Version: 0.9.80.dev.260506

后端：
1. LLM 独立服务与统一计费出口落地：新增 `cmd/llm`、`client/llm` 与 `services/llm/rpc`，补齐 BillingContext、CreditBalanceGuard、价格规则解析、stream usage 归集与 `credit.charge.requested` outbox 发布，active-scheduler / agent / course / memory / gateway fallback 全部改走 llm zrpc，不再各自本地初始化模型。
2. TokenStore 收口为 Credit 权威账本：新增 credit account / ledger / product / order / price-rule / reward-rule 能力与 Redis 快照缓存，扩展 tokenstore rpc/client 支撑余额快照、消耗看板、商品、订单、流水、价格规则和奖励规则，并接入 LLM charge 事件消费完成 Credit 扣费落账。
3. 计费旧链路下线与网关切口切换：`/token-store` 语义整体切到 `/credit-store`，agent chat 移除旧 TokenQuotaGuard，userauth 的 CheckTokenQuota / AdjustTokenUsage 改为废弃，聊天历史落库不再同步旧 token 额度账本，course 图片解析请求补 user_id 进入新计费口径。

前端：
4. 计划广场从 mock 数据切到真实接口：新增 forum api/types，首页支持真实列表、标签、搜索、防抖、点赞、导入和发布计划，详情页补齐帖子详情、评论树、回复和删除评论链路，同时补上“至少一个标签”的前后端约束与默认标签兜底。
5. 商店页切到 Credit 体系并重做展示：顶部改为余额 + Credit/Token 消耗看板，支持 24h/7d/30d/all 周期切换；套餐区展示原价与当前价；历史区改为当前用户 Credit 流水并支持查看更多，整体视觉和交互同步收口。

仓库：
6. 配置与本地启动体系补齐 llm / outbox 编排：`config.example.yaml` 增加 llm rpc 和统一 outbox service 配置，`dev-common.ps1` 把 llm 纳入多服务依赖并自动建 Kafka topic，`docker-compose.yml` 同步初始化 agent/task/memory/active-scheduler/notification/taskclass-forum/llm/token-store 全量 outbox topic。

backend/services/memory/internal/orchestrator/llm_decision_orchestrator.go

@@ -41,6 +41,7 @@ func NewLLMDecisionOrchestrator(client *llmservice.Client, cfg memorymodel.Confi
 // 3. 不做最终决策，最终动作由确定性汇总逻辑产出。
 func (o *LLMDecisionOrchestrator) Compare(
 	ctx context.Context,
+	billing llmservice.BillingContext,
 	fact memorymodel.NormalizedFact,
 	candidate memorymodel.CandidateSnapshot,
 ) (*memorymodel.ComparisonResult, error) {
@@ -53,10 +54,11 @@ func (o *LLMDecisionOrchestrator) Compare(
 	userPrompt := buildDecisionCompareUserPrompt(fact, candidate)
 
 	messages := llmservice.BuildSystemUserMessages(systemPrompt, nil, userPrompt)
+	invokeCtx := llmservice.WithBillingContext(ctx, billing)
 
 	// 2. 调用 LLM 做结构化输出，温度用低值保证判断稳定。
 	resp, _, err := llmservice.GenerateJSON[decisionCompareResponse](
-		ctx,
+		invokeCtx,
 		o.client,
 		messages,
 		llmservice.GenerateOptions{

backend/services/memory/internal/orchestrator/llm_write_orchestrator.go

@@ -59,9 +59,10 @@ func (o *LLMWriteOrchestrator) ExtractFacts(ctx context.Context, payload memorym
 		nil,
 		buildMemoryExtractUserPrompt(payload),
 	)
+	invokeCtx := llmservice.WithBillingContext(ctx, buildMemoryExtractBillingContext(payload))
 
 	resp, rawResult, err := llmservice.GenerateJSON[memoryExtractResponse](
-		ctx,
+		invokeCtx,
 		o.client,
 		messages,
 		llmservice.GenerateOptions{
@@ -329,3 +330,18 @@ func truncateForLog(raw *llmservice.TextResult) string {
 	}
 	return text[:200] + "..."
 }
+
+func buildMemoryExtractBillingContext(payload memorymodel.ExtractJobPayload) llmservice.BillingContext {
+	requestID := strings.TrimSpace(payload.TraceID)
+	if requestID == "" {
+		requestID = fmt.Sprintf("memory_extract:%d:%s:%d", payload.UserID, strings.TrimSpace(payload.ConversationID), payload.SourceMessageID)
+	}
+	return llmservice.BillingContext{
+		UserID:         uint64(payload.UserID),
+		EventID:        requestID,
+		Scene:          "memory_extract",
+		RequestID:      requestID,
+		ConversationID: strings.TrimSpace(payload.ConversationID),
+		ModelAlias:     "memory_extract",
+	}
+}

backend/services/memory/internal/worker/decision_flow.go

@@ -3,7 +3,9 @@ package worker
 import (
 	"context"
 	"fmt"
+	"strings"
 
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	memoryrepo "github.com/LoveLosita/smartflow/backend/services/memory/internal/repo"
 	memoryutils "github.com/LoveLosita/smartflow/backend/services/memory/internal/utils"
 	memorymodel "github.com/LoveLosita/smartflow/backend/services/memory/model"
@@ -144,7 +146,7 @@ func (r *Runner) executeDecisionForFact(
 	}
 
 	// Step 3: 逐对 LLM 比对。
-	comparisons := r.compareWithCandidates(ctx, fact, candidates)
+	comparisons := r.compareWithCandidates(ctx, payload, fact, candidates)
 
 	// Step 4: 确定性汇总。
 	decision := memoryutils.AggregateComparisons(fact, comparisons, candidates)
@@ -298,6 +300,7 @@ func (r *Runner) recallCandidatesFromMySQL(
 // 3. 无候选或决策编排器为空时返回空切片，上层直接走 ADD 路径。
 func (r *Runner) compareWithCandidates(
 	ctx context.Context,
+	payload memorymodel.ExtractJobPayload,
 	fact memorymodel.NormalizedFact,
 	candidates []memorymodel.CandidateSnapshot,
 ) []memorymodel.ComparisonResult {
@@ -307,7 +310,7 @@ func (r *Runner) compareWithCandidates(
 
 	comparisons := make([]memorymodel.ComparisonResult, 0, len(candidates))
 	for _, candidate := range candidates {
-		compResult, err := r.decisionOrchestrator.Compare(ctx, fact, candidate)
+		compResult, err := r.decisionOrchestrator.Compare(ctx, buildMemoryDecisionBillingContext(payload, fact, candidate), fact, candidate)
 		if err != nil {
 			// LLM 调用失败 → 视为 unrelated，不影响其他候选。
 			if r.logger != nil {
@@ -335,6 +338,26 @@ func (r *Runner) compareWithCandidates(
 	return comparisons
 }
 
+func buildMemoryDecisionBillingContext(
+	payload memorymodel.ExtractJobPayload,
+	fact memorymodel.NormalizedFact,
+	candidate memorymodel.CandidateSnapshot,
+) llmservice.BillingContext {
+	requestID := strings.TrimSpace(payload.TraceID)
+	if requestID == "" {
+		requestID = fmt.Sprintf("memory_decision:%d:%s:%d", payload.UserID, strings.TrimSpace(payload.ConversationID), payload.SourceMessageID)
+	}
+	eventID := fmt.Sprintf("%s:%d:%s", requestID, candidate.MemoryID, fact.ContentHash)
+	return llmservice.BillingContext{
+		UserID:         uint64(payload.UserID),
+		EventID:        eventID,
+		Scene:          "memory_decision_compare",
+		RequestID:      requestID,
+		ConversationID: strings.TrimSpace(payload.ConversationID),
+		ModelAlias:     "memory_decision_compare",
+	}
+}
+
 // collectActionOutcome 汇总单个动作结果到全局 outcome。
 func (r *Runner) collectActionOutcome(outcome *DecisionFlowOutcome, actionOutcome *ApplyActionOutcome) {
 	if actionOutcome == nil {