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

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,
-			Ephemeral1hInputTokens: oaiUsage.ClaudeCacheCreation1hTokens,
+			Ephemeral5mInputTokens: cacheCreation5m,
+			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
-		// 可能为非标准的 OpenAI 响应，判断是否已经完成
-		if info.ClaudeConvertInfo.Done {
+		// Some OpenAI-compatible upstreams end with a usage-only SSE chunk.
+		oaiUsage := openAIResponse.Usage
+		if oaiUsage == nil {
+			oaiUsage = info.ClaudeConvertInfo.Usage
+		}
+		if oaiUsage != nil {
 			stopOpenBlocks()
-			oaiUsage := info.ClaudeConvertInfo.Usage
-			if oaiUsage != nil {
-				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
-					Type:  "message_delta",
-					Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
-					Delta: &dto.ClaudeMediaMessage{
-						StopReason: common.GetPointer[string](stopReasonOpenAI2Claude(info.FinishReason)),
-					},
-				})
+			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](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