutput []byte, requestModelForFrame string) {
+func (m *openAIWSPassthroughUsageMeta) updateFromResponseCreate(policyOutput []byte, mappedModel string, requestModelForFrame string) {
 	if m == nil {
 		return
 	}
 	m.serviceTier.Store(extractOpenAIServiceTierFromBody(policyOutput))
-	m.reasoningEffort.Store(extractOpenAIReasoningEffortFromBody(policyOutput, requestModelForFrame))
+	m.reasoningEffort.Store(extractOpenAIReasoningEffortFromBody(policyOutput, mappedModel, requestModelForFrame))
 }
 
 func openAIWSPassthroughRequestModelForFrame(payload []byte) string {
@@ -311,7 +311,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough(
 	// 因此使用 atomic.Pointer[string] 在 filter（runClientToUpstream
 	// goroutine）和 OnTurnComplete / final result（runUpstreamToClient
 	// goroutine）之间同步当前 turn 的 usage metadata。
-	usageMeta.initFromFirstFrame(firstClientMessage)
+	usageMeta.initFromFirstFrame(firstClientMessage, capturedSessionModel)
 	promptCacheKey := strings.TrimSpace(gjson.GetBytes(firstClientMessage, "prompt_cache_key").String())
 
 	wsURL, err := s.buildOpenAIResponsesWSURL(account)
@@ -455,7 +455,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough(
 			//     service_tier 时按 default 处理，billing 应如实反映。
 			if policyErr == nil && blocked == nil &&
 				strings.TrimSpace(gjson.GetBytes(payload, "type").String()) == "response.create" {
-				usageMeta.updateFromResponseCreate(out, requestModelForThisFrame)
+				usageMeta.updateFromResponseCreate(out, model, requestModelForThisFrame)
 			}
 			return out, blocked, policyErr
 		},

backend/internal/service/openai_ws_v2_passthrough_adapter_effort_test.go

@@ -0,0 +1,42 @@
+//go:build unit
+
+package service
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestWSPassthroughUsageMeta_InitFromFirstFrame_MappedModelCandidate(t *testing.T) {
+	body := []byte(`{"type":"response.create","model":"sol","reasoning":{"effort":"max"}}`)
+
+	meta := newOpenAIWSPassthroughUsageMeta("sol", body)
+	meta.initFromFirstFrame(body, "gpt-5.6-sol")
+
+	got := meta.reasoningEffort.Load()
+	require.NotNil(t, got, "reasoning effort should be set")
+	require.Equal(t, "max", *got, "mapped model gpt-5.6-sol should preserve max")
+}
+
+func TestWSPassthroughUsageMeta_InitFromFirstFrame_NonGPT56FallsBackToXHigh(t *testing.T) {
+	body := []byte(`{"type":"response.create","model":"gpt-5.4","reasoning":{"effort":"max"}}`)
+
+	meta := newOpenAIWSPassthroughUsageMeta("gpt-5.4", body)
+	meta.initFromFirstFrame(body, "gpt-5.4")
+
+	got := meta.reasoningEffort.Load()
+	require.NotNil(t, got)
+	require.Equal(t, "xhigh", *got, "non-5.6 model should normalize max to xhigh")
+}
+
+func TestWSPassthroughUsageMeta_UpdateFromResponseCreate_MappedModelCandidate(t *testing.T) {
+	body := []byte(`{"type":"response.create","model":"sol","reasoning":{"effort":"max"}}`)
+
+	meta := newOpenAIWSPassthroughUsageMeta("sol", body)
+	meta.updateFromResponseCreate(body, "gpt-5.6-sol", "sol")
+
+	got := meta.reasoningEffort.Load()
+	require.NotNil(t, got)
+	require.Equal(t, "max", *got, "mapped model should preserve max on multi-turn update")
+}