github lbjlaq/Antigravity-Manager v5.0.4-beta.6
Antigravity Tools v5.0.4-beta.6
on GitHub

pre-release40 minutes ago
  • [Authoritative Data-Plane Ban Probe Refactor & Long-Interval Cooldown Loop] (Ref #3863, #3865, #3875, #3895, Thanks to @jeikl, @cubelikeplayDaniel):
    • Data-Plane Inference Probe Aligned with Authentic Antigravity Packets: Upgraded ban probe to the latest official baseline model gemini-3.8-flash-low (deprecating gemini-2.5-flash). Structured the probe via wrap_request with complete sessionId, systemInstruction, generationConfig.thinkingConfig (LOW), official requestId, and userAgent envelopes, preventing Google WAF from rejecting incomplete probe structures with 429 RESOURCE_EXHAUSTED.
    • Strict candidates Entity Verification & Dual-Endpoint Failover: Backed ban probes with dual endpoints (daily-cloudcode-pa and cloudcode-pa), enforcing Connection: close to prevent reusing severed TCP keepalive sockets after 403 blocks. Required candidates structure in 2xx responses to eliminate proxy intercept error pages or empty 200s from falsely clearing bans.
    • Ban Authority Convergence & Unknown-State Interception: Converged all quota refresh pathways into fetch_quota_with_retry. When probe returns Unknown or network timeouts occur, banned or Hard-403 accounts strictly bypass the quota summary endpoint, preventing 200 responses from washing away ban states.
    • Smart Long-Interval Cooldown Eliminating 429 Throttling: Healthy accounts bypass inference probes during automated background refreshes (probing at most once every 6 hours), while banned accounts cool down for 1 hour. Full probes run only upon manual card clicks, imports, or startup, eradicating 429 rate limits caused by excessive probing.
    • Zero Traffic Log Pollution: Excluded ban probes from ProxyRequestLog and monitor tables, keeping the dashboard exclusively dedicated to real user chat traffic.
  • [Eliminate Quota Disparity Unbinding in Sticky Session & 60s Window to Preserve KV Cache] (PR #3895, Ref #3639, #3893, Thanks to @cubelikeplayDaniel):
    • Preserve Prefix Cache Across Positive Quota: Completely removed quota_disparity_migration and migration_min_quota across Mode A (Sticky Session) and Mode B (60s Window). Accounts with quota > 0% are unconditionally reused without premature hopping, preserving expensive prefill KV cache in 100k~300k+ token sessions. (PR #3895)
    • Strict Failover Boundaries & Runtime Ban Escape: Constrained unbinding to 0% exhaustion, upstream 429s, or runtime 403 / TOS bans. Added test_sticky_session_unbinds_and_fails_over_when_bound_account_forbidden_or_banned to guarantee seamless failover to backup accounts upon runtime bans. (PR #3895)
  • [Configurable Sliding-Window Probe Concurrency, 1s Startup & Silent Streaming UI Sync] (Ref #3726, #3871, Thanks to @jeikl, @cubelikeplayDaniel):
    • Abolish Account Threshold & Introduce Concurrent Probes Configuration: Replaced the 15-account hard limit with refresh_concurrency (default 5, range 1~20) and refresh_slide_delay_secs (default 0.2s), configurable in both settings and the refresh dropdown menu.
    • 1-Second Fast Startup Probing: Reduced startup delay from 5s to 1s, safely verifying all accounts through the sliding window.
    • Silent Streaming UI Updates: Broadcasts progress as each account finishes, paired with 100ms debouncing and silent store hydration to eliminate screen jitter, full-page loading spinners, and toast flooding.
    • Eradicate Rapid-Fire Refresh Bug: Fixed an issue in Accounts.tsx where single account refresh triggered refreshQuota 3 times sequentially.
  • [Comprehensive Multi-Language & Protocol Badge Alignment]:
    • Synchronized sliding delay, concurrency, and startup refresh descriptions across 12 languages (AR, EN, ES, JA, KO, MY, PT, RU, TR, VI, ZH, ZH-TW).
  • [上游权威数据面封禁探针重构与智能长时冷却闭环] (Ref #3863, #3865, #3875, #3895, Thanks to @jeikl, @cubelikeplayDaniel):
    • 数据面推理探针与标准 Antigravity 报文对齐: 封禁探针全面升级为最新官方基准模型 gemini-3.8-flash-low（彻底淘汰废弃的 gemini-2.5-flash），通过 wrap_request 生成具备完整 sessionId、systemInstruction、generationConfig.thinkingConfig (LOW)、官方 requestId 与 userAgent 的正规请求信封，杜绝因结构残缺或废弃模型被 Google WAF 判定异常并抛出 429 RESOURCE_EXHAUSTED。
    • 严格 candidates 实体校验与双端点容灾: 探针采用 daily-cloudcode-pa 与 cloudcode-pa 双端点互备，请求显式携带 Connection: close 避免底层复用被 Google 403 掐断的残缺 TCP 连接引发假 Unknown / 连接重置；2xx 响应严格校验 candidates 实体，杜绝代理拦截错误页或空 200 误解封。
    • 解封权限严格收敛与未知状态阻断: 所有刷新入口统一收敛至 fetch_quota_with_retry，探针遇到网络超时或未知状态（Unknown）时，已封禁或 Hard 403 账号坚决阻断日配额接口，绝不被配额接口的 200 满配额反向冲刷洗白。
    • 智能长时冷却彻底根除 429 限流: 正常健康账号在后台自动定时刷新时跳过推理探针（至少 6 小时才顺便核验一次），已封禁账号后台冷却 1 小时，单账号手动刷新、导入新号与启动时才强制探测，彻底解决高频刷新消耗推理配额导致的 429 限流。
    • 彻底剔除流量日志记录: 封禁探针作为内部健康检测，完全不进入流量监控数据库和看板，彻底根除 /internal/ban_probe 刷屏，还监控面板纯净业务视图。
  • [粘性会话与 60s 窗口彻底移除配额悬殊解绑，锁定最大化 KV 缓存命中率] (PR #3895, Ref #3639, #3893, Thanks to @cubelikeplayDaniel):
    • 坚守前缀缓存与零悬殊跳号: 彻底移除模式 A（粘性会话）与模式 B（60s 窗口）中的 quota_disparity_migration（配额悬殊主动迁移）与 migration_min_quota 计算。只要绑定账号配额 > 0% 且未被限流/封禁，无条件优先复用，杜绝长会话（100k~300k+ tokens）因号池存在高额度账号频繁跳号导致昂贵的前缀预填充缓存（KV Cache）反复失效。(PR #3895)
    • 严密容灾边界与中途封禁无缝转移: 换号时机严格收敛于真正 0% 耗尽、429 配额用尽，或账号在运行中遭遇模型级 403 / TOS 封禁；补齐 test_sticky_session_unbinds_and_fails_over_when_bound_account_forbidden_or_banned 防御用例，确保中途封禁依然秒级解绑、无缝切换可用备选账号续杯。(PR #3895)
  • [滑动窗口并发探测容量可配置化、毫秒级启动与静默流式刷新上屏] (Ref #3726, #3871, Thanks to @jeikl, @cubelikeplayDaniel):
    • 废除账号上限门槛，引入「同时探测账号数」配置: 移除冷启动“超过 15 个不刷新”硬门槛，新增 refresh_concurrency（同时探测账号数，默认 5，范围 1~20）与 refresh_slide_delay_secs（滑动窗口延迟，默认 0.2 秒）配置项，在刷新下拉菜单二级面板与设置页均可自由调节。
    • 秒级启动探测: 启动等待时间由 5 秒缩减至 1 秒，配合滑动窗口安全完成全池探测。
    • 静默平滑流式刷新上屏: 后端滑动窗口每完成一个账号立即推送事件，前端采用 100ms 防抖与 silent 静默更新模式，不触发全屏 Loading 转圈、不闪烁、不弹窗，账号列表顺滑流式更新最新状态。
    • 根治前端连击手误: 修复 Accounts.tsx 单账号刷新连续调用 3 次 refreshQuota 的遗留问题，杜绝多线程并发读写状态竞态。
  • [多语言与协议标签全方位对齐]:
    • 12 种语言（中、繁、英、日、韩、西、葡、俄、土、越、阿、马）完整同步滑动延迟、同时探测账号数与启动刷新文案。

What's Changed

Full Changelog: v5.0.4-beta.5...v5.0.4-beta.6

Check out latest releases or
releases around lbjlaq/Antigravity-Manager v5.0.4-beta.6

Don't miss a new Antigravity-Manager release

NewReleases is sending notifications on new releases.

Get notifications