Add Docker support with CI/CD workflow and usage instructions

- Added `.github/workflows/docker-image.yml` for automated Docker image build and push on version tags.
- Created `Dockerfile` to containerize the application.
- Updated README with instructions for running the application using Docker.

.github/workflows/docker-image.yml

@@ -0,0 +1,42 @@
+name: docker-image
+
+on:
+  push:
+    tags:
+      - v*
+
+env:
+  APP_NAME: CLIProxyAPI
+  DOCKERHUB_REPO: eceasy/cli-proxy-api
+
+jobs:
+  docker:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Generate App Version
+        run: echo APP_VERSION=`git describe --tags --always` >> $GITHUB_ENV
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: |
+            linux/amd64
+            linux/arm64
+          push: true
+          build-args: |
+            APP_NAME=${{ env.APP_NAME }}
+            APP_VERSION=${{ env.APP_VERSION }}
+          tags: |
+            ${{ env.DOCKERHUB_REPO }}:latest
+            ${{ env.DOCKERHUB_REPO }}:${{ env.APP_VERSION }}

Dockerfile

@@ -0,0 +1,23 @@
+FROM golang:1.24-alpine AS builder
+
+WORKDIR /app
+
+COPY go.mod go.sum ./
+
+RUN go mod download
+
+COPY . .
+
+RUN CGO_ENABLED=0 GOOS=linux go build -o ./CLIProxyAPI ./cmd/server/
+
+FROM alpine:3.22.0
+
+RUN mkdir /CLIProxyAPI
+
+COPY --from=builder ./app/CLIProxyAPI /CLIProxyAPI/CLIProxyAPI
+
+WORKDIR /CLIProxyAPI
+
+EXPOSE 8317
+
+CMD ["./CLIProxyAPI"]

README.md

@@ -1,10 +1,10 @@
 # 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)
@@ -136,7 +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
 
@@ -208,6 +208,24 @@ 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.
+
+## Run with Docker
+
+Run the following command to login: 
+
+```bash
+docker run --rm -p 8085:8085 -v /path/to/your/config.yaml:/CLIProxyAPI/config.yaml -v /path/to/your/auth-dir:/root/.cli-proxy-api eceasy/cli-proxy-api:latest /CLIProxyAPI/CLIProxyAPI --login
+```
+
+Run the following command to start the server:
+
+```bash
+docker run --rm -p 8317:8317 -v /path/to/your/config.yaml:/CLIProxyAPI/config.yaml -v /path/to/your/auth-dir:/root/.cli-proxy-api eceasy/cli-proxy-api:latest
+```
+
 ## Contributing
 
 Contributions are welcome! Please feel free to submit a Pull Request.

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

@@ -12,10 +12,21 @@ import (
 	"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" {

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",

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
@@ -140,7 +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",
 			})
@@ -159,7 +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"
@@ -20,11 +23,15 @@ func PrepareRequest(rawJson []byte) (string, *client.Content, []client.Content,
 		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()
@@ -35,21 +42,26 @@ func PrepareRequest(rawJson []byte) (string, *client.Content, []client.Content,
 				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()
 					}
 
-					// drop the timestamp from the tool call ID
+					// 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
 				}
@@ -124,25 +136,33 @@ func PrepareRequest(rawJson []byte) (string, *client.Content, []client.Content,
 					}
 					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 {
 								parts = append(parts, client.Part{
@@ -153,17 +173,22 @@ func PrepareRequest(rawJson []byte) (string, *client.Content, []client.Content,
 								})
 							}
 						}
+
+						// 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})
 						}
 					}
@@ -197,3 +222,323 @@ func PrepareRequest(rawJson []byte) (string, *client.Content, []client.Content,
 
 	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 +1,7 @@
 package translator
 
 import (
+	"bytes"
 	"fmt"
 	"time"
 
@@ -188,3 +189,194 @@ func ConvertCliToOpenAINonStream(rawJson []byte, unixTimestamp int64, isGlAPIKey
 
 	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

@@ -311,6 +311,7 @@ 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))}
 	}
 
@@ -411,130 +412,192 @@ 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, systemInstruction *Content, 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: true,
+					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(), // Assuming ProjectID is available
+			"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)
 
-		// log.Debug(string(byteRequestBody))
-
+		// 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)
 		}
 
-		// log.Debug(string(byteRequestBody))
+		// 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
+						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
+					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
+			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()
-			// log.Printf("Received stream chunk: %s", line)
+			// 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:]
+				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 {
-			// log.Println(err)
+			// 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) {
-	rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
+	if c.glAPIKey == "" {
+		rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
+	}
 
 	modelResult := gjson.GetBytes(rawJson, "model")
 	model := modelResult.String()
@@ -584,7 +647,9 @@ func (c *Client) SendRawMessageStream(ctx context.Context, rawJson []byte) (<-ch
 		defer close(errChan)
 		defer close(dataChan)
 
-		rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
+		if c.glAPIKey == "" {
+			rawJson, _ = sjson.SetBytes(rawJson, "project", c.GetProjectID())
+		}
 
 		modelResult := gjson.GetBytes(rawJson, "model")
 		model := modelResult.String()