- [Large Payload Upload Stall Decoupling and Thinking QoS False Positive Fix] (Ref #3763, Fixes #3772, Ref #3768, Ref #3781, PR #3784, Thanks to @cubelikeplayDaniel):
- Upload Stall Decoupling & Floor Guard: Completely decoupled
stall_secsfrom adaptiveupload_budget_secs, ensuring stall detection never shrinks to 3~5s and strictly respectsmax connect_secsto avoid false stall aborts during TLS/proxy handshakes; enforced a 15s budget floor for payloads ≥256KB. (Ref #3763, PR #3784, Thanks to @cubelikeplayDaniel) - Upload Terminal State & Replay Cascade Suppression: Explicitly marked
UpstreamTerminal::Uploadterminal state when all primary and fallback endpoints fail during upload, immediately returning 504 to prevent blind cross-account retries from exhausting healthy account quotas. (Ref #3763, PR #3784, Thanks to @cubelikeplayDaniel) - Unified Pipeline-Level Thinking Recognition: Implemented
payload_enables_thinkingunderpipeline::qosto inspect convertedgenerationConfig.thinkingConfigpayloads directly while matchingargon,-high,o1/o3/o4, andv4.0+patterns, fundamentally preventing thinking models likegemini-3.8-flash-highfrom being misjudged with a 20s non-thinking threshold during multi-turn chats. (Fixes #3772, Ref #3768, PR #3784, Thanks to @cubelikeplayDaniel) - Adaptive Security Prompt Sanitization: Expanded
RE_SECURITY_CTF_POLICYto seamlessly match security instructions missing theauthorizedprefix, extending fast-path pre-filters and multi-turn user prompt unit tests. (Ref #3781, PR #3784)
- Upload Stall Decoupling & Floor Guard: Completely decoupled
- [大请求上传超时解耦与思考模型小黑屋误封修复] (Ref #3763, Fixes #3772, Ref #3768, Ref #3781, PR #3784, Thanks to @cubelikeplayDaniel):
- 上传停滞超时解耦与保底: 彻底解耦
stall_secs与自适应upload_budget_secs,停滞超时不再随体积计算缩水至 3~5s,并强制保证绝不低于连接超时 (max connect_secs),杜绝 TLS 握手及代理抖动误触发停滞中断;对 ≥256KB 请求引入 15s 上传保底时间。(Ref #3763, PR #3784, Thanks to @cubelikeplayDaniel) - 端点耗尽终态防穿透重放: 当请求在上传阶段耗尽全部首选与备用端点时,明确标记
UpstreamTerminal::Upload终态并直接向客户端返回 504 截断,杜绝中等请求因未标记终态引发跨账号盲目轮换并耗死正常账号。(Ref #3763, PR #3784, Thanks to @cubelikeplayDaniel) - 流水线深度思考特征统一收敛: 在
pipeline::qos下实现统一的payload_enables_thinking,优先依据转出报文的generationConfig.thinkingConfig真实指令进行判定,并通配argon、-high、o1/o3/o4及v4.0+模型语义,从根源上消除gemini-3.8-flash-high等思考模型在多轮会话中被误判为非思考请求而按 20s 阈值连击误关小黑屋的问题。(Fixes #3772, Ref #3768, PR #3784, Thanks to @cubelikeplayDaniel) - 安全策略提示词自适应清洗: 扩展
RE_SECURITY_CTF_POLICY兼容无authorized前缀的提示词注入,并在前置快速过滤器与多行 user turn 中补全正则覆盖与自动化单测。(Ref #3781, PR #3784)
- 上传停滞超时解耦与保底: 彻底解耦
What's Changed
- fix(proxy,qos): 修复大请求上传超时挤压重放与思考模型小黑屋误封 (Ref #3763, Fixes #3772, Ref #3768, Ref #3781, Ref #3783) by @cubelikeplayDaniel in #3784
Full Changelog: v5.0.2-beta.0...v5.0.2-beta.1