feat: add token caching middleware and max history messages config

- Strip additionalProperties and $schema from tool parameter schemas
  (Gemini API rejects these unsupported JSON Schema fields)
- Add thoughtSignature to functionCall parts for multi-turn tool calling
  (Gemini 3 models require thought signatures on function call history)
- Ensure functionResponse.response is always a JSON object, never an
  array or primitive (wrap non-objects in {"result": ...})
- Resolve tool names from tool_call_id or positional index when the
  tool message doesn't carry a name field
- Add functionCall parsing in streaming responses (emitGeminiChunk)
  to properly relay tool calls via SSE
- Add error logging for Gemini stream failures with request body dump
- Add unit tests for schema cleaning and streaming tool call emission

internal/config/config.go

@@ -19,10 +19,11 @@ type Config struct {
 }
 
 type ServerConfig struct {
-	Port            int      `mapstructure:"port"`
-	Host            string   `mapstructure:"host"`
-	AuthTokens      []string `mapstructure:"auth_tokens"`
-	WSAllowedOrigin string   `mapstructure:"ws_allowed_origin"`
+	Port               int      `mapstructure:"port"`
+	Host               string   `mapstructure:"host"`
+	AuthTokens         []string `mapstructure:"auth_tokens"`
+	WSAllowedOrigin    string   `mapstructure:"ws_allowed_origin"`
+	MaxHistoryMessages int      `mapstructure:"max_history_messages"`
 }
 
 type DatabaseConfig struct {
@@ -96,6 +97,7 @@ func Load() (*Config, error) {
 	v.SetDefault("server.port", 8080)
 	v.SetDefault("server.host", "0.0.0.0")
 	v.SetDefault("server.auth_tokens", []string{})
+	v.SetDefault("server.max_history_messages", 0)
 	v.SetDefault("database.path", "./data/llm_proxy.db")
 	v.SetDefault("database.max_connections", 10)
 
@@ -142,6 +144,7 @@ func Load() (*Config, error) {
 	v.BindEnv("encryption_key", "LLM_PROXY__ENCRYPTION_KEY")
 	v.BindEnv("server.port", "LLM_PROXY__SERVER__PORT")
 	v.BindEnv("server.host", "LLM_PROXY__SERVER__HOST")
+	v.BindEnv("server.max_history_messages", "LLM_PROXY__SERVER__MAX_HISTORY_MESSAGES")
 	v.BindEnv("providers.ollama.enabled", "LLM_PROXY__PROVIDERS__OLLAMA__ENABLED")
 	v.BindEnv("providers.ollama.base_url", "LLM_PROXY__PROVIDERS__OLLAMA__BASE_URL")
 	v.BindEnv("providers.ollama.models", "LLM_PROXY__PROVIDERS__OLLAMA__MODELS")
@@ -174,6 +177,9 @@ func Load() (*Config, error) {
 		cfg.Server.Host = host
 
 	}
+	if maxHistory := os.Getenv("LLM_PROXY__SERVER__MAX_HISTORY_MESSAGES"); maxHistory != "" {
+		fmt.Sscanf(maxHistory, "%d", &cfg.Server.MaxHistoryMessages)
+	}
 
 	// Ollama overrides
 	if enabled := os.Getenv("LLM_PROXY__PROVIDERS__OLLAMA__ENABLED"); enabled != "" {

internal/middleware/auth.go

@@ -4,6 +4,8 @@ import (
 	"log"
 	"net/http"
 	"strings"
+	"sync"
+	"time"
 
 	"gophergate/internal/db"
 	"gophergate/internal/models"
@@ -11,6 +13,15 @@ import (
 	"github.com/gin-gonic/gin"
 )
 
+type tokenCacheEntry struct {
+	clientID  string
+	expiredAt time.Time
+}
+
+var (
+	tokenCache sync.Map // map[string]tokenCacheEntry
+)
+
 func AuthMiddleware(database *db.DB, requireAuth bool) gin.HandlerFunc {
 	return func(c *gin.Context) {
 		authHeader := c.GetHeader("Authorization")
@@ -52,11 +63,39 @@ func AuthMiddleware(database *db.DB, requireAuth bool) gin.HandlerFunc {
 			return
 		}
 
-		// Try to resolve client from database with a read-only SELECT
+		// Try to resolve client from cache first
 		var clientID string
-		err := database.Get(&clientID, "SELECT client_id FROM client_tokens WHERE token = ? AND is_active = 1", token)
+		var dbErr error
+		now := time.Now()
+		if cached, ok := tokenCache.Load(token); ok {
+			entry := cached.(tokenCacheEntry)
+			if now.Before(entry.expiredAt) {
+				clientID = entry.clientID
+			}
+		}
 
-		if err == nil {
+		// If cache miss (or expired), query database
+		if clientID == "" {
+			var fetchedID string
+			dbErr = database.Get(&fetchedID, "SELECT client_id FROM client_tokens WHERE token = ? AND is_active = 1", token)
+			if dbErr == nil {
+				clientID = fetchedID
+				// Cache valid token for 10 seconds
+				tokenCache.Store(token, tokenCacheEntry{
+					clientID:  clientID,
+					expiredAt: now.Add(10 * time.Second),
+				})
+			} else {
+				// If error (invalid/inactive token), cache negative result for 2 seconds
+				// to avoid hammering SQLite on repeated invalid requests
+				tokenCache.Store(token, tokenCacheEntry{
+					clientID:  "",
+					expiredAt: now.Add(2 * time.Second),
+				})
+			}
+		}
+
+		if clientID != "" {
 			c.Set("auth", models.AuthInfo{
 				Token:    token,
 				ClientID: clientID,
@@ -76,7 +115,11 @@ func AuthMiddleware(database *db.DB, requireAuth bool) gin.HandlerFunc {
 			if len(token) > 8 {
 				maskedToken = token[:3] + "••••" + token[len(token)-4:]
 			}
-			log.Printf("Token not found, inactive or error in DB: %s (err: %v)", maskedToken, err)
+			if dbErr != nil {
+				log.Printf("Token not found, inactive or error in DB: %s (err: %v)", maskedToken, dbErr)
+			} else {
+				log.Printf("Token not found or inactive: %s", maskedToken)
+			}
 			c.AbortWithStatusJSON(http.StatusUnauthorized, gin.H{
 				"error": gin.H{
 					"message": "Invalid or inactive client token.",

internal/providers/gemini.go

@@ -59,6 +59,7 @@ type GeminiPart struct {
 	InlineData       *GeminiInlineData       `json:"inlineData,omitempty"`
 	FunctionCall     *GeminiFunctionCall     `json:"functionCall,omitempty"`
 	FunctionResponse *GeminiFunctionResponse `json:"functionResponse,omitempty"`
+	ThoughtSignature string                  `json:"thoughtSignature,omitempty"`
 }
 
 type GeminiInlineData struct {
@@ -231,6 +232,7 @@ func (p *GeminiProvider) ChatCompletion(ctx context.Context, req *models.Unified
 						Name: tc.Function.Name,
 						Args: json.RawMessage(tc.Function.Arguments),
 					},
+					ThoughtSignature: "skip_thought_signature_validator",
 				})
 			}
 			contents = append(contents, GeminiContent{Role: "model", Parts: parts})
@@ -256,16 +258,9 @@ func (p *GeminiProvider) ChatCompletion(ctx context.Context, req *models.Unified
 				if len(m.Content) > 0 {
 					text = m.Content[0].Text
 				}
-				name := "unknown_function"
-				if m.Name != nil {
-					name = *m.Name
-				}
+				name := resolveToolName(m, msg.ToolCalls, j-i-1)
 
-				var responseObj interface{}
-				if err := json.Unmarshal([]byte(text), &responseObj); err != nil {
-					responseObj = map[string]interface{}{"result": text}
-				}
-				respBytes, _ := json.Marshal(responseObj)
+				respBytes := ensureJSONObject(text)
 
 				functionParts = append(functionParts, GeminiPart{
 					FunctionResponse: &GeminiFunctionResponse{
@@ -353,7 +348,9 @@ func (p *GeminiProvider) ChatCompletion(ctx context.Context, req *models.Unified
 		geminiTool := GeminiTool{FunctionDeclarations: []models.FunctionDef{}}
 		for _, t := range req.Tools {
 			if t.Type == "function" {
-				geminiTool.FunctionDeclarations = append(geminiTool.FunctionDeclarations, t.Function)
+				funcDef := t.Function
+				funcDef.Parameters = cleanParametersSchema(funcDef.Parameters)
+				geminiTool.FunctionDeclarations = append(geminiTool.FunctionDeclarations, funcDef)
 			}
 		}
 		if len(geminiTool.FunctionDeclarations) > 0 {
@@ -509,6 +506,7 @@ func (p *GeminiProvider) ChatCompletionStream(ctx context.Context, req *models.U
 						Name: tc.Function.Name,
 						Args: json.RawMessage(tc.Function.Arguments),
 					},
+					ThoughtSignature: "skip_thought_signature_validator",
 				})
 			}
 			contents = append(contents, GeminiContent{Role: "model", Parts: parts})
@@ -522,16 +520,9 @@ func (p *GeminiProvider) ChatCompletionStream(ctx context.Context, req *models.U
 				if len(m.Content) > 0 {
 					text = m.Content[0].Text
 				}
-				name := "unknown_function"
-				if m.Name != nil {
-					name = *m.Name
-				}
+				name := resolveToolName(m, msg.ToolCalls, j-i-1)
 
-				var responseObj interface{}
-				if err := json.Unmarshal([]byte(text), &responseObj); err != nil {
-					responseObj = map[string]interface{}{"result": text}
-				}
-				respBytes, _ := json.Marshal(responseObj)
+				respBytes := ensureJSONObject(text)
 
 				functionParts = append(functionParts, GeminiPart{
 					FunctionResponse: &GeminiFunctionResponse{
@@ -610,7 +601,9 @@ func (p *GeminiProvider) ChatCompletionStream(ctx context.Context, req *models.U
 		geminiTool := GeminiTool{FunctionDeclarations: []models.FunctionDef{}}
 		for _, t := range req.Tools {
 			if t.Type == "function" {
-				geminiTool.FunctionDeclarations = append(geminiTool.FunctionDeclarations, t.Function)
+				funcDef := t.Function
+				funcDef.Parameters = cleanParametersSchema(funcDef.Parameters)
+				geminiTool.FunctionDeclarations = append(geminiTool.FunctionDeclarations, funcDef)
 			}
 		}
 		if len(geminiTool.FunctionDeclarations) > 0 {
@@ -652,6 +645,9 @@ func (p *GeminiProvider) ChatCompletionStream(ctx context.Context, req *models.U
 				msg = string(body)
 			}
 		}
+		fmt.Printf("[Gemini-Stream] API Error %d: %s\n", resp.StatusCode(), msg)
+		reqJSON, _ := json.Marshal(body)
+		fmt.Printf("[Gemini-Stream] Request Body: %s\n", string(reqJSON))
 		return nil, fmt.Errorf("Gemini API error (%d): %s", resp.StatusCode(), msg)
 	}
 
@@ -668,3 +664,87 @@ func uint32Ptr(v uint32) *uint32 {
 	}
 	return nil
 }
+
+func cleanGeminiSchema(schema map[string]interface{}) {
+	delete(schema, "additionalProperties")
+	delete(schema, "$schema")
+	for _, v := range schema {
+		if subMap, ok := v.(map[string]interface{}); ok {
+			cleanGeminiSchema(subMap)
+		} else if subList, ok := v.([]interface{}); ok {
+			for _, item := range subList {
+				if itemMap, ok := item.(map[string]interface{}); ok {
+					cleanGeminiSchema(itemMap)
+				}
+			}
+		}
+	}
+}
+
+func cleanParametersSchema(raw json.RawMessage) json.RawMessage {
+	if len(raw) == 0 {
+		return raw
+	}
+	var schema map[string]interface{}
+	if err := json.Unmarshal(raw, &schema); err != nil {
+		return raw
+	}
+	cleanGeminiSchema(schema)
+	cleaned, err := json.Marshal(schema)
+	if err != nil {
+		return raw
+	}
+	return json.RawMessage(cleaned)
+}
+
+// ensureJSONObject ensures that a tool result string is serialized as a JSON object
+// for Gemini's functionResponse.response field, which must always be an object.
+// If the text is a JSON array, primitive, or invalid JSON, it gets wrapped in {"result": ...}.
+func ensureJSONObject(text string) []byte {
+	if text == "" {
+		b, _ := json.Marshal(map[string]interface{}{"result": ""})
+		return b
+	}
+
+	var parsed interface{}
+	if err := json.Unmarshal([]byte(text), &parsed); err != nil {
+		// Not valid JSON — wrap as string
+		b, _ := json.Marshal(map[string]interface{}{"result": text})
+		return b
+	}
+
+	// Only allow map types through; wrap everything else
+	if _, ok := parsed.(map[string]interface{}); ok {
+		return []byte(text)
+	}
+
+	// It's a JSON array, number, string, bool, or null — wrap it
+	b, _ := json.Marshal(map[string]interface{}{"result": parsed})
+	return b
+}
+
+// resolveToolName determines the function name for a tool response message.
+// It tries: 1) the Name field on the tool message, 2) matching by ToolCallID
+// against the preceding assistant's tool calls, 3) positional index match.
+func resolveToolName(toolMsg models.UnifiedMessage, toolCalls []models.ToolCall, posIndex int) string {
+	if toolMsg.Name != nil && *toolMsg.Name != "" {
+		return *toolMsg.Name
+	}
+
+	// Try to match by tool_call_id
+	if toolMsg.ToolCallID != nil && *toolMsg.ToolCallID != "" {
+		for _, tc := range toolCalls {
+			if tc.ID == *toolMsg.ToolCallID {
+				return tc.Function.Name
+			}
+		}
+	}
+
+	// Positional fallback
+	if posIndex >= 0 && posIndex < len(toolCalls) {
+		return toolCalls[posIndex].Function.Name
+	}
+
+	return "unknown_function"
+}
+

internal/providers/gemini_test.go

@@ -0,0 +1,266 @@
+package providers
+
+import (
+	"encoding/json"
+	"reflect"
+	"strings"
+	"testing"
+
+	"gophergate/internal/models"
+)
+
+func TestCleanParametersSchema(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected string
+	}{
+		{
+			name: "flat additionalProperties and $schema",
+			input: `{
+				"$schema": "http://json-schema.org/draft-07/schema#",
+				"type": "object",
+				"properties": {
+					"query": {
+						"type": "string"
+					}
+				},
+				"required": ["query"],
+				"additionalProperties": false
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"query": {
+						"type": "string"
+					}
+				},
+				"required": ["query"]
+			}`,
+		},
+		{
+			name: "nested additionalProperties",
+			input: `{
+				"type": "object",
+				"properties": {
+					"user": {
+						"type": "object",
+						"properties": {
+							"id": {
+								"type": "integer"
+							}
+						},
+						"additionalProperties": false
+					}
+				},
+				"additionalProperties": true
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"user": {
+						"type": "object",
+						"properties": {
+							"id": {
+								"type": "integer"
+							}
+						}
+					}
+				}
+			}`,
+		},
+		{
+			name: "nested additionalProperties in array items",
+			input: `{
+				"type": "object",
+				"properties": {
+					"items": {
+						"type": "array",
+						"items": {
+							"type": "object",
+							"properties": {
+								"name": {
+									"type": "string"
+								}
+							},
+							"additionalProperties": false
+						}
+					}
+				}
+			}`,
+			expected: `{
+				"type": "object",
+				"properties": {
+					"items": {
+						"type": "array",
+						"items": {
+							"type": "object",
+							"properties": {
+								"name": {
+									"type": "string"
+								}
+							}
+						}
+					}
+				}
+			}`,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			inputRaw := json.RawMessage(tc.input)
+			cleanedRaw := cleanParametersSchema(inputRaw)
+
+			var cleanedMap, expectedMap map[string]interface{}
+			if err := json.Unmarshal(cleanedRaw, &cleanedMap); err != nil {
+				t.Fatalf("failed to unmarshal cleaned: %v", err)
+			}
+			if err := json.Unmarshal([]byte(tc.expected), &expectedMap); err != nil {
+				t.Fatalf("failed to unmarshal expected: %v", err)
+			}
+
+			if !reflect.DeepEqual(cleanedMap, expectedMap) {
+				t.Errorf("expected %v, got %v", expectedMap, cleanedMap)
+			}
+		})
+	}
+}
+
+func TestEmitGeminiChunk_ToolCalls(t *testing.T) {
+	ch := make(chan *models.ChatCompletionStreamResponse, 1)
+	defer close(ch)
+
+	chunk := &geminiStreamChunk{}
+	chunk.Candidates = []struct {
+		Content struct {
+			Parts []struct {
+				Text         string `json:"text,omitempty"`
+				Thought      string `json:"thought,omitempty"`
+				FunctionCall *struct {
+					Name string          `json:"name"`
+					Args json.RawMessage `json:"args"`
+				} `json:"functionCall,omitempty"`
+			} `json:"parts"`
+		} `json:"content"`
+		FinishReason string `json:"finishReason"`
+	}{
+		{
+			Content: struct {
+				Parts []struct {
+					Text         string `json:"text,omitempty"`
+					Thought      string `json:"thought,omitempty"`
+					FunctionCall *struct {
+						Name string          `json:"name"`
+						Args json.RawMessage `json:"args"`
+					} `json:"functionCall,omitempty"`
+				} `json:"parts"`
+			}{
+				Parts: []struct {
+					Text         string `json:"text,omitempty"`
+					Thought      string `json:"thought,omitempty"`
+					FunctionCall *struct {
+						Name string          `json:"name"`
+						Args json.RawMessage `json:"args"`
+					} `json:"functionCall,omitempty"`
+				}{
+					{
+						FunctionCall: &struct {
+							Name string          `json:"name"`
+							Args json.RawMessage `json:"args"`
+						}{
+							Name: "google_search",
+							Args: json.RawMessage(`{"query": "test"}`),
+						},
+					},
+				},
+			},
+			FinishReason: "",
+		},
+	}
+
+	emitted := emitGeminiChunk(ch, chunk, "gemini-3-flash-preview")
+	if !emitted {
+		t.Fatalf("expected emitGeminiChunk to return true")
+	}
+
+	select {
+	case resp := <-ch:
+		if len(resp.Choices) != 1 {
+			t.Fatalf("expected 1 choice, got %d", len(resp.Choices))
+		}
+		choice := resp.Choices[0]
+		if choice.FinishReason == nil || *choice.FinishReason != "tool_calls" {
+			t.Errorf("expected finish_reason 'tool_calls', got %v", choice.FinishReason)
+		}
+		if len(choice.Delta.ToolCalls) != 1 {
+			t.Fatalf("expected 1 tool call in delta, got %d", len(choice.Delta.ToolCalls))
+		}
+		tc := choice.Delta.ToolCalls[0]
+		if tc.ID == nil || *tc.ID != "call_google_search" {
+			t.Errorf("expected ID 'call_google_search', got %v", tc.ID)
+		}
+		if tc.Function == nil || tc.Function.Name == nil || *tc.Function.Name != "google_search" {
+			t.Errorf("expected function name 'google_search', got %v", tc.Function)
+		}
+		if tc.Function == nil || tc.Function.Arguments == nil || *tc.Function.Arguments != `{"query": "test"}` {
+			t.Errorf("expected arguments '{\"query\": \"test\"}', got %v", tc.Function)
+		}
+	default:
+		t.Fatalf("expected response on channel")
+	}
+}
+
+func TestStreamGeminiJSONArrayStream(t *testing.T) {
+	input := `[
+		{
+			"candidates": [
+				{
+					"content": {
+						"parts": [
+							{"text": "Hello"}
+						]
+					}
+				}
+			]
+		},
+		{
+			"candidates": [
+				{
+					"content": {
+						"parts": [
+							{"text": " world"}
+						]
+					}
+				}
+			],
+			"usageMetadata": {
+				"promptTokenCount": 5,
+				"candidatesTokenCount": 2,
+				"totalTokenCount": 7
+			}
+		}
+	]`
+
+	ch := make(chan *models.ChatCompletionStreamResponse, 5)
+	defer close(ch)
+
+	streamGeminiJSONArrayStream(strings.NewReader(input), ch, "gemini-2.5-flash")
+
+	var texts []string
+	for i := 0; i < 2; i++ {
+		select {
+		case resp := <-ch:
+			if len(resp.Choices) > 0 && resp.Choices[0].Delta.Content != nil {
+				texts = append(texts, *resp.Choices[0].Delta.Content)
+			}
+		default:
+			t.Fatalf("expected chunk %d", i+1)
+		}
+	}
+
+	joined := strings.Join(texts, "")
+	if joined != "Hello world" {
+		t.Errorf("expected 'Hello world', got %q", joined)
+	}
+}

internal/providers/helpers.go

@@ -409,8 +409,12 @@ type geminiStreamChunk struct {
 	Candidates []struct {
 		Content struct {
 			Parts []struct {
-				Text    string `json:"text,omitempty"`
-				Thought string `json:"thought,omitempty"`
+				Text         string `json:"text,omitempty"`
+				Thought      string `json:"thought,omitempty"`
+				FunctionCall *struct {
+					Name string          `json:"name"`
+					Args json.RawMessage `json:"args"`
+				} `json:"functionCall,omitempty"`
 			} `json:"parts"`
 		} `json:"content"`
 		FinishReason string `json:"finishReason"`
@@ -433,6 +437,7 @@ func emitGeminiChunk(ch chan<- *models.ChatCompletionStreamResponse, chunk *gemi
 	content := ""
 	var reasoning *string
 	var finishReason *string
+	var toolCalls []models.ToolCallDelta
 	if len(chunk.Candidates) > 0 {
 		for _, p := range chunk.Candidates[0].Content.Parts {
 			if p.Text != "" {
@@ -444,8 +449,26 @@ func emitGeminiChunk(ch chan<- *models.ChatCompletionStreamResponse, chunk *gemi
 				}
 				*reasoning += p.Thought
 			}
+			if p.FunctionCall != nil {
+				name := p.FunctionCall.Name
+				args := string(p.FunctionCall.Args)
+				tcID := fmt.Sprintf("call_%s", name)
+				tcType := "function"
+				toolCalls = append(toolCalls, models.ToolCallDelta{
+					Index: uint32(len(toolCalls)),
+					ID:    &tcID,
+					Type:  &tcType,
+					Function: &models.FunctionCallDelta{
+						Name:      &name,
+						Arguments: &args,
+					},
+				})
+			}
 		}
 		fr := strings.ToLower(chunk.Candidates[0].FinishReason)
+		if len(toolCalls) > 0 && fr == "" {
+			fr = "tool_calls"
+		}
 		finishReason = &fr
 	}
 
@@ -460,6 +483,7 @@ func emitGeminiChunk(ch chan<- *models.ChatCompletionStreamResponse, chunk *gemi
 				Delta: models.ChatStreamDelta{
 					Content:          &content,
 					ReasoningContent: reasoning,
+					ToolCalls:        toolCalls,
 				},
 				FinishReason: finishReason,
 			},
@@ -499,9 +523,12 @@ func StreamGemini(ctx io.ReadCloser, model string) (<-chan *models.ChatCompletio
 		first := string(peek[:n])
 
 		if first[0] == '[' {
-			// JSON array format
-			rest, _ := io.ReadAll(ctx)
-			streamGeminiJSONArray(append([]byte(first), rest...), ch, model)
+			// JSON array format — stream parse it in real-time
+			combined := io.MultiReader(
+				strings.NewReader(string(peek[:n])),
+				ctx,
+			)
+			streamGeminiJSONArrayStream(combined, ch, model)
 			return
 		} else if strings.HasPrefix(first, "data:") || strings.HasPrefix(first, "data: ") {
 			// SSE format — pre-pend the peeked bytes then run SSE scanner
@@ -524,35 +551,48 @@ func StreamGemini(ctx io.ReadCloser, model string) (<-chan *models.ChatCompletio
 	return ch, nil
 }
 
-// readAll reads remaining bytes from a reader (keeps the function signature simple
-// for the JSON array fallback path).
-func readAll(r io.Reader) []byte {
-	b, _ := io.ReadAll(r)
-	return b
-}
+func streamGeminiJSONArrayStream(r io.Reader, ch chan<- *models.ChatCompletionStreamResponse, model string) {
+	dec := json.NewDecoder(r)
 
-func streamGeminiJSONArray(data []byte, ch chan<- *models.ChatCompletionStreamResponse, model string) {
-	var chunks []geminiStreamChunk
-	if err := json.Unmarshal(data, &chunks); err != nil {
-		fmt.Printf("[Gemini-Stream] JSON array parse error: %v\n", err)
+	// Read open bracket '['
+	t, err := dec.Token()
+	if err != nil {
+		fmt.Printf("[Gemini-Stream] JSON array token error: %v\n", err)
 		return
 	}
-	// Track the last chunk with usage for the final emission
+	delim, ok := t.(json.Delim)
+	if !ok || delim != '[' {
+		fmt.Printf("[Gemini-Stream] JSON array expected '[', got %v\n", t)
+		return
+	}
+
 	var lastUsage *geminiStreamChunk
-	for i := range chunks {
-		if chunks[i].UsageMetadata.TotalTokenCount > 0 {
-			lastUsage = &chunks[i]
+
+	// Read array elements
+	for dec.More() {
+		var chunk geminiStreamChunk
+		if err := dec.Decode(&chunk); err != nil {
+			fmt.Printf("[Gemini-Stream] JSON array decode error: %v\n", err)
+			return
+		}
+
+		if chunk.UsageMetadata.TotalTokenCount > 0 {
+			temp := chunk
+			lastUsage = &temp
+		}
+
+		// Emit content-bearing chunks immediately
+		if len(chunk.Candidates) > 0 {
+			emitGeminiChunk(ch, &chunk, model)
 		}
 	}
-	if lastUsage != nil {
-		// Emit a synthetic final chunk with usage data
-		if len(lastUsage.Candidates) == 0 && lastUsage.UsageMetadata.TotalTokenCount > 0 {
-			emitGeminiChunk(ch, lastUsage, model)
-		}
-	}
-	// Also emit each content-bearing chunk
-	for i := range chunks {
-		emitGeminiChunk(ch, &chunks[i], model)
+
+	// Read close bracket ']'
+	_, _ = dec.Token()
+
+	// Emit synthetic final chunk with usage if we collected it and it was not yet emitted
+	if lastUsage != nil && len(lastUsage.Candidates) == 0 && lastUsage.UsageMetadata.TotalTokenCount > 0 {
+		emitGeminiChunk(ch, lastUsage, model)
 	}
 }
 

internal/server/routing_test.go

@@ -1,6 +1,10 @@
 package server
 
-import "testing"
+import (
+	"testing"
+
+	"gophergate/internal/router"
+)
 
 func TestIsSoftwareDevelopment(t *testing.T) {
 	tests := []struct {
@@ -29,3 +33,68 @@ func TestIsSoftwareDevelopment(t *testing.T) {
 		}
 	}
 }
+
+func TestGetRouteCtxTags(t *testing.T) {
+	s := &Server{}
+	tests := []struct {
+		name         string
+		routeCtx     *router.RouteContext
+		expectedTags []string
+		mustContain  []string
+		mustExclude  []string
+	}{
+		{
+			name: "Standard query with tools",
+			routeCtx: &router.RouteContext{
+				UserMessage:         "Search the web for weather in Paris",
+				RequiresToolCalling: true,
+			},
+			mustExclude: []string{"tool-heavy", "swe-bench"},
+		},
+		{
+			name: "Coding query with tools",
+			routeCtx: &router.RouteContext{
+				UserMessage:           "Write a python script to parse logs",
+				RequiresToolCalling:   true,
+				IsSoftwareDevelopment: true,
+			},
+			mustContain: []string{"tool-heavy", "swe-bench"},
+		},
+		{
+			name: "Agent query with tools",
+			routeCtx: &router.RouteContext{
+				UserMessage:         "agent please orchestrate the multi-agent task",
+				RequiresToolCalling: true,
+			},
+			mustContain: []string{"tool-heavy", "swe-bench"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			tags := s.getRouteCtxTags(tt.routeCtx)
+
+			for _, expected := range tt.mustContain {
+				found := false
+				for _, tag := range tags {
+					if tag == expected {
+						found = true
+						break
+					}
+				}
+				if !found {
+					t.Errorf("expected tag %q to be present, but was not in %v", expected, tags)
+				}
+			}
+
+			for _, excluded := range tt.mustExclude {
+				for _, tag := range tags {
+					if tag == excluded {
+						t.Errorf("expected tag %q to be excluded, but was found in %v", excluded, tags)
+					}
+				}
+			}
+		})
+	}
+}
+

internal/server/server.go

@@ -574,6 +574,31 @@ func (s *Server) handleChatCompletions(c *gin.Context) {
 		return
 	}
 
+	// Prune message history to sliding window if configured
+	if s.cfg.Server.MaxHistoryMessages > 0 && len(req.Messages) > s.cfg.Server.MaxHistoryMessages {
+		var systemMsgs []models.ChatMessage
+		var otherMsgs []models.ChatMessage
+		for _, msg := range req.Messages {
+			if msg.Role == "system" {
+				systemMsgs = append(systemMsgs, msg)
+			} else {
+				otherMsgs = append(otherMsgs, msg)
+			}
+		}
+
+		keepCount := s.cfg.Server.MaxHistoryMessages - len(systemMsgs)
+		if keepCount < 1 {
+			keepCount = 1
+		}
+		if len(otherMsgs) > keepCount {
+			startIndex := len(otherMsgs) - keepCount
+			otherMsgs = otherMsgs[startIndex:]
+			log.Printf("[DEBUG] Pruned message history: kept %d system messages and last %d messages (total %d out of %d)", 
+				len(systemMsgs), len(otherMsgs), len(systemMsgs)+len(otherMsgs), len(req.Messages))
+			req.Messages = append(systemMsgs, otherMsgs...)
+		}
+	}
+
 	// Strip common prefixes and prepare model ID
 	modelID := req.Model
 	prefixes := []string{"gemini/", "google/", "openai/", "deepseek/", "moonshot/", "grok/", "ollama/", "xiaomi/"}
@@ -1119,7 +1144,15 @@ func (s *Server) getRouteCtxTags(routeCtx *router.RouteContext) []string {
 		}
 	}
 
-	if routeCtx.RequiresToolCalling {
+	hasHeavyLogic := false
+	for _, tag := range tags {
+		if tag == "heavy-logic" {
+			hasHeavyLogic = true
+			break
+		}
+	}
+
+	if routeCtx.RequiresToolCalling && (routeCtx.IsSoftwareDevelopment || hasHeavyLogic) {
 		tags = append(tags, "tool-heavy", "multi-step-agent", "swe-bench")
 	}