Merge pull request #8094 from wucm667/feat/issue-8092-configurable-min-output

feat(apicompat): configure minimum output token floor

backend/internal/config/config.go

@@ -979,6 +979,9 @@ const (
 
 // GatewayConfig API网关相关配置
 type GatewayConfig struct {
+	// MinOutputTokens is the minimum max_output_tokens used when
+	// converting Chat Completions or Anthropic requests to Responses.
+	MinOutputTokens int `mapstructure:"min_output_tokens"`
 	// 等待上游响应头的超时时间（秒），0表示无超时
 	// 注意：这不影响流式数据传输，只控制等待响应头的时间
 	ResponseHeaderTimeout int `mapstructure:"response_header_timeout"`
@@ -2035,6 +2038,7 @@ func configureConfigSource(setConfigFile, addConfigPath func(string)) {
 
 func setDefaults() {
 	viper.SetDefault("run_mode", RunModeStandard)
+	viper.SetDefault("gateway.min_output_tokens", 128)
 	viper.SetDefault("simple_mode.auto_create_default_groups", true)
 	viper.SetDefault("simple_mode_key_rate_limit_enabled", false)
 
@@ -2707,6 +2711,9 @@ func setEnvReachableDefaults() {
 }
 
 func (c *Config) Validate() error {
+	if c.Gateway.MinOutputTokens < 16 {
+		return fmt.Errorf("gateway.min_output_tokens must be at least 16")
+	}
 	forwardedClientIPHeaders, err := NormalizeForwardedClientIPHeaders(c.Security.ForwardedClientIPHeaders)
 	if err != nil {
 		return fmt.Errorf("security.forwarded_client_ip_headers: %w", err)

backend/internal/config/config_test.go

@@ -31,6 +31,29 @@ func TestLoadDefaultModelsListReadMaxBytes(t *testing.T) {
 	require.Equal(t, DefaultModelsListReadMaxBytes, cfg.Gateway.ModelsListReadMaxBytes)
 }
 
+func TestLoadDefaultMinOutputTokens(t *testing.T) {
+	resetViperWithJWTSecret(t)
+	cfg, err := Load()
+	require.NoError(t, err)
+	require.Equal(t, 128, cfg.Gateway.MinOutputTokens)
+}
+
+func TestLoadMinOutputTokensFromEnv(t *testing.T) {
+	resetViperWithJWTSecret(t)
+	t.Setenv("GATEWAY_MIN_OUTPUT_TOKENS", "16")
+	cfg, err := Load()
+	require.NoError(t, err)
+	require.Equal(t, 16, cfg.Gateway.MinOutputTokens)
+}
+
+func TestValidateMinOutputTokensMinimum(t *testing.T) {
+	resetViperWithJWTSecret(t)
+	cfg, err := Load()
+	require.NoError(t, err)
+	cfg.Gateway.MinOutputTokens = 15
+	require.EqualError(t, cfg.Validate(), "gateway.min_output_tokens must be at least 16")
+}
+
 func TestLoadTimezonePrecedence(t *testing.T) {
 	tests := []struct {
 		name         string

backend/internal/pkg/apicompat/anthropic_to_responses.go

@@ -13,6 +13,12 @@ import (
 // Chat Completions intermediary round-trip (e.g. thinking, cache_control,
 // structured system prompts).
 func AnthropicToResponses(req *AnthropicRequest) (*ResponsesRequest, error) {
+	return AnthropicToResponsesWithMinOutputTokens(req, DefaultMinMaxOutputTokens)
+}
+
+// AnthropicToResponsesWithMinOutputTokens converts an Anthropic request using
+// the supplied max_output_tokens floor.
+func AnthropicToResponsesWithMinOutputTokens(req *AnthropicRequest, minOutputTokens int) (*ResponsesRequest, error) {
 	if err := openai.ValidateGPT61SolReasoningEffort(req.Model, anthropicReasoningEffort(req)); err != nil {
 		return nil, err
 	}
@@ -50,8 +56,8 @@ func AnthropicToResponses(req *AnthropicRequest) (*ResponsesRequest, error) {
 
 	if req.MaxTokens > 0 {
 		v := req.MaxTokens
-		if v < minMaxOutputTokens {
-			v = minMaxOutputTokens
+		if v < minOutputTokens {
+			v = minOutputTokens
 		}
 		out.MaxOutputTokens = &v
 	}

backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge.go

@@ -40,6 +40,12 @@ import (
 // ResponsesToChatCompletionsRequest but avoids materializing the intermediate
 // ResponsesRequest and the extra marshal/unmarshal cycle.
 func AnthropicToChatCompletionsRequest(req *AnthropicRequest) (*ChatCompletionsRequest, error) {
+	return AnthropicToChatCompletionsRequestWithMinOutputTokens(req, DefaultMinMaxOutputTokens)
+}
+
+// AnthropicToChatCompletionsRequestWithMinOutputTokens converts an Anthropic
+// Messages request using the supplied max_completion_tokens floor.
+func AnthropicToChatCompletionsRequestWithMinOutputTokens(req *AnthropicRequest, minOutputTokens int) (*ChatCompletionsRequest, error) {
 	if req == nil {
 		return nil, fmt.Errorf("anthropic request is nil")
 	}
@@ -63,8 +69,8 @@ func AnthropicToChatCompletionsRequest(req *AnthropicRequest) (*ChatCompletionsR
 
 	if req.MaxTokens > 0 {
 		v := req.MaxTokens
-		if v < minMaxOutputTokens {
-			v = minMaxOutputTokens
+		if v < minOutputTokens {
+			v = minOutputTokens
 		}
 		out.MaxCompletionTokens = &v
 	}

backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge_test.go

@@ -267,6 +267,43 @@ func TestAnthropicToChatCompletionsRequest_MaxTokensFloor(t *testing.T) {
 	require.Equal(t, minMaxOutputTokens, *out.MaxCompletionTokens)
 }
 
+func TestAnthropicToChatCompletionsRequestWithMinOutputTokens_ConfiguredFloor(t *testing.T) {
+	req := &AnthropicRequest{
+		Model:     "claude-sonnet-4-20250514",
+		MaxTokens: 16,
+		Messages:  []AnthropicMessage{{Role: "user", Content: json.RawMessage(`"hi"`)}},
+	}
+
+	out, err := AnthropicToChatCompletionsRequestWithMinOutputTokens(req, 64)
+	require.NoError(t, err)
+	require.NotNil(t, out.MaxCompletionTokens)
+	require.Equal(t, 64, *out.MaxCompletionTokens)
+}
+
+func TestAnthropicToResponses_ConfiguredMaxTokensFloor(t *testing.T) {
+	for _, input := range []int{16, 64, 127} {
+		req := &AnthropicRequest{
+			Model:     "claude-sonnet-4-20250514",
+			MaxTokens: input,
+			Messages:  []AnthropicMessage{{Role: "user", Content: json.RawMessage(`"hi"`)}},
+		}
+		out, err := AnthropicToResponsesWithMinOutputTokens(req, 16)
+		require.NoError(t, err)
+		require.NotNil(t, out.MaxOutputTokens)
+		require.Equal(t, input, *out.MaxOutputTokens)
+	}
+
+	req := &AnthropicRequest{
+		Model:     "claude-sonnet-4-20250514",
+		MaxTokens: 1,
+		Messages:  []AnthropicMessage{{Role: "user", Content: json.RawMessage(`"hi"`)}},
+	}
+	out, err := AnthropicToResponsesWithMinOutputTokens(req, 16)
+	require.NoError(t, err)
+	require.NotNil(t, out.MaxOutputTokens)
+	require.Equal(t, 16, *out.MaxOutputTokens)
+}
+
 func TestAnthropicToChatCompletionsRequest_ReasoningEffortMapping(t *testing.T) {
 	req := &AnthropicRequest{
 		Model:        "gpt-5.4",

backend/internal/pkg/apicompat/chatcompletions_responses_test.go

@@ -2,6 +2,7 @@ package apicompat
 
 import (
 	"encoding/json"
+	"fmt"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -290,6 +291,46 @@ func mustMarshalJSON(t *testing.T, value any) []byte {
 }
 
 func TestChatCompletionsToResponses_MaxTokens(t *testing.T) {
+	t.Run("default floor", func(t *testing.T) {
+		maxTokens := 1
+		req := &ChatCompletionsRequest{
+			Model:     "gpt-4o",
+			MaxTokens: &maxTokens,
+			Messages:  []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+		}
+		resp, err := ChatCompletionsToResponses(req)
+		require.NoError(t, err)
+		require.NotNil(t, resp.MaxOutputTokens)
+		assert.Equal(t, DefaultMinMaxOutputTokens, *resp.MaxOutputTokens)
+	})
+
+	t.Run("configured floor", func(t *testing.T) {
+		for _, input := range []int{16, 64, 127, 128} {
+			input := input
+			t.Run(fmt.Sprintf("input_%d", input), func(t *testing.T) {
+				req := &ChatCompletionsRequest{
+					Model:     "gpt-4o",
+					MaxTokens: &input,
+					Messages:  []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+				}
+				resp, err := ChatCompletionsToResponsesWithMinOutputTokens(req, 16)
+				require.NoError(t, err)
+				require.NotNil(t, resp.MaxOutputTokens)
+				assert.Equal(t, input, *resp.MaxOutputTokens)
+			})
+		}
+		input := 1
+		req := &ChatCompletionsRequest{
+			Model:     "gpt-4o",
+			MaxTokens: &input,
+			Messages:  []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+		}
+		resp, err := ChatCompletionsToResponsesWithMinOutputTokens(req, 16)
+		require.NoError(t, err)
+		require.NotNil(t, resp.MaxOutputTokens)
+		assert.Equal(t, 16, *resp.MaxOutputTokens)
+	})
+
 	t.Run("max_tokens", func(t *testing.T) {
 		maxTokens := 100
 		req := &ChatCompletionsRequest{
@@ -318,6 +359,21 @@ func TestChatCompletionsToResponses_MaxTokens(t *testing.T) {
 		require.NotNil(t, resp.MaxOutputTokens)
 		assert.Equal(t, 500, *resp.MaxOutputTokens)
 	})
+
+	t.Run("max_completion_tokens_preferred before configured floor", func(t *testing.T) {
+		maxTokens := 500
+		maxCompletion := 1
+		req := &ChatCompletionsRequest{
+			Model:               "gpt-4o",
+			MaxTokens:           &maxTokens,
+			MaxCompletionTokens: &maxCompletion,
+			Messages:            []ChatMessage{{Role: "user", Content: json.RawMessage(`"Hi"`)}},
+		}
+		resp, err := ChatCompletionsToResponsesWithMinOutputTokens(req, 16)
+		require.NoError(t, err)
+		require.NotNil(t, resp.MaxOutputTokens)
+		assert.Equal(t, 16, *resp.MaxOutputTokens)
+	})
 }
 
 func TestChatCompletionsToResponses_ReasoningEffort(t *testing.T) {

backend/internal/pkg/apicompat/chatcompletions_to_responses.go

@@ -18,6 +18,12 @@ type chatMessageContent struct {
 // true. store is always false and reasoning.encrypted_content is always
 // included so that the response translator has full context.
 func ChatCompletionsToResponses(req *ChatCompletionsRequest) (*ResponsesRequest, error) {
+	return ChatCompletionsToResponsesWithMinOutputTokens(req, DefaultMinMaxOutputTokens)
+}
+
+// ChatCompletionsToResponsesWithMinOutputTokens converts a Chat Completions
+// request using the supplied max_output_tokens floor.
+func ChatCompletionsToResponsesWithMinOutputTokens(req *ChatCompletionsRequest, minOutputTokens int) (*ResponsesRequest, error) {
 	if err := openai.ValidateGPT61SolReasoningEffort(req.Model, req.ReasoningEffort); err != nil {
 		return nil, err
 	}
@@ -62,8 +68,8 @@ func ChatCompletionsToResponses(req *ChatCompletionsRequest) (*ResponsesRequest,
 	}
 	if maxTokens > 0 {
 		v := maxTokens
-		if v < minMaxOutputTokens {
-			v = minMaxOutputTokens
+		if v < minOutputTokens {
+			v = minOutputTokens
 		}
 		out.MaxOutputTokens = &v
 	}

backend/internal/pkg/apicompat/types.go

@@ -860,6 +860,9 @@ func (d ChatDelta) reasoningText() *string {
 // Shared constants
 // ---------------------------------------------------------------------------
 
-// minMaxOutputTokens is the floor for max_output_tokens in a Responses request.
-// Very small values may cause upstream API errors, so we enforce a minimum.
-const minMaxOutputTokens = 128
+// DefaultMinMaxOutputTokens is the default floor for max_output_tokens in a
+// Responses request. Very small values may cause upstream API errors, so we
+// enforce a minimum.
+const DefaultMinMaxOutputTokens = 128
+
+const minMaxOutputTokens = DefaultMinMaxOutputTokens

backend/internal/service/antigravity_gateway_compat.go

@@ -71,7 +71,7 @@ func (s *AntigravityGatewayService) ForwardAsChatCompletions(
 		return nil, s.writeAntigravityCompatError(c, http.StatusBadRequest, "invalid_request_error", "model is required")
 	}
 
-	responsesRequest, err := apicompat.ChatCompletionsToResponses(&request)
+	responsesRequest, err := chatCompletionsToResponses(settingServiceConfig(s.settingService), &request)
 	if err != nil {
 		return nil, s.writeAntigravityCompatError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
 	}

backend/internal/service/apicompat_config.go

@@ -0,0 +1,28 @@
+package service
+
+import (
+	"github.com/Wei-Shaw/sub2api/internal/config"
+	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
+)
+
+func configuredMinOutputTokens(cfg *config.Config) int {
+	if cfg != nil && cfg.Gateway.MinOutputTokens >= 16 {
+		return cfg.Gateway.MinOutputTokens
+	}
+	return apicompat.DefaultMinMaxOutputTokens
+}
+
+func chatCompletionsToResponses(cfg *config.Config, req *apicompat.ChatCompletionsRequest) (*apicompat.ResponsesRequest, error) {
+	return apicompat.ChatCompletionsToResponsesWithMinOutputTokens(req, configuredMinOutputTokens(cfg))
+}
+
+func anthropicToResponses(cfg *config.Config, req *apicompat.AnthropicRequest) (*apicompat.ResponsesRequest, error) {
+	return apicompat.AnthropicToResponsesWithMinOutputTokens(req, configuredMinOutputTokens(cfg))
+}
+
+func settingServiceConfig(settingService *SettingService) *config.Config {
+	if settingService == nil {
+		return nil
+	}
+	return settingService.cfg
+}

backend/internal/service/gateway_forward_as_chat_completions.go

@@ -44,7 +44,7 @@ func (s *GatewayService) ForwardAsChatCompletions(
 	clientStream := ccReq.Stream
 
 	// 2. Convert CC → Responses → Anthropic (chained conversion)
-	responsesReq, err := apicompat.ChatCompletionsToResponses(&ccReq)
+	responsesReq, err := chatCompletionsToResponses(s.cfg, &ccReq)
 	if err != nil {
 		return nil, fmt.Errorf("convert chat completions to responses: %w", err)
 	}

backend/internal/service/gateway_forward_as_responses.go

@@ -37,7 +37,7 @@ func (s *GatewayService) ForwardAsResponses(
 ) (*ForwardResult, error) {
 	startTime := time.Now()
 
-	normalizedBody, normalized, err := normalizeOpenAIResponsesLegacyIngress(body)
+	normalizedBody, normalized, err := normalizeOpenAIResponsesLegacyIngressWithConfig(s.cfg, body)
 	if err != nil {
 		return nil, err
 	}

backend/internal/service/gemini_chat_completions_compat_service.go

@@ -42,7 +42,7 @@ func (s *GeminiMessagesCompatService) ForwardAsChatCompletions(
 	clientStream := ccReq.Stream
 	includeUsage := ccReq.StreamOptions != nil && ccReq.StreamOptions.IncludeUsage
 
-	responsesReq, err := apicompat.ChatCompletionsToResponses(&ccReq)
+	responsesReq, err := chatCompletionsToResponses(s.cfg, &ccReq)
 	if err != nil {
 		return nil, s.writeChatCompletionsError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
 	}

backend/internal/service/openai_gateway_chat_completions.go

@@ -231,7 +231,7 @@ func (s *OpenAIGatewayService) forwardAsChatCompletions(
 		// Normal path: convert Chat Completions → Responses.
 		// ChatCompletionsToResponses always sets Stream=true (upstream always streams).
 		chatReq.Model = upstreamModel
-		responsesReq, err = apicompat.ChatCompletionsToResponses(&chatReq)
+		responsesReq, err = chatCompletionsToResponses(s.cfg, &chatReq)
 		if err != nil {
 			return nil, fmt.Errorf("convert chat completions to responses: %w", err)
 		}

backend/internal/service/openai_gateway_chat_completions_anthropic_native.go

@@ -57,7 +57,7 @@ func (s *OpenAIGatewayService) forwardChatCompletionsViaNativeAnthropic(
 	clientStream := ccReq.Stream
 
 	// 2. Convert CC → Responses → Anthropic (chained conversion)
-	responsesReq, err := apicompat.ChatCompletionsToResponses(&ccReq)
+	responsesReq, err := chatCompletionsToResponses(s.cfg, &ccReq)
 	if err != nil {
 		writeChatCompletionsError(c, http.StatusBadRequest, "invalid_request_error", "Failed to convert request")
 		return nil, fmt.Errorf("convert chat completions to responses: %w", err)

backend/internal/service/openai_gateway_forward.go

@@ -65,7 +65,7 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 	if normalized {
 		body = normalizedBody
 	}
-	legacyIngressBody, legacyIngressChanged, legacyIngressErr := normalizeOpenAIResponsesLegacyIngress(body)
+	legacyIngressBody, legacyIngressChanged, legacyIngressErr := normalizeOpenAIResponsesLegacyIngressWithConfig(s.cfg, body)
 	if legacyIngressErr != nil {
 		return nil, legacyIngressErr
 	}

backend/internal/service/openai_gateway_grok_chat_bridge.go

@@ -541,7 +541,7 @@ func (s *OpenAIGatewayService) forwardGrokChatCompletionsViaResponses(
 		return s.forwardAsRawChatCompletions(ctx, c, account, body, defaultMappedModel)
 	}
 
-	responsesReq, err := apicompat.ChatCompletionsToResponses(&chatReq)
+	responsesReq, err := chatCompletionsToResponses(s.cfg, &chatReq)
 	if err != nil {
 		return nil, fmt.Errorf("convert grok chat completions to responses: %w", err)
 	}

backend/internal/service/openai_gateway_messages.go

@@ -144,7 +144,7 @@ func (s *OpenAIGatewayService) ForwardAsAnthropic(
 
 	// 3. Convert Anthropic → Responses after compatibility-only replay guard.
 	anthropicReq.Model = upstreamModel
-	responsesReq, err := apicompat.AnthropicToResponses(&anthropicReq)
+	responsesReq, err := anthropicToResponses(s.cfg, &anthropicReq)
 	if err != nil {
 		return nil, fmt.Errorf("convert anthropic to responses: %w", err)
 	}

backend/internal/service/openai_gateway_messages_chat_fallback.go

@@ -52,7 +52,10 @@ func (s *OpenAIGatewayService) forwardAnthropicViaRawChatCompletions(
 	clientStream := anthropicReq.Stream
 
 	// 2. Anthropic → Chat Completions (direct, no Responses intermediary)
-	chatReq, err := apicompat.AnthropicToChatCompletionsRequest(&anthropicReq)
+	chatReq, err := apicompat.AnthropicToChatCompletionsRequestWithMinOutputTokens(
+		&anthropicReq,
+		configuredMinOutputTokens(s.cfg),
+	)
 	if err != nil {
 		writeAnthropicError(c, http.StatusBadRequest, "invalid_request_error", err.Error())
 		return nil, fmt.Errorf("convert anthropic to chat completions: %w", err)

backend/internal/service/openai_gateway_passthrough.go

@@ -214,6 +214,7 @@ func (s *OpenAIGatewayService) forwardOpenAIPassthrough(
 	if account != nil && account.IsOpenAI() {
 		responsesLite := isOpenAIResponsesLiteHeader(c.GetHeader(responsesLiteHeader)) || isOpenAIResponsesLiteWebSocketPayload(body)
 		normalizedBody, normalized, normalizeErr := normalizeOpenAIResponsesCompatibilityBodyWithOptions(body, account, openAIResponsesCompatibilityOptions{
+			Config:        s.cfg,
 			ResponsesLite: responsesLite,
 			Compact:       isOpenAIResponsesCompactPath(c),
 		})

backend/internal/service/openai_gateway_request_body.go

@@ -11,6 +11,7 @@ import (
 	"strings"
 	"unicode/utf8"
 
+	"github.com/Wei-Shaw/sub2api/internal/config"
 	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
 	"github.com/Wei-Shaw/sub2api/internal/pkg/ctxkey"
 	"github.com/Wei-Shaw/sub2api/internal/util/urlvalidator"
@@ -1433,7 +1434,12 @@ func normalizeOpenAIResponsesWebSocketCompatibilityBody(body []byte, account *Ac
 	return normalizeOpenAIResponsesCompatibilityBodyWithOptions(body, account, openAIResponsesCompatibilityOptions{ResponsesLite: responsesLite})
 }
 
+func normalizeOpenAIResponsesWebSocketCompatibilityBodyWithConfig(cfg *config.Config, body []byte, account *Account, responsesLite bool) ([]byte, bool, error) {
+	return normalizeOpenAIResponsesCompatibilityBodyWithOptions(body, account, openAIResponsesCompatibilityOptions{Config: cfg, ResponsesLite: responsesLite})
+}
+
 type openAIResponsesCompatibilityOptions struct {
+	Config        *config.Config
 	ResponsesLite bool
 	// Compact marks the /responses/compact wire shape, which is left as-is
 	// by request-shape compatibility rewrites such as web_search history.
@@ -1449,7 +1455,7 @@ func normalizeOpenAIResponsesCompatibilityBodyWithOptions(body []byte, account *
 	changed := false
 	if account.IsOpenAIOAuthLike() {
 		var err error
-		normalized, changed, err = normalizeOpenAIResponsesLegacyIngress(body)
+		normalized, changed, err = normalizeOpenAIResponsesLegacyIngressWithConfig(opts.Config, body)
 		if err != nil {
 			return body, false, err
 		}

backend/internal/service/openai_responses_ingress_compat.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/Wei-Shaw/sub2api/internal/config"
 	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
 	"github.com/tidwall/gjson"
 )
@@ -13,6 +14,10 @@ import (
 // payloads that a few Responses clients still emit. Native Responses input is
 // always authoritative because this path has no separate full-replay attempt.
 func normalizeOpenAIResponsesLegacyIngress(body []byte) ([]byte, bool, error) {
+	return normalizeOpenAIResponsesLegacyIngressWithConfig(nil, body)
+}
+
+func normalizeOpenAIResponsesLegacyIngressWithConfig(cfg *config.Config, body []byte) ([]byte, bool, error) {
 	if len(body) == 0 {
 		return body, false, nil
 	}
@@ -43,7 +48,7 @@ func normalizeOpenAIResponsesLegacyIngress(body []byte) ([]byte, bool, error) {
 	if hasMessages {
 		messages, messagesAreArray := messagesValue.([]any)
 		if messagesAreArray && len(messages) > 0 {
-			legacy, err := convertLegacyResponsesMessages(body)
+			legacy, err := convertLegacyResponsesMessages(cfg, body)
 			if err != nil {
 				return body, false, err
 			}
@@ -105,13 +110,13 @@ type convertedLegacyResponsesMessages struct {
 	hasReasoningEffort bool
 }
 
-func convertLegacyResponsesMessages(body []byte) (convertedLegacyResponsesMessages, error) {
+func convertLegacyResponsesMessages(cfg *config.Config, body []byte) (convertedLegacyResponsesMessages, error) {
 	var converted convertedLegacyResponsesMessages
 	var chatRequest apicompat.ChatCompletionsRequest
 	if err := json.Unmarshal(body, &chatRequest); err != nil {
 		return converted, fmt.Errorf("normalize legacy Responses messages: %w", err)
 	}
-	responsesRequest, err := apicompat.ChatCompletionsToResponses(&chatRequest)
+	responsesRequest, err := chatCompletionsToResponses(cfg, &chatRequest)
 	if err != nil {
 		return converted, fmt.Errorf("normalize legacy Responses messages: %w", err)
 	}

backend/internal/service/openai_ws_forwarder_ingress.go

@@ -269,7 +269,7 @@ func (s *OpenAIGatewayService) ProxyResponsesWebSocketFromClient(
 			normalized = next
 		}
 		responsesLite := isOpenAIResponsesLiteWebSocketPayload(normalized)
-		if compatibilityBody, compatibilityChanged, compatibilityErr := normalizeOpenAIResponsesWebSocketCompatibilityBody(normalized, account, responsesLite); compatibilityErr != nil {
+		if compatibilityBody, compatibilityChanged, compatibilityErr := normalizeOpenAIResponsesWebSocketCompatibilityBodyWithConfig(s.cfg, normalized, account, responsesLite); compatibilityErr != nil {
 			return openAIWSClientPayload{}, NewOpenAIWSClientCloseError(coderws.StatusPolicyViolation, "invalid websocket request payload", compatibilityErr)
 		} else if compatibilityChanged {
 			normalized = compatibilityBody

backend/internal/service/openai_ws_v2_passthrough_adapter.go

@@ -746,7 +746,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough(
 		firstClientMessage = s.ReplaceModelInBody(firstClientMessage, capturedSessionModel)
 	}
 	firstMessageResponsesLite := isOpenAIResponsesLiteWebSocketPayload(firstClientMessage)
-	if normalized, compatibilityChanged, normalizeErr := normalizeOpenAIResponsesWebSocketCompatibilityBody(firstClientMessage, account, firstMessageResponsesLite); normalizeErr != nil {
+	if normalized, compatibilityChanged, normalizeErr := normalizeOpenAIResponsesWebSocketCompatibilityBodyWithConfig(s.cfg, firstClientMessage, account, firstMessageResponsesLite); normalizeErr != nil {
 		return fmt.Errorf("normalize first websocket response.create: %w", normalizeErr)
 	} else if compatibilityChanged {
 		firstClientMessage = normalized
@@ -998,7 +998,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough(
 			}
 			responsesLite := isResponseCreate && isOpenAIResponsesLiteWebSocketPayload(payload)
 			if isResponseCreate {
-				if normalized, compatibilityChanged, normalizeErr := normalizeOpenAIResponsesWebSocketCompatibilityBody(payload, account, responsesLite); normalizeErr != nil {
+				if normalized, compatibilityChanged, normalizeErr := normalizeOpenAIResponsesWebSocketCompatibilityBodyWithConfig(s.cfg, payload, account, responsesLite); normalizeErr != nil {
 					return payload, nil, NewOpenAIWSClientCloseError(coderws.StatusPolicyViolation, "invalid websocket request payload", normalizeErr)
 				} else if compatibilityChanged {
 					payload = normalized