Propagate upstream token usage across Gemini usage metadata

API.en.md

@@ -267,6 +267,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` models emit `delta.reasoning_content`
 - Text emits `delta.content`
 - Last chunk includes `finish_reason` and `usage`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 #### Tool Calls
 
@@ -535,6 +536,7 @@ Returns SSE (`text/event-stream`), each chunk as `data: <json>`:
 - regular text: incremental text chunks
 - `tools` mode: buffered and emitted as `functionCall` at finalize phase
 - final chunk: includes `finishReason: "STOP"` and `usageMetadata`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 ---
 

API.md

@@ -267,6 +267,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` 模型输出 `delta.reasoning_content`
 - 普通文本输出 `delta.content`
 - 最后一段包含 `finish_reason` 和 `usage`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 #### Tool Calls
 
@@ -541,6 +542,7 @@ data: {"type":"message_stop"}
 - 常规文本：持续返回增量文本 chunk
 - `tools` 场景：会缓冲并在结束时输出 `functionCall` 结构
 - 结束 chunk：包含 `finishReason: "STOP"` 与 `usageMetadata`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 ---
 

internal/adapter/gemini/handler_generate.go

@@ -149,14 +149,15 @@ func (h *Handler) handleNonStreamGenerateContent(w http.ResponseWriter, resp *ht
 		cleanVisibleOutput(result.Thinking, stripReferenceMarkers),
 		cleanVisibleOutput(result.Text, stripReferenceMarkers),
 		toolNames,
+		result.PromptTokens,
 		result.OutputTokens,
 	))
 }
 
 //nolint:unused // retained for native Gemini non-stream handling path.
-func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string, outputTokens int) map[string]any {
+func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string, promptTokens, outputTokens int) map[string]any {
 	parts := buildGeminiPartsFromFinal(finalText, finalThinking, toolNames)
-	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText, outputTokens)
+	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText, promptTokens, outputTokens)
 	return map[string]any{
 		"candidates": []map[string]any{
 			{
@@ -174,8 +175,10 @@ func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, final
 }
 
 //nolint:unused // retained for native Gemini non-stream handling path.
-func buildGeminiUsage(finalPrompt, finalThinking, finalText string, outputTokens int) map[string]any {
-	promptTokens := util.EstimateTokens(finalPrompt)
+func buildGeminiUsage(finalPrompt, finalThinking, finalText string, promptTokens, outputTokens int) map[string]any {
+	if promptTokens <= 0 {
+		promptTokens = util.EstimateTokens(finalPrompt)
+	}
 	reasoningTokens := util.EstimateTokens(finalThinking)
 	completionTokens := util.EstimateTokens(finalText)
 	if outputTokens > 0 {

internal/adapter/gemini/handler_stream_runtime.go

@@ -67,6 +67,7 @@ type geminiStreamRuntime struct {
 
 	thinking     strings.Builder
 	text         strings.Builder
+	promptTokens int
 	outputTokens int
 }
 
@@ -112,6 +113,9 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
+	if parsed.PromptTokens > 0 {
+		s.promptTokens = parsed.PromptTokens
+	}
 	if parsed.OutputTokens > 0 {
 		s.outputTokens = parsed.OutputTokens
 	}
@@ -198,6 +202,6 @@ func (s *geminiStreamRuntime) finalize() {
 			},
 		},
 		"modelVersion":  s.model,
-		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText, s.outputTokens),
+		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText, s.promptTokens, s.outputTokens),
 	})
 }

internal/adapter/gemini/handler_test.go

@@ -296,6 +296,32 @@ func TestGenerateContentOpenAIProxyErrorUsesGeminiEnvelope(t *testing.T) {
 	}
 }
 
+func TestBuildGeminiUsageOverridesPromptAndOutputTokensWhenProvided(t *testing.T) {
+	usage := buildGeminiUsage("prompt", "thinking", "answer", 11, 29)
+	if got, _ := usage["promptTokenCount"].(int); got != 11 {
+		t.Fatalf("expected promptTokenCount=11, got %#v", usage["promptTokenCount"])
+	}
+	if got, _ := usage["candidatesTokenCount"].(int); got != 29 {
+		t.Fatalf("expected candidatesTokenCount=29, got %#v", usage["candidatesTokenCount"])
+	}
+	if got, _ := usage["totalTokenCount"].(int); got != 40 {
+		t.Fatalf("expected totalTokenCount=40, got %#v", usage["totalTokenCount"])
+	}
+}
+
+func TestBuildGeminiUsageFallsBackToEstimateWhenNoUpstreamUsage(t *testing.T) {
+	usage := buildGeminiUsage("abcdef", "", "ghijkl", 0, 0)
+	if got, _ := usage["promptTokenCount"].(int); got <= 0 {
+		t.Fatalf("expected positive promptTokenCount estimate, got %#v", usage["promptTokenCount"])
+	}
+	if got, _ := usage["candidatesTokenCount"].(int); got <= 0 {
+		t.Fatalf("expected positive candidatesTokenCount estimate, got %#v", usage["candidatesTokenCount"])
+	}
+	if got, _ := usage["totalTokenCount"].(int); got <= 0 {
+		t.Fatalf("expected positive totalTokenCount estimate, got %#v", usage["totalTokenCount"])
+	}
+}
+
 func extractGeminiSSEFrames(t *testing.T, body string) []map[string]any {
 	t.Helper()
 	scanner := bufio.NewScanner(strings.NewReader(body))