- [Adaptive QoS Scheduling System with Pure Peer Calibration & Ladder Probing] (Fixes #3720, PR #3725):
- Pure TTFT Transient Extraction: Leverages the pure elapsed duration from request dispatch to upstream first-chunk arrival as the core physical metric, completely stripping output length and streaming transfer noise to detect silent QoS tiering demotion. (Fixes #3720, PR #3725)
- 5-Stage Pure Peer Baseline & Anti-False-Positive Shield: Peer baselines strictly exclude self, heterogeneous subscription tiers, disabled accounts, 429 rate-limited/zero-quota accounts, Jailed, Probing, Suspicious, and Half-Open accounts to completely eliminate baseline poisoning; introduces 35s dynamic threshold compensation for deep reasoning models to prevent misjudging thinking latency; applies global outage brakes (pausing penalties when pure peer average exceeds 20s) and an 8s~20s hysteresis deadband to eradicate state boundary jitter. (Fixes #3720, PR #3725)
- Transient Tagging & Immediate Session Handover: Automatically tags an account as
Suspiciouson its first strike (instantly excluding it from peer average calculation while activating drain mode to pause new session allocation); on 2 strikes within a 5-minute sliding window, immediately isolates the account into jail cooldown and forcibly unbinds its active sticky session; the P2C scheduler dynamically selects a pure healthy account and re-welds the sticky session, restoring 100% Prompt Cache efficiency. (Fixes #3720, PR #3725) - Background Ladder Probing & Half-Open Anti-Flapping: Implements an escalating cooldown schedule (2m -> 5m -> 10m -> 15m -> 20m -> 30m) with cooldown timers strictly reset to 0 upon entering each new tier; background daemon issues lightweight synthetic "Hello" probes (<5 tokens) with an 8-second hard timeout abort; passing probes transition the account to
Half-Openprobation, requiring 2 consecutive real requests under 8s to achieve full health; 30-minute penalty retention directly inherits elevated tiers upon relapse. (Fixes #3720, PR #3725) - Cross-UI Status Transparency & 12-Language i18n: Dynamically displays yellow status badges (
suspicious,jailed,probing,half_open) across Account Cards, Tables, Rows, and ProxyMonitor traffic logs with multilingual tooltip hints, automatically disappearing upon full recovery; fully backed by pure monolingual entries across all 12 supported language catalogs. (Fixes #3720, PR #3725)
- [CI Preview Release Asset Cap Protection]:
- Automated Stale Preview Asset Cleanup: Configured
removeArtifacts: truein the GitHub ActionsSync Preview Channel Releasestep, resolving an issue where accumulated pre-release assets reached GitHub's 1000-asset limit per release and caused publishing failures; ensures the preview release mirror only retains the latest version assets andupdater.json.
- Automated Stale Preview Asset Cleanup: Configured
- [基于纯净同侪校准与阶梯探针的自适应 QoS 调度系统] (Fixes #3720, PR #3725):
- 真实首包延迟 (TTFT) 瞬态感知: 以发出请求至收到上游首个有效 Chunk 的纯粹时间为核心指标，彻底排除输出代码量与流式传输的干扰，作为判定静默 QoS 分桶降级的权威物理依据。(Fixes #3720, PR #3725)
- 五重纯净同侪基线比对与防误杀体系: 同类账号基线严格排除自身、异构 Tier、本地禁用、429 限流/零配额、小黑屋(Jailed)、探测中(Probing)、瞬态可疑(Suspicious)以及缓刑(HalfOpen)账号，彻底杜绝劣币毒化基准；为深度思考模型提供 35s 动态时间补偿，杜绝推理耗时误杀；设置全局雪崩制动刹车（同行均值 >= 20s 时不滥杀）与 8s~20s 迟滞死区消除边界震荡。(Fixes #3720, PR #3725)
- 瞬态打标与会话即刻解绑 (Session Handover): 首次超标打上
Suspicious瞬态可疑标记（同行算均值瞬间排除它，调度器开启“只出不进”保护）；5 分钟滑动窗口内累计 2 次超标判定处于降级池，立即强制解绑粘性会话切断死锁灌水；由 P2C 调度器在纯健康号池中重新选号并立即焊死会话绑定，后续请求稳定享受 100% Prompt Cache 极速响应。(Fixes #3720, PR #3725) - 后台阶梯轻量自愈探针与半开防抖机制: 阶梯冷却梯度为 2m -> 5m -> 10m -> 15m -> 20m -> 30m，每次升级等待时钟严格从 0 重新计时；到达节点后台发起纯文本“你好”极简探针（8 秒硬超时中止，消耗 < 5 tokens）；探针通过后进入
Half-Open缓刑期，经受住连续 2 次真实请求考验（TTFT < 8s）才满血出狱；引入 30 分钟惩罚记忆，复发违规直接继承高阶梯防止高频进出。(Fixes #3720, PR #3725) - 多端状态透明化与 12 语种多语言支持: 在账号管理（卡片、表格、移动行）动态展示黄色状态标签（疑似被限速、限速小黑屋中、解限探测中、观察缓刑中），恢复健康后自动消除；在调用记录/流量日志（ProxyMonitor）表格的账号列小巧注入黄色 QoS 状态徽章与多语言浮层提示；完整覆盖 12 种单语语言字典。(Fixes #3720, PR #3725)
- [CI 预览发布资产配额防爆仓治理]:
- 旧版预览资产自动清理: 在 GitHub Actions 的
Sync Preview Channel Release步骤中配置removeArtifacts: true，解决历史预发布版本资产不断堆积导致触达单 Release 1000 个文件上限并引发发布失败的问题，确保每次发布预览版时 preview 镜像 Release 仅保留最新版本的资产与updater.json。
- 旧版预览资产自动清理: 在 GitHub Actions 的
What's Changed
Full Changelog: v4.9.9-beta.3...v4.9.9-beta.5