支持 Nano Banana Pro（Gemini Image Pro）

adapter/openai/processor.go

@@ -80,8 +80,27 @@ func getChoices(form *ChatStreamResponse) *globals.Chunk {
 
 	choice := form.Choices[0].Delta
 
+	content := choice.Content
+
+	// handle gemini multi-modal inline data (base64 image)
+	if len(content) == 0 && choice.MultiModal != nil {
+		for _, item := range *choice.MultiModal {
+			if item.InlineData == nil || len(item.InlineData.Data) == 0 {
+				continue
+			}
+
+			mime := item.InlineData.MimeType
+			if len(mime) == 0 {
+				mime = "image/png"
+			}
+			// use markdown image to let frontend直接渲染而不是显示base64文本
+			content = fmt.Sprintf("![image](data:%s;base64,%s)", mime, item.InlineData.Data)
+			break
+		}
+	}
+
 	return &globals.Chunk{
-		Content:      choice.Content,
+		Content:      content,
 		ToolCall:     choice.ToolCalls,
 		FunctionCall: choice.FunctionCall,
 	}
@@ -128,7 +147,7 @@ func (c *ChatInstance) ProcessLine(data string, isCompletionType bool) (*globals
 	}
 
 	if form := processChatErrorResponse(data); form != nil {
-		return &globals.Chunk{Content: ""}, errors.New(fmt.Sprintf("openai error: %s (type: %s)", form.Error.Message, form.Error.Type))
+		return &globals.Chunk{Content: ""}, fmt.Errorf("openai error: %s (type: %s)", form.Error.Message, form.Error.Type)
 	}
 
 	globals.Warn(fmt.Sprintf("openai error: cannot parse chat completion response: %s", data))

globals/types.go

@@ -10,6 +10,7 @@ type Message struct {
 	ToolCallId       *string       `json:"tool_call_id,omitempty"`      // only `tool` role
 	ToolCalls        *ToolCalls    `json:"tool_calls,omitempty"`        // only `assistant` role
 	ReasoningContent *string       `json:"reasoning_content,omitempty"` // only for deepseek reasoner models
+	MultiModal       *MultiModal   `json:"multi_mod_content,omitempty"` // gemini inline data (e.g., image base64)
 }
 
 type Chunk struct {
@@ -18,6 +19,17 @@ type Chunk struct {
 	FunctionCall *FunctionCall `json:"function_call,omitempty"`
 }
 
+type InlineData struct {
+	Data     string `json:"data"`
+	MimeType string `json:"mime_type"`
+}
+
+type MultiModalContent struct {
+	InlineData *InlineData `json:"inline_data,omitempty"`
+}
+
+type MultiModal []MultiModalContent
+
 type ChatSegmentResponse struct {
 	Conversation int64   `json:"conversation"`
 	Quota        float32 `json:"quota"`

utils/scanner.go

@@ -94,7 +94,10 @@ func EventScanner(props *EventScannerProps, config ...globals.ProxyConfig) *Even
 }
 
 func processFullSSE(body io.ReadCloser, callback func(string) error) *EventScannerError {
+	// allow large SSE lines (e.g., image/base64 payloads)
 	scanner := bufio.NewScanner(body)
+	buf := make([]byte, 0, 1024*1024)
+	scanner.Buffer(buf, 16*1024*1024)
 	var eventType, eventData string
 	var buffer strings.Builder
 
@@ -164,11 +167,18 @@ func processFullSSE(body io.ReadCloser, callback func(string) error) *EventScann
 		}
 	}
 
+	if err := scanner.Err(); err != nil {
+		return &EventScannerError{Error: err}
+	}
+
 	return nil
 }
 
 func processLegacySSE(body io.ReadCloser, callback func(string) error) *EventScannerError {
+	// default scanner buffer (64kb) is not enough for image/base64 chunks
 	scanner := bufio.NewScanner(body)
+	buf := make([]byte, 0, 1024*1024)
+	scanner.Buffer(buf, 16*1024*1024)
 	scanner.Split(func(data []byte, atEOF bool) (advance int, token []byte, err error) {
 		if atEOF && len(data) == 0 {
 			// when EOF and empty data
@@ -219,5 +229,9 @@ func processLegacySSE(body io.ReadCloser, callback func(string) error) *EventSca
 		}
 	}
 
+	if err := scanner.Err(); err != nil {
+		return &EventScannerError{Error: err}
+	}
+
 	return nil
 }