feat: complete Kthena inference service workflows

backend/cmd/gorm-gen/curd/generate.go

@@ -54,6 +54,9 @@ func main() {
 		model.PrequeueConfig{},
 		model.QueueQuotaLimit{},
 		model.UserBanRecord{},
+		model.KthenaChatSession{},
+		model.KthenaChatMessage{},
+		model.KthenaInferenceTemplate{},
 	)
 
 	// 执行并生成代码

backend/cmd/gorm-gen/models/migrate.go

@@ -142,6 +142,36 @@ func modelDownloadSubmissionMigration() *gormigrate.Migration {
 	}
 }
 
+func kthenaChatMigration() *gormigrate.Migration {
+	return &gormigrate.Migration{
+		ID: "202608021430",
+		Migrate: func(tx *gorm.DB) error {
+			if err := createTableIfMissing(tx, &model.KthenaChatSession{}); err != nil {
+				return err
+			}
+			return createTableIfMissing(tx, &model.KthenaChatMessage{})
+		},
+		Rollback: func(tx *gorm.DB) error {
+			if err := dropTableIfPresent(tx, &model.KthenaChatMessage{}); err != nil {
+				return err
+			}
+			return dropTableIfPresent(tx, &model.KthenaChatSession{})
+		},
+	}
+}
+
+func kthenaInferenceTemplateMigration() *gormigrate.Migration {
+	return &gormigrate.Migration{
+		ID: "202608021500",
+		Migrate: func(tx *gorm.DB) error {
+			return createTableIfMissing(tx, &model.KthenaInferenceTemplate{})
+		},
+		Rollback: func(tx *gorm.DB) error {
+			return dropTableIfPresent(tx, &model.KthenaInferenceTemplate{})
+		},
+	}
+}
+
 func createTableIfMissing(db *gorm.DB, value any) error {
 	if db.Migrator().HasTable(value) {
 		return nil
@@ -1572,6 +1602,8 @@ func main() {
 				return dropColumnIfPresent(tx, "users", &User{}, "BannedTimestamp")
 			},
 		},
+		kthenaChatMigration(),
+		kthenaInferenceTemplateMigration(),
 		modelDownloadSubmissionMigration(),
 	})
 
@@ -1609,6 +1641,9 @@ func main() {
 			&model.PrequeueConfig{},
 			&model.QueueQuotaLimit{},
 			&model.UserBanRecord{},
+			&model.KthenaChatSession{},
+			&model.KthenaChatMessage{},
+			&model.KthenaInferenceTemplate{},
 		)
 		if err != nil {
 			return err

backend/cmd/gorm-gen/models/migrate_test.go

@@ -126,3 +126,59 @@ func TestModelDownloadSubmissionMigrationAndRollback(t *testing.T) {
 		t.Fatal("model download submission table remains after rollback")
 	}
 }
+
+func TestKthenaChatMigrationAndRollback(t *testing.T) {
+	db, err := gorm.Open(sqlite.Open("file:kthena_chat_migration?mode=memory&cache=shared"), &gorm.Config{})
+	if err != nil {
+		t.Fatalf("open sqlite: %v", err)
+	}
+	migration := kthenaChatMigration()
+	if err := migration.Migrate(db); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	if err := migration.Migrate(db); err != nil {
+		t.Fatalf("idempotent migrate: %v", err)
+	}
+	for _, table := range []any{&model.KthenaChatSession{}, &model.KthenaChatMessage{}} {
+		if !db.Migrator().HasTable(table) {
+			t.Fatalf("missing migrated table %T", table)
+		}
+	}
+	if err := migration.Rollback(db); err != nil {
+		t.Fatalf("rollback: %v", err)
+	}
+	if err := migration.Rollback(db); err != nil {
+		t.Fatalf("idempotent rollback: %v", err)
+	}
+	for _, table := range []any{&model.KthenaChatSession{}, &model.KthenaChatMessage{}} {
+		if db.Migrator().HasTable(table) {
+			t.Fatalf("table remains after rollback: %T", table)
+		}
+	}
+}
+
+func TestKthenaInferenceTemplateMigrationAndRollback(t *testing.T) {
+	db, err := gorm.Open(sqlite.Open("file:kthena_inference_template_migration?mode=memory&cache=shared"), &gorm.Config{})
+	if err != nil {
+		t.Fatalf("open sqlite: %v", err)
+	}
+	migration := kthenaInferenceTemplateMigration()
+	if err := migration.Migrate(db); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	if err := migration.Migrate(db); err != nil {
+		t.Fatalf("idempotent migrate: %v", err)
+	}
+	if !db.Migrator().HasTable(&model.KthenaInferenceTemplate{}) {
+		t.Fatal("Kthena inference template table is missing")
+	}
+	if err := migration.Rollback(db); err != nil {
+		t.Fatalf("rollback: %v", err)
+	}
+	if err := migration.Rollback(db); err != nil {
+		t.Fatalf("idempotent rollback: %v", err)
+	}
+	if db.Migrator().HasTable(&model.KthenaInferenceTemplate{}) {
+		t.Fatal("Kthena inference template table remains after rollback")
+	}
+}

backend/dao/model/kthena_chat.go

@@ -0,0 +1,63 @@
+package model
+
+import (
+	"time"
+
+	"gorm.io/datatypes"
+)
+
+// KthenaChatSession stores one user's conversation with one routed Kthena
+// deployment. ClientSessionID is deliberately an application-level UUID: the
+// web client can keep using its existing local UUID while the database keeps a
+// compact numeric primary key for message joins.
+//
+// A session is scoped by user, account, namespace, deployment and served
+// model. The scope is derived from the authorized ModelBooster object rather
+// than from a client supplied model name.
+//
+//nolint:lll // Composite GORM index declarations must stay in a single struct tag.
+type KthenaChatSession struct {
+	ID uint `gorm:"primarykey"`
+
+	UserID    uint   `gorm:"not null;uniqueIndex:idx_kthena_chat_session_scope,priority:1;index:idx_kthena_chat_session_list,priority:1;comment:会话所属用户ID"`
+	AccountID uint   `gorm:"not null;uniqueIndex:idx_kthena_chat_session_scope,priority:2;index:idx_kthena_chat_session_list,priority:2;comment:会话所属账户ID"`
+	Username  string `gorm:"type:varchar(64);not null;comment:创建会话时的用户名快照"`
+
+	Namespace   string `gorm:"type:varchar(63);not null;uniqueIndex:idx_kthena_chat_session_scope,priority:3;index:idx_kthena_chat_session_list,priority:3;comment:模型部署命名空间"`
+	ServiceName string `gorm:"type:varchar(63);not null;uniqueIndex:idx_kthena_chat_session_scope,priority:4;index:idx_kthena_chat_session_list,priority:4;comment:ModelBooster名称"`
+	ModelName   string `gorm:"type:varchar(256);not null;uniqueIndex:idx_kthena_chat_session_scope,priority:5;index:idx_kthena_chat_session_list,priority:5;comment:路由模型名快照"`
+	BackendType string `gorm:"type:varchar(64);not null;comment:推理后端快照"`
+
+	ClientSessionID string     `gorm:"type:varchar(128);not null;uniqueIndex:idx_kthena_chat_session_scope,priority:6;comment:客户端会话UUID"`
+	Title           string     `gorm:"type:varchar(256);not null;default:'';comment:会话标题"`
+	MessageCount    int        `gorm:"not null;default:0;comment:消息数"`
+	LastMessageAt   *time.Time `gorm:"index:idx_kthena_chat_session_list,priority:6;comment:最后一条消息时间"`
+	CreatedAt       time.Time  `gorm:"not null;index:idx_kthena_chat_session_list,priority:7"`
+	UpdatedAt       time.Time  `gorm:"not null;index:idx_kthena_chat_session_list,priority:8"`
+}
+
+func (KthenaChatSession) TableName() string {
+	return "kthena_chat_sessions"
+}
+
+// KthenaChatMessage is an ordered message in a KthenaChatSession. ResponseJSON
+// stores the original non-streaming OpenAI-compatible completion for safe
+// idempotent retries of a client turn.
+//
+//nolint:lll // Composite GORM index declarations must stay in a single struct tag.
+type KthenaChatMessage struct {
+	ID        uint   `gorm:"primarykey"`
+	SessionID uint   `gorm:"not null;uniqueIndex:idx_kthena_chat_message_sequence,priority:1;uniqueIndex:idx_kthena_chat_message_turn,priority:1;index:idx_kthena_chat_message_session,priority:1;comment:会话ID"`
+	Sequence  int    `gorm:"not null;uniqueIndex:idx_kthena_chat_message_sequence,priority:2;comment:会话内消息顺序"`
+	Role      string `gorm:"type:varchar(32);not null;comment:OpenAI消息角色"`
+	Content   string `gorm:"type:text;not null;comment:消息正文"`
+	// ClientTurnID is populated only for the user message of an atomic turn.
+	// Nullable values keep normal CRUD message replacement unconstrained.
+	ClientTurnID *string        `gorm:"type:varchar(128);uniqueIndex:idx_kthena_chat_message_turn,priority:2;comment:客户端幂等请求ID"`
+	ResponseJSON datatypes.JSON `gorm:"type:jsonb;comment:助手消息对应的原始模型响应"`
+	CreatedAt    time.Time      `gorm:"not null;index:idx_kthena_chat_message_session,priority:2"`
+}
+
+func (KthenaChatMessage) TableName() string {
+	return "kthena_chat_messages"
+}

backend/dao/model/kthena_inference_template.go

@@ -0,0 +1,29 @@
+package model
+
+import (
+	"gorm.io/datatypes"
+	"gorm.io/gorm"
+)
+
+// KthenaInferenceTemplate is a private reusable deployment preset. It is
+// intentionally scoped by both user and account, so a template never becomes
+// visible when a browser switches identity or account context.
+//
+// Config keeps the complete form payload as JSON. The deployment API remains
+// authoritative when a template is eventually submitted, while templates can
+// evolve without a schema migration for every new form field.
+//
+//nolint:lll // Composite GORM index declarations must stay in a single struct tag.
+type KthenaInferenceTemplate struct {
+	gorm.Model
+
+	UserID      uint           `gorm:"not null;uniqueIndex:idx_kthena_inference_template_scope,priority:1;index:idx_kthena_inference_template_list,priority:1;comment:模板所属用户ID"`
+	AccountID   uint           `gorm:"not null;uniqueIndex:idx_kthena_inference_template_scope,priority:2;index:idx_kthena_inference_template_list,priority:2;comment:模板所属账户ID"`
+	Name        string         `gorm:"type:varchar(64);not null;uniqueIndex:idx_kthena_inference_template_scope,priority:3;comment:模板名称"`
+	Description string         `gorm:"type:varchar(512);not null;default:'';comment:模板说明"`
+	Config      datatypes.JSON `gorm:"type:jsonb;not null;comment:模型部署表单配置"`
+}
+
+func (KthenaInferenceTemplate) TableName() string {
+	return "kthena_inference_templates"
+}

backend/dao/model/system_config.go

@@ -15,7 +15,8 @@ const (
 	ConfigKeyLLMModelName = "LLM_MODEL_NAME"
 
 	// 功能开关配置键
-	ConfigKeyEnableGpuAnalysis = "ENABLE_GPU_ANALYSIS" // 值: "true" or "false"
+	ConfigKeyEnableGpuAnalysis     = "ENABLE_GPU_ANALYSIS"     // 值: "true" or "false"
+	ConfigKeyEnableKthenaInference = "ENABLE_KTHENA_INFERENCE" // 值: "true" or "false"
 
 	// Billing 功能与调度配置键
 	ConfigKeyEnableBillingFeature                     = "ENABLE_BILLING_FEATURE"
@@ -48,6 +49,7 @@ var DefaultConfigKeys = []string{
 	ConfigKeyLLMAPIKey,
 	ConfigKeyLLMModelName,
 	ConfigKeyEnableGpuAnalysis,
+	ConfigKeyEnableKthenaInference,
 	ConfigKeyEnableBillingFeature,
 	ConfigKeyEnableBillingActive,
 	ConfigKeyEnableRunningSettlement,