Add ClaudeMessages handler for SSE-compatible chat completions

- Introduced `ClaudeMessages` to handle Claude-compatible streaming chat completions.
- Implemented client rotation, quota management, and dynamic model name mapping for better load balancing and resource utilization.
- Enhanced response streaming with real-time chunking and Claude format conversion.
- Added error handling for quota exhaustion, client disconnections, and backend failures.

README.md

@@ -1,15 +1,17 @@
 # CLI Proxy API
 
-A proxy server that provides an OpenAI-compatible API interface for CLI. This allows you to use CLI models with tools and libraries designed for the OpenAI API.
+A proxy server that provides an OpenAI/Gemini/Claude compatible API interface for CLI. This allows you to use CLI models with tools and libraries designed for the OpenAI/Gemini/Claude API.
 
 ## Features
 
-- OpenAI-compatible API endpoints for CLI models
+- OpenAI/Gemini/Claude compatible API endpoints for CLI models
 - Support for both streaming and non-streaming responses
 - Function calling/tools support
 - Multimodal input support (text and images)
 - Multiple account support with load balancing
 - Simple CLI authentication flow
+- Support for Generative Language API Key
+- Support Gemini CLI with multiple account load balancing
 
 ## Installation
 
@@ -134,7 +136,7 @@ console.log(response.choices[0].message.content);
 
 - gemini-2.5-pro
 - gemini-2.5-flash
-- And various preview versions
+- And it automates switching to various preview versions
 
 ## Configuration
 
@@ -146,13 +148,14 @@ The server uses a YAML configuration file (`config.yaml`) located in the project
 
 ### Configuration Options
 
-| Parameter   | Type     | Default            | Description                                                                                  |
-|-------------|----------|--------------------|----------------------------------------------------------------------------------------------|
-| `port`      | integer  | 8317               | The port number on which the server will listen                                              |
-| `auth_dir`  | string   | "~/.cli-proxy-api" | Directory where authentication tokens are stored. Supports using `~` for home directory      |
-| `proxy-url` | string   | ""                 | Proxy url, support socks5/http/https protocol, example: socks5://user:pass@192.168.1.1:1080/ |
-| `debug`     | boolean  | false              | Enable debug mode for verbose logging                                                        |
-| `api_keys`  | string[] | []                 | List of API keys that can be used to authenticate requests                                   |
+| Parameter                     | Type     | Default            | Description                                                                                  |
+|-------------------------------|----------|--------------------|----------------------------------------------------------------------------------------------|
+| `port`                        | integer  | 8317               | The port number on which the server will listen                                              |
+| `auth-dir`                    | string   | "~/.cli-proxy-api" | Directory where authentication tokens are stored. Supports using `~` for home directory      |
+| `proxy-url`                   | string   | ""                 | Proxy url, support socks5/http/https protocol, example: socks5://user:pass@192.168.1.1:1080/ |
+| `debug`                       | boolean  | false              | Enable debug mode for verbose logging                                                        |
+| `api-keys`                    | string[] | []                 | List of API keys that can be used to authenticate requests                                   |
+| `generative-language-api-key` | string[] | []                 | List of Generative Language API keys                                                         |
 
 ### Example Configuration File
 
@@ -161,29 +164,54 @@ The server uses a YAML configuration file (`config.yaml`) located in the project
 port: 8317
 
 # Authentication directory (supports ~ for home directory)
-auth_dir: "~/.cli-proxy-api"
+auth-dir: "~/.cli-proxy-api"
 
 # Enable debug logging
 debug: false
 
 # API keys for authentication
-api_keys:
+api-keys:
   - "your-api-key-1"
   - "your-api-key-2"
+
+# API keys for official Generative Language API
+generative-language-api-key:
+  - "AIzaSy...01"
+  - "AIzaSy...02"
+  - "AIzaSy...03"
+  - "AIzaSy...04"
 ```
 
 ### Authentication Directory
 
-The `auth_dir` parameter specifies where authentication tokens are stored. When you run the login command, the application will create JSON files in this directory containing the authentication tokens for your Google accounts. Multiple accounts can be used for load balancing.
+The `auth-dir` parameter specifies where authentication tokens are stored. When you run the login command, the application will create JSON files in this directory containing the authentication tokens for your Google accounts. Multiple accounts can be used for load balancing.
 
 ### API Keys
 
-The `api_keys` parameter allows you to define a list of API keys that can be used to authenticate requests to your proxy server. When making requests to the API, you can include one of these keys in the `Authorization` header:
+The `api-keys` parameter allows you to define a list of API keys that can be used to authenticate requests to your proxy server. When making requests to the API, you can include one of these keys in the `Authorization` header:
 
 ```
 Authorization: Bearer your-api-key-1
 ```
 
+### Official Generative Language API
+
+The `generative-language-api-key` parameter allows you to define a list of API keys that can be used to authenticate requests to the official Generative Language API.
+
+## Gemini CLI with multiple account load balancing
+
+Start CLI Proxy API server, and then set the `CODE_ASSIST_ENDPOINT` environment variable to the URL of the CLI Proxy API server.
+
+```bash
+export CODE_ASSIST_ENDPOINT="http://127.0.0.1:8317"
+```
+
+The server will relay the `loadCodeAssist`, `onboardUser`, and `countTokens` requests. And automatically load balance the text generation requests between the multiple accounts.
+
+> [!NOTE]  
+> This feature only allows local access because I couldn't find a way to authenticate the requests.   
+> I hardcoded `127.0.0.1` into the load balancing.
+
 ## Contributing
 
 Contributions are welcome! Please feel free to submit a Pull Request.

config.yaml

@@ -7,4 +7,9 @@ quota-exceeded:
   switch-preview-model: true
 api-keys:
   - "12345"
-  - "23456"
+  - "23456"
+generative-language-api-key:
+  - "AIzaSy...01"
+  - "AIzaSy...02"
+  - "AIzaSy...03"
+  - "AIzaSy...04"

internal/api/claude-code-handlers.go

@@ -0,0 +1,234 @@
+package api
+
+import (
+	"context"
+	"fmt"
+	"github.com/gin-gonic/gin"
+	"github.com/luispater/CLIProxyAPI/internal/api/translator"
+	"github.com/luispater/CLIProxyAPI/internal/client"
+	log "github.com/sirupsen/logrus"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// ClaudeMessages handles Claude-compatible streaming chat completions.
+// This function implements a sophisticated client rotation and quota management system
+// to ensure high availability and optimal resource utilization across multiple backend clients.
+func (h *APIHandlers) ClaudeMessages(c *gin.Context) {
+	// Extract raw JSON data from the incoming request
+	rawJson, err := c.GetRawData()
+	// If data retrieval fails, return a 400 Bad Request error.
+	if err != nil {
+		c.JSON(http.StatusBadRequest, ErrorResponse{
+			Error: ErrorDetail{
+				Message: fmt.Sprintf("Invalid request: %v", err),
+				Type:    "invalid_request_error",
+			},
+		})
+		return
+	}
+
+	// Set up Server-Sent Events (SSE) headers for streaming response
+	// These headers are essential for maintaining a persistent connection
+	// and enabling real-time streaming of chat completions
+	c.Header("Content-Type", "text/event-stream")
+	c.Header("Cache-Control", "no-cache")
+	c.Header("Connection", "keep-alive")
+	c.Header("Access-Control-Allow-Origin", "*")
+
+	// Get the http.Flusher interface to manually flush the response.
+	// This is crucial for streaming as it allows immediate sending of data chunks
+	flusher, ok := c.Writer.(http.Flusher)
+	if !ok {
+		c.JSON(http.StatusInternalServerError, ErrorResponse{
+			Error: ErrorDetail{
+				Message: "Streaming not supported",
+				Type:    "server_error",
+			},
+		})
+		return
+	}
+
+	// Parse and prepare the Claude request, extracting model name, system instructions,
+	// conversation contents, and available tools from the raw JSON
+	modelName, systemInstruction, contents, tools := translator.PrepareClaudeRequest(rawJson)
+
+	// Map Claude model names to corresponding Gemini models
+	// This allows the proxy to handle Claude API calls using Gemini backend
+	if modelName == "claude-sonnet-4-20250514" {
+		modelName = "gemini-2.5-pro"
+	} else if modelName == "claude-3-5-haiku-20241022" {
+		modelName = "gemini-2.5-flash"
+	}
+
+	// Create a cancellable context for the backend client request
+	// This allows proper cleanup and cancellation of ongoing requests
+	cliCtx, cliCancel := context.WithCancel(context.Background())
+	var cliClient *client.Client
+	defer func() {
+		// Ensure the client's mutex is unlocked on function exit.
+		// This prevents deadlocks and ensures proper resource cleanup
+		if cliClient != nil {
+			cliClient.RequestMutex.Unlock()
+		}
+	}()
+
+	// Main client rotation loop with quota management
+	// This loop implements a sophisticated load balancing and failover mechanism
+outLoop:
+	for {
+		// Thread-safe client index rotation to distribute load evenly
+		// This ensures fair usage across all available clients
+		mutex.Lock()
+		startIndex := lastUsedClientIndex
+		currentIndex := (startIndex + 1) % len(h.cliClients)
+		lastUsedClientIndex = currentIndex
+		mutex.Unlock()
+
+		// Build a list of available clients, starting from the next client in rotation
+		// This implements round-robin load balancing while filtering out quota-exceeded clients
+		reorderedClients := make([]*client.Client, 0)
+		for i := 0; i < len(h.cliClients); i++ {
+			cliClient = h.cliClients[(startIndex+1+i)%len(h.cliClients)]
+
+			// Skip clients that have exceeded their quota for the requested model
+			if cliClient.IsModelQuotaExceeded(modelName) {
+				// Log different messages based on authentication method (API key vs account)
+				if cliClient.GetGenerativeLanguageAPIKey() == "" {
+					log.Debugf("Model %s is quota exceeded for account %s, project id: %s", modelName, cliClient.GetEmail(), cliClient.GetProjectID())
+				} else {
+					log.Debugf("Model %s is quota exceeded for generative language API Key: %s", modelName, cliClient.GetGenerativeLanguageAPIKey())
+				}
+
+				cliClient = nil
+				continue
+			}
+			reorderedClients = append(reorderedClients, cliClient)
+		}
+
+		// If all clients have exceeded quota, return a 429 Too Many Requests error
+		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
+		}
+
+		// Attempt to acquire a lock on an available client using non-blocking TryLock
+		// This prevents blocking when a client is busy with another request
+		locked := false
+		for i := 0; i < len(reorderedClients); i++ {
+			cliClient = reorderedClients[i]
+			if cliClient.RequestMutex.TryLock() {
+				locked = true
+				break
+			}
+		}
+
+		// If no client is immediately available, fall back to blocking on the first client
+		// This ensures the request will eventually be processed
+		if !locked {
+			cliClient = h.cliClients[0]
+			cliClient.RequestMutex.Lock()
+		}
+
+		// Determine the authentication method being used by the selected client
+		// This affects how responses are formatted and logged
+		isGlAPIKey := false
+		if glAPIKey := cliClient.GetGenerativeLanguageAPIKey(); glAPIKey != "" {
+			log.Debugf("Request use generative language API Key: %s", glAPIKey)
+			isGlAPIKey = true
+		} else {
+			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
+		}
+		// Initiate streaming communication with the backend client
+		// This returns two channels: one for response chunks and one for errors
+
+		includeThoughts := false
+		if userAgent, hasKey := c.Request.Header["User-Agent"]; hasKey {
+			includeThoughts = !strings.Contains(userAgent[0], "claude-cli")
+		}
+
+		respChan, errChan := cliClient.SendMessageStream(cliCtx, rawJson, modelName, systemInstruction, contents, tools, includeThoughts)
+
+		// Track response state for proper Claude format conversion
+		hasFirstResponse := false
+		responseType := 0
+		responseIndex := 0
+
+		// Main streaming loop - handles multiple concurrent events using Go channels
+		// This select statement manages four different types of events simultaneously
+		for {
+			select {
+			// Case 1: Handle client disconnection
+			// Detects when the HTTP client has disconnected and cleans up resources
+			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 to prevent resource leaks
+					return
+				}
+
+			// Case 2: Process incoming response chunks from the backend
+			// This handles the actual streaming data from the AI model
+			case chunk, okStream := <-respChan:
+				if !okStream {
+					// Stream has ended - send the final message_stop event
+					// This follows the Claude API specification for stream termination
+					_, _ = c.Writer.Write([]byte(`event: message_stop`))
+					_, _ = c.Writer.Write([]byte("\n"))
+					_, _ = c.Writer.Write([]byte(`data: {"type":"message_stop"}`))
+					_, _ = c.Writer.Write([]byte("\n\n\n"))
+
+					flusher.Flush()
+					cliCancel()
+					return
+				} else {
+					// Convert the backend response to Claude-compatible format
+					// This translation layer ensures API compatibility
+					claudeFormat := translator.ConvertCliToClaude(chunk, isGlAPIKey, hasFirstResponse, &responseType, &responseIndex)
+					if claudeFormat != "" {
+						_, _ = c.Writer.Write([]byte(claudeFormat))
+						flusher.Flush() // Immediately send the chunk to the client
+					}
+					hasFirstResponse = true
+				}
+
+			// Case 3: Handle errors from the backend
+			// This manages various error conditions and implements retry logic
+			case errInfo, okError := <-errChan:
+				if okError {
+					// Special handling for quota exceeded errors
+					// If configured, attempt to switch to a different project/client
+					if errInfo.StatusCode == 429 && h.cfg.QuotaExceeded.SwitchProject {
+						continue outLoop // Restart the client selection process
+					} else {
+						// Forward other errors directly to the client
+						c.Status(errInfo.StatusCode)
+						_, _ = fmt.Fprint(c.Writer, errInfo.Error.Error())
+						flusher.Flush()
+						cliCancel()
+					}
+					return
+				}
+
+			// Case 4: Send periodic keep-alive signals
+			// Prevents connection timeouts during long-running requests
+			case <-time.After(500 * time.Millisecond):
+				if hasFirstResponse {
+					// Send a ping event to maintain the connection
+					// This is especially important for slow AI model responses
+					output := "event: ping\n"
+					output = output + `data: {"type": "ping"}`
+					output = output + "\n\n\n"
+					_, _ = c.Writer.Write([]byte(output))
+
+					flusher.Flush()
+				}
+			}
+		}
+	}
+
+}

internal/api/cli-handlers.go

@@ -0,0 +1,239 @@
+package api
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"github.com/gin-gonic/gin"
+	"github.com/luispater/CLIProxyAPI/internal/client"
+	"github.com/luispater/CLIProxyAPI/internal/util"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+)
+
+func (h *APIHandlers) CLIHandler(c *gin.Context) {
+	if !strings.HasPrefix(c.Request.RemoteAddr, "127.0.0.1:") {
+		c.JSON(http.StatusForbidden, ErrorResponse{
+			Error: ErrorDetail{
+				Message: "CLI reply only allow local access",
+				Type:    "forbidden",
+			},
+		})
+		return
+	}
+
+	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
+		}
+
+		httpClient, err := util.SetProxy(h.cfg, &http.Client{})
+		if err != nil {
+			log.Fatalf("set proxy failed: %v", err)
+		}
+
+		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 {
+		var errorResponse *client.ErrorMessage
+		cliClient, errorResponse = h.getClient(modelName)
+		if errorResponse != nil {
+			c.Status(errorResponse.StatusCode)
+			_, _ = fmt.Fprint(c.Writer, errorResponse.Error)
+			flusher.Flush()
+			cliCancel()
+			return
+		}
+
+		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 {
+		var errorResponse *client.ErrorMessage
+		cliClient, errorResponse = h.getClient(modelName)
+		if errorResponse != nil {
+			c.Status(errorResponse.StatusCode)
+			_, _ = fmt.Fprint(c.Writer, errorResponse.Error)
+			cliCancel()
+			return
+		}
+
+		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/gemini-handlers.go

@@ -0,0 +1,242 @@
+package api
+
+import (
+	"context"
+	"fmt"
+	"github.com/gin-gonic/gin"
+	"github.com/luispater/CLIProxyAPI/internal/api/translator"
+	"github.com/luispater/CLIProxyAPI/internal/client"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+	"net/http"
+	"strings"
+	"time"
+)
+
+func (h *APIHandlers) GeminiHandler(c *gin.Context) {
+	var person struct {
+		Action string `uri:"action" binding:"required"`
+	}
+	if err := c.ShouldBindUri(&person); err != nil {
+		c.JSON(http.StatusBadRequest, ErrorResponse{
+			Error: ErrorDetail{
+				Message: fmt.Sprintf("Invalid request: %v", err),
+				Type:    "invalid_request_error",
+			},
+		})
+		return
+	}
+	action := strings.Split(person.Action, ":")
+	if len(action) != 2 {
+		c.JSON(http.StatusNotFound, ErrorResponse{
+			Error: ErrorDetail{
+				Message: fmt.Sprintf("%s not found.", c.Request.URL.Path),
+				Type:    "invalid_request_error",
+			},
+		})
+		return
+	}
+
+	modelName := action[0]
+	method := action[1]
+	rawJson, _ := c.GetRawData()
+	rawJson, _ = sjson.SetBytes(rawJson, "model", []byte(modelName))
+
+	if method == "generateContent" {
+		h.geminiGenerateContent(c, rawJson)
+	} else if method == "streamGenerateContent" {
+		h.geminiStreamGenerateContent(c, rawJson)
+	}
+}
+
+func (h *APIHandlers) geminiStreamGenerateContent(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 {
+		var errorResponse *client.ErrorMessage
+		cliClient, errorResponse = h.getClient(modelName)
+		if errorResponse != nil {
+			c.Status(errorResponse.StatusCode)
+			_, _ = fmt.Fprint(c.Writer, errorResponse.Error)
+			flusher.Flush()
+			cliCancel()
+			return
+		}
+
+		template := `{"project":"","request":{},"model":""}`
+		template, _ = sjson.SetRaw(template, "request", string(rawJson))
+		template, _ = sjson.Set(template, "model", gjson.Get(template, "request.model").String())
+		template, _ = sjson.Delete(template, "request.model")
+
+		template, errFixCLIToolResponse := translator.FixCLIToolResponse(template)
+		if errFixCLIToolResponse != nil {
+			c.JSON(http.StatusInternalServerError, ErrorResponse{
+				Error: ErrorDetail{
+					Message: errFixCLIToolResponse.Error(),
+					Type:    "server_error",
+				},
+			})
+			cliCancel()
+			return
+		}
+
+		systemInstructionResult := gjson.Get(template, "request.system_instruction")
+		if systemInstructionResult.Exists() {
+			template, _ = sjson.SetRaw(template, "request.systemInstruction", systemInstructionResult.Raw)
+			template, _ = sjson.Delete(template, "request.system_instruction")
+		}
+		rawJson = []byte(template)
+
+		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)
+		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 {
+					if cliClient.GetGenerativeLanguageAPIKey() == "" {
+						responseResult := gjson.GetBytes(chunk, "response")
+						if responseResult.Exists() {
+							chunk = []byte(responseResult.Raw)
+						}
+					}
+					_, _ = 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):
+			}
+		}
+	}
+}
+
+func (h *APIHandlers) geminiGenerateContent(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 {
+		var errorResponse *client.ErrorMessage
+		cliClient, errorResponse = h.getClient(modelName)
+		if errorResponse != nil {
+			c.Status(errorResponse.StatusCode)
+			_, _ = fmt.Fprint(c.Writer, errorResponse.Error)
+			cliCancel()
+			return
+		}
+
+		template := `{"project":"","request":{},"model":""}`
+		template, _ = sjson.SetRaw(template, "request", string(rawJson))
+		template, _ = sjson.Set(template, "model", gjson.Get(template, "request.model").String())
+		template, _ = sjson.Delete(template, "request.model")
+
+		template, errFixCLIToolResponse := translator.FixCLIToolResponse(template)
+		if errFixCLIToolResponse != nil {
+			c.JSON(http.StatusInternalServerError, ErrorResponse{
+				Error: ErrorDetail{
+					Message: errFixCLIToolResponse.Error(),
+					Type:    "server_error",
+				},
+			})
+			cliCancel()
+			return
+		}
+
+		systemInstructionResult := gjson.Get(template, "request.system_instruction")
+		if systemInstructionResult.Exists() {
+			template, _ = sjson.SetRaw(template, "request.systemInstruction", systemInstructionResult.Raw)
+			template, _ = sjson.Delete(template, "request.system_instruction")
+		}
+		rawJson = []byte(template)
+
+		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 {
+			if cliClient.GetGenerativeLanguageAPIKey() == "" {
+				responseResult := gjson.GetBytes(resp, "response")
+				if responseResult.Exists() {
+					resp = []byte(responseResult.Raw)
+				}
+			}
+			_, _ = c.Writer.Write(resp)
+			cliCancel()
+			break
+		}
+	}
+}

internal/api/handlers.go

@@ -41,46 +41,6 @@ func NewAPIHandlers(cliClients []*client.Client, cfg *config.Config) *APIHandler
 func (h *APIHandlers) Models(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{
 		"data": []map[string]any{
-			{
-				"id":                    "gemini-2.5-pro-preview-05-06",
-				"object":                "model",
-				"version":               "2.5-preview-05-06",
-				"name":                  "Gemini 2.5 Pro Preview 05-06",
-				"description":           "Preview release (May 6th, 2025) of Gemini 2.5 Pro",
-				"context_length":        1048576,
-				"max_completion_tokens": 65536,
-				"supported_parameters": []string{
-					"tools",
-					"temperature",
-					"top_p",
-					"top_k",
-				},
-				"temperature":    1,
-				"topP":           0.95,
-				"topK":           64,
-				"maxTemperature": 2,
-				"thinking":       true,
-			},
-			{
-				"id":                    "gemini-2.5-pro-preview-06-05",
-				"object":                "model",
-				"version":               "2.5-preview-06-05",
-				"name":                  "Gemini 2.5 Pro Preview 06-05",
-				"description":           "Preview release (June 5th, 2025) of Gemini 2.5 Pro",
-				"context_length":        1048576,
-				"max_completion_tokens": 65536,
-				"supported_parameters": []string{
-					"tools",
-					"temperature",
-					"top_p",
-					"top_k",
-				},
-				"temperature":    1,
-				"topP":           0.95,
-				"topK":           64,
-				"maxTemperature": 2,
-				"thinking":       true,
-			},
 			{
 				"id":                    "gemini-2.5-pro",
 				"object":                "model",
@@ -101,46 +61,6 @@ func (h *APIHandlers) Models(c *gin.Context) {
 				"maxTemperature": 2,
 				"thinking":       true,
 			},
-			{
-				"id":                    "gemini-2.5-flash-preview-04-17",
-				"object":                "model",
-				"version":               "2.5-preview-04-17",
-				"name":                  "Gemini 2.5 Flash Preview 04-17",
-				"description":           "Preview release (April 17th, 2025) of Gemini 2.5 Flash",
-				"context_length":        1048576,
-				"max_completion_tokens": 65536,
-				"supported_parameters": []string{
-					"tools",
-					"temperature",
-					"top_p",
-					"top_k",
-				},
-				"temperature":    1,
-				"topP":           0.95,
-				"topK":           64,
-				"maxTemperature": 2,
-				"thinking":       true,
-			},
-			{
-				"id":                    "gemini-2.5-flash-preview-05-20",
-				"object":                "model",
-				"version":               "2.5-preview-05-20",
-				"name":                  "Gemini 2.5 Flash Preview 05-20",
-				"description":           "Preview release (April 17th, 2025) of Gemini 2.5 Flash",
-				"context_length":        1048576,
-				"max_completion_tokens": 65536,
-				"supported_parameters": []string{
-					"tools",
-					"temperature",
-					"top_p",
-					"top_k",
-				},
-				"temperature":    1,
-				"topP":           0.95,
-				"topK":           64,
-				"maxTemperature": 2,
-				"thinking":       true,
-			},
 			{
 				"id":                    "gemini-2.5-flash",
 				"object":                "model",
@@ -165,6 +85,48 @@ func (h *APIHandlers) Models(c *gin.Context) {
 	})
 }
 
+func (h *APIHandlers) getClient(modelName string) (*client.Client, *client.ErrorMessage) {
+	var cliClient *client.Client
+
+	// 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 {
+		return nil, &client.ErrorMessage{StatusCode: 429, Error: fmt.Errorf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, modelName)}
+	}
+
+	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()
+	}
+
+	return cliClient, nil
+}
+
 // ChatCompletions handles the /v1/chat/completions endpoint.
 // It determines whether the request is for a streaming or non-streaming response
 // and calls the appropriate handler.
@@ -196,19 +158,7 @@ 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() {
@@ -218,64 +168,37 @@ func (h *APIHandlers) handleNonStreamingResponse(c *gin.Context, rawJson []byte)
 	}()
 
 	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()
+		var errorResponse *client.ErrorMessage
+		cliClient, errorResponse = h.getClient(modelName)
+		if errorResponse != nil {
+			c.Status(errorResponse.StatusCode)
+			_, _ = fmt.Fprint(c.Writer, errorResponse.Error)
 			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()
+		isGlAPIKey := false
+		if glAPIKey := cliClient.GetGenerativeLanguageAPIKey(); glAPIKey != "" {
+			log.Debugf("Request use generative language API Key: %s", glAPIKey)
+			isGlAPIKey = true
+		} else {
+			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
 		}
 
-		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)
+			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
@@ -303,7 +226,7 @@ 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() {
@@ -315,49 +238,25 @@ func (h *APIHandlers) handleStreamingResponse(c *gin.Context, rawJson []byte) {
 
 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))
+		var errorResponse *client.ErrorMessage
+		cliClient, errorResponse = h.getClient(modelName)
+		if errorResponse != nil {
+			c.Status(errorResponse.StatusCode)
+			_, _ = fmt.Fprint(c.Writer, errorResponse.Error)
 			flusher.Flush()
 			cliCancel()
 			return
 		}
 
-		locked := false
-		for i := 0; i < len(reorderedClients); i++ {
-			cliClient = reorderedClients[i]
-			if cliClient.RequestMutex.TryLock() {
-				locked = true
-				break
-			}
+		isGlAPIKey := false
+		if glAPIKey := cliClient.GetGenerativeLanguageAPIKey(); glAPIKey != "" {
+			log.Debugf("Request use generative language API Key: %s", glAPIKey)
+			isGlAPIKey = true
+		} else {
+			log.Debugf("Request use account: %s, project id: %s", cliClient.GetEmail(), cliClient.GetProjectID())
 		}
-		if !locked {
-			cliClient = h.cliClients[0]
-			cliClient.RequestMutex.Lock()
-		}
-
-		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 {
@@ -379,7 +278,7 @@ outLoop:
 				} else {
 					// Convert the chunk to OpenAI format and send it to the client.
 					hasFirstResponse = true
-					openAIFormat := translator.ConvertCliToOpenAI(chunk)
+					openAIFormat := translator.ConvertCliToOpenAI(chunk, time.Now().Unix(), isGlAPIKey)
 					if openAIFormat != "" {
 						_, _ = fmt.Fprintf(c.Writer, "data: %s\n\n", openAIFormat)
 						flusher.Flush()

internal/api/server.go

@@ -68,6 +68,15 @@ func (s *Server) setupRoutes() {
 	{
 		v1.GET("/models", s.handlers.Models)
 		v1.POST("/chat/completions", s.handlers.ChatCompletions)
+		v1.POST("/messages", s.handlers.ClaudeMessages)
+	}
+
+	// Gemini compatible API routes
+	v1beta := s.engine.Group("/v1beta")
+	v1beta.Use(AuthMiddleware(s.cfg))
+	{
+		v1beta.GET("/models", s.handlers.Models)
+		v1beta.POST("/models/:action", s.handlers.GeminiHandler)
 	}
 
 	// Root endpoint
@@ -81,6 +90,8 @@ func (s *Server) setupRoutes() {
 			},
 		})
 	})
+	s.engine.POST("/v1internal:method", s.handlers.CLIHandler)
+
 }
 
 // Start begins listening for and serving HTTP requests.
@@ -138,7 +149,9 @@ func AuthMiddleware(cfg *config.Config) gin.HandlerFunc {
 
 		// Get the Authorization header
 		authHeader := c.GetHeader("Authorization")
-		if authHeader == "" {
+		authHeaderGoogle := c.GetHeader("X-Goog-Api-Key")
+		authHeaderAnthropic := c.GetHeader("X-Api-Key")
+		if authHeader == "" && authHeaderGoogle == "" && authHeaderAnthropic == "" {
 			c.AbortWithStatusJSON(http.StatusUnauthorized, gin.H{
 				"error": "Missing API key",
 			})
@@ -157,7 +170,7 @@ func AuthMiddleware(cfg *config.Config) gin.HandlerFunc {
 		// Find the API key in the in-memory list
 		var foundKey string
 		for i := range cfg.ApiKeys {
-			if cfg.ApiKeys[i] == apiKey {
+			if cfg.ApiKeys[i] == apiKey || cfg.ApiKeys[i] == authHeaderGoogle || cfg.ApiKeys[i] == authHeaderAnthropic {
 				foundKey = cfg.ApiKeys[i]
 				break
 			}

internal/api/translator/request.go

@@ -1,7 +1,10 @@
 package translator
 
 import (
+	"bytes"
 	"encoding/json"
+	"fmt"
+	"github.com/tidwall/sjson"
 	"strings"
 
 	"github.com/luispater/CLIProxyAPI/internal/client"
@@ -12,7 +15,7 @@ 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")
@@ -20,9 +23,52 @@ func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDecl
 		modelName = modelResult.String()
 	}
 
-	// Process the array of messages.
+	// Initialize data structures for processing conversation messages
+	// contents: stores the processed conversation history
+	// systemInstruction: stores system-level instructions separate from conversation
 	contents := make([]client.Content, 0)
+	var systemInstruction *client.Content
 	messagesResult := gjson.GetBytes(rawJson, "messages")
+
+	// Pre-process tool responses to create a lookup map
+	// This first pass collects all tool responses so they can be matched with their corresponding calls
+	toolItems := make(map[string]*client.FunctionResponse)
+	if messagesResult.IsArray() {
+		messagesResults := messagesResult.Array()
+		for i := 0; i < len(messagesResults); i++ {
+			messageResult := messagesResults[i]
+			roleResult := messageResult.Get("role")
+			if roleResult.Type != gjson.String {
+				continue
+			}
+			contentResult := messageResult.Get("content")
+
+			// Extract tool responses for later matching with function calls
+			if roleResult.String() == "tool" {
+				toolCallID := messageResult.Get("tool_call_id").String()
+				if toolCallID != "" {
+					var responseData string
+					// Handle both string and object-based tool response formats
+					if contentResult.Type == gjson.String {
+						responseData = contentResult.String()
+					} else if contentResult.IsObject() && contentResult.Get("type").String() == "text" {
+						responseData = contentResult.Get("text").String()
+					}
+
+					// Clean up tool call ID by removing timestamp suffix
+					// This normalizes IDs for consistent matching between calls and responses
+					toolCallIDs := strings.Split(toolCallID, "-")
+					strings.Join(toolCallIDs, "-")
+					newToolCallID := strings.Join(toolCallIDs[:len(toolCallIDs)-1], "-")
+
+					// Create function response object with normalized ID and response data
+					functionResponse := client.FunctionResponse{Name: newToolCallID, Response: map[string]interface{}{"result": responseData}}
+					toolItems[toolCallID] = &functionResponse
+				}
+			}
+		}
+	}
+
 	if messagesResult.IsArray() {
 		messagesResults := messagesResult.Array()
 		for i := 0; i < len(messagesResults); i++ {
@@ -37,13 +83,11 @@ 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.
@@ -92,46 +136,63 @@ func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDecl
 					}
 					contents = append(contents, client.Content{Role: "user", Parts: parts})
 				}
-			// Assistant messages can contain text or tool calls.
+			// Assistant messages can contain text responses or tool calls
+			// In the internal format, assistant messages are converted to "model" role
 			case "assistant":
 				if contentResult.Type == gjson.String {
+					// Simple text response from the assistant
 					contents = append(contents, client.Content{Role: "model", Parts: []client.Part{{Text: contentResult.String()}}})
 				} else if !contentResult.Exists() || contentResult.Type == gjson.Null {
-					// Handle tool calls made by the assistant.
+					// Handle complex tool calls made by the assistant
+					// This processes function calls and matches them with their responses
+					functionIDs := make([]string, 0)
 					toolCallsResult := messageResult.Get("tool_calls")
 					if toolCallsResult.IsArray() {
+						parts := make([]client.Part, 0)
 						tcsResult := toolCallsResult.Array()
+
+						// Process each tool call in the assistant's message
 						for j := 0; j < len(tcsResult); j++ {
 							tcResult := tcsResult[j]
+
+							// Extract function call details
+							functionID := tcResult.Get("id").String()
+							functionIDs = append(functionIDs, functionID)
+
 							functionName := tcResult.Get("function.name").String()
 							functionArgs := tcResult.Get("function.arguments").String()
+
+							// Parse function arguments from JSON string to map
 							var args map[string]any
 							if err := json.Unmarshal([]byte(functionArgs), &args); err == nil {
-								contents = append(contents, client.Content{
-									Role: "model", Parts: []client.Part{{
-										FunctionCall: &client.FunctionCall{
-											Name: functionName,
-											Args: args,
-										},
-									}},
+								parts = append(parts, client.Part{
+									FunctionCall: &client.FunctionCall{
+										Name: functionName,
+										Args: args,
+									},
 								})
 							}
 						}
+
+						// Add the model's function calls to the conversation
+						if len(parts) > 0 {
+							contents = append(contents, client.Content{
+								Role: "model", Parts: parts,
+							})
+
+							// Create a separate tool response message with the collected responses
+							// This matches function calls with their corresponding responses
+							toolParts := make([]client.Part, 0)
+							for _, functionID := range functionIDs {
+								if functionResponse, ok := toolItems[functionID]; ok {
+									toolParts = append(toolParts, client.Part{FunctionResponse: functionResponse})
+								}
+							}
+							// Add the tool responses as a separate message in the conversation
+							contents = append(contents, client.Content{Role: "tool", Parts: toolParts})
+						}
 					}
 				}
-			// Tool messages contain the output of a tool call.
-			case "tool":
-				toolCallID := messageResult.Get("tool_call_id").String()
-				if toolCallID != "" {
-					var responseData string
-					if contentResult.Type == gjson.String {
-						responseData = contentResult.String()
-					} else if contentResult.IsObject() && contentResult.Get("type").String() == "text" {
-						responseData = contentResult.Get("text").String()
-					}
-					functionResponse := client.FunctionResponse{Name: toolCallID, Response: map[string]interface{}{"result": responseData}}
-					contents = append(contents, client.Content{Role: "tool", Parts: []client.Part{{FunctionResponse: &functionResponse}}})
-				}
 			}
 		}
 	}
@@ -159,5 +220,325 @@ func PrepareRequest(rawJson []byte) (string, []client.Content, []client.ToolDecl
 		tools = make([]client.ToolDeclaration, 0)
 	}
 
-	return modelName, contents, tools
+	return modelName, systemInstruction, contents, tools
+}
+
+// FunctionCallGroup represents a group of function calls and their responses
+type FunctionCallGroup struct {
+	ModelContent    map[string]interface{}
+	FunctionCalls   []gjson.Result
+	ResponsesNeeded int
+}
+
+// FixCLIToolResponse performs sophisticated tool response format conversion and grouping.
+// This function transforms the CLI tool response format by intelligently grouping function calls
+// with their corresponding responses, ensuring proper conversation flow and API compatibility.
+// It converts from a linear format (1.json) to a grouped format (2.json) where function calls
+// and their responses are properly associated and structured.
+func FixCLIToolResponse(input string) (string, error) {
+	// Parse the input JSON to extract the conversation structure
+	parsed := gjson.Parse(input)
+
+	// Extract the contents array which contains the conversation messages
+	contents := parsed.Get("request.contents")
+	if !contents.Exists() {
+		return input, fmt.Errorf("contents not found in input")
+	}
+
+	// Initialize data structures for processing and grouping
+	var newContents []interface{}          // Final processed contents array
+	var pendingGroups []*FunctionCallGroup // Groups awaiting completion with responses
+	var collectedResponses []gjson.Result  // Standalone responses to be matched
+
+	// Process each content object in the conversation
+	// This iterates through messages and groups function calls with their responses
+	contents.ForEach(func(key, value gjson.Result) bool {
+		role := value.Get("role").String()
+		parts := value.Get("parts")
+
+		// Check if this content has function responses
+		var responsePartsInThisContent []gjson.Result
+		parts.ForEach(func(_, part gjson.Result) bool {
+			if part.Get("functionResponse").Exists() {
+				responsePartsInThisContent = append(responsePartsInThisContent, part)
+			}
+			return true
+		})
+
+		// If this content has function responses, collect them
+		if len(responsePartsInThisContent) > 0 {
+			collectedResponses = append(collectedResponses, responsePartsInThisContent...)
+
+			// Check if any pending groups can be satisfied
+			for i := len(pendingGroups) - 1; i >= 0; i-- {
+				group := pendingGroups[i]
+				if len(collectedResponses) >= group.ResponsesNeeded {
+					// Take the needed responses for this group
+					groupResponses := collectedResponses[:group.ResponsesNeeded]
+					collectedResponses = collectedResponses[group.ResponsesNeeded:]
+
+					// Create merged function response content
+					var responseParts []interface{}
+					for _, response := range groupResponses {
+						var responseMap map[string]interface{}
+						errUnmarshal := json.Unmarshal([]byte(response.Raw), &responseMap)
+						if errUnmarshal != nil {
+							log.Warnf("failed to unmarshal function response: %v\n", errUnmarshal)
+							continue
+						}
+						responseParts = append(responseParts, responseMap)
+					}
+
+					if len(responseParts) > 0 {
+						functionResponseContent := map[string]interface{}{
+							"parts": responseParts,
+							"role":  "function",
+						}
+						newContents = append(newContents, functionResponseContent)
+					}
+
+					// Remove this group as it's been satisfied
+					pendingGroups = append(pendingGroups[:i], pendingGroups[i+1:]...)
+					break
+				}
+			}
+
+			return true // Skip adding this content, responses are merged
+		}
+
+		// If this is a model with function calls, create a new group
+		if role == "model" {
+			var functionCallsInThisModel []gjson.Result
+			parts.ForEach(func(_, part gjson.Result) bool {
+				if part.Get("functionCall").Exists() {
+					functionCallsInThisModel = append(functionCallsInThisModel, part)
+				}
+				return true
+			})
+
+			if len(functionCallsInThisModel) > 0 {
+				// Add the model content
+				var contentMap map[string]interface{}
+				errUnmarshal := json.Unmarshal([]byte(value.Raw), &contentMap)
+				if errUnmarshal != nil {
+					log.Warnf("failed to unmarshal model content: %v\n", errUnmarshal)
+					return true
+				}
+				newContents = append(newContents, contentMap)
+
+				// Create a new group for tracking responses
+				group := &FunctionCallGroup{
+					ModelContent:    contentMap,
+					FunctionCalls:   functionCallsInThisModel,
+					ResponsesNeeded: len(functionCallsInThisModel),
+				}
+				pendingGroups = append(pendingGroups, group)
+			} else {
+				// Regular model content without function calls
+				var contentMap map[string]interface{}
+				errUnmarshal := json.Unmarshal([]byte(value.Raw), &contentMap)
+				if errUnmarshal != nil {
+					log.Warnf("failed to unmarshal content: %v\n", errUnmarshal)
+					return true
+				}
+				newContents = append(newContents, contentMap)
+			}
+		} else {
+			// Non-model content (user, etc.)
+			var contentMap map[string]interface{}
+			errUnmarshal := json.Unmarshal([]byte(value.Raw), &contentMap)
+			if errUnmarshal != nil {
+				log.Warnf("failed to unmarshal content: %v\n", errUnmarshal)
+				return true
+			}
+			newContents = append(newContents, contentMap)
+		}
+
+		return true
+	})
+
+	// Handle any remaining pending groups with remaining responses
+	for _, group := range pendingGroups {
+		if len(collectedResponses) >= group.ResponsesNeeded {
+			groupResponses := collectedResponses[:group.ResponsesNeeded]
+			collectedResponses = collectedResponses[group.ResponsesNeeded:]
+
+			var responseParts []interface{}
+			for _, response := range groupResponses {
+				var responseMap map[string]interface{}
+				errUnmarshal := json.Unmarshal([]byte(response.Raw), &responseMap)
+				if errUnmarshal != nil {
+					log.Warnf("failed to unmarshal function response: %v\n", errUnmarshal)
+					continue
+				}
+				responseParts = append(responseParts, responseMap)
+			}
+
+			if len(responseParts) > 0 {
+				functionResponseContent := map[string]interface{}{
+					"parts": responseParts,
+					"role":  "function",
+				}
+				newContents = append(newContents, functionResponseContent)
+			}
+		}
+	}
+
+	// Update the original JSON with the new contents
+	result := input
+	newContentsJSON, _ := json.Marshal(newContents)
+	result, _ = sjson.Set(result, "request.contents", json.RawMessage(newContentsJSON))
+
+	return result, nil
+}
+
+func PrepareClaudeRequest(rawJson []byte) (string, *client.Content, []client.Content, []client.ToolDeclaration) {
+	var pathsToDelete []string
+	root := gjson.ParseBytes(rawJson)
+	walk(root, "", "additionalProperties", &pathsToDelete)
+	walk(root, "", "$schema", &pathsToDelete)
+
+	var err error
+	for _, p := range pathsToDelete {
+		rawJson, err = sjson.DeleteBytes(rawJson, p)
+		if err != nil {
+			continue
+		}
+	}
+	rawJson = bytes.Replace(rawJson, []byte(`"url":{"type":"string","format":"uri",`), []byte(`"url":{"type":"string",`), -1)
+
+	// log.Debug(string(rawJson))
+	modelName := "gemini-2.5-pro"
+	modelResult := gjson.GetBytes(rawJson, "model")
+	if modelResult.Type == gjson.String {
+		modelName = modelResult.String()
+	}
+
+	contents := make([]client.Content, 0)
+
+	var systemInstruction *client.Content
+
+	systemResult := gjson.GetBytes(rawJson, "system")
+	if systemResult.IsArray() {
+		systemResults := systemResult.Array()
+		systemInstruction = &client.Content{Role: "user", Parts: []client.Part{}}
+		for i := 0; i < len(systemResults); i++ {
+			systemPromptResult := systemResults[i]
+			systemTypePromptResult := systemPromptResult.Get("type")
+			if systemTypePromptResult.Type == gjson.String && systemTypePromptResult.String() == "text" {
+				systemPrompt := systemPromptResult.Get("text").String()
+				systemPart := client.Part{Text: systemPrompt}
+				systemInstruction.Parts = append(systemInstruction.Parts, systemPart)
+			}
+		}
+		if len(systemInstruction.Parts) == 0 {
+			systemInstruction = nil
+		}
+	}
+
+	messagesResult := gjson.GetBytes(rawJson, "messages")
+	if messagesResult.IsArray() {
+		messageResults := messagesResult.Array()
+		for i := 0; i < len(messageResults); i++ {
+			messageResult := messageResults[i]
+			roleResult := messageResult.Get("role")
+			if roleResult.Type != gjson.String {
+				continue
+			}
+			role := roleResult.String()
+			if role == "assistant" {
+				role = "model"
+			}
+			clientContent := client.Content{Role: role, Parts: []client.Part{}}
+
+			contentsResult := messageResult.Get("content")
+			if contentsResult.IsArray() {
+				contentResults := contentsResult.Array()
+				for j := 0; j < len(contentResults); j++ {
+					contentResult := contentResults[j]
+					contentTypeResult := contentResult.Get("type")
+					if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "text" {
+						prompt := contentResult.Get("text").String()
+						clientContent.Parts = append(clientContent.Parts, client.Part{Text: prompt})
+					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_use" {
+						functionName := contentResult.Get("name").String()
+						functionArgs := contentResult.Get("input").String()
+						var args map[string]any
+						if err = json.Unmarshal([]byte(functionArgs), &args); err == nil {
+							clientContent.Parts = append(clientContent.Parts, client.Part{
+								FunctionCall: &client.FunctionCall{
+									Name: functionName,
+									Args: args,
+								},
+							})
+						}
+					} else if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "tool_result" {
+						toolCallID := contentResult.Get("tool_use_id").String()
+						if toolCallID != "" {
+							funcName := toolCallID
+							toolCallIDs := strings.Split(toolCallID, "-")
+							if len(toolCallIDs) > 1 {
+								funcName = strings.Join(toolCallIDs[0:len(toolCallIDs)-1], "-")
+							}
+							responseData := contentResult.Get("content").String()
+							functionResponse := client.FunctionResponse{Name: funcName, Response: map[string]interface{}{"result": responseData}}
+							clientContent.Parts = append(clientContent.Parts, client.Part{FunctionResponse: &functionResponse})
+						}
+					}
+				}
+				contents = append(contents, clientContent)
+			} else if contentsResult.Type == gjson.String {
+				prompt := contentsResult.String()
+				contents = append(contents, client.Content{Role: role, Parts: []client.Part{{Text: prompt}}})
+			}
+		}
+	}
+
+	var tools []client.ToolDeclaration
+	toolsResult := gjson.GetBytes(rawJson, "tools")
+	if toolsResult.IsArray() {
+		tools = make([]client.ToolDeclaration, 1)
+		tools[0].FunctionDeclarations = make([]any, 0)
+		toolsResults := toolsResult.Array()
+		for i := 0; i < len(toolsResults); i++ {
+			toolResult := toolsResults[i]
+			inputSchemaResult := toolResult.Get("input_schema")
+			if inputSchemaResult.Exists() && inputSchemaResult.IsObject() {
+				inputSchema := inputSchemaResult.Raw
+				inputSchema, _ = sjson.Delete(inputSchema, "additionalProperties")
+				inputSchema, _ = sjson.Delete(inputSchema, "$schema")
+
+				tool, _ := sjson.Delete(toolResult.Raw, "input_schema")
+				tool, _ = sjson.SetRaw(tool, "parameters", inputSchema)
+				var toolDeclaration any
+				if err = json.Unmarshal([]byte(tool), &toolDeclaration); err == nil {
+					tools[0].FunctionDeclarations = append(tools[0].FunctionDeclarations, toolDeclaration)
+				}
+			}
+		}
+	} else {
+		tools = make([]client.ToolDeclaration, 0)
+	}
+
+	return modelName, systemInstruction, contents, tools
+}
+
+func walk(value gjson.Result, path, field string, pathsToDelete *[]string) {
+	switch value.Type {
+	case gjson.JSON:
+		value.ForEach(func(key, val gjson.Result) bool {
+			var childPath string
+			if path == "" {
+				childPath = key.String()
+			} else {
+				childPath = path + "." + key.String()
+			}
+			if key.String() == field {
+				*pathsToDelete = append(*pathsToDelete, childPath)
+			}
+			walk(val, childPath, field, pathsToDelete)
+			return true
+		})
+	case gjson.String, gjson.Number, gjson.True, gjson.False, gjson.Null:
+	}
 }

internal/api/translator/response.go

@@ -1,6 +1,8 @@
 package translator
 
 import (
+	"bytes"
+	"fmt"
 	"time"
 
 	"github.com/tidwall/gjson"
@@ -10,7 +12,11 @@ import (
 // ConvertCliToOpenAI translates a single chunk of a streaming response from the
 // backend client format to the OpenAI Server-Sent Events (SSE) format.
 // It returns an empty string if the chunk contains no useful data.
-func ConvertCliToOpenAI(rawJson []byte) string {
+func ConvertCliToOpenAI(rawJson []byte, unixTimestamp int64, isGlAPIKey bool) string {
+	if isGlAPIKey {
+		rawJson, _ = sjson.SetRawBytes(rawJson, "response", rawJson)
+	}
+
 	// Initialize the OpenAI SSE template.
 	template := `{"id":"","object":"chat.completion.chunk","created":12345,"model":"model","choices":[{"index":0,"delta":{"role":null,"content":null,"reasoning_content":null,"tool_calls":null},"finish_reason":null,"native_finish_reason":null}]}`
 
@@ -22,11 +28,12 @@ func ConvertCliToOpenAI(rawJson []byte) string {
 	// Extract and set the creation timestamp.
 	if createTimeResult := gjson.GetBytes(rawJson, "response.createTime"); createTimeResult.Exists() {
 		t, err := time.Parse(time.RFC3339Nano, createTimeResult.String())
-		unixTimestamp := time.Now().Unix()
 		if err == nil {
 			unixTimestamp = t.Unix()
 		}
 		template, _ = sjson.Set(template, "created", unixTimestamp)
+	} else {
+		template, _ = sjson.Set(template, "created", unixTimestamp)
 	}
 
 	// Extract and set the response ID.
@@ -57,32 +64,40 @@ func ConvertCliToOpenAI(rawJson []byte) string {
 	}
 
 	// Process the main content part of the response.
-	partResult := gjson.GetBytes(rawJson, "response.candidates.0.content.parts.0")
-	partTextResult := partResult.Get("text")
-	functionCallResult := partResult.Get("functionCall")
+	partsResult := gjson.GetBytes(rawJson, "response.candidates.0.content.parts")
+	if partsResult.IsArray() {
+		partResults := partsResult.Array()
+		for i := 0; i < len(partResults); i++ {
+			partResult := partResults[i]
+			partTextResult := partResult.Get("text")
+			functionCallResult := partResult.Get("functionCall")
 
-	if partTextResult.Exists() {
-		// Handle text content, distinguishing between regular content and reasoning/thoughts.
-		if partResult.Get("thought").Bool() {
-			template, _ = sjson.Set(template, "choices.0.delta.reasoning_content", partTextResult.String())
-		} else {
-			template, _ = sjson.Set(template, "choices.0.delta.content", partTextResult.String())
+			if partTextResult.Exists() {
+				// Handle text content, distinguishing between regular content and reasoning/thoughts.
+				if partResult.Get("thought").Bool() {
+					template, _ = sjson.Set(template, "choices.0.delta.reasoning_content", partTextResult.String())
+				} else {
+					template, _ = sjson.Set(template, "choices.0.delta.content", partTextResult.String())
+				}
+				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+			} else if functionCallResult.Exists() {
+				// Handle function call content.
+				toolCallsResult := gjson.Get(template, "choices.0.delta.tool_calls")
+				if !toolCallsResult.Exists() || !toolCallsResult.IsArray() {
+					template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
+				}
+
+				functionCallTemplate := `{"id": "","type": "function","function": {"name": "","arguments": ""}}`
+				fcName := functionCallResult.Get("name").String()
+				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.name", fcName)
+				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
+					functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.arguments", fcArgsResult.Raw)
+				}
+				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+				template, _ = sjson.SetRaw(template, "choices.0.message.tool_calls.-1", functionCallTemplate)
+			}
 		}
-		template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-	} else if functionCallResult.Exists() {
-		// Handle function call content.
-		functionCallTemplate := `[{"id": "","type": "function","function": {"name": "","arguments": ""}}]`
-		fcName := functionCallResult.Get("name").String()
-		functionCallTemplate, _ = sjson.Set(functionCallTemplate, "0.id", fcName)
-		functionCallTemplate, _ = sjson.Set(functionCallTemplate, "0.function.name", fcName)
-		if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
-			functionCallTemplate, _ = sjson.Set(functionCallTemplate, "0.function.arguments", fcArgsResult.Raw)
-		}
-		template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-		template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", functionCallTemplate)
-	} else {
-		// If no usable content is found, return an empty string.
-		return ""
 	}
 
 	return template
@@ -90,19 +105,25 @@ func ConvertCliToOpenAI(rawJson []byte) string {
 
 // ConvertCliToOpenAINonStream aggregates response from the backend client
 // convert a single, non-streaming OpenAI-compatible JSON response.
-func ConvertCliToOpenAINonStream(rawJson []byte) string {
+func ConvertCliToOpenAINonStream(rawJson []byte, unixTimestamp int64, isGlAPIKey bool) string {
+	if isGlAPIKey {
+		rawJson, _ = sjson.SetRawBytes(rawJson, "response", rawJson)
+	}
 	template := `{"id":"","object":"chat.completion","created":123456,"model":"model","choices":[{"index":0,"message":{"role":"assistant","content":null,"reasoning_content":null,"tool_calls":null},"finish_reason":null,"native_finish_reason":null}]}`
 	if modelVersionResult := gjson.GetBytes(rawJson, "response.modelVersion"); modelVersionResult.Exists() {
 		template, _ = sjson.Set(template, "model", modelVersionResult.String())
 	}
+
 	if createTimeResult := gjson.GetBytes(rawJson, "response.createTime"); createTimeResult.Exists() {
 		t, err := time.Parse(time.RFC3339Nano, createTimeResult.String())
-		unixTimestamp := time.Now().Unix()
 		if err == nil {
 			unixTimestamp = t.Unix()
 		}
 		template, _ = sjson.Set(template, "created", unixTimestamp)
+	} else {
+		template, _ = sjson.Set(template, "created", unixTimestamp)
 	}
+
 	if responseIdResult := gjson.GetBytes(rawJson, "response.responseId"); responseIdResult.Exists() {
 		template, _ = sjson.Set(template, "id", responseIdResult.String())
 	}
@@ -152,7 +173,7 @@ func ConvertCliToOpenAINonStream(rawJson []byte) string {
 				}
 				functionCallItemTemplate := `{"id": "","type": "function","function": {"name": "","arguments": ""}}`
 				fcName := functionCallResult.Get("name").String()
-				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", fcName)
+				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
 				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", fcName)
 				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
 					functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", fcArgsResult.Raw)
@@ -168,3 +189,194 @@ func ConvertCliToOpenAINonStream(rawJson []byte) string {
 
 	return template
 }
+
+// ConvertCliToClaude performs sophisticated streaming response format conversion.
+// This function implements a complex state machine that translates backend client responses
+// into Claude-compatible Server-Sent Events (SSE) format. It manages different response types
+// and handles state transitions between content blocks, thinking processes, and function calls.
+//
+// Response type states: 0=none, 1=content, 2=thinking, 3=function
+// The function maintains state across multiple calls to ensure proper SSE event sequencing.
+func ConvertCliToClaude(rawJson []byte, isGlAPIKey, hasFirstResponse bool, responseType, responseIndex *int) string {
+	// Normalize the response format for different API key types
+	// Generative Language API keys have a different response structure
+	if isGlAPIKey {
+		rawJson, _ = sjson.SetRawBytes(rawJson, "response", rawJson)
+	}
+
+	// Track whether tools are being used in this response chunk
+	usedTool := false
+	output := ""
+
+	// Initialize the streaming session with a message_start event
+	// This is only sent for the very first response chunk
+	if !hasFirstResponse {
+		output = "event: message_start\n"
+
+		// Create the initial message structure with default values
+		// This follows the Claude API specification for streaming message initialization
+		messageStartTemplate := `{"type": "message_start", "message": {"id": "msg_1nZdL29xx5MUA1yADyHTEsnR8uuvGzszyY", "type": "message", "role": "assistant", "content": [], "model": "claude-3-5-sonnet-20241022", "stop_reason": null, "stop_sequence": null, "usage": {"input_tokens": 0, "output_tokens": 0}}}`
+
+		// Override default values with actual response metadata if available
+		if modelVersionResult := gjson.GetBytes(rawJson, "response.modelVersion"); modelVersionResult.Exists() {
+			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.model", modelVersionResult.String())
+		}
+		if responseIdResult := gjson.GetBytes(rawJson, "response.responseId"); responseIdResult.Exists() {
+			messageStartTemplate, _ = sjson.Set(messageStartTemplate, "message.id", responseIdResult.String())
+		}
+		output = output + fmt.Sprintf("data: %s\n\n\n", messageStartTemplate)
+	}
+
+	// Process the response parts array from the backend client
+	// Each part can contain text content, thinking content, or function calls
+	partsResult := gjson.GetBytes(rawJson, "response.candidates.0.content.parts")
+	if partsResult.IsArray() {
+		partResults := partsResult.Array()
+		for i := 0; i < len(partResults); i++ {
+			partResult := partResults[i]
+
+			// Extract the different types of content from each part
+			partTextResult := partResult.Get("text")
+			functionCallResult := partResult.Get("functionCall")
+
+			// Handle text content (both regular content and thinking)
+			if partTextResult.Exists() {
+				// Process thinking content (internal reasoning)
+				if partResult.Get("thought").Bool() {
+					// Continue existing thinking block
+					if *responseType == 2 {
+						output = output + "event: content_block_delta\n"
+						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, *responseIndex), "delta.thinking", partTextResult.String())
+						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+					} else {
+						// Transition from another state to thinking
+						// First, close any existing content block
+						if *responseType != 0 {
+							if *responseType == 2 {
+								output = output + "event: content_block_delta\n"
+								output = output + fmt.Sprintf(`data: {"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":null}}`, *responseIndex)
+								output = output + "\n\n\n"
+							}
+							output = output + "event: content_block_stop\n"
+							output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, *responseIndex)
+							output = output + "\n\n\n"
+							*responseIndex++
+						}
+
+						// Start a new thinking content block
+						output = output + "event: content_block_start\n"
+						output = output + fmt.Sprintf(`data: {"type":"content_block_start","index":%d,"content_block":{"type":"thinking","thinking":""}}`, *responseIndex)
+						output = output + "\n\n\n"
+						output = output + "event: content_block_delta\n"
+						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"thinking_delta","thinking":""}}`, *responseIndex), "delta.thinking", partTextResult.String())
+						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						*responseType = 2 // Set state to thinking
+					}
+				} else {
+					// Process regular text content (user-visible output)
+					// Continue existing text block
+					if *responseType == 1 {
+						output = output + "event: content_block_delta\n"
+						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, *responseIndex), "delta.text", partTextResult.String())
+						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+					} else {
+						// Transition from another state to text content
+						// First, close any existing content block
+						if *responseType != 0 {
+							if *responseType == 2 {
+								output = output + "event: content_block_delta\n"
+								output = output + fmt.Sprintf(`data: {"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":null}}`, *responseIndex)
+								output = output + "\n\n\n"
+							}
+							output = output + "event: content_block_stop\n"
+							output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, *responseIndex)
+							output = output + "\n\n\n"
+							*responseIndex++
+						}
+
+						// Start a new text content block
+						output = output + "event: content_block_start\n"
+						output = output + fmt.Sprintf(`data: {"type":"content_block_start","index":%d,"content_block":{"type":"text","text":""}}`, *responseIndex)
+						output = output + "\n\n\n"
+						output = output + "event: content_block_delta\n"
+						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"text_delta","text":""}}`, *responseIndex), "delta.text", partTextResult.String())
+						output = output + fmt.Sprintf("data: %s\n\n\n", data)
+						*responseType = 1 // Set state to content
+					}
+				}
+			} else if functionCallResult.Exists() {
+				// Handle function/tool calls from the AI model
+				// This processes tool usage requests and formats them for Claude API compatibility
+				usedTool = true
+				fcName := functionCallResult.Get("name").String()
+
+				// Handle state transitions when switching to function calls
+				// Close any existing function call block first
+				if *responseType == 3 {
+					output = output + "event: content_block_stop\n"
+					output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, *responseIndex)
+					output = output + "\n\n\n"
+					*responseIndex++
+					*responseType = 0
+				}
+
+				// Special handling for thinking state transition
+				if *responseType == 2 {
+					output = output + "event: content_block_delta\n"
+					output = output + fmt.Sprintf(`data: {"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":null}}`, *responseIndex)
+					output = output + "\n\n\n"
+				}
+
+				// Close any other existing content block
+				if *responseType != 0 {
+					output = output + "event: content_block_stop\n"
+					output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, *responseIndex)
+					output = output + "\n\n\n"
+					*responseIndex++
+				}
+
+				// Start a new tool use content block
+				// This creates the structure for a function call in Claude format
+				output = output + "event: content_block_start\n"
+
+				// Create the tool use block with unique ID and function details
+				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, *responseIndex)
+				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d", fcName, time.Now().UnixNano()))
+				data, _ = sjson.Set(data, "content_block.name", fcName)
+				output = output + fmt.Sprintf("data: %s\n\n\n", data)
+
+				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
+					output = output + "event: content_block_delta\n"
+					data, _ = sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"input_json_delta","partial_json":""}}`, *responseIndex), "delta.partial_json", fcArgsResult.Raw)
+					output = output + fmt.Sprintf("data: %s\n\n\n", data)
+				}
+				*responseType = 3
+			}
+		}
+	}
+
+	usageResult := gjson.GetBytes(rawJson, "response.usageMetadata")
+	if usageResult.Exists() && bytes.Contains(rawJson, []byte(`"finishReason"`)) {
+		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
+			output = output + "event: content_block_stop\n"
+			output = output + fmt.Sprintf(`data: {"type":"content_block_stop","index":%d}`, *responseIndex)
+			output = output + "\n\n\n"
+
+			output = output + "event: message_delta\n"
+			output = output + `data: `
+
+			template := `{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+			if usedTool {
+				template = `{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
+			}
+
+			thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
+			template, _ = sjson.Set(template, "usage.output_tokens", candidatesTokenCountResult.Int()+thoughtsTokenCount)
+			template, _ = sjson.Set(template, "usage.input_tokens", usageResult.Get("promptTokenCount").Int())
+
+			output = output + template + "\n\n\n"
+		}
+	}
+
+	return output
+}

internal/client/client.go

@@ -27,6 +27,9 @@ const (
 	codeAssistEndpoint = "https://cloudcode-pa.googleapis.com"
 	apiVersion         = "v1internal"
 	pluginVersion      = "0.1.9"
+
+	glEndPoint   = "https://generativelanguage.googleapis.com/"
+	glApiVersion = "v1beta"
 )
 
 var (
@@ -43,15 +46,21 @@ type Client struct {
 	tokenStorage       *auth.TokenStorage
 	cfg                *config.Config
 	modelQuotaExceeded map[string]*time.Time
+	glAPIKey           string
 }
 
 // NewClient creates a new CLI API client.
-func NewClient(httpClient *http.Client, ts *auth.TokenStorage, cfg *config.Config) *Client {
+func NewClient(httpClient *http.Client, ts *auth.TokenStorage, cfg *config.Config, glAPIKey ...string) *Client {
+	var glKey string
+	if len(glAPIKey) > 0 {
+		glKey = glAPIKey[0]
+	}
 	return &Client{
 		httpClient:         httpClient,
 		tokenStorage:       ts,
 		cfg:                cfg,
 		modelQuotaExceeded: make(map[string]*time.Time),
+		glAPIKey:           glKey,
 	}
 }
 
@@ -80,7 +89,14 @@ func (c *Client) GetEmail() string {
 }
 
 func (c *Client) GetProjectID() string {
-	return c.tokenStorage.ProjectID
+	if c.tokenStorage != nil {
+		return c.tokenStorage.ProjectID
+	}
+	return ""
+}
+
+func (c *Client) GetGenerativeLanguageAPIKey() string {
+	return c.glAPIKey
 }
 
 // SetupUser performs the initial user onboarding and setup.
@@ -196,6 +212,7 @@ func (c *Client) makeAPIRequest(ctx context.Context, endpoint, method string, bo
 	metadataStr := getClientMetadataString()
 	req.Header.Set("Content-Type", "application/json")
 	req.Header.Set("User-Agent", getUserAgent())
+	req.Header.Set("X-Goog-Api-Client", "gl-node/22.17.0")
 	req.Header.Set("Client-Metadata", metadataStr)
 	req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", token.AccessToken))
 
@@ -235,35 +252,56 @@ func (c *Client) APIRequest(ctx context.Context, endpoint string, body interface
 			return nil, &ErrorMessage{500, fmt.Errorf("failed to marshal request body: %w", err)}
 		}
 	}
+
+	var url string
+	if c.glAPIKey == "" {
+		// Add alt=sse for streaming
+		url = fmt.Sprintf("%s/%s:%s", codeAssistEndpoint, apiVersion, endpoint)
+		if stream {
+			url = url + "?alt=sse"
+		}
+	} else {
+		modelResult := gjson.GetBytes(jsonBody, "model")
+		url = fmt.Sprintf("%s/%s/models/%s:%s", glEndPoint, glApiVersion, modelResult.String(), endpoint)
+		if stream {
+			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))
 	reqBody := bytes.NewBuffer(jsonBody)
 
-	// Add alt=sse for streaming
-	url := fmt.Sprintf("%s/%s:%s", codeAssistEndpoint, apiVersion, endpoint)
-	if stream {
-		url = url + "?alt=sse"
-	}
-
 	req, err := http.NewRequestWithContext(ctx, "POST", url, reqBody)
 	if err != nil {
-		return nil, &ErrorMessage{500, fmt.Errorf("failed to create request: %w", err)}
-	}
-
-	token, err := c.httpClient.Transport.(*oauth2.Transport).Source.Token()
-	if err != nil {
-		return nil, &ErrorMessage{500, fmt.Errorf("failed to get token: %w", err)}
+		return nil, &ErrorMessage{500, fmt.Errorf("failed to create request: %v", err)}
 	}
 
 	// Set headers
 	metadataStr := getClientMetadataString()
 	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("User-Agent", getUserAgent())
-	req.Header.Set("Client-Metadata", metadataStr)
-	req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", token.AccessToken))
+	if c.glAPIKey == "" {
+		token, errToken := c.httpClient.Transport.(*oauth2.Transport).Source.Token()
+		if errToken != nil {
+			return nil, &ErrorMessage{500, fmt.Errorf("failed to get token: %v", errToken)}
+		}
+		req.Header.Set("User-Agent", getUserAgent())
+		req.Header.Set("X-Goog-Api-Client", "gl-node/22.17.0")
+		req.Header.Set("Client-Metadata", metadataStr)
+		req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", token.AccessToken))
+	} else {
+		req.Header.Set("x-goog-api-key", c.glAPIKey)
+	}
 
 	resp, err := c.httpClient.Do(req)
 	if err != nil {
-		return nil, &ErrorMessage{500, fmt.Errorf("failed to execute request: %w", err)}
+		return nil, &ErrorMessage{500, fmt.Errorf("failed to execute request: %v", err)}
 	}
 
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
@@ -273,15 +311,15 @@ func (c *Client) APIRequest(ctx context.Context, endpoint string, body interface
 			}
 		}()
 		bodyBytes, _ := io.ReadAll(resp.Body)
-
+		// log.Debug(string(jsonBody))
 		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{
@@ -290,10 +328,13 @@ func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string,
 			},
 		},
 	}
+
+	request.SystemInstruction = systemInstruction
+
 	request.Tools = tools
 
 	requestBody := map[string]interface{}{
-		"project": c.tokenStorage.ProjectID, // Assuming ProjectID is available
+		"project": c.GetProjectID(), // Assuming ProjectID is available
 		"request": request,
 		"model":   model,
 	}
@@ -337,7 +378,7 @@ func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string,
 	// log.Debug(string(byteRequestBody))
 	for {
 		if c.isModelQuotaExceeded(modelName) {
-			if c.cfg.QuotaExceeded.SwitchPreviewModel {
+			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)
@@ -356,7 +397,7 @@ func (c *Client) SendMessage(ctx context.Context, rawJson []byte, model string,
 			if err.StatusCode == 429 {
 				now := time.Now()
 				c.modelQuotaExceeded[modelName] = &now
-				if c.cfg.QuotaExceeded.SwitchPreviewModel {
+				if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
 					continue
 				}
 			}
@@ -371,8 +412,234 @@ 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) {
+// SendMessageStream handles streaming conversational turns with comprehensive parameter management.
+// This function implements a sophisticated streaming system that supports tool calls, reasoning modes,
+// quota management, and automatic model fallback. It returns two channels for asynchronous communication:
+// one for streaming response data and another for error handling.
+func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model string, systemInstruction *Content, contents []Content, tools []ToolDeclaration, includeThoughts ...bool) (<-chan []byte, <-chan *ErrorMessage) {
+	// Define the data prefix used in Server-Sent Events streaming format
+	dataTag := []byte("data: ")
+
+	// Create channels for asynchronous communication
+	// errChan: delivers error messages during streaming
+	// dataChan: delivers response data chunks
+	errChan := make(chan *ErrorMessage)
+	dataChan := make(chan []byte)
+
+	// Launch a goroutine to handle the streaming process asynchronously
+	// This allows the function to return immediately while processing continues in the background
+	go func() {
+		// Ensure channels are properly closed when the goroutine exits
+		defer close(errChan)
+		defer close(dataChan)
+
+		// Configure thinking/reasoning capabilities
+		// Default to including thoughts unless explicitly disabled
+		includeThoughtsFlag := true
+		if len(includeThoughts) > 0 {
+			includeThoughtsFlag = includeThoughts[0]
+		}
+
+		// Build the base request structure for the Gemini API
+		// This includes conversation contents and generation configuration
+		request := GenerateContentRequest{
+			Contents: contents,
+			GenerationConfig: GenerationConfig{
+				ThinkingConfig: GenerationConfigThinkingConfig{
+					IncludeThoughts: includeThoughtsFlag,
+				},
+			},
+		}
+
+		// Add system instructions if provided
+		// System instructions guide the AI's behavior and response style
+		request.SystemInstruction = systemInstruction
+
+		// Add available tools for function calling capabilities
+		// Tools allow the AI to perform actions beyond text generation
+		request.Tools = tools
+
+		// Construct the complete request body with project context
+		// The project ID is essential for proper API routing and billing
+		requestBody := map[string]interface{}{
+			"project": c.GetProjectID(), // Project ID for API routing and quota management
+			"request": request,
+			"model":   model,
+		}
+
+		// Serialize the request body to JSON for API transmission
+		byteRequestBody, _ := json.Marshal(requestBody)
+
+		// Parse and configure reasoning effort levels from the original request
+		// This maps Claude-style reasoning effort parameters to Gemini's thinking budget system
+		reasoningEffortResult := gjson.GetBytes(rawJson, "reasoning_effort")
+		if reasoningEffortResult.String() == "none" {
+			// Disable thinking entirely for fastest responses
+			byteRequestBody, _ = sjson.DeleteBytes(byteRequestBody, "request.generationConfig.thinkingConfig.include_thoughts")
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 0)
+		} else if reasoningEffortResult.String() == "auto" {
+			// Let the model decide the appropriate thinking budget automatically
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
+		} else if reasoningEffortResult.String() == "low" {
+			// Minimal thinking for simple tasks (1KB thinking budget)
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 1024)
+		} else if reasoningEffortResult.String() == "medium" {
+			// Moderate thinking for complex tasks (8KB thinking budget)
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 8192)
+		} else if reasoningEffortResult.String() == "high" {
+			// Maximum thinking for very complex tasks (24KB thinking budget)
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 24576)
+		} else {
+			// Default to automatic thinking budget if no specific level is provided
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
+		}
+
+		// Configure temperature parameter for response randomness control
+		// Temperature affects the creativity vs consistency trade-off in responses
+		temperatureResult := gjson.GetBytes(rawJson, "temperature")
+		if temperatureResult.Exists() && temperatureResult.Type == gjson.Number {
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.temperature", temperatureResult.Num)
+		}
+
+		// Configure top-p parameter for nucleus sampling
+		// Controls the cumulative probability threshold for token selection
+		topPResult := gjson.GetBytes(rawJson, "top_p")
+		if topPResult.Exists() && topPResult.Type == gjson.Number {
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.topP", topPResult.Num)
+		}
+
+		// Configure top-k parameter for limiting token candidates
+		// Restricts the model to consider only the top K most likely tokens
+		topKResult := gjson.GetBytes(rawJson, "top_k")
+		if topKResult.Exists() && topKResult.Type == gjson.Number {
+			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.topK", topKResult.Num)
+		}
+
+		// Initialize model name for quota management and potential fallback
+		modelName := model
+		var stream io.ReadCloser
+
+		// Quota management and model fallback loop
+		// This loop handles quota exceeded scenarios and automatic model switching
+		for {
+			// Check if the current model has exceeded its quota
+			if c.isModelQuotaExceeded(modelName) {
+				// Attempt to switch to a preview model if configured and using account auth
+				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)
+						// Update the request body with the new model name
+						byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "model", modelName)
+						continue // Retry with the preview model
+					}
+				}
+				// If no fallback is available, return a quota exceeded error
+				errChan <- &ErrorMessage{
+					StatusCode: 429,
+					Error:      fmt.Errorf(`{"error":{"code":429,"message":"All the models of '%s' are quota exceeded","status":"RESOURCE_EXHAUSTED"}}`, model),
+				}
+				return
+			}
+
+			// Attempt to establish a streaming connection with the API
+			var err *ErrorMessage
+			stream, err = c.APIRequest(ctx, "streamGenerateContent", byteRequestBody, true)
+			if err != nil {
+				// Handle quota exceeded errors by marking the model and potentially retrying
+				if err.StatusCode == 429 {
+					now := time.Now()
+					c.modelQuotaExceeded[modelName] = &now // Mark model as quota exceeded
+					// If preview model switching is enabled, retry the loop
+					if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
+						continue
+					}
+				}
+				// Forward other errors to the error channel
+				errChan <- err
+				return
+			}
+			// Clear any previous quota exceeded status for this model
+			delete(c.modelQuotaExceeded, modelName)
+			break // Successfully established connection, exit the retry loop
+		}
+
+		// Process the streaming response using a scanner
+		// This handles the Server-Sent Events format from the API
+		scanner := bufio.NewScanner(stream)
+		for scanner.Scan() {
+			line := scanner.Bytes()
+			// Filter and forward only data lines (those prefixed with "data: ")
+			// This extracts the actual JSON content from the SSE format
+			if bytes.HasPrefix(line, dataTag) {
+				dataChan <- line[6:] // Remove "data: " prefix and send the JSON content
+			}
+		}
+
+		// Handle any scanning errors that occurred during stream processing
+		if errScanner := scanner.Err(); errScanner != nil {
+			// Send a 500 Internal Server Error for scanning failures
+			errChan <- &ErrorMessage{500, errScanner}
+			_ = stream.Close()
+			return
+		}
+
+		// Ensure the stream is properly closed to prevent resource leaks
+		_ = stream.Close()
+	}()
+
+	// Return the channels immediately for asynchronous communication
+	// The caller can read from these channels while the goroutine processes the request
+	return dataChan, errChan
+}
+
+// SendRawMessage handles a single conversational turn, including tool calls.
+func (c *Client) SendRawMessage(ctx context.Context, rawJson []byte) ([]byte, *ErrorMessage) {
+	if c.glAPIKey == "" {
+		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)
@@ -380,67 +647,21 @@ func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model st
 		defer close(errChan)
 		defer close(dataChan)
 
-		request := GenerateContentRequest{
-			Contents: contents,
-			GenerationConfig: GenerationConfig{
-				ThinkingConfig: GenerationConfigThinkingConfig{
-					IncludeThoughts: true,
-				},
-			},
-		}
-		request.Tools = tools
-
-		requestBody := map[string]interface{}{
-			"project": c.tokenStorage.ProjectID, // Assuming ProjectID is available
-			"request": request,
-			"model":   model,
+		if c.glAPIKey == "" {
+			rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
 		}
 
-		byteRequestBody, _ := json.Marshal(requestBody)
-
-		// log.Debug(string(byteRequestBody))
-
-		reasoningEffortResult := gjson.GetBytes(rawJson, "reasoning_effort")
-		if reasoningEffortResult.String() == "none" {
-			byteRequestBody, _ = sjson.DeleteBytes(byteRequestBody, "request.generationConfig.thinkingConfig.include_thoughts")
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 0)
-		} else if reasoningEffortResult.String() == "auto" {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-		} else if reasoningEffortResult.String() == "low" {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 1024)
-		} else if reasoningEffortResult.String() == "medium" {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 8192)
-		} else if reasoningEffortResult.String() == "high" {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", 24576)
-		} else {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-		}
-
-		temperatureResult := gjson.GetBytes(rawJson, "temperature")
-		if temperatureResult.Exists() && temperatureResult.Type == gjson.Number {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.temperature", temperatureResult.Num)
-		}
-
-		topPResult := gjson.GetBytes(rawJson, "top_p")
-		if topPResult.Exists() && topPResult.Type == gjson.Number {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.topP", topPResult.Num)
-		}
-
-		topKResult := gjson.GetBytes(rawJson, "top_k")
-		if topKResult.Exists() && topKResult.Type == gjson.Number {
-			byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "request.generationConfig.topK", topKResult.Num)
-		}
-
-		// log.Debug(string(byteRequestBody))
+		modelResult := gjson.GetBytes(rawJson, "model")
+		model := modelResult.String()
 		modelName := model
 		var stream io.ReadCloser
 		for {
 			if c.isModelQuotaExceeded(modelName) {
-				if c.cfg.QuotaExceeded.SwitchPreviewModel {
+				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)
-						byteRequestBody, _ = sjson.SetBytes(byteRequestBody, "model", modelName)
+						rawJson, _ = sjson.SetBytes(rawJson, "model", modelName)
 						continue
 					}
 				}
@@ -451,12 +672,12 @@ func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model st
 				return
 			}
 			var err *ErrorMessage
-			stream, err = c.APIRequest(ctx, "streamGenerateContent", byteRequestBody, true)
+			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 {
+					if c.cfg.QuotaExceeded.SwitchPreviewModel && c.glAPIKey == "" {
 						continue
 					}
 				}
@@ -470,14 +691,12 @@ func (c *Client) SendMessageStream(ctx context.Context, rawJson []byte, model st
 		scanner := bufio.NewScanner(stream)
 		for scanner.Scan() {
 			line := scanner.Bytes()
-			// log.Printf("Received stream chunk: %s", line)
 			if bytes.HasPrefix(line, dataTag) {
 				dataChan <- line[6:]
 			}
 		}
 
 		if errScanner := scanner.Err(); errScanner != nil {
-			// log.Println(err)
 			errChan <- &ErrorMessage{500, errScanner}
 			_ = stream.Close()
 			return
@@ -629,10 +848,10 @@ func (c *Client) SaveTokenToFile() error {
 // such as IDE type, platform, and plugin version.
 func getClientMetadata() map[string]string {
 	return map[string]string{
-		"ideType":       "IDE_UNSPECIFIED",
-		"platform":      getPlatform(),
-		"pluginType":    "GEMINI",
-		"pluginVersion": pluginVersion,
+		"ideType":    "IDE_UNSPECIFIED",
+		"platform":   "PLATFORM_UNSPECIFIED",
+		"pluginType": "GEMINI",
+		// "pluginVersion": pluginVersion,
 	}
 }
 
@@ -649,7 +868,8 @@ func getClientMetadataString() string {
 
 // getUserAgent constructs the User-Agent string for HTTP requests.
 func getUserAgent() string {
-	return fmt.Sprintf("GeminiCLI/%s (%s; %s)", pluginVersion, runtime.GOOS, runtime.GOARCH)
+	// return fmt.Sprintf("GeminiCLI/%s (%s; %s)", pluginVersion, runtime.GOOS, runtime.GOARCH)
+	return "google-api-nodejs-client/9.15.1"
 }
 
 // getPlatform determines the operating system and architecture and formats

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.

internal/cmd/run.go

@@ -7,8 +7,10 @@ import (
 	"github.com/luispater/CLIProxyAPI/internal/auth"
 	"github.com/luispater/CLIProxyAPI/internal/client"
 	"github.com/luispater/CLIProxyAPI/internal/config"
+	"github.com/luispater/CLIProxyAPI/internal/util"
 	log "github.com/sirupsen/logrus"
 	"io/fs"
+	"net/http"
 	"os"
 	"os/signal"
 	"path/filepath"
@@ -64,6 +66,19 @@ func StartService(cfg *config.Config) {
 		log.Fatalf("Error walking auth directory: %v", err)
 	}
 
+	if len(cfg.GlAPIKey) > 0 {
+		for i := 0; i < len(cfg.GlAPIKey); i++ {
+			httpClient, errSetProxy := util.SetProxy(cfg, &http.Client{})
+			if errSetProxy != nil {
+				log.Fatalf("set proxy failed: %v", errSetProxy)
+			}
+
+			log.Debug("Initializing with Generative Language API key...")
+			cliClient := client.NewClient(httpClient, nil, cfg, cfg.GlAPIKey[i])
+			cliClients = append(cliClients, cliClient)
+		}
+	}
+
 	// Create and start the API server with the pool of clients.
 	apiServer := api.NewServer(cfg, cliClients)
 	log.Infof("Starting API server on port %d", cfg.Port)

internal/config/config.go

@@ -20,6 +20,8 @@ type Config struct {
 	ApiKeys []string `yaml:"api-keys"`
 	// QuotaExceeded defines the behavior when a quota is exceeded.
 	QuotaExceeded ConfigQuotaExceeded `yaml:"quota-exceeded"`
+	// GlAPIKey is the API key for the generative language API.
+	GlAPIKey []string `yaml:"generative-language-api-key"`
 }
 
 type ConfigQuotaExceeded struct {

internal/util/proxy.go

@@ -0,0 +1,37 @@
+package util
+
+import (
+	"context"
+	"github.com/luispater/CLIProxyAPI/internal/config"
+	"golang.org/x/net/proxy"
+	"net"
+	"net/http"
+	"net/url"
+)
+
+func SetProxy(cfg *config.Config, httpClient *http.Client) (*http.Client, error) {
+	var transport *http.Transport
+	proxyURL, errParse := url.Parse(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 {
+				return nil, 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)}
+		}
+	}
+	if transport != nil {
+		httpClient.Transport = transport
+	}
+	return httpClient, nil
+}