gproxy v3.0.1

v3.0.1

Faster settlement, catalogue and audit fixes, and a repaired v2 usage import.

English

Fixed

• Manual "import from upstream" in the console now asks the provider even when auto_refresh_models is off; the switch only governs what a client's model listing triggers (#214).
• OpenAI-compatible catalogues that label the entry model instead of id are parsed again (#215).
• The request audit and usage pages no longer freeze the range end at the moment the page opened: applying a search follows the clock unless the end was edited (#217).
• v2 usage rows written with the { "dimensions", "quantities" } envelope are unwrapped on import, and the dry run rejects metrics v3 cannot aggregate instead of importing them (#218). A store already imported by 3.0.0 must be re-imported from the backup the upgrade kept beside the database.
• A bootstrap administrator key that starts with sk- is accepted as a bearer token.

Changed

• Settlement leaves the response path on native hosts: the usage row lands after the reply, through a bounded backlog that stops the process outrunning its ledger, and the SQLite backend folds queued settlements into grouped commits. Decrypted credentials are served from memory. Measured on one machine, the gateway adds 0.21 ms to a request and settles more than 27,000 requests per second; see the Performance page.
• Streams and websockets reserve their settlement slot when they open, so a burst of streams cannot grow memory without bound.
• Credential health observations that do not change state are persisted at most once every thirty seconds.

简体中文

修复

• 控制台的"从上游导入模型"在关闭 auto_refresh_models 后也会请求上游；该开关只控制客户端请求模型列表时是否顺带刷新（#214）。
• 兼容 OpenAI 的模型目录若用 model 而非 id 标识条目，现在能正确解析（#215）。
• 请求审计与用量统计页面不再把结束时间冻结在打开页面的时刻：未手动修改结束时间时，每次应用筛选都跟随当前时间（#217）。
• v2 以 { "dimensions", "quantities" } 信封写入的 usage 记录在导入时被拍平；dry run 会拒绝 v3 无法汇总的 metrics，而不是照样导入（#218）。已用 3.0.0 导入过的库需要从升级时留在数据库旁边的备份重新导入。
• 以 sk- 开头的引导管理员密钥可以作为 Bearer 令牌使用。

变更

• 原生宿主的结算不再占用响应路径：用量行在响应发出后落库，经过一个有上界的积压队列，进程不会跑得比账本快；SQLite 后端把排队的结算合并成批量提交。解密后的凭据从内存提供。单机实测网关为每个请求增加 0.21 ms，每秒可结算 27,000 个以上的请求，详见文档的性能页。
• 流式与 WebSocket 在打开时就预留结算名额，突发的流不会让内存无限增长。
• 状态未变化的凭据健康观察最多每三十秒持久化一次。