Add system instruction support and enhance internal API handlers

- Introduced `SystemInstruction` field in `PrepareRequest` and `GenerateContentRequest` for better message parsing.
- Updated `SendMessage` and `SendMessageStream` to handle system instructions in client API calls.
- Enhanced error handling and manual flushing logic in response flows.
- Added new internal API endpoints `/v1internal:generateContent` and `/v1internal:streamGenerateContent`.
- Improved proxy handling and transport logic in HTTP client initialization.

internal/api/handlers.go

@@ -1,6 +1,7 @@
 package api
 
 import (
+	"bytes"
 	"context"
 	"fmt"
 	"github.com/luispater/CLIProxyAPI/internal/api/translator"
@@ -8,7 +9,12 @@ import (
 	"github.com/luispater/CLIProxyAPI/internal/config"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+	"golang.org/x/net/proxy"
+	"io"
+	"net"
 	"net/http"
+	"net/url"
 	"sync"
 	"time"
 
@@ -196,19 +202,7 @@ func (h *APIHandlers) ChatCompletions(c *gin.Context) {
 func (h *APIHandlers) handleNonStreamingResponse(c *gin.Context, rawJson []byte) {
 	c.Header("Content-Type", "application/json")
 
-	// Handle streaming manually
-	flusher, ok := c.Writer.(http.Flusher)
-	if !ok {
-		c.JSON(http.StatusInternalServerError, ErrorResponse{
-			Error: ErrorDetail{
-				Message: "Streaming not supported",
-				Type:    "server_error",
-			},
-		})
-		return
-	}
-
-	modelName, contents, tools := translator.PrepareRequest(rawJson)
+	modelName, systemInstruction, contents, tools := translator.PrepareRequest(rawJson)
 	cliCtx, cliCancel := context.WithCancel(context.Background())
 	var cliClient *client.Client
 	defer func() {
@@ -239,8 +233,7 @@ func (h *APIHandlers) handleNonStreamingResponse(c *gin.Context, rawJson []byte)
 
 		if len(reorderedClients) == 0 {
 			c.Status(429)
-			_, _ = fmt.Fprint(c.Writer, fmt.Sprintf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, modelName))
-			flusher.Flush()
+			_, _ = c.Writer.Write([]byte(fmt.Sprintf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, modelName)))
 			cliCancel()
 			return
 		}
@@ -266,22 +259,20 @@ func (h *APIHandlers) handleNonStreamingResponse(c *gin.Context, rawJson []byte)
 			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
 		}
 
-		resp, err := cliClient.SendMessage(cliCtx, rawJson, modelName, contents, tools)
+		resp, err := cliClient.SendMessage(cliCtx, rawJson, modelName, systemInstruction, contents, tools)
 		if err != nil {
 			if err.StatusCode == 429 && h.cfg.QuotaExceeded.SwitchProject {
 				continue
 			} else {
 				c.Status(err.StatusCode)
-				_, _ = fmt.Fprint(c.Writer, err.Error.Error())
-				flusher.Flush()
+				_, _ = c.Writer.Write([]byte(err.Error.Error()))
 				cliCancel()
 			}
 			break
 		} else {
 			openAIFormat := translator.ConvertCliToOpenAINonStream(resp, time.Now().Unix(), isGlAPIKey)
 			if openAIFormat != "" {
-				_, _ = fmt.Fprintf(c.Writer, "data: %s\n\n", openAIFormat)
-				flusher.Flush()
+				_, _ = c.Writer.Write([]byte(openAIFormat))
 			}
 			cliCancel()
 			break
@@ -309,7 +300,7 @@ func (h *APIHandlers) handleStreamingResponse(c *gin.Context, rawJson []byte) {
 	}
 
 	// Prepare the request for the backend client.
-	modelName, contents, tools := translator.PrepareRequest(rawJson)
+	modelName, systemInstruction, contents, tools := translator.PrepareRequest(rawJson)
 	cliCtx, cliCancel := context.WithCancel(context.Background())
 	var cliClient *client.Client
 	defer func() {
@@ -369,7 +360,7 @@ outLoop:
 			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
 		}
 		// Send the message and receive response chunks and errors via channels.
-		respChan, errChan := cliClient.SendMessageStream(cliCtx, rawJson, modelName, contents, tools)
+		respChan, errChan := cliClient.SendMessageStream(cliCtx, rawJson, modelName, systemInstruction, contents, tools)
 		hasFirstResponse := false
 		for {
 			select {
@@ -420,3 +411,295 @@ outLoop:
 		}
 	}
 }
+
+func (h *APIHandlers) Internal(c *gin.Context) {
+	rawJson, _ := c.GetRawData()
+	requestRawURI := c.Request.URL.Path
+	if requestRawURI == "/v1internal:generateContent" {
+		h.internalGenerateContent(c, rawJson)
+	} else if requestRawURI == "/v1internal:streamGenerateContent" {
+		h.internalStreamGenerateContent(c, rawJson)
+	} else {
+		reqBody := bytes.NewBuffer(rawJson)
+		req, err := http.NewRequest("POST", fmt.Sprintf("https://cloudcode-pa.googleapis.com%s", c.Request.URL.RequestURI()), reqBody)
+		if err != nil {
+			c.JSON(http.StatusBadRequest, ErrorResponse{
+				Error: ErrorDetail{
+					Message: fmt.Sprintf("Invalid request: %v", err),
+					Type:    "invalid_request_error",
+				},
+			})
+			return
+		}
+		for key, value := range c.Request.Header {
+			req.Header[key] = value
+		}
+
+		var transport *http.Transport
+		proxyURL, errParse := url.Parse(h.cfg.ProxyUrl)
+		if errParse == nil {
+			if proxyURL.Scheme == "socks5" {
+				username := proxyURL.User.Username()
+				password, _ := proxyURL.User.Password()
+				proxyAuth := &proxy.Auth{User: username, Password: password}
+				dialer, errSOCKS5 := proxy.SOCKS5("tcp", proxyURL.Host, proxyAuth, proxy.Direct)
+				if errSOCKS5 != nil {
+					log.Fatalf("create SOCKS5 dialer failed: %v", errSOCKS5)
+				}
+				transport = &http.Transport{
+					DialContext: func(ctx context.Context, network, addr string) (net.Conn, error) {
+						return dialer.Dial(network, addr)
+					},
+				}
+			} else if proxyURL.Scheme == "http" || proxyURL.Scheme == "https" {
+				transport = &http.Transport{Proxy: http.ProxyURL(proxyURL)}
+			}
+		}
+		httpClient := &http.Client{}
+		if transport != nil {
+			httpClient.Transport = transport
+		}
+
+		resp, err := httpClient.Do(req)
+		if err != nil {
+			c.JSON(http.StatusBadRequest, ErrorResponse{
+				Error: ErrorDetail{
+					Message: fmt.Sprintf("Invalid request: %v", err),
+					Type:    "invalid_request_error",
+				},
+			})
+			return
+		}
+
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			defer func() {
+				if err = resp.Body.Close(); err != nil {
+					log.Printf("warn: failed to close response body: %v", err)
+				}
+			}()
+			bodyBytes, _ := io.ReadAll(resp.Body)
+
+			c.JSON(http.StatusBadRequest, ErrorResponse{
+				Error: ErrorDetail{
+					Message: string(bodyBytes),
+					Type:    "invalid_request_error",
+				},
+			})
+			return
+		}
+
+		defer func() {
+			_ = resp.Body.Close()
+		}()
+
+		for key, value := range resp.Header {
+			c.Header(key, value[0])
+		}
+		output, err := io.ReadAll(resp.Body)
+		if err != nil {
+			log.Errorf("Failed to read response body: %v", err)
+			return
+		}
+		_, _ = c.Writer.Write(output)
+	}
+}
+
+func (h *APIHandlers) internalStreamGenerateContent(c *gin.Context, rawJson []byte) {
+	// Get the http.Flusher interface to manually flush the response.
+	flusher, ok := c.Writer.(http.Flusher)
+	if !ok {
+		c.JSON(http.StatusInternalServerError, ErrorResponse{
+			Error: ErrorDetail{
+				Message: "Streaming not supported",
+				Type:    "server_error",
+			},
+		})
+		return
+	}
+
+	modelResult := gjson.GetBytes(rawJson, "model")
+	modelName := modelResult.String()
+
+	cliCtx, cliCancel := context.WithCancel(context.Background())
+	var cliClient *client.Client
+	defer func() {
+		// Ensure the client's mutex is unlocked on function exit.
+		if cliClient != nil {
+			cliClient.RequestMutex.Unlock()
+		}
+	}()
+
+outLoop:
+	for {
+		// Lock the mutex to update the last used client index
+		mutex.Lock()
+		startIndex := lastUsedClientIndex
+		currentIndex := (startIndex + 1) % len(h.cliClients)
+		lastUsedClientIndex = currentIndex
+		mutex.Unlock()
+
+		// Reorder the client to start from the last used index
+		reorderedClients := make([]*client.Client, 0)
+		for i := 0; i < len(h.cliClients); i++ {
+			cliClient = h.cliClients[(startIndex+1+i)%len(h.cliClients)]
+			if cliClient.IsModelQuotaExceeded(modelName) {
+				log.Debugf("Model %s is quota exceeded for account %s, project id: %s", modelName, cliClient.GetEmail(), cliClient.GetProjectID())
+				cliClient = nil
+				continue
+			}
+			reorderedClients = append(reorderedClients, cliClient)
+		}
+
+		if len(reorderedClients) == 0 {
+			c.Status(429)
+			_, _ = fmt.Fprint(c.Writer, fmt.Sprintf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, modelName))
+			flusher.Flush()
+			cliCancel()
+			return
+		}
+
+		locked := false
+		for i := 0; i < len(reorderedClients); i++ {
+			cliClient = reorderedClients[i]
+			if cliClient.RequestMutex.TryLock() {
+				locked = true
+				break
+			}
+		}
+		if !locked {
+			cliClient = h.cliClients[0]
+			cliClient.RequestMutex.Lock()
+		}
+
+		if glAPIKey := cliClient.GetGenerativeLanguageAPIKey(); glAPIKey != "" {
+			log.Debugf("Request use generative language API Key: %s", glAPIKey)
+		} else {
+			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
+		}
+		// Send the message and receive response chunks and errors via channels.
+		respChan, errChan := cliClient.SendRawMessageStream(cliCtx, rawJson)
+		hasFirstResponse := false
+		for {
+			select {
+			// Handle client disconnection.
+			case <-c.Request.Context().Done():
+				if c.Request.Context().Err().Error() == "context canceled" {
+					log.Debugf("Client disconnected: %v", c.Request.Context().Err())
+					cliCancel() // Cancel the backend request.
+					return
+				}
+			// Process incoming response chunks.
+			case chunk, okStream := <-respChan:
+				if !okStream {
+					cliCancel()
+					return
+				} else {
+					hasFirstResponse = true
+					if cliClient.GetGenerativeLanguageAPIKey() != "" {
+						chunk, _ = sjson.SetRawBytes(chunk, "response", chunk)
+					}
+					_, _ = c.Writer.Write([]byte("data: "))
+					_, _ = c.Writer.Write(chunk)
+					_, _ = c.Writer.Write([]byte("\n\n"))
+					flusher.Flush()
+				}
+			// Handle errors from the backend.
+			case err, okError := <-errChan:
+				if okError {
+					if err.StatusCode == 429 && h.cfg.QuotaExceeded.SwitchProject {
+						continue outLoop
+					} else {
+						c.Status(err.StatusCode)
+						_, _ = fmt.Fprint(c.Writer, err.Error.Error())
+						flusher.Flush()
+						cliCancel()
+					}
+					return
+				}
+			// Send a keep-alive signal to the client.
+			case <-time.After(500 * time.Millisecond):
+				if hasFirstResponse {
+					_, _ = c.Writer.Write([]byte("\n"))
+					flusher.Flush()
+				}
+			}
+		}
+	}
+}
+
+func (h *APIHandlers) internalGenerateContent(c *gin.Context, rawJson []byte) {
+	c.Header("Content-Type", "application/json")
+
+	modelResult := gjson.GetBytes(rawJson, "model")
+	modelName := modelResult.String()
+	cliCtx, cliCancel := context.WithCancel(context.Background())
+	var cliClient *client.Client
+	defer func() {
+		if cliClient != nil {
+			cliClient.RequestMutex.Unlock()
+		}
+	}()
+
+	for {
+		// Lock the mutex to update the last used client index
+		mutex.Lock()
+		startIndex := lastUsedClientIndex
+		currentIndex := (startIndex + 1) % len(h.cliClients)
+		lastUsedClientIndex = currentIndex
+		mutex.Unlock()
+
+		// Reorder the client to start from the last used index
+		reorderedClients := make([]*client.Client, 0)
+		for i := 0; i < len(h.cliClients); i++ {
+			cliClient = h.cliClients[(startIndex+1+i)%len(h.cliClients)]
+			if cliClient.IsModelQuotaExceeded(modelName) {
+				log.Debugf("Model %s is quota exceeded for account %s, project id: %s", modelName, cliClient.GetEmail(), cliClient.GetProjectID())
+				cliClient = nil
+				continue
+			}
+			reorderedClients = append(reorderedClients, cliClient)
+		}
+
+		if len(reorderedClients) == 0 {
+			c.Status(429)
+			_, _ = c.Writer.Write([]byte(fmt.Sprintf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, modelName)))
+			cliCancel()
+			return
+		}
+
+		locked := false
+		for i := 0; i < len(reorderedClients); i++ {
+			cliClient = reorderedClients[i]
+			if cliClient.RequestMutex.TryLock() {
+				locked = true
+				break
+			}
+		}
+		if !locked {
+			cliClient = h.cliClients[0]
+			cliClient.RequestMutex.Lock()
+		}
+
+		if glAPIKey := cliClient.GetGenerativeLanguageAPIKey(); glAPIKey != "" {
+			log.Debugf("Request use generative language API Key: %s", glAPIKey)
+		} else {
+			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
+		}
+
+		resp, err := cliClient.SendRawMessage(cliCtx, rawJson)
+		if err != nil {
+			if err.StatusCode == 429 && h.cfg.QuotaExceeded.SwitchProject {
+				continue
+			} else {
+				c.Status(err.StatusCode)
+				_, _ = c.Writer.Write([]byte(err.Error.Error()))
+				cliCancel()
+			}
+			break
+		} else {
+			_, _ = c.Writer.Write(resp)
+			cliCancel()
+			break
+		}
+	}
+}

internal/api/server.go

@@ -81,6 +81,8 @@ func (s *Server) setupRoutes() {
 			},
 		})
 	})
+	s.engine.POST("/v1internal:method", s.handlers.Internal)
+
 }
 
 // Start begins listening for and serving HTTP requests.

internal/api/translator/request.go

@@ -12,7 +12,7 @@ import (
 // PrepareRequest translates a raw JSON request from an OpenAI-compatible format
 // to the internal format expected by the backend client. It parses messages,
 // roles, content types (text, image, file), and tool calls.
-func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDeclaration) {
+func PrepareRequest(rawJson []byte) (string, *client.Content, []client.Content, []client.ToolDeclaration) {
 	// Extract the model name from the request, defaulting to "gemini-2.5-pro".
 	modelName := "gemini-2.5-pro"
 	modelResult := gjson.GetBytes(rawJson, "model")
@@ -22,6 +22,7 @@ func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDecl
 
 	// Process the array of messages.
 	contents := make([]client.Content, 0)
+	var systemInstruction *client.Content
 	messagesResult := gjson.GetBytes(rawJson, "messages")
 	if messagesResult.IsArray() {
 		messagesResults := messagesResult.Array()
@@ -37,13 +38,11 @@ func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDecl
 			// System messages are converted to a user message followed by a model's acknowledgment.
 			case "system":
 				if contentResult.Type == gjson.String {
-					contents = append(contents, client.Content{Role: "user", Parts: []client.Part{{Text: contentResult.String()}}})
-					contents = append(contents, client.Content{Role: "model", Parts: []client.Part{{Text: "Understood. I will follow these instructions and use my tools to assist you."}}})
+					systemInstruction = &client.Content{Role: "user", Parts: []client.Part{{Text: contentResult.String()}}}
 				} else if contentResult.IsObject() {
 					// Handle object-based system messages.
 					if contentResult.Get("type").String() == "text" {
-						contents = append(contents, client.Content{Role: "user", Parts: []client.Part{{Text: contentResult.Get("text").String()}}})
-						contents = append(contents, client.Content{Role: "model", Parts: []client.Part{{Text: "Understood. I will follow these instructions and use my tools to assist you."}}})
+						systemInstruction = &client.Content{Role: "user", Parts: []client.Part{{Text: contentResult.Get("text").String()}}}
 					}
 				}
 			// User messages can contain simple text or a multi-part body.
@@ -159,5 +158,5 @@ func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDecl
 		tools = make([]client.ToolDeclaration, 0)
 	}
 
-	return modelName, contents, tools
+	return modelName, systemInstruction, contents, tools
 }

internal/client/client.go

@@ -266,6 +266,12 @@ func (c *Client) APIRequest(ctx context.Context, endpoint string, body interface
 			url = url + "?alt=sse"
 		}
 		jsonBody = []byte(gjson.GetBytes(jsonBody, "request").Raw)
+		systemInstructionResult := gjson.GetBytes(jsonBody, "systemInstruction")
+		if systemInstructionResult.Exists() {
+			jsonBody, _ = sjson.SetRawBytes(jsonBody, "system_instruction", []byte(systemInstructionResult.Raw))
+			jsonBody, _ = sjson.DeleteBytes(jsonBody, "systemInstruction")
+			jsonBody, _ = sjson.DeleteBytes(jsonBody, "session_id")
+		}
 	}
 
 	// log.Debug(string(jsonBody))
@@ -303,15 +309,14 @@ func (c *Client) APIRequest(ctx context.Context, endpoint string, body interface
 			}
 		}()
 		bodyBytes, _ := io.ReadAll(resp.Body)
-
 		return nil, &ErrorMessage{resp.StatusCode, fmt.Errorf(string(bodyBytes))}
 	}
 
 	return resp.Body, nil
 }
 
-// SendMessageStream handles a single conversational turn, including tool calls.
-func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string, contents []Content, tools []ToolDeclaration) ([]byte, *ErrorMessage) {
+// SendMessage handles a single conversational turn, including tool calls.
+func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string, systemInstruction *Content, contents []Content, tools []ToolDeclaration) ([]byte, *ErrorMessage) {
 	request := GenerateContentRequest{
 		Contents: contents,
 		GenerationConfig: GenerationConfig{
@@ -320,6 +325,9 @@ func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string,
 			},
 		},
 	}
+
+	request.SystemInstruction = systemInstruction
+
 	request.Tools = tools
 
 	requestBody := map[string]interface{}{
@@ -402,7 +410,7 @@ func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string,
 }
 
 // SendMessageStream handles a single conversational turn, including tool calls.
-func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model string, contents []Content, tools []ToolDeclaration) (<-chan []byte, <-chan *ErrorMessage) {
+func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model string, systemInstruction *Content, contents []Content, tools []ToolDeclaration) (<-chan []byte, <-chan *ErrorMessage) {
 	dataTag := []byte("data: ")
 	errChan := make(chan *ErrorMessage)
 	dataChan := make(chan []byte)
@@ -418,6 +426,9 @@ func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model st
 				},
 			},
 		}
+
+		request.SystemInstruction = systemInstruction
+
 		request.Tools = tools
 
 		requestBody := map[string]interface{}{
@@ -519,6 +530,117 @@ func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model st
 	return dataChan, errChan
 }
 
+// SendRawMessage handles a single conversational turn, including tool calls.
+func (c *Client) SendRawMessage(ctx context.Context, rawJson []byte) ([]byte, *ErrorMessage) {
+	rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
+
+	modelResult := gjson.GetBytes(rawJson, "model")
+	model := modelResult.String()
+	modelName := model
+	for {
+		if c.isModelQuotaExceeded(modelName) {
+			if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
+				modelName = c.getPreviewModel(model)
+				if modelName != "" {
+					log.Debugf("Model %s is quota exceeded. Switch to preview model %s", model, modelName)
+					rawJson, _ = sjson.SetBytes(rawJson, "model", modelName)
+					continue
+				}
+			}
+			return nil, &ErrorMessage{
+				StatusCode: 429,
+				Error:      fmt.Errorf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, model),
+			}
+		}
+
+		respBody, err := c.APIRequest(ctx, "generateContent", rawJson, false)
+		if err != nil {
+			if err.StatusCode == 429 {
+				now := time.Now()
+				c.modelQuotaExceeded[modelName] = &now
+				if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
+					continue
+				}
+			}
+			return nil, err
+		}
+		delete(c.modelQuotaExceeded, modelName)
+		bodyBytes, errReadAll := io.ReadAll(respBody)
+		if errReadAll != nil {
+			return nil, &ErrorMessage{StatusCode: 500, Error: errReadAll}
+		}
+		return bodyBytes, nil
+	}
+}
+
+// SendRawMessageStream handles a single conversational turn, including tool calls.
+func (c *Client) SendRawMessageStream(ctx context.Context, rawJson []byte) (<-chan []byte, <-chan *ErrorMessage) {
+	dataTag := []byte("data: ")
+	errChan := make(chan *ErrorMessage)
+	dataChan := make(chan []byte)
+	go func() {
+		defer close(errChan)
+		defer close(dataChan)
+
+		rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
+
+		modelResult := gjson.GetBytes(rawJson, "model")
+		model := modelResult.String()
+		modelName := model
+		var stream io.ReadCloser
+		for {
+			if c.isModelQuotaExceeded(modelName) {
+				if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
+					modelName = c.getPreviewModel(model)
+					if modelName != "" {
+						log.Debugf("Model %s is quota exceeded. Switch to preview model %s", model, modelName)
+						rawJson, _ = sjson.SetBytes(rawJson, "model", modelName)
+						continue
+					}
+				}
+				errChan <- &ErrorMessage{
+					StatusCode: 429,
+					Error:      fmt.Errorf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, model),
+				}
+				return
+			}
+			var err *ErrorMessage
+			stream, err = c.APIRequest(ctx, "streamGenerateContent", rawJson, true)
+			if err != nil {
+				if err.StatusCode == 429 {
+					now := time.Now()
+					c.modelQuotaExceeded[modelName] = &now
+					if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
+						continue
+					}
+				}
+				errChan <- err
+				return
+			}
+			delete(c.modelQuotaExceeded, modelName)
+			break
+		}
+
+		scanner := bufio.NewScanner(stream)
+		for scanner.Scan() {
+			line := scanner.Bytes()
+			if bytes.HasPrefix(line, dataTag) {
+				dataChan <- line[6:]
+			}
+		}
+
+		if errScanner := scanner.Err(); errScanner != nil {
+			errChan <- &ErrorMessage{500, errScanner}
+			_ = stream.Close()
+			return
+		}
+
+		_ = stream.Close()
+	}()
+
+	return dataChan, errChan
+}
+
 func (c *Client) isModelQuotaExceeded(model string) bool {
 	if lastExceededTime, hasKey := c.modelQuotaExceeded[model]; hasKey {
 		duration := time.Now().Sub(*lastExceededTime)

internal/client/models.go

@@ -64,9 +64,10 @@ type FunctionResponse struct {
 
 // GenerateContentRequest is the top-level request structure for the streamGenerateContent endpoint.
 type GenerateContentRequest struct {
-	Contents         []Content         `json:"contents"`
-	Tools            []ToolDeclaration `json:"tools,omitempty"`
-	GenerationConfig `json:"generationConfig"`
+	SystemInstruction *Content          `json:"systemInstruction,omitempty"`
+	Contents          []Content         `json:"contents"`
+	Tools             []ToolDeclaration `json:"tools,omitempty"`
+	GenerationConfig  `json:"generationConfig"`
 }
 
 // GenerationConfig defines parameters that control the model's generation behavior.