Merge pull request #181 from taetaetae/fix/kiro-compaction-tool-use-content

fix(kiro): handle tool_use in content array for compaction requests

config.example.yaml

@@ -40,6 +40,11 @@ api-keys:
 # Enable debug logging
 debug: false
 
+# Enable pprof HTTP debug server (host:port). Keep it bound to localhost for safety.
+pprof:
+  enable: false
+  addr: "127.0.0.1:8316"
+
 # When true, disable high-overhead HTTP middleware features to reduce per-request memory usage under high concurrency.
 commercial-mode: false
 

internal/config/config.go

@@ -18,7 +18,10 @@ import (
 	"gopkg.in/yaml.v3"
 )
 
-const DefaultPanelGitHubRepository = "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+const (
+	DefaultPanelGitHubRepository = "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+	DefaultPprofAddr             = "127.0.0.1:8316"
+)
 
 // Config represents the application's configuration, loaded from a YAML file.
 type Config struct {
@@ -41,6 +44,9 @@ type Config struct {
 	// Debug enables or disables debug-level logging and other debug features.
 	Debug bool `yaml:"debug" json:"debug"`
 
+	// Pprof config controls the optional pprof HTTP debug server.
+	Pprof PprofConfig `yaml:"pprof" json:"pprof"`
+
 	// CommercialMode disables high-overhead HTTP middleware features to minimize per-request memory usage.
 	CommercialMode bool `yaml:"commercial-mode" json:"commercial-mode"`
 
@@ -134,6 +140,14 @@ type TLSConfig struct {
 	Key string `yaml:"key" json:"key"`
 }
 
+// PprofConfig holds pprof HTTP server settings.
+type PprofConfig struct {
+	// Enable toggles the pprof HTTP debug server.
+	Enable bool `yaml:"enable" json:"enable"`
+	// Addr is the host:port address for the pprof HTTP server.
+	Addr string `yaml:"addr" json:"addr"`
+}
+
 // RemoteManagement holds management API configuration under 'remote-management'.
 type RemoteManagement struct {
 	// AllowRemote toggles remote (non-localhost) access to management API.
@@ -556,6 +570,8 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	cfg.ErrorLogsMaxFiles = 10
 	cfg.UsageStatisticsEnabled = false
 	cfg.DisableCooling = false
+	cfg.Pprof.Enable = false
+	cfg.Pprof.Addr = DefaultPprofAddr
 	cfg.AmpCode.RestrictManagementToLocalhost = false // Default to false: API key auth is sufficient
 	cfg.RemoteManagement.PanelGitHubRepository = DefaultPanelGitHubRepository
 	cfg.IncognitoBrowser = false // Default to normal browser (AWS uses incognito by force)
@@ -599,6 +615,11 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 		cfg.RemoteManagement.PanelGitHubRepository = DefaultPanelGitHubRepository
 	}
 
+	cfg.Pprof.Addr = strings.TrimSpace(cfg.Pprof.Addr)
+	if cfg.Pprof.Addr == "" {
+		cfg.Pprof.Addr = DefaultPprofAddr
+	}
+
 	if cfg.LogsMaxTotalSizeMB < 0 {
 		cfg.LogsMaxTotalSizeMB = 0
 	}

internal/logging/global_logger.go

@@ -132,7 +132,10 @@ func ResolveLogDirectory(cfg *config.Config) string {
 		return logDir
 	}
 	if !isDirWritable(logDir) {
-		authDir := strings.TrimSpace(cfg.AuthDir)
+		authDir, err := util.ResolveAuthDir(cfg.AuthDir)
+		if err != nil {
+			log.Warnf("Failed to resolve auth-dir %q for log directory: %v", cfg.AuthDir, err)
+		}
 		if authDir != "" {
 			logDir = filepath.Join(authDir, "logs")
 		}

internal/registry/model_definitions_static_data.go

@@ -803,6 +803,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: iFlowThinkingSupport},
 		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
 		{ID: "iflow-rome-30ba3b", DisplayName: "iFlow-ROME", Description: "iFlow Rome 30BA3B model", Created: 1736899200},
+		{ID: "kimi-k2.5", DisplayName: "Kimi-K2.5", Description: "Moonshot Kimi K2.5", Created: 1769443200, Thinking: iFlowThinkingSupport},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {

internal/runtime/executor/antigravity_executor.go

@@ -1280,51 +1280,40 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 	payload = geminiToAntigravity(modelName, payload, projectID)
 	payload, _ = sjson.SetBytes(payload, "model", modelName)
 
-	if strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high") {
+	useAntigravitySchema := strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high")
-		strJSON := string(payload)
+	payloadStr := string(payload)
-		paths := make([]string, 0)
+	paths := make([]string, 0)
-		util.Walk(gjson.ParseBytes(payload), "", "parametersJsonSchema", &paths)
+	util.Walk(gjson.Parse(payloadStr), "", "parametersJsonSchema", &paths)
-		for _, p := range paths {
+	for _, p := range paths {
-			strJSON, _ = util.RenameKey(strJSON, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
+		payloadStr, _ = util.RenameKey(payloadStr, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
-		}
-
-		// Use the centralized schema cleaner to handle unsupported keywords,
-		// const->enum conversion, and flattening of types/anyOf.
-		strJSON = util.CleanJSONSchemaForAntigravity(strJSON)
-		payload = []byte(strJSON)
-	} else {
-		strJSON := string(payload)
-		paths := make([]string, 0)
-		util.Walk(gjson.Parse(strJSON), "", "parametersJsonSchema", &paths)
-		for _, p := range paths {
-			strJSON, _ = util.RenameKey(strJSON, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
-		}
-		// Clean tool schemas for Gemini to remove unsupported JSON Schema keywords
-		// without adding empty-schema placeholders.
-		strJSON = util.CleanJSONSchemaForGemini(strJSON)
-		payload = []byte(strJSON)
 	}
 
-	if strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high") {
+	if useAntigravitySchema {
-		systemInstructionPartsResult := gjson.GetBytes(payload, "request.systemInstruction.parts")
+		payloadStr = util.CleanJSONSchemaForAntigravity(payloadStr)
-		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.role", "user")
+	} else {
-		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.0.text", systemInstruction)
+		payloadStr = util.CleanJSONSchemaForGemini(payloadStr)
-		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
+	}
+
+	if useAntigravitySchema {
+		systemInstructionPartsResult := gjson.Get(payloadStr, "request.systemInstruction.parts")
+		payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.role", "user")
+		payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.parts.0.text", systemInstruction)
+		payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
 
 		if systemInstructionPartsResult.Exists() && systemInstructionPartsResult.IsArray() {
 			for _, partResult := range systemInstructionPartsResult.Array() {
-				payload, _ = sjson.SetRawBytes(payload, "request.systemInstruction.parts.-1", []byte(partResult.Raw))
+				payloadStr, _ = sjson.SetRaw(payloadStr, "request.systemInstruction.parts.-1", partResult.Raw)
 			}
 		}
 	}
 
 	if strings.Contains(modelName, "claude") {
-		payload, _ = sjson.SetBytes(payload, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
+		payloadStr, _ = sjson.Set(payloadStr, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
 	} else {
-		payload, _ = sjson.DeleteBytes(payload, "request.generationConfig.maxOutputTokens")
+		payloadStr, _ = sjson.Delete(payloadStr, "request.generationConfig.maxOutputTokens")
 	}
 
-	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), bytes.NewReader(payload))
+	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), strings.NewReader(payloadStr))
 	if errReq != nil {
 		return nil, errReq
 	}
@@ -1346,11 +1335,15 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 		authLabel = auth.Label
 		authType, authValue = auth.AccountInfo()
 	}
+	var payloadLog []byte
+	if e.cfg != nil && e.cfg.RequestLog {
+		payloadLog = []byte(payloadStr)
+	}
 	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 		URL:       requestURL.String(),
 		Method:    http.MethodPost,
 		Headers:   httpReq.Header.Clone(),
-		Body:      payload,
+		Body:      payloadLog,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,

internal/thinking/apply.go

@@ -388,7 +388,12 @@ func extractGeminiConfig(body []byte, provider string) ThinkingConfig {
 	}
 
 	// Check thinkingLevel first (Gemini 3 format takes precedence)
-	if level := gjson.GetBytes(body, prefix+".thinkingLevel"); level.Exists() {
+	level := gjson.GetBytes(body, prefix+".thinkingLevel")
+	if !level.Exists() {
+		// Google official Gemini Python SDK sends snake_case field names
+		level = gjson.GetBytes(body, prefix+".thinking_level")
+	}
+	if level.Exists() {
 		value := level.String()
 		switch value {
 		case "none":
@@ -401,7 +406,12 @@ func extractGeminiConfig(body []byte, provider string) ThinkingConfig {
 	}
 
 	// Check thinkingBudget (Gemini 2.5 format)
-	if budget := gjson.GetBytes(body, prefix+".thinkingBudget"); budget.Exists() {
+	budget := gjson.GetBytes(body, prefix+".thinkingBudget")
+	if !budget.Exists() {
+		// Google official Gemini Python SDK sends snake_case field names
+		budget = gjson.GetBytes(body, prefix+".thinking_budget")
+	}
+	if budget.Exists() {
 		value := int(budget.Int())
 		switch value {
 		case 0:

internal/thinking/provider/antigravity/apply.go

@@ -94,8 +94,10 @@ func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig, m
 }
 
 func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
@@ -114,28 +116,30 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+
+	// Respect user's explicit includeThoughts setting from original body; default to true if not set
+	// Support both camelCase and snake_case variants
+	includeThoughts := true
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	}
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo, isClaude bool) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
-	includeThoughts := false
-	switch config.Mode {
-	case thinking.ModeNone:
-		includeThoughts = false
-	case thinking.ModeAuto:
-		includeThoughts = true
-	default:
-		includeThoughts = budget > 0
-	}
 
-	// Apply Claude-specific constraints
+	// Apply Claude-specific constraints first to get the final budget value
 	if isClaude && modelInfo != nil {
 		budget, result = a.normalizeClaudeBudget(budget, result, modelInfo)
 		// Check if budget was removed entirely
@@ -144,6 +148,37 @@ func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig,
 		}
 	}
 
+	// For ModeNone, always set includeThoughts to false regardless of user setting.
+	// This ensures that when user requests budget=0 (disable thinking output),
+	// the includeThoughts is correctly set to false even if budget is clamped to min.
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
+		return result, nil
+	}
+
+	// Determine includeThoughts: respect user's explicit setting from original body if provided
+	// Support both camelCase and snake_case variants
+	var includeThoughts bool
+	var userSetIncludeThoughts bool
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	}
+
+	if !userSetIncludeThoughts {
+		// No explicit setting, use default logic based on mode
+		switch config.Mode {
+		case thinking.ModeAuto:
+			includeThoughts = true
+		default:
+			includeThoughts = budget > 0
+		}
+	}
+
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil

internal/thinking/provider/gemini/apply.go

@@ -118,8 +118,10 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
 	// ValidateConfig sets config.Level to the lowest level when ModeNone + Budget > 0.
 
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingBudget")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_budget")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_level")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
 
@@ -138,29 +140,58 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", level)
-	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", true)
+
+	// Respect user's explicit includeThoughts setting from original body; default to true if not set
+	// Support both camelCase and snake_case variants
+	includeThoughts := true
+	if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	} else if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	}
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingLevel")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_level")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_budget")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
-	// ModeNone semantics:
+
-	//   - ModeNone + Budget=0: completely disable thinking
+	// For ModeNone, always set includeThoughts to false regardless of user setting.
-	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
+	// This ensures that when user requests budget=0 (disable thinking output),
-	// When ZeroAllowed=false, ValidateConfig clamps Budget to Min while preserving ModeNone.
+	// the includeThoughts is correctly set to false even if budget is clamped to min.
-	includeThoughts := false
+	if config.Mode == thinking.ModeNone {
-	switch config.Mode {
+		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)
-	case thinking.ModeNone:
+		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", false)
-		includeThoughts = false
+		return result, nil
-	case thinking.ModeAuto:
+	}
-		includeThoughts = true
+
-	default:
+	// Determine includeThoughts: respect user's explicit setting from original body if provided
-		includeThoughts = budget > 0
+	// Support both camelCase and snake_case variants
+	var includeThoughts bool
+	var userSetIncludeThoughts bool
+	if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	} else if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	}
+
+	if !userSetIncludeThoughts {
+		// No explicit setting, use default logic based on mode
+		switch config.Mode {
+		case thinking.ModeAuto:
+			includeThoughts = true
+		default:
+			includeThoughts = budget > 0
+		}
 	}
 
 	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)

internal/thinking/provider/geminicli/apply.go

@@ -79,8 +79,10 @@ func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig) (
 }
 
 func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
@@ -99,25 +101,58 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+
+	// Respect user's explicit includeThoughts setting from original body; default to true if not set
+	// Support both camelCase and snake_case variants
+	includeThoughts := true
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	}
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
-	includeThoughts := false
+
-	switch config.Mode {
+	// For ModeNone, always set includeThoughts to false regardless of user setting.
-	case thinking.ModeNone:
+	// This ensures that when user requests budget=0 (disable thinking output),
-		includeThoughts = false
+	// the includeThoughts is correctly set to false even if budget is clamped to min.
-	case thinking.ModeAuto:
+	if config.Mode == thinking.ModeNone {
-		includeThoughts = true
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-	default:
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
-		includeThoughts = budget > 0
+		return result, nil
+	}
+
+	// Determine includeThoughts: respect user's explicit setting from original body if provided
+	// Support both camelCase and snake_case variants
+	var includeThoughts bool
+	var userSetIncludeThoughts bool
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	}
+
+	if !userSetIncludeThoughts {
+		// No explicit setting, use default logic based on mode
+		switch config.Mode {
+		case thinking.ModeAuto:
+			includeThoughts = true
+		default:
+			includeThoughts = budget > 0
+		}
 	}
 
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)

internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go

@@ -3,7 +3,6 @@
 package chat_completions
 
 import (
-	"bytes"
 	"fmt"
 	"strings"
 
@@ -28,7 +27,7 @@ const geminiCLIFunctionThoughtSignature = "skip_thought_signature_validator"
 // Returns:
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base envelope (no default thinkingConfig)
 	out := []byte(`{"project":"","request":{"contents":[]},"model":"gemini-2.5-pro"}`)
 

internal/translator/claude/gemini/claude_gemini_request.go

@@ -116,7 +116,11 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 		// Include thoughts configuration for reasoning process visibility
 		// Translator only does format conversion, ApplyThinking handles model capability validation.
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+			thinkingLevel := thinkingConfig.Get("thinkingLevel")
+			if !thinkingLevel.Exists() {
+				thinkingLevel = thinkingConfig.Get("thinking_level")
+			}
+			if thinkingLevel.Exists() {
 				level := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
 				switch level {
 				case "":
@@ -132,23 +136,29 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
 					}
 				}
-			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+			} else {
-				budget := int(thinkingBudget.Int())
+				thinkingBudget := thinkingConfig.Get("thinkingBudget")
-				switch budget {
+				if !thinkingBudget.Exists() {
-				case 0:
+					thinkingBudget = thinkingConfig.Get("thinking_budget")
-					out, _ = sjson.Set(out, "thinking.type", "disabled")
+				}
-					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				if thinkingBudget.Exists() {
-				case -1:
+					budget := int(thinkingBudget.Int())
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					switch budget {
-					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					case 0:
-				default:
+						out, _ = sjson.Set(out, "thinking.type", "disabled")
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
-					out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					case -1:
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					default:
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					}
+				} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
 				}
-			} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-				out, _ = sjson.Set(out, "thinking.type", "enabled")
-			} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-				out, _ = sjson.Set(out, "thinking.type", "enabled")
 			}
 		}
 	}

internal/translator/codex/gemini/codex_gemini_request.go

@@ -243,19 +243,30 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	out, _ = sjson.Set(out, "parallel_tool_calls", true)
 
 	// Convert Gemini thinkingConfig to Codex reasoning.effort.
+	// Note: Google official Python SDK sends snake_case fields (thinking_level/thinking_budget).
 	effortSet := false
 	if genConfig := root.Get("generationConfig"); genConfig.Exists() {
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+			thinkingLevel := thinkingConfig.Get("thinkingLevel")
+			if !thinkingLevel.Exists() {
+				thinkingLevel = thinkingConfig.Get("thinking_level")
+			}
+			if thinkingLevel.Exists() {
 				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
 				if effort != "" {
 					out, _ = sjson.Set(out, "reasoning.effort", effort)
 					effortSet = true
 				}
-			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+			} else {
-				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+				thinkingBudget := thinkingConfig.Get("thinkingBudget")
-					out, _ = sjson.Set(out, "reasoning.effort", effort)
+				if !thinkingBudget.Exists() {
-					effortSet = true
+					thinkingBudget = thinkingConfig.Get("thinking_budget")
+				}
+				if thinkingBudget.Exists() {
+					if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+						out, _ = sjson.Set(out, "reasoning.effort", effort)
+						effortSet = true
+					}
 				}
 			}
 		}

internal/translator/gemini-cli/claude/gemini-cli_claude_request.go

@@ -116,6 +116,19 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 						part, _ = sjson.Set(part, "functionResponse.name", funcName)
 						part, _ = sjson.Set(part, "functionResponse.response.result", responseData)
 						contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+
+					case "image":
+						source := contentResult.Get("source")
+						if source.Get("type").String() == "base64" {
+							mimeType := source.Get("media_type").String()
+							data := source.Get("data").String()
+							if mimeType != "" && data != "" {
+								part := `{"inlineData":{"mime_type":"","data":""}}`
+								part, _ = sjson.Set(part, "inlineData.mime_type", mimeType)
+								part, _ = sjson.Set(part, "inlineData.data", data)
+								contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+							}
+						}
 					}
 					return true
 				})

internal/translator/kiro/openai/kiro_openai_request.go

@@ -659,13 +659,36 @@ func buildAssistantMessageFromOpenAI(msg gjson.Result) KiroAssistantResponseMess
 		contentBuilder.WriteString(content.String())
 	} else if content.IsArray() {
 		for _, part := range content.Array() {
-			if part.Get("type").String() == "text" {
+			partType := part.Get("type").String()
+			switch partType {
+			case "text":
 				contentBuilder.WriteString(part.Get("text").String())
+			case "tool_use":
+				// Handle tool_use in content array (Anthropic/OpenCode format)
+				// This is different from OpenAI's tool_calls format
+				toolUseID := part.Get("id").String()
+				toolName := part.Get("name").String()
+				inputData := part.Get("input")
+
+				inputMap := make(map[string]interface{})
+				if inputData.Exists() && inputData.IsObject() {
+					inputData.ForEach(func(key, value gjson.Result) bool {
+						inputMap[key.String()] = value.Value()
+						return true
+					})
+				}
+
+				toolUses = append(toolUses, KiroToolUse{
+					ToolUseID: toolUseID,
+					Name:      toolName,
+					Input:     inputMap,
+				})
+				log.Debugf("kiro-openai: extracted tool_use from content array: %s", toolName)
 			}
 		}
 	}
 
-	// Handle tool_calls
+	// Handle tool_calls (OpenAI format)
 	toolCalls := msg.Get("tool_calls")
 	if toolCalls.IsArray() {
 		for _, tc := range toolCalls.Array() {

internal/translator/openai/gemini/openai_gemini_request.go

@@ -83,16 +83,27 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 		}
 
 		// Map Gemini thinkingConfig to OpenAI reasoning_effort.
-		// Always perform conversion to support allowCompat models that may not be in registry
+		// Always perform conversion to support allowCompat models that may not be in registry.
+		// Note: Google official Python SDK sends snake_case fields (thinking_level/thinking_budget).
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+			thinkingLevel := thinkingConfig.Get("thinkingLevel")
+			if !thinkingLevel.Exists() {
+				thinkingLevel = thinkingConfig.Get("thinking_level")
+			}
+			if thinkingLevel.Exists() {
 				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
 				if effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
-			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+			} else {
-				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+				thinkingBudget := thinkingConfig.Get("thinkingBudget")
-					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				if !thinkingBudget.Exists() {
+					thinkingBudget = thinkingConfig.Get("thinking_budget")
+				}
+				if thinkingBudget.Exists() {
+					if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+						out, _ = sjson.Set(out, "reasoning_effort", effort)
+					}
 				}
 			}
 		}

internal/util/gemini_schema.go

@@ -667,6 +667,9 @@ func orDefault(val, def string) string {
 }
 
 func escapeGJSONPathKey(key string) string {
+	if strings.IndexAny(key, ".*?") == -1 {
+		return key
+	}
 	return gjsonPathKeyReplacer.Replace(key)
 }
 

internal/util/translator.go

@@ -6,7 +6,6 @@ package util
 import (
 	"bytes"
 	"fmt"
-	"strings"
 
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -33,15 +32,15 @@ func Walk(value gjson.Result, path, field string, paths *[]string) {
 			// . -> \.
 			// * -> \*
 			// ? -> \?
-			var keyReplacer = strings.NewReplacer(".", "\\.", "*", "\\*", "?", "\\?")
+			keyStr := key.String()
-			safeKey := keyReplacer.Replace(key.String())
+			safeKey := escapeGJSONPathKey(keyStr)
 
 			if path == "" {
 				childPath = safeKey
 			} else {
 				childPath = path + "." + safeKey
 			}
-			if key.String() == field {
+			if keyStr == field {
 				*paths = append(*paths, childPath)
 			}
 			Walk(val, childPath, field, paths)
@@ -87,15 +86,6 @@ func RenameKey(jsonStr, oldKeyPath, newKeyPath string) (string, error) {
 	return finalJson, nil
 }
 
-func DeleteKey(jsonStr, keyName string) string {
-	paths := make([]string, 0)
-	Walk(gjson.Parse(jsonStr), "", keyName, &paths)
-	for _, p := range paths {
-		jsonStr, _ = sjson.Delete(jsonStr, p)
-	}
-	return jsonStr
-}
-
 // FixJSON converts non-standard JSON that uses single quotes for strings into
 // RFC 8259-compliant JSON by converting those single-quoted strings to
 // double-quoted strings with proper escaping.

internal/watcher/clients.go

@@ -6,6 +6,7 @@ import (
 	"context"
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json"
 	"fmt"
 	"io/fs"
 	"os"
@@ -15,6 +16,7 @@ import (
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	log "github.com/sirupsen/logrus"
 )
@@ -72,6 +74,7 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 		w.clientsMutex.Lock()
 
 		w.lastAuthHashes = make(map[string]string)
+		w.lastAuthContents = make(map[string]*coreauth.Auth)
 		if resolvedAuthDir, errResolveAuthDir := util.ResolveAuthDir(cfg.AuthDir); errResolveAuthDir != nil {
 			log.Errorf("failed to resolve auth directory for hash cache: %v", errResolveAuthDir)
 		} else if resolvedAuthDir != "" {
@@ -84,6 +87,11 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 						sum := sha256.Sum256(data)
 						normalizedPath := w.normalizeAuthPath(path)
 						w.lastAuthHashes[normalizedPath] = hex.EncodeToString(sum[:])
+						// Parse and cache auth content for future diff comparisons
+						var auth coreauth.Auth
+						if errParse := json.Unmarshal(data, &auth); errParse == nil {
+							w.lastAuthContents[normalizedPath] = &auth
+						}
 					}
 				}
 				return nil
@@ -127,6 +135,13 @@ func (w *Watcher) addOrUpdateClient(path string) {
 	curHash := hex.EncodeToString(sum[:])
 	normalized := w.normalizeAuthPath(path)
 
+	// Parse new auth content for diff comparison
+	var newAuth coreauth.Auth
+	if errParse := json.Unmarshal(data, &newAuth); errParse != nil {
+		log.Errorf("failed to parse auth file %s: %v", filepath.Base(path), errParse)
+		return
+	}
+
 	w.clientsMutex.Lock()
 
 	cfg := w.config
@@ -141,7 +156,26 @@ func (w *Watcher) addOrUpdateClient(path string) {
 		return
 	}
 
+	// Get old auth for diff comparison
+	var oldAuth *coreauth.Auth
+	if w.lastAuthContents != nil {
+		oldAuth = w.lastAuthContents[normalized]
+	}
+
+	// Compute and log field changes
+	if changes := diff.BuildAuthChangeDetails(oldAuth, &newAuth); len(changes) > 0 {
+		log.Debugf("auth field changes for %s:", filepath.Base(path))
+		for _, c := range changes {
+			log.Debugf("  %s", c)
+		}
+	}
+
+	// Update caches
 	w.lastAuthHashes[normalized] = curHash
+	if w.lastAuthContents == nil {
+		w.lastAuthContents = make(map[string]*coreauth.Auth)
+	}
+	w.lastAuthContents[normalized] = &newAuth
 
 	w.clientsMutex.Unlock() // Unlock before the callback
 
@@ -160,6 +194,7 @@ func (w *Watcher) removeClient(path string) {
 
 	cfg := w.config
 	delete(w.lastAuthHashes, normalized)
+	delete(w.lastAuthContents, normalized)
 
 	w.clientsMutex.Unlock() // Release the lock before the callback
 

internal/watcher/diff/auth_diff.go

@@ -0,0 +1,44 @@
+// auth_diff.go computes human-readable diffs for auth file field changes.
+package diff
+
+import (
+	"fmt"
+	"strings"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+// BuildAuthChangeDetails computes a redacted, human-readable list of auth field changes.
+// Only prefix, proxy_url, and disabled fields are tracked; sensitive data is never printed.
+func BuildAuthChangeDetails(oldAuth, newAuth *coreauth.Auth) []string {
+	changes := make([]string, 0, 3)
+
+	// Handle nil cases by using empty Auth as default
+	if oldAuth == nil {
+		oldAuth = &coreauth.Auth{}
+	}
+	if newAuth == nil {
+		return changes
+	}
+
+	// Compare prefix
+	oldPrefix := strings.TrimSpace(oldAuth.Prefix)
+	newPrefix := strings.TrimSpace(newAuth.Prefix)
+	if oldPrefix != newPrefix {
+		changes = append(changes, fmt.Sprintf("prefix: %s -> %s", oldPrefix, newPrefix))
+	}
+
+	// Compare proxy_url (redacted)
+	oldProxy := strings.TrimSpace(oldAuth.ProxyURL)
+	newProxy := strings.TrimSpace(newAuth.ProxyURL)
+	if oldProxy != newProxy {
+		changes = append(changes, fmt.Sprintf("proxy_url: %s -> %s", formatProxyURL(oldProxy), formatProxyURL(newProxy)))
+	}
+
+	// Compare disabled
+	if oldAuth.Disabled != newAuth.Disabled {
+		changes = append(changes, fmt.Sprintf("disabled: %t -> %t", oldAuth.Disabled, newAuth.Disabled))
+	}
+
+	return changes
+}

internal/watcher/diff/config_diff.go

@@ -27,6 +27,12 @@ func BuildConfigChangeDetails(oldCfg, newCfg *config.Config) []string {
 	if oldCfg.Debug != newCfg.Debug {
 		changes = append(changes, fmt.Sprintf("debug: %t -> %t", oldCfg.Debug, newCfg.Debug))
 	}
+	if oldCfg.Pprof.Enable != newCfg.Pprof.Enable {
+		changes = append(changes, fmt.Sprintf("pprof.enable: %t -> %t", oldCfg.Pprof.Enable, newCfg.Pprof.Enable))
+	}
+	if strings.TrimSpace(oldCfg.Pprof.Addr) != strings.TrimSpace(newCfg.Pprof.Addr) {
+		changes = append(changes, fmt.Sprintf("pprof.addr: %s -> %s", strings.TrimSpace(oldCfg.Pprof.Addr), strings.TrimSpace(newCfg.Pprof.Addr)))
+	}
 	if oldCfg.LoggingToFile != newCfg.LoggingToFile {
 		changes = append(changes, fmt.Sprintf("logging-to-file: %t -> %t", oldCfg.LoggingToFile, newCfg.LoggingToFile))
 	}

internal/watcher/watcher.go

@@ -38,6 +38,7 @@ type Watcher struct {
 	reloadCallback    func(*config.Config)
 	watcher           *fsnotify.Watcher
 	lastAuthHashes    map[string]string
+	lastAuthContents  map[string]*coreauth.Auth
 	lastRemoveTimes   map[string]time.Time
 	lastConfigHash    string
 	authQueue         chan<- AuthUpdate

sdk/api/handlers/handlers.go

@@ -155,20 +155,6 @@ func requestExecutionMetadata(ctx context.Context) map[string]any {
 	return map[string]any{idempotencyKeyMetadataKey: key}
 }
 
-func mergeMetadata(base, overlay map[string]any) map[string]any {
-	if len(base) == 0 && len(overlay) == 0 {
-		return nil
-	}
-	out := make(map[string]any, len(base)+len(overlay))
-	for k, v := range base {
-		out[k] = v
-	}
-	for k, v := range overlay {
-		out[k] = v
-	}
-	return out
-}
-
 // BaseAPIHandler contains the handlers for API endpoints.
 // It holds a pool of clients to interact with the backend service and manages
 // load balancing, client selection, and configuration.
@@ -399,7 +385,7 @@ func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType
 	opts := coreexecutor.Options{
 		Stream:          false,
 		Alt:             alt,
-		OriginalRequest: cloneBytes(rawJSON),
+		OriginalRequest: rawJSON,
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
@@ -438,7 +424,7 @@ func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handle
 	opts := coreexecutor.Options{
 		Stream:          false,
 		Alt:             alt,
-		OriginalRequest: cloneBytes(rawJSON),
+		OriginalRequest: rawJSON,
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
@@ -480,7 +466,7 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 	opts := coreexecutor.Options{
 		Stream:          true,
 		Alt:             alt,
-		OriginalRequest: cloneBytes(rawJSON),
+		OriginalRequest: rawJSON,
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
@@ -669,17 +655,6 @@ func cloneBytes(src []byte) []byte {
 	return dst
 }
 
-func cloneMetadata(src map[string]any) map[string]any {
-	if len(src) == 0 {
-		return nil
-	}
-	dst := make(map[string]any, len(src))
-	for k, v := range src {
-		dst[k] = v
-	}
-	return dst
-}
-
 // WriteErrorResponse writes an error message to the response writer using the HTTP status embedded in the message.
 func (h *BaseAPIHandler) WriteErrorResponse(c *gin.Context, msg *interfaces.ErrorMessage) {
 	status := http.StatusInternalServerError

sdk/cliproxy/pprof_server.go

@@ -0,0 +1,163 @@
+package cliproxy
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/pprof"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	log "github.com/sirupsen/logrus"
+)
+
+type pprofServer struct {
+	mu      sync.Mutex
+	server  *http.Server
+	addr    string
+	enabled bool
+}
+
+func newPprofServer() *pprofServer {
+	return &pprofServer{}
+}
+
+func (s *Service) applyPprofConfig(cfg *config.Config) {
+	if s == nil || cfg == nil {
+		return
+	}
+	if s.pprofServer == nil {
+		s.pprofServer = newPprofServer()
+	}
+	s.pprofServer.Apply(cfg)
+}
+
+func (s *Service) shutdownPprof(ctx context.Context) error {
+	if s == nil || s.pprofServer == nil {
+		return nil
+	}
+	return s.pprofServer.Shutdown(ctx)
+}
+
+func (p *pprofServer) Apply(cfg *config.Config) {
+	if p == nil || cfg == nil {
+		return
+	}
+	addr := strings.TrimSpace(cfg.Pprof.Addr)
+	if addr == "" {
+		addr = config.DefaultPprofAddr
+	}
+	enabled := cfg.Pprof.Enable
+
+	p.mu.Lock()
+	currentServer := p.server
+	currentAddr := p.addr
+	p.addr = addr
+	p.enabled = enabled
+	if !enabled {
+		p.server = nil
+		p.mu.Unlock()
+		if currentServer != nil {
+			p.stopServer(currentServer, currentAddr, "disabled")
+		}
+		return
+	}
+	if currentServer != nil && currentAddr == addr {
+		p.mu.Unlock()
+		return
+	}
+	p.server = nil
+	p.mu.Unlock()
+
+	if currentServer != nil {
+		p.stopServer(currentServer, currentAddr, "restarted")
+	}
+
+	p.startServer(addr)
+}
+
+func (p *pprofServer) Shutdown(ctx context.Context) error {
+	if p == nil {
+		return nil
+	}
+	p.mu.Lock()
+	currentServer := p.server
+	currentAddr := p.addr
+	p.server = nil
+	p.enabled = false
+	p.mu.Unlock()
+
+	if currentServer == nil {
+		return nil
+	}
+	return p.stopServerWithContext(ctx, currentServer, currentAddr, "shutdown")
+}
+
+func (p *pprofServer) startServer(addr string) {
+	mux := newPprofMux()
+	server := &http.Server{
+		Addr:              addr,
+		Handler:           mux,
+		ReadHeaderTimeout: 5 * time.Second,
+	}
+
+	p.mu.Lock()
+	if !p.enabled || p.addr != addr || p.server != nil {
+		p.mu.Unlock()
+		return
+	}
+	p.server = server
+	p.mu.Unlock()
+
+	log.Infof("pprof server starting on %s", addr)
+	go func() {
+		if errServe := server.ListenAndServe(); errServe != nil && !errors.Is(errServe, http.ErrServerClosed) {
+			log.Errorf("pprof server failed on %s: %v", addr, errServe)
+			p.mu.Lock()
+			if p.server == server {
+				p.server = nil
+			}
+			p.mu.Unlock()
+		}
+	}()
+}
+
+func (p *pprofServer) stopServer(server *http.Server, addr string, reason string) {
+	_ = p.stopServerWithContext(context.Background(), server, addr, reason)
+}
+
+func (p *pprofServer) stopServerWithContext(ctx context.Context, server *http.Server, addr string, reason string) error {
+	if server == nil {
+		return nil
+	}
+	stopCtx := ctx
+	if stopCtx == nil {
+		stopCtx = context.Background()
+	}
+	stopCtx, cancel := context.WithTimeout(stopCtx, 5*time.Second)
+	defer cancel()
+	if errStop := server.Shutdown(stopCtx); errStop != nil {
+		log.Errorf("pprof server stop failed on %s: %v", addr, errStop)
+		return errStop
+	}
+	log.Infof("pprof server stopped on %s (%s)", addr, reason)
+	return nil
+}
+
+func newPprofMux() *http.ServeMux {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/debug/pprof/", pprof.Index)
+	mux.HandleFunc("/debug/pprof/cmdline", pprof.Cmdline)
+	mux.HandleFunc("/debug/pprof/profile", pprof.Profile)
+	mux.HandleFunc("/debug/pprof/symbol", pprof.Symbol)
+	mux.HandleFunc("/debug/pprof/trace", pprof.Trace)
+	mux.Handle("/debug/pprof/allocs", pprof.Handler("allocs"))
+	mux.Handle("/debug/pprof/block", pprof.Handler("block"))
+	mux.Handle("/debug/pprof/goroutine", pprof.Handler("goroutine"))
+	mux.Handle("/debug/pprof/heap", pprof.Handler("heap"))
+	mux.Handle("/debug/pprof/mutex", pprof.Handler("mutex"))
+	mux.Handle("/debug/pprof/threadcreate", pprof.Handler("threadcreate"))
+	return mux
+}

sdk/cliproxy/service.go

@@ -58,6 +58,9 @@ type Service struct {
 	// server is the HTTP API server instance.
 	server *api.Server
 
+	// pprofServer manages the optional pprof HTTP debug server.
+	pprofServer *pprofServer
+
 	// serverErr channel for server startup/shutdown errors.
 	serverErr chan error
 
@@ -281,27 +284,42 @@ func (s *Service) wsOnDisconnected(channelID string, reason error) {
 }
 
 func (s *Service) applyCoreAuthAddOrUpdate(ctx context.Context, auth *coreauth.Auth) {
-	if s == nil || auth == nil || auth.ID == "" {
+	if s == nil || s.coreManager == nil || auth == nil || auth.ID == "" {
-		return
-	}
-	if s.coreManager == nil {
 		return
 	}
 	auth = auth.Clone()
 	s.ensureExecutorsForAuth(auth)
-	s.registerModelsForAuth(auth)
+
-	if existing, ok := s.coreManager.GetByID(auth.ID); ok && existing != nil {
+	// IMPORTANT: Update coreManager FIRST, before model registration.
+	// This ensures that configuration changes (proxy_url, prefix, etc.) take effect
+	// immediately for API calls, rather than waiting for model registration to complete.
+	// Model registration may involve network calls (e.g., FetchAntigravityModels) that
+	// could timeout if the new proxy_url is unreachable.
+	op := "register"
+	var err error
+	if existing, ok := s.coreManager.GetByID(auth.ID); ok {
 		auth.CreatedAt = existing.CreatedAt
 		auth.LastRefreshedAt = existing.LastRefreshedAt
 		auth.NextRefreshAfter = existing.NextRefreshAfter
-		if _, err := s.coreManager.Update(ctx, auth); err != nil {
+		op = "update"
-			log.Errorf("failed to update auth %s: %v", auth.ID, err)
+		_, err = s.coreManager.Update(ctx, auth)
+	} else {
+		_, err = s.coreManager.Register(ctx, auth)
+	}
+	if err != nil {
+		log.Errorf("failed to %s auth %s: %v", op, auth.ID, err)
+		current, ok := s.coreManager.GetByID(auth.ID)
+		if !ok || current.Disabled {
+			GlobalModelRegistry().UnregisterClient(auth.ID)
+			return
 		}
-		return
+		auth = current
-	}
-	if _, err := s.coreManager.Register(ctx, auth); err != nil {
-		log.Errorf("failed to register auth %s: %v", auth.ID, err)
 	}
+
+	// Register models after auth is updated in coreManager.
+	// This operation may block on network calls, but the auth configuration
+	// is already effective at this point.
+	s.registerModelsForAuth(auth)
 }
 
 func (s *Service) applyCoreAuthRemoval(ctx context.Context, id string) {
@@ -516,6 +534,8 @@ func (s *Service) Run(ctx context.Context) error {
 	time.Sleep(100 * time.Millisecond)
 	fmt.Printf("API server started successfully on: %s:%d\n", s.cfg.Host, s.cfg.Port)
 
+	s.applyPprofConfig(s.cfg)
+
 	if s.hooks.OnAfterStart != nil {
 		s.hooks.OnAfterStart(s)
 	}
@@ -561,6 +581,7 @@ func (s *Service) Run(ctx context.Context) error {
 		}
 
 		s.applyRetryConfig(newCfg)
+		s.applyPprofConfig(newCfg)
 		if s.server != nil {
 			s.server.UpdateClients(newCfg)
 		}
@@ -666,6 +687,13 @@ func (s *Service) Shutdown(ctx context.Context) error {
 			s.authQueueStop = nil
 		}
 
+		if errShutdownPprof := s.shutdownPprof(ctx); errShutdownPprof != nil {
+			log.Errorf("failed to stop pprof server: %v", errShutdownPprof)
+			if shutdownErr == nil {
+				shutdownErr = errShutdownPprof
+			}
+		}
+
 		// no legacy clients to persist
 
 		if s.server != nil {

test/builtin_tools_translation_test.go

@@ -33,7 +33,7 @@ func TestOpenAIToCodex_PreservesBuiltinTools(t *testing.T) {
 	}
 }
 
-func TestOpenAIResponsesToOpenAI_PreservesBuiltinTools(t *testing.T) {
+func TestOpenAIResponsesToOpenAI_IgnoresBuiltinTools(t *testing.T) {
 	in := []byte(`{
 		"model":"gpt-5",
 		"input":[{"role":"user","content":[{"type":"input_text","text":"hi"}]}],
@@ -42,13 +42,7 @@ func TestOpenAIResponsesToOpenAI_PreservesBuiltinTools(t *testing.T) {
 
 	out := sdktranslator.TranslateRequest(sdktranslator.FormatOpenAIResponse, sdktranslator.FormatOpenAI, "gpt-5", in, false)
 
-	if got := gjson.GetBytes(out, "tools.#").Int(); got != 1 {
+	if got := gjson.GetBytes(out, "tools.#").Int(); got != 0 {
-		t.Fatalf("expected 1 tool, got %d: %s", got, string(out))
+		t.Fatalf("expected 0 tools (builtin tools not supported in Chat Completions), got %d: %s", got, string(out))
-	}
-	if got := gjson.GetBytes(out, "tools.0.type").String(); got != "web_search" {
-		t.Fatalf("expected tools[0].type=web_search, got %q: %s", got, string(out))
-	}
-	if got := gjson.GetBytes(out, "tools.0.search_context_size").String(); got != "low" {
-		t.Fatalf("expected tools[0].search_context_size=low, got %q: %s", got, string(out))
 	}
 }