Skip to content

UsageMeter v0.11.2

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 09 Aug 15:13
v0.11.2
4ce0e86

[0.11.2] - 2026-08-09

Added

  • Upstream Model Probe: Gateway profiles now support model-list discovery and single-model availability probes against the upstream (persisted, capped at 200 models); probe requests bypass the forwarder so they never count toward usage or trip the circuit breaker

Changed

  • Gateway & Proxy URL Shortening: Gateway profile ids use short random ids (p + 6 hex), handle ids are unified to 8-hex, and proxy URLs use the /umg/<tool>/s/<id> format — old gateway-<hex>-<hex> and legacy URL formats remain parseable and route correctly
  • SSE Parser Consolidation: The three duplicated SSE streaming loops now share one SseEventReader driver in sse.rs
  • Reasoning Token Semantics: opencode / hermes token accounting now uniformly folds reasoning into output tokens with a single merge point

Fixed

  • Gateway Double-Counting: Gateway records now reconcile with local session scans by message id, so the same physical request no longer produces two usage facts
  • Codex Proxy Double-Counting: Gateway Codex records now join the fuzzy-matching pool, eliminating duplicate usage facts when the same request appears both via the gateway and in local Codex scans
  • Streaming Idle Timeout: Streaming responses are now reaped after 300s of upstream silence (configurable, explicit 0 still disables), preventing leaked connections
  • Legacy Idle Timeout Migration: Existing users whose settings defaulted to a 0s idle timeout are migrated to 300s (settingsVersion 3, persisted so the migration applies once across all load paths)
  • Request Body Stall Timeout: read_body_limited now enforces a 60s per-chunk idle timeout (408), preventing requests that stall after declaring a Content-Length from hanging the upstream
  • Gateway Protocol Mismatch: Requests whose body does not match the profile's protocol now get a clear 400 instead of an opaque upstream error
  • Startup DB Concurrency: Proxy startup database initialization is serialized to avoid concurrent-init races

新增

  • 上游模型探测：网关 profile 支持对上游做模型列表发现与单模型可用性探测（持久化保存，上限 200 个模型）；探测请求不经过转发器，因此不计入用量、不触发熔断

变更

  • 网关与代理 URL 精简：Gateway profile id 改为短随机 id（p + 6 位 hex），handle id 统一为 8 位 hex，代理 URL 采用 /umg/<tool>/s/<id> 新格式——旧 gateway-<hex>-<hex> 与历史 URL 格式仍可解析并正确路由
  • SSE 解析收敛：三套重复的 SSE 流式循环统一为 sse.rs 中的 SseEventReader 公共驱动
  • Reasoning Token 口径：opencode / hermes 的 token 统计统一将 reasoning 并入 output tokens，仅保留单一合并点

修复

  • 网关重复计数：网关记录按 message id 与本地会话扫描对账，同一物理请求不再产生两条用量事实
  • Codex 代理重复计数：网关 Codex 记录纳入模糊匹配池，消除同一请求同时出现在网关与本地 Codex 扫描时的重复用量事实
  • 流式空闲超时：流式响应在上游静默 300 秒后被回收（可配置，显式 0 仍可关闭），避免连接泄漏
  • 旧版空闲超时迁移：存量用户默认 0 秒空闲超时的设置自动迁移为 300 秒（settingsVersion 3，写盘固化使迁移在所有加载路径上一次生效）
  • 请求体停滞超时：read_body_limited 逐 chunk 增加 60 秒空闲超时（408），防止声明 Content-Length 后断流的请求挂起上游
  • 网关协议不匹配：请求体与 profile 协议不符时返回明确的 400 错误，而非晦涩的上游报错
  • 启动期数据库并发：代理启动期数据库初始化改为串行执行，避免并发初始化竞态
Assets 17
Loading