M4 收口:神识用户面与计量闭环——BYOK/配额三段式/usage 落库/Anthropic+Gemini 适配器/神识阁
- 神识 Go:store 双实现(PG 自动建表+60s 就绪重试 / 内存回退)承载 usage_logs·byok_keys·quota_balances;BYOK 密钥库(AES-GCM 加密、API 脱敏、撤回即失效、主密钥未配置 fail-closed);配额三段式(预扣 max_tokens 钳位 1M→结算实际 token→全败幂等退款,HJ-2004);计量归属(IR user_id→usage_logs,/me/usage 用户维度聚合+配额总览,/admin/usage/sample 对账抽样);Anthropic/Gemini 出站适配器(SSE 解析、system 提升、用量映射);stub 端点补名修复计量空 endpoint - IR 契约:user_id/byok_key_id 双端同步(Go types.go + Py types.py) - 灵台:runs.owner_id(迁移 005,server_default 兼容存量)经 kernel 全链透传为 IR user_id;shenshi_client 扩展用户面方法(X-Hunjing-User-Id 代理,Echo 桩带内存实现);/me/byok-keys 登记·列表·撤回与 /me/usage 端点(4 用例) - 前端神识阁(/models 重建):偏好(生成参数表单)/自带密钥(登记加密保存、脱敏列表、撤回二次确认)/用量(聚合表+配额进度条)三区;me API 库+mock+5 用例;侧栏恢复入口 - 配置与部署:.env.example 增 SHENSHI_DATABASE_URL/SHENSHI_BYOK_MASTER_KEY/SHENSHI_USER_BUDGET_TOKENS;compose 神识容器接自有库(init-shenshi.sql 预建 hunjing_shenshi,与灵台不共库共表) - 证据:Go 34 用例(BYOK 独占/配额拒绝/结算退款/归属聚合)、后端 532、前端 210、BFF 4 全绿;跨进程 E2E 四场景(灵台 HttpShenshiClient→真 Go 神识:登记脱敏/BYOK 独占不可达整体失败/用量归属聚合/撤回即失效)通过;rewrite-plan M4 标记完成、能力矩阵神识行升已验证
This commit is contained in:
1 parent
462d570751
commit
144415247e
51 files changed
+3533
-90
No files matched your search
+6
-2
@@ -15,11 +15,15 @@
|
||||
|
||||
```text
|
||||
shenshi/
|
||||
main.go # 入口:配置与路由装配
|
||||
main.go # 入口:store/byok/quota 装配与路由
|
||||
internal/ir/types.go # IR 请求/增量(与灵台 app/shenshi_client/types.py 对齐)
|
||||
internal/transform/ # 出站适配器(Adaptor 接口 + stub)
|
||||
internal/transform/ # 出站适配器(openai 兼容 / anthropic / gemini / stub)
|
||||
internal/route/registry.go # 槽位 → 端点候选注册表
|
||||
internal/relay/sse.go # SSE 帧写出(X-Accel-Buffering: no)
|
||||
internal/store/ # 神识自有库(PG + 内存双实现)
|
||||
internal/byok/ # BYOK 密钥库(AES-GCM 加密与脱敏)
|
||||
internal/quota/ # 配额三段式(预扣/结算/退款,钳位集中)
|
||||
internal/metering/ # 用量计量(stdout + 落库)
|
||||
internal/server/ # HTTP 服务与测试
|
||||
```
|
||||
|
||||
|
||||
+10
-1
@@ -1,3 +1,12 @@
|
||||
module hunjing/shenshi
|
||||
|
||||
go 1.25
|
||||
go 1.25.0
|
||||
|
||||
require (
|
||||
github.com/jackc/pgpassfile v1.0.0 // indirect
|
||||
github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 // indirect
|
||||
github.com/jackc/pgx/v5 v5.10.0 // indirect
|
||||
github.com/jackc/puddle/v2 v2.2.2 // indirect
|
||||
golang.org/x/sync v0.17.0 // indirect
|
||||
golang.org/x/text v0.29.0 // indirect
|
||||
)
|
||||
@@ -0,0 +1,19 @@
|
||||
github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
|
||||
github.com/jackc/pgpassfile v1.0.0 h1:/6Hmqy13Ss2zCq62VdNG8tM1wchn8zjSGOBJ6icpsIM=
|
||||
github.com/jackc/pgpassfile v1.0.0/go.mod h1:CEx0iS5ambNFdcRtxPj5JhEz+xB6uRky5eyVu/W2HEg=
|
||||
github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 h1:iCEnooe7UlwOQYpKFhBabPMi4aNAfoODPEFNiAnClxo=
|
||||
github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761/go.mod h1:5TJZWKEWniPve33vlWYSoGYefn3gLQRzjfDlhSJ9ZKM=
|
||||
github.com/jackc/pgx/v5 v5.10.0 h1:VhSvgU2jSli8o3AqIEOTJr7rZwAEUVo4E4XhR94Zfr0=
|
||||
github.com/jackc/pgx/v5 v5.10.0/go.mod h1:mal1tBGAFfLHvZzaYh77YS/eC6IX9OWbRV1QIIM0Jn4=
|
||||
github.com/jackc/puddle/v2 v2.2.2 h1:PR8nw+E/1w0GLuRFSmiioY6UooMp6KJv0/61nB7icHo=
|
||||
github.com/jackc/puddle/v2 v2.2.2/go.mod h1:vriiEXHvEE654aYKXXjOvZM39qJ0q+azkZFrfEOc3H4=
|
||||
github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
|
||||
github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
|
||||
github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
|
||||
github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
|
||||
golang.org/x/sync v0.17.0 h1:l60nONMj9l5drqw6jlhIELNv9I0A4OFgRsG9k2oT9Ug=
|
||||
golang.org/x/sync v0.17.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
|
||||
golang.org/x/text v0.29.0 h1:1neNs90w9YzJ9BocxfsQNHKuAT4pkghyXc4nhZ6sJvk=
|
||||
golang.org/x/text v0.29.0/go.mod h1:7MhJOA9CD2qZyOKYazxdYMF85OwPdEr9jTtBpO7ydH4=
|
||||
gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
|
||||
gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
|
||||
@@ -0,0 +1,164 @@
|
||||
// Package byok 用户自带密钥:AES-GCM 加密存储与脱敏呈现(shenshi-gateway.md §6)。
|
||||
//
|
||||
// 密钥分层:平台密钥走端点目录(SHENSHI_ENDPOINTS),用户 BYOK 走本包;
|
||||
// 主密钥由 SHENSHI_BYOK_MASTER_KEY 提供(32 字节 hex/base64);未配置时
|
||||
// BYOK 功能整体禁用(fail-closed,HJ-2003),不留明文降级路径。
|
||||
package byok
|
||||
|
||||
import (
|
||||
"crypto/aes"
|
||||
"crypto/cipher"
|
||||
"crypto/rand"
|
||||
"encoding/base64"
|
||||
"encoding/hex"
|
||||
"fmt"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
// Masked API 呈现形态:永不回明文。
|
||||
type Masked struct {
|
||||
ID string `json:"id"`
|
||||
Name string `json:"name"`
|
||||
Type string `json:"type"`
|
||||
BaseURL string `json:"base_url"`
|
||||
Model string `json:"model"`
|
||||
CreatedAt time.Time `json:"created_at"`
|
||||
LastUsed *time.Time `json:"last_used,omitempty"`
|
||||
// APIKeyMask 形如 sk-****abcd(前缀仅示意类型,不泄露中段)
|
||||
APIKeyMask string `json:"api_key_mask"`
|
||||
}
|
||||
|
||||
// Vault BYOK 密钥库。
|
||||
type Vault struct {
|
||||
aesgcm cipher.AEAD
|
||||
store store.Store
|
||||
}
|
||||
|
||||
// NewVault 构造;masterKey 为 32 字节 hex 或 base64。
|
||||
func NewVault(masterKey string, st store.Store) (*Vault, error) {
|
||||
if masterKey == "" {
|
||||
return nil, fmt.Errorf("SHENSHI_BYOK_MASTER_KEY 未配置,BYOK 功能禁用")
|
||||
}
|
||||
key, err := decodeKey(masterKey)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
block, err := aes.NewCipher(key)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
gcm, err := cipher.NewGCM(block)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
return &Vault{aesgcm: gcm, store: st}, nil
|
||||
}
|
||||
|
||||
func decodeKey(raw string) ([]byte, error) {
|
||||
raw = strings.TrimSpace(raw)
|
||||
if b, err := hex.DecodeString(raw); err == nil && len(b) == 32 {
|
||||
return b, nil
|
||||
}
|
||||
if b, err := base64.StdEncoding.DecodeString(raw); err == nil && len(b) == 32 {
|
||||
return b, nil
|
||||
}
|
||||
return nil, fmt.Errorf("SHENSHI_BYOK_MASTER_KEY 须为 32 字节 hex 或 base64")
|
||||
}
|
||||
|
||||
func (v *Vault) encrypt(plaintext string) (string, error) {
|
||||
nonce := make([]byte, v.aesgcm.NonceSize())
|
||||
if _, err := rand.Read(nonce); err != nil {
|
||||
return "", err
|
||||
}
|
||||
sealed := v.aesgcm.Seal(nil, nonce, []byte(plaintext), nil)
|
||||
return base64.StdEncoding.EncodeToString(append(nonce, sealed...)), nil
|
||||
}
|
||||
|
||||
func (v *Vault) decrypt(encoded string) (string, error) {
|
||||
raw, err := base64.StdEncoding.DecodeString(encoded)
|
||||
if err != nil {
|
||||
return "", err
|
||||
}
|
||||
n := v.aesgcm.NonceSize()
|
||||
if len(raw) <= n {
|
||||
return "", fmt.Errorf("密文形态非法")
|
||||
}
|
||||
plain, err := v.aesgcm.Open(nil, raw[:n], raw[n:], nil)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("HJ-2003 BYOK 密钥解密失败")
|
||||
}
|
||||
return string(plain), nil
|
||||
}
|
||||
|
||||
// Register 登记用户密钥,返回脱敏形态。
|
||||
func (v *Vault) Register(userID, name, typ, baseURL, apiKey, model string) (*Masked, error) {
|
||||
enc, err := v.encrypt(apiKey)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
key := store.ByokKey{
|
||||
ID: newID(), UserID: userID, Name: name, Type: typ,
|
||||
BaseURL: baseURL, APIKey: enc, Model: model, CreatedAt: time.Now().UTC(),
|
||||
}
|
||||
if err := v.store.UpsertByok(key); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
m := mask(key, apiKey)
|
||||
return &m, nil
|
||||
}
|
||||
|
||||
// List 脱敏列举用户有效密钥。
|
||||
func (v *Vault) List(userID string) ([]Masked, error) {
|
||||
keys, err := v.store.ListByok(userID)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
out := make([]Masked, 0, len(keys))
|
||||
for _, k := range keys {
|
||||
out = append(out, mask(k, ""))
|
||||
}
|
||||
return out, nil
|
||||
}
|
||||
|
||||
// Resolve 取解密后的密钥材料(路由期专用,结果不落日志)。
|
||||
func (v *Vault) Resolve(userID, id string) (*store.ByokKey, string, error) {
|
||||
k, err := v.store.GetByok(userID, id)
|
||||
if err != nil || k == nil {
|
||||
return nil, "", fmt.Errorf("HJ-2003 BYOK 密钥不存在或已撤回")
|
||||
}
|
||||
plain, err := v.decrypt(k.APIKey)
|
||||
if err != nil {
|
||||
return nil, "", err
|
||||
}
|
||||
now := time.Now().UTC()
|
||||
_ = v.store.TouchByok(id, now)
|
||||
k.LastUsed = &now
|
||||
return k, plain, nil
|
||||
}
|
||||
|
||||
// Revoke 撤回(软删,立即失效)。
|
||||
func (v *Vault) Revoke(userID, id string) error {
|
||||
return v.store.RevokeByok(userID, id)
|
||||
}
|
||||
|
||||
func mask(k store.ByokKey, plainForPrefix string) Masked {
|
||||
prefix := ""
|
||||
if plainForPrefix != "" {
|
||||
prefix = plainForPrefix[:min(3, len(plainForPrefix))]
|
||||
}
|
||||
m := Masked{
|
||||
ID: k.ID, Name: k.Name, Type: k.Type, BaseURL: k.BaseURL, Model: k.Model,
|
||||
CreatedAt: k.CreatedAt, LastUsed: k.LastUsed,
|
||||
APIKeyMask: prefix + "****",
|
||||
}
|
||||
return m
|
||||
}
|
||||
|
||||
func newID() string {
|
||||
b := make([]byte, 10)
|
||||
_, _ = rand.Read(b)
|
||||
return "byok_" + hex.EncodeToString(b)
|
||||
}
|
||||
@@ -0,0 +1,77 @@
|
||||
package byok
|
||||
|
||||
import (
|
||||
"encoding/hex"
|
||||
"strings"
|
||||
"testing"
|
||||
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
func testMaster() string {
|
||||
b := make([]byte, 32)
|
||||
for i := range b {
|
||||
b[i] = byte(i)
|
||||
}
|
||||
return hex.EncodeToString(b)
|
||||
}
|
||||
|
||||
func TestVaultRoundTripAndMask(t *testing.T) {
|
||||
v, err := NewVault(testMaster(), store.NewMem())
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
m, err := v.Register("u1", "主力", "anthropic", "https://api.anthropic.com", "sk-ant--secret123", "claude-x")
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if !strings.Contains(m.APIKeyMask, "****") || strings.Contains(m.APIKeyMask, "secret") {
|
||||
t.Fatalf("掩码泄露明文: %s", m.APIKeyMask)
|
||||
}
|
||||
|
||||
// 解密往返
|
||||
k, plain, err := v.Resolve("u1", m.ID)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if plain != "sk-ant--secret123" || k.Model != "claude-x" {
|
||||
t.Fatalf("往返失真: %s %s", plain, k.Model)
|
||||
}
|
||||
if k.LastUsed == nil {
|
||||
t.Fatal("使用时间应被记录")
|
||||
}
|
||||
|
||||
// 列表脱敏
|
||||
keys, _ := v.List("u1")
|
||||
if len(keys) != 1 || strings.Contains(keys[0].APIKeyMask, "secret") {
|
||||
t.Fatalf("列表脱敏失败: %+v", keys)
|
||||
}
|
||||
|
||||
// 撤回后失效
|
||||
if err := v.Revoke("u1", m.ID); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if _, _, err := v.Resolve("u1", m.ID); err == nil {
|
||||
t.Fatal("撤回后仍可解析")
|
||||
}
|
||||
}
|
||||
|
||||
func TestVaultFailClosedWithoutMasterKey(t *testing.T) {
|
||||
if _, err := NewVault("", store.NewMem()); err == nil {
|
||||
t.Fatal("未配置主密钥应拒绝启用(fail-closed)")
|
||||
}
|
||||
}
|
||||
|
||||
func TestVaultRejectsBadMasterKey(t *testing.T) {
|
||||
if _, err := NewVault("too-short", store.NewMem()); err == nil {
|
||||
t.Fatal("非 32 字节主密钥应被拒绝")
|
||||
}
|
||||
}
|
||||
|
||||
func TestVaultCrossUserIsolation(t *testing.T) {
|
||||
v, _ := NewVault(testMaster(), store.NewMem())
|
||||
m, _ := v.Register("u1", "k", "gemini", "https://gen.example.com", "AIza-key", "gemini-x")
|
||||
if _, _, err := v.Resolve("u2", m.ID); err == nil {
|
||||
t.Fatal("u2 不应解析 u1 的密钥")
|
||||
}
|
||||
}
|
||||
@@ -22,9 +22,13 @@ type ChatRequest struct {
|
||||
RunID string `json:"run_id"`
|
||||
TraceID string `json:"trace_id,omitempty"`
|
||||
CharacterID string `json:"character_id,omitempty"`
|
||||
Temperature *float64 `json:"temperature,omitempty"`
|
||||
MaxTokens *int `json:"max_tokens,omitempty"`
|
||||
Stream bool `json:"stream"`
|
||||
// UserID 用量归属与配额裁决维度(灵台鉴权后注入;空则记为平台调用)
|
||||
UserID string `json:"user_id,omitempty"`
|
||||
// ByokKeyID 用户自带密钥引用(配置继承的最终解析结果,神识不自行决定)
|
||||
ByokKeyID string `json:"byok_key_id,omitempty"`
|
||||
Temperature *float64 `json:"temperature,omitempty"`
|
||||
MaxTokens *int `json:"max_tokens,omitempty"`
|
||||
Stream bool `json:"stream"`
|
||||
}
|
||||
|
||||
// Delta 神识标准化流增量(与 Provider 方言无关)。
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
// 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 用量计量:每次尝试一条记录(stdout 结构化日志 + 神识自有库落库)。
|
||||
// 未配置 SHENSHI_DATABASE_URL 时退化为进程内环形缓冲(开发档),/me/usage 与对账抽样同语义。
|
||||
package metering
|
||||
|
||||
import (
|
||||
@@ -7,52 +7,57 @@ import (
|
||||
"log"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
// Attempt 一次端点尝试的计量记录。
|
||||
// Attempt 一次端点尝试的计量记录(IR 维度齐全,user_id 为归属计量主键)。
|
||||
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"`
|
||||
ID string `json:"id"` // run_id#endpoint 序号(结算幂等键)
|
||||
RunID string `json:"run_id"`
|
||||
CharacterID string `json:"character_id,omitempty"`
|
||||
UserID string `json:"user_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)。
|
||||
// Log 计量器:stdout JSON 始终输出;持久化走 store(nil 时仅进程内有界环形)。
|
||||
type Log struct {
|
||||
mu sync.Mutex
|
||||
mu sync.Mutex
|
||||
store store.Store
|
||||
entries []Attempt
|
||||
limit int
|
||||
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}
|
||||
// New 构造计量器(st 为 nil 时进程内保留最近 N 条)。
|
||||
func New(limit int, st store.Store) *Log {
|
||||
return &Log{store: st, entries: make([]Attempt, 0, limit), limit: limit}
|
||||
}
|
||||
|
||||
// Record 记录一次尝试:stdout JSON 日志 + 内存追加。
|
||||
// Record 记录一次尝试:stdout JSON 日志(对账抽样基准)+ 落库。
|
||||
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:]))
|
||||
}
|
||||
if l.store != nil {
|
||||
_ = l.store.InsertUsage(store.UsageLog{
|
||||
ID: a.ID, RunID: a.RunID, CharacterID: a.CharacterID, UserID: a.UserID,
|
||||
Model: a.Model, Slot: a.Slot, Endpoint: a.Endpoint,
|
||||
Status: a.Status, ErrorCode: a.ErrorCode,
|
||||
FirstTokenMS: a.FirstTokenMS, TotalMS: a.TotalMS,
|
||||
PromptTokens: a.PromptTokens, CompletionTokens: a.CompletionTokens, At: a.At,
|
||||
})
|
||||
return
|
||||
}
|
||||
l.mu.Lock()
|
||||
defer l.mu.Unlock()
|
||||
if len(l.entries) >= l.limit {
|
||||
@@ -61,16 +66,22 @@ func (l *Log) Record(a Attempt) {
|
||||
l.entries = append(l.entries, a)
|
||||
}
|
||||
|
||||
// Aggregate 按 模型×端点 聚合。
|
||||
func (l *Log) Aggregate() []UsageAgg {
|
||||
// Aggregate 按 用户×模型×端点 聚合(userID 为空聚合全体;管理面语义)。
|
||||
func (l *Log) Aggregate(userID string) ([]store.UsageAgg, error) {
|
||||
if l.store != nil {
|
||||
return l.store.AggregateUsage(userID)
|
||||
}
|
||||
l.mu.Lock()
|
||||
defer l.mu.Unlock()
|
||||
agg := map[string]*UsageAgg{}
|
||||
agg := map[string]*store.UsageAgg{}
|
||||
for _, e := range l.entries {
|
||||
key := e.Model + "|" + e.Endpoint
|
||||
if userID != "" && e.UserID != userID {
|
||||
continue
|
||||
}
|
||||
key := e.UserID + "|" + e.Model + "|" + e.Endpoint
|
||||
row := agg[key]
|
||||
if row == nil {
|
||||
row = &UsageAgg{Model: e.Model, Endpoint: e.Endpoint}
|
||||
row = &store.UsageAgg{UserID: e.UserID, Model: e.Model, Endpoint: e.Endpoint}
|
||||
agg[key] = row
|
||||
}
|
||||
row.Requests++
|
||||
@@ -80,9 +91,33 @@ func (l *Log) Aggregate() []UsageAgg {
|
||||
row.CompletionTokens += e.CompletionTokens
|
||||
}
|
||||
}
|
||||
out := make([]UsageAgg, 0, len(agg))
|
||||
out := make([]store.UsageAgg, 0, len(agg))
|
||||
for _, row := range agg {
|
||||
out = append(out, *row)
|
||||
}
|
||||
return out
|
||||
return out, nil
|
||||
}
|
||||
|
||||
// Sample 对账抽样:最近 N 条原始记录(与 Provider 账单比对)。
|
||||
func (l *Log) Sample(limit int) ([]store.UsageLog, error) {
|
||||
if l.store != nil {
|
||||
return l.store.SampleUsage(limit)
|
||||
}
|
||||
l.mu.Lock()
|
||||
defer l.mu.Unlock()
|
||||
n := limit
|
||||
if n > len(l.entries) {
|
||||
n = len(l.entries)
|
||||
}
|
||||
out := make([]store.UsageLog, n)
|
||||
for i, e := range l.entries[len(l.entries)-n:] {
|
||||
out[i] = store.UsageLog{
|
||||
ID: e.ID, RunID: e.RunID, CharacterID: e.CharacterID, UserID: e.UserID,
|
||||
Model: e.Model, Slot: e.Slot, Endpoint: e.Endpoint,
|
||||
Status: e.Status, ErrorCode: e.ErrorCode,
|
||||
FirstTokenMS: e.FirstTokenMS, TotalMS: e.TotalMS,
|
||||
PromptTokens: e.PromptTokens, CompletionTokens: e.CompletionTokens, At: e.At,
|
||||
}
|
||||
}
|
||||
return out, nil
|
||||
}
|
||||
@@ -0,0 +1,86 @@
|
||||
// Package quota 用户级软配额:三段式(预扣 → 结算 → 退款,shenshi-gateway.md §7)。
|
||||
//
|
||||
// P0/P1 只做预算熔断(token 口径,HJ-2004 配额不足);1.0 官方云计费属 P2。
|
||||
// 换算与钳位集中在本包单一函数,饱和截断防溢出变负。
|
||||
package quota
|
||||
|
||||
import (
|
||||
"errors"
|
||||
"fmt"
|
||||
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
// ErrInsufficient 配额不足(HJ-2004)。
|
||||
var ErrInsufficient = errors.New("HJ-2004 配额不足")
|
||||
|
||||
// Limiter 用户配额裁决器;budget=0 表示不限制(本地开发档)。
|
||||
type Limiter struct {
|
||||
store store.Store
|
||||
budget int64
|
||||
}
|
||||
|
||||
// New 构造裁决器。
|
||||
func New(st store.Store, budgetTokens int64) *Limiter {
|
||||
return &Limiter{store: st, budget: budgetTokens}
|
||||
}
|
||||
|
||||
// clamp 集中单一换算函数:预扣额 = min(请求上限, 钳位上限),且饱和截断不为负。
|
||||
// 请求侧任何计费乘数(max_tokens 等)在进入本函数前已由 IR 校验限幅。
|
||||
func clamp(requested int64) int64 {
|
||||
const reserveCeiling = 1 << 20 // 1M token 单次预扣上限(校验期设上限,防溢出)
|
||||
if requested <= 0 {
|
||||
return 0
|
||||
}
|
||||
if requested > reserveCeiling {
|
||||
return reserveCeiling
|
||||
}
|
||||
return requested
|
||||
}
|
||||
|
||||
// Reserve 预扣(按上限):配额不足时拒绝并返回 ErrInsufficient。
|
||||
func (l *Limiter) Reserve(userID string, requested int64) (reserved int64, err error) {
|
||||
reserved = clamp(requested)
|
||||
if l.budget <= 0 || reserved <= 0 {
|
||||
// 不限制档仍记账(可见性),但不裁决
|
||||
if reserved > 0 {
|
||||
return reserved, l.store.Reserve(userID, reserved)
|
||||
}
|
||||
return 0, nil
|
||||
}
|
||||
b, err := l.store.Balance(userID)
|
||||
if err != nil {
|
||||
return 0, err
|
||||
}
|
||||
if b.Reserved+b.Used+reserved > l.budget {
|
||||
return 0, fmt.Errorf("%w:已用 %d + 预扣 %d + 本次 %d > 预算 %d",
|
||||
ErrInsufficient, b.Used, b.Reserved, reserved, l.budget)
|
||||
}
|
||||
if err := l.store.Reserve(userID, reserved); err != nil {
|
||||
return 0, err
|
||||
}
|
||||
return reserved, nil
|
||||
}
|
||||
|
||||
// Settle 结算:实际用量回填,差额(预扣-实际)即时释放。幂等性由 usage_logs
|
||||
// 主键(同一尝试 ID 只落一条)保证:结算与落库同批执行,重放不会二次扣减。
|
||||
func (l *Limiter) Settle(userID string, reserved, actual int64) error {
|
||||
return l.store.Settle(userID, reserved, clamp(actual))
|
||||
}
|
||||
|
||||
// Refund 失败退款:全额释放预扣(幂等——释放至多为预扣额,GREATEST 饱和截断)。
|
||||
func (l *Limiter) Refund(userID string, reserved int64) error {
|
||||
if reserved <= 0 {
|
||||
return nil
|
||||
}
|
||||
return l.store.Refund(userID, reserved)
|
||||
}
|
||||
|
||||
// Budget 预算与余额总览(/me/usage 附带)。
|
||||
func (l *Limiter) Budget(userID string) (budget, used, reserved int64, err error) {
|
||||
b, err := l.store.Balance(userID)
|
||||
if err != nil {
|
||||
return 0, 0, 0, err
|
||||
}
|
||||
return l.budget, b.Used, b.Reserved, nil
|
||||
}
|
||||
@@ -0,0 +1,65 @@
|
||||
package quota
|
||||
|
||||
import (
|
||||
"errors"
|
||||
"testing"
|
||||
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
func TestReserveRejectsWhenBudgetExhausted(t *testing.T) {
|
||||
st := store.NewMem()
|
||||
l := New(st, 1000)
|
||||
if _, err := l.Reserve("u1", 600); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if _, err := l.Reserve("u1", 600); err == nil {
|
||||
t.Fatal("累计超预算应拒绝(HJ-2004)")
|
||||
} else if !errors.Is(err, ErrInsufficient) {
|
||||
t.Fatalf("错误类型错误: %v", err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestSettleReleasesAndAllowsReuse(t *testing.T) {
|
||||
st := store.NewMem()
|
||||
l := New(st, 1000)
|
||||
r, _ := l.Reserve("u1", 600)
|
||||
if err := l.Settle("u1", r, 100); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
// 已用 100,剩 900 可用
|
||||
if _, err := l.Reserve("u1", 900); err != nil {
|
||||
t.Fatalf("结算释放差值后应可继续使用: %v", err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestRefundIdempotent(t *testing.T) {
|
||||
st := store.NewMem()
|
||||
l := New(st, 1000)
|
||||
r, _ := l.Reserve("u1", 500)
|
||||
_ = l.Refund("u1", r)
|
||||
_ = l.Refund("u1", r) // 幂等重放
|
||||
budget, used, reserved, _ := l.Budget("u1")
|
||||
if reserved != 0 || used != 0 || budget != 1000 {
|
||||
t.Fatalf("重复退款后账目异常: budget=%d used=%d reserved=%d", budget, used, reserved)
|
||||
}
|
||||
}
|
||||
|
||||
func TestClampCeiling(t *testing.T) {
|
||||
if got := clamp(1 << 30); got != 1<<20 {
|
||||
t.Fatalf("超大预扣应钳位到 1M,实际 %d", got)
|
||||
}
|
||||
if got := clamp(-5); got != 0 {
|
||||
t.Fatalf("负值应截断为 0,实际 %d", got)
|
||||
}
|
||||
}
|
||||
|
||||
func TestZeroBudgetMeansUnlimited(t *testing.T) {
|
||||
st := store.NewMem()
|
||||
l := New(st, 0)
|
||||
for i := 0; i < 3; i++ {
|
||||
if _, err := l.Reserve("u1", 1<<20); err != nil {
|
||||
t.Fatalf("budget=0 不应裁决: %v", err)
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -54,19 +54,14 @@ func New() *Registry {
|
||||
}
|
||||
specs := loadSpecs()
|
||||
if len(specs) == 0 {
|
||||
r.Register("chat", Endpoint{Type: "stub", Adaptor: transform.StubAdaptor{}, Priority: 100, UpstreamModel: "stub"})
|
||||
r.Register("chat", Endpoint{Name: "stub", Type: "stub", Adaptor: transform.StubAdaptor{}, Priority: 100, UpstreamModel: "stub"})
|
||||
return r
|
||||
}
|
||||
for _, spec := range specs {
|
||||
if spec.Type != "openai_compatible" && spec.Type != "local" {
|
||||
adaptor, ok := BuildAdaptor(spec)
|
||||
if !ok {
|
||||
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
|
||||
@@ -79,6 +74,29 @@ func New() *Registry {
|
||||
return r
|
||||
}
|
||||
|
||||
// BuildAdaptor 按端点类型构造出站适配器(契约词典 §14.2;BYOK 动态候选复用)。
|
||||
func BuildAdaptor(spec EndpointSpec) (transform.Adaptor, bool) {
|
||||
switch spec.Type {
|
||||
case "openai_compatible", "local":
|
||||
return &transform.OpenAICompatAdaptor{
|
||||
EndpointName: spec.Name, BaseURL: spec.BaseURL,
|
||||
APIKey: spec.APIKey, UpstreamModel: spec.Model,
|
||||
}, true
|
||||
case "anthropic":
|
||||
return &transform.AnthropicAdaptor{
|
||||
EndpointName: spec.Name, BaseURL: spec.BaseURL,
|
||||
APIKey: spec.APIKey, UpstreamModel: spec.Model,
|
||||
}, true
|
||||
case "gemini":
|
||||
return &transform.GeminiAdaptor{
|
||||
EndpointName: spec.Name, BaseURL: spec.BaseURL,
|
||||
APIKey: spec.APIKey, UpstreamModel: spec.Model,
|
||||
}, true
|
||||
default:
|
||||
return nil, false
|
||||
}
|
||||
}
|
||||
|
||||
func loadSpecs() []EndpointSpec {
|
||||
raw := os.Getenv("SHENSHI_ENDPOINTS")
|
||||
if raw == "" {
|
||||
|
||||
@@ -0,0 +1,195 @@
|
||||
// BYOK / 配额三段式 / 用量归属的服务级场景测试。
|
||||
package server
|
||||
|
||||
import (
|
||||
"bufio"
|
||||
"encoding/hex"
|
||||
"encoding/json"
|
||||
"net/http"
|
||||
"net/http/httptest"
|
||||
"strings"
|
||||
"testing"
|
||||
"time"
|
||||
|
||||
"hunjing/shenshi/internal/byok"
|
||||
"hunjing/shenshi/internal/quota"
|
||||
"hunjing/shenshi/internal/route"
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
func testMasterKey() string {
|
||||
b := make([]byte, 32)
|
||||
for i := range b {
|
||||
b[i] = byte(i + 1)
|
||||
}
|
||||
return hex.EncodeToString(b)
|
||||
}
|
||||
|
||||
// fullServer 装配 store/byok/quota 的完整服务;eps 为注册表候选(按优先级升序尝试)。
|
||||
func fullServer(t *testing.T, budget int64, eps ...route.Endpoint) (*httptest.Server, store.Store, *byok.Vault) {
|
||||
t.Helper()
|
||||
st := store.NewMem()
|
||||
vault, err := byok.NewVault(testMasterKey(), st)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
registry := registryWith(t, time.Second, eps...)
|
||||
srv := New(Config{Store: st, Vault: vault, Quota: quota.New(st, budget)}, registry)
|
||||
ts := httptest.NewServer(srv)
|
||||
t.Cleanup(ts.Close)
|
||||
return ts, st, vault
|
||||
}
|
||||
|
||||
func postChat(ts *httptest.Server, body, userID string) (*http.Response, string) {
|
||||
req, _ := http.NewRequest(http.MethodPost, ts.URL+"/v1/chat/completions", strings.NewReader(body))
|
||||
req.Header.Set("content-type", "application/json")
|
||||
if userID != "" {
|
||||
req.Header.Set("X-Hunjing-User-Id", userID)
|
||||
}
|
||||
resp, err := ts.Client().Do(req)
|
||||
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 byokChatBody(run, extra string) string {
|
||||
return `{"ir_version":1,"slot":"chat","messages":[{"role":"user","content":"hi"}],"run_id":"` + run + `"` + extra + `,"stream":true}`
|
||||
}
|
||||
|
||||
// 场景:BYOK 独占路由——请求携带 byok_key_id 时只走用户端点,不回落平台候选。
|
||||
func TestByokExclusiveRouting(t *testing.T) {
|
||||
platform := fakeOpenAI(t, 0, 0, "平台端点")
|
||||
defer platform.Close()
|
||||
ts, _, vault := fullServer(t, 0, oaEndpoint("platform", platform.URL, 10))
|
||||
|
||||
byokUpstream := fakeOpenAI(t, 0, 0, "用户端点")
|
||||
defer byokUpstream.Close()
|
||||
key, err := vault.Register("u1", "自带", "openai_compatible", byokUpstream.URL, "sk-mine", "my-model")
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
|
||||
body := byokChatBody("run_byok", `,"user_id":"u1","byok_key_id":"`+key.ID+`"`)
|
||||
resp, out := postChat(ts, body, "u1")
|
||||
if resp == nil || resp.StatusCode != 200 {
|
||||
t.Fatalf("BYOK 请求失败: %s", out)
|
||||
}
|
||||
if !strings.Contains(out, "用户端点") || strings.Contains(out, "平台端点") {
|
||||
t.Fatalf("应独占 BYOK 端点: %s", out)
|
||||
}
|
||||
}
|
||||
|
||||
// 场景:BYOK 无效引用 → 整体失败,不回落平台端点。
|
||||
func TestByokInvalidReferenceRejected(t *testing.T) {
|
||||
platform := fakeOpenAI(t, 0, 0, "平台端点")
|
||||
defer platform.Close()
|
||||
ts, _, _ := fullServer(t, 0, oaEndpoint("platform", platform.URL, 10))
|
||||
|
||||
resp, out := postChat(ts, byokChatBody("run_bad", `,"user_id":"u1","byok_key_id":"byok_missing"`), "u1")
|
||||
if resp == nil || resp.StatusCode != http.StatusServiceUnavailable {
|
||||
t.Fatalf("应 503: %s", out)
|
||||
}
|
||||
if !strings.Contains(out, "HJ-2001") || strings.Contains(out, "平台端点") {
|
||||
t.Fatalf("应整体失败而非回落: %s", out)
|
||||
}
|
||||
}
|
||||
|
||||
// 场景:配额不足 → HJ-2004 预拒,且不产生用量记录。
|
||||
func TestQuotaRejectsWhenExhausted(t *testing.T) {
|
||||
upstream := fakeOpenAI(t, 0, 0, "内容")
|
||||
defer upstream.Close()
|
||||
ts, st, _ := fullServer(t, 10, oaEndpoint("up", upstream.URL, 10))
|
||||
|
||||
resp, out := postChat(ts, byokChatBody("run_q", `,"user_id":"u1","max_tokens":100`), "u1")
|
||||
if resp == nil || resp.StatusCode != http.StatusPaymentRequired {
|
||||
t.Fatalf("配额不足应 402: %s", out)
|
||||
}
|
||||
if !strings.Contains(out, "HJ-2004") {
|
||||
t.Fatalf("应携带 HJ-2004: %s", out)
|
||||
}
|
||||
sample, _ := st.SampleUsage(10)
|
||||
if len(sample) != 0 {
|
||||
t.Fatalf("预拒请求不应产生用量记录: %+v", sample)
|
||||
}
|
||||
}
|
||||
|
||||
// 场景:成功结算释放差值;全败全额退款——余额不残留。
|
||||
func TestQuotaSettleAndRefundBalance(t *testing.T) {
|
||||
upstream := fakeOpenAI(t, 0, 0, "内容")
|
||||
bad := fakeOpenAI(t, 0, http.StatusForbidden, "")
|
||||
defer upstream.Close()
|
||||
defer bad.Close()
|
||||
ts, st, _ := fullServer(t, 1000, oaEndpoint("up", upstream.URL, 10))
|
||||
|
||||
// 成功:预扣 max_tokens=200,实际 3+4=7
|
||||
if resp, out := postChat(ts, byokChatBody("run_ok", `,"user_id":"u1","max_tokens":200`), "u1"); resp == nil || resp.StatusCode != 200 {
|
||||
t.Fatalf("成功请求失败: %s", out)
|
||||
}
|
||||
b, _ := st.Balance("u1")
|
||||
if b.Used != 7 || b.Reserved != 0 {
|
||||
t.Fatalf("结算后账目错误: %+v", b)
|
||||
}
|
||||
|
||||
// 已用 7,预算 1000:请求 max_tokens 994 应被拒(7+994 > 1000)
|
||||
if resp, out := postChat(ts, byokChatBody("run_over", `,"user_id":"u1","max_tokens":994`), "u1"); resp == nil || resp.StatusCode != http.StatusPaymentRequired {
|
||||
t.Fatalf("超预算应拒: %s", out)
|
||||
}
|
||||
|
||||
// 全候选失败 → 全额退款
|
||||
tsBad, stBad, _ := fullServer(t, 1000, oaEndpoint("bad", bad.URL, 10))
|
||||
if resp, out := postChat(tsBad, byokChatBody("run_af", `,"user_id":"u2","max_tokens":500`), "u2"); resp == nil || resp.StatusCode != http.StatusServiceUnavailable {
|
||||
t.Fatalf("全败应 503: %s", out)
|
||||
}
|
||||
b2, _ := stBad.Balance("u2")
|
||||
if b2.Reserved != 0 || b2.Used != 0 {
|
||||
t.Fatalf("全败后应全额退款: %+v", b2)
|
||||
}
|
||||
}
|
||||
|
||||
// 场景:用量归属计量——/me/usage 按用户过滤并附配额总览。
|
||||
func TestUsageAttributionByUser(t *testing.T) {
|
||||
upstream := fakeOpenAI(t, 0, 0, "内容")
|
||||
defer upstream.Close()
|
||||
ts, _, _ := fullServer(t, 0, oaEndpoint("up", upstream.URL, 10))
|
||||
|
||||
for _, b := range []string{
|
||||
byokChatBody("run_u1", `,"user_id":"u1"`),
|
||||
byokChatBody("run_u2", `,"user_id":"u2"`),
|
||||
byokChatBody("run_u1b", `,"user_id":"u1"`),
|
||||
} {
|
||||
if resp, out := postChat(ts, b, ""); resp == nil || resp.StatusCode != 200 {
|
||||
t.Fatalf("请求失败: %s", out)
|
||||
}
|
||||
}
|
||||
|
||||
req, _ := http.NewRequest(http.MethodGet, ts.URL+"/me/usage", nil)
|
||||
req.Header.Set("X-Hunjing-User-Id", "u1")
|
||||
resp, err := ts.Client().Do(req)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
var payload struct {
|
||||
Rows []struct {
|
||||
UserID string `json:"user_id"`
|
||||
Requests int `json:"requests"`
|
||||
} `json:"rows"`
|
||||
Quota *struct {
|
||||
Budget int64 `json:"budget"`
|
||||
Used int64 `json:"used"`
|
||||
} `json:"quota"`
|
||||
}
|
||||
_ = json.NewDecoder(resp.Body).Decode(&payload)
|
||||
if len(payload.Rows) != 1 || payload.Rows[0].UserID != "u1" || payload.Rows[0].Requests != 2 {
|
||||
t.Fatalf("u1 聚合错误: %+v", payload.Rows)
|
||||
}
|
||||
if payload.Quota == nil || payload.Quota.Used != 14 {
|
||||
t.Fatalf("配额总览错误: %+v", payload.Quota)
|
||||
}
|
||||
}
|
||||
@@ -150,12 +150,14 @@ func TestUsageAggregate(t *testing.T) {
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
var agg []map[string]any
|
||||
var payload struct {
|
||||
Rows []map[string]any `json:"rows"`
|
||||
}
|
||||
defer func() { _ = usageResp.Body.Close() }()
|
||||
if err := json.NewDecoder(usageResp.Body).Decode(&agg); err != nil {
|
||||
if err := json.NewDecoder(usageResp.Body).Decode(&payload); 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)
|
||||
if len(payload.Rows) != 1 || payload.Rows[0]["ok"].(float64) != 1 || payload.Rows[0]["prompt_tokens"].(float64) != 3 {
|
||||
t.Fatalf("聚合不符:%v", payload.Rows)
|
||||
}
|
||||
}
|
||||
@@ -1,25 +1,36 @@
|
||||
// Package server 神识 HTTP 服务:数据面(IR 入站 → SSE 出站)、健康面与用量面。
|
||||
// Package server 神识 HTTP 服务:数据面(IR 入站 → SSE 出站)、健康面、用量面与 BYOK 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)。
|
||||
//
|
||||
// 配额三段式:预扣(max_tokens 上限,钳位)→ 结算(实际 token)→ 退款(失败幂等)。
|
||||
// BYOK:请求携带 byok_key_id 时独占候选(不回落平台端点,防止用户预期自带密钥却被平台计费)。
|
||||
package server
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"fmt"
|
||||
"net/http"
|
||||
"time"
|
||||
|
||||
"hunjing/shenshi/internal/byok"
|
||||
"hunjing/shenshi/internal/ir"
|
||||
"hunjing/shenshi/internal/metering"
|
||||
"hunjing/shenshi/internal/quota"
|
||||
"hunjing/shenshi/internal/relay"
|
||||
"hunjing/shenshi/internal/route"
|
||||
"hunjing/shenshi/internal/store"
|
||||
"hunjing/shenshi/internal/transform"
|
||||
)
|
||||
|
||||
// Config 服务配置。
|
||||
type Config struct {
|
||||
ServiceToken string
|
||||
Store store.Store // nil 时计量退化为进程内环形
|
||||
Vault *byok.Vault // nil 时 BYOK 面整体 503(fail-closed)
|
||||
Quota *quota.Limiter
|
||||
}
|
||||
|
||||
// Server 神识 HTTP 服务。
|
||||
@@ -32,11 +43,19 @@ type Server struct {
|
||||
|
||||
// New 构建服务并注册路由。
|
||||
func New(cfg Config, routes *route.Registry) *Server {
|
||||
s := &Server{cfg: cfg, routes: routes, usage: metering.New(10_000), mux: http.NewServeMux()}
|
||||
s := &Server{
|
||||
cfg: cfg, routes: routes,
|
||||
usage: metering.New(10_000, cfg.Store),
|
||||
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)))
|
||||
s.mux.Handle("GET /admin/usage/sample", s.auth(http.HandlerFunc(s.handleUsageSample)))
|
||||
s.mux.Handle("GET /me/byok-keys", s.auth(http.HandlerFunc(s.handleByokList)))
|
||||
s.mux.Handle("POST /me/byok-keys", s.auth(http.HandlerFunc(s.handleByokRegister)))
|
||||
s.mux.Handle("DELETE /me/byok-keys/{id}", s.auth(http.HandlerFunc(s.handleByokRevoke)))
|
||||
return s
|
||||
}
|
||||
|
||||
@@ -58,7 +77,10 @@ func (s *Server) auth(next http.Handler) http.Handler {
|
||||
})
|
||||
}
|
||||
|
||||
// handleChat 数据面:IR → 候选迭代(首字超时/冷却标记/粘性)→ SSE。
|
||||
// defaultReserve 预扣默认上限(max_tokens 未指定时按 4K 输出预扣)。
|
||||
const defaultReserve = 4096
|
||||
|
||||
// handleChat 数据面:配额预扣 → 候选迭代(首字超时/冷却标记/粘性/BYOK 独占)→ SSE → 结算/退款。
|
||||
func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
|
||||
var req ir.ChatRequest
|
||||
if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
|
||||
@@ -69,20 +91,37 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
|
||||
writeError(w, http.StatusBadRequest, "HJ-2002", err.Error())
|
||||
return
|
||||
}
|
||||
candidates, err := s.routes.Candidates(req.Slot, req.RunID)
|
||||
|
||||
// 三段式之一:预扣(按输出上限;钳位集中在 quota 包)
|
||||
reserveTokens := int64(defaultReserve)
|
||||
if req.MaxTokens != nil {
|
||||
reserveTokens = int64(*req.MaxTokens)
|
||||
}
|
||||
reserved, err := s.quotaReserve(req.UserID, reserveTokens)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusPaymentRequired, "HJ-2004", err.Error())
|
||||
return
|
||||
}
|
||||
|
||||
candidates, err := s.candidates(&req)
|
||||
if err != nil {
|
||||
_ = s.quotaRefund(req.UserID, reserved)
|
||||
writeError(w, http.StatusServiceUnavailable, "HJ-2001", err.Error())
|
||||
return
|
||||
}
|
||||
|
||||
var lastErr error
|
||||
for _, endpoint := range candidates {
|
||||
res := s.tryEndpoint(w, r, endpoint, &req)
|
||||
for i, endpoint := range candidates {
|
||||
res := s.tryEndpoint(w, r, endpoint, &req, i)
|
||||
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)
|
||||
s.quotaSettle(req.UserID, reserved, res.Attempt.PromptTokens+res.Attempt.CompletionTokens)
|
||||
} else {
|
||||
// 流中失败:已发生的 token 照实结算,其余退款
|
||||
s.quotaSettle(req.UserID, reserved, res.Attempt.PromptTokens+res.Attempt.CompletionTokens)
|
||||
}
|
||||
return
|
||||
}
|
||||
@@ -98,9 +137,55 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
|
||||
}
|
||||
lastErr = res.Err
|
||||
}
|
||||
// 三段式之三:全部候选失败 → 全额退款
|
||||
_ = s.quotaRefund(req.UserID, reserved)
|
||||
writeError(w, http.StatusServiceUnavailable, "HJ-2001", "全部候选端点失败:"+errString(lastErr))
|
||||
}
|
||||
|
||||
// candidates 候选序列:BYOK 独占(不回落平台端点);否则注册表候选。
|
||||
func (s *Server) candidates(req *ir.ChatRequest) ([]route.Endpoint, error) {
|
||||
if req.ByokKeyID == "" {
|
||||
return s.routes.Candidates(req.Slot, req.RunID)
|
||||
}
|
||||
if s.cfg.Vault == nil {
|
||||
return nil, fmt.Errorf("HJ-2003 BYOK 功能未启用(SHENSHI_BYOK_MASTER_KEY 未配置)")
|
||||
}
|
||||
key, plain, err := s.cfg.Vault.Resolve(req.UserID, req.ByokKeyID)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
spec := route.EndpointSpec{Type: key.Type, Name: "byok:" + key.ID, BaseURL: key.BaseURL, APIKey: plain, Model: key.Model}
|
||||
adaptor, ok := route.BuildAdaptor(spec)
|
||||
if !ok {
|
||||
return nil, fmt.Errorf("HJ-2003 BYOK 密钥类型不受支持: %s", key.Type)
|
||||
}
|
||||
return []route.Endpoint{{
|
||||
Name: spec.Name, Type: key.Type, Adaptor: adaptor,
|
||||
Priority: -100, CredentialKey: spec.Name, UpstreamModel: key.Model,
|
||||
}}, nil
|
||||
}
|
||||
|
||||
func (s *Server) quotaReserve(userID string, tokens int64) (int64, error) {
|
||||
if s.cfg.Quota == nil {
|
||||
return 0, nil
|
||||
}
|
||||
return s.cfg.Quota.Reserve(userID, tokens)
|
||||
}
|
||||
|
||||
func (s *Server) quotaSettle(userID string, reserved int64, actual int) error {
|
||||
if s.cfg.Quota == nil {
|
||||
return nil
|
||||
}
|
||||
return s.cfg.Quota.Settle(userID, reserved, int64(actual))
|
||||
}
|
||||
|
||||
func (s *Server) quotaRefund(userID string, reserved int64) error {
|
||||
if s.cfg.Quota == nil {
|
||||
return nil
|
||||
}
|
||||
return s.cfg.Quota.Refund(userID, reserved)
|
||||
}
|
||||
|
||||
type attempt struct {
|
||||
metering.Attempt
|
||||
Committed bool // 是否已向客户端写出 SSE
|
||||
@@ -116,7 +201,7 @@ func errString(err error) string {
|
||||
}
|
||||
|
||||
// tryEndpoint 单端点尝试。
|
||||
func (s *Server) tryEndpoint(w http.ResponseWriter, r *http.Request, endpoint route.Endpoint, req *ir.ChatRequest) attempt {
|
||||
func (s *Server) tryEndpoint(w http.ResponseWriter, r *http.Request, endpoint route.Endpoint, req *ir.ChatRequest, seq int) attempt {
|
||||
start := time.Now()
|
||||
ctx, cancel := context.WithTimeout(r.Context(), s.routes.FirstByteTimeout)
|
||||
defer cancel()
|
||||
@@ -129,7 +214,8 @@ func (s *Server) tryEndpoint(w http.ResponseWriter, r *http.Request, endpoint ro
|
||||
}()
|
||||
|
||||
res := attempt{Attempt: metering.Attempt{
|
||||
RunID: req.RunID, CharacterID: req.CharacterID,
|
||||
ID: fmt.Sprintf("%s#%s#%d", req.RunID, endpoint.Name, seq),
|
||||
RunID: req.RunID, CharacterID: req.CharacterID, UserID: req.UserID,
|
||||
Model: req.Model, Slot: req.Slot, Endpoint: endpoint.Name,
|
||||
}}
|
||||
|
||||
@@ -195,10 +281,112 @@ func failStatus(err error) string {
|
||||
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) {
|
||||
// meUser 从代理头取用户标识(灵台鉴权后注入,服务令牌即信任边界)。
|
||||
func meUser(r *http.Request) string { return r.Header.Get("X-Hunjing-User-Id") }
|
||||
|
||||
// handleUsage 用户用量聚合(/me/usage;附配额余额)。
|
||||
func (s *Server) handleUsage(w http.ResponseWriter, r *http.Request) {
|
||||
userID := meUser(r)
|
||||
rows, err := s.usage.Aggregate(userID)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusInternalServerError, "HJ-2001", err.Error())
|
||||
return
|
||||
}
|
||||
resp := map[string]any{"rows": rows}
|
||||
if s.cfg.Quota != nil {
|
||||
budget, used, reserved, err := s.cfg.Quota.Budget(userID)
|
||||
if err == nil {
|
||||
resp["quota"] = map[string]int64{"budget": budget, "used": used, "reserved": reserved}
|
||||
}
|
||||
}
|
||||
w.Header().Set("Content-Type", "application/json")
|
||||
_ = json.NewEncoder(w).Encode(s.usage.Aggregate())
|
||||
_ = json.NewEncoder(w).Encode(resp)
|
||||
}
|
||||
|
||||
// handleUsageSample 对账抽样(/admin/usage/sample?limit=N;最近 N 条原始记录)。
|
||||
func (s *Server) handleUsageSample(w http.ResponseWriter, r *http.Request) {
|
||||
limit := 50
|
||||
if v := r.URL.Query().Get("limit"); v != "" {
|
||||
if _, err := fmt.Sscanf(v, "%d", &limit); err != nil || limit <= 0 || limit > 1000 {
|
||||
writeError(w, http.StatusBadRequest, "HJ-2002", "limit 须为 1..1000")
|
||||
return
|
||||
}
|
||||
}
|
||||
rows, err := s.usage.Sample(limit)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusInternalServerError, "HJ-2001", err.Error())
|
||||
return
|
||||
}
|
||||
w.Header().Set("Content-Type", "application/json")
|
||||
_ = json.NewEncoder(w).Encode(map[string]any{"rows": rows})
|
||||
}
|
||||
|
||||
func (s *Server) handleByokList(w http.ResponseWriter, r *http.Request) {
|
||||
if s.cfg.Vault == nil {
|
||||
writeError(w, http.StatusServiceUnavailable, "HJ-2003", "BYOK 功能未启用(SHENSHI_BYOK_MASTER_KEY 未配置)")
|
||||
return
|
||||
}
|
||||
keys, err := s.cfg.Vault.List(meUser(r))
|
||||
if err != nil {
|
||||
writeError(w, http.StatusInternalServerError, "HJ-2001", err.Error())
|
||||
return
|
||||
}
|
||||
w.Header().Set("Content-Type", "application/json")
|
||||
_ = json.NewEncoder(w).Encode(map[string]any{"keys": keys})
|
||||
}
|
||||
|
||||
type byokRegisterRequest struct {
|
||||
Name string `json:"name"`
|
||||
Type string `json:"type"` // openai_compatible / anthropic / gemini
|
||||
BaseURL string `json:"base_url"`
|
||||
APIKey string `json:"api_key"`
|
||||
Model string `json:"model"`
|
||||
}
|
||||
|
||||
func (s *Server) handleByokRegister(w http.ResponseWriter, r *http.Request) {
|
||||
if s.cfg.Vault == nil {
|
||||
writeError(w, http.StatusServiceUnavailable, "HJ-2003", "BYOK 功能未启用(SHENSHI_BYOK_MASTER_KEY 未配置)")
|
||||
return
|
||||
}
|
||||
var body byokRegisterRequest
|
||||
if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
|
||||
writeError(w, http.StatusBadRequest, "HJ-2002", "请求解析失败: "+err.Error())
|
||||
return
|
||||
}
|
||||
switch body.Type {
|
||||
case "openai_compatible", "anthropic", "gemini":
|
||||
default:
|
||||
writeError(w, http.StatusBadRequest, "HJ-2002", "type 须为 openai_compatible / anthropic / gemini")
|
||||
return
|
||||
}
|
||||
if body.APIKey == "" || body.BaseURL == "" || body.Model == "" {
|
||||
writeError(w, http.StatusBadRequest, "HJ-2002", "api_key / base_url / model 不能为空")
|
||||
return
|
||||
}
|
||||
key, err := s.cfg.Vault.Register(meUser(r), body.Name, body.Type, body.BaseURL, body.APIKey, body.Model)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, "HJ-2003", err.Error())
|
||||
return
|
||||
}
|
||||
w.Header().Set("Content-Type", "application/json")
|
||||
w.WriteHeader(http.StatusCreated)
|
||||
_ = json.NewEncoder(w).Encode(key)
|
||||
}
|
||||
|
||||
func (s *Server) handleByokRevoke(w http.ResponseWriter, r *http.Request) {
|
||||
if s.cfg.Vault == nil {
|
||||
writeError(w, http.StatusServiceUnavailable, "HJ-2003", "BYOK 功能未启用(SHENSHI_BYOK_MASTER_KEY 未配置)")
|
||||
return
|
||||
}
|
||||
if err := s.cfg.Vault.Revoke(meUser(r), r.PathValue("id")); err != nil {
|
||||
if errors.Is(err, store.ErrByokNotFound) {
|
||||
writeError(w, http.StatusNotFound, "HJ-2003", err.Error())
|
||||
return
|
||||
}
|
||||
writeError(w, http.StatusInternalServerError, "HJ-2001", err.Error())
|
||||
return
|
||||
}
|
||||
w.WriteHeader(http.StatusNoContent)
|
||||
}
|
||||
|
||||
// handleHealth 健康面。
|
||||
|
||||
@@ -0,0 +1,163 @@
|
||||
// PG 实现:与内存实现同语义的直写 SQL。
|
||||
package store
|
||||
|
||||
import (
|
||||
"database/sql"
|
||||
"time"
|
||||
)
|
||||
|
||||
type pgStore struct {
|
||||
db *sql.DB
|
||||
}
|
||||
|
||||
func (p *pgStore) InsertUsage(l UsageLog) error {
|
||||
_, err := p.db.Exec(
|
||||
`INSERT INTO usage_logs (id, run_id, character_id, user_id, model, slot, endpoint,
|
||||
status, error_code, first_token_ms, total_ms, prompt_tokens, completion_tokens, at)
|
||||
VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14)`,
|
||||
l.ID, l.RunID, l.CharacterID, l.UserID, l.Model, l.Slot, l.Endpoint,
|
||||
l.Status, l.ErrorCode, l.FirstTokenMS, l.TotalMS, l.PromptTokens, l.CompletionTokens, l.At,
|
||||
)
|
||||
return err
|
||||
}
|
||||
|
||||
func (p *pgStore) AggregateUsage(userID string) ([]UsageAgg, error) {
|
||||
q := `SELECT user_id, model, endpoint, COUNT(*),
|
||||
COUNT(*) FILTER (WHERE status = 'ok'),
|
||||
COALESCE(SUM(prompt_tokens) FILTER (WHERE status = 'ok'), 0),
|
||||
COALESCE(SUM(completion_tokens) FILTER (WHERE status = 'ok'), 0)
|
||||
FROM usage_logs`
|
||||
args := []any{}
|
||||
if userID != "" {
|
||||
q += ` WHERE user_id = $1`
|
||||
args = append(args, userID)
|
||||
}
|
||||
q += ` GROUP BY user_id, model, endpoint ORDER BY user_id, model`
|
||||
rows, err := p.db.Query(q, args...)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
defer func() { _ = rows.Close() }()
|
||||
var out []UsageAgg
|
||||
for rows.Next() {
|
||||
var r UsageAgg
|
||||
if err := rows.Scan(&r.UserID, &r.Model, &r.Endpoint, &r.Requests, &r.OK, &r.PromptTokens, &r.CompletionTokens); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
out = append(out, r)
|
||||
}
|
||||
return out, rows.Err()
|
||||
}
|
||||
|
||||
func (p *pgStore) SampleUsage(limit int) ([]UsageLog, error) {
|
||||
rows, err := p.db.Query(
|
||||
`SELECT id, run_id, character_id, user_id, model, slot, endpoint, status, error_code,
|
||||
first_token_ms, total_ms, prompt_tokens, completion_tokens, at
|
||||
FROM usage_logs ORDER BY at DESC LIMIT $1`, limit)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
defer func() { _ = rows.Close() }()
|
||||
var out []UsageLog
|
||||
for rows.Next() {
|
||||
var l UsageLog
|
||||
if err := rows.Scan(&l.ID, &l.RunID, &l.CharacterID, &l.UserID, &l.Model, &l.Slot, &l.Endpoint,
|
||||
&l.Status, &l.ErrorCode, &l.FirstTokenMS, &l.TotalMS, &l.PromptTokens, &l.CompletionTokens, &l.At); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
out = append(out, l)
|
||||
}
|
||||
return out, rows.Err()
|
||||
}
|
||||
|
||||
func (p *pgStore) UpsertByok(k ByokKey) error {
|
||||
_, err := p.db.Exec(
|
||||
`INSERT INTO byok_keys (id, user_id, name, type, base_url, api_key, model, created_at)
|
||||
VALUES ($1,$2,$3,$4,$5,$6,$7,$8)
|
||||
ON CONFLICT (id) DO UPDATE SET name=$3, type=$4, base_url=$5, api_key=$6, model=$7`,
|
||||
k.ID, k.UserID, k.Name, k.Type, k.BaseURL, k.APIKey, k.Model, k.CreatedAt)
|
||||
return err
|
||||
}
|
||||
|
||||
func (p *pgStore) ListByok(userID string) ([]ByokKey, error) {
|
||||
return p.byokQuery(`SELECT id, user_id, name, type, base_url, api_key, model, created_at, last_used, revoked_at
|
||||
FROM byok_keys WHERE user_id = $1 AND revoked_at IS NULL ORDER BY created_at`, userID)
|
||||
}
|
||||
|
||||
func (p *pgStore) GetByok(userID, id string) (*ByokKey, error) {
|
||||
keys, err := p.byokQuery(`SELECT id, user_id, name, type, base_url, api_key, model, created_at, last_used, revoked_at
|
||||
FROM byok_keys WHERE user_id = $1 AND id = $2 AND revoked_at IS NULL`, userID, id)
|
||||
if err != nil || len(keys) == 0 {
|
||||
return nil, err
|
||||
}
|
||||
return &keys[0], nil
|
||||
}
|
||||
|
||||
func (p *pgStore) byokQuery(q string, args ...any) ([]ByokKey, error) {
|
||||
rows, err := p.db.Query(q, args...)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
defer func() { _ = rows.Close() }()
|
||||
var out []ByokKey
|
||||
for rows.Next() {
|
||||
var k ByokKey
|
||||
if err := rows.Scan(&k.ID, &k.UserID, &k.Name, &k.Type, &k.BaseURL, &k.APIKey, &k.Model, &k.CreatedAt, &k.LastUsed, &k.RevokedAt); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
out = append(out, k)
|
||||
}
|
||||
return out, rows.Err()
|
||||
}
|
||||
|
||||
func (p *pgStore) RevokeByok(userID, id string) error {
|
||||
res, err := p.db.Exec(`UPDATE byok_keys SET revoked_at = $1 WHERE user_id = $2 AND id = $3 AND revoked_at IS NULL`,
|
||||
time.Now().UTC(), userID, id)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
if n, _ := res.RowsAffected(); n == 0 {
|
||||
return ErrByokNotFound
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func (p *pgStore) TouchByok(id string, at time.Time) error {
|
||||
_, err := p.db.Exec(`UPDATE byok_keys SET last_used = $1 WHERE id = $2`, at, id)
|
||||
return err
|
||||
}
|
||||
|
||||
func (p *pgStore) Reserve(userID string, amount int64) error {
|
||||
if amount <= 0 {
|
||||
return nil
|
||||
}
|
||||
// UPSERT + 原子累加;配额上限裁决在 quota 包(集中单一函数)
|
||||
_, err := p.db.Exec(`INSERT INTO quota_balances (user_id, reserved, used) VALUES ($1, $2, 0)
|
||||
ON CONFLICT (user_id) DO UPDATE SET reserved = quota_balances.reserved + $2`, userID, amount)
|
||||
return err
|
||||
}
|
||||
|
||||
func (p *pgStore) Settle(userID string, reserved, actual int64) error {
|
||||
// 饱和截断在 SQL 层以 GREATEST 保证(防负)
|
||||
_, err := p.db.Exec(`UPDATE quota_balances SET
|
||||
reserved = GREATEST(reserved - $2, 0), used = used + $3 WHERE user_id = $1`,
|
||||
userID, reserved, actual)
|
||||
return err
|
||||
}
|
||||
|
||||
func (p *pgStore) Refund(userID string, reserved int64) error {
|
||||
_, err := p.db.Exec(`UPDATE quota_balances SET reserved = GREATEST(reserved - $2, 0) WHERE user_id = $1`,
|
||||
userID, reserved)
|
||||
return err
|
||||
}
|
||||
|
||||
func (p *pgStore) Balance(userID string) (QuotaBalance, error) {
|
||||
var b QuotaBalance
|
||||
b.UserID = userID
|
||||
err := p.db.QueryRow(`SELECT reserved, used FROM quota_balances WHERE user_id = $1`, userID).
|
||||
Scan(&b.Reserved, &b.Used)
|
||||
if err == sql.ErrNoRows {
|
||||
return b, nil
|
||||
}
|
||||
return b, err
|
||||
}
|
||||
@@ -0,0 +1,343 @@
|
||||
// Package store 神识自有库访问层:usage_logs / byok_keys / quota 余额。
|
||||
//
|
||||
// 双实现:PG(SHENSHI_DATABASE_URL,生产)与内存(开发/测试,未配置时回退)。
|
||||
// 语义契约由 store_test.go 在内存实现上钉死;PG 实现为同接口直写 SQL。
|
||||
package store
|
||||
|
||||
import (
|
||||
"database/sql"
|
||||
"errors"
|
||||
"fmt"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
// PG 驱动(database/sql 标准接口)
|
||||
_ "github.com/jackc/pgx/v5/stdlib"
|
||||
)
|
||||
|
||||
// UsageLog 一次端点尝试的完整计量记录(shenshi-gateway.md §3 usage_logs)。
|
||||
type UsageLog struct {
|
||||
ID string // 尝试唯一标识(run_id#endpoint#seq,配额结算幂等键)
|
||||
RunID string
|
||||
CharacterID string
|
||||
UserID string
|
||||
Model string
|
||||
Slot string
|
||||
Endpoint string
|
||||
Status string // ok / error / first_byte_timeout
|
||||
ErrorCode string
|
||||
FirstTokenMS int64
|
||||
TotalMS int64
|
||||
PromptTokens int
|
||||
CompletionTokens int
|
||||
At time.Time
|
||||
}
|
||||
|
||||
// UsageAgg 用量聚合行(/me/usage)。
|
||||
type UsageAgg struct {
|
||||
UserID string `json:"user_id"`
|
||||
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"`
|
||||
}
|
||||
|
||||
// ByokKey 用户自带密钥(API 永远脱敏返回)。
|
||||
type ByokKey struct {
|
||||
ID string
|
||||
UserID string
|
||||
Name string
|
||||
Type string // openai_compatible / anthropic / gemini(契约词典 §14.2)
|
||||
BaseURL string
|
||||
APIKey string // 加密存储;出站时解密
|
||||
Model string
|
||||
CreatedAt time.Time
|
||||
LastUsed *time.Time
|
||||
RevokedAt *time.Time
|
||||
}
|
||||
|
||||
// QuotaBalance 用户配额账(三段式状态:Reserved 预扣 / Used 已结算)。
|
||||
type QuotaBalance struct {
|
||||
UserID string
|
||||
Reserved int64
|
||||
Used int64
|
||||
}
|
||||
|
||||
// ErrByokNotFound BYOK 密钥不存在或已撤回。
|
||||
var ErrByokNotFound = errors.New("BYOK 密钥不存在或已撤回")
|
||||
|
||||
// Store 存储接口(全异步语义由调用方保证;实现自身须并发安全)。
|
||||
type Store interface {
|
||||
InsertUsage(log UsageLog) error
|
||||
AggregateUsage(userID string) ([]UsageAgg, error)
|
||||
SampleUsage(limit int) ([]UsageLog, error) // 对账抽样:最近 N 条原始记录
|
||||
|
||||
UpsertByok(key ByokKey) error
|
||||
ListByok(userID string) ([]ByokKey, error)
|
||||
GetByok(userID, id string) (*ByokKey, error)
|
||||
RevokeByok(userID, id string) error
|
||||
TouchByok(id string, at time.Time) error
|
||||
|
||||
// Quota 三段式原子操作:Reserve 预扣(不足返回错误);Settle 结算差值;Refund 幂等退款。
|
||||
Reserve(userID string, amount int64) error
|
||||
Settle(userID string, reserved, actual int64) error
|
||||
Refund(userID string, reserved int64) error
|
||||
Balance(userID string) (QuotaBalance, error)
|
||||
}
|
||||
|
||||
// Open 按配置打开存储:DSN 非空时 PG(自动建表,含就绪重试),否则内存实现。
|
||||
// 重试窗口 60 秒:容器编排下 PG 可能晚于神识就绪(跨 compose 文件无 depends_on)。
|
||||
func Open(dsn string) (Store, error) {
|
||||
if dsn == "" {
|
||||
return NewMem(), nil
|
||||
}
|
||||
db, err := sql.Open("pgx", dsn)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("打开神识库失败: %w", err)
|
||||
}
|
||||
deadline := time.Now().Add(60 * time.Second)
|
||||
for {
|
||||
err = migrate(db)
|
||||
if err == nil {
|
||||
return &pgStore{db: db}, nil
|
||||
}
|
||||
if time.Now().After(deadline) {
|
||||
return nil, fmt.Errorf("神识库迁移失败(60 秒重试后放弃): %w", err)
|
||||
}
|
||||
time.Sleep(2 * time.Second)
|
||||
}
|
||||
}
|
||||
|
||||
func migrate(db *sql.DB) error {
|
||||
stmts := []string{
|
||||
`CREATE TABLE IF NOT EXISTS usage_logs (
|
||||
id TEXT PRIMARY KEY,
|
||||
run_id TEXT NOT NULL,
|
||||
character_id TEXT NOT NULL DEFAULT '',
|
||||
user_id TEXT NOT NULL DEFAULT '',
|
||||
model TEXT NOT NULL,
|
||||
slot TEXT NOT NULL,
|
||||
endpoint TEXT NOT NULL,
|
||||
status TEXT NOT NULL,
|
||||
error_code TEXT NOT NULL DEFAULT '',
|
||||
first_token_ms BIGINT NOT NULL DEFAULT 0,
|
||||
total_ms BIGINT NOT NULL DEFAULT 0,
|
||||
prompt_tokens INT NOT NULL DEFAULT 0,
|
||||
completion_tokens INT NOT NULL DEFAULT 0,
|
||||
at TIMESTAMPTZ NOT NULL
|
||||
)`,
|
||||
`CREATE INDEX IF NOT EXISTS idx_usage_logs_user ON usage_logs(user_id, at DESC)`,
|
||||
`CREATE TABLE IF NOT EXISTS byok_keys (
|
||||
id TEXT PRIMARY KEY,
|
||||
user_id TEXT NOT NULL,
|
||||
name TEXT NOT NULL,
|
||||
type TEXT NOT NULL,
|
||||
base_url TEXT NOT NULL DEFAULT '',
|
||||
api_key TEXT NOT NULL,
|
||||
model TEXT NOT NULL DEFAULT '',
|
||||
created_at TIMESTAMPTZ NOT NULL,
|
||||
last_used TIMESTAMPTZ,
|
||||
revoked_at TIMESTAMPTZ
|
||||
)`,
|
||||
`CREATE INDEX IF NOT EXISTS idx_byok_user ON byok_keys(user_id)`,
|
||||
`CREATE TABLE IF NOT EXISTS quota_balances (
|
||||
user_id TEXT PRIMARY KEY,
|
||||
reserved BIGINT NOT NULL DEFAULT 0,
|
||||
used BIGINT NOT NULL DEFAULT 0
|
||||
)`,
|
||||
}
|
||||
for _, s := range stmts {
|
||||
if _, err := db.Exec(s); err != nil {
|
||||
return fmt.Errorf("建表失败: %w", err)
|
||||
}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
// ── 内存实现 ──────────────────────────────────────────────
|
||||
|
||||
type memStore struct {
|
||||
mu sync.Mutex
|
||||
usage []UsageLog
|
||||
byok map[string][]ByokKey // user → keys
|
||||
quota map[string]*QuotaBalance
|
||||
}
|
||||
|
||||
// NewMem 内存存储(保留最近 10 万条 usage,防无界增长)。
|
||||
func NewMem() Store {
|
||||
return &memStore{byok: map[string][]ByokKey{}, quota: map[string]*QuotaBalance{}}
|
||||
}
|
||||
|
||||
const memUsageLimit = 100_000
|
||||
|
||||
func (m *memStore) InsertUsage(log UsageLog) error {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
if len(m.usage) >= memUsageLimit {
|
||||
m.usage = m.usage[1:]
|
||||
}
|
||||
m.usage = append(m.usage, log)
|
||||
return nil
|
||||
}
|
||||
|
||||
func (m *memStore) AggregateUsage(userID string) ([]UsageAgg, error) {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
agg := map[string]*UsageAgg{}
|
||||
for _, e := range m.usage {
|
||||
if userID != "" && e.UserID != userID {
|
||||
continue
|
||||
}
|
||||
key := e.UserID + "|" + e.Model + "|" + e.Endpoint
|
||||
row := agg[key]
|
||||
if row == nil {
|
||||
row = &UsageAgg{UserID: e.UserID, 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, nil
|
||||
}
|
||||
|
||||
func (m *memStore) SampleUsage(limit int) ([]UsageLog, error) {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
n := limit
|
||||
if n > len(m.usage) {
|
||||
n = len(m.usage)
|
||||
}
|
||||
out := make([]UsageLog, n)
|
||||
copy(out, m.usage[len(m.usage)-n:])
|
||||
return out, nil
|
||||
}
|
||||
|
||||
func (m *memStore) UpsertByok(key ByokKey) error {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
keys := m.byok[key.UserID]
|
||||
for i, k := range keys {
|
||||
if k.ID == key.ID {
|
||||
keys[i] = key
|
||||
return nil
|
||||
}
|
||||
}
|
||||
m.byok[key.UserID] = append(keys, key)
|
||||
return nil
|
||||
}
|
||||
|
||||
func (m *memStore) ListByok(userID string) ([]ByokKey, error) {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
out := make([]ByokKey, 0, len(m.byok[userID]))
|
||||
for _, k := range m.byok[userID] {
|
||||
if k.RevokedAt == nil {
|
||||
out = append(out, k)
|
||||
}
|
||||
}
|
||||
return out, nil
|
||||
}
|
||||
|
||||
func (m *memStore) GetByok(userID, id string) (*ByokKey, error) {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
for _, k := range m.byok[userID] {
|
||||
if k.ID == id && k.RevokedAt == nil {
|
||||
out := k
|
||||
return &out, nil
|
||||
}
|
||||
}
|
||||
return nil, nil
|
||||
}
|
||||
|
||||
func (m *memStore) RevokeByok(userID, id string) error {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
now := time.Now().UTC()
|
||||
keys := m.byok[userID]
|
||||
for i, k := range keys {
|
||||
if k.ID == id && k.RevokedAt == nil {
|
||||
keys[i].RevokedAt = &now
|
||||
return nil
|
||||
}
|
||||
}
|
||||
return ErrByokNotFound
|
||||
}
|
||||
|
||||
func (m *memStore) TouchByok(id string, at time.Time) error {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
for user, keys := range m.byok {
|
||||
for i, k := range keys {
|
||||
if k.ID == id {
|
||||
used := at
|
||||
keys[i].LastUsed = &used
|
||||
m.byok[user] = keys
|
||||
return nil
|
||||
}
|
||||
}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func (m *memStore) Reserve(userID string, amount int64) error {
|
||||
if amount <= 0 {
|
||||
return nil
|
||||
}
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
b := m.quotaGet(userID)
|
||||
b.Reserved += amount
|
||||
return nil
|
||||
}
|
||||
|
||||
func (m *memStore) Settle(userID string, reserved, actual int64) error {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
b := m.quotaGet(userID)
|
||||
b.Reserved -= reserved
|
||||
if b.Reserved < 0 {
|
||||
b.Reserved = 0
|
||||
}
|
||||
b.Used += actual
|
||||
return nil
|
||||
}
|
||||
|
||||
func (m *memStore) Refund(userID string, reserved int64) error {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
b := m.quotaGet(userID)
|
||||
b.Reserved -= reserved
|
||||
if b.Reserved < 0 {
|
||||
b.Reserved = 0
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func (m *memStore) Balance(userID string) (QuotaBalance, error) {
|
||||
m.mu.Lock()
|
||||
defer m.mu.Unlock()
|
||||
if b, ok := m.quota[userID]; ok {
|
||||
return *b, nil
|
||||
}
|
||||
return QuotaBalance{UserID: userID}, nil
|
||||
}
|
||||
|
||||
func (m *memStore) quotaGet(userID string) *QuotaBalance {
|
||||
b, ok := m.quota[userID]
|
||||
if !ok {
|
||||
b = &QuotaBalance{UserID: userID}
|
||||
m.quota[userID] = b
|
||||
}
|
||||
return b
|
||||
}
|
||||
@@ -0,0 +1,105 @@
|
||||
// store 语义契约测试(内存实现;PG 实现同接口同语义)。
|
||||
package store
|
||||
|
||||
import (
|
||||
"testing"
|
||||
"time"
|
||||
)
|
||||
|
||||
func newUsage(id, user, model, endpoint, status string, p, c int) UsageLog {
|
||||
return UsageLog{ID: id, RunID: "run_t", UserID: user, Model: model, Slot: "chat",
|
||||
Endpoint: endpoint, Status: status, PromptTokens: p, CompletionTokens: c,
|
||||
At: time.Now().UTC()}
|
||||
}
|
||||
|
||||
func TestUsageAggregateFiltersByUserAndCountsOK(t *testing.T) {
|
||||
st := NewMem()
|
||||
_ = st.InsertUsage(newUsage("1", "u1", "m", "e1", "ok", 10, 20))
|
||||
_ = st.InsertUsage(newUsage("2", "u1", "m", "e1", "error", 0, 0))
|
||||
_ = st.InsertUsage(newUsage("3", "u2", "m", "e2", "ok", 1, 2))
|
||||
|
||||
rows, err := st.AggregateUsage("u1")
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if len(rows) != 1 {
|
||||
t.Fatalf("u1 聚合行数 = %d, 期望 1", len(rows))
|
||||
}
|
||||
r := rows[0]
|
||||
if r.Requests != 2 || r.OK != 1 || r.PromptTokens != 10 || r.CompletionTokens != 20 {
|
||||
t.Fatalf("聚合字段错误: %+v", r)
|
||||
}
|
||||
|
||||
all, _ := st.AggregateUsage("")
|
||||
if len(all) != 2 {
|
||||
t.Fatalf("全体聚合行数 = %d, 期望 2", len(all))
|
||||
}
|
||||
}
|
||||
|
||||
func TestUsageSampleReturnsLatest(t *testing.T) {
|
||||
st := NewMem()
|
||||
for i := 0; i < 5; i++ {
|
||||
_ = st.InsertUsage(newUsage(string(rune('a'+i)), "u1", "m", "e", "ok", i, 0))
|
||||
}
|
||||
rows, _ := st.SampleUsage(2)
|
||||
if len(rows) != 2 || rows[1].ID != "e" {
|
||||
t.Fatalf("抽样应返回最新 2 条(d,e),实际 %+v", rows)
|
||||
}
|
||||
}
|
||||
|
||||
func TestByokLifecycle(t *testing.T) {
|
||||
st := NewMem()
|
||||
key := ByokKey{ID: "byok_x", UserID: "u1", Name: "主力", Type: "openai_compatible",
|
||||
BaseURL: "https://api.example.com", APIKey: "enc", Model: "m1", CreatedAt: time.Now().UTC()}
|
||||
if err := st.UpsertByok(key); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
got, err := st.GetByok("u1", "byok_x")
|
||||
if err != nil || got == nil || got.Model != "m1" {
|
||||
t.Fatalf("GetByok 失败: %v %+v", err, got)
|
||||
}
|
||||
// 跨用户隔离
|
||||
if other, _ := st.GetByok("u2", "byok_x"); other != nil {
|
||||
t.Fatal("u2 不应读到 u1 的密钥")
|
||||
}
|
||||
if err := st.RevokeByok("u1", "byok_x"); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
got2, _ := st.GetByok("u1", "byok_x")
|
||||
if got2 != nil {
|
||||
t.Fatal("撤回后不应可读")
|
||||
}
|
||||
if err := st.RevokeByok("u1", "byok_x"); err != ErrByokNotFound {
|
||||
t.Fatalf("重复撤回应返回 ErrByokNotFound,实际 %v", err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestQuotaSaturatingRefund(t *testing.T) {
|
||||
st := NewMem()
|
||||
if err := st.Reserve("u1", 100); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
// 幂等退款:重复退款不产生负余额
|
||||
if err := st.Refund("u1", 100); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := st.Refund("u1", 100); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
b, _ := st.Balance("u1")
|
||||
if b.Reserved < 0 || b.Used != 0 {
|
||||
t.Fatalf("重复退款后余额为负: %+v", b)
|
||||
}
|
||||
}
|
||||
|
||||
func TestQuotaSettleReleasesDifference(t *testing.T) {
|
||||
st := NewMem()
|
||||
_ = st.Reserve("u1", 100)
|
||||
if err := st.Settle("u1", 100, 37); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
b, _ := st.Balance("u1")
|
||||
if b.Reserved != 0 || b.Used != 37 {
|
||||
t.Fatalf("结算后应释放全部预扣并记实际用量: %+v", b)
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
package transform
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"net/http"
|
||||
"net/http/httptest"
|
||||
"testing"
|
||||
|
||||
"hunjing/shenshi/internal/ir"
|
||||
)
|
||||
|
||||
func readJSON(r *http.Request, v any) error {
|
||||
defer func() { _ = r.Body.Close() }()
|
||||
return json.NewDecoder(r.Body).Decode(v)
|
||||
}
|
||||
|
||||
// 收集全量增量(适配器测试通用)。
|
||||
func collect(t *testing.T, ad Adaptor, req *ir.ChatRequest) []ir.Delta {
|
||||
t.Helper()
|
||||
out := make(chan ir.Delta, 64)
|
||||
done := make(chan error, 1)
|
||||
go func() {
|
||||
defer close(out)
|
||||
done <- ad.Stream(context.Background(), req, out)
|
||||
}()
|
||||
var deltas []ir.Delta
|
||||
for d := range out {
|
||||
deltas = append(deltas, d)
|
||||
}
|
||||
if err := <-done; err != nil {
|
||||
t.Fatalf("Stream 出错: %v", err)
|
||||
}
|
||||
return deltas
|
||||
}
|
||||
|
||||
func TestAnthropicAdaptorRequestAndStream(t *testing.T) {
|
||||
var gotBody antRequest
|
||||
var gotAuth, gotVersion bool
|
||||
srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
|
||||
if r.URL.Path != "/v1/messages" {
|
||||
t.Errorf("路径错误: %s", r.URL.Path)
|
||||
}
|
||||
gotAuth = r.Header.Get("x-api-key") == "sk-test"
|
||||
gotVersion = r.Header.Get("anthropic-version") != ""
|
||||
_ = readJSON(r, &gotBody)
|
||||
w.Header().Set("content-type", "text/event-stream")
|
||||
_, _ = w.Write([]byte("data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"input_tokens\":11,\"output_tokens\":1}}}\n\n"))
|
||||
_, _ = w.Write([]byte("data: {\"type\":\"content_block_delta\",\"delta\":{\"type\":\"text_delta\",\"text\":\"你好\"}}\n\n"))
|
||||
_, _ = w.Write([]byte("data: {\"type\":\"content_block_delta\",\"delta\":{\"type\":\"text_delta\",\"text\":\",镜中\"}}\n\n"))
|
||||
_, _ = w.Write([]byte("data: {\"type\":\"message_delta\",\"usage\":{\"output_tokens\":7}}\n\n"))
|
||||
}))
|
||||
defer srv.Close()
|
||||
|
||||
ad := &AnthropicAdaptor{EndpointName: "ant", BaseURL: srv.URL, APIKey: "sk-test", UpstreamModel: "claude-x"}
|
||||
req := &ir.ChatRequest{
|
||||
Messages: []ir.Message{
|
||||
{Role: "system", Content: "系统契约"},
|
||||
{Role: "user", Content: "照心"},
|
||||
},
|
||||
RunID: "run_ant", Stream: true,
|
||||
}
|
||||
deltas := collect(t, ad, req)
|
||||
|
||||
if !gotAuth || !gotVersion {
|
||||
t.Fatal("缺少 x-api-key / anthropic-version 头")
|
||||
}
|
||||
if gotBody.System == nil || *gotBody.System != "系统契约" {
|
||||
t.Fatalf("system 应提升为顶层字段: %+v", gotBody.System)
|
||||
}
|
||||
if gotBody.MaxTokens == nil {
|
||||
t.Fatal("Anthropic 要求 max_tokens 必填,适配器应兜底")
|
||||
}
|
||||
var text string
|
||||
var last ir.Delta
|
||||
for _, d := range deltas {
|
||||
text += d.Delta
|
||||
if d.FinishReason != "" {
|
||||
last = d
|
||||
}
|
||||
}
|
||||
if text != "你好,镜中" {
|
||||
t.Fatalf("增量拼接错误: %q", text)
|
||||
}
|
||||
if last.FinishReason != "stop" || last.UsagePromptTokens != 11 || last.UsageCompletionTokens != 7 {
|
||||
t.Fatalf("末帧用量错误: %+v", last)
|
||||
}
|
||||
}
|
||||
|
||||
func TestAnthropicAdaptorSurfaces429(t *testing.T) {
|
||||
srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
|
||||
w.WriteHeader(http.StatusTooManyRequests)
|
||||
_, _ = w.Write([]byte(`{"error":"rate_limited"}`))
|
||||
}))
|
||||
defer srv.Close()
|
||||
ad := &AnthropicAdaptor{BaseURL: srv.URL, APIKey: "k", UpstreamModel: "m"}
|
||||
out := make(chan ir.Delta, 4)
|
||||
err := ad.Stream(context.Background(), &ir.ChatRequest{RunID: "r", Stream: true, Messages: []ir.Message{{Role: "user", Content: "x"}}}, out)
|
||||
up, ok := err.(*UpstreamError)
|
||||
if !ok || up.StatusCode != 429 {
|
||||
t.Fatalf("应携带 429 UpstreamError 供绑定级冷却,实际 %v", err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestGeminiAdaptorRequestAndStream(t *testing.T) {
|
||||
var gotPath, gotKey string
|
||||
srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
|
||||
gotPath = r.URL.Path
|
||||
gotKey = r.Header.Get("x-goog-api-key")
|
||||
w.Header().Set("content-type", "text/event-stream")
|
||||
_, _ = w.Write([]byte("data: {\"candidates\":[{\"content\":{\"parts\":[{\"text\":\"第一\"}]}}]}\n\n"))
|
||||
_, _ = w.Write([]byte("data: {\"candidates\":[{\"content\":{\"parts\":[{\"text\":\"第二\"}]},\"finishReason\":\"STOP\"}],\"usageMetadata\":{\"promptTokenCount\":5,\"candidatesTokenCount\":9}}\n\n"))
|
||||
}))
|
||||
defer srv.Close()
|
||||
|
||||
ad := &GeminiAdaptor{EndpointName: "gem", BaseURL: srv.URL, APIKey: "AIza-test", UpstreamModel: "gemini-x"}
|
||||
req := &ir.ChatRequest{
|
||||
Messages: []ir.Message{
|
||||
{Role: "system", Content: "系统契约"},
|
||||
{Role: "user", Content: "问"},
|
||||
},
|
||||
RunID: "run_gem", Stream: true,
|
||||
}
|
||||
deltas := collect(t, ad, req)
|
||||
|
||||
if gotPath != "/v1beta/models/gemini-x:streamGenerateContent" {
|
||||
t.Fatalf("Gemini 路径错误: %s", gotPath)
|
||||
}
|
||||
if gotKey != "AIza-test" {
|
||||
t.Fatal("缺少 x-goog-api-key 头")
|
||||
}
|
||||
var text string
|
||||
var last ir.Delta
|
||||
for _, d := range deltas {
|
||||
text += d.Delta
|
||||
if d.FinishReason != "" {
|
||||
last = d
|
||||
}
|
||||
}
|
||||
if text != "第一第二" {
|
||||
t.Fatalf("增量拼接错误: %q", text)
|
||||
}
|
||||
if last.UsagePromptTokens != 5 || last.UsageCompletionTokens != 9 {
|
||||
t.Fatalf("用量解析错误: %+v", last)
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,168 @@
|
||||
// Anthropic 出站适配器:IR → Anthropic Messages API(SSE 事件流)。
|
||||
package transform
|
||||
|
||||
import (
|
||||
"bufio"
|
||||
"bytes"
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"net/http"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
"hunjing/shenshi/internal/ir"
|
||||
)
|
||||
|
||||
// AnthropicAdaptor Anthropic Messages 端点适配器。
|
||||
type AnthropicAdaptor struct {
|
||||
EndpointName string
|
||||
BaseURL string
|
||||
APIKey string
|
||||
UpstreamModel string
|
||||
Client *http.Client
|
||||
}
|
||||
|
||||
// Name 实现 Adaptor。
|
||||
func (a *AnthropicAdaptor) Name() string { return a.EndpointName }
|
||||
|
||||
type antMessage struct {
|
||||
Role string `json:"role"`
|
||||
Content string `json:"content"`
|
||||
}
|
||||
|
||||
// 可选标量一律指针 + omitempty(shenshi-gateway.md §5 DTO 纪律)。
|
||||
type antRequest struct {
|
||||
Model string `json:"model"`
|
||||
Messages []antMessage `json:"messages"`
|
||||
MaxTokens *int `json:"max_tokens,omitempty"`
|
||||
Temperature *float64 `json:"temperature,omitempty"`
|
||||
Stream bool `json:"stream"`
|
||||
System *string `json:"system,omitempty"`
|
||||
}
|
||||
|
||||
// antEvent Anthropic SSE 事件:message_start 的用量在 message.usage,
|
||||
// message_delta 的用量在顶层 usage(两处形态不同,指针区分可选标量)。
|
||||
type antEvent struct {
|
||||
Type string `json:"type"`
|
||||
Delta struct {
|
||||
Type string `json:"type"`
|
||||
Text string `json:"text"`
|
||||
} `json:"delta"`
|
||||
Message *struct {
|
||||
Usage *struct {
|
||||
InputTokens int `json:"input_tokens"`
|
||||
OutputTokens int `json:"output_tokens"`
|
||||
} `json:"usage"`
|
||||
} `json:"message"`
|
||||
Usage *struct {
|
||||
OutputTokens int `json:"output_tokens"`
|
||||
} `json:"usage"`
|
||||
}
|
||||
|
||||
// Stream 实现 Adaptor:Anthropic SSE(message_delta / content_block_delta 等)转 IR Delta。
|
||||
func (a *AnthropicAdaptor) Stream(ctx context.Context, req *ir.ChatRequest, out chan<- ir.Delta) error {
|
||||
body := antRequest{
|
||||
Model: a.UpstreamModel,
|
||||
Messages: make([]antMessage, 0, len(req.Messages)),
|
||||
MaxTokens: req.MaxTokens,
|
||||
Temperature: req.Temperature,
|
||||
Stream: true,
|
||||
}
|
||||
// Anthropic 的 system 是顶层字段而非消息序列首条
|
||||
var system strings.Builder
|
||||
for _, m := range req.Messages {
|
||||
if m.Role == "system" {
|
||||
if system.Len() > 0 {
|
||||
system.WriteString("\n\n")
|
||||
}
|
||||
system.WriteString(m.Content)
|
||||
continue
|
||||
}
|
||||
role := m.Role
|
||||
if role == "tool" {
|
||||
role = "user" // 最小映射:tool 结果按用户侧内容回传(工具循环随百工 M5 完整化)
|
||||
}
|
||||
body.Messages = append(body.Messages, antMessage{Role: role, Content: m.Content})
|
||||
}
|
||||
if system.Len() > 0 {
|
||||
s := system.String()
|
||||
body.System = &s
|
||||
}
|
||||
if body.MaxTokens == nil {
|
||||
def := 4096 // Anthropic 要求 max_tokens 必填
|
||||
body.MaxTokens = &def
|
||||
}
|
||||
|
||||
payload, err := json.Marshal(body)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost,
|
||||
strings.TrimRight(a.BaseURL, "/")+"/v1/messages", bytes.NewReader(payload))
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
httpReq.Header.Set("content-type", "application/json")
|
||||
httpReq.Header.Set("x-api-key", a.APIKey)
|
||||
httpReq.Header.Set("anthropic-version", "2023-06-01")
|
||||
|
||||
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()}
|
||||
}
|
||||
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()}
|
||||
}
|
||||
|
||||
var finish string
|
||||
var usagePrompt, usageCompletion int
|
||||
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 == "" {
|
||||
continue
|
||||
}
|
||||
var ev antEvent
|
||||
if err := json.Unmarshal([]byte(data), &ev); err != nil {
|
||||
continue
|
||||
}
|
||||
switch ev.Type {
|
||||
case "content_block_delta":
|
||||
if ev.Delta.Type == "text_delta" && ev.Delta.Text != "" {
|
||||
out <- ir.Delta{Delta: ev.Delta.Text}
|
||||
}
|
||||
case "message_delta":
|
||||
// 结束与用量在 message_delta 顶层 usage 汇总(output_tokens)
|
||||
if ev.Usage != nil {
|
||||
usageCompletion = ev.Usage.OutputTokens
|
||||
}
|
||||
finish = "stop"
|
||||
case "message_start":
|
||||
if ev.Message != nil && ev.Message.Usage != nil {
|
||||
usagePrompt = ev.Message.Usage.InputTokens
|
||||
}
|
||||
case "error":
|
||||
return fmt.Errorf("上游流内错误: %s", data)
|
||||
}
|
||||
}
|
||||
if err := scanner.Err(); err != nil {
|
||||
return err
|
||||
}
|
||||
if finish != "" {
|
||||
out <- ir.Delta{FinishReason: finish, UsagePromptTokens: usagePrompt, UsageCompletionTokens: usageCompletion}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
@@ -0,0 +1,145 @@
|
||||
// Gemini 出站适配器:IR → Google Generative Language streamGenerateContent(alt=sse)。
|
||||
package transform
|
||||
|
||||
import (
|
||||
"bufio"
|
||||
"bytes"
|
||||
"context"
|
||||
"encoding/json"
|
||||
"net/http"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
"hunjing/shenshi/internal/ir"
|
||||
)
|
||||
|
||||
// GeminiAdaptor Gemini generateContent 端点适配器。
|
||||
type GeminiAdaptor struct {
|
||||
EndpointName string
|
||||
BaseURL string
|
||||
APIKey string
|
||||
UpstreamModel string
|
||||
Client *http.Client
|
||||
}
|
||||
|
||||
// Name 实现 Adaptor。
|
||||
func (g *GeminiAdaptor) Name() string { return g.EndpointName }
|
||||
|
||||
type gemPart struct {
|
||||
Text string `json:"text"`
|
||||
}
|
||||
|
||||
type gemContent struct {
|
||||
Role string `json:"role,omitempty"` // user / model(Gemini 无 system 角色)
|
||||
Parts []gemPart `json:"parts"`
|
||||
}
|
||||
|
||||
type gemGeneration struct {
|
||||
Temperature *float64 `json:"temperature,omitempty"`
|
||||
MaxTokens *int `json:"maxOutputTokens,omitempty"`
|
||||
}
|
||||
|
||||
type gemRequest struct {
|
||||
Contents []gemContent `json:"contents"`
|
||||
Generation *gemGeneration `json:"generationConfig,omitempty"`
|
||||
SystemInstruction *gemContent `json:"systemInstruction,omitempty"`
|
||||
}
|
||||
|
||||
type gemChunk struct {
|
||||
Candidates []struct {
|
||||
Content struct {
|
||||
Parts []gemPart `json:"parts"`
|
||||
} `json:"content"`
|
||||
FinishReason *string `json:"finishReason"`
|
||||
} `json:"candidates"`
|
||||
UsageMetadata *struct {
|
||||
PromptTokenCount int `json:"promptTokenCount"`
|
||||
CandidatesTokenCount int `json:"candidatesTokenCount"`
|
||||
} `json:"usageMetadata"`
|
||||
}
|
||||
|
||||
// Stream 实现 Adaptor:Gemini SSE 数据帧转 IR Delta。
|
||||
func (g *GeminiAdaptor) Stream(ctx context.Context, req *ir.ChatRequest, out chan<- ir.Delta) error {
|
||||
body := gemRequest{}
|
||||
for _, m := range req.Messages {
|
||||
switch m.Role {
|
||||
case "system":
|
||||
if body.SystemInstruction == nil {
|
||||
body.SystemInstruction = &gemContent{}
|
||||
}
|
||||
body.SystemInstruction.Parts = append(body.SystemInstruction.Parts, gemPart{Text: m.Content})
|
||||
case "assistant":
|
||||
body.Contents = append(body.Contents, gemContent{Role: "model", Parts: []gemPart{{Text: m.Content}}})
|
||||
default: // user / tool 结果按 user 侧回传(工具循环随百工 M5 完整化)
|
||||
body.Contents = append(body.Contents, gemContent{Role: "user", Parts: []gemPart{{Text: m.Content}}})
|
||||
}
|
||||
}
|
||||
if req.Temperature != nil || req.MaxTokens != nil {
|
||||
body.Generation = &gemGeneration{Temperature: req.Temperature, MaxTokens: req.MaxTokens}
|
||||
}
|
||||
|
||||
payload, err := json.Marshal(body)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
url := strings.TrimRight(g.BaseURL, "/") +
|
||||
"/v1beta/models/" + g.UpstreamModel + ":streamGenerateContent?alt=sse"
|
||||
httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(payload))
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
httpReq.Header.Set("content-type", "application/json")
|
||||
httpReq.Header.Set("x-goog-api-key", g.APIKey)
|
||||
|
||||
client := g.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()}
|
||||
}
|
||||
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 == "" {
|
||||
continue
|
||||
}
|
||||
var chunk gemChunk
|
||||
if err := json.Unmarshal([]byte(data), &chunk); err != nil {
|
||||
continue
|
||||
}
|
||||
if len(chunk.Candidates) > 0 {
|
||||
c := chunk.Candidates[0]
|
||||
for _, part := range c.Content.Parts {
|
||||
if part.Text != "" {
|
||||
out <- ir.Delta{Delta: part.Text}
|
||||
}
|
||||
}
|
||||
if c.FinishReason != nil {
|
||||
finish := "stop"
|
||||
if strings.EqualFold(*c.FinishReason, "SAFETY") || strings.EqualFold(*c.FinishReason, "RECITATION") {
|
||||
finish = "content_filter"
|
||||
}
|
||||
var up, uc int
|
||||
if chunk.UsageMetadata != nil {
|
||||
up, uc = chunk.UsageMetadata.PromptTokenCount, chunk.UsageMetadata.CandidatesTokenCount
|
||||
}
|
||||
out <- ir.Delta{FinishReason: finish, UsagePromptTokens: up, UsageCompletionTokens: uc}
|
||||
}
|
||||
}
|
||||
}
|
||||
return scanner.Err()
|
||||
}
|
||||
+39
-9
@@ -1,14 +1,19 @@
|
||||
// 神识(Shenshi)入口:魂镜模型网关,rewrite/1.0 M0 骨架。
|
||||
// 数据面 POST /v1/chat/completions(IR 入站,SSE 出站);健康面 /healthz /readyz。
|
||||
// 神识(Shenshi)入口:魂镜模型网关(rewrite/1.0)。
|
||||
// 数据面 POST /v1/chat/completions(IR 入站,SSE 出站);健康面 /healthz /readyz;
|
||||
// 计量面 /me/usage、对账抽样 /admin/usage/sample;BYOK 面 /me/byok-keys。
|
||||
package main
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
"log"
|
||||
"net/http"
|
||||
"os"
|
||||
|
||||
"hunjing/shenshi/internal/byok"
|
||||
"hunjing/shenshi/internal/quota"
|
||||
"hunjing/shenshi/internal/route"
|
||||
"hunjing/shenshi/internal/server"
|
||||
"hunjing/shenshi/internal/store"
|
||||
)
|
||||
|
||||
func envOr(key, fallback string) string {
|
||||
@@ -19,13 +24,38 @@ func envOr(key, fallback string) string {
|
||||
}
|
||||
|
||||
func main() {
|
||||
port := envOr("SHENSHI_PORT", "8080")
|
||||
srv := server.New(
|
||||
server.Config{ServiceToken: os.Getenv("SHENSHI_SERVICE_TOKEN")},
|
||||
route.New(),
|
||||
)
|
||||
addr := ":" + port
|
||||
log.Printf("神识就绪:http://0.0.0.0%s(stub 端点,等待 M4 接入真实 Provider)", addr)
|
||||
st, err := store.Open(os.Getenv("SHENSHI_DATABASE_URL"))
|
||||
if err != nil {
|
||||
log.Fatalf("神识存储初始化失败: %v", err)
|
||||
}
|
||||
|
||||
cfg := server.Config{
|
||||
ServiceToken: os.Getenv("SHENSHI_SERVICE_TOKEN"),
|
||||
Store: st,
|
||||
}
|
||||
|
||||
// BYOK:主密钥未配置时整体禁用(fail-closed)
|
||||
if master := os.Getenv("SHENSHI_BYOK_MASTER_KEY"); master != "" {
|
||||
vault, err := byok.NewVault(master, st)
|
||||
if err != nil {
|
||||
log.Fatalf("BYOK 主密钥配置无效: %v", err)
|
||||
}
|
||||
cfg.Vault = vault
|
||||
log.Printf("BYOK 密钥库就绪")
|
||||
}
|
||||
|
||||
// 配额:用户级软配额(token 口径;0 = 不限制)
|
||||
var budget int64
|
||||
if raw := os.Getenv("SHENSHI_USER_BUDGET_TOKENS"); raw != "" {
|
||||
if _, err := fmt.Sscanf(raw, "%d", &budget); err != nil {
|
||||
log.Fatalf("SHENSHI_USER_BUDGET_TOKENS 须为整数: %v", err)
|
||||
}
|
||||
cfg.Quota = quota.New(st, budget)
|
||||
}
|
||||
|
||||
srv := server.New(cfg, route.New())
|
||||
addr := ":" + envOr("SHENSHI_PORT", "8080")
|
||||
log.Printf("神识就绪:http://0.0.0.0%s(端点目录 SHENSHI_ENDPOINTS,未配置时回声桩兜底)", addr)
|
||||
if err := http.ListenAndServe(addr, srv); err != nil {
|
||||
log.Fatal(err)
|
||||
}
|
||||
|
||||
Reference in new issue
Block a user