feat: 会话亲和粘住缓存渠道；透传 opaque 字段；用量记思考/1h 缓存写/实际模型

会话亲和（proxy/affinity.go）：同一会话粘住上次应答的渠道，命中上游提示词
缓存而非被加权随机打散后全价重算。会话身份 = 密钥 + 模型 + 请求文本稳定前缀
摘要，不依赖各客户端会话头；TTL 默认 5m（OT_PROXY_AFFINITY_TTL，0=关闭），
成功应答刷新。亲和渠道必须在候选集内（健康/未冷却/绑定），否则回退加权随机。
proxyOne 增加 served 返回值区分「应答成功」与「已终结但失败」。

opaque 透传（convert）：chat↔responses 互转原样携带 metadata、service_tier、
parallel_tool_calls。此前逐字段重建请求体会丢它们，而中转上游会校验/使用。

用量细化（proxy+store）：usage_logs 新增 reasoning_tokens（展示口径，不重复
计费）、cache_write_1h_tokens（Anthropic 1h 缓存写按 2× 输入价，此前统一
1.25× 少算 60%）、served_model（上游实际应答模型，映射/聚合商场景对账）。
schema v6→v7。

验证：go vet ./... 干净；go test ./... 与 go test -race ./... 全绿；
新增测试覆盖亲和粘性与冷却回退、opaque 双向透传、reasoning 不重复计费、
1h/5m 缓存写分开计价、served_model 提取（非流式与流式）。

server/internal/cli/serve.go

@@ -39,6 +39,7 @@ func serve() error {
 		FailoverBudget:  cfg.Proxy.FailoverBudget,
 		StreamHeartbeat: cfg.Proxy.StreamHeartbeat,
 		FirstTokenWait:  cfg.Proxy.FirstTokenWait,
+		AffinityTTL:     cfg.Proxy.AffinityTTL,
 	})
 	router := api.NewRouter(a, gw)
 

server/internal/config/config.go

@@ -82,6 +82,7 @@ type ProxyConfig struct {
 	FailoverBudget      time.Duration // 故障转移给"第二个及以后的渠道"留的时间预算，0=不限
 	StreamHeartbeat     time.Duration // 流式空闲多久发一次 SSE 注释心跳，0=关闭
 	FirstTokenWait      time.Duration // 流式放行前等内容的最长时间，0=不限；超时未出内容换下一个渠道
+	AffinityTTL         time.Duration // 会话亲和 TTL，0=关闭；同一会话粘住上次应答的渠道以命中上游缓存
 	HealthInterval      time.Duration // 渠道健康检查周期
 	HealthFailThreshold int           // 连续失败 N 次进 cooldown
 	LogRaw              bool          // 记录管理员原始请求体+响应到 usage_logs（调试用，默认关）
@@ -150,6 +151,7 @@ func Load() (*Config, error) {
 	v.SetDefault("proxy.failover_budget", "30s")
 	v.SetDefault("proxy.stream_heartbeat", "15s")
 	v.SetDefault("proxy.first_token_wait", "0s")
+	v.SetDefault("proxy.affinity_ttl", "5m")
 	v.SetDefault("proxy.health_interval", "60s")
 	v.SetDefault("proxy.health_fail_threshold", 2)
 	v.SetDefault("proxy.log_raw", false)
@@ -200,6 +202,7 @@ func Load() (*Config, error) {
 			FailoverBudget:      v.GetDuration("proxy.failover_budget"),
 			StreamHeartbeat:     v.GetDuration("proxy.stream_heartbeat"),
 			FirstTokenWait:      v.GetDuration("proxy.first_token_wait"),
+			AffinityTTL:         v.GetDuration("proxy.affinity_ttl"),
 			HealthInterval:      v.GetDuration("proxy.health_interval"),
 			HealthFailThreshold: v.GetInt("proxy.health_fail_threshold"),
 			LogRaw:              v.GetBool("proxy.log_raw"),

server/internal/proxy/affinity.go

@@ -0,0 +1,84 @@
+package proxy
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"fmt"
+	"sync"
+	"time"
+)
+
+// 会话亲和（借鉴 magpie 的 cache-economics 思路，做服务端可用的简化版）：
+// 同一会话粘住上次应答的渠道，让上游的提示词缓存（整段上下文每轮重发）被再次
+// 命中，而不是被加权随机打散后全价重算。
+//
+// 会话身份 = 密钥 + 模型 + 请求文本稳定前缀的摘要。编码代理每轮重发全量上下文，
+// 系统提示与最早的用户消息不变，取前缀即可区分会话；不需要解析各家客户端的
+// 会话头。命中亲和的渠道仍必须在候选集里（健康、未冷却、认识该模型）——
+// 冷却或下线后自然回退到加权随机。
+//
+// TTL 默认 5 分钟（对齐 Anthropic 提示词缓存窗口），每次成功应答刷新；
+// 上游不报缓存的（OpenAI 自动前缀缓存）同样受益：粘住就是省。
+
+const (
+	// affinityMax 亲和表上限：超过先清过期，仍超则不再新增（防御性，正常流量到不了）。
+	affinityMax = 8192
+	// affinityPrefix 会话身份摘要取请求文本的前缀字节数。
+	affinityPrefix = 512
+)
+
+type affinityTable struct {
+	mu sync.Mutex
+	m  map[string]affinityEntry
+}
+
+type affinityEntry struct {
+	channelID uint64
+	until     time.Time
+}
+
+func newAffinityTable() *affinityTable {
+	return &affinityTable{m: map[string]affinityEntry{}}
+}
+
+// get 返回该会话上次应答的渠道；过期条目顺手清理。
+func (t *affinityTable) get(key string) (uint64, bool) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	e, ok := t.m[key]
+	if !ok {
+		return 0, false
+	}
+	if time.Now().After(e.until) {
+		delete(t.m, key)
+		return 0, false
+	}
+	return e.channelID, true
+}
+
+// put 记录（或刷新）会话→渠道的绑定。
+func (t *affinityTable) put(key string, channelID uint64, ttl time.Duration) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if len(t.m) >= affinityMax {
+		now := time.Now()
+		for k, e := range t.m {
+			if now.After(e.until) {
+				delete(t.m, k)
+			}
+		}
+		if len(t.m) >= affinityMax {
+			return // 表满且全是活的：放弃本次记录，旧绑定继续服务
+		}
+	}
+	t.m[key] = affinityEntry{channelID: channelID, until: time.Now().Add(ttl)}
+}
+
+// affinityKey 会话身份：密钥 + 模型 + 请求文本稳定前缀的摘要。
+func affinityKey(keyID uint64, model, firstText string) string {
+	if len(firstText) > affinityPrefix {
+		firstText = firstText[:affinityPrefix]
+	}
+	h := sha256.Sum256([]byte(firstText))
+	return fmt.Sprintf("%d|%s|%s", keyID, model, hex.EncodeToString(h[:8]))
+}

server/internal/proxy/affinity_test.go

@@ -0,0 +1,87 @@
+package proxy
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/openteam/server/internal/channel"
+	"github.com/openteam/server/internal/store"
+)
+
+// 会话亲和：同一会话（同密钥+模型+同首条消息）粘住上次应答的渠道——
+// 即使加权随机此时会选别的渠道；不同会话不受影响，照常负载均衡。
+func TestAffinitySticksToAnsweringChannel(t *testing.T) {
+	var ch1Hits, ch2Hits atomic.Int64
+	ch1 := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		ch1Hits.Add(1)
+		_, _ = io.WriteString(w, `{"id":"ok","choices":[{"message":{"role":"assistant","content":"from-1"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`)
+	}))
+	defer ch1.Close()
+	ch2 := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		ch2Hits.Add(1)
+		_, _ = io.WriteString(w, `{"id":"ok","choices":[{"message":{"role":"assistant","content":"from-2"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`)
+	}))
+	defer ch2.Close()
+
+	h := newHarness(t, ch1.URL, 0, true, nil)
+	h.gw.affinityTTL = 5 * time.Minute
+	addChannel(t, h, "c2", ch2.URL, 1)
+	pushWeightTo(t, h, "c1", 1000)
+
+	sameConv := `{"model":"m1","messages":[{"role":"user","content":"会话的第一条消息"}]}`
+	// 第一轮：权重压在渠道1，应答来自渠道1，亲和绑定建立
+	if w := h.run(t, nil, sameConv); w.Code != http.StatusOK || !strings.Contains(w.Body.String(), "from-1") {
+		t.Fatalf("第一轮应来自渠道1：%d %s", w.Code, w.Body.String())
+	}
+	// 反转权重：加权随机此刻会选渠道2
+	pushWeightTo(t, h, "c1", 1)
+	pushWeightTo(t, h, "c2", 1000)
+
+	// 同一会话第二轮：仍应粘在渠道1（缓存值得保留）
+	if w := h.run(t, nil, sameConv); !strings.Contains(w.Body.String(), "from-1") {
+		t.Fatalf("同一会话应粘住渠道1：%s", w.Body.String())
+	}
+	// 不同会话（不同首条消息）：走加权随机，落到渠道2
+	otherConv := `{"model":"m1","messages":[{"role":"user","content":"另一个会话的开场"}]}`
+	if w := h.run(t, nil, otherConv); !strings.Contains(w.Body.String(), "from-2") {
+		t.Fatalf("不同会话应走加权随机落到渠道2：%s", w.Body.String())
+	}
+	if ch1Hits.Load() != 2 || ch2Hits.Load() != 1 {
+		t.Fatalf("命中次数不符：ch1=%d ch2=%d", ch1Hits.Load(), ch2Hits.Load())
+	}
+}
+
+// 亲和渠道进入运行期冷却后自动失效，回退加权随机；不会把请求发给冷却中的渠道。
+func TestAffinityRespectsCooldown(t *testing.T) {
+	ch1 := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Error("冷却中的渠道1 不应收到请求")
+		_, _ = io.WriteString(w, `{"id":"bad"}`)
+	}))
+	defer ch1.Close()
+	ch2 := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.WriteString(w, `{"id":"ok","choices":[{"message":{"role":"assistant","content":"from-2"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`)
+	}))
+	defer ch2.Close()
+
+	h := newHarness(t, ch1.URL, 0, true, nil)
+	h.gw.affinityTTL = 5 * time.Minute
+	addChannel(t, h, "c2", ch2.URL, 1)
+
+	// 预置亲和绑定指向渠道1，随后让渠道1 进运行期冷却
+	var ch1row store.Channel
+	h.db.Where("name = ?", "c1").First(&ch1row)
+	h.gw.aff.put(affinityKey(1, "m1", "会话的第一条消息"), ch1row.ID, 5*time.Minute)
+	for i := 0; i < 3; i++ {
+		h.gw.ch.ReportFailure(&ch1row, channel.FailRate, 0)
+	}
+
+	w := h.run(t, nil, `{"model":"m1","messages":[{"role":"user","content":"会话的第一条消息"}]}`)
+	if !strings.Contains(w.Body.String(), "from-2") {
+		t.Fatalf("冷却后亲和应失效并切换渠道2：%d %s", w.Code, w.Body.String())
+	}
+}

server/internal/proxy/convert/convert_test.go

@@ -1136,3 +1136,49 @@ func TestStreamResponsesToChatTextStillWorks(t *testing.T) {
 		t.Fatalf("missing finish chunk: %s", out)
 	}
 }
+
+// OpenAI 两协议同形的 opaque 字段（metadata / service_tier / parallel_tool_calls）
+// 在 chat↔responses 互转时原样透传——上游可能是另一层中转，会校验/使用它们。
+func TestChatToResponsesCarriesOpaque(t *testing.T) {
+	body := `{"model":"m1","messages":[{"role":"user","content":"hi"}],"service_tier":"priority","metadata":{"session_id":"s1"},"parallel_tool_calls":false}`
+	out, err := ConvertRequest([]byte(body), ProtoChat, ProtoResponses)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if json.Unmarshal(out, &m) != nil {
+		t.Fatal("invalid json")
+	}
+	if m["service_tier"] != "priority" {
+		t.Fatalf("service_tier 丢失：%v", m["service_tier"])
+	}
+	md, _ := m["metadata"].(map[string]any)
+	if md == nil || md["session_id"] != "s1" {
+		t.Fatalf("metadata 丢失：%v", m["metadata"])
+	}
+	if m["parallel_tool_calls"] != false {
+		t.Fatalf("parallel_tool_calls 丢失：%v", m["parallel_tool_calls"])
+	}
+}
+
+func TestResponsesToChatCarriesOpaque(t *testing.T) {
+	body := `{"model":"m1","input":"hi","service_tier":"flex","metadata":{"installation_id":"i9"},"parallel_tool_calls":true}`
+	out, err := ConvertRequest([]byte(body), ProtoResponses, ProtoChat)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if json.Unmarshal(out, &m) != nil {
+		t.Fatal("invalid json")
+	}
+	if m["service_tier"] != "flex" {
+		t.Fatalf("service_tier 丢失：%v", m["service_tier"])
+	}
+	md, _ := m["metadata"].(map[string]any)
+	if md == nil || md["installation_id"] != "i9" {
+		t.Fatalf("metadata 丢失：%v", m["metadata"])
+	}
+	if m["parallel_tool_calls"] != true {
+		t.Fatalf("parallel_tool_calls 丢失：%v", m["parallel_tool_calls"])
+	}
+}

server/internal/proxy/convert/json_chat.go

@@ -39,6 +39,12 @@ type chatReq struct {
 	MaxTokens   *int       `json:"max_tokens"`
 	Stop        []string   `json:"stop"`
 	Stream      bool       `json:"stream"`
+	// OpenAI 两协议同形的 opaque 字段：转换到 Responses 时原样透传。
+	// 上游可能是另一层中转，会校验/使用它们（service_tier 影响优先级，
+	// metadata 承载 Codex 等客户端的会话/安装标识）。
+	ServiceTier       string            `json:"service_tier"`
+	Metadata          map[string]string `json:"metadata"`
+	ParallelToolCalls *bool             `json:"parallel_tool_calls"`
 }
 
 // chatToMessagesReq 将 OpenAI Chat 请求转为 Anthropic Messages 请求。

server/internal/proxy/convert/json_responses.go

@@ -27,6 +27,13 @@ func responsesToChatReq(body []byte) ([]byte, error) {
 	if v, ok := m["max_output_tokens"]; ok {
 		out["max_tokens"] = v
 	}
+	// OpenAI 两协议同形的 opaque 字段原样透传：metadata（map[string]string）、
+	// service_tier、parallel_tool_calls——丢了它们，中转上游的校验与路由会失效
+	for _, k := range []string{"metadata", "service_tier", "parallel_tool_calls"} {
+		if v := rawJSON(m, k); v != nil {
+			out[k] = json.RawMessage(v)
+		}
+	}
 
 	var msgs []any
 	if ins := str(rawJSON(m, "instructions")); ins != "" {
@@ -199,6 +206,16 @@ func chatToResponsesReq(body []byte) ([]byte, error) {
 	if req.MaxTokens != nil {
 		out["max_output_tokens"] = *req.MaxTokens
 	}
+	// OpenAI 两协议同形的 opaque 字段原样透传（见 chatReq 注释）
+	if req.ServiceTier != "" {
+		out["service_tier"] = req.ServiceTier
+	}
+	if len(req.Metadata) > 0 {
+		out["metadata"] = req.Metadata
+	}
+	if req.ParallelToolCalls != nil {
+		out["parallel_tool_calls"] = *req.ParallelToolCalls
+	}
 
 	var system []string
 	var input []any

server/internal/proxy/gateway.go

@@ -50,8 +50,11 @@ type Gateway struct {
 	streamHeartbeat time.Duration
 	// 流式放行前等多久内容（0=不限制）。超时未出内容取消本次尝试换下一个渠道。
 	firstTokenWait time.Duration
-	tpMu           sync.Mutex
-	tps            map[time.Duration]*http.Transport
+	// 会话亲和 TTL（0=关闭）：同一会话粘住上次应答的渠道，保住上游提示词缓存。
+	affinityTTL time.Duration
+	aff         *affinityTable
+	tpMu        sync.Mutex
+	tps         map[time.Duration]*http.Transport
 
 	policyMu sync.Mutex
 	policy   modelPolicy
@@ -167,6 +170,7 @@ type Options struct {
 	FailoverBudget  time.Duration // 故障转移重试的时间预算
 	StreamHeartbeat time.Duration // 流式空闲心跳间隔
 	FirstTokenWait  time.Duration // 流式放行前等内容的最长时间（0=不限）
+	AffinityTTL     time.Duration // 会话亲和 TTL（0=关闭）
 }
 
 func NewGateway(db *gorm.DB, enc *crypto.Encryptor, rec *usage.Recorder, lim *ratelimit.Limiter, o Options) *Gateway {
@@ -183,6 +187,8 @@ func NewGateway(db *gorm.DB, enc *crypto.Encryptor, rec *usage.Recorder, lim *ra
 		failoverBudget:  o.FailoverBudget,
 		streamHeartbeat: o.StreamHeartbeat,
 		firstTokenWait:  o.FirstTokenWait,
+		affinityTTL:     o.AffinityTTL,
+		aff:             newAffinityTable(),
 		tps:             map[time.Duration]*http.Transport{},
 	}
 }

server/internal/proxy/passthrough.go

@@ -197,13 +197,28 @@ func (g *Gateway) doProxy(c *gin.Context, cands []channel.Candidate, clientProto
 		g.recordError(c, nil, nil, now(), "no_available_channel")
 		return
 	}
-	// 加权随机选择起始渠道
-	picked := g.ch.Pick(cands)
-	startIdx := 0
-	for i, cand := range cands {
-		if cand.Channel.ID == picked.ID {
-			startIdx = i
-			break
+	// 会话亲和：上次应答过该会话的渠道优先（候选集已过滤健康/冷却/绑定），
+	// 让上游提示词缓存被再次命中；命中不了（首次/冷却/下线）走加权随机。
+	affKey := g.affinityKeyFor(c)
+	startIdx := -1
+	if affKey != "" {
+		if chID, ok := g.aff.get(affKey); ok {
+			for i, cand := range cands {
+				if cand.Channel.ID == chID {
+					startIdx = i
+					break
+				}
+			}
+		}
+	}
+	if startIdx < 0 {
+		// 加权随机选择起始渠道
+		picked := g.ch.Pick(cands)
+		for i, cand := range cands {
+			if cand.Channel.ID == picked.ID {
+				startIdx = i
+				break
+			}
 		}
 	}
 
@@ -253,9 +268,12 @@ func (g *Gateway) doProxy(c *gin.Context, cands []channel.Candidate, clientProto
 		}
 		// 失败/成功计数由 proxyOne 就地报：只有它读得到上游响应头（429 的 Retry-After）
 		lastCh = ch
-		done, retry, st, b := g.proxyOne(c, ch, plan, stream, sink, sh, clientProto)
+		done, served, retry, st, b := g.proxyOne(c, ch, plan, stream, sink, sh, clientProto)
 		release()
 		if done {
+			if served && affKey != "" {
+				g.aff.put(affKey, ch.ID, g.affinityTTL)
+			}
 			return
 		}
 		if !retry {
@@ -283,6 +301,23 @@ func (g *Gateway) doProxy(c *gin.Context, cands []channel.Candidate, clientProto
 	g.recordError(c, lastCh, nil, now(), "all_channels_failed")
 }
 
+// affinityKeyFor 计算本请求的会话亲和键；亲和关闭或缺少身份要素时返回空。
+func (g *Gateway) affinityKeyFor(c *gin.Context) string {
+	if g.affinityTTL <= 0 {
+		return ""
+	}
+	kid, ok := c.Get(CtxKeyID)
+	kID, ok2 := kid.(uint64)
+	if !ok || !ok2 {
+		return ""
+	}
+	mn, _ := c.Get("model_name")
+	model, _ := mn.(string)
+	text, _ := c.Get("est_input_text")
+	firstText, _ := text.(string)
+	return affinityKey(kID, model, firstText)
+}
+
 // streamErrorEvent 把错误构造成客户端协议的 SSE 错误事件（响应头已写出时的唯一交付方式）。
 func streamErrorEvent(proto string, status int, msg string) []byte {
 	errObj := map[string]any{"message": msg, "type": "api_error", "code": "upstream_error", "status": status}
@@ -301,13 +336,14 @@ func streamErrorEvent(proto string, status int, msg string) []byte {
 
 // proxyOne 对单个渠道执行一次代理。
 // 返回：done=本次请求已终结（客户端响应已写出或已记账，停止故障转移）；
-// retry=是否可安全换渠道重试；status+respBody=失败信息。
+// served=该渠道成功应答（会话亲和据此绑定）；retry=是否可安全换渠道重试；
+// status+respBody=失败信息。
 // sh 为流式容错缓冲（非流式为 nil）；sink 为请求级记账 sink。
-func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan, stream bool, sink *usageSink, sh *streamHold, clientProto string) (done bool, retry bool, status int, respBody []byte) {
+func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan, stream bool, sink *usageSink, sh *streamHold, clientProto string) (done bool, served bool, retry bool, status int, respBody []byte) {
 	upKey, err := g.ch.UpstreamKey(ch)
 	if err != nil {
 		g.ch.ReportFailure(ch, channel.FailNet, 0)
-		return false, true, http.StatusInternalServerError, []byte("failed to decrypt channel key")
+		return false, false, true, http.StatusInternalServerError, []byte("failed to decrypt channel key")
 	}
 
 	upBody := plan.body
@@ -351,7 +387,7 @@ func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan
 
 	req, err := http.NewRequestWithContext(tryCtx, http.MethodPost, upstreamURL(ch, plan.proto, plan.path), bytes.NewReader(upBody))
 	if err != nil {
-		return false, false, http.StatusInternalServerError, []byte("failed to build upstream request")
+		return false, false, false, http.StatusInternalServerError, []byte("failed to build upstream request")
 	}
 	req.Header.Set("Content-Type", "application/json")
 	req.Header.Set("Authorization", "Bearer "+upKey)
@@ -380,10 +416,10 @@ func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan
 		// 客户端断开导致的取消：就地记账并终结本次请求（不计入渠道失败，也不换渠道）
 		if c.Request.Context().Err() != nil {
 			g.finishUsage(c, ch, start, store.UsageStatusCanceled, "client_disconnect")
-			return true, false, 0, nil
+			return true, false, false, 0, nil
 		}
 		g.ch.ReportFailure(ch, channel.FailNet, 0) // 网络错误/超时
-		return false, true, http.StatusBadGateway, []byte("upstream request failed: " + err.Error())
+		return false, false, true, http.StatusBadGateway, []byte("upstream request failed: " + err.Error())
 	}
 	defer resp.Body.Close()
 
@@ -394,9 +430,9 @@ func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan
 		kind, retryable := classifyUpstreamError(resp.StatusCode, errBody)
 		if retryable {
 			g.ch.ReportFailure(ch, kind, retryAfter(resp.Header))
-			return false, true, resp.StatusCode, errBody
+			return false, false, true, resp.StatusCode, errBody
 		}
-		return false, false, resp.StatusCode, errBody
+		return false, false, false, resp.StatusCode, errBody
 	}
 	g.ch.ReportSuccess(ch)
 
@@ -411,7 +447,7 @@ func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan
 		if stream {
 			asked, got = "流式", "整包响应"
 		}
-		return false, true, http.StatusBadGateway,
+		return false, false, true, http.StatusBadGateway,
 			[]byte(ch.Name + " answered a " + asked + " request with " + got + " (didNotStream)")
 	}
 
@@ -425,25 +461,28 @@ func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan
 // copyAndCapture 非流式：先读完整上游响应再写客户端——响应头与正文都在"确定成功"后
 // 才写出，读失败（网络错误/超时）发生在写出之前，可换渠道重试。
 // 此前实现先写 200 头再读，超时只能给客户端留一个 200 空响应。
-func (g *Gateway) copyAndCapture(c *gin.Context, ch *store.Channel, resp *http.Response, start time.Time, bodyConv func([]byte) ([]byte, error), sink *usageSink) (done bool, retry bool, status int, respBody []byte) {
+func (g *Gateway) copyAndCapture(c *gin.Context, ch *store.Channel, resp *http.Response, start time.Time, bodyConv func([]byte) ([]byte, error), sink *usageSink) (done bool, served bool, retry bool, status int, respBody []byte) {
 	data, err := io.ReadAll(resp.Body)
 	if err != nil {
 		switch {
 		case c.Request.Context().Err() != nil:
 			// 客户端断开：就地记账终结
 			g.finishUsage(c, ch, start, store.UsageStatusCanceled, "client_disconnect")
-			return true, false, 0, nil
+			return true, false, false, 0, nil
 		case errors.Is(err, context.DeadlineExceeded):
 			g.ch.ReportFailure(ch, channel.FailNet, 0)
-			return false, true, http.StatusBadGateway, []byte(ch.Name + " upstream timeout")
+			return false, false, true, http.StatusBadGateway, []byte(ch.Name + " upstream timeout")
 		default:
 			g.ch.ReportFailure(ch, channel.FailNet, 0)
-			return false, true, http.StatusBadGateway, []byte(ch.Name + " read error: " + err.Error())
+			return false, false, true, http.StatusBadGateway, []byte(ch.Name + " read error: " + err.Error())
 		}
 	}
 	if usageRaw := extractUsage(data); usageRaw != nil && sink != nil {
 		sink.push(usageRaw)
 	}
+	if sink != nil && sink.servedModel == "" {
+		sink.servedModel = extractServedModel(data)
+	}
 	out := data
 	if bodyConv != nil {
 		if converted, cerr := bodyConv(data); cerr == nil {
@@ -457,7 +496,7 @@ func (g *Gateway) copyAndCapture(c *gin.Context, ch *store.Channel, resp *http.R
 		c.Set("raw_response", string(data)) // 上游原始响应（未转换）
 	}
 	g.finishUsage(c, ch, start, store.UsageStatusSuccess, "")
-	return true, false, 0, nil
+	return true, true, false, 0, nil
 }
 
 // streamHoldCopy 流式转发：边读上游 SSE 边经 hold 写客户端（放行前缓冲）。
@@ -467,7 +506,7 @@ func (g *Gateway) copyAndCapture(c *gin.Context, ch *store.Channel, resp *http.R
 // 失败渠道的 message_start 用量不能计入最终账单。
 func (g *Gateway) streamHoldCopy(c *gin.Context, sh *streamHold, ch *store.Channel, r io.Reader, start time.Time,
 	lineConv func([]byte) []byte, clientProto string, trySink, reqSink *usageSink, onRelease func(), tryCtx context.Context,
-) (done bool, retry bool, status int, respBody []byte) {
+) (done bool, served bool, retry bool, status int, respBody []byte) {
 	// 原始响应捕获：仅管理员且开关开启（raw_request 已 set）时累积上游原始行
 	_, capture := c.Get("raw_request")
 	var rawResp strings.Builder
@@ -479,9 +518,9 @@ func (g *Gateway) streamHoldCopy(c *gin.Context, sh *streamHold, ch *store.Chann
 
 	scanner := newSSEScanner(r)
 	releaseSeen := false
-	fail := func(msg string, kind channel.FailKind) (bool, bool, int, []byte) {
+	fail := func(msg string, kind channel.FailKind) (bool, bool, bool, int, []byte) {
 		g.ch.ReportFailure(ch, kind, 0)
-		return false, true, http.StatusBadGateway, []byte(ch.Name + ": " + msg)
+		return false, false, true, http.StatusBadGateway, []byte(ch.Name + ": " + msg)
 	}
 	for {
 		line, err := scanner.Next()
@@ -504,7 +543,7 @@ func (g *Gateway) streamHoldCopy(c *gin.Context, sh *streamHold, ch *store.Chann
 					// 客户端意外断开：按已生成部分收费（canceled）
 					commitRaw()
 					g.finishUsage(c, ch, start, store.UsageStatusCanceled, "client_disconnect")
-					return true, false, 0, nil
+					return true, false, false, 0, nil
 				}
 				if trySink != nil {
 					trySink.outputText += sseContentText(out)
@@ -520,6 +559,11 @@ func (g *Gateway) streamHoldCopy(c *gin.Context, sh *streamHold, ch *store.Chann
 			if usageRaw := scanUsage(line); usageRaw != nil && trySink != nil {
 				trySink.push(usageRaw)
 			}
+			if trySink != nil && trySink.servedModel == "" && bytes.Contains(line, []byte(`"model"`)) {
+				if payload, ok := sseDataPayload(line); ok {
+					trySink.servedModel = extractServedModel([]byte(payload))
+				}
+			}
 		}
 		if err != nil {
 			if releaseSeen {
@@ -533,12 +577,12 @@ func (g *Gateway) streamHoldCopy(c *gin.Context, sh *streamHold, ch *store.Chann
 				} else {
 					g.recordError(c, ch, nil, start, "stream_read_error")
 				}
-				return true, false, 0, nil
+				return true, err == io.EOF, false, 0, nil
 			}
 			// 放行前失败：换渠道重试
 			if c.Request.Context().Err() != nil {
 				g.finishUsage(c, ch, start, store.UsageStatusCanceled, "client_disconnect")
-				return true, false, 0, nil
+				return true, false, false, 0, nil
 			}
 			if tryCtx != nil && tryCtx.Err() != nil {
 				return fail(fmt.Sprintf("first token timeout (no content in %s)", g.firstTokenWait), channel.FailNet)
@@ -563,6 +607,22 @@ type usageShape struct {
 	TotalTokens              int64 `json:"total_tokens"`
 	CacheReadInputTokens     int64 `json:"cache_read_input_tokens"`
 	CacheCreationInputTokens int64 `json:"cache_creation_input_tokens"`
+
+	// Reasoning：思考 token（chat 的 completion_tokens_details / responses 的
+	// output_tokens_details）。上游已把它计入 completion/output，只做展示口径，不重复计费。
+	CompletionDetails *usageTokenDetails `json:"completion_tokens_details,omitempty"`
+	OutputDetails     *usageTokenDetails `json:"output_tokens_details,omitempty"`
+	// Anthropic 的 cache_creation 细分：1h 写价是输入价 2×、5m 是 1.25×，必须拆开算钱
+	CacheCreation *usageCacheCreation `json:"cache_creation,omitempty"`
+}
+
+type usageTokenDetails struct {
+	ReasoningTokens int64 `json:"reasoning_tokens"`
+}
+
+type usageCacheCreation struct {
+	Ephemeral5m int64 `json:"ephemeral_5m_input_tokens"`
+	Ephemeral1h int64 `json:"ephemeral_1h_input_tokens"`
 }
 
 // extractUsage 从完整响应体提取 usage 子对象（chat / responses / messages）。
@@ -676,11 +736,45 @@ func (s *sseScanner) Next() ([]byte, error) {
 // ---------------------------------------------------------------------------
 // 记账
 
+// extractServedModel 从响应体（或流式事件载荷）提取上游实际应答的模型名：
+// chat/responses/messages 三协议都在顶层带 model；流式事件在 message.model /
+// response.model。模型映射、聚合商场景下它与请求的模型名不同，落库便于对账。
+func extractServedModel(data []byte) string {
+	var m map[string]json.RawMessage
+	if json.Unmarshal(data, &m) != nil {
+		return ""
+	}
+	if raw, ok := m["model"]; ok {
+		var s string
+		if json.Unmarshal(raw, &s) == nil && s != "" {
+			return s
+		}
+	}
+	for _, key := range []string{"response", "message"} {
+		raw, ok := m[key]
+		if !ok {
+			continue
+		}
+		var sub map[string]json.RawMessage
+		if json.Unmarshal(raw, &sub) != nil {
+			continue
+		}
+		if mraw, ok := sub["model"]; ok {
+			var s string
+			if json.Unmarshal(mraw, &s) == nil && s != "" {
+				return s
+			}
+		}
+	}
+	return ""
+}
+
 // usageSink 累积多次 usage：合并各事件字段（message_start 给 input，message_delta 给 output）。
 // outputText 累积已转发内容文本，用于流式中断时按 tiktoken 估算输出 token。
 type usageSink struct {
-	us         usageShape
-	outputText string
+	us          usageShape
+	outputText  string
+	servedModel string // 上游实际应答的模型名（模型映射/聚合商场景下与请求的不同）
 }
 
 func (u *usageSink) push(raw json.RawMessage) {
@@ -718,6 +812,15 @@ func (u *usageSink) push(raw json.RawMessage) {
 	if t.CacheCreationInputTokens > 0 {
 		u.us.CacheCreationInputTokens = t.CacheCreationInputTokens
 	}
+	if t.CompletionDetails != nil {
+		u.us.CompletionDetails = t.CompletionDetails
+	}
+	if t.OutputDetails != nil {
+		u.us.OutputDetails = t.OutputDetails
+	}
+	if t.CacheCreation != nil {
+		u.us.CacheCreation = t.CacheCreation
+	}
 }
 
 // Shape 返回合并后的用量。
@@ -748,6 +851,9 @@ func (u *usageSink) merge(o *usageSink) {
 		u.us.CacheCreationInputTokens = o.us.CacheCreationInputTokens
 	}
 	u.outputText += o.outputText
+	if o.servedModel != "" && u.servedModel == "" {
+		u.servedModel = o.servedModel
+	}
 }
 
 // finishUsage 落账：计算成本并异步写入。
@@ -769,9 +875,30 @@ func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time
 	out := us.CompletionTokens + us.OutputTokens
 	cacheRead := us.CacheReadInputTokens
 	cacheCreate := us.CacheCreationInputTokens
+	reasoning := int64(0)
+	if us.CompletionDetails != nil {
+		reasoning += us.CompletionDetails.ReasoningTokens
+	}
+	if us.OutputDetails != nil {
+		reasoning += us.OutputDetails.ReasoningTokens
+	}
+	// Anthropic 缓存写细分：1h 是输入价 2×、5m 是 1.25×，倍率 1.6。
+	// CacheWritePriceOr() 是 5m 口径（缺省输入价 ×1.25），1h 按同比例放大。
+	cache1h := int64(0)
+	if us.CacheCreation != nil && us.CacheCreation.Ephemeral1h > 0 {
+		cache1h = us.CacheCreation.Ephemeral1h
+		if cache1h > cacheCreate {
+			cache1h = cacheCreate
+		}
+	}
+	cache5m := cacheCreate - cache1h
 
 	modelName, _ := c.Get("model_name")
 	mn, _ := modelName.(string)
+	servedModel := ""
+	if sink != nil {
+		servedModel = sink.servedModel
+	}
 
 	// 上游未返回真实 usage 时估算（tiktoken）：
 	//   - 流式中断（canceled）：上游最终 usage 可能未返回
@@ -800,10 +927,12 @@ func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time
 		// 计价口径：in=非缓存输入、cacheRead=缓存读、cacheCreate=缓存写（Anthropic 语义，
 		// messages 流式 message_delta 的 input_tokens 即非缓存部分）。
 		// 缓存写价未单独配置时按输入价 1.25×（Anthropic 5m 口径）。
+		cwp := model.CacheWritePriceOr()
 		cost = float64(in)/1e6*model.InputPrice +
 			float64(out)/1e6*model.OutputPrice +
 			float64(cacheRead)/1e6*model.CacheReadPrice +
-			float64(cacheCreate)/1e6*model.CacheWritePriceOr()
+			float64(cache5m)/1e6*cwp +
+			float64(cache1h)/1e6*cwp*1.6
 	} else {
 		// 模型在记账时已被删除（请求进行中被下架）：无法定价，按 0 记账留痕。
 		// 虚构单价扣用户的钱既不可审计也无法对账。
@@ -871,6 +1000,9 @@ func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time
 		OutputPrice:         model.OutputPrice,
 		CacheReadPrice:      model.CacheReadPrice,
 		Cost:                cost,
+		ReasoningTokens:     reasoning,
+		CacheWrite1hTokens:  cache1h,
+		ServedModel:         servedModel,
 		LatencyMS:           latency,
 		Status:              status,
 		ErrorCode:           errCodePtr,
@@ -880,8 +1012,8 @@ func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time
 	})
 
 	// 请求级一行日志：usage_logs 落库是异步的且可能丢，排障时日志是唯一即时依据
-	log.Printf("proxy: trace=%s user=%d key=%d model=%q channel=%d proto=%s in=%d out=%d cache_r=%d cache_w=%d cost=%.6f status=%s http=%d latency=%dms err=%s",
-		traceStr, uidVal, kidVal, mn, chID, p, in, out, cacheRead, cacheCreate, cost, status,
+	log.Printf("proxy: trace=%s user=%d key=%d model=%q served=%q channel=%d proto=%s in=%d out=%d r=%d cache_r=%d cache_w=%d(1h=%d) cost=%.6f status=%s http=%d latency=%dms err=%s",
+		traceStr, uidVal, kidVal, mn, servedModel, chID, p, in, out, reasoning, cacheRead, cacheCreate, cache1h, cost, status,
 		c.Writer.Status(), latency, errCode)
 }
 

server/internal/proxy/usage_refine_test.go

@@ -0,0 +1,114 @@
+package proxy
+
+import (
+	"math"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/openteam/server/internal/store"
+)
+
+// runPath 走完整链路（Auth → Handle），路径与请求体由调用方给。
+func runPath(t *testing.T, h *harness, path, body string) *httptest.ResponseRecorder {
+	t.Helper()
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	req := httptest.NewRequest(http.MethodPost, path, strings.NewReader(body))
+	req.Header.Set("Authorization", "Bearer "+h.key)
+	c.Request = req
+	h.gw.Auth(c)
+	if !c.IsAborted() {
+		h.gw.Handle(c)
+	}
+	return w
+}
+
+// usage 细化：reasoning token 与上游实际应答模型落库；计费口径不变
+// （reasoning 已含在 completion_tokens 内，不重复计费）。
+func TestUsageReasoningAndServedModelRecorded(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"model":"gpt-x-actual","choices":[{"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}],"usage":{"prompt_tokens":10,"completion_tokens":5,"total_tokens":15,"completion_tokens_details":{"reasoning_tokens":3}}}`))
+	}))
+	defer srv.Close()
+
+	h := newHarness(t, srv.URL, 0, true, nil)
+	w := h.run(t, nil, `{"model":"m1","messages":[{"role":"user","content":"hi"}]}`)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code=%d body=%s", w.Code, w.Body.String())
+	}
+	h.rec.Close()
+	var logs []store.UsageLog
+	h.db.Find(&logs)
+	if len(logs) != 1 {
+		t.Fatalf("应有一条记账，实际 %d", len(logs))
+	}
+	l := logs[0]
+	if l.ServedModel != "gpt-x-actual" {
+		t.Fatalf("served_model 应记录上游应答的模型名，实际 %q", l.ServedModel)
+	}
+	if l.ReasoningTokens != 3 {
+		t.Fatalf("reasoning_tokens 应为 3，实际 %d", l.ReasoningTokens)
+	}
+	// 计费不含 reasoning 加成：10×0.15/1e6 + 5×0.6/1e6 = 4.5e-6
+	if math.Abs(l.Cost-4.5e-6) > 1e-12 {
+		t.Fatalf("cost 应为 4.5e-6（reasoning 不重复计费），实际 %g", l.Cost)
+	}
+}
+
+// Anthropic 缓存写细分：1h 与 5m 分开计价（1h = 2× 输入价，5m = 1.25×）。
+func TestAnthropicCache1hBillingSplit(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","model":"claude-x","role":"assistant","content":[{"type":"text","text":"hi"}],"stop_reason":"end_turn","usage":{"input_tokens":100,"output_tokens":10,"cache_read_input_tokens":200,"cache_creation_input_tokens":300,"cache_creation":{"ephemeral_5m_input_tokens":100,"ephemeral_1h_input_tokens":200}}}`))
+	}))
+	defer srv.Close()
+
+	h := newHarness(t, srv.URL, 0, true, nil)
+	// 渠道改为 Anthropic 形态，模型定价：输入 1、输出 2、缓存读 0.1、缓存写未配置（1.25× 输入）
+	var ch1 store.Channel
+	if err := h.db.Where("name = ?", "c1").First(&ch1).Error; err != nil {
+		t.Fatal(err)
+	}
+	ch1.Formats = []string{store.FormatMessages}
+	if err := h.db.Save(&ch1).Error; err != nil {
+		t.Fatal(err)
+	}
+	if err := h.db.Model(&store.Model{}).Where("name = ?", "m1").
+		Updates(map[string]any{"input_price": 1.0, "output_price": 2.0, "cache_read_price": 0.1}).Error; err != nil {
+		t.Fatal(err)
+	}
+
+	w := runPath(t, h, "/v1/messages", `{"model":"m1","max_tokens":16,"messages":[{"role":"user","content":"hi"}]}`)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code=%d body=%s", w.Code, w.Body.String())
+	}
+	h.rec.Close()
+	var logs []store.UsageLog
+	h.db.Find(&logs)
+	if len(logs) != 1 {
+		t.Fatalf("应有一条记账，实际 %d", len(logs))
+	}
+	l := logs[0]
+	if l.CacheWrite1hTokens != 200 {
+		t.Fatalf("cache_write_1h_tokens 应为 200，实际 %d", l.CacheWrite1hTokens)
+	}
+	if l.ServedModel != "claude-x" {
+		t.Fatalf("served_model 应为 claude-x，实际 %q", l.ServedModel)
+	}
+	// 100/1e6*1 + 10/1e6*2 + 200/1e6*0.1 + 100/1e6*1.25 + 200/1e6*(1.25*1.6) = 6.65e-4
+	want := 100.0/1e6*1.0 + 10.0/1e6*2.0 + 200.0/1e6*0.1 + 100.0/1e6*1.25 + 200.0/1e6*2.0
+	if math.Abs(l.Cost-want) > 1e-9 {
+		t.Fatalf("cost 应为 %g（1h/5m 分开计价），实际 %g", want, l.Cost)
+	}
+	// 落库 input 为含缓存的总量口径
+	if l.InputTokens != 600 || l.OutputTokens != 10 {
+		t.Fatalf("in=%d out=%d，期望 600/10", l.InputTokens, l.OutputTokens)
+	}
+	if !strings.Contains(w.Body.String(), `"text":"hi"`) {
+		t.Fatalf("响应应原样透传：%s", w.Body.String())
+	}
+}

server/internal/store/db.go

@@ -25,7 +25,8 @@ import (
 // v5：usage_logs 补 key_id/channel_id 索引与 (user,created)/(channel,created) 复合索引。
 // v6：修正 v5 复合索引列序——AutoMigrate 按名跳过已存在的索引，旧库上的
 // idx_user_created 是 user_id 单列，光改 struct tag 不会重建，需删掉重迁一次。
-const schemaVersion = 6
+// v7：usage_logs 加 reasoning_tokens / cache_write_1h_tokens / served_model 三列。
+const schemaVersion = 7
 
 // Open 打开数据库连接并自动迁移。
 // 开发默认 SQLite（自动追加 WAL 与 busy_timeout，见 withSQLitePragmas），生产可切 postgres。

server/internal/store/models.go

@@ -196,6 +196,9 @@ type UsageLog struct {
 	OutputTokens        int64   `json:"output_tokens"`
 	CacheReadTokens     int64   `json:"cache_read_tokens"`
 	CacheCreationTokens int64   `json:"cache_creation_tokens"`
+	ReasoningTokens     int64   `json:"reasoning_tokens"`                           // 思考 token（上游已计入 output，展示口径不重复计费）
+	CacheWrite1hTokens  int64   `json:"cache_write_1h_tokens"`                      // Anthropic 1h 缓存写 token（计价 2× 输入）
+	ServedModel         string  `gorm:"size:128" json:"served_model"`               // 上游实际应答的模型名（映射/聚合商场景与请求名不同）
 	InputPrice          float64 `gorm:"type:numeric(20,8)" json:"input_price"`      // 快照
 	OutputPrice         float64 `gorm:"type:numeric(20,8)" json:"output_price"`     // 快照
 	CacheReadPrice      float64 `gorm:"type:numeric(20,8)" json:"cache_read_price"` // 快照