Commit Graph
76 Commits
Author SHA1 Message Date
William DrakeandClaude Opus 5.5 52c123e6f4 fix(billing): hold inflight reservation until billing lands; price estimate like billing; wire remaining endpoints
Address gaps found in adversarial verification of the in-flight balance
reservation:

- Early release: the reservation is now a ref-counted handle carried on the
  request context. Billing tasks submitted via submit*UsageRecordTask take a
  ref and return it only after RecordUsage has run; with reservation enabled
  the balance cache is deducted synchronously inside the billing task, so the
  reservation is released only after admission can see the deduction. Dropped
  tasks return their ref immediately. The reservation is renewed every ttl/3
  while the handler runs; after the handler returns it is no longer renewed,
  so the post-handler hold is bounded by one TTL.
- Estimate: use the same billing model (channel mapping + BillingModelSource,
  max over requested/mapped when ambiguous) and ModelPricingResolver (group
  ModelPricing incl. wildcards, channel per_request/image/video/token
  overrides), default/group/user rate + peak multiplier, plus image, video,
  audio and search surcharge kinds. Unpriced requests fail open with a
  throttled log; new billing.inflight_reservation.fail_closed_on_unpriced
  (default false) rejects them instead.
- Wire images, embeddings, web/x search, alpha search, grok media generation,
  grok realtime/voice, OpenAI-platform /v1/messages and ResponsesWebSocket
  (session-level reservation with renewal).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-28 16:14:43 -04:00
William DrakeandClaude Opus 5.5 43f88c99aa fix(billing): reserve in-flight balance to prevent concurrent overdraft
Balance-mode admission only checked cached balance > threshold, so N
concurrent expensive requests could all pass and overdraw the user.
Add a Redis in-flight reservation: at admission estimate a conservative
per-request cost (input estimate + output price x max_tokens, times rate
multiplier) and atomically require cached_balance - inflight >= estimate
(first in-flight request always admitted). Reservations are released via
defer on every path and expire by TTL. Fails open on estimation / Redis
errors. Configurable via billing.inflight_reservation (enabled by default).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-28 14:20:49 -04:00
wucm667 5e4958c889 fix(openai): schedule accounts with mapped model 2026-09-03 10:51:12 +08:00
刘佳林 aa7a811e66 feat(groups): add deny-or-downgrade control for reasoning effort ceiling
OpenAI/Composite groups can now reject requests that exceed the max
reasoning effort instead of always rewriting them down to the ceiling.
Existing groups keep the previous auto-downgrade behavior.
2026-09-01 10:15:45 +08:00
刘佳林 11ada80d58 feat(usage): 使用记录展示映射前的推理强度
用量行新增 requested_reasoning_effort,记录分组策略与模型族改写前的客户端请求值;推理强度列与模型列一致,主行显示请求值,有映射时用 ↳ 展示实际转发值。
2026-08-25 15:10:52 +08:00
alfadb f06bf181d2 feat(openai): support Fast mode service_tier across responses/chat/WS paths
- Accept fast|priority (canonical priority), flex|auto|default|scale on
  /v1/responses and /v1/chat/completions; reject unknown/empty/non-string
  with HTTP 400; omitted and null stay compatible.
- Propagate service_tier through JSON/SSE, Responses<->Chat conversions,
  fallback paths and HTTP->upstream WebSocket bridge.
- Billing prefers the upstream terminal tier; the outbound (policy-
  transformed) tier is used only when upstream omits the field.
  Explicit upstream default bills Standard even when Fast was requested.
- Pricing: Fast premium 2x Standard for gpt-5.6-sol/terra/luna and
  gpt-5.4; 2.5x for gpt-5.5; channel FastMultiplier stays authoritative.
- Live verification (official Codex 0.149.0 + gateway, HTTP & WS):
  upstream ChatGPT backend may return terminal default even when the
  account catalog advertises priority; billing follows the actual tier.
2026-08-24 11:52:48 +08:00
IanShaw 16b15e870d 修复 5888 与 5925 的同号重试语义冲突 2026-08-20 18:16:38 -07:00
IanShaw027 7c53a842e4 合并上游 main,保留 5925 的 Grok 重试上限与 5888 的协议兼容修复
同账号重试采用次数上限加 deadline;畸形 tools 在出站前删除;compaction 422 与结构化错误扫描一并保留。
2026-08-21 08:49:44 +08:00
IanShaw c628b3eea7 让 Grok stream idle 重试上限作用于主路径 2026-08-20 07:07:55 -07:00
IanShaw ad26172b83 完善 Grok 限流冷却与用量兼容 2026-08-20 02:32:43 -07:00
IanShaw c374ff2951 完善 OpenAI 网关切换与运维错误语义 2026-08-19 23:13:40 -07:00
IanShaw cf3577a3c5 fix(openai): harden Responses compatibility 2026-08-19 09:03:37 -07:00
shaw 10c8b70203 fix(cn-providers): 修复 CN 分组五项功能缺陷(调度闸门/计费/断开漏记/count_tokens/403)
对已合并 PR #5666 + 分组入口放行后的全量功能审计发现的 P0/P1 修复,
全部先经代码与厂商文档实证再实施:

1. /v1/messages 调度闸门(P0):sanitizeGroupMessagesDispatchFields 对非
   openai 平台恒置 AllowMessagesDispatch=false,而闸门豁免名单只有 grok,
   CN 分组经正常途径创建后恒 403——原生 Anthropic 直通(Claude Code 主用例)
   完全不可达。修复:闸门对 CN 与 grok 同语义豁免;count_tokens 处的内联
   裸检查统一走同一 helper;ResolveMessagesDispatchModel 对 CN 早退,避免
   openai 专属的 gpt-5.x 默认映射发给 CN 上游。

2. 计费候选链(P0):候选链兜底含客户端原始模型名,配合 getFallbackPricing
   的 claude→Sonnet 统一兜底,映射的 CN 模型无价时 CN 流量会按 Claude 原价
   (数倍~数十倍)静默误计,且 usage 日志显示 claude-* 名无从察觉。修复:
   CN 账号的 claude-* 候选仅在显式分组/渠道定价时放行;候选全滤空时按
   ErrModelPricingUnavailable 走零成本+告警落账(顺带修复原空候选错误会
   丢弃整条 usage 记录的次生问题)。

3. 断开/中断漏记(P0,#5148 对齐,惠及 openai 平台):messages/responses/
   chat_completions 三个 handler 的错误路径此前在 err!=nil 时丢弃携带的
   部分 result——客户端断开排水后的完整 usage 被丢,payg 上游照常计费而
   平台漏记(anthropic 网关早有同修复,openai 网关缺失)。修复:错误路径
   result 非空时照常提交 usage;failover 错误恒 result=nil 无重复计费。
   Responses×anthropic 流式转换器同时改为断开后继续排水至流自然结束
   (末尾 message_delta 的 output_tokens 不再丢),finalize 帧补工具名反转
   与客户端工具还原、仅在客户端仍连接时写出。

4. count_tokens(P1,证据修正):经实证三家 Anthropic 兼容层均无
   /v1/messages/count_tokens(DeepSeek 官方文档无此端点且注明
   anthropic-version 被忽略;OpenModel 标注该端点 Anthropic only),
   anthropic 协议转发上游=常态 404,且错误处置缺模型上下文会把不计费的
   探测放大成整账号停调。修复:CN 全协议一律本地 tiktoken 估算(与 Grok
   同方案),删除上游转发死代码。

5. 403 处置(P1):CN 此前落入通用 handleAuthError,单次 HTML 403(CDN/
   代理拦截页)即永久禁用,且 403 在 failover 集里会逐账号重放连环禁用
   整组。修复:CN 与 openai 同口径——HTML 豁免 + 3 次累计 + 临时冷却。

新增回归测试 8 项:闸门豁免(含 openai 仍受控断言)、CN 调度映射空返回、
候选过滤三态、空候选零成本落账、断开排水 usage 完整性、HTML-403 零处罚、
结构化 403 首次临时停调。handler/service 全包测试通过。
2026-08-17 17:28:53 +08:00
cyh 74fcdf3d42 fix(openai): preserve exponential capacity backoff 2026-08-08 01:02:02 +08:00
Wesley Liddick 27e8f69a9e Merge pull request #5171 from heathermhuang/codex/composite-reasoning-policy
feat(composite): enforce reasoning effort policy
2026-08-03 11:28:38 +08:00
shaw dec47e8fae fix(profit-control): stop leaking profit policy, close veto livelock, restore passthrough turn pricing
审计修复,逐条如下。

H1 利润策略泄露给所有普通用户
  profit_control_enabled / profit_min_margin / profit_safety_buffer 从
  dto.Group 移到 dto.AdminGroup(后者内嵌前者),赋值相应从
  groupFromServiceBase 移到 GroupFromServiceAdmin;前端 TS 同步从 Group 移到
  AdminGroup。dto.Group 是 GET /api/v1/groups/available 的响应体,该响应本就带
  rate_multiplier,相乘即可反推运营方上游采购成本上限。
  api_contract_test.go 的 /groups/available golden JSON 回滚这三个字段,并把
  fixture 改成非零值(require.JSONEq 是精确比对,缺字段即失败)。
  新增 dto 层边界测试:普通用户 DTO 不含三字段、管理员 DTO 仍含。

M1 利润终检 continue 与 failover 503 退避互动产生活锁
  FailoverState 新增 profitVetoedAccountIDs / profitVetoCount 与
  RecordProfitVeto():加入排除集 + 计数,达 maxProfitVetoAttempts(10) 返回
  FailoverExhausted。HandleSelectionExhausted 的 503 清空分支改为清空后把利润
  否决的账号放回排除集;若排除集已全部由利润否决贡献,清空不会带来任何新候选,
  直接判定耗尽(否则 SwitchCount 永不前进、退避条件永远成立,每 2s 空转一轮)。
  五个 handler 否决点(gateway_handler ×2 / responses / chat_completions /
  gemini_v1beta)改为经 RecordProfitVeto 决策,耗尽时按无可用账号终止。
  回归测试钉死:503 之后持续利润否决必须有限步终止且不 spin;未启用利润控制的
  请求退避语义完全不变。

M2 排队等槽后才终检,延迟可放大到 N × WaitPlan.Timeout
  OpenAI 侧选号循环(自有 failedAccountIDs map,非 FailoverState)新增
  recordOpenAIProfitVeto + handleOpenAIProfitVetoExhausted,共用同一上限语义。
  覆盖 responses / messages-dispatch / chat_completions / alpha_search /
  embeddings / images / grok_media 七处,以及 WS 两处否决分支。

M4 ws_v2 透传 ingress 绕过 per-turn 重定价(选方案 B:最小止血)
  透传 relay 只回调 AfterTurn、没有任何 turn 起始回调,hooks.BeforeTurn 永远
  不触发,而 handler 把 turnPricingAt 初始化成建连时刻 ⇒ 透传连接全部 turn 按
  建连时刻的高峰因子结算,客户端峰前建连保活即可全程谷价——正是本 PR 想堵的
  漏洞。改为 openAIWSTurnPricing 零值起步、只由 BeforeTurn 冻结;透传路径保持
  零值,RecordUsage 回退记录时刻,与引入利润控制前的基线一致。
  未选方案 A(给透传补 turn 起始回调):passthrough_relay.go 是 #5167 刚修过的
  取消传播/close frame 时序敏感区;且 BeforeTurn 还承担 turn>1 的并发槽位抢占,
  接进去等于给透传连接引入 per-turn 抢槽,风险远超本次修复范围。透传仍有建连时
  的准入门,只是没有 turn 级复核,已在两处注释写明。
  测试:service 层钉死透传 ingress 不触发 BeforeTurn(含失败时的复核指引),
  handler 层钉死零值语义与逐 turn 覆盖。

M5 装门读分组走了带账号计数聚合的 GetByID
  SchedulerSnapshotService 新增 GetGroupByIDLite,openai/gateway 两处装门改用
  之。门只需要平台/倍率/利润/高峰字段,且该查询发生在「是否启用利润控制」判定
  之前,未启用的分组同样付代价。两个测试 stub 的 GetByID 改成 panic 守卫。

M6 认证快照注释与真实读取路径相反
  门解析优先取 ctxkey.Group,而它就是本快照物化出来的对象,直连流量走的正是这
  条路。改正注释,与 api_key_repo.go 投影处的说明对齐,避免后人照旧注释删列。

M3 rate_multiplier 为 nil 时利润门 fail-closed(不改行为,加护栏)
  保留 fail-closed。补 repository 层测试钉死账号调度快照的 full/metadata 两份
  payload 都必须保留 RateMultiplier(含 0 值),漏列在 CI 就红。

L1 迁移号注释 191 / 191-192 改为实际的 192/193。
L2 admin group Create 的利润配置预校验改用与 CreateGroup 一致的归一化平台
   (新增 service.NormalizeGroupPlatform,两边共用)。保留预校验而非删除:
   service 层返回的是无类型 error,经 ErrorFrom 会变成 500,删掉会把合法的
   400 降级成 500。
L3 前端利润校验的上界改为判定换算后的小数(后端按小数校验 [0,1)),
   99.999% 会四舍五入进位成 1.0 而被后端 400;i18n en/zh 同步改为 0-99.99。
L4 clampProfitControlThreshold / profitControlOverThreshold 抽为共用函数,
   线上装门/否决点与 profit-preview 不再各自实现,附边界语义测试。
L5 profit-preview 补「默认 D 有账号但最低有效 D 归零」的告警(两档都为 0 由
   既有告警覆盖,不重复)。
2026-08-01 22:39:33 +08:00
Brisbanehuang fad2f215e8 fix(profit-control): decouple gate from image intent, propagate gate via selection, restore eager sticky fallback
Review fixes for the profit-control feature commit:

- Image intent no longer disables the profit gate. The shared /v1/responses
  handler previously skipped the pricing context (and therefore the gate)
  whenever the platform-wide image intent predicate matched, which includes
  Codex's passive image_gen namespace declaration: any client could disable
  admission control for anthropic/gemini/antigravity groups by declaring a
  namespace tool in the request body. Both /v1/responses paths now always
  install the token pricing context; image intent only drives capability
  routing and image billing. Mixed token+image requests stay token-gated;
  only dedicated media endpoints remain out of scope.
- Out-of-scope paths are now explicitly suppressed instead of implicitly
  ungated: Grok media (billed by media multipliers; also prevents in-flight
  video lookups from turning into spurious 404s), OpenAI-group count_tokens
  (unbilled), and Live calls (duration billed) carry a suppress marker that
  every install point honors, including the defensive scheduler-entry
  install.
- The gate resolved during selection now travels back to handlers on the
  AccountSelectionResult. The shared gateway installed the gate only on a
  scheduler-local context, so handler-side post-slot terminal rechecks and
  post-admission sticky binding were no-ops for anthropic/gemini/
  antigravity/shared-grok requests (and for composite-routed member groups
  on the OpenAI path). Handlers re-apply the carried gate via
  ContextWithSelectionProfitGate before the terminal recheck and binding;
  the WS acquired-selection branch gained the previously missing recheck.
- Sticky binding semantics restored for ungated traffic:
  BindStickySessionAfterProfitAdmission falls back to the official eager
  bind when no gate is installed (wait paths lost their only binding point
  otherwise), reads the pre-existing binding at bind time only when gated
  (removes the unconditional per-request Redis read the feature added to
  the shared handlers), and the legacy engine's three selection-time
  binding writes are skipped under a gate so a terminally vetoed account
  can no longer become the new sticky target.
- Responses WS connections re-freeze pricingAt and re-resolve the gate at
  every turn (BeforeTurn) and bill each turn with its own instant, closing
  the connect-at-valley/bill-at-valley window; a turn that fails the
  recheck closes the connection so the client reselects on reconnect.
- Terminal recheck no longer swaps a DB-fresh selected account for a
  staler snapshot object (UpdatedAt guard), and observer counters are
  documented as per-evaluation.
- Migrations renumbered 191/192 -> 192/193 after the passkey migration
  landed upstream as 191.

New regressions: selection-carried gate propagation (control group proves
the pre-fix no-op), image intent not disabling the shared gate, eager
binding fallback without a gate (both services), gated
read-failure/sentinel-miss binding semantics, legacy-engine deferred
binding under a gate with eager behavior preserved ungated, turn-level
pricing refresh (config re-resolution, scheduled-group precedence,
suppress, mid-connection disable), and suppress-marker coverage for the
request pricing context.
2026-08-01 22:39:32 +08:00
Brisbanehuang 20ad5ec506 feat(scheduler): per-group profit control for token account admission
Group pricing (rate multiplier, peak windows, per-user overrides) and
account cost (accounts.rate_multiplier) already live side by side, but
nothing stops the scheduler from handing a request to an account whose
cost multiplier exceeds what the group's pricing can profitably serve.
Add an opt-in per-group profit gate that filters scheduling candidates
by a margin rule, while ordering, scoring, stickiness and breakers keep
working unchanged among qualified accounts.

Admission rule: an account qualifies iff U <= D * (1 - min_margin -
safety_buffer) within a small relative epsilon, where U is
accounts.rate_multiplier (0 is legal; missing/negative/NaN/Inf are
conservatively rejected as invalid) and D is the requester's effective
downstream multiplier (user-group override ?? group default, times the
group peak factor) frozen at the request's pricing instant.

- groups gain profit_control_enabled / profit_min_margin /
  profit_safety_buffer (migration 191); the durable auth-cache
  invalidation trigger additionally watches the profit and pricing
  columns (migration 192) so out-of-band group edits cannot leave
  stale auth snapshots; GetByKeyForAuth explicitly projects the new
  columns and the API-key auth snapshot version is bumped to force a
  refresh of pre-existing snapshots
- request-level pricing instant: token entry points install pricingAt
  into ctx; the profit threshold D and the RecordUsage peak factor
  read the same instant, so one request never changes price mid-flight
  across waits/retries/failover (media and unwired paths keep the
  existing record-time semantics)
- the gate covers token requests on openai, anthropic, gemini, grok
  and antigravity groups: OpenAI-family handlers via
  WithOpenAIRequestPricingContext (responses incl. WS bridge, chat
  completions, messages, embeddings, alpha search), the shared gateway
  via WithGatewayTokenRequestPricing (messages, chat completions,
  responses, gemini model actions); composite groups cannot enable it
  directly; image/video/models/usage/count_tokens stay ungated and an
  explicit image-generation intent suppresses the gate end to end
- post-slot recheck: after a slot is acquired the account is re-read
  via SchedulerSnapshotService.GetAccount (scheduler cache, then DB;
  only when both fail the check fails open with WARN + metric); a
  vetoed account releases its slot and joins the request's exclusion
  set for reselection; sticky bindings are written only after the
  final check passes, and an over-threshold sticky account is skipped,
  not deleted, so it comes back once its rate recovers
- sticky-session cache contract: GatewayCache.GetSessionAccountID now
  returns ErrStickySessionNotFound on a miss (mapped from redis.Nil in
  the repository implementation, mirroring ErrRefreshTokenNotFound) so
  the profit sticky path can distinguish "no binding yet" from a real
  read failure without importing the cache driver in service code
- cross-group re-entry (composite parent -> member group) resolves the
  gate against the member group and clears a stale parent gate instead
  of letting a foreign threshold veto accounts
- per-platform/group activity counters (installs, threshold vetoes,
  invalid-rate vetoes, refresh failures) for observability
- admin UI: profit-control section on the five platforms' group forms
  with percent input, validation and platform-switch reset; group
  create/update/duplicate normalize and validate the config at a
  single choke point
- cmd/profit-preview: offline what-if tool that replays the production
  admission semantics over an exported config/account/override/model
  dump, reports per-model admitted-account counts under the default
  and the worst-case (lowest user override) D, and surfaces probe-sync
  staleness as warnings without affecting admission

Tests: service unit coverage for gate resolution/veto/threshold
epsilon/pricing instant/suppress marker/scheduler filtering and
post-slot recheck (incl. -race on the profit surface), unit-tagged
handler slot-recheck and capability-mapping regressions, sqlmock and
real-PostgreSQL integration regressions for the GetByKeyForAuth
projection and the migration-192 trigger watch list, API contract
update, and frontend specs for the five-platform form helpers.
2026-08-01 22:39:31 +08:00
Heatherm Huang d60f4e442b feat(composite): enforce reasoning effort policy 2026-08-01 21:15:44 +08:00
Edison42 1c0cb24c7e feat(usage): persist client session identifiers 2026-07-24 01:22:34 +08:00
Heatherm Huang 1c7959d0de fix: allow composite grok chat completions 2026-07-23 09:20:52 +08:00
Heatherm Huang 3a683fff55 Fix composite route alias attribution 2026-07-23 09:20:52 +08:00
Heatherm Huang ebc1028771 Add composite group routing 2026-07-23 09:19:24 +08:00
zhaozewu 6af622c340 feat(groups): add OpenAI reasoning policy
Persist reasoning ceilings and exact mappings for OpenAI groups, enforce them across HTTP and WebSocket forwarding, and invalidate cached auth snapshots.
2026-07-21 11:02:43 +08:00
haruka bd0f2d6405 fix: report chat stream transport failures 2026-07-18 03:24:04 +08:00
mt21625457 d11bdb13f5 feat(security-audit): add OpenAI-compatible prompt auditing 2026-07-17 00:39:39 +08:00
shaw 0408bdb34f Merge remote-tracking branch 'origin/main' into feat/upstream-rate-scheduling
# Conflicts:
#	backend/internal/handler/openai_gateway_handler.go
2026-07-16 10:31:21 +08:00
Wesley Liddick f3138ceb43 Merge pull request #4384 from wp-a/fix/openai-model-scoped-transient-cooldown
[codex] scope OpenAI transient cooldowns by model
2026-07-16 09:33:09 +08:00
Tian Lee 90ee85f3ef feat: 按上游计费倍率调度 OpenAI 账号 2026-07-16 00:40:46 +08:00
王鹏 40b8f04a6a fix(openai): scope transient cooldowns by model 2026-07-15 22:52:35 +08:00
wucm667 2fe7df9b81 fix(openai): reject image models on chat completions 2026-07-15 22:22:51 +08:00
shaw a0593b0bf8 fix(gateway): 客户端断开后 failover 静默终止,不再误报 502 账号耗尽
上游请求经 detachUpstreamContext(WithoutCancel) 有意脱离客户端取消(保
计费),但 failover 循环仍用原始 c.Request.Context() 重新选号:客户端
断开后上游返回 520 等可 failover 错误时,重新选号必然得到 context
canceled,被误判为账号耗尽,记录并返回通用 502。

修复:客户端已断开 ⇒ failover 静默终止。

- 新增 failoverClientGone(c):请求 ctx 已取消时先停 compact 心跳
  (建立 happens-before,对齐其它终结路径),响应未提交则标 499
  (statusClientClosedRequest,与并发槽取消路径同惯例)
- 7 个 OpenAI 内联 failover 循环(Responses/Messages/chat_completions/
  embeddings/images/grok_media/alpha_search)加双 guard:换号前 +
  选号失败分支入口;guard 位于 ReportOpenAIAccountScheduleResult(false)
  之后、RecordOpenAIAccountSwitch/池模式重试之前,账号健康副作用
  (service 层 detached ctx)不受影响
- FailoverState.HandleFailoverError/HandleSelectionExhausted 入口加
  ctx.Err() 检查返回 FailoverCanceled,取消不再改动 failover 状态;
  全部 10 个 FailoverCanceled 分支统一调用 failoverClientGone 归类 499
- 上游 detach 与计费设计不变;真实上游 520 事件仍完整落 ops
  (面板显示码 COALESCE(upstream_status_code,status_code)=520,
  错误率/告警口径不变)

测试:新增 openai_responses_failover_cancel_test.go 复现 issue 场景
(520+取消 ⇒ 不切号、499、无 502 终态)+ 在线客户端对照(正常切换、
耗尽 502);failover_loop_test.go 补入口取消用例并修正取消语义断言。

Fixes #4257
2026-07-15 09:29:05 +08:00
Heatherm Huang b32b815e46 fix(grok): harden OAuth pool recovery 2026-07-14 14:55:30 +08:00
Heatherm Huang 343390057d fix(grok): fail over OAuth credential errors safely 2026-07-14 14:55:30 +08:00
Heatherm Huang 8a22dc7347 fix(grok): diagnose unavailable models by platform 2026-07-13 10:11:34 +08:00
superman2003 7050070aa3 fix(grok): route cacheable chat requests via responses 2026-07-12 11:50:56 +08:00
InCerry 53a5c45bd8 fix(gateway): cap lenient json normalization
Fixes #3540
2026-07-09 11:15:52 +08:00
li 40c563c4ae fix(gateway): 记录请求体解析失败的真实原因,不再吞错
400 "Failed to parse request body" 此前丢弃底层错误,无法区分
JSON 真非法、还是 body 被截断/被中间件提前消费。

- 服务层 invalid json 错误增补 len/offset/非法字符信息
  (仅诊断元数据,不含 body 内容,可安全 wrap);
- handler 层新增 logRequestBodyParseFailure,向服务端日志输出
  底层错误 + body 长度 + 转义后的 head/tail 片段(各 256B),
  客户端响应文案保持不变;
- 接入全部 9 处入站解析点(messages/count_tokens/responses/
  chat_completions/embeddings,Anthropic 与 OpenAI 网关)。

Fixes #3715
2026-07-07 13:53:39 +08:00
shaw 0fd2e9216d fix(scheduler): 修复 OpenAI 高级调度器审计发现的正确性与性能问题
针对 #3692 合并后审计发现的问题集中修复:

- previous_response_id 剥离条件改为按 call_id 全覆盖校验,
  部分可重建的工具续链不再被误剥离(不受开关门控的行为回归)
- 粘性加权回退路径补分组归属校验并清理失效绑定,杜绝跨分组账号泄漏
- 账号列表页:无 OpenAI 账号时跳过分数计算、过滤池限定 openai 平台、
  负载批查合并为账号并集一次查询,消除全表扫描与 Redis N+1
- 订阅优先模式下常规池不可用时回退订阅池等待计划,
  busy-but-waitable 的订阅账号不再导致请求硬失败
- TopK/权重 DB 覆盖显式受总开关门控,与兄弟子开关语义一致
- 前端未分组 OpenAI 账号回退展示基础分,不再显示 "-"
- ListAllWithFilters 等能力正式进入 AccountRepository/AdminService 接口,
  移除匿名接口断言与静默降级;负载批查失败补 warn 日志
- SelectAccountWithSchedulerForCapability 增加显式 previousResponseCanMove
  参数,移除 "previous_response_can_move" 魔法字符串哨兵
- 设置写入路径补"基础权重不得全为零"聚合校验;
  运行时设置批量读取失败的降级路径覆盖全部键并留痕
2026-07-06 11:43:16 +08:00
mxyhi 82553c4dca fix(openai): preserve quota platform in usage billing 2026-06-28 11:13:09 +08:00
Wesley Liddick 2fc4fef847 Merge pull request #3310 from heathermhuang/codex/grok-subscription-support
feat: add grok subscription support
2026-06-26 15:41:52 +08:00
shaw fcd3bc1272 fix: return 404 model_not_found instead of 503 when no account supports the model 2026-06-26 15:38:06 +08:00
Heatherm Huang 39be1ec97f feat: add grok subscription support 2026-06-26 10:36:09 +08:00
FjlI5 bab8a9a93e fix(openai): log /v1/chat/completions upstream endpoint for chat-only API-key accounts
DeriveUpstreamEndpoint maps every OpenAI-platform request to /v1/responses, but
API-key accounts whose upstream only speaks Chat Completions
(!ShouldUseResponsesAPI) are forwarded directly to /v1/chat/completions. The
messages, responses and cyber-policy recording sites derived the endpoint via the
bare GetUpstreamEndpoint, so usage/ops records mislabeled those requests as
/v1/responses. Generalize the existing resolveRawCCUpstreamEndpoint into
resolveOpenAIUpstreamEndpoint and use it at every OpenAI recording site, matching
the already-correct chat-completions client path.
2026-06-18 00:29:04 +08:00
DaydreamCodingandClaude Opus 4.8 b62b573f7f feat(openai): cyber_policy 硬阻断全链路透传、审计与计费
上游对单次请求下发 error.code=cyber_policy 硬阻断时,网关在所有端点
(/v1/responses、/v1/chat/completions、/v1/messages、WebSocket)及流式/
非流式路径下,将该结果原样透传给客户端,绝不 failover、换号或同步拦截;
命中后异步完成审计与计费:

- 风控中心记录 cyber_policy 留痕并发送通知邮件,落库先于发信,SMTP 阻塞
  不影响留痕
- ops 错误请求记录,状态码对齐客户端实际接收(流式 200 / 非流式 400)
- 用量明细标记 request_type=cyber,按上游真实 token 计费,HTTP 与
  WebSocket 计费口径统一,零 token 命中不误扣
- 会话级自动屏蔽(管理员开关,默认关):命中的会话在可配 TTL 内本地拦截
  不再发往上游,仅屏蔽该会话不影响同 Key 其他会话
- 封号计数排除开关:可选让 cyber 命中不计入自动封号,命中当次不判定且
  历史行在违规计数中一并排除

WebSocket 多轮连接下 cyber 标记按 turn 生命周期管理,逐轮独立检测与记录;
透传的错误响应不被兜底逻辑追加内容污染。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-12 01:47:01 +08:00
Wesley Liddick 651b24571c Merge pull request #3000 from Pluviobyte/codex/sub2api-stream-field-validation
fix: validate stream field type across OpenAI-compatible handlers
2026-06-06 15:02:06 +08:00
whatIsNextToTheMoon 5bd3d90434 fix(openai): preserve upstream response.failed errors 2026-06-04 11:17:22 +00:00
Pluviobyte 3571b082fb fix: validate stream field type 2026-06-03 14:18:12 +08:00
Wesley Liddick 69e7c4db30 Merge pull request #2865 from wey-gu/feat/usage-request-context
fix(gateway): preserve usage request context
2026-05-29 16:21:59 +08:00
shaw ed1b57c597 fix(openai): gate routing by endpoint capability 2026-05-29 08:58:10 +08:00