fix: emit claude message_delta for usage-only final stream chunk

relay/channel/claude/message_delta_usage_patch_test.go

@@ -85,7 +85,7 @@ func TestBuildMessageDeltaPatchUsage(t *testing.T) {
 		require.EqualValues(t, 50, usage.CacheCreationInputTokens)
 		require.EqualValues(t, 53, usage.OutputTokens)
 		require.NotNil(t, usage.CacheCreation)
-		require.EqualValues(t, 10, usage.CacheCreation.Ephemeral5mInputTokens)
+		require.EqualValues(t, 30, usage.CacheCreation.Ephemeral5mInputTokens)
 		require.EqualValues(t, 20, usage.CacheCreation.Ephemeral1hInputTokens)
 	})
 
@@ -108,4 +108,22 @@ func TestBuildMessageDeltaPatchUsage(t *testing.T) {
 		require.EqualValues(t, 7, usage.CacheReadInputTokens)
 		require.EqualValues(t, 6, usage.CacheCreationInputTokens)
 	})
+
+	t.Run("default aggregate cache creation to 5m when split missing", func(t *testing.T) {
+		claudeResponse := &dto.ClaudeResponse{Usage: &dto.ClaudeUsage{
+			OutputTokens:             53,
+			CacheCreationInputTokens: 50,
+		}}
+		claudeInfo := &ClaudeResponseInfo{Usage: &dto.Usage{
+			PromptTokensDetails: dto.InputTokenDetails{
+				CachedCreationTokens: 50,
+			},
+		}}
+
+		usage := buildMessageDeltaPatchUsage(claudeResponse, claudeInfo)
+		require.NotNil(t, usage)
+		require.NotNil(t, usage.CacheCreation)
+		require.EqualValues(t, 50, usage.CacheCreation.Ephemeral5mInputTokens)
+		require.EqualValues(t, 0, usage.CacheCreation.Ephemeral1hInputTokens)
+	})
 }

relay/channel/claude/relay-claude.go

@@ -648,6 +648,11 @@ func buildOpenAIStyleUsageFromClaudeUsage(usage *dto.Usage) dto.Usage {
 		return dto.Usage{}
 	}
 	clone := *usage
+	clone.ClaudeCacheCreation5mTokens, clone.ClaudeCacheCreation1hTokens = service.NormalizeCacheCreationSplit(
+		usage.PromptTokensDetails.CachedCreationTokens,
+		usage.ClaudeCacheCreation5mTokens,
+		usage.ClaudeCacheCreation1hTokens,
+	)
 	cacheCreationTokens := cacheCreationTokensForOpenAIUsage(usage)
 	totalInputTokens := usage.PromptTokens + usage.PromptTokensDetails.CachedTokens + cacheCreationTokens
 	clone.PromptTokens = totalInputTokens
@@ -677,11 +682,26 @@ func buildMessageDeltaPatchUsage(claudeResponse *dto.ClaudeResponse, claudeInfo
 	if usage.CacheCreationInputTokens == 0 && claudeInfo.Usage.PromptTokensDetails.CachedCreationTokens > 0 {
 		usage.CacheCreationInputTokens = claudeInfo.Usage.PromptTokensDetails.CachedCreationTokens
 	}
-	if usage.CacheCreation == nil && (claudeInfo.Usage.ClaudeCacheCreation5mTokens > 0 || claudeInfo.Usage.ClaudeCacheCreation1hTokens > 0) {
+	cacheCreation5m := 0
-		usage.CacheCreation = &dto.ClaudeCacheCreationUsage{
+	cacheCreation1h := 0
-			Ephemeral5mInputTokens: claudeInfo.Usage.ClaudeCacheCreation5mTokens,
+	if usage.CacheCreation != nil {
-			Ephemeral1hInputTokens: claudeInfo.Usage.ClaudeCacheCreation1hTokens,
+		cacheCreation5m = usage.CacheCreation.Ephemeral5mInputTokens
+		cacheCreation1h = usage.CacheCreation.Ephemeral1hInputTokens
+	} else {
+		cacheCreation5m = claudeInfo.Usage.ClaudeCacheCreation5mTokens
+		cacheCreation1h = claudeInfo.Usage.ClaudeCacheCreation1hTokens
 	}
+	cacheCreation5m, cacheCreation1h = service.NormalizeCacheCreationSplit(
+		usage.CacheCreationInputTokens,
+		cacheCreation5m,
+		cacheCreation1h,
+	)
+	if usage.CacheCreation == nil && (cacheCreation5m > 0 || cacheCreation1h > 0) {
+		usage.CacheCreation = &dto.ClaudeCacheCreationUsage{}
+	}
+	if usage.CacheCreation != nil {
+		usage.CacheCreation.Ephemeral5mInputTokens = cacheCreation5m
+		usage.CacheCreation.Ephemeral1hInputTokens = cacheCreation1h
 	}
 	return usage
 }

relay/channel/claude/relay_claude_test.go

@@ -258,6 +258,23 @@ func TestBuildOpenAIStyleUsageFromClaudeUsagePreservesCacheCreationRemainder(t *
 	}
 }
 
+func TestBuildOpenAIStyleUsageFromClaudeUsageDefaultsAggregateCacheCreationTo5m(t *testing.T) {
+	usage := &dto.Usage{
+		PromptTokens:     100,
+		CompletionTokens: 20,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens:         30,
+			CachedCreationTokens: 50,
+		},
+		UsageSemantic: "anthropic",
+	}
+
+	openAIUsage := buildOpenAIStyleUsageFromClaudeUsage(usage)
+
+	require.Equal(t, 50, openAIUsage.ClaudeCacheCreation5mTokens)
+	require.Equal(t, 0, openAIUsage.ClaudeCacheCreation1hTokens)
+}
+
 func TestRequestOpenAI2ClaudeMessage_IgnoresUnsupportedFileContent(t *testing.T) {
 	request := dto.GeneralOpenAIRequest{
 		Model: "claude-3-5-sonnet",

service/convert.go

@@ -227,21 +227,31 @@ func buildClaudeUsageFromOpenAIUsage(oaiUsage *dto.Usage) *dto.ClaudeUsage {
 	if oaiUsage == nil {
 		return nil
 	}
+	cacheCreation5m, cacheCreation1h := NormalizeCacheCreationSplit(
+		oaiUsage.PromptTokensDetails.CachedCreationTokens,
+		oaiUsage.ClaudeCacheCreation5mTokens,
+		oaiUsage.ClaudeCacheCreation1hTokens,
+	)
 	usage := &dto.ClaudeUsage{
 		InputTokens:              oaiUsage.PromptTokens,
 		OutputTokens:             oaiUsage.CompletionTokens,
 		CacheCreationInputTokens: oaiUsage.PromptTokensDetails.CachedCreationTokens,
 		CacheReadInputTokens:     oaiUsage.PromptTokensDetails.CachedTokens,
 	}
-	if oaiUsage.ClaudeCacheCreation5mTokens > 0 || oaiUsage.ClaudeCacheCreation1hTokens > 0 {
+	if cacheCreation5m > 0 || cacheCreation1h > 0 {
 		usage.CacheCreation = &dto.ClaudeCacheCreationUsage{
-			Ephemeral5mInputTokens: oaiUsage.ClaudeCacheCreation5mTokens,
+			Ephemeral5mInputTokens: cacheCreation5m,
-			Ephemeral1hInputTokens: oaiUsage.ClaudeCacheCreation1hTokens,
+			Ephemeral1hInputTokens: cacheCreation1h,
 		}
 	}
 	return usage
 }
 
+func NormalizeCacheCreationSplit(totalTokens int, tokens5m int, tokens1h int) (int, int) {
+	remainder := lo.Max([]int{totalTokens - tokens5m - tokens1h, 0})
+	return tokens5m + remainder, tokens1h
+}
+
 func StreamResponseOpenAI2Claude(openAIResponse *dto.ChatCompletionsStreamResponse, info *relaycommon.RelayInfo) []*dto.ClaudeResponse {
 	if info.ClaudeConvertInfo.Done {
 		return nil
@@ -426,23 +436,28 @@ func StreamResponseOpenAI2Claude(openAIResponse *dto.ChatCompletionsStreamRespon
 	}
 
 	if len(openAIResponse.Choices) == 0 {
-		// no choices
+		// Some OpenAI-compatible upstreams end with a usage-only SSE chunk.
-		// 可能为非标准的 OpenAI 响应，判断是否已经完成
+		oaiUsage := openAIResponse.Usage
-		if info.ClaudeConvertInfo.Done {
+		if oaiUsage == nil {
-			stopOpenBlocks()
+			oaiUsage = info.ClaudeConvertInfo.Usage
-			oaiUsage := info.ClaudeConvertInfo.Usage
+		}
 		if oaiUsage != nil {
+			stopOpenBlocks()
+			stopReason := stopReasonOpenAI2Claude(info.FinishReason)
+			if stopReason == "" {
+				stopReason = "end_turn"
+			}
 			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
 				Type:  "message_delta",
 				Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
 				Delta: &dto.ClaudeMediaMessage{
-						StopReason: common.GetPointer[string](stopReasonOpenAI2Claude(info.FinishReason)),
+					StopReason: common.GetPointer[string](stopReason),
 				},
 			})
-			}
 			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
 				Type: "message_stop",
 			})
+			info.ClaudeConvertInfo.Done = true
 		}
 		return claudeResponses
 	} else {
@@ -450,6 +465,13 @@ func StreamResponseOpenAI2Claude(openAIResponse *dto.ChatCompletionsStreamRespon
 		doneChunk := chosenChoice.FinishReason != nil && *chosenChoice.FinishReason != ""
 		if doneChunk {
 			info.FinishReason = *chosenChoice.FinishReason
+			oaiUsage := openAIResponse.Usage
+			if oaiUsage == nil {
+				oaiUsage = info.ClaudeConvertInfo.Usage
+				// Some upstreams emit finish_reason first, then send a final usage-only chunk.
+				// Defer closing until usage is available so the final message_delta carries it.
+				return claudeResponses
+			}
 		}
 
 		var claudeResponse dto.ClaudeResponse