模块介绍 / 今日规划 / 下班日报 / 团队摘要 生成成功后没有上报，也几乎不写「AI」日志。

流结束时的 usage 用了非阻塞发送，通道一满就被丢掉；很多服务商默认还不在流里带 usage。
现在每次真正打到模型的调用都会：

在本地日志（分类 AI，消息 AI 调用完成）写下 scene / prompt / completion / cached
登录后同步上报到后台日汇总

service/ai/ai_test.go

@@ -56,6 +56,51 @@ func TestChatStreamSSE(t *testing.T) {
 	}
 }
 
+func TestParseUsageCachedTokens(t *testing.T) {
+	u := parseUsage(&chatUsage{
+		PromptTokens:         100,
+		CompletionTokens:     20,
+		PromptCacheHitTokens: 80,
+	})
+	if u == nil || u.CachedTokens != 80 || u.PromptTokens != 100 {
+		t.Fatalf("deepseek cache field: %+v", u)
+	}
+	u = parseUsage(&chatUsage{
+		PromptTokens:     90,
+		CompletionTokens: 10,
+		PromptTokensDetails: &struct {
+			CachedTokens int64 `json:"cached_tokens"`
+		}{CachedTokens: 64},
+	})
+	if u == nil || u.CachedTokens != 64 {
+		t.Fatalf("openai cached_tokens: %+v", u)
+	}
+}
+
+func TestChatStreamSendsUsage(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.Write([]byte("data: {\"choices\":[{\"delta\":{\"content\":\"ok\"}}]}\n\n"))
+		w.Write([]byte("data: {\"usage\":{\"prompt_tokens\":12,\"completion_tokens\":3,\"prompt_cache_hit_tokens\":8}}\n\n"))
+		w.Write([]byte("data: [DONE]\n\n"))
+	}))
+	defer srv.Close()
+	p := &openAICompatible{name: "test", baseURL: srv.URL, model: "m", apiKey: "k", client: srv.Client()}
+	stream, e := p.ChatStream(context.Background(), []Message{{Role: "user", Content: "hi"}})
+	if e != nil {
+		t.Fatal(e)
+	}
+	var usage *Usage
+	for c := range stream {
+		if c.Usage != nil {
+			usage = c.Usage
+		}
+	}
+	if usage == nil || usage.PromptTokens != 12 || usage.CachedTokens != 8 {
+		t.Fatalf("usage not delivered: %+v", usage)
+	}
+}
+
 func TestChatStreamAuthError(t *testing.T) {
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(http.StatusUnauthorized)

service/ai/openai.go

@@ -26,9 +26,14 @@ type openAICompatible struct {
 func (p *openAICompatible) Name() string { return p.name }
 
 type chatRequest struct {
-	Model    string    `json:"model"`
-	Messages []Message `json:"messages"`
-	Stream   bool      `json:"stream"`
+	Model         string         `json:"model"`
+	Messages      []Message      `json:"messages"`
+	Stream        bool           `json:"stream"`
+	StreamOptions *streamOptions `json:"stream_options,omitempty"`
+}
+
+type streamOptions struct {
+	IncludeUsage bool `json:"include_usage"`
 }
 
 type chatDelta struct {
@@ -37,17 +42,42 @@ type chatDelta struct {
 			Content string `json:"content"`
 		} `json:"delta"`
 	} `json:"choices"`
-	Usage *struct {
-		PromptTokens     int64 `json:"prompt_tokens"`
-		CompletionTokens int64 `json:"completion_tokens"`
-	} `json:"usage"`
+	Usage *chatUsage `json:"usage"`
 	Error *struct {
 		Message string `json:"message"`
 	} `json:"error"`
 }
 
+// chatUsage 兼容 OpenAI / DeepSeek / 星火：缓存命中字段名不统一。
+type chatUsage struct {
+	PromptTokens         int64 `json:"prompt_tokens"`
+	CompletionTokens     int64 `json:"completion_tokens"`
+	PromptCacheHitTokens int64 `json:"prompt_cache_hit_tokens"`
+	PromptTokensDetails  *struct {
+		CachedTokens int64 `json:"cached_tokens"`
+	} `json:"prompt_tokens_details"`
+}
+
+func parseUsage(u *chatUsage) *Usage {
+	if u == nil {
+		return nil
+	}
+	cached := u.PromptCacheHitTokens
+	if u.PromptTokensDetails != nil && u.PromptTokensDetails.CachedTokens > cached {
+		cached = u.PromptTokensDetails.CachedTokens
+	}
+	return &Usage{
+		PromptTokens:     u.PromptTokens,
+		CompletionTokens: u.CompletionTokens,
+		CachedTokens:     cached,
+	}
+}
+
 func (p *openAICompatible) ChatStream(ctx context.Context, messages []Message) (<-chan Chunk, error) {
-	body, e := json.Marshal(chatRequest{Model: p.model, Messages: messages, Stream: true})
+	body, e := json.Marshal(chatRequest{
+		Model: p.model, Messages: messages, Stream: true,
+		StreamOptions: &streamOptions{IncludeUsage: true},
+	})
 	if e != nil {
 		return nil, e
 	}
@@ -100,10 +130,7 @@ func (p *openAICompatible) ChatStream(ctx context.Context, messages []Message) (
 				return
 			}
 			if d.Usage != nil {
-				lastUsage = &Usage{
-					PromptTokens:     d.Usage.PromptTokens,
-					CompletionTokens: d.Usage.CompletionTokens,
-				}
+				lastUsage = parseUsage(d.Usage)
 			}
 			if len(d.Choices) > 0 && d.Choices[0].Delta.Content != "" {
 				completionChars += len(d.Choices[0].Delta.Content)
@@ -119,7 +146,6 @@ func (p *openAICompatible) ChatStream(ctx context.Context, messages []Message) (
 			return
 		}
 		if lastUsage == nil {
-			// 粗估：约 4 字符 ≈ 1 token
 			lastUsage = &Usage{
 				PromptTokens:     int64((promptChars + 3) / 4),
 				CompletionTokens: int64((completionChars + 3) / 4),
@@ -128,7 +154,7 @@ func (p *openAICompatible) ChatStream(ctx context.Context, messages []Message) (
 		}
 		select {
 		case out <- Chunk{Usage: lastUsage}:
-		default:
+		case <-ctx.Done():
 		}
 	}()
 	return out, nil

service/ai/provider.go

@@ -22,9 +22,23 @@ type Chunk struct {
 type Usage struct {
 	PromptTokens     int64
 	CompletionTokens int64
+	CachedTokens     int64 // 提示缓存命中（服务商返回时才有）
 	Estimated        bool
 }
 
+// EstimateUsage 在服务商未回 usage 时按字符粗估（约 4 字符 ≈ 1 token）。
+func EstimateUsage(messages []Message, completion string) *Usage {
+	n := 0
+	for _, m := range messages {
+		n += len(m.Content)
+	}
+	return &Usage{
+		PromptTokens:     int64((n + 3) / 4),
+		CompletionTokens: int64((len(completion) + 3) / 4),
+		Estimated:        true,
+	}
+}
+
 // Provider 是 AI 服务商的统一抽象。
 type Provider interface {
 	Name() string