- [Official Hub Fingerprint Upgrade to 2.32.1 & Full Integration of Orchestrator (Argon)] (Ref #3745, #3937, PR #3949, Thanks to @cubelikeplayDaniel):
- Upgrade Official Hub Fingerprint Baseline to 2.32.1: Synchronized Hub baseline constants, User-Agent synthesis, and pipeline policies across the gateway to
antigravity/hub/2.32.1 (aidev_client; os_type=windows; arch=amd64; cl=1021296516), preventing outdated client rejections. (PR #3949) - Full Pipeline Integration of
gemini-pro-orchestrator-g1-pro: Aligned official model specifications (MODEL_PLACEHOLDER_M668, Orchestrator Argon, 1M context, 65,536 max output, thinking level channel) across model catalog, variant routing, CLI synchronization, and QoS streaming watchdog; integrated into Argon bucket isolation (Ref #3745) to ensure scheduling is restricted exclusively to Pro and Ultra accounts holding dedicated Argon weekly quota. (Ref #3745, PR #3949) - Exclude Internal Worker Models: Filtered Google internal worker model
gemini-flash-worker-agyglobally, preventing leakage into public recommended lists or quota dashboards. (PR #3949) - Unify Argon Capability & Row-Click Family Expansion: Unified
isArgonModelandisSupportedDisplayModellogic across account cards, tables, details modal, and Mini View; enabled direct row-clicking on model families in the Supported Models panel to expand or collapse variants. (Ref #3937, PR #3949)
- Upgrade Official Hub Fingerprint Baseline to 2.32.1: Synchronized Hub baseline constants, User-Agent synthesis, and pipeline policies across the gateway to
- [Hierarchical Model Catalog, -online Suffix Stripping & Dropdown Fix] (Fixes #3937, PR #3938, Thanks to @cubelikeplayDaniel):
- Strip
-onlineSuffix in Gateway Routing: Normalized incoming requests with-onlinesuffixes (e.g.gemini-3.8-flash-online) to base models, resolving 404 errors during web-search-augmented calls. (Fixes #3937, PR #3938) - Hierarchical Family View & Bare Testing: Restructured supported models display with tier grouping, default bare model connectivity testing, official Gemini icons, and custom dropdowns resolving native select rendering glitches. (Fixes #3937, PR #3938)
- Strip
- [ThinkingStore Plain-Text Signature Protection & Compaction Boundary Hardening] (Ref #3614, #3944, PR #3944, #3945, Thanks to @mclinyeshi):
- Protect Thought Signatures Across Plain-Text Turns: Prevented consecutive plain-text chat turns without tool calls from overwriting existing
thoughtSignatureentries in ThinkingStore. (PR #3944) - Tighten Compaction Boundary Matching: Refined compaction slicing and micro-folding boundary detection to eliminate spurious prompt cache resets. (Ref #3614, PR #3945)
- Protect Thought Signatures Across Plain-Text Turns: Prevented consecutive plain-text chat turns without tool calls from overwriting existing
- [CacheFirst Transient 5xx Account Protection & QoS TTFT Cache-Miss Alignment] (Ref #3639, #3821, #3893, #3899, #3935, PR #3942, #3943, Thanks to @mclinyeshi):
- Preserve CacheFirst Sticky Bindings on Transient 500/503/529: Maintained sticky account sessions through upstream transient server errors to safeguard high-value KV prompt cache assets. (PR #3943)
- Reconcile QoS TTFT Against Upstream usageMetadata: Adjusted time-to-first-token thresholds based on upstream cache miss metadata, avoiding false QoS penalty lockouts during cache warmups. (PR #3942)
- [Add Android ARM64 (Termux) Headless Server to Release Pipeline] (Fixes #3818, PR #3946, Thanks to @cubelikeplayDaniel):
- [官方 Hub 指纹升级至 2.32.1 与 Orchestrator (Argon) 编排者模型全链路接入] (Ref #3745, #3937, PR #3949, Thanks to @cubelikeplayDaniel):
- 官方 Hub 指纹基准升级至 2.32.1: 将常量定义、User-Agent 生成、版本同步与网关流水线策略基准线全量更新为
antigravity/hub/2.32.1 (aidev_client; os_type=windows; arch=amd64; cl=1021296516)，消除旧版本客户端拦截风险。(PR #3949) - 全链路集成
gemini-pro-orchestrator-g1-pro编排者模型: 对齐官方模型规格（MODEL_PLACEHOLDER_M668，Orchestrator Argon，1M 上下文、65,536 最大输出与思考通道），接入网关模型目录、路由映射、CLI 同步及 QoS 流式看门狗；将其纳入高阶 Argon 配额隔离与智能探测机制（Ref #3745），仅限具备 Argon 独立周配额的 Pro/Ultra 账号调度。(Ref #3745, PR #3949) - 内部 Worker 模型全链路排除: 统一过滤 Google 内部专用模型
gemini-flash-worker-agy，杜绝其泄露至公开推荐列表与配额看板。(PR #3949) - 前端 Argon 能力统合与模型组整行展开: 统合
isArgonModel与isSupportedDisplayModel判定，覆盖账号卡片、表格、详情弹窗与 Mini 悬浮窗；在「支持模型与集成」面板支持直接点击模型家族整行展开或折叠子档位。(Ref #3937, PR #3949)
- 官方 Hub 指纹基准升级至 2.32.1: 将常量定义、User-Agent 生成、版本同步与网关流水线策略基准线全量更新为
- [多级模型展示优化、-online 联网后缀剥离与下拉菜单交互修复] (Fixes #3937, PR #3938, Thanks to @cubelikeplayDaniel):
- [ThinkingStore 纯文本轮次签名保护与边界匹配加固] (Ref #3614, #3944, PR #3944, #3945, Thanks to @mclinyeshi):
- [CacheFirst 粘性会话瞬态 5xx 保护与 QoS TTFT 缓存未命中对齐] (Ref #3639, #3821, #3893, #3899, #3935, PR #3942, #3943, Thanks to @mclinyeshi):
- [Release 工作流新增 Android ARM64 (Termux) 架构支持] (Fixes #3818, PR #3946, Thanks to @cubelikeplayDaniel):
What's Changed
- fix(proxy,ui): strip -online suffix, hierarchical model catalog with default bare testing, Gemini icon & dropdown fix (Fixes #3937) by @cubelikeplayDaniel in #3938
- ci(release): build Android ARM64 headless server and attach release assets (Fixes #3818) by @cubelikeplayDaniel in #3946
- fix(proxy,thinking_store): tighten compaction boundary matching to prevent false cache resets by @mclinyeshi in #3945
- fix(proxy): prevent identical plain-text ThinkingStore turns from overwriting thoughtSignature by @mclinyeshi in #3944
- fix(proxy,pipeline): preserve CacheFirst bound account on transient 500/503/529 and protect sticky binding (Ref #3639, #3821, #3893, #3899, #3935) by @mclinyeshi in #3943
- fix(qos,pipeline): reconcile QoS TTFT against upstream usageMetadata to avoid false strikes on cache misses by @mclinyeshi in #3942
- feat(models,proxy,ui): upgrade Hub UA to 2.32.1 and integrate gemini-pro-orchestrator-g1-pro (Ref #3745, #3937) by @cubelikeplayDaniel in #3949
- chore(release): bump version to 5.0.4-beta.10 and update changelog (Fixes #3818, #3937, Ref #3614, #3639, #3745, #3821, #3893, #3899, #3935, #3944, Thanks to @cubelikeplayDaniel, @mclinyeshi) by @cubelikeplayDaniel in #3950
Full Changelog: v5.0.4-beta.9...v5.0.4-beta.10