fix(openai): bridge oversized websocket requests

backend/internal/config/config.go

@@ -885,6 +885,12 @@ type GatewayOpenAIWSConfig struct {
 	StoreDisabledForceNewConn bool `mapstructure:"store_disabled_force_new_conn"`
 	// PrewarmGenerateEnabled: 是否启用 WSv2 generate=false 预热（默认 false）
 	PrewarmGenerateEnabled bool `mapstructure:"prewarm_generate_enabled"`
+	// ClientReadLimitBytes: 入站客户端 WS 单帧读取上限。
+	ClientReadLimitBytes int64 `mapstructure:"client_read_limit_bytes"`
+	// HTTPBridgeEnabled: 首包过大时，保持客户端 WS，改用 HTTP Responses 上游。
+	HTTPBridgeEnabled bool `mapstructure:"http_bridge_enabled"`
+	// HTTPBridgeThresholdBytes: 触发 HTTP bridge 的入站 WS payload 阈值。
+	HTTPBridgeThresholdBytes int64 `mapstructure:"http_bridge_threshold_bytes"`
 
 	// Feature 开关：v2 优先于 v1
 	ResponsesWebsockets   bool `mapstructure:"responses_websockets"`
@@ -1806,6 +1812,9 @@ func setDefaults() {
 	viper.SetDefault("gateway.openai_ws.store_disabled_conn_mode", "strict")
 	viper.SetDefault("gateway.openai_ws.store_disabled_force_new_conn", true)
 	viper.SetDefault("gateway.openai_ws.prewarm_generate_enabled", false)
+	viper.SetDefault("gateway.openai_ws.client_read_limit_bytes", 64*1024*1024)
+	viper.SetDefault("gateway.openai_ws.http_bridge_enabled", true)
+	viper.SetDefault("gateway.openai_ws.http_bridge_threshold_bytes", 15*1024*1024)
 	viper.SetDefault("gateway.openai_ws.responses_websockets", false)
 	viper.SetDefault("gateway.openai_ws.responses_websockets_v2", true)
 	viper.SetDefault("gateway.openai_ws.max_conns_per_account", 128)
@@ -2543,6 +2552,15 @@ func (c *Config) Validate() error {
 	if c.Gateway.OpenAIWS.PrewarmCooldownMS < 0 {
 		return fmt.Errorf("gateway.openai_ws.prewarm_cooldown_ms must be non-negative")
 	}
+	if c.Gateway.OpenAIWS.ClientReadLimitBytes <= 0 {
+		return fmt.Errorf("gateway.openai_ws.client_read_limit_bytes must be positive")
+	}
+	if c.Gateway.OpenAIWS.HTTPBridgeThresholdBytes < 0 {
+		return fmt.Errorf("gateway.openai_ws.http_bridge_threshold_bytes must be non-negative")
+	}
+	if c.Gateway.OpenAIWS.HTTPBridgeEnabled && c.Gateway.OpenAIWS.HTTPBridgeThresholdBytes == 0 {
+		return fmt.Errorf("gateway.openai_ws.http_bridge_threshold_bytes must be positive when http_bridge_enabled is true")
+	}
 	if c.Gateway.OpenAIWS.FallbackCooldownSeconds < 0 {
 		return fmt.Errorf("gateway.openai_ws.fallback_cooldown_seconds must be non-negative")
 	}

backend/internal/config/config_test.go

@@ -134,6 +134,15 @@ func TestLoadDefaultOpenAIWSConfig(t *testing.T) {
 	if cfg.Gateway.OpenAIWS.PrewarmCooldownMS != 300 {
 		t.Fatalf("Gateway.OpenAIWS.PrewarmCooldownMS = %d, want 300", cfg.Gateway.OpenAIWS.PrewarmCooldownMS)
 	}
+	if cfg.Gateway.OpenAIWS.ClientReadLimitBytes != 64*1024*1024 {
+		t.Fatalf("Gateway.OpenAIWS.ClientReadLimitBytes = %d, want %d", cfg.Gateway.OpenAIWS.ClientReadLimitBytes, 64*1024*1024)
+	}
+	if !cfg.Gateway.OpenAIWS.HTTPBridgeEnabled {
+		t.Fatalf("Gateway.OpenAIWS.HTTPBridgeEnabled = false, want true")
+	}
+	if cfg.Gateway.OpenAIWS.HTTPBridgeThresholdBytes != 15*1024*1024 {
+		t.Fatalf("Gateway.OpenAIWS.HTTPBridgeThresholdBytes = %d, want %d", cfg.Gateway.OpenAIWS.HTTPBridgeThresholdBytes, 15*1024*1024)
+	}
 	if cfg.Gateway.OpenAIWS.RetryBackoffInitialMS != 120 {
 		t.Fatalf("Gateway.OpenAIWS.RetryBackoffInitialMS = %d, want 120", cfg.Gateway.OpenAIWS.RetryBackoffInitialMS)
 	}

backend/internal/handler/openai_gateway_handler.go

@@ -1167,7 +1167,7 @@ func (h *OpenAIGatewayHandler) ResponsesWebSocket(c *gin.Context) {
 	defer func() {
 		_ = wsConn.CloseNow()
 	}()
-	wsConn.SetReadLimit(16 * 1024 * 1024)
+	wsConn.SetReadLimit(service.ResolveOpenAIWSClientReadLimitBytes(h.cfg))
 
 	ctx := c.Request.Context()
 	readCtx, cancel := context.WithTimeout(ctx, 30*time.Second)

backend/internal/service/openai_gateway_service.go

@@ -256,6 +256,9 @@ type OpenAIForwardResult struct {
 	ImageOutputSizes   []string
 	ImageSizeSource    string
 	ImageSizeBreakdown map[string]int
+
+	wsReplayInput       []json.RawMessage
+	wsReplayInputExists bool
 }
 
 type OpenAIWSRetryMetricsSnapshot struct {