Files
sub2api/backend/internal/service/upstream_response_model.go
T
shaw b689e5b401 fix(billing): harden response-model billing and repair its test fixtures
按上游响应模型计费的准入过宽、且自带用例必然失败,本次一并修复。

严格化准入
- 新增 PricingService.GetIdentifiedModelPricing / BillingService.HasIdentifiedTokenPricing:
  只接受价格表中能被确定性识别的条目(精确名、已知拼写变体、去掉日期版本后缀),
  不再接受 getFallbackPricing / matchByModelFamily 按子串猜出的系列兜底价。
  此前上游只要自报一个含 "haiku" 的编造名字就能被判定"已定价",把账单压到最便宜的
  系列价(实测 claude-opus-4.8 基线 $0.0019250 → $0.0000963,20 倍少收)。
  GetModelPricing 的对外行为不变,仅把前三步查找抽成共用函数。
- 图片 / 视频 / 网页搜索请求不再走响应模型覆盖:这些路径按张、按秒、按次定价,
  与准入检查所验的 token 价不是同一套价格表。
- 准入判断抽成 responseModelBillingDeclaration,两条计费主干共用同一套规则。

正确性与可观测性
- 去掉 recordUsageCore 中 billingModel 的无效赋值(ineffassign 已启用,会让
  golangci-lint 直接失败),改由日志表达实际生效的计费基准。
- 补 cost != nil 守卫,与 OpenAI 侧及本文件既有写法对齐。
- 每次实际生效的基准切换记一条 billing.response_model_applied,少收可审计。
- 修正 upstreamResponseModelObserver 上"冲突仅用于诊断、永不影响计费"的过期注释。

测试
- gpt-5.1 与 gpt-5.5 实际共用同一条 gpt-5.4 价格,夹具"价格必须不同"的前置断言
  必然失败,OpenAI 侧 3 个用例(含 4 个子用例)从未跑通;改用 gpt-5.4-nano /
  gpt-5.5。Anthropic 侧 claude-opus-4 不是价格表精确条目,改用 claude-opus-4.8。
- 夹具增加"必须可被确定性识别"的前置断言,避免用例被更靠前的门挡掉而失去判别力。
- 新增:准入规则表驱动用例、可识别性判定用例、编造家族名在两条主干上均被拒的用例。

前端
- 选择该模式时提示"计费基准以上游自报模型为准,只降不升,仅对可信上游启用"。
2026-08-10 18:45:15 +08:00

181 lines
4.9 KiB
Go

package service
import (
"strings"
"github.com/gin-gonic/gin"
"github.com/tidwall/gjson"
)
const (
upstreamResponseModelObserverContextKey = "upstream_response_model_observer"
upstreamResponseModelMaxLength = 200
)
// upstreamResponseModelObserver tracks one forwarding attempt (or one WS turn).
// A terminal declaration wins over an earlier declaration; otherwise the first
// declaration is retained. Observation never affects the forwarding path.
//
// Billing normally ignores the observed model as well; the only exception is a
// channel explicitly configured with billing_model_source = response_model,
// where a conflict flag makes billing fall back to the baseline model
// (see responseModelBillingDeclaration).
type upstreamResponseModelObserver struct {
first string
terminal string
conflict bool
}
func (o *upstreamResponseModelObserver) Observe(model string, terminal bool) {
model = normalizeObservedUpstreamResponseModel(model)
if model == "" {
return
}
current := o.Model()
if current != "" && !strings.EqualFold(current, model) {
o.conflict = true
}
if terminal {
o.terminal = model
return
}
if o.first == "" {
o.first = model
}
}
func normalizeObservedUpstreamResponseModel(model string) string {
model = strings.TrimSpace(model)
if model == "" {
return ""
}
runes := []rune(model)
if len(runes) > upstreamResponseModelMaxLength {
model = string(runes[:upstreamResponseModelMaxLength])
}
return model
}
func (o *upstreamResponseModelObserver) ObserveOpenAI(payload []byte, eventType string) {
model := firstValidTrimmedGJSONModel(payload, "response.model", "model")
o.Observe(model, isUpstreamResponseModelTerminalEvent(eventType))
}
func (o *upstreamResponseModelObserver) ObserveAnthropic(payload []byte) {
model := firstValidTrimmedGJSONModel(payload, "message.model", "model")
o.Observe(model, false)
}
func (o *upstreamResponseModelObserver) ObserveGemini(payload []byte) {
model := firstValidTrimmedGJSONModel(
payload,
"modelVersion",
"response.modelVersion",
"response.response.modelVersion",
)
// Gemini streaming has no universal terminal event carrying modelVersion;
// treating each declaration as terminal retains the latest chunk.
o.Observe(model, true)
}
func (o *upstreamResponseModelObserver) Model() string {
if o == nil {
return ""
}
if o.terminal != "" {
return o.terminal
}
return o.first
}
func (o *upstreamResponseModelObserver) Conflict() bool {
return o != nil && o.conflict
}
func beginUpstreamResponseModelObservation(c *gin.Context) *upstreamResponseModelObserver {
observer := &upstreamResponseModelObserver{}
if c != nil {
c.Set(upstreamResponseModelObserverContextKey, observer)
}
return observer
}
func upstreamResponseModelObserverFromContext(c *gin.Context) *upstreamResponseModelObserver {
if c == nil {
return nil
}
value, ok := c.Get(upstreamResponseModelObserverContextKey)
if !ok {
return nil
}
observer, _ := value.(*upstreamResponseModelObserver)
return observer
}
func observedUpstreamResponseModel(c *gin.Context) string {
return upstreamResponseModelObserverFromContext(c).Model()
}
func observedUpstreamResponseModelConflict(c *gin.Context) bool {
return upstreamResponseModelObserverFromContext(c).Conflict()
}
func observeOpenAISSEBody(observer *upstreamResponseModelObserver, body string) {
if observer == nil || strings.TrimSpace(body) == "" {
return
}
forEachOpenAISSEDataPayload(body, func(payload []byte) {
eventType := strings.TrimSpace(gjson.GetBytes(payload, "type").String())
observer.ObserveOpenAI(payload, eventType)
})
}
func firstValidTrimmedGJSONModel(payload []byte, paths ...string) string {
if len(payload) == 0 {
return ""
}
for _, path := range paths {
value := gjson.GetBytes(payload, path)
if !value.Exists() || value.Type != gjson.String {
continue
}
if model := strings.TrimSpace(value.String()); model != "" {
// Validate only after finding a candidate. This avoids a full validation
// pass on the common model-free delta path while still rejecting malformed
// payloads that appear to declare a model.
if !gjson.ValidBytes(payload) {
return ""
}
return model
}
}
return ""
}
func isUpstreamResponseModelTerminalEvent(eventType string) bool {
switch strings.TrimSpace(eventType) {
case "response.completed", "response.done", "response.failed", "response.incomplete", "response.cancelled", "response.canceled":
return true
default:
return false
}
}
func upstreamModelMismatch(sentModel, responseModel string) *bool {
responseModel = strings.TrimSpace(responseModel)
if responseModel == "" {
return nil
}
sentModel = strings.TrimSpace(sentModel)
mismatch := sentModel == "" || !strings.EqualFold(sentModel, responseModel)
return &mismatch
}
func upstreamSentModel(requestedModel, upstreamModel string) string {
sentModel := strings.TrimSpace(upstreamModel)
if sentModel == "" {
sentModel = strings.TrimSpace(requestedModel)
}
return sentModel
}