Version: 0.6.6.dev.260317
✨ feat(task,agent): 新增任务完成接口,并打通聊天全链路 Token 记账 - ✅ 新增“标记任务为完成”接口,并补充幂等保护,避免重复完成导致状态污染 - 📊 为聊天链路补充 Token 统计能力: - 流式主对话链路直接读取模型 `usage` - Agent 链路通过 `Eino callback + ctx` 聚合 `Generate usage` - 在流式场景下补齐缺失的 `usage` 数据 - 🧾 按口径 B 完成 Token 落库: - 用户消息 `token` 记为 `0` - 助手消息记录本轮总 `token` - 持久化时同步更新 `chat_histories.tokens_consumed`、`agent_chats.tokens_total`、`users.token_usage` - 🔄 异步标题生成产生的 Token 通过 Outbox 事件完成账本增量调整,保证统计口径一致 - 📝 同步更新 `AGENTS.md` 与 `.gitignore` - 📚 小幅更新 README 说明文档
This commit is contained in:
@@ -32,6 +32,11 @@ type AgentService struct {
|
||||
// 这里通过依赖注入把“模型、仓储、缓存、异步持久化通道”统一交给服务层管理,
|
||||
// 便于后续在单测中替换实现,或在启动流程中按环境切换配置。
|
||||
func NewAgentService(aiHub *inits.AIHub, repo *dao.AgentDAO, taskRepo *dao.TaskDAO, agentRedis *dao.AgentCache, eventPublisher outboxinfra.EventPublisher) *AgentService {
|
||||
// 全局注册一次 token 采集 callback:
|
||||
// 1. 只注册一次,避免重复处理;
|
||||
// 2. 只有带 RequestTokenMeter 的请求上下文才会真正累加。
|
||||
ensureTokenMeterCallbackRegistered()
|
||||
|
||||
return &AgentService{
|
||||
AIHub: aiHub,
|
||||
repo: repo,
|
||||
@@ -76,7 +81,7 @@ func (s *AgentService) PersistChatHistory(ctx context.Context, payload model.Cha
|
||||
// 1. 未注入事件发布器时(例如本地极简环境),直接同步写 DB。
|
||||
// 这样可以保证功能不依赖 Kafka 也能跑通。
|
||||
if s.eventPublisher == nil {
|
||||
return s.repo.SaveChatHistory(ctx, payload.UserID, payload.ConversationID, payload.Role, payload.Message)
|
||||
return s.repo.SaveChatHistory(ctx, payload.UserID, payload.ConversationID, payload.Role, payload.Message, payload.TokensConsumed)
|
||||
}
|
||||
// 2. 已启用异步总线时,只发布“持久化请求事件”,不在请求路径阻塞 Kafka。
|
||||
// 2.1 发布成功仅代表“事件安全入队”,实际落库由消费者异步完成。
|
||||
@@ -167,12 +172,23 @@ func (s *AgentService) runNormalChatFlow(
|
||||
|
||||
// 6. 执行真正的流式聊天。
|
||||
// fullText 用于后续写 Redis/持久化,outChan 用于把流片段实时推给前端。
|
||||
fullText, streamErr := chat.StreamChat(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, chatHistory, outChan, traceID, chatID, requestStart)
|
||||
fullText, streamUsage, streamErr := chat.StreamChat(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, chatHistory, outChan, traceID, chatID, requestStart)
|
||||
if streamErr != nil {
|
||||
pushErrNonBlocking(errChan, streamErr)
|
||||
return
|
||||
}
|
||||
|
||||
// 6.1 流式 usage 并入请求级 token 统计器:
|
||||
// 6.1.1 route/quicknote/taskquery 等 Generate 调用由 callback 自动累加;
|
||||
// 6.1.2 主对话 Stream usage 在这里手动补齐。
|
||||
addSchemaUsageIntoRequest(ctx, streamUsage)
|
||||
requestTokenSnapshot := snapshotRequestTokenMeter(ctx)
|
||||
requestTotalTokens := requestTokenSnapshot.TotalTokens
|
||||
if requestTotalTokens <= 0 && streamUsage != nil {
|
||||
// 兜底:若 callback/meter 未生效,至少使用流式 usage 保底记账。
|
||||
requestTotalTokens = normalizeUsageTotal(streamUsage.TotalTokens, streamUsage.PromptTokens, streamUsage.CompletionTokens)
|
||||
}
|
||||
|
||||
// 7. 后置持久化(用户消息):
|
||||
// 7.1 先写 Redis,保证“最新会话上下文”可立即用于下一轮推理;
|
||||
// 7.2 再走可靠持久化入口(outbox 或同步 DB)。
|
||||
@@ -185,6 +201,8 @@ func (s *AgentService) runNormalChatFlow(
|
||||
ConversationID: chatID,
|
||||
Role: "user",
|
||||
Message: userMessage,
|
||||
// 口径B:用户消息固定记 0;本轮总 token 统一记在助手消息。
|
||||
TokensConsumed: 0,
|
||||
}); err != nil {
|
||||
pushErrNonBlocking(errChan, err)
|
||||
return
|
||||
@@ -204,6 +222,8 @@ func (s *AgentService) runNormalChatFlow(
|
||||
ConversationID: chatID,
|
||||
Role: "assistant",
|
||||
Message: fullText,
|
||||
// 口径B:助手消息记录“本轮请求总 token”。
|
||||
TokensConsumed: requestTotalTokens,
|
||||
}); saveErr != nil {
|
||||
pushErrNonBlocking(errChan, saveErr)
|
||||
}
|
||||
@@ -223,13 +243,16 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
outChan := make(chan string, 8)
|
||||
errChan := make(chan error, 1)
|
||||
|
||||
// 0. 初始化“请求级 token 统计器”,用于聚合本次请求所有模型开销。
|
||||
requestCtx, _ := withRequestTokenMeter(ctx)
|
||||
|
||||
// 1) 规范会话 ID,选择模型。
|
||||
chatID = normalizeConversationID(chatID)
|
||||
selectedModel, resolvedModelName := s.pickChatModel(modelName)
|
||||
|
||||
// 2) 确保会话存在(优先缓存,必要时回源 DB 并创建)。
|
||||
// 2.1 先查 Redis 会话标记,命中则可跳过 DB 存在性校验。
|
||||
result, err := s.agentCache.GetConversationStatus(ctx, chatID)
|
||||
result, err := s.agentCache.GetConversationStatus(requestCtx, chatID)
|
||||
if err != nil {
|
||||
errChan <- err
|
||||
close(outChan)
|
||||
@@ -238,7 +261,7 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
}
|
||||
if !result {
|
||||
// 2.2 缓存未命中时回源 DB:确认会话是否存在。
|
||||
innerResult, ifErr := s.repo.IfChatExists(ctx, userID, chatID)
|
||||
innerResult, ifErr := s.repo.IfChatExists(requestCtx, userID, chatID)
|
||||
if ifErr != nil {
|
||||
errChan <- ifErr
|
||||
close(outChan)
|
||||
@@ -255,7 +278,7 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
}
|
||||
}
|
||||
// 2.4 补写 Redis 会话标记,优化下次访问。
|
||||
if err = s.agentCache.SetConversationStatus(ctx, chatID); err != nil {
|
||||
if err = s.agentCache.SetConversationStatus(requestCtx, chatID); err != nil {
|
||||
log.Printf("设置会话状态缓存失败 chat=%s: %v", chatID, err)
|
||||
}
|
||||
}
|
||||
@@ -269,11 +292,11 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
defer close(outChan)
|
||||
|
||||
// 3.1 先走轻量路由,拿到统一 action。
|
||||
routing := s.decideActionRouting(ctx, selectedModel, userMessage)
|
||||
routing := s.decideActionRouting(requestCtx, selectedModel, userMessage)
|
||||
|
||||
// 3.2 chat:直接走普通聊天主链路。
|
||||
if routing.Action == route.ActionChat {
|
||||
s.runNormalChatFlow(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
s.runNormalChatFlow(requestCtx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
return
|
||||
}
|
||||
|
||||
@@ -284,7 +307,7 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
// 3.4 quick_note_create:执行随口记 graph。
|
||||
if routing.Action == route.ActionQuickNoteCreate {
|
||||
quickHandled, quickState, quickErr := s.tryHandleQuickNoteWithGraph(
|
||||
ctx,
|
||||
requestCtx,
|
||||
selectedModel,
|
||||
userMessage,
|
||||
userID,
|
||||
@@ -301,14 +324,15 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
if quickHandled {
|
||||
// 3.4.1 随口记处理成功:组织最终回复并按 OpenAI 兼容格式输出。
|
||||
progress.Emit("quick_note.reply.polishing", "正在结合你的话题润色回复。")
|
||||
quickReply := buildQuickNoteFinalReply(ctx, selectedModel, userMessage, quickState)
|
||||
quickReply := buildQuickNoteFinalReply(requestCtx, selectedModel, userMessage, quickState)
|
||||
if emitErr := emitSingleAssistantCompletion(outChan, resolvedModelName, quickReply); emitErr != nil {
|
||||
pushErrNonBlocking(errChan, emitErr)
|
||||
return
|
||||
}
|
||||
|
||||
// 3.4.2 对随口记回复执行统一后置持久化(Redis + outbox/DB)。
|
||||
s.persistChatAfterReply(ctx, userID, chatID, userMessage, quickReply, errChan)
|
||||
requestTotalTokens := snapshotRequestTokenMeter(requestCtx).TotalTokens
|
||||
s.persistChatAfterReply(requestCtx, userID, chatID, userMessage, quickReply, 0, requestTotalTokens, errChan)
|
||||
// 3.4.3 随口记链路同样异步生成会话标题(仅首次写入)。
|
||||
s.ensureConversationTitleAsync(userID, chatID)
|
||||
return
|
||||
@@ -316,18 +340,18 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
|
||||
// 3.4.4 路由误判或 graph 判定非随口记时,回落普通聊天,保证“能聊”。
|
||||
progress.Emit("quick_note.fallback", "当前输入不是随口记请求,切换到普通对话。")
|
||||
s.runNormalChatFlow(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
s.runNormalChatFlow(requestCtx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
return
|
||||
}
|
||||
|
||||
// 3.5 task_query:执行任务查询 tool-calling。
|
||||
if routing.Action == route.ActionTaskQuery {
|
||||
reply, queryErr := s.runTaskQueryFlow(ctx, selectedModel, userMessage, userID, progress.Emit)
|
||||
reply, queryErr := s.runTaskQueryFlow(requestCtx, selectedModel, userMessage, userID, progress.Emit)
|
||||
if queryErr != nil {
|
||||
// 3.5.1 任务查询失败时回退普通聊天,避免请求直接中断。
|
||||
log.Printf("任务查询 tool-calling 执行失败,回退普通聊天 trace_id=%s chat_id=%s err=%v", traceID, chatID, queryErr)
|
||||
progress.Emit("task_query.fallback", "任务查询暂不可用,先切回普通对话。")
|
||||
s.runNormalChatFlow(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
s.runNormalChatFlow(requestCtx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
return
|
||||
}
|
||||
|
||||
@@ -336,13 +360,14 @@ func (s *AgentService) AgentChat(ctx context.Context, userMessage string, ifThin
|
||||
pushErrNonBlocking(errChan, emitErr)
|
||||
return
|
||||
}
|
||||
s.persistChatAfterReply(ctx, userID, chatID, userMessage, reply, errChan)
|
||||
requestTotalTokens := snapshotRequestTokenMeter(requestCtx).TotalTokens
|
||||
s.persistChatAfterReply(requestCtx, userID, chatID, userMessage, reply, 0, requestTotalTokens, errChan)
|
||||
s.ensureConversationTitleAsync(userID, chatID)
|
||||
return
|
||||
}
|
||||
|
||||
// 3.6 未知 action 兜底:走普通聊天,保证可用性。
|
||||
s.runNormalChatFlow(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
s.runNormalChatFlow(requestCtx, selectedModel, resolvedModelName, userMessage, ifThinking, userID, chatID, traceID, requestStart, outChan, errChan)
|
||||
}()
|
||||
|
||||
return outChan, errChan
|
||||
|
||||
@@ -10,6 +10,7 @@ import (
|
||||
|
||||
"github.com/LoveLosita/smartflow/backend/model"
|
||||
"github.com/LoveLosita/smartflow/backend/respond"
|
||||
eventsvc "github.com/LoveLosita/smartflow/backend/service/events"
|
||||
"github.com/cloudwego/eino-ext/components/model/ark"
|
||||
einoModel "github.com/cloudwego/eino/components/model"
|
||||
"github.com/cloudwego/eino/schema"
|
||||
@@ -32,6 +33,9 @@ const (
|
||||
conversationListDefaultPageSize = 20
|
||||
// conversationListMaxPageSize 是会话列表单页上限,避免超大分页压垮数据库。
|
||||
conversationListMaxPageSize = 100
|
||||
// conversationTitleTokenAdjustReason 是“标题异步生成 token 账本调整”原因码。
|
||||
// 用于日志和后续审计归因。
|
||||
conversationTitleTokenAdjustReason = "conversation_title_async"
|
||||
)
|
||||
|
||||
const conversationTitlePrompt = `你是 SmartFlow 的会话标题生成器。
|
||||
@@ -190,7 +194,7 @@ func (s *AgentService) ensureConversationTitleAsync(userID int, chatID string) {
|
||||
}
|
||||
|
||||
// 4. 调用模型生成标题,并做格式清洗。
|
||||
generated, err := s.generateConversationTitle(ctx, history)
|
||||
generated, titleTokens, err := s.generateConversationTitle(ctx, history)
|
||||
if err != nil {
|
||||
log.Printf("异步生成会话标题失败(模型生成失败) chat=%s err=%v", chatID, err)
|
||||
return
|
||||
@@ -199,6 +203,28 @@ func (s *AgentService) ensureConversationTitleAsync(userID int, chatID string) {
|
||||
return
|
||||
}
|
||||
|
||||
// 4.1 标题生成成功后,把本次异步模型 token 记账:
|
||||
// 4.1.1 启用 outbox 时走 adjust 事件,异步可靠入账;
|
||||
// 4.1.2 未启用 outbox 时走同步兜底,直接更新账本。
|
||||
if titleTokens > 0 {
|
||||
if s.eventPublisher != nil {
|
||||
publishErr := eventsvc.PublishChatTokenUsageAdjustRequested(ctx, s.eventPublisher, model.ChatTokenUsageAdjustPayload{
|
||||
UserID: userID,
|
||||
ConversationID: chatID,
|
||||
TokensDelta: titleTokens,
|
||||
Reason: conversationTitleTokenAdjustReason,
|
||||
TriggeredAt: time.Now(),
|
||||
})
|
||||
if publishErr != nil {
|
||||
log.Printf("异步标题 token 记账事件发布失败 chat=%s tokens=%d err=%v", chatID, titleTokens, publishErr)
|
||||
}
|
||||
} else {
|
||||
if adjustErr := s.repo.AdjustTokenUsage(ctx, userID, chatID, titleTokens); adjustErr != nil {
|
||||
log.Printf("异步标题 token 同步记账失败 chat=%s tokens=%d err=%v", chatID, titleTokens, adjustErr)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 5. 只在标题仍为空时写入,保证并发幂等。
|
||||
if err = s.repo.UpdateConversationTitleIfEmpty(ctx, userID, chatID, generated); err != nil {
|
||||
log.Printf("异步生成会话标题失败(写库失败) chat=%s err=%v", chatID, err)
|
||||
@@ -207,17 +233,17 @@ func (s *AgentService) ensureConversationTitleAsync(userID int, chatID string) {
|
||||
}
|
||||
|
||||
// generateConversationTitle 使用聊天模型从近期历史生成标题。
|
||||
func (s *AgentService) generateConversationTitle(ctx context.Context, history []*schema.Message) (string, error) {
|
||||
func (s *AgentService) generateConversationTitle(ctx context.Context, history []*schema.Message) (string, int, error) {
|
||||
modelInst := s.pickTitleModel()
|
||||
if modelInst == nil {
|
||||
return "", fmt.Errorf("标题生成模型未初始化")
|
||||
return "", 0, fmt.Errorf("标题生成模型未初始化")
|
||||
}
|
||||
|
||||
// 1. 只取最近 N 条,降低 token 并聚焦当前会话主题。
|
||||
trimmed := tailMessages(history, conversationTitleHistoryLimit)
|
||||
prompt := buildConversationTitleUserPrompt(trimmed)
|
||||
if strings.TrimSpace(prompt) == "" {
|
||||
return "", fmt.Errorf("缺少可用历史内容")
|
||||
return "", 0, fmt.Errorf("缺少可用历史内容")
|
||||
}
|
||||
|
||||
messages := []*schema.Message{
|
||||
@@ -232,12 +258,22 @@ func (s *AgentService) generateConversationTitle(ctx context.Context, history []
|
||||
einoModel.WithMaxTokens(40),
|
||||
)
|
||||
if err != nil {
|
||||
return "", err
|
||||
return "", 0, err
|
||||
}
|
||||
if resp == nil {
|
||||
return "", fmt.Errorf("标题生成模型返回为空")
|
||||
return "", 0, fmt.Errorf("标题生成模型返回为空")
|
||||
}
|
||||
return normalizeConversationTitle(resp.Content), nil
|
||||
|
||||
// 2.1 标题链路的 token 从模型响应 usage 中提取;缺失则按 0 处理,不影响主流程。
|
||||
titleTokens := 0
|
||||
if resp.ResponseMeta != nil && resp.ResponseMeta.Usage != nil {
|
||||
titleTokens = normalizeUsageTotal(
|
||||
resp.ResponseMeta.Usage.TotalTokens,
|
||||
resp.ResponseMeta.Usage.PromptTokens,
|
||||
resp.ResponseMeta.Usage.CompletionTokens,
|
||||
)
|
||||
}
|
||||
return normalizeConversationTitle(resp.Content), titleTokens, nil
|
||||
}
|
||||
|
||||
// pickTitleModel 选择用于标题生成的模型。
|
||||
|
||||
@@ -345,6 +345,8 @@ func (s *AgentService) persistChatAfterReply(
|
||||
chatID string,
|
||||
userMessage string,
|
||||
assistantReply string,
|
||||
userTokens int,
|
||||
assistantTokens int,
|
||||
errChan chan error,
|
||||
) {
|
||||
// 1. 先把用户消息写入 Redis,保证会话上下文“马上可见”。
|
||||
@@ -358,6 +360,7 @@ func (s *AgentService) persistChatAfterReply(
|
||||
ConversationID: chatID,
|
||||
Role: "user",
|
||||
Message: userMessage,
|
||||
TokensConsumed: userTokens,
|
||||
}); err != nil {
|
||||
pushErrNonBlocking(errChan, err)
|
||||
return
|
||||
@@ -374,6 +377,7 @@ func (s *AgentService) persistChatAfterReply(
|
||||
ConversationID: chatID,
|
||||
Role: "assistant",
|
||||
Message: assistantReply,
|
||||
TokensConsumed: assistantTokens,
|
||||
}); err != nil {
|
||||
pushErrNonBlocking(errChan, err)
|
||||
}
|
||||
|
||||
145
backend/service/agentsvc/token_meter.go
Normal file
145
backend/service/agentsvc/token_meter.go
Normal file
@@ -0,0 +1,145 @@
|
||||
package agentsvc
|
||||
|
||||
import (
|
||||
"context"
|
||||
"sync"
|
||||
|
||||
einoCallbacks "github.com/cloudwego/eino/callbacks"
|
||||
einoModel "github.com/cloudwego/eino/components/model"
|
||||
"github.com/cloudwego/eino/schema"
|
||||
templatecb "github.com/cloudwego/eino/utils/callbacks"
|
||||
)
|
||||
|
||||
type requestTokenMeterCtxKey struct{}
|
||||
|
||||
// RequestTokenMeter 是“单次请求级”的 token 统计容器。
|
||||
//
|
||||
// 设计目标:
|
||||
// 1. 聚合本次请求内所有模型调用 token(路由/图节点/流式主对话);
|
||||
// 2. 线程安全,允许在同一请求内被多个链路节点并发累加;
|
||||
// 3. 最终由服务层一次性读取快照并写入持久化。
|
||||
type RequestTokenMeter struct {
|
||||
mu sync.Mutex
|
||||
|
||||
promptTokens int
|
||||
completionTokens int
|
||||
totalTokens int
|
||||
}
|
||||
|
||||
// RequestTokenMeterSnapshot 是 RequestTokenMeter 的只读快照。
|
||||
type RequestTokenMeterSnapshot struct {
|
||||
PromptTokens int
|
||||
CompletionTokens int
|
||||
TotalTokens int
|
||||
}
|
||||
|
||||
var registerTokenMeterCallbackOnce sync.Once
|
||||
|
||||
// ensureTokenMeterCallbackRegistered 注册一次全局 ChatModel callback。
|
||||
//
|
||||
// 说明:
|
||||
// 1. callback 只负责“采集并累加 token”,不做业务决策;
|
||||
// 2. 仅当 ctx 里存在 RequestTokenMeter 时才会生效;
|
||||
// 3. 采用 once,避免在测试/多次构造服务时重复注册。
|
||||
func ensureTokenMeterCallbackRegistered() {
|
||||
registerTokenMeterCallbackOnce.Do(func() {
|
||||
handler := templatecb.NewHandlerHelper().
|
||||
ChatModel(&templatecb.ModelCallbackHandler{
|
||||
OnEnd: func(ctx context.Context, _ *einoCallbacks.RunInfo, output *einoModel.CallbackOutput) context.Context {
|
||||
if output == nil || output.TokenUsage == nil {
|
||||
return ctx
|
||||
}
|
||||
addModelUsageIntoRequest(ctx, output.TokenUsage)
|
||||
return ctx
|
||||
},
|
||||
}).
|
||||
Handler()
|
||||
einoCallbacks.AppendGlobalHandlers(handler)
|
||||
})
|
||||
}
|
||||
|
||||
// withRequestTokenMeter 创建并挂载“请求级 token 统计器”。
|
||||
func withRequestTokenMeter(ctx context.Context) (context.Context, *RequestTokenMeter) {
|
||||
meter := &RequestTokenMeter{}
|
||||
return context.WithValue(ctx, requestTokenMeterCtxKey{}, meter), meter
|
||||
}
|
||||
|
||||
// getRequestTokenMeter 读取请求上下文中的 token 统计器。
|
||||
func getRequestTokenMeter(ctx context.Context) *RequestTokenMeter {
|
||||
if ctx == nil {
|
||||
return nil
|
||||
}
|
||||
meter, _ := ctx.Value(requestTokenMeterCtxKey{}).(*RequestTokenMeter)
|
||||
return meter
|
||||
}
|
||||
|
||||
// addSchemaUsageIntoRequest 把 schema usage 累加到请求级统计器。
|
||||
func addSchemaUsageIntoRequest(ctx context.Context, usage *schema.TokenUsage) {
|
||||
if usage == nil {
|
||||
return
|
||||
}
|
||||
addTokenUsageValues(ctx, usage.PromptTokens, usage.CompletionTokens, normalizeUsageTotal(usage.TotalTokens, usage.PromptTokens, usage.CompletionTokens))
|
||||
}
|
||||
|
||||
// addModelUsageIntoRequest 把 Eino model callback usage 累加到请求级统计器。
|
||||
func addModelUsageIntoRequest(ctx context.Context, usage *einoModel.TokenUsage) {
|
||||
if usage == nil {
|
||||
return
|
||||
}
|
||||
addTokenUsageValues(ctx, usage.PromptTokens, usage.CompletionTokens, normalizeUsageTotal(usage.TotalTokens, usage.PromptTokens, usage.CompletionTokens))
|
||||
}
|
||||
|
||||
// addTokenUsageValues 统一累加 token 数值。
|
||||
func addTokenUsageValues(ctx context.Context, promptTokens, completionTokens, totalTokens int) {
|
||||
meter := getRequestTokenMeter(ctx)
|
||||
if meter == nil {
|
||||
return
|
||||
}
|
||||
|
||||
if promptTokens < 0 {
|
||||
promptTokens = 0
|
||||
}
|
||||
if completionTokens < 0 {
|
||||
completionTokens = 0
|
||||
}
|
||||
if totalTokens < 0 {
|
||||
totalTokens = 0
|
||||
}
|
||||
|
||||
meter.mu.Lock()
|
||||
defer meter.mu.Unlock()
|
||||
meter.promptTokens += promptTokens
|
||||
meter.completionTokens += completionTokens
|
||||
meter.totalTokens += totalTokens
|
||||
}
|
||||
|
||||
// snapshotRequestTokenMeter 获取请求级 token 统计快照。
|
||||
func snapshotRequestTokenMeter(ctx context.Context) RequestTokenMeterSnapshot {
|
||||
meter := getRequestTokenMeter(ctx)
|
||||
if meter == nil {
|
||||
return RequestTokenMeterSnapshot{}
|
||||
}
|
||||
meter.mu.Lock()
|
||||
defer meter.mu.Unlock()
|
||||
return RequestTokenMeterSnapshot{
|
||||
PromptTokens: meter.promptTokens,
|
||||
CompletionTokens: meter.completionTokens,
|
||||
TotalTokens: meter.totalTokens,
|
||||
}
|
||||
}
|
||||
|
||||
// normalizeUsageTotal 统一 total token 口径。
|
||||
//
|
||||
// 规则:
|
||||
// 1. 模型返回 total>0 时优先使用 total;
|
||||
// 2. total 缺失时使用 prompt+completion 回退。
|
||||
func normalizeUsageTotal(totalTokens, promptTokens, completionTokens int) int {
|
||||
if totalTokens > 0 {
|
||||
return totalTokens
|
||||
}
|
||||
sum := promptTokens + completionTokens
|
||||
if sum < 0 {
|
||||
return 0
|
||||
}
|
||||
return sum
|
||||
}
|
||||
@@ -68,6 +68,7 @@ func RegisterChatHistoryPersistHandler(
|
||||
payload.ConversationID,
|
||||
payload.Role,
|
||||
payload.Message,
|
||||
payload.TokensConsumed,
|
||||
)
|
||||
})
|
||||
}
|
||||
|
||||
101
backend/service/events/chat_token_usage_adjust.go
Normal file
101
backend/service/events/chat_token_usage_adjust.go
Normal file
@@ -0,0 +1,101 @@
|
||||
package events
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"strconv"
|
||||
"time"
|
||||
|
||||
"github.com/LoveLosita/smartflow/backend/dao"
|
||||
kafkabus "github.com/LoveLosita/smartflow/backend/infra/kafka"
|
||||
outboxinfra "github.com/LoveLosita/smartflow/backend/infra/outbox"
|
||||
"github.com/LoveLosita/smartflow/backend/model"
|
||||
"gorm.io/gorm"
|
||||
)
|
||||
|
||||
const (
|
||||
// EventTypeChatTokenUsageAdjustRequested 是“会话 token 账本增量调整”事件类型。
|
||||
//
|
||||
// 命名约束:
|
||||
// 1. 仅表达业务语义,不泄露 outbox/kafka 实现细节;
|
||||
// 2. 作为稳定路由键长期保留,后续演进优先通过 event_version。
|
||||
EventTypeChatTokenUsageAdjustRequested = "chat.token.usage.adjust.requested"
|
||||
)
|
||||
|
||||
// RegisterChatTokenUsageAdjustHandler 注册“会话 token 账本增量调整”消费者。
|
||||
//
|
||||
// 职责边界:
|
||||
// 1. 只处理 token 调整事件,不处理聊天正文落库;
|
||||
// 2. 通过 outbox 统一消费事务入口,保证“业务成功 + consumed 推进”原子一致;
|
||||
// 3. 非法载荷直接标记 dead,避免无意义重试。
|
||||
func RegisterChatTokenUsageAdjustHandler(
|
||||
bus *outboxinfra.EventBus,
|
||||
outboxRepo *outboxinfra.Repository,
|
||||
repoManager *dao.RepoManager,
|
||||
) error {
|
||||
if bus == nil {
|
||||
return errors.New("event bus is nil")
|
||||
}
|
||||
if outboxRepo == nil {
|
||||
return errors.New("outbox repository is nil")
|
||||
}
|
||||
if repoManager == nil {
|
||||
return errors.New("repo manager is nil")
|
||||
}
|
||||
|
||||
handler := func(ctx context.Context, envelope kafkabus.Envelope) error {
|
||||
var payload model.ChatTokenUsageAdjustPayload
|
||||
if unmarshalErr := json.Unmarshal(envelope.Payload, &payload); unmarshalErr != nil {
|
||||
_ = outboxRepo.MarkDead(ctx, envelope.OutboxID, "解析会话 token 调整载荷失败: "+unmarshalErr.Error())
|
||||
return nil
|
||||
}
|
||||
|
||||
if payload.UserID <= 0 || payload.TokensDelta <= 0 || payload.ConversationID == "" {
|
||||
_ = outboxRepo.MarkDead(ctx, envelope.OutboxID, "会话 token 调整载荷无效: user_id/conversation_id/tokens_delta 非法")
|
||||
return nil
|
||||
}
|
||||
|
||||
return outboxRepo.ConsumeAndMarkConsumed(ctx, envelope.OutboxID, func(tx *gorm.DB) error {
|
||||
txM := repoManager.WithTx(tx)
|
||||
return txM.Agent.AdjustTokenUsageInTx(ctx, payload.UserID, payload.ConversationID, payload.TokensDelta)
|
||||
})
|
||||
}
|
||||
|
||||
return bus.RegisterEventHandler(EventTypeChatTokenUsageAdjustRequested, handler)
|
||||
}
|
||||
|
||||
// PublishChatTokenUsageAdjustRequested 发布“会话 token 账本增量调整”事件。
|
||||
//
|
||||
// 说明:
|
||||
// 1. 只保证“写入 outbox 成功”,不等待消费完成;
|
||||
// 2. 业务层只传 DTO,不关心 outbox/kafka 协议细节。
|
||||
func PublishChatTokenUsageAdjustRequested(
|
||||
ctx context.Context,
|
||||
publisher outboxinfra.EventPublisher,
|
||||
payload model.ChatTokenUsageAdjustPayload,
|
||||
) error {
|
||||
if publisher == nil {
|
||||
return errors.New("event publisher is nil")
|
||||
}
|
||||
if payload.UserID <= 0 {
|
||||
return errors.New("invalid user_id")
|
||||
}
|
||||
if payload.TokensDelta <= 0 {
|
||||
return errors.New("invalid tokens_delta")
|
||||
}
|
||||
if payload.ConversationID == "" {
|
||||
return errors.New("invalid conversation_id")
|
||||
}
|
||||
if payload.TriggeredAt.IsZero() {
|
||||
payload.TriggeredAt = time.Now()
|
||||
}
|
||||
|
||||
return publisher.Publish(ctx, outboxinfra.PublishRequest{
|
||||
EventType: EventTypeChatTokenUsageAdjustRequested,
|
||||
EventVersion: outboxinfra.DefaultEventVersion,
|
||||
MessageKey: payload.ConversationID,
|
||||
AggregateID: strconv.Itoa(payload.UserID) + ":" + payload.ConversationID,
|
||||
Payload: payload,
|
||||
})
|
||||
}
|
||||
@@ -14,6 +14,7 @@ import (
|
||||
"github.com/LoveLosita/smartflow/backend/respond"
|
||||
eventsvc "github.com/LoveLosita/smartflow/backend/service/events"
|
||||
"github.com/go-redis/redis/v8"
|
||||
"gorm.io/gorm"
|
||||
)
|
||||
|
||||
const (
|
||||
@@ -72,6 +73,46 @@ func (ts *TaskService) AddTask(ctx context.Context, req *model.UserAddTaskReques
|
||||
return response, nil
|
||||
}
|
||||
|
||||
// CompleteTask 将用户指定任务标记为“已完成”。
|
||||
//
|
||||
// 职责边界:
|
||||
// 1. 负责入参校验与业务错误映射;
|
||||
// 2. 负责调用 DAO 执行状态更新;
|
||||
// 3. 不负责幂等键校验(幂等由中间件处理);
|
||||
// 4. 不负责缓存删除细节(缓存删除由 GORM cache_deleter 回调触发)。
|
||||
func (ts *TaskService) CompleteTask(ctx context.Context, req *model.UserCompleteTaskRequest, userID int) (*model.UserCompleteTaskResponse, error) {
|
||||
// 1. 参数兜底:请求体为空、非法 user 或非法 task_id 直接返回业务错误。
|
||||
if req == nil || userID <= 0 || req.TaskID <= 0 {
|
||||
return nil, respond.WrongTaskID
|
||||
}
|
||||
|
||||
// 2. 调用 DAO 执行“查询 + 必要时更新”。
|
||||
updatedTask, alreadyCompleted, err := ts.dao.CompleteTaskByID(ctx, userID, req.TaskID)
|
||||
if err != nil {
|
||||
// 2.1 任务不存在或不属于当前用户时,统一映射为 WrongTaskID。
|
||||
if errors.Is(err, gorm.ErrRecordNotFound) {
|
||||
return nil, respond.WrongTaskID
|
||||
}
|
||||
// 2.2 其余数据库异常向上透传,交由统一错误处理器返回 500。
|
||||
return nil, err
|
||||
}
|
||||
if updatedTask == nil {
|
||||
// 3. 极端防御:DAO 不应返回 nil,若发生则视为内部异常。
|
||||
return nil, errors.New("complete task succeeded but task is nil")
|
||||
}
|
||||
|
||||
// 4. 构造响应:
|
||||
// 4.1 already_completed=true 表示本次命中幂等,不影响最终成功状态;
|
||||
// 4.2 is_completed 始终为 true,便于前端直接刷新状态。
|
||||
resp := &model.UserCompleteTaskResponse{
|
||||
TaskID: updatedTask.ID,
|
||||
IsCompleted: true,
|
||||
AlreadyCompleted: alreadyCompleted,
|
||||
Status: "completed",
|
||||
}
|
||||
return resp, nil
|
||||
}
|
||||
|
||||
// GetUserTasks 获取用户任务列表(含“读时紧急性派生”与“异步平移触发”)。
|
||||
//
|
||||
// 核心流程(步骤化):
|
||||
|
||||
Reference in New Issue
Block a user