Merge branch 'router-for-me:main' into main

fix: add default copilot claude model aliases for oauth routing

config.example.yaml

@@ -73,6 +73,10 @@ proxy-url: ''
 # When true, unprefixed model requests only use credentials without a prefix (except when prefix == model name).
 force-model-prefix: false
 
+# When true, forward filtered upstream response headers to downstream clients.
+# Default is false (disabled).
+passthrough-headers: false
+
 # Number of times to retry a request. Retries will occur if the HTTP response code is 403, 408, 500, 502, 503, or 504.
 request-retry: 3
 
@@ -160,6 +164,7 @@ nonstream-keepalive-interval: 0
 #       sensitive-words:             # optional: words to obfuscate with zero-width characters
 #         - "API"
 #         - "proxy"
+#       cache-user-id: true          # optional: default is false; set true to reuse cached user_id per API key instead of generating a random one each request
 
 # Default headers for Claude API requests. Update when Claude Code releases new versions.
 # These are used as fallbacks when the client does not send its own headers.

examples/custom-provider/main.go

@@ -159,13 +159,13 @@ func (MyExecutor) CountTokens(context.Context, *coreauth.Auth, clipexec.Request,
 	return clipexec.Response{}, errors.New("count tokens not implemented")
 }
 
-func (MyExecutor) ExecuteStream(ctx context.Context, a *coreauth.Auth, req clipexec.Request, opts clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+func (MyExecutor) ExecuteStream(ctx context.Context, a *coreauth.Auth, req clipexec.Request, opts clipexec.Options) (*clipexec.StreamResult, error) {
 	ch := make(chan clipexec.StreamChunk, 1)
 	go func() {
 		defer close(ch)
 		ch <- clipexec.StreamChunk{Payload: []byte("data: {\"ok\":true}\n\n")}
 	}()
-	return ch, nil
+	return &clipexec.StreamResult{Chunks: ch}, nil
 }
 
 func (MyExecutor) Refresh(ctx context.Context, a *coreauth.Auth) (*coreauth.Auth, error) {

examples/http-request/main.go

@@ -58,7 +58,7 @@ func (EchoExecutor) Execute(context.Context, *coreauth.Auth, clipexec.Request, c
 	return clipexec.Response{}, errors.New("echo executor: Execute not implemented")
 }
 
-func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (*clipexec.StreamResult, error) {
 	return nil, errors.New("echo executor: ExecuteStream not implemented")
 }
 

internal/auth/claude/anthropic_auth.go

@@ -20,7 +20,7 @@ import (
 // OAuth configuration constants for Claude/Anthropic
 const (
 	AuthURL     = "https://claude.ai/oauth/authorize"
-	TokenURL    = "https://console.anthropic.com/v1/oauth/token"
+	TokenURL    = "https://api.anthropic.com/v1/oauth/token"
 	ClientID    = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
 	RedirectURI = "http://localhost:54545/callback"
 )

internal/config/config.go

@@ -314,6 +314,10 @@ type CloakConfig struct {
 	// SensitiveWords is a list of words to obfuscate with zero-width characters.
 	// This can help bypass certain content filters.
 	SensitiveWords []string `yaml:"sensitive-words,omitempty" json:"sensitive-words,omitempty"`
+
+	// CacheUserID controls whether Claude user_id values are cached per API key.
+	// When false, a fresh random user_id is generated for every request.
+	CacheUserID *bool `yaml:"cache-user-id,omitempty" json:"cache-user-id,omitempty"`
 }
 
 // ClaudeKey represents the configuration for a Claude API key,
@@ -759,22 +763,24 @@ func (cfg *Config) SanitizeOAuthModelAlias() {
 		return
 	}
 
-	// Inject default Kiro aliases if no user-configured kiro aliases exist
+	// Inject channel defaults when the channel is absent in user config.
+	// Presence is checked case-insensitively and includes explicit nil/empty markers.
 	if cfg.OAuthModelAlias == nil {
 		cfg.OAuthModelAlias = make(map[string][]OAuthModelAlias)
 	}
-	if _, hasKiro := cfg.OAuthModelAlias["kiro"]; !hasKiro {
-		// Check case-insensitive too
-		found := false
+	hasChannel := func(channel string) bool {
 		for k := range cfg.OAuthModelAlias {
-			if strings.EqualFold(strings.TrimSpace(k), "kiro") {
-				found = true
-				break
+			if strings.EqualFold(strings.TrimSpace(k), channel) {
+				return true
 			}
 		}
-		if !found {
-			cfg.OAuthModelAlias["kiro"] = defaultKiroAliases()
-		}
+		return false
+	}
+	if !hasChannel("kiro") {
+		cfg.OAuthModelAlias["kiro"] = defaultKiroAliases()
+	}
+	if !hasChannel("github-copilot") {
+		cfg.OAuthModelAlias["github-copilot"] = defaultGitHubCopilotAliases()
 	}
 
 	if len(cfg.OAuthModelAlias) == 0 {

internal/config/oauth_model_alias_migration.go

@@ -42,6 +42,21 @@ func defaultKiroAliases() []OAuthModelAlias {
 	}
 }
 
+// defaultGitHubCopilotAliases returns default oauth-model-alias entries that
+// expose Claude hyphen-style IDs for GitHub Copilot Claude models.
+// This keeps compatibility with clients (e.g. Claude Code) that use
+// Anthropic-style model IDs like "claude-opus-4-6".
+func defaultGitHubCopilotAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		{Name: "claude-haiku-4.5", Alias: "claude-haiku-4-5", Fork: true},
+		{Name: "claude-opus-4.1", Alias: "claude-opus-4-1", Fork: true},
+		{Name: "claude-opus-4.5", Alias: "claude-opus-4-5", Fork: true},
+		{Name: "claude-opus-4.6", Alias: "claude-opus-4-6", Fork: true},
+		{Name: "claude-sonnet-4.5", Alias: "claude-sonnet-4-5", Fork: true},
+		{Name: "claude-sonnet-4.6", Alias: "claude-sonnet-4-6", Fork: true},
+	}
+}
+
 // defaultAntigravityAliases returns the default oauth-model-alias configuration
 // for the antigravity channel when neither field exists.
 func defaultAntigravityAliases() []OAuthModelAlias {

internal/config/oauth_model_alias_test.go

@@ -107,6 +107,44 @@ func TestSanitizeOAuthModelAlias_InjectsDefaultKiroAliases(t *testing.T) {
 	}
 }
 
+func TestSanitizeOAuthModelAlias_InjectsDefaultGitHubCopilotAliases(t *testing.T) {
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"codex": {
+				{Name: "gpt-5", Alias: "g5"},
+			},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	copilotAliases := cfg.OAuthModelAlias["github-copilot"]
+	if len(copilotAliases) == 0 {
+		t.Fatal("expected default github-copilot aliases to be injected")
+	}
+
+	aliasSet := make(map[string]bool, len(copilotAliases))
+	for _, a := range copilotAliases {
+		aliasSet[a.Alias] = true
+		if !a.Fork {
+			t.Fatalf("expected all default github-copilot aliases to have fork=true, got fork=false for %q", a.Alias)
+		}
+	}
+	expectedAliases := []string{
+		"claude-haiku-4-5",
+		"claude-opus-4-1",
+		"claude-opus-4-5",
+		"claude-opus-4-6",
+		"claude-sonnet-4-5",
+		"claude-sonnet-4-6",
+	}
+	for _, expected := range expectedAliases {
+		if !aliasSet[expected] {
+			t.Fatalf("expected default github-copilot alias %q to be present", expected)
+		}
+	}
+}
+
 func TestSanitizeOAuthModelAlias_DoesNotOverrideUserKiroAliases(t *testing.T) {
 	// When user has configured kiro aliases, defaults should NOT be injected
 	cfg := &Config{
@@ -128,6 +166,26 @@ func TestSanitizeOAuthModelAlias_DoesNotOverrideUserKiroAliases(t *testing.T) {
 	}
 }
 
+func TestSanitizeOAuthModelAlias_DoesNotOverrideUserGitHubCopilotAliases(t *testing.T) {
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"github-copilot": {
+				{Name: "claude-opus-4.6", Alias: "my-opus", Fork: true},
+			},
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	copilotAliases := cfg.OAuthModelAlias["github-copilot"]
+	if len(copilotAliases) != 1 {
+		t.Fatalf("expected 1 user-configured github-copilot alias, got %d", len(copilotAliases))
+	}
+	if copilotAliases[0].Alias != "my-opus" {
+		t.Fatalf("expected user alias to be preserved, got %q", copilotAliases[0].Alias)
+	}
+}
+
 func TestSanitizeOAuthModelAlias_DoesNotReinjectAfterExplicitDeletion(t *testing.T) {
 	// When user explicitly deletes kiro aliases (key exists with nil value),
 	// defaults should NOT be re-injected on subsequent sanitize calls (#222).
@@ -154,6 +212,24 @@ func TestSanitizeOAuthModelAlias_DoesNotReinjectAfterExplicitDeletion(t *testing
 	}
 }
 
+func TestSanitizeOAuthModelAlias_GitHubCopilotDoesNotReinjectAfterExplicitDeletion(t *testing.T) {
+	cfg := &Config{
+		OAuthModelAlias: map[string][]OAuthModelAlias{
+			"github-copilot": nil, // explicitly deleted
+		},
+	}
+
+	cfg.SanitizeOAuthModelAlias()
+
+	copilotAliases := cfg.OAuthModelAlias["github-copilot"]
+	if len(copilotAliases) != 0 {
+		t.Fatalf("expected github-copilot aliases to remain empty after explicit deletion, got %d aliases", len(copilotAliases))
+	}
+	if _, exists := cfg.OAuthModelAlias["github-copilot"]; !exists {
+		t.Fatal("expected github-copilot key to be preserved as nil marker after sanitization")
+	}
+}
+
 func TestSanitizeOAuthModelAlias_DoesNotReinjectAfterExplicitDeletionEmpty(t *testing.T) {
 	// Same as above but with empty slice instead of nil (PUT with empty body).
 	cfg := &Config{

internal/config/sdk_config.go

@@ -20,6 +20,10 @@ type SDKConfig struct {
 	// APIKeys is a list of keys for authenticating clients to this proxy server.
 	APIKeys []string `yaml:"api-keys" json:"api-keys"`
 
+	// PassthroughHeaders controls whether upstream response headers are forwarded to downstream clients.
+	// Default is false (disabled).
+	PassthroughHeaders bool `yaml:"passthrough-headers" json:"passthrough-headers"`
+
 	// Streaming configures server-side streaming behavior (keep-alives and safe bootstrap retries).
 	Streaming StreamingConfig `yaml:"streaming" json:"streaming"`
 

internal/registry/model_definitions.go

@@ -129,7 +129,19 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 // These models are available through the GitHub Copilot API at api.githubcopilot.com.
 func GetGitHubCopilotModels() []*ModelInfo {
 	now := int64(1732752000) // 2024-11-27
-	return []*ModelInfo{
+	gpt4oEntries := []struct {
+		ID          string
+		DisplayName string
+		Description string
+	}{
+		{ID: "gpt-4o-2024-11-20", DisplayName: "GPT-4o (2024-11-20)", Description: "OpenAI GPT-4o 2024-11-20 via GitHub Copilot"},
+		{ID: "gpt-4o-2024-08-06", DisplayName: "GPT-4o (2024-08-06)", Description: "OpenAI GPT-4o 2024-08-06 via GitHub Copilot"},
+		{ID: "gpt-4o-2024-05-13", DisplayName: "GPT-4o (2024-05-13)", Description: "OpenAI GPT-4o 2024-05-13 via GitHub Copilot"},
+		{ID: "gpt-4o", DisplayName: "GPT-4o", Description: "OpenAI GPT-4o via GitHub Copilot"},
+		{ID: "gpt-4-o-preview", DisplayName: "GPT-4-o Preview", Description: "OpenAI GPT-4-o Preview via GitHub Copilot"},
+	}
+
+	models := []*ModelInfo{
 		{
 			ID:                  "gpt-4.1",
 			Object:              "model",
@@ -141,6 +153,23 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       128000,
 			MaxCompletionTokens: 16384,
 		},
+	}
+
+	for _, entry := range gpt4oEntries {
+		models = append(models, &ModelInfo{
+			ID:                  entry.ID,
+			Object:              "model",
+			Created:             now,
+			OwnedBy:             "github-copilot",
+			Type:                "github-copilot",
+			DisplayName:         entry.DisplayName,
+			Description:         entry.Description,
+			ContextLength:       128000,
+			MaxCompletionTokens: 16384,
+		})
+	}
+
+	return append(models, []*ModelInfo{
 		{
 			ID:                  "gpt-5",
 			Object:              "model",
@@ -377,6 +406,17 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			ContextLength:       1048576,
 			MaxCompletionTokens: 65536,
 		},
+		{
+			ID:                  "gemini-3.1-pro-preview",
+			Object:              "model",
+			Created:             now,
+			OwnedBy:             "github-copilot",
+			Type:                "github-copilot",
+			DisplayName:         "Gemini 3.1 Pro (Preview)",
+			Description:         "Google Gemini 3.1 Pro Preview via GitHub Copilot",
+			ContextLength:       1048576,
+			MaxCompletionTokens: 65536,
+		},
 		{
 			ID:                  "gemini-3-flash-preview",
 			Object:              "model",
@@ -411,7 +451,7 @@ func GetGitHubCopilotModels() []*ModelInfo {
 			MaxCompletionTokens: 16384,
 			SupportedEndpoints:  []string{"/chat/completions", "/responses"},
 		},
-	}
+	}...)
 }
 
 // GetKiroModels returns the Kiro (AWS CodeWhisperer) model definitions

internal/registry/model_definitions_static_data.go

@@ -196,6 +196,21 @@ func GetGeminiModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",
@@ -306,6 +321,21 @@ func GetGeminiVertexModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
 			Object:                     "model",
@@ -448,6 +478,21 @@ func GetGeminiCLIModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",
@@ -529,6 +574,21 @@ func GetAIStudioModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",

internal/runtime/executor/aistudio_executor.go

@@ -164,12 +164,12 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 	reporter.publish(ctx, parseGeminiUsage(wsResp.Body))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, body.toFormat, opts.SourceFormat, req.Model, opts.OriginalRequest, translatedReq, wsResp.Body, &param)
-	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out))}
+	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out)), Headers: wsResp.Headers.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming request to the AI Studio API.
-func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -254,7 +254,6 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 		return nil, statusErr{code: firstEvent.Status, msg: body.String()}
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func(first wsrelay.StreamEvent) {
 		defer close(out)
 		var param any
@@ -318,7 +317,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 			}
 		}
 	}(firstEvent)
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: firstEvent.Headers.Clone(), Chunks: out}, nil
 }
 
 // CountTokens counts tokens for the given request using the AI Studio API.

internal/runtime/executor/antigravity_executor.go

@@ -232,7 +232,7 @@ attemptLoop:
 			reporter.publish(ctx, parseAntigravityUsage(bodyBytes))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bodyBytes, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			resp = cliproxyexecutor.Response{Payload: []byte(converted), Headers: httpResp.Header.Clone()}
 			reporter.ensurePublished(ctx)
 			return resp, nil
 		}
@@ -436,7 +436,7 @@ attemptLoop:
 			reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, resp.Payload, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			resp = cliproxyexecutor.Response{Payload: []byte(converted), Headers: httpResp.Header.Clone()}
 			reporter.ensurePublished(ctx)
 
 			return resp, nil
@@ -645,7 +645,7 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 }
 
 // ExecuteStream performs a streaming request to the Antigravity API.
-func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -775,7 +775,6 @@ attemptLoop:
 			}
 
 			out := make(chan cliproxyexecutor.StreamChunk)
-			stream = out
 			go func(resp *http.Response) {
 				defer close(out)
 				defer func() {
@@ -820,7 +819,7 @@ attemptLoop:
 					reporter.ensurePublished(ctx)
 				}
 			}(httpResp)
-			return stream, nil
+			return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 		}
 
 		switch {
@@ -968,7 +967,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 		if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
 			count := gjson.GetBytes(bodyBytes, "totalTokens").Int()
 			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, bodyBytes)
-			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+			return cliproxyexecutor.Response{Payload: []byte(translated), Headers: httpResp.Header.Clone()}, nil
 		}
 
 		lastStatus = httpResp.StatusCode

internal/runtime/executor/claude_executor.go

@@ -117,7 +117,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 
 	// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
 	// based on client type and configuration.
-	body = applyCloaking(ctx, e.cfg, auth, body, baseModel)
+	body = applyCloaking(ctx, e.cfg, auth, body, baseModel, apiKey)
 
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
@@ -223,11 +223,11 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		data,
 		&param,
 	)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -258,7 +258,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 
 	// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
 	// based on client type and configuration.
-	body = applyCloaking(ctx, e.cfg, auth, body, baseModel)
+	body = applyCloaking(ctx, e.cfg, auth, body, baseModel, apiKey)
 
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
@@ -330,7 +330,6 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -399,7 +398,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
@@ -488,7 +487,7 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "input_tokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: resp.Header.Clone()}, nil
 }
 
 func (e *ClaudeExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
@@ -983,10 +982,10 @@ func getClientUserAgent(ctx context.Context) string {
 }
 
 // getCloakConfigFromAuth extracts cloak configuration from auth attributes.
-// Returns (cloakMode, strictMode, sensitiveWords).
-func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string) {
+// Returns (cloakMode, strictMode, sensitiveWords, cacheUserID).
+func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string, bool) {
 	if auth == nil || auth.Attributes == nil {
-		return "auto", false, nil
+		return "auto", false, nil, false
 	}
 
 	cloakMode := auth.Attributes["cloak_mode"]
@@ -1004,7 +1003,9 @@ func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string) {
 		}
 	}
 
-	return cloakMode, strictMode, sensitiveWords
+	cacheUserID := strings.EqualFold(strings.TrimSpace(auth.Attributes["cloak_cache_user_id"]), "true")
+
+	return cloakMode, strictMode, sensitiveWords, cacheUserID
 }
 
 // resolveClaudeKeyCloakConfig finds the matching ClaudeKey config and returns its CloakConfig.
@@ -1037,16 +1038,24 @@ func resolveClaudeKeyCloakConfig(cfg *config.Config, auth *cliproxyauth.Auth) *c
 }
 
 // injectFakeUserID generates and injects a fake user ID into the request metadata.
-func injectFakeUserID(payload []byte) []byte {
+// When useCache is false, a new user ID is generated for every call.
+func injectFakeUserID(payload []byte, apiKey string, useCache bool) []byte {
+	generateID := func() string {
+		if useCache {
+			return cachedUserID(apiKey)
+		}
+		return generateFakeUserID()
+	}
+
 	metadata := gjson.GetBytes(payload, "metadata")
 	if !metadata.Exists() {
-		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateFakeUserID())
+		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateID())
 		return payload
 	}
 
 	existingUserID := gjson.GetBytes(payload, "metadata.user_id").String()
 	if existingUserID == "" || !isValidUserID(existingUserID) {
-		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateFakeUserID())
+		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateID())
 	}
 	return payload
 }
@@ -1083,7 +1092,7 @@ func checkSystemInstructionsWithMode(payload []byte, strictMode bool) []byte {
 
 // applyCloaking applies cloaking transformations to the payload based on config and client.
 // Cloaking includes: system prompt injection, fake user ID, and sensitive word obfuscation.
-func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, payload []byte, model string) []byte {
+func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, payload []byte, model string, apiKey string) []byte {
 	clientUserAgent := getClientUserAgent(ctx)
 
 	// Get cloak config from ClaudeKey configuration
@@ -1093,16 +1102,20 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 	var cloakMode string
 	var strictMode bool
 	var sensitiveWords []string
+	var cacheUserID bool
 
 	if cloakCfg != nil {
 		cloakMode = cloakCfg.Mode
 		strictMode = cloakCfg.StrictMode
 		sensitiveWords = cloakCfg.SensitiveWords
+		if cloakCfg.CacheUserID != nil {
+			cacheUserID = *cloakCfg.CacheUserID
+		}
 	}
 
 	// Fallback to auth attributes if no config found
 	if cloakMode == "" {
-		attrMode, attrStrict, attrWords := getCloakConfigFromAuth(auth)
+		attrMode, attrStrict, attrWords, attrCache := getCloakConfigFromAuth(auth)
 		cloakMode = attrMode
 		if !strictMode {
 			strictMode = attrStrict
@@ -1110,6 +1123,12 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 		if len(sensitiveWords) == 0 {
 			sensitiveWords = attrWords
 		}
+		if cloakCfg == nil || cloakCfg.CacheUserID == nil {
+			cacheUserID = attrCache
+		}
+	} else if cloakCfg == nil || cloakCfg.CacheUserID == nil {
+		_, _, _, attrCache := getCloakConfigFromAuth(auth)
+		cacheUserID = attrCache
 	}
 
 	// Determine if cloaking should be applied
@@ -1123,7 +1142,7 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 	}
 
 	// Inject fake user ID
-	payload = injectFakeUserID(payload)
+	payload = injectFakeUserID(payload, apiKey, cacheUserID)
 
 	// Apply sensitive word obfuscation
 	if len(sensitiveWords) > 0 {

internal/runtime/executor/claude_executor_test.go

@@ -2,9 +2,18 @@ package executor
 
 import (
 	"bytes"
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
 	"testing"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
 	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
 func TestApplyClaudeToolPrefix(t *testing.T) {
@@ -199,6 +208,119 @@ func TestApplyClaudeToolPrefix_NestedToolReference(t *testing.T) {
 	}
 }
 
+func TestClaudeExecutor_ReusesUserIDAcrossModelsWhenCacheEnabled(t *testing.T) {
+	resetUserIDCache()
+
+	var userIDs []string
+	var requestModels []string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		userID := gjson.GetBytes(body, "metadata.user_id").String()
+		model := gjson.GetBytes(body, "model").String()
+		userIDs = append(userIDs, userID)
+		requestModels = append(requestModels, model)
+		t.Logf("HTTP Server received request: model=%s, user_id=%s, url=%s", model, userID, r.URL.String())
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","type":"message","model":"claude-3-5-sonnet","role":"assistant","content":[{"type":"text","text":"ok"}],"usage":{"input_tokens":1,"output_tokens":1}}`))
+	}))
+	defer server.Close()
+
+	t.Logf("End-to-end test: Fake HTTP server started at %s", server.URL)
+
+	cacheEnabled := true
+	executor := NewClaudeExecutor(&config.Config{
+		ClaudeKey: []config.ClaudeKey{
+			{
+				APIKey:  "key-123",
+				BaseURL: server.URL,
+				Cloak: &config.CloakConfig{
+					CacheUserID: &cacheEnabled,
+				},
+			},
+		},
+	})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+	models := []string{"claude-3-5-sonnet", "claude-3-5-haiku"}
+	for _, model := range models {
+		t.Logf("Sending request for model: %s", model)
+		modelPayload, _ := sjson.SetBytes(payload, "model", model)
+		if _, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   model,
+			Payload: modelPayload,
+		}, cliproxyexecutor.Options{
+			SourceFormat: sdktranslator.FromString("claude"),
+		}); err != nil {
+			t.Fatalf("Execute(%s) error: %v", model, err)
+		}
+	}
+
+	if len(userIDs) != 2 {
+		t.Fatalf("expected 2 requests, got %d", len(userIDs))
+	}
+	if userIDs[0] == "" || userIDs[1] == "" {
+		t.Fatal("expected user_id to be populated")
+	}
+	t.Logf("user_id[0] (model=%s): %s", requestModels[0], userIDs[0])
+	t.Logf("user_id[1] (model=%s): %s", requestModels[1], userIDs[1])
+	if userIDs[0] != userIDs[1] {
+		t.Fatalf("expected user_id to be reused across models, got %q and %q", userIDs[0], userIDs[1])
+	}
+	if !isValidUserID(userIDs[0]) {
+		t.Fatalf("user_id %q is not valid", userIDs[0])
+	}
+	t.Logf("✓ End-to-end test passed: Same user_id (%s) was used for both models", userIDs[0])
+}
+
+func TestClaudeExecutor_GeneratesNewUserIDByDefault(t *testing.T) {
+	resetUserIDCache()
+
+	var userIDs []string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		userIDs = append(userIDs, gjson.GetBytes(body, "metadata.user_id").String())
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","type":"message","model":"claude-3-5-sonnet","role":"assistant","content":[{"type":"text","text":"ok"}],"usage":{"input_tokens":1,"output_tokens":1}}`))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	for i := 0; i < 2; i++ {
+		if _, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   "claude-3-5-sonnet",
+			Payload: payload,
+		}, cliproxyexecutor.Options{
+			SourceFormat: sdktranslator.FromString("claude"),
+		}); err != nil {
+			t.Fatalf("Execute call %d error: %v", i, err)
+		}
+	}
+
+	if len(userIDs) != 2 {
+		t.Fatalf("expected 2 requests, got %d", len(userIDs))
+	}
+	if userIDs[0] == "" || userIDs[1] == "" {
+		t.Fatal("expected user_id to be populated")
+	}
+	if userIDs[0] == userIDs[1] {
+		t.Fatalf("expected user_id to change when caching is not enabled, got identical values %q", userIDs[0])
+	}
+	if !isValidUserID(userIDs[0]) || !isValidUserID(userIDs[1]) {
+		t.Fatalf("user_ids should be valid, got %q and %q", userIDs[0], userIDs[1])
+	}
+}
+
 func TestStripClaudeToolPrefixFromResponse_NestedToolReference(t *testing.T) {
 	input := []byte(`{"content":[{"type":"tool_result","tool_use_id":"toolu_123","content":[{"type":"tool_reference","tool_name":"proxy_mcp__nia__manage_resource"}]}]}`)
 	out := stripClaudeToolPrefixFromResponse(input, "proxy_")

internal/runtime/executor/codex_executor.go

@@ -183,7 +183,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 
 		var param any
 		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, line, &param)
-		resp = cliproxyexecutor.Response{Payload: []byte(out)}
+		resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 		return resp, nil
 	}
 	err = statusErr{code: 408, msg: "stream error: stream disconnected before completion: stream closed before response.completed"}
@@ -273,11 +273,11 @@ func (e *CodexExecutor) executeCompact(ctx context.Context, auth *cliproxyauth.A
 	reporter.ensurePublished(ctx)
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusBadRequest, msg: "streaming not supported for /responses/compact"}
 	}
@@ -362,7 +362,6 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -397,7 +396,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/codex_websockets_executor.go

@@ -363,7 +363,7 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 	}
 }
 
-func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	log.Debugf("Executing Codex Websockets stream request with auth ID: %s, model: %s", auth.ID, req.Model)
 	if ctx == nil {
 		ctx = context.Background()
@@ -436,7 +436,9 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 	})
 
 	conn, respHS, errDial := e.ensureUpstreamConn(ctx, auth, sess, authID, wsURL, wsHeaders)
+	var upstreamHeaders http.Header
 	if respHS != nil {
+		upstreamHeaders = respHS.Header.Clone()
 		recordAPIResponseMetadata(ctx, e.cfg, respHS.StatusCode, respHS.Header.Clone())
 	}
 	if errDial != nil {
@@ -516,7 +518,6 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 	markCodexWebsocketCreateSent(sess, conn, wsReqBody)
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		terminateReason := "completed"
 		var terminateErr error
@@ -627,7 +628,7 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 		}
 	}()
 
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: upstreamHeaders, Chunks: out}, nil
 }
 
 func (e *CodexWebsocketsExecutor) dialCodexWebsocket(ctx context.Context, auth *cliproxyauth.Auth, wsURL string, headers http.Header) (*websocket.Conn, *http.Response, error) {
@@ -1343,7 +1344,7 @@ func (e *CodexAutoExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	return e.httpExec.Execute(ctx, auth, req, opts)
 }
 
-func (e *CodexAutoExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (e *CodexAutoExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if e == nil || e.httpExec == nil || e.wsExec == nil {
 		return nil, fmt.Errorf("codex auto executor: executor is nil")
 	}

internal/runtime/executor/gemini_cli_executor.go

@@ -225,7 +225,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 			reporter.publish(ctx, parseGeminiCLIUsage(data))
 			var param any
 			out := sdktranslator.TranslateNonStream(respCtx, to, from, attemptModel, opts.OriginalRequest, payload, data, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(out)}
+			resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 			return resp, nil
 		}
 
@@ -256,7 +256,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 }
 
 // ExecuteStream performs a streaming request to the Gemini CLI API.
-func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -382,7 +382,6 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 		}
 
 		out := make(chan cliproxyexecutor.StreamChunk)
-		stream = out
 		go func(resp *http.Response, reqBody []byte, attemptModel string) {
 			defer close(out)
 			defer func() {
@@ -441,7 +440,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			}
 		}(httpResp, append([]byte(nil), payload...), attemptModel)
 
-		return stream, nil
+		return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 	}
 
 	if len(lastBody) > 0 {
@@ -546,7 +545,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 		if resp.StatusCode >= 200 && resp.StatusCode < 300 {
 			count := gjson.GetBytes(data, "totalTokens").Int()
 			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, data)
-			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+			return cliproxyexecutor.Response{Payload: []byte(translated), Headers: resp.Header.Clone()}, nil
 		}
 		lastStatus = resp.StatusCode
 		lastBody = append([]byte(nil), data...)

internal/runtime/executor/gemini_executor.go

@@ -205,12 +205,12 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming request to the Gemini API.
-func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -298,7 +298,6 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -335,7 +334,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // CountTokens counts tokens for the given request using the Gemini API.
@@ -416,7 +415,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(translated), Headers: resp.Header.Clone()}, nil
 }
 
 // Refresh refreshes the authentication credentials (no-op for Gemini API key).

internal/runtime/executor/gemini_vertex_executor.go

@@ -253,7 +253,7 @@ func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 }
 
 // ExecuteStream performs a streaming request to the Vertex AI API.
-func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -419,7 +419,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	to := sdktranslator.FromString("gemini")
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
@@ -524,12 +524,12 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // executeStreamWithServiceAccount handles streaming authentication using service account credentials.
-func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -618,7 +618,6 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -650,11 +649,11 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // executeStreamWithAPIKey handles streaming authentication using API key credentials.
-func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -743,7 +742,6 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -775,7 +773,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // countTokensWithServiceAccount counts tokens using service account credentials.
@@ -859,7 +857,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}, nil
 }
 
 // countTokensWithAPIKey handles token counting using API key credentials.
@@ -943,7 +941,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}, nil
 }
 
 // vertexCreds extracts project, location and raw service account JSON from auth metadata.

internal/runtime/executor/github_copilot_executor.go

@@ -35,12 +35,12 @@ const (
 	maxScannerBufferSize = 20_971_520
 
 	// Copilot API header values.
-	copilotUserAgent      = "GitHubCopilotChat/0.35.0"
-	copilotEditorVersion  = "vscode/1.107.0"
-	copilotPluginVersion  = "copilot-chat/0.35.0"
-	copilotIntegrationID  = "vscode-chat"
-	copilotOpenAIIntent   = "conversation-panel"
-	copilotGitHubAPIVer   = "2025-04-01"
+	copilotUserAgent     = "GitHubCopilotChat/0.35.0"
+	copilotEditorVersion = "vscode/1.107.0"
+	copilotPluginVersion = "copilot-chat/0.35.0"
+	copilotIntegrationID = "vscode-chat"
+	copilotOpenAIIntent  = "conversation-panel"
+	copilotGitHubAPIVer  = "2025-04-01"
 )
 
 // GitHubCopilotExecutor handles requests to the GitHub Copilot API.
@@ -232,7 +232,7 @@ func (e *GitHubCopilotExecutor) Execute(ctx context.Context, auth *cliproxyauth.
 }
 
 // ExecuteStream handles streaming requests to GitHub Copilot.
-func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	apiToken, baseURL, errToken := e.ensureAPIToken(ctx, auth)
 	if errToken != nil {
 		return nil, errToken
@@ -341,7 +341,6 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 
 	go func() {
 		defer close(out)
@@ -394,7 +393,10 @@ func (e *GitHubCopilotExecutor) ExecuteStream(ctx context.Context, auth *cliprox
 		}
 	}()
 
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{
+		Headers: httpResp.Header.Clone(),
+		Chunks:  out,
+	}, nil
 }
 
 // CountTokens is not supported for GitHub Copilot.

internal/runtime/executor/iflow_executor.go

@@ -169,12 +169,12 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming chat completion request.
-func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -262,7 +262,6 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -294,7 +293,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		reporter.ensurePublished(ctx)
 	}()
 
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/kilo_executor.go

@@ -168,7 +168,7 @@ func (e *KiloExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 }
 
 // ExecuteStream performs a streaming request.
-func (e *KiloExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *KiloExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -253,7 +253,6 @@ func (e *KiloExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer httpResp.Body.Close()
@@ -286,7 +285,10 @@ func (e *KiloExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		reporter.ensurePublished(ctx)
 	}()
 
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{
+		Headers: httpResp.Header.Clone(),
+		Chunks:  out,
+	}, nil
 }
 
 // Refresh validates the Kilo token.
@@ -456,4 +458,3 @@ func FetchKiloModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *config.C
 
 	return allModels
 }
-

internal/runtime/executor/kimi_executor.go

@@ -161,12 +161,12 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming chat completion request to Kimi.
-func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	from := opts.SourceFormat
 	if from.String() == "claude" {
 		auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
@@ -253,7 +253,6 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -285,7 +284,7 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // CountTokens estimates token count for Kimi requests.

internal/runtime/executor/kiro_executor.go

@@ -1053,7 +1053,7 @@ func (e *KiroExecutor) executeWithRetry(ctx context.Context, auth *cliproxyauth.
 
 // ExecuteStream handles streaming requests to Kiro API.
 // Supports automatic token refresh on 401/403 errors and quota fallback on 429.
-func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	accessToken, profileArn := kiroCredentials(auth)
 	if accessToken == "" {
 		return nil, fmt.Errorf("kiro: access token not found in auth")
@@ -1110,7 +1110,11 @@ func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	// Route to MCP endpoint instead of normal Kiro API
 	if kiroclaude.HasWebSearchTool(req.Payload) {
 		log.Infof("kiro: detected pure web_search request, routing to MCP endpoint")
-		return e.handleWebSearchStream(ctx, auth, req, opts, accessToken, profileArn)
+		streamWebSearch, errWebSearch := e.handleWebSearchStream(ctx, auth, req, opts, accessToken, profileArn)
+		if errWebSearch != nil {
+			return nil, errWebSearch
+		}
+		return &cliproxyexecutor.StreamResult{Chunks: streamWebSearch}, nil
 	}
 
 	reporter := newUsageReporter(ctx, e.Identifier(), req.Model, auth)
@@ -1128,7 +1132,11 @@ func (e *KiroExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 
 	// Execute stream with retry on 401/403 and 429 (quota exhausted)
 	// Note: currentOrigin and kiroPayload are built inside executeStreamWithRetry for each endpoint
-	return e.executeStreamWithRetry(ctx, auth, req, opts, accessToken, effectiveProfileArn, nil, body, from, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey)
+	streamKiro, errStreamKiro := e.executeStreamWithRetry(ctx, auth, req, opts, accessToken, effectiveProfileArn, nil, body, from, reporter, "", kiroModelID, isAgentic, isChatOnly, tokenKey)
+	if errStreamKiro != nil {
+		return nil, errStreamKiro
+	}
+	return &cliproxyexecutor.StreamResult{Chunks: streamKiro}, nil
 }
 
 // executeStreamWithRetry performs the streaming HTTP request with automatic retry on auth errors.
@@ -2543,6 +2551,7 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 	isThinkingBlockOpen := false                   // Track if thinking content block SSE event is open
 	thinkingBlockIndex := -1                       // Index of the thinking content block
 	var accumulatedThinkingContent strings.Builder // Accumulate thinking content for token counting
+	hasOfficialReasoningEvent := false             // Disable tag parsing after official reasoning events appear
 
 	// Buffer for handling partial tag matches at chunk boundaries
 	var pendingContent strings.Builder // Buffer content that might be part of a tag
@@ -2978,6 +2987,31 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 					lastUsageUpdateTime = time.Now()
 				}
 
+				if hasOfficialReasoningEvent {
+					processText := strings.TrimSpace(strings.ReplaceAll(strings.ReplaceAll(contentDelta, kirocommon.ThinkingStartTag, ""), kirocommon.ThinkingEndTag, ""))
+					if processText != "" {
+						if !isTextBlockOpen {
+							contentBlockIndex++
+							isTextBlockOpen = true
+							blockStart := kiroclaude.BuildClaudeContentBlockStartEvent(contentBlockIndex, "text", "", "")
+							sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, blockStart, &translatorParam)
+							for _, chunk := range sseData {
+								if chunk != "" {
+									out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
+								}
+							}
+						}
+						claudeEvent := kiroclaude.BuildClaudeStreamEvent(processText, contentBlockIndex)
+						sseData := sdktranslator.TranslateStream(ctx, sdktranslator.FromString("kiro"), targetFormat, model, originalReq, claudeBody, claudeEvent, &translatorParam)
+						for _, chunk := range sseData {
+							if chunk != "" {
+								out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunk + "\n\n")}
+							}
+						}
+					}
+					continue
+				}
+
 				// TAG-BASED THINKING PARSING: Parse <thinking> tags from content
 				// Combine pending content with new content for processing
 				pendingContent.WriteString(contentDelta)
@@ -3256,6 +3290,7 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 			}
 
 			if thinkingText != "" {
+				hasOfficialReasoningEvent = true
 				// Close text block if open before starting thinking block
 				if isTextBlockOpen && contentBlockIndex >= 0 {
 					blockStop := kiroclaude.BuildClaudeContentBlockStopEvent(contentBlockIndex)

internal/runtime/executor/openai_compat_executor.go

@@ -172,11 +172,11 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	// Translate response back to source format when needed
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, body, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -258,7 +258,6 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -298,7 +297,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		// Ensure we record the request if no usage chunk was ever seen
 		reporter.ensurePublished(ctx)
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/qwen_executor.go

@@ -150,11 +150,11 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -236,7 +236,6 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -268,7 +267,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/user_id_cache.go

@@ -0,0 +1,89 @@
+package executor
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"sync"
+	"time"
+)
+
+type userIDCacheEntry struct {
+	value  string
+	expire time.Time
+}
+
+var (
+	userIDCache            = make(map[string]userIDCacheEntry)
+	userIDCacheMu          sync.RWMutex
+	userIDCacheCleanupOnce sync.Once
+)
+
+const (
+	userIDTTL                = time.Hour
+	userIDCacheCleanupPeriod = 15 * time.Minute
+)
+
+func startUserIDCacheCleanup() {
+	go func() {
+		ticker := time.NewTicker(userIDCacheCleanupPeriod)
+		defer ticker.Stop()
+		for range ticker.C {
+			purgeExpiredUserIDs()
+		}
+	}()
+}
+
+func purgeExpiredUserIDs() {
+	now := time.Now()
+	userIDCacheMu.Lock()
+	for key, entry := range userIDCache {
+		if !entry.expire.After(now) {
+			delete(userIDCache, key)
+		}
+	}
+	userIDCacheMu.Unlock()
+}
+
+func userIDCacheKey(apiKey string) string {
+	sum := sha256.Sum256([]byte(apiKey))
+	return hex.EncodeToString(sum[:])
+}
+
+func cachedUserID(apiKey string) string {
+	if apiKey == "" {
+		return generateFakeUserID()
+	}
+
+	userIDCacheCleanupOnce.Do(startUserIDCacheCleanup)
+
+	key := userIDCacheKey(apiKey)
+	now := time.Now()
+
+	userIDCacheMu.RLock()
+	entry, ok := userIDCache[key]
+	valid := ok && entry.value != "" && entry.expire.After(now) && isValidUserID(entry.value)
+	userIDCacheMu.RUnlock()
+	if valid {
+		userIDCacheMu.Lock()
+		entry = userIDCache[key]
+		if entry.value != "" && entry.expire.After(now) && isValidUserID(entry.value) {
+			entry.expire = now.Add(userIDTTL)
+			userIDCache[key] = entry
+			userIDCacheMu.Unlock()
+			return entry.value
+		}
+		userIDCacheMu.Unlock()
+	}
+
+	newID := generateFakeUserID()
+
+	userIDCacheMu.Lock()
+	entry, ok = userIDCache[key]
+	if !ok || entry.value == "" || !entry.expire.After(now) || !isValidUserID(entry.value) {
+		entry.value = newID
+	}
+	entry.expire = now.Add(userIDTTL)
+	userIDCache[key] = entry
+	userIDCacheMu.Unlock()
+	return entry.value
+}

internal/runtime/executor/user_id_cache_test.go

@@ -0,0 +1,86 @@
+package executor
+
+import (
+	"testing"
+	"time"
+)
+
+func resetUserIDCache() {
+	userIDCacheMu.Lock()
+	userIDCache = make(map[string]userIDCacheEntry)
+	userIDCacheMu.Unlock()
+}
+
+func TestCachedUserID_ReusesWithinTTL(t *testing.T) {
+	resetUserIDCache()
+
+	first := cachedUserID("api-key-1")
+	second := cachedUserID("api-key-1")
+
+	if first == "" {
+		t.Fatal("expected generated user_id to be non-empty")
+	}
+	if first != second {
+		t.Fatalf("expected cached user_id to be reused, got %q and %q", first, second)
+	}
+}
+
+func TestCachedUserID_ExpiresAfterTTL(t *testing.T) {
+	resetUserIDCache()
+
+	expiredID := cachedUserID("api-key-expired")
+	cacheKey := userIDCacheKey("api-key-expired")
+	userIDCacheMu.Lock()
+	userIDCache[cacheKey] = userIDCacheEntry{
+		value:  expiredID,
+		expire: time.Now().Add(-time.Minute),
+	}
+	userIDCacheMu.Unlock()
+
+	newID := cachedUserID("api-key-expired")
+	if newID == expiredID {
+		t.Fatalf("expected expired user_id to be replaced, got %q", newID)
+	}
+	if newID == "" {
+		t.Fatal("expected regenerated user_id to be non-empty")
+	}
+}
+
+func TestCachedUserID_IsScopedByAPIKey(t *testing.T) {
+	resetUserIDCache()
+
+	first := cachedUserID("api-key-1")
+	second := cachedUserID("api-key-2")
+
+	if first == second {
+		t.Fatalf("expected different API keys to have different user_ids, got %q", first)
+	}
+}
+
+func TestCachedUserID_RenewsTTLOnHit(t *testing.T) {
+	resetUserIDCache()
+
+	key := "api-key-renew"
+	id := cachedUserID(key)
+	cacheKey := userIDCacheKey(key)
+
+	soon := time.Now()
+	userIDCacheMu.Lock()
+	userIDCache[cacheKey] = userIDCacheEntry{
+		value:  id,
+		expire: soon.Add(2 * time.Second),
+	}
+	userIDCacheMu.Unlock()
+
+	if refreshed := cachedUserID(key); refreshed != id {
+		t.Fatalf("expected cached user_id to be reused before expiry, got %q", refreshed)
+	}
+
+	userIDCacheMu.RLock()
+	entry := userIDCache[cacheKey]
+	userIDCacheMu.RUnlock()
+
+	if entry.expire.Sub(soon) < 30*time.Minute {
+		t.Fatalf("expected TTL to renew, got %v remaining", entry.expire.Sub(soon))
+	}
+}

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -231,8 +231,12 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 
 							} else if functionResponseResult.IsObject() {
 								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
-							} else {
+							} else if functionResponseResult.Raw != "" {
 								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
+							} else {
+								// Content field is missing entirely — .Raw is empty which
+								// causes sjson.SetRaw to produce invalid JSON (e.g. "result":}).
+								functionResponseJSON, _ = sjson.Set(functionResponseJSON, "response.result", "")
 							}
 
 							partJSON := `{}`

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -661,6 +661,85 @@ func TestConvertClaudeRequestToAntigravity_ThinkingOnly_NoHint(t *testing.T) {
 	}
 }
 
+func TestConvertClaudeRequestToAntigravity_ToolResultNoContent(t *testing.T) {
+	// Bug repro: tool_result with no content field produces invalid JSON
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "MyTool-123-456",
+						"name": "MyTool",
+						"input": {"key": "value"}
+					}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "MyTool-123-456"
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, true)
+	outputStr := string(output)
+
+	if !gjson.Valid(outputStr) {
+		t.Errorf("Result is not valid JSON:\n%s", outputStr)
+	}
+
+	// Verify the functionResponse has a valid result value
+	fr := gjson.Get(outputStr, "request.contents.1.parts.0.functionResponse.response.result")
+	if !fr.Exists() {
+		t.Error("functionResponse.response.result should exist")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolResultNullContent(t *testing.T) {
+	// Bug repro: tool_result with null content produces invalid JSON
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "MyTool-123-456",
+						"name": "MyTool",
+						"input": {"key": "value"}
+					}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "MyTool-123-456",
+						"content": null
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, true)
+	outputStr := string(output)
+
+	if !gjson.Valid(outputStr) {
+		t.Errorf("Result is not valid JSON:\n%s", outputStr)
+	}
+}
+
 func TestConvertClaudeRequestToAntigravity_ToolAndThinking_NoExistingSystem(t *testing.T) {
 	// When tools + thinking but no system instruction, should create one with hint
 	inputJSON := []byte(`{

internal/translator/codex/claude/codex_claude_response.go

@@ -22,8 +22,9 @@ var (
 
 // ConvertCodexResponseToClaudeParams holds parameters for response conversion.
 type ConvertCodexResponseToClaudeParams struct {
-	HasToolCall bool
-	BlockIndex  int
+	HasToolCall              bool
+	BlockIndex               int
+	HasReceivedArgumentsDelta bool
 }
 
 // ConvertCodexResponseToClaude performs sophisticated streaming response format conversion.
@@ -137,6 +138,7 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		itemType := itemResult.Get("type").String()
 		if itemType == "function_call" {
 			(*param).(*ConvertCodexResponseToClaudeParams).HasToolCall = true
+			(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta = false
 			template = `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
 			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 			template, _ = sjson.Set(template, "content_block.id", itemResult.Get("call_id").String())
@@ -171,12 +173,29 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 			output += fmt.Sprintf("data: %s\n\n", template)
 		}
 	} else if typeStr == "response.function_call_arguments.delta" {
+		(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta = true
 		template = `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
 		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 		template, _ = sjson.Set(template, "delta.partial_json", rootResult.Get("delta").String())
 
 		output += "event: content_block_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
+	} else if typeStr == "response.function_call_arguments.done" {
+		// Some models (e.g. gpt-5.3-codex-spark) send function call arguments
+		// in a single "done" event without preceding "delta" events.
+		// Emit the full arguments as a single input_json_delta so the
+		// downstream Claude client receives the complete tool input.
+		// When delta events were already received, skip to avoid duplicating arguments.
+		if !(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta {
+			if args := rootResult.Get("arguments").String(); args != "" {
+				template = `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
+				template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
+				template, _ = sjson.Set(template, "delta.partial_json", args)
+
+				output += "event: content_block_delta\n"
+				output += fmt.Sprintf("data: %s\n\n", template)
+			}
+		}
 	}
 
 	return []string{output}

internal/translator/kiro/claude/kiro_claude_request.go

@@ -243,13 +243,11 @@ func BuildKiroPayload(claudeBody []byte, modelID, profileArn, origin string, isA
 	// Process messages and build history
 	history, currentUserMsg, currentToolResults := processMessages(messages, modelID, origin)
 
-	// Build content with system prompt (only on first turn to avoid re-injection)
+	// Build content with system prompt.
+	// Keep thinking tags on subsequent turns so multi-turn Claude sessions
+	// continue to emit reasoning events.
 	if currentUserMsg != nil {
-		effectiveSystemPrompt := systemPrompt
-		if len(history) > 0 {
-			effectiveSystemPrompt = "" // Don't re-inject on subsequent turns
-		}
-		currentUserMsg.Content = buildFinalContent(currentUserMsg.Content, effectiveSystemPrompt, currentToolResults)
+		currentUserMsg.Content = buildFinalContent(currentUserMsg.Content, systemPrompt, currentToolResults)
 
 		// Deduplicate currentToolResults
 		currentToolResults = deduplicateToolResults(currentToolResults)
@@ -475,6 +473,15 @@ func IsThinkingEnabledWithHeaders(body []byte, headers http.Header) bool {
 		}
 	}
 
+	// Check model name directly for thinking hints.
+	// This enables thinking variants even when clients don't send explicit thinking fields.
+	model := strings.TrimSpace(gjson.GetBytes(body, "model").String())
+	modelLower := strings.ToLower(model)
+	if strings.Contains(modelLower, "thinking") || strings.Contains(modelLower, "-reason") {
+		log.Debugf("kiro: thinking mode enabled via model name hint: %s", model)
+		return true
+	}
+
 	log.Debugf("kiro: IsThinkingEnabled returning false (no thinking mode detected)")
 	return false
 }

internal/translator/kiro/openai/kiro_openai_request.go

@@ -234,16 +234,16 @@ func BuildKiroPayloadFromOpenAI(openaiBody []byte, modelID, profileArn, origin s
 	// Kiro API supports official thinking/reasoning mode via <thinking_mode> tag.
 	// When set to "enabled", Kiro returns reasoning content as official reasoningContentEvent
 	// rather than inline <thinking> tags in assistantResponseEvent.
-	// We use a high max_thinking_length to allow extensive reasoning.
+	// Use a conservative thinking budget to reduce latency/cost spikes in long sessions.
 	if thinkingEnabled {
 		thinkingHint := `<thinking_mode>enabled</thinking_mode>
-<max_thinking_length>200000</max_thinking_length>`
+<max_thinking_length>16000</max_thinking_length>`
 		if systemPrompt != "" {
 			systemPrompt = thinkingHint + "\n\n" + systemPrompt
 		} else {
 			systemPrompt = thinkingHint
 		}
-		log.Debugf("kiro-openai: injected thinking prompt (official mode)")
+		log.Infof("kiro-openai: injected thinking prompt (official mode), has_tools: %v", len(kiroTools) > 0)
 	}
 
 	// Process messages and build history
@@ -831,7 +831,6 @@ func hasThinkingTagInBody(body []byte) bool {
 	return strings.Contains(bodyStr, "<thinking_mode>") || strings.Contains(bodyStr, "<max_thinking_length>")
 }
 
-
 // extractToolChoiceHint extracts tool_choice from OpenAI request and returns a system prompt hint.
 // OpenAI tool_choice values:
 // - "none": Don't use any tools

sdk/api/handlers/claude/code_handlers.go

@@ -112,12 +112,13 @@ func (h *ClaudeCodeAPIHandler) ClaudeCountTokens(c *gin.Context) {
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 
-	resp, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -165,7 +166,7 @@ func (h *ClaudeCodeAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSO
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
@@ -194,6 +195,7 @@ func (h *ClaudeCodeAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSO
 		}
 	}
 
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -225,7 +227,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 	// This allows proper cleanup and cancellation of ongoing requests
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
 		c.Header("Cache-Control", "no-cache")
@@ -257,6 +259,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 			if !ok {
 				// Stream closed without data? Send DONE or just headers.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				flusher.Flush()
 				cliCancel(nil)
 				return
@@ -264,6 +267,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 
 			// Success! Set headers now.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write the first chunk
 			if len(chunk) > 0 {

sdk/api/handlers/gemini/gemini-cli_handlers.go

@@ -159,7 +159,8 @@ func (h *GeminiCLIAPIHandler) handleInternalStreamGenerateContent(c *gin.Context
 	modelName := modelResult.String()
 
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	h.forwardCLIStream(c, flusher, "", func(err error) { cliCancel(err) }, dataChan, errChan)
 	return
 }
@@ -172,12 +173,13 @@ func (h *GeminiCLIAPIHandler) handleInternalGenerateContent(c *gin.Context, rawJ
 	modelName := modelResult.String()
 
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }

sdk/api/handlers/gemini/gemini_handlers.go

@@ -188,7 +188,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 	}
 
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -223,6 +223,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 				if alt == "" {
 					setSSEHeaders()
 				}
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				flusher.Flush()
 				cliCancel(nil)
 				return
@@ -232,6 +233,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 			if alt == "" {
 				setSSEHeaders()
 			}
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write first chunk
 			if alt == "" {
@@ -262,12 +264,13 @@ func (h *GeminiAPIHandler) handleCountTokens(c *gin.Context, modelName string, r
 	c.Header("Content-Type", "application/json")
 	alt := h.GetAlt(c)
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -286,13 +289,14 @@ func (h *GeminiAPIHandler) handleGenerateContent(c *gin.Context, modelName strin
 	alt := h.GetAlt(c)
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }

sdk/api/handlers/handlers.go

@@ -179,6 +179,12 @@ func StreamingBootstrapRetries(cfg *config.SDKConfig) int {
 	return retries
 }
 
+// PassthroughHeadersEnabled returns whether upstream response headers should be forwarded to clients.
+// Default is false.
+func PassthroughHeadersEnabled(cfg *config.SDKConfig) bool {
+	return cfg != nil && cfg.PassthroughHeaders
+}
+
 func requestExecutionMetadata(ctx context.Context) map[string]any {
 	// Idempotency-Key is an optional client-supplied header used to correlate retries.
 	// It is forwarded as execution metadata; when absent we generate a UUID.
@@ -462,10 +468,10 @@ func appendAPIResponse(c *gin.Context, data []byte) {
 
 // ExecuteWithAuthManager executes a non-streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, *interfaces.ErrorMessage) {
+func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, http.Header, *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
-		return nil, errMsg
+		return nil, nil, errMsg
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -498,17 +504,20 @@ func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType
 				addon = hdr.Clone()
 			}
 		}
-		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
+		return nil, nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return resp.Payload, nil
+	if !PassthroughHeadersEnabled(h.Cfg) {
+		return resp.Payload, nil, nil
+	}
+	return resp.Payload, FilterUpstreamHeaders(resp.Headers), nil
 }
 
 // ExecuteCountWithAuthManager executes a non-streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, *interfaces.ErrorMessage) {
+func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, http.Header, *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
-		return nil, errMsg
+		return nil, nil, errMsg
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -541,20 +550,24 @@ func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handle
 				addon = hdr.Clone()
 			}
 		}
-		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
+		return nil, nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return resp.Payload, nil
+	if !PassthroughHeadersEnabled(h.Cfg) {
+		return resp.Payload, nil, nil
+	}
+	return resp.Payload, FilterUpstreamHeaders(resp.Headers), nil
 }
 
 // ExecuteStreamWithAuthManager executes a streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) (<-chan []byte, <-chan *interfaces.ErrorMessage) {
+// The returned http.Header carries upstream response headers captured before streaming begins.
+func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) (<-chan []byte, http.Header, <-chan *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
 		errChan := make(chan *interfaces.ErrorMessage, 1)
 		errChan <- errMsg
 		close(errChan)
-		return nil, errChan
+		return nil, nil, errChan
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -573,7 +586,7 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
-	chunks, err := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
+	streamResult, err := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
 	if err != nil {
 		errChan := make(chan *interfaces.ErrorMessage, 1)
 		status := http.StatusInternalServerError
@@ -590,8 +603,19 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		}
 		errChan <- &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 		close(errChan)
-		return nil, errChan
+		return nil, nil, errChan
 	}
+	passthroughHeadersEnabled := PassthroughHeadersEnabled(h.Cfg)
+	// Capture upstream headers from the initial connection synchronously before the goroutine starts.
+	// Keep a mutable map so bootstrap retries can replace it before first payload is sent.
+	var upstreamHeaders http.Header
+	if passthroughHeadersEnabled {
+		upstreamHeaders = cloneHeader(FilterUpstreamHeaders(streamResult.Headers))
+		if upstreamHeaders == nil {
+			upstreamHeaders = make(http.Header)
+		}
+	}
+	chunks := streamResult.Chunks
 	dataChan := make(chan []byte)
 	errChan := make(chan *interfaces.ErrorMessage, 1)
 	go func() {
@@ -665,9 +689,12 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 					if !sentPayload {
 						if bootstrapRetries < maxBootstrapRetries && bootstrapEligible(streamErr) {
 							bootstrapRetries++
-							retryChunks, retryErr := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
+							retryResult, retryErr := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
 							if retryErr == nil {
-								chunks = retryChunks
+								if passthroughHeadersEnabled {
+									replaceHeader(upstreamHeaders, FilterUpstreamHeaders(retryResult.Headers))
+								}
+								chunks = retryResult.Chunks
 								continue outer
 							}
 							streamErr = retryErr
@@ -698,7 +725,7 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 			}
 		}
 	}()
-	return dataChan, errChan
+	return dataChan, upstreamHeaders, errChan
 }
 
 func statusFromError(err error) int {
@@ -758,13 +785,33 @@ func cloneBytes(src []byte) []byte {
 	return dst
 }
 
+func cloneHeader(src http.Header) http.Header {
+	if src == nil {
+		return nil
+	}
+	dst := make(http.Header, len(src))
+	for key, values := range src {
+		dst[key] = append([]string(nil), values...)
+	}
+	return dst
+}
+
+func replaceHeader(dst http.Header, src http.Header) {
+	for key := range dst {
+		delete(dst, key)
+	}
+	for key, values := range src {
+		dst[key] = append([]string(nil), values...)
+	}
+}
+
 // 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
 	if msg != nil && msg.StatusCode > 0 {
 		status = msg.StatusCode
 	}
-	if msg != nil && msg.Addon != nil {
+	if msg != nil && msg.Addon != nil && PassthroughHeadersEnabled(h.Cfg) {
 		for key, values := range msg.Addon {
 			if len(values) == 0 {
 				continue

sdk/api/handlers/handlers_error_response_test.go

@@ -0,0 +1,68 @@
+package handlers
+
+import (
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestWriteErrorResponse_AddonHeadersDisabledByDefault(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodGet, "/", nil)
+
+	handler := NewBaseAPIHandlers(nil, nil)
+	handler.WriteErrorResponse(c, &interfaces.ErrorMessage{
+		StatusCode: http.StatusTooManyRequests,
+		Error:      errors.New("rate limit"),
+		Addon: http.Header{
+			"Retry-After":  {"30"},
+			"X-Request-Id": {"req-1"},
+		},
+	})
+
+	if recorder.Code != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want %d", recorder.Code, http.StatusTooManyRequests)
+	}
+	if got := recorder.Header().Get("Retry-After"); got != "" {
+		t.Fatalf("Retry-After should be empty when passthrough is disabled, got %q", got)
+	}
+	if got := recorder.Header().Get("X-Request-Id"); got != "" {
+		t.Fatalf("X-Request-Id should be empty when passthrough is disabled, got %q", got)
+	}
+}
+
+func TestWriteErrorResponse_AddonHeadersEnabled(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodGet, "/", nil)
+	c.Writer.Header().Set("X-Request-Id", "old-value")
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{PassthroughHeaders: true}, nil)
+	handler.WriteErrorResponse(c, &interfaces.ErrorMessage{
+		StatusCode: http.StatusTooManyRequests,
+		Error:      errors.New("rate limit"),
+		Addon: http.Header{
+			"Retry-After":  {"30"},
+			"X-Request-Id": {"new-1", "new-2"},
+		},
+	})
+
+	if recorder.Code != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want %d", recorder.Code, http.StatusTooManyRequests)
+	}
+	if got := recorder.Header().Get("Retry-After"); got != "30" {
+		t.Fatalf("Retry-After = %q, want %q", got, "30")
+	}
+	if got := recorder.Header().Values("X-Request-Id"); !reflect.DeepEqual(got, []string{"new-1", "new-2"}) {
+		t.Fatalf("X-Request-Id = %#v, want %#v", got, []string{"new-1", "new-2"})
+	}
+}

sdk/api/handlers/handlers_stream_bootstrap_test.go

@@ -23,7 +23,7 @@ func (e *failOnceStreamExecutor) Execute(context.Context, *coreauth.Auth, coreex
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }
 
-func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	e.mu.Lock()
 	e.calls++
 	call := e.calls
@@ -40,12 +40,18 @@ func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth,
 			},
 		}
 		close(ch)
-		return ch, nil
+		return &coreexecutor.StreamResult{
+			Headers: http.Header{"X-Upstream-Attempt": {"1"}},
+			Chunks:  ch,
+		}, nil
 	}
 
 	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{
+		Headers: http.Header{"X-Upstream-Attempt": {"2"}},
+		Chunks:  ch,
+	}, nil
 }
 
 func (e *failOnceStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -81,7 +87,7 @@ func (e *payloadThenErrorStreamExecutor) Execute(context.Context, *coreauth.Auth
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }
 
-func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	e.mu.Lock()
 	e.calls++
 	e.mu.Unlock()
@@ -97,7 +103,7 @@ func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreaut
 		},
 	}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
 }
 
 func (e *payloadThenErrorStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -134,7 +140,7 @@ func (e *authAwareStreamExecutor) Execute(context.Context, *coreauth.Auth, coree
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }
 
-func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *coreauth.Auth, req coreexecutor.Request, opts coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *coreauth.Auth, req coreexecutor.Request, opts coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	_ = ctx
 	_ = req
 	_ = opts
@@ -160,12 +166,12 @@ func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *corea
 			},
 		}
 		close(ch)
-		return ch, nil
+		return &coreexecutor.StreamResult{Chunks: ch}, nil
 	}
 
 	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
 }
 
 func (e *authAwareStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -231,11 +237,12 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	})
 
 	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		PassthroughHeaders: true,
 		Streaming: sdkconfig.StreamingConfig{
 			BootstrapRetries: 1,
 		},
 	}, manager)
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, upstreamHeaders, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -257,6 +264,70 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	if executor.Calls() != 2 {
 		t.Fatalf("expected 2 stream attempts, got %d", executor.Calls())
 	}
+	upstreamAttemptHeader := upstreamHeaders.Get("X-Upstream-Attempt")
+	if upstreamAttemptHeader != "2" {
+		t.Fatalf("expected upstream header from retry attempt, got %q", upstreamAttemptHeader)
+	}
+}
+
+func TestExecuteStreamWithAuthManager_HeaderPassthroughDisabledByDefault(t *testing.T) {
+	executor := &failOnceStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth1 := &coreauth.Auth{
+		ID:       "auth1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test1@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth1); err != nil {
+		t.Fatalf("manager.Register(auth1): %v", err)
+	}
+
+	auth2 := &coreauth.Auth{
+		ID:       "auth2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test2@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth2); err != nil {
+		t.Fatalf("manager.Register(auth2): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth1.ID, auth1.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	registry.GetGlobalRegistry().RegisterClient(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth1.ID)
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 1,
+		},
+	}, manager)
+	dataChan, upstreamHeaders, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+	for msg := range errChan {
+		if msg != nil {
+			t.Fatalf("unexpected error: %+v", msg)
+		}
+	}
+
+	if string(got) != "ok" {
+		t.Fatalf("expected payload ok, got %q", string(got))
+	}
+	if upstreamHeaders != nil {
+		t.Fatalf("expected nil upstream headers when passthrough is disabled, got %#v", upstreamHeaders)
+	}
 }
 
 func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
@@ -296,7 +367,7 @@ func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
 			BootstrapRetries: 1,
 		},
 	}, manager)
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -367,7 +438,7 @@ func TestExecuteStreamWithAuthManager_PinnedAuthKeepsSameUpstream(t *testing.T)
 		},
 	}, manager)
 	ctx := WithPinnedAuthID(context.Background(), "auth1")
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -431,7 +502,7 @@ func TestExecuteStreamWithAuthManager_SelectedAuthCallbackReceivesAuthID(t *test
 	ctx := WithSelectedAuthIDCallback(context.Background(), func(authID string) {
 		selectedAuthID = authID
 	})
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}

sdk/api/handlers/header_filter.go

@@ -0,0 +1,80 @@
+package handlers
+
+import (
+	"net/http"
+	"strings"
+)
+
+// hopByHopHeaders lists RFC 7230 Section 6.1 hop-by-hop headers that MUST NOT
+// be forwarded by proxies, plus security-sensitive headers that should not leak.
+var hopByHopHeaders = map[string]struct{}{
+	// RFC 7230 hop-by-hop
+	"Connection":          {},
+	"Keep-Alive":          {},
+	"Proxy-Authenticate":  {},
+	"Proxy-Authorization": {},
+	"Te":                  {},
+	"Trailer":             {},
+	"Transfer-Encoding":   {},
+	"Upgrade":             {},
+	// Security-sensitive
+	"Set-Cookie": {},
+	// CPA-managed (set by handlers, not upstream)
+	"Content-Length":   {},
+	"Content-Encoding": {},
+}
+
+// FilterUpstreamHeaders returns a copy of src with hop-by-hop and security-sensitive
+// headers removed. Returns nil if src is nil or empty after filtering.
+func FilterUpstreamHeaders(src http.Header) http.Header {
+	if src == nil {
+		return nil
+	}
+	connectionScoped := connectionScopedHeaders(src)
+	dst := make(http.Header)
+	for key, values := range src {
+		canonicalKey := http.CanonicalHeaderKey(key)
+		if _, blocked := hopByHopHeaders[canonicalKey]; blocked {
+			continue
+		}
+		if _, scoped := connectionScoped[canonicalKey]; scoped {
+			continue
+		}
+		dst[key] = values
+	}
+	if len(dst) == 0 {
+		return nil
+	}
+	return dst
+}
+
+func connectionScopedHeaders(src http.Header) map[string]struct{} {
+	scoped := make(map[string]struct{})
+	for _, rawValue := range src.Values("Connection") {
+		for _, token := range strings.Split(rawValue, ",") {
+			headerName := strings.TrimSpace(token)
+			if headerName == "" {
+				continue
+			}
+			scoped[http.CanonicalHeaderKey(headerName)] = struct{}{}
+		}
+	}
+	return scoped
+}
+
+// WriteUpstreamHeaders writes filtered upstream headers to the gin response writer.
+// Headers already set by CPA (e.g., Content-Type) are NOT overwritten.
+func WriteUpstreamHeaders(dst http.Header, src http.Header) {
+	if src == nil {
+		return
+	}
+	for key, values := range src {
+		// Don't overwrite headers already set by CPA handlers
+		if dst.Get(key) != "" {
+			continue
+		}
+		for _, v := range values {
+			dst.Add(key, v)
+		}
+	}
+}

sdk/api/handlers/header_filter_test.go

@@ -0,0 +1,55 @@
+package handlers
+
+import (
+	"net/http"
+	"testing"
+)
+
+func TestFilterUpstreamHeaders_RemovesConnectionScopedHeaders(t *testing.T) {
+	src := http.Header{}
+	src.Add("Connection", "keep-alive, x-hop-a, x-hop-b")
+	src.Add("Connection", "x-hop-c")
+	src.Set("Keep-Alive", "timeout=5")
+	src.Set("X-Hop-A", "a")
+	src.Set("X-Hop-B", "b")
+	src.Set("X-Hop-C", "c")
+	src.Set("X-Request-Id", "req-1")
+	src.Set("Set-Cookie", "session=secret")
+
+	filtered := FilterUpstreamHeaders(src)
+	if filtered == nil {
+		t.Fatalf("expected filtered headers, got nil")
+	}
+
+	requestID := filtered.Get("X-Request-Id")
+	if requestID != "req-1" {
+		t.Fatalf("expected X-Request-Id to be preserved, got %q", requestID)
+	}
+
+	blockedHeaderKeys := []string{
+		"Connection",
+		"Keep-Alive",
+		"X-Hop-A",
+		"X-Hop-B",
+		"X-Hop-C",
+		"Set-Cookie",
+	}
+	for _, key := range blockedHeaderKeys {
+		value := filtered.Get(key)
+		if value != "" {
+			t.Fatalf("expected %s to be removed, got %q", key, value)
+		}
+	}
+}
+
+func TestFilterUpstreamHeaders_ReturnsNilWhenAllHeadersBlocked(t *testing.T) {
+	src := http.Header{}
+	src.Add("Connection", "x-hop-a")
+	src.Set("X-Hop-A", "a")
+	src.Set("Set-Cookie", "session=secret")
+
+	filtered := FilterUpstreamHeaders(src)
+	if filtered != nil {
+		t.Fatalf("expected nil when all headers are filtered, got %#v", filtered)
+	}
+}

sdk/api/handlers/openai/openai_handlers.go

@@ -519,12 +519,13 @@ func (h *OpenAIAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSON []
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -534,12 +535,13 @@ func (h *OpenAIAPIHandler) handleNonStreamingResponseViaResponses(c *gin.Context
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	converted := convertResponsesObjectToChatCompletion(cliCtx, modelName, originalChatJSON, rawJSON, resp)
 	if converted == nil {
 		h.WriteErrorResponse(c, &interfaces.ErrorMessage{
@@ -575,7 +577,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -608,6 +610,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 			if !ok {
 				// Stream closed without data? Send DONE or just headers.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -616,6 +619,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 
 			// Success! Commit to streaming headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			_, _ = fmt.Fprintf(c.Writer, "data: %s\n\n", string(chunk))
 			flusher.Flush()
@@ -641,7 +645,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponseViaResponses(c *gin.Context, r
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenaiResponse, modelName, rawJSON, h.GetAlt(c))
 	var param any
 
 	setSSEHeaders := func() {
@@ -672,6 +676,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponseViaResponses(c *gin.Context, r
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -679,6 +684,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponseViaResponses(c *gin.Context, r
 			}
 
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 			writeConvertedResponsesChunk(c, cliCtx, modelName, originalChatJSON, rawJSON, chunk, &param)
 			flusher.Flush()
 
@@ -704,13 +710,14 @@ func (h *OpenAIAPIHandler) handleCompletionsNonStreamingResponse(c *gin.Context,
 	modelName := gjson.GetBytes(chatCompletionsJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	completionsResp := convertChatCompletionsResponseToCompletions(resp)
 	_, _ = c.Writer.Write(completionsResp)
 	cliCancel()
@@ -741,7 +748,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 
 	modelName := gjson.GetBytes(chatCompletionsJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -772,6 +779,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -780,6 +788,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 
 			// Success! Set headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write the first chunk
 			converted := convertChatCompletionsStreamChunkToCompletions(chunk)

sdk/api/handlers/openai/openai_responses_compact_test.go

@@ -31,7 +31,7 @@ func (e *compactCaptureExecutor) Execute(ctx context.Context, auth *coreauth.Aut
 	return coreexecutor.Response{Payload: []byte(`{"ok":true}`)}, nil
 }
 
-func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	return nil, errors.New("not implemented")
 }
 

sdk/api/handlers/openai/openai_responses_handlers.go

@@ -139,13 +139,14 @@ func (h *OpenAIResponsesAPIHandler) Compact(c *gin.Context) {
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -164,13 +165,14 @@ func (h *OpenAIResponsesAPIHandler) handleNonStreamingResponse(c *gin.Context, r
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
 
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -180,12 +182,13 @@ func (h *OpenAIResponsesAPIHandler) handleNonStreamingResponseViaChat(c *gin.Con
 
 	modelName := gjson.GetBytes(chatJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	var param any
 	converted := responsesconverter.ConvertOpenAIChatCompletionsResponseToOpenAIResponsesNonStream(cliCtx, modelName, originalResponsesJSON, originalResponsesJSON, resp, &param)
 	if converted == "" {
@@ -223,7 +226,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 	// New core execution path
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -256,6 +259,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 			if !ok {
 				// Stream closed without data? Send headers and done.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = c.Writer.Write([]byte("\n"))
 				flusher.Flush()
 				cliCancel(nil)
@@ -264,6 +268,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 
 			// Success! Set headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write first chunk logic (matching forwardResponsesStream)
 			if bytes.HasPrefix(chunk, []byte("event:")) {
@@ -294,7 +299,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponseViaChat(c *gin.Contex
 
 	modelName := gjson.GetBytes(chatJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, OpenAI, modelName, chatJSON, "")
 	var param any
 
 	setSSEHeaders := func() {
@@ -324,6 +329,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponseViaChat(c *gin.Contex
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = c.Writer.Write([]byte("\n"))
 				flusher.Flush()
 				cliCancel(nil)
@@ -331,6 +337,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponseViaChat(c *gin.Contex
 			}
 
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 			writeChatAsResponsesChunk(c, cliCtx, modelName, originalResponsesJSON, chunk, &param)
 			flusher.Flush()
 

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -153,7 +153,7 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 				pinnedAuthID = strings.TrimSpace(authID)
 			})
 		}
-		dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, requestJSON, "")
+		dataChan, _, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, requestJSON, "")
 
 		completedOutput, errForward := h.forwardResponsesWebsocket(c, conn, cliCancel, dataChan, errChan, &wsBodyLog, passthroughSessionID)
 		if errForward != nil {

sdk/cliproxy/auth/conductor.go

@@ -30,8 +30,9 @@ type ProviderExecutor interface {
 	Identifier() string
 	// Execute handles non-streaming execution and returns the provider response payload.
 	Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error)
-	// ExecuteStream handles streaming execution and returns a channel of provider chunks.
-	ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error)
+	// ExecuteStream handles streaming execution and returns a StreamResult containing
+	// upstream headers and a channel of provider chunks.
+	ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error)
 	// Refresh attempts to refresh provider credentials and returns the updated auth state.
 	Refresh(ctx context.Context, auth *Auth) (*Auth, error)
 	// CountTokens returns the token count for the given request.
@@ -558,7 +559,7 @@ func (m *Manager) ExecuteCount(ctx context.Context, providers []string, req clip
 
 // ExecuteStream performs a streaming execution using the configured selector and executor.
 // It supports multiple providers for the same model and round-robins the starting provider per model.
-func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	normalized := m.normalizeProviders(providers)
 	if len(normalized) == 0 {
 		return nil, &Error{Code: "provider_not_found", Message: "no provider supplied"}
@@ -568,9 +569,9 @@ func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cli
 
 	var lastErr error
 	for attempt := 0; ; attempt++ {
-		chunks, errStream := m.executeStreamMixedOnce(ctx, normalized, req, opts)
+		result, errStream := m.executeStreamMixedOnce(ctx, normalized, req, opts)
 		if errStream == nil {
-			return chunks, nil
+			return result, nil
 		}
 		lastErr = errStream
 		wait, shouldRetry := m.shouldRetryAfterError(errStream, attempt, normalized, req.Model, maxWait)
@@ -699,7 +700,7 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 	}
 }
 
-func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if len(providers) == 0 {
 		return nil, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
@@ -730,7 +731,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 		execReq.Model = rewriteModelForAuth(routeModel, auth)
 		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
 		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		chunks, errStream := executor.ExecuteStream(execCtx, auth, execReq, opts)
+		streamResult, errStream := executor.ExecuteStream(execCtx, auth, execReq, opts)
 		if errStream != nil {
 			if errCtx := execCtx.Err(); errCtx != nil {
 				return nil, errCtx
@@ -778,8 +779,11 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 			if !failed {
 				m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: true})
 			}
-		}(execCtx, auth.Clone(), provider, chunks)
-		return out, nil
+		}(execCtx, auth.Clone(), provider, streamResult.Chunks)
+		return &cliproxyexecutor.StreamResult{
+			Headers: streamResult.Headers,
+			Chunks:  out,
+		}, nil
 	}
 }
 

sdk/cliproxy/auth/conductor_executor_replace_test.go

@@ -24,10 +24,10 @@ func (e *replaceAwareExecutor) Execute(context.Context, *Auth, cliproxyexecutor.
 	return cliproxyexecutor.Response{}, nil
 }
 
-func (e *replaceAwareExecutor) ExecuteStream(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (e *replaceAwareExecutor) ExecuteStream(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	ch := make(chan cliproxyexecutor.StreamChunk)
 	close(ch)
-	return ch, nil
+	return &cliproxyexecutor.StreamResult{Chunks: ch}, nil
 }
 
 func (e *replaceAwareExecutor) Refresh(_ context.Context, auth *Auth) (*Auth, error) {
@@ -89,7 +89,11 @@ func TestManagerExecutorReturnsRegisteredExecutor(t *testing.T) {
 	if !okResolved {
 		t.Fatal("expected registered executor to be found")
 	}
-	if resolved != current {
+	resolvedExecutor, okResolvedExecutor := resolved.(*replaceAwareExecutor)
+	if !okResolvedExecutor {
+		t.Fatalf("expected resolved executor type %T, got %T", current, resolved)
+	}
+	if resolvedExecutor != current {
 		t.Fatal("expected resolved executor to match registered executor")
 	}
 

sdk/cliproxy/executor/types.go

@@ -57,6 +57,8 @@ type Response struct {
 	Payload []byte
 	// Metadata exposes optional structured data for translators.
 	Metadata map[string]any
+	// Headers carries upstream HTTP response headers for passthrough to clients.
+	Headers http.Header
 }
 
 // StreamChunk represents a single streaming payload unit emitted by provider executors.
@@ -67,6 +69,15 @@ type StreamChunk struct {
 	Err error
 }
 
+// StreamResult wraps the streaming response, providing both the chunk channel
+// and the upstream HTTP response headers captured before streaming begins.
+type StreamResult struct {
+	// Headers carries upstream HTTP response headers from the initial connection.
+	Headers http.Header
+	// Chunks is the channel of streaming payload units.
+	Chunks <-chan StreamChunk
+}
+
 // StatusError represents an error that carries an HTTP-like status code.
 // Provider executors should implement this when possible to enable
 // better auth state updates on failures (e.g., 401/402/429).

sdk/cliproxy/service_oauth_model_alias_test.go

@@ -90,3 +90,26 @@ func TestApplyOAuthModelAlias_ForkAddsMultipleAliases(t *testing.T) {
 		t.Fatalf("expected forked model name %q, got %q", "models/g5-2", out[2].Name)
 	}
 }
+
+func TestApplyOAuthModelAlias_DefaultGitHubCopilotAliasViaSanitize(t *testing.T) {
+	cfg := &config.Config{}
+	cfg.SanitizeOAuthModelAlias()
+
+	models := []*ModelInfo{
+		{ID: "claude-opus-4.6", Name: "models/claude-opus-4.6"},
+	}
+
+	out := applyOAuthModelAlias(cfg, "github-copilot", "oauth", models)
+	if len(out) != 2 {
+		t.Fatalf("expected 2 models (original + default alias), got %d", len(out))
+	}
+	if out[0].ID != "claude-opus-4.6" {
+		t.Fatalf("expected first model id %q, got %q", "claude-opus-4.6", out[0].ID)
+	}
+	if out[1].ID != "claude-opus-4-6" {
+		t.Fatalf("expected second model id %q, got %q", "claude-opus-4-6", out[1].ID)
+	}
+	if out[1].Name != "models/claude-opus-4-6" {
+		t.Fatalf("expected aliased model name %q, got %q", "models/claude-opus-4-6", out[1].Name)
+	}
+}