M4a：神识全量网关能力——OpenAI 兼容适配器（元炁共用）、候选路由+凭证/绑定级冷却+首字超时切道+粘性会话、用量计量与 /me/usage，故障注入测试四场景

shenshi/internal/metering/usage.go

@@ -0,0 +1,88 @@
+// Package metering 用量计量：每次尝试一条记录（stdout 结构化日志 + 进程内聚合）。
+// 生产部署接 PG（shenshi-gateway.md §7）；M4a 先以日志与内存聚合支撑 /me/usage 与对账抽样。
+package metering
+
+import (
+	"encoding/json"
+	"log"
+	"sync"
+	"time"
+)
+
+// Attempt 一次端点尝试的计量记录。
+type Attempt struct {
+	RunID            string    `json:"run_id"`
+	CharacterID      string    `json:"character_id,omitempty"`
+	Model            string    `json:"model"`
+	Slot             string    `json:"slot"`
+	Endpoint         string    `json:"endpoint"`
+	Status           string    `json:"status"` // ok / error / first_byte_timeout
+	ErrorCode        string    `json:"error_code,omitempty"`
+	FirstTokenMS     int64     `json:"first_token_ms"`
+	TotalMS          int64     `json:"total_ms"`
+	PromptTokens     int       `json:"prompt_tokens"`
+	CompletionTokens int       `json:"completion_tokens"`
+	At               time.Time `json:"at"`
+}
+
+// UsageAgg /me/usage 聚合行。
+type UsageAgg struct {
+	Model            string `json:"model"`
+	Endpoint         string `json:"endpoint"`
+	Requests         int    `json:"requests"`
+	OK               int    `json:"ok"`
+	PromptTokens     int    `json:"prompt_tokens"`
+	CompletionTokens int    `json:"completion_tokens"`
+}
+
+// Log 进程内计量器（有界环形，避免无界增长）。
+type Log struct {
+	mu      sync.Mutex
+	entries []Attempt
+	limit   int
+}
+
+// New 构造计量器（保留最近 N 条）。
+func New(limit int) *Log {
+	return &Log{entries: make([]Attempt, 0, limit), limit: limit}
+}
+
+// Record 记录一次尝试：stdout JSON 日志 + 内存追加。
+func (l *Log) Record(a Attempt) {
+	a.At = time.Now().UTC()
+	if raw, err := json.Marshal(a); err == nil {
+		log.Printf(`{"kind":"shenshi_usage",%s`, string(raw[1:]))
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	if len(l.entries) >= l.limit {
+		l.entries = l.entries[1:]
+	}
+	l.entries = append(l.entries, a)
+}
+
+// Aggregate 按 模型×端点 聚合。
+func (l *Log) Aggregate() []UsageAgg {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	agg := map[string]*UsageAgg{}
+	for _, e := range l.entries {
+		key := e.Model + "|" + e.Endpoint
+		row := agg[key]
+		if row == nil {
+			row = &UsageAgg{Model: e.Model, Endpoint: e.Endpoint}
+			agg[key] = row
+		}
+		row.Requests++
+		if e.Status == "ok" {
+			row.OK++
+			row.PromptTokens += e.PromptTokens
+			row.CompletionTokens += e.CompletionTokens
+		}
+	}
+	out := make([]UsageAgg, 0, len(agg))
+	for _, row := range agg {
+		out = append(out, *row)
+	}
+	return out
+}

shenshi/internal/route/failover.go

@@ -0,0 +1,75 @@
+// 候选路由与冷却：凭证级（401/403）与绑定级（429/5xx/超时）冷却，粘性会话与故障迭代。
+package route
+
+import (
+	"sync"
+	"time"
+)
+
+// 冷却时长：凭证错误长冷却，限流/超时短冷却（避免整渠道误杀，shenshi-gateway.md §4）。
+const (
+	credentialCooldown = 10 * time.Minute
+	bindingCooldown    = 30 * time.Second
+)
+
+// Cooldowns 进程内冷却状态：key 形如 "cred:<endpoint>:<keyhash>" / "bind:<endpoint>:<model>"。
+type Cooldowns struct {
+	mu    sync.Mutex
+	until map[string]time.Time
+}
+
+// NewCooldowns 构造冷却表。
+func NewCooldowns() *Cooldowns {
+	return &Cooldowns{until: make(map[string]time.Time)}
+}
+
+// Mark 凭证级（credential=true）或绑定级冷却。
+func (c *Cooldowns) Mark(key string, credential bool) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	duration := bindingCooldown
+	if credential {
+		duration = credentialCooldown
+	}
+	c.until[key] = time.Now().Add(duration)
+}
+
+// Active 该冷却键当前是否生效。
+func (c *Cooldowns) Active(key string) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	t, ok := c.until[key]
+	return ok && time.Now().Before(t)
+}
+
+// Sticky 粘性会话：同一 run 连续请求优先同端点（进程内，run 生命周期短）。
+type Sticky struct {
+	mu    sync.Mutex
+	binds map[string]string
+}
+
+// NewSticky 构造粘性表。
+func NewSticky() *Sticky {
+	return &Sticky{binds: make(map[string]string)}
+}
+
+// Remember 记录 run → 端点绑定。
+func (s *Sticky) Remember(runID, endpointName string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.binds[runID] = endpointName
+}
+
+// Recall 取 run 的粘性端点（无则空串）。
+func (s *Sticky) Recall(runID string) string {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.binds[runID]
+}
+
+// Forget run 结束后清理。
+func (s *Sticky) Forget(runID string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	delete(s.binds, runID)
+}

shenshi/internal/route/registry.go

@@ -1,46 +1,150 @@
 // Package route 维护逻辑模型/能力槽位到候选端点的路由注册表。
-// M0 为单一 stub 端点的静态表；候选迭代、健康冷却与故障切换按
-// docs/architecture/shenshi-gateway.md §4 于 M4 落地。
+// 端点来源：环境 SHENSHI_ENDPOINTS（JSON 数组）；未配置时回落 stub（骨架形态）。
+// 冷却粒度（shenshi-gateway.md §4）：401/403 只冷却凭证；429/5xx/超时只冷却该端点绑定。
 package route
 
 import (
+	"encoding/json"
 	"fmt"
+	"os"
+	"sort"
+	"time"
 
 	"hunjing/shenshi/internal/transform"
 )
 
-// Endpoint 一个可路由的出站端点：类型 + 适配器实例。
+// Endpoint 一个可路由的出站端点。
 type Endpoint struct {
-	Type     string
+	Name     string
+	Type     string // openai_compatible / local / stub（契约词典 §14.2）
 	Adaptor  transform.Adaptor
 	Priority int
+	// 凭证指纹（凭证级冷却键）；stub/openai 兼容端点各自填充
+	CredentialKey string
+	// 绑定级冷却键的模型名部分
+	UpstreamModel string
+}
+
+// EndpointSpec 环境配置形态。
+type EndpointSpec struct {
+	Type     string `json:"type"`
+	Name     string `json:"name"`
+	BaseURL  string `json:"base_url"`
+	APIKey   string `json:"api_key"`
+	Model    string `json:"model"`
+	Priority int    `json:"priority"`
 }
 
 // Registry 槽位 → 候选端点列表（按 Priority 升序尝试）。
 type Registry struct {
 	bySlot map[string][]Endpoint
+	// Cooldowns/FirstByteTimeoutMS 由 server 层共享使用
+	Cooldowns        *Cooldowns
+	Sticky           *Sticky
+	FirstByteTimeout time.Duration
 }
 
-// New 构建注册表并登记默认端点。
+// New 构建注册表：读 SHENSHI_ENDPOINTS（JSON），无配置时登记 stub。
 func New() *Registry {
-	r := &Registry{bySlot: make(map[string][]Endpoint)}
-	r.Register("chat", Endpoint{Type: "stub", Adaptor: transform.StubAdaptor{}, Priority: 100})
+	r := &Registry{
+		bySlot:           make(map[string][]Endpoint),
+		Cooldowns:        NewCooldowns(),
+		Sticky:           NewSticky(),
+		FirstByteTimeout: 10 * time.Second,
+	}
+	specs := loadSpecs()
+	if len(specs) == 0 {
+		r.Register("chat", Endpoint{Type: "stub", Adaptor: transform.StubAdaptor{}, Priority: 100, UpstreamModel: "stub"})
+		return r
+	}
+	for _, spec := range specs {
+		if spec.Type != "openai_compatible" && spec.Type != "local" {
+			continue
+		}
+		adaptor := &transform.OpenAICompatAdaptor{
+			EndpointName:  spec.Name,
+			BaseURL:       spec.BaseURL,
+			APIKey:        spec.APIKey,
+			UpstreamModel: spec.Model,
+		}
+		cred := spec.Name
+		if spec.APIKey != "" {
+			cred = spec.Name + ":" + fmt.Sprintf("%x", len(spec.APIKey)) // 不落密钥原文，仅指纹
+		}
+		r.Register("chat", Endpoint{
+			Name: spec.Name, Type: spec.Type, Adaptor: adaptor,
+			Priority: spec.Priority, CredentialKey: cred, UpstreamModel: spec.Model,
+		})
+	}
 	return r
 }
 
-// Register 登记端点到指定能力槽位（契约词典 §14.1 model_slot 取值）。
+func loadSpecs() []EndpointSpec {
+	raw := os.Getenv("SHENSHI_ENDPOINTS")
+	if raw == "" {
+		return nil
+	}
+	var specs []EndpointSpec
+	if err := json.Unmarshal([]byte(raw), &specs); err != nil {
+		return nil
+	}
+	return specs
+}
+
+// Reset 清空槽位（测试装配用）。
+func (r *Registry) Reset(slot string) { delete(r.bySlot, slot) }
+
+// Register 登记端点到指定能力槽位（契约词典 §14.1）。
 func (r *Registry) Register(slot string, ep Endpoint) {
 	r.bySlot[slot] = append(r.bySlot[slot], ep)
 }
 
-// Resolve 返回槽位的候选端点；无候选即错误（HJ-2001 语义）。
-func (r *Registry) Resolve(slot string) ([]Endpoint, error) {
+// Candidates 返回槽位候选（Priority 升序；粘性端点置首；冷却中端点排后不剔除——
+// 全部冷却时仍可尝试最优先者，避免死锁）。
+func (r *Registry) Candidates(slot, runID string) ([]Endpoint, error) {
 	if slot == "" {
 		slot = "chat"
 	}
-	eps := r.bySlot[slot]
+	eps := append([]Endpoint(nil), r.bySlot[slot]...)
 	if len(eps) == 0 {
 		return nil, fmt.Errorf("槽位 %s 无可用端点", slot)
 	}
+	sort.SliceStable(eps, func(i, j int) bool { return eps[i].Priority < eps[j].Priority })
+	if runID != "" {
+		if stickyName := r.Sticky.Recall(runID); stickyName != "" {
+			for i, ep := range eps {
+				if ep.Name == stickyName {
+					eps = append([]Endpoint{ep}, append(eps[:i:i], eps[i+1:]...)...)
+					break
+				}
+			}
+		}
+	}
+	// 冷却中的端点移到队尾（保留兜底能力）
+	cooling := 0
+	for i := 0; i < len(eps)-cooling; {
+		if r.Cooldowns.Active(bindingKey(eps[i])) || r.Cooldowns.Active(credKey(eps[i])) {
+			ep := eps[i]
+			eps = append(append(eps[:i:i], eps[i+1:]...), ep)
+			cooling++
+			continue
+		}
+		i++
+	}
 	return eps, nil
 }
+
+func bindingKey(ep Endpoint) string {
+	return "bind:" + ep.Name + ":" + ep.UpstreamModel
+}
+
+func credKey(ep Endpoint) string {
+	if ep.CredentialKey == "" {
+		return "cred:" + ep.Name
+	}
+	return "cred:" + ep.CredentialKey
+}
+
+// BindingCooldownKey / CredentialCooldownKey 供 server 层标记冷却。
+func BindingCooldownKey(ep Endpoint) string    { return bindingKey(ep) }
+func CredentialCooldownKey(ep Endpoint) string { return credKey(ep) }

shenshi/internal/server/failover_test.go

@@ -0,0 +1,161 @@
+package server
+
+import (
+	"bufio"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"hunjing/shenshi/internal/route"
+	"hunjing/shenshi/internal/transform"
+)
+
+// fakeOpenAI 构造一个 OpenAI 兼容 SSE 假上游：可控延迟/状态码/内容。
+func fakeOpenAI(t *testing.T, delay time.Duration, status int, content string) *httptest.Server {
+	t.Helper()
+	return httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		if status != 0 {
+			w.WriteHeader(status)
+			_, _ = w.Write([]byte(`{"error":{"message":"boom"}}`))
+			return
+		}
+		if delay > 0 {
+			time.Sleep(delay)
+		}
+		w.Header().Set("content-type", "text/event-stream")
+		flusher := w.(http.Flusher)
+		_, _ = w.Write([]byte("data: {\"choices\":[{\"delta\":{\"content\":\"" + content + "\"}}]}\n\n"))
+		flusher.Flush()
+		_, _ = w.Write([]byte("data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":3,\"completion_tokens\":4}}\n\n"))
+		flusher.Flush()
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+	}))
+}
+
+func registryWith(t *testing.T, firstByte time.Duration, eps ...route.Endpoint) *route.Registry {
+	t.Helper()
+	r := route.New()
+	r.FirstByteTimeout = firstByte
+	// 清掉默认 stub，仅保留测试端点
+	r.Reset("chat")
+	for _, ep := range eps {
+		r.Register("chat", ep)
+	}
+	return r
+}
+
+func oaEndpoint(name, baseURL string, priority int) route.Endpoint {
+	return route.Endpoint{
+		Name: name, Type: "openai_compatible", Priority: priority,
+		Adaptor: &transform.OpenAICompatAdaptor{EndpointName: name, BaseURL: baseURL, UpstreamModel: "test-model"},
+	}
+}
+
+func foPostChat(ts *httptest.Server, body string) (*http.Response, string) {
+	resp, err := ts.Client().Post(ts.URL+"/v1/chat/completions", "application/json", strings.NewReader(body))
+	if err != nil {
+		return nil, ""
+	}
+	var sb strings.Builder
+	scanner := bufio.NewScanner(resp.Body)
+	for scanner.Scan() {
+		sb.WriteString(scanner.Text())
+		sb.WriteString("\n")
+	}
+	return resp, sb.String()
+}
+
+func foChatBody(run string) string {
+	return `{"ir_version":1,"slot":"chat","messages":[{"role":"user","content":"你好"}],"run_id":"` + run + `","stream":true}`
+}
+
+// 场景：首端点首字超时 → 切道到次端点成功（不向客户端暴露失败）。
+func TestFailoverOnFirstByteTimeout(t *testing.T) {
+	slow := fakeOpenAI(t, 500*time.Millisecond, 0, "慢端点")
+	fast := fakeOpenAI(t, 0, 0, "快端点")
+	defer slow.Close()
+	defer fast.Close()
+
+	srv := New(Config{}, registryWith(t, 50*time.Millisecond, oaEndpoint("slow", slow.URL, 10), oaEndpoint("fast", fast.URL, 20)))
+	ts := httptest.NewServer(srv)
+	defer ts.Close()
+
+	resp, body := foPostChat(ts, foChatBody("run_fo"))
+	if resp == nil || resp.StatusCode != http.StatusOK {
+		t.Fatalf("状态码异常：%v", resp)
+	}
+	if !strings.Contains(body, "快端点") || strings.Contains(body, "慢端点") {
+		t.Fatalf("应切道到快端点：%s", body)
+	}
+	if !strings.Contains(body, `"finish_reason":"stop"`) || !strings.Contains(body, "data: [DONE]") {
+		t.Fatalf("缺少终帧/DONE：%s", body)
+	}
+}
+
+// 场景：首端点 401 → 凭证级冷却；下一请求直接跳过它走次端点。
+func TestCredentialCooldownSkipsEndpoint(t *testing.T) {
+	bad := fakeOpenAI(t, 0, http.StatusUnauthorized, "")
+	good := fakeOpenAI(t, 0, 0, "好端点")
+	defer bad.Close()
+	defer good.Close()
+
+	srv := New(Config{}, registryWith(t, time.Second, oaEndpoint("bad", bad.URL, 10), oaEndpoint("good", good.URL, 20)))
+	ts := httptest.NewServer(srv)
+	defer ts.Close()
+
+	// 第一次：bad 401 → 冷却 → 切 good
+	resp, body := foPostChat(ts, foChatBody("run_a"))
+	if resp == nil || resp.StatusCode != http.StatusOK || !strings.Contains(body, "好端点") {
+		t.Fatalf("首次应切到好端点：%v %s", resp, body)
+	}
+	// 第二次：bad 仍在冷却，直接命中 good（若未跳过会再花一次 401 往返——这里断言成功即可）
+	resp2, body2 := foPostChat(ts, foChatBody("run_b"))
+	if resp2 == nil || resp2.StatusCode != http.StatusOK || !strings.Contains(body2, "好端点") {
+		t.Fatalf("冷却后应直达好端点：%v", resp2)
+	}
+}
+
+// 场景：全部候选失败且未写字节 → HJ-2001 JSON（不是半截 SSE）。
+func TestAllCandidatesFailedReturnsJsonError(t *testing.T) {
+	bad := fakeOpenAI(t, 0, http.StatusBadGateway, "")
+	defer bad.Close()
+	srv := New(Config{}, registryWith(t, time.Second, oaEndpoint("bad", bad.URL, 10)))
+	ts := httptest.NewServer(srv)
+	defer ts.Close()
+
+	resp, body := foPostChat(ts, foChatBody("run_x"))
+	if resp.StatusCode != http.StatusServiceUnavailable {
+		t.Fatalf("状态码 = %d", resp.StatusCode)
+	}
+	if !strings.Contains(body, "HJ-2001") || strings.Contains(body, "text/event-stream") {
+		t.Fatalf("应为 JSON 错误而非 SSE：%s", body)
+	}
+}
+
+// 场景：用量聚合端点返回 ok 计数与 token。
+func TestUsageAggregate(t *testing.T) {
+	good := fakeOpenAI(t, 0, 0, "内容")
+	defer good.Close()
+	srv := New(Config{}, registryWith(t, time.Second, oaEndpoint("good", good.URL, 10)))
+	ts := httptest.NewServer(srv)
+	defer ts.Close()
+
+	if resp, _ := foPostChat(ts, foChatBody("run_u")); resp == nil || resp.StatusCode != 200 {
+		t.Fatal("请求失败")
+	}
+	usageResp, err := ts.Client().Get(ts.URL + "/me/usage")
+	if err != nil {
+		t.Fatal(err)
+	}
+	var agg []map[string]any
+	defer func() { _ = usageResp.Body.Close() }()
+	if err := json.NewDecoder(usageResp.Body).Decode(&agg); err != nil {
+		t.Fatal(err)
+	}
+	if len(agg) != 1 || agg[0]["ok"].(float64) != 1 || agg[0]["prompt_tokens"].(float64) != 3 {
+		t.Fatalf("聚合不符：%v", agg)
+	}
+}

shenshi/internal/server/server.go

@@ -1,33 +1,42 @@
-// Package server 神识 HTTP 服务：数据面（IR 入站 → SSE 出站）与健康面。
+// Package server 神识 HTTP 服务：数据面（IR 入站 → SSE 出站）、健康面与用量面。
+//
+// 故障切换红线：首帧到达前不提交 SSE 响应——失败可整体换道并返回 JSON 错误；
+// 已写出字节后不再换道，如实以 error 终帧收尾（shenshi-gateway.md §4/§5）。
 package server
 
 import (
+	"context"
 	"encoding/json"
 	"net/http"
+	"time"
 
 	"hunjing/shenshi/internal/ir"
+	"hunjing/shenshi/internal/metering"
 	"hunjing/shenshi/internal/relay"
 	"hunjing/shenshi/internal/route"
+	"hunjing/shenshi/internal/transform"
 )
 
 // Config 服务配置。
 type Config struct {
-	ServiceToken string // 灵台→神识服务令牌；为空时开发档不校验
+	ServiceToken string
 }
 
 // Server 神识 HTTP 服务。
 type Server struct {
 	cfg    Config
 	routes *route.Registry
+	usage  *metering.Log
 	mux    *http.ServeMux
 }
 
 // New 构建服务并注册路由。
 func New(cfg Config, routes *route.Registry) *Server {
-	s := &Server{cfg: cfg, routes: routes, mux: http.NewServeMux()}
+	s := &Server{cfg: cfg, routes: routes, usage: metering.New(10_000), mux: http.NewServeMux()}
 	s.mux.HandleFunc("GET /healthz", s.handleHealth)
 	s.mux.HandleFunc("GET /readyz", s.handleHealth)
 	s.mux.Handle("POST /v1/chat/completions", s.auth(http.HandlerFunc(s.handleChat)))
+	s.mux.Handle("GET /me/usage", s.auth(http.HandlerFunc(s.handleUsage)))
 	return s
 }
 
@@ -41,8 +50,7 @@ func (s *Server) auth(next http.Handler) http.Handler {
 			next.ServeHTTP(w, r)
 			return
 		}
-		credential := r.Header.Get("Authorization")
-		if credential != "Bearer "+s.cfg.ServiceToken {
+		if r.Header.Get("Authorization") != "Bearer "+s.cfg.ServiceToken {
 			writeError(w, http.StatusUnauthorized, "HJ-2003", "服务令牌无效")
 			return
 		}
@@ -50,7 +58,7 @@ func (s *Server) auth(next http.Handler) http.Handler {
 	})
 }
 
-// handleChat 数据面：解析 IR → 路由 → 适配器流出 SSE。
+// handleChat 数据面：IR → 候选迭代（首字超时/冷却标记/粘性）→ SSE。
 func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
 	var req ir.ChatRequest
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
@@ -61,35 +69,136 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusBadRequest, "HJ-2002", err.Error())
 		return
 	}
-	endpoints, err := s.routes.Resolve(req.Slot)
+	candidates, err := s.routes.Candidates(req.Slot, req.RunID)
 	if err != nil {
 		writeError(w, http.StatusServiceUnavailable, "HJ-2001", err.Error())
 		return
 	}
 
-	relay.SSEHeaders(w)
-	w.WriteHeader(http.StatusOK)
+	var lastErr error
+	for _, endpoint := range candidates {
+		res := s.tryEndpoint(w, r, endpoint, &req)
+		s.usage.Record(res.Attempt)
+		if res.Committed {
+			// 已向客户端写出：无论成败都不能再换道
+			if res.Status == "ok" {
+				s.routes.Sticky.Remember(req.RunID, endpoint.Name)
+			}
+			return
+		}
+		// 未提交：分类冷却后换下一候选
+		if res.Upstream != nil {
+			if res.Upstream.StatusCode == 401 || res.Upstream.StatusCode == 403 {
+				s.routes.Cooldowns.Mark(route.CredentialCooldownKey(endpoint), true)
+			} else {
+				s.routes.Cooldowns.Mark(route.BindingCooldownKey(endpoint), false)
+			}
+		} else {
+			s.routes.Cooldowns.Mark(route.BindingCooldownKey(endpoint), false)
+		}
+		lastErr = res.Err
+	}
+	writeError(w, http.StatusServiceUnavailable, "HJ-2001", "全部候选端点失败："+errString(lastErr))
+}
+
+type attempt struct {
+	metering.Attempt
+	Committed bool                     // 是否已向客户端写出 SSE
+	Upstream  *transform.UpstreamError // 上游分类（可空）
+	Err       error                    // 原始错误
+}
+
+func errString(err error) string {
+	if err == nil {
+		return "未知错误"
+	}
+	return err.Error()
+}
+
+// tryEndpoint 单端点尝试。
+func (s *Server) tryEndpoint(w http.ResponseWriter, r *http.Request, endpoint route.Endpoint, req *ir.ChatRequest) attempt {
+	start := time.Now()
+	ctx, cancel := context.WithTimeout(r.Context(), s.routes.FirstByteTimeout)
+	defer cancel()
 
 	deltas := make(chan ir.Delta, 16)
-	streamErr := make(chan error, 1)
+	errCh := make(chan error, 1)
 	go func() {
 		defer close(deltas)
-		streamErr <- endpoints[0].Adaptor.Stream(r.Context(), &req, deltas)
+		errCh <- endpoint.Adaptor.Stream(ctx, req, deltas)
 	}()
+
+	res := attempt{Attempt: metering.Attempt{
+		RunID: req.RunID, CharacterID: req.CharacterID,
+		Model: req.Model, Slot: req.Slot, Endpoint: endpoint.Name,
+	}}
+
+	// 阶段一：等首帧或失败——此阶段未提交响应，可整体回退
+	var first ir.Delta
+	select {
+	case d, ok := <-deltas:
+		if !ok {
+			res.Err = <-errCh
+			res.Upstream, _ = res.Err.(*transform.UpstreamError)
+			res.Status, res.TotalMS = failStatus(res.Err), time.Since(start).Milliseconds()
+			return res
+		}
+		first = d
+	case <-ctx.Done():
+		res.Status, res.Err = "first_byte_timeout", ctx.Err()
+		res.TotalMS = time.Since(start).Milliseconds()
+		return res
+	}
+
+	// 阶段二：首帧已到——提交 SSE，流式转发
+	relay.SSEHeaders(w)
+	w.WriteHeader(http.StatusOK)
+	res.Committed = true
+	res.FirstTokenMS = time.Since(start).Milliseconds()
+	_ = relay.WriteDelta(w, first)
+
 	for d := range deltas {
-		if err := relay.WriteDelta(w, d); err != nil {
-			return // 客户端断开：取消上游由 r.Context() 传播
+		if d.Delta != "" || d.FinishReason != "" {
+			if err := relay.WriteDelta(w, d); err != nil {
+				break // 客户端断开
+			}
+		}
+		if d.FinishReason != "" {
+			res.Attempt.PromptTokens = d.UsagePromptTokens
+			res.Attempt.CompletionTokens = d.UsageCompletionTokens
 		}
 	}
-	if err := <-streamErr; err != nil {
-		// 流已开始，只能如实中断：写一条错误帧后终止（不伪装 [DONE]）
-		_ = relay.WriteDelta(w, ir.Delta{FinishReason: "error", Delta: ""})
+	err := <-errCh
+	res.TotalMS = time.Since(start).Milliseconds()
+	if err == nil {
 		_ = relay.WriteDone(w)
-		return
+		res.Status = "ok"
+		return res
 	}
-	if err := relay.WriteDone(w); err != nil {
-		return
+	// 流中失败：如实以 error 终帧收尾，不伪装完成
+	res.Err = err
+	res.Upstream, _ = err.(*transform.UpstreamError)
+	res.Status = "error"
+	res.ErrorCode = "stream_written"
+	_ = relay.WriteDelta(w, ir.Delta{FinishReason: "error"})
+	_ = relay.WriteDone(w)
+	return res
+}
+
+func failStatus(err error) string {
+	if up, ok := err.(*transform.UpstreamError); ok {
+		if up.StatusCode == 0 {
+			return "timeout_or_network"
+		}
+		return "upstream_error"
 	}
+	return "error"
+}
+
+// handleUsage 用量聚合（/me/usage；BYOK 归属维度随 M4b 接入）。
+func (s *Server) handleUsage(w http.ResponseWriter, _ *http.Request) {
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(s.usage.Aggregate())
 }
 
 // handleHealth 健康面。

shenshi/internal/transform/openai.go

@@ -0,0 +1,138 @@
+// OpenAI 兼容出站适配器：IR → OpenAI chat/completions（元炁亦走此协议，type=local 复用）。
+package transform
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"hunjing/shenshi/internal/ir"
+)
+
+// UpstreamError 上游错误：携带 HTTP 状态，供路由层分类冷却（凭证级/绑定级）。
+type UpstreamError struct {
+	StatusCode int
+	Message    string
+}
+
+func (e *UpstreamError) Error() string {
+	return fmt.Sprintf("上游 %d: %s", e.StatusCode, e.Message)
+}
+
+// OpenAICompatAdaptor OpenAI 兼容端点适配器（云厂商兼容面 + 元炁本地端点）。
+type OpenAICompatAdaptor struct {
+	EndpointName string // 端点标识（usage 日志与冷却键）
+	BaseURL      string
+	APIKey       string
+	// UpstreamModel 该端点承载的上游模型名（逻辑模型 → 端点绑定的组成部分）
+	UpstreamModel string
+	Client        *http.Client
+}
+
+// Name 实现 Adaptor。
+func (a *OpenAICompatAdaptor) Name() string { return a.EndpointName }
+
+type oaMessage struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
+type oaRequest struct {
+	Model       string      `json:"model"`
+	Messages    []oaMessage `json:"messages"`
+	Stream      bool        `json:"stream"`
+	Temperature *float64    `json:"temperature,omitempty"`
+	MaxTokens   *int        `json:"max_tokens,omitempty"`
+}
+
+type oaChunk struct {
+	Choices []struct {
+		Delta struct {
+			Content string `json:"content"`
+		} `json:"delta"`
+		FinishReason *string `json:"finish_reason"`
+	} `json:"choices"`
+	Usage *struct {
+		PromptTokens     int `json:"prompt_tokens"`
+		CompletionTokens int `json:"completion_tokens"`
+	} `json:"usage"`
+}
+
+// Stream 实现 Adaptor：SSE 逐帧转 IR Delta。首字节超时由调用方 context 控制。
+func (a *OpenAICompatAdaptor) Stream(ctx context.Context, req *ir.ChatRequest, out chan<- ir.Delta) error {
+	body := oaRequest{
+		Model:    a.UpstreamModel,
+		Messages: make([]oaMessage, 0, len(req.Messages)),
+		Stream:   true,
+	}
+	for _, m := range req.Messages {
+		body.Messages = append(body.Messages, oaMessage{Role: m.Role, Content: m.Content})
+	}
+	body.Temperature = req.Temperature
+	body.MaxTokens = req.MaxTokens
+
+	payload, err := json.Marshal(body)
+	if err != nil {
+		return err
+	}
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, strings.TrimRight(a.BaseURL, "/")+"/v1/chat/completions", bytes.NewReader(payload))
+	if err != nil {
+		return err
+	}
+	httpReq.Header.Set("content-type", "application/json")
+	if a.APIKey != "" {
+		httpReq.Header.Set("authorization", "Bearer "+a.APIKey)
+	}
+
+	client := a.Client
+	if client == nil {
+		client = &http.Client{Timeout: 120 * time.Second}
+	}
+	resp, err := client.Do(httpReq)
+	if err != nil {
+		return &UpstreamError{StatusCode: 0, Message: err.Error()} // 0 = 网络/超时（绑定级冷却）
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode >= 400 {
+		buf := new(bytes.Buffer)
+		_, _ = buf.ReadFrom(resp.Body)
+		return &UpstreamError{StatusCode: resp.StatusCode, Message: buf.String()}
+	}
+
+	scanner := bufio.NewScanner(resp.Body)
+	scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
+	for scanner.Scan() {
+		line := scanner.Text()
+		if !strings.HasPrefix(line, "data:") {
+			continue
+		}
+		data := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
+		if data == "" || data == "[DONE]" {
+			continue
+		}
+		var chunk oaChunk
+		if err := json.Unmarshal([]byte(data), &chunk); err != nil {
+			continue
+		}
+		if len(chunk.Choices) > 0 {
+			c := chunk.Choices[0]
+			if c.Delta.Content != "" {
+				out <- ir.Delta{Delta: c.Delta.Content}
+			}
+			if c.FinishReason != nil {
+				finish := *c.FinishReason
+				usagePrompt, usageCompletion := 0, 0
+				if chunk.Usage != nil {
+					usagePrompt, usageCompletion = chunk.Usage.PromptTokens, chunk.Usage.CompletionTokens
+				}
+				out <- ir.Delta{FinishReason: finish, UsagePromptTokens: usagePrompt, UsageCompletionTokens: usageCompletion}
+			}
+		}
+	}
+	return scanner.Err()
+}