github lbjlaq/Antigravity-Manager v4.9.7-beta.8
Antigravity Tools v4.9.7-beta.8
on GitHub

pre-release4 hours ago
  • [Token Stats Overview & Deep Insights Dashboard, Full-Stack Thinking (Reasoning) Stats & Unified Dark Theme]:
    • Full-Stack Thinking / Reasoning Token Analytics:
      • Database & Proxy Integration: Seamlessly migrated proxy logs (proxy_db) and aggregate tables (token_stats) to track reasoning_tokens. Extracted model thinking token consumption automatically from streaming SSE chunks and non-streaming responses across all protocols.
      • Multi-Dimensional Aggregations: Extended hourly, daily, weekly, all-time, account-level, and model-level statistical endpoints to return aggregated reasoning tokens and average reasoning intensity.
      • Overview Dashboard Integration: Prominently featured reasoning tokens with a dedicated purple badge and output proportion (Reasoning: XX (YY%)) in the main usage overview card, along with average reasoning throughput per request in the J23 efficiency card header.
    • "Overview" Dashboard View Integration & Polishing: Introduced the new Overview dashboard featuring a 360° radial ray gauge, core model breakdown with tooltip inspection, 24-hour activity bar pulse, J23 efficiency gauge, and account leaderboard with single-model highlight focus.
    • Dark Theme Background & Component Elevation Unification: Eradicated fragmented hardcoded dark backgrounds (dark:bg-[#0a0a0c], dark:bg-[#121316], dark:bg-[#1c1c1e], etc.), aligning entirely with global design tokens (card containers dark:bg-base-100, pills & segmented controls dark:bg-base-200, borders dark:border-base-200/300), eliminating seams against sticky navigation bars and harmonizing Classic vs new dashboard visual tones.
    • Full 12-Language Localization for Token Stats: Fully synchronized token_stats i18n keys across all 12 supported locales (English, Spanish, Russian, Arabic, Japanese, Korean, Portuguese, Turkish, Vietnamese, Burmese, Simplified & Traditional Chinese), decoupling hardcoded week formatting.
    • Chart Tooltips & Custom Pricing Modal Dark Tuning: Optimized dark mode backgrounds, borders, and input controls for chart tooltip overlays and the custom pricing modal (PricingModal), ensuring seamless tactile feedback across both light and dark modes.
  • [Intelligent Multi-Tier Scheduling: Weekly Quota EDF, Zero-Quota Session Auto-Failover, Cross-Model Isolation & Claude Compaction Hardening] (PR #3640, PR #3641, PR #3642, PR #3644, PR #3645, PR #3647, PR #3650, Thanks to @modded-soldier-9, @cubelikeplayDaniel):
    • Weekly Quota Earliest Deadline First Scheduling (EDF / Use-It-Or-Lose-It): Accurately extracted Google 7d/weekly bucket quota reset timestamps, introducing Earliest Deadline First rules (Priority 0.5) to multi-account rotation and P2C random selection. Prioritizes accounts approaching their reset window (weekly reset <= 24h or 5h bucket <= 2h with quota > 0%) to consume expiring capacity before scheduled resets, completely preventing usable allowances from going to waste under rigid high-quota-first policies. (PR #3645, Fixes #549, Ref #900, Thanks to @modded-soldier-9)
    • Continuous Session & Preferred Account Zero-Quota Auto-Failover: Added proactive quota awareness to Sticky Session and Preferred Account selection. When a bound account hits 0% quota for the requested model while healthy accounts remain available, atomically unbinds and re-routes requests before atomically rebinding the session upon success, preventing hard socket breaks and client interruptions during continuous AI Agent coding. (PR #3642, Fixes #3639, Ref #3509, Thanks to @modded-soldier-9)
    • Cross-Model Quota Lockout Isolation: Completely refactored get_quota_reset_time and fetch_and_lock_with_realtime_quota to strictly match normalized model IDs and only apply long-horizon reset lockouts when the target model's quota is truly exhausted (percentage <= 0). Prevents transient 429s on healthy models from inheriting days-long lockout timestamps from unrelated exhausted models (such as Claude Opus), eliminating false 503 Service Unavailable errors and pool starvation. (PR #3640, Fixes #3627, Thanks to @modded-soldier-9)
    • Imagen 3 Micro-gap In-Place GraceRetry & Transient Cooldown Rescue: Introduced micro-gap in-place recovery (delay <= 2500ms) during the first multi-account rotation round for Google image generation models, eradicating rapid 50ms cascading sweeps that exhausted whole pools within 300ms. Also introduced queue-level cooldown wait rescue in get_image_token within remaining timeout budgets, dramatically boosting image generation success rates. (PR #3641, Fixes #3329, Fixes #3409, Thanks to @modded-soldier-9)
    • Claude Cowork Deep Compact State Machine & Continuation Tag Hardening: Adjusted the sliding immunity lease window (COMPACTION_IMMUNITY_WINDOW_SECS) to 90 seconds to stably accommodate long reasoning models and complex multi-step MCP toolchains, completely resolving Double-400 false alarm loops. Expanded continuation scanning depth to the first 5 turns to pierce through multiple preceding <system-reminder> and environment detail blocks, while recognizing /compact. and /compact <args> command variants. (PR #3644, PR #3647, Ref #3576, Thanks to @modded-soldier-9, @cubelikeplayDaniel)
    • Claude Code Session Usage Scaling Bypass & Native 95% Auto-Compaction Self-Healing: Unified is_code_mode detection logic to precisely recognize Claude Code CLI and Desktop Code Tab; forcibly bypassed usage scaling for Code mode requests when enable_usage_scaling is enabled to reflect raw physical token usage. Completely eradicates false low-usage signals that pacified client auto-compaction watchdogs, preventing sessions from failing to trigger /compact and abruptly disconnecting upon exceeding context limits. (PR #3650, Fixes #867, Fixes #1066, Fixes #1079, Thanks to @cubelikeplayDaniel)
  • [Gemini Thinking Tier Alias Normalization & Inert Padding Degrade] (PR #3646, PR #3649, Fixes #3629, Ref #2208, Thanks to @ThermalEng, @modded-soldier-9):
    • Normalize Pro Thinking Tier Aliases to Upstream Physical ID: Reused the unified model compatibility mapping to ensure bare gemini-3.1-pro routed to high, medium fallback, or default tier aliases, as well as explicit gemini-3.1-pro-high, are all normalized to the real upstream physical model ID gemini-pro-agent, completely resolving 400 INVALID_ARGUMENT errors when querying gemini-3.1-pro-high upstream. (PR #3646, Ref #2208, Thanks to @ThermalEng)
    • Degrade Gemini Fallback User Padding to Inert Punctuation: Downgraded fallback placeholder text (TRANSIT_DEFENSE_FALLBACK_TEXT) in ensure_gemini_payload_ends_with_user from natural language "ok go on" to inert punctuation "." for empty/substanceless user turns. Satisfies Gemini protocol non-empty parts constraints while completely preventing AI Coding Agents (such as Cline, Claude Code, OpenCode) from misinterpreting synthetic text as explicit user approval, permanently securing human-in-the-loop gates and eliminating loop cycles. (PR #3649, Fixes #3629, Thanks to @modded-soldier-9)
    • Zero Thinking Budget Inbound Hard Cutoff & Soft Target Documentation: When zero thinking budget is explicitly configured or derived via adaptive fast-response compression, strictly inject both includeThoughts: false and thinkingBudget: 0, achieving 100% parity with Google's native IDE protocol to completely extinguish upstream reasoning chains; additionally clarified in the settings UI across all 12 languages that upstream Gemini treats positive thinkingBudget as a heuristic soft target that can dynamically expand under complex prompts. (Ref #3648)
  • [HTTP API Field Restoration, Quiet Startup, Missing storage.json Self-Healing & Admin Hot-Sync] (PR #3651, Fixes #3631, Fixes #3630, Fixes #1411, Fixes #3270, Fixes #3275, Fixes #3280, Fixes #3260, Thanks to @modded-soldier-9):
    • Full HTTP / Headless API Metadata Serialization: Fully supplemented validation_url, validation_blocked*, priority, disabled*, proxy_disabled*, and protected_models metadata fields across AccountResponse in modules/http_api.rs, eliminating dead code in external panels and headless consumers to ensure unblock URLs and account statuses are transparently exposed. (PR #3651, Fixes #3631, Thanks to @modded-soldier-9)
    • Validation Blocked State Complete Sanitization: Concurrently cleared validation_blocked_until, validation_url, validation_blocked_reason, and reset quota.is_forbidden = false within clear_validation_blocked in modules/account.rs, relaxing cleanup guard conditions to eradicate circular lockout states caused by residual flags after manual unblocking. (PR #3651, Fixes #3630, Thanks to @modded-soldier-9)
    • Comprehensive Quiet Startup Arguments Support: Expanded apply_login_launch_policy CLI parameter checks to detect --silent, --hidden, -s, and /minimized in addition to --minimized, enabling seamless background tray launch from Windows Task Scheduler, batch scripts, and desktop shortcuts. (PR #3651, Fixes #1411, Thanks to @modded-soldier-9)
    • Self-Healing Missing storage.json on Clean Install Profile Switching: Enhanced get_storage_path and write_profile in modules/device.rs to automatically create parent directories and initialize fresh device profile configurations when IDE directories exist without a pre-generated storage.json, eliminating storage_json_not_found errors on initial switch. (PR #3651, Fixes #3270, Thanks to @modded-soldier-9)
    • Dual Codex / ChatGPT CLI Directory Synchronization: Accommodated path alterations following Codex's rebranding to ChatGPT, seamlessly synchronizing configuration files whenever .chatgpt and .codex directories exist concurrently or individually. (PR #3651, Fixes #3275, Thanks to @modded-soldier-9)
    • Admin Server Hot Configuration Sync: Extended save_config in commands/mod.rs to hot-reload model filtering, proxy policies, and circuit breaker settings for admin_server.axum_server, ensuring proxy rules take immediate effect even when running the management dashboard standalone. (PR #3651, Fixes #3280, Fixes #3285, Thanks to @modded-soldier-9)
    • Extended Linux RPM Installation Path Detection: Added standard /usr/lib64 and /usr/lib paths to Linux IDE executable detection in modules/process.rs for Fedora, openSUSE, and RHEL distributions. (PR #3651, Fixes #3260, Thanks to @modded-soldier-9)
  • [Token 统计「总览看板」与「深度洞察」视界接入、全链路思考量 (Thinking) 统计与全景深色主题统一]:
    • 全链路思考量 (Thinking / Reasoning Tokens) 统计支持:
      • 后端与存储全量对齐: 代理层（proxy_db 与 token_stats 数据库）全面扩展 reasoning_tokens 字段并支持平滑迁移，在 SSE 流式及非流式响应中自动从 usageMetadata 与各协议返回中提取模型深度思考 Token 消耗并持久化。
      • 全维度聚合与查询支持: 小时/天/周/全部历史、账号维度与模型维度的统计接口均扩展思考量聚合，提供精确的思考 Token 规模及均次思考指标支持。
      • 总览看板深度集成: 「用量调度全景」卡片常驻呈现思考 Token 专属紫色指标与输出占比（思: XX (YY%)），下半区「效能指数 · J23」标题栏同步展示单次请求思考量密度（~XX 思 / req），全 12 国语言无缝支持。
    • 「总览看板」全新视界与体验调优: 新增「总览看板」模式，集成 360° 辐射彩针发光仪、核心模型生态分布、24 小时律动立柱、J23 效能刻度仪表与活跃账号排行榜；支持单模型高亮聚焦与交互联动。
    • 深色主题背景与组件层级统一规范: 彻底清除硬编码深色背景（dark:bg-[#0a0a0c]、dark:bg-[#121316]、dark:bg-[#1c1c1e] 等），全面统一对齐系统全局设计规范（卡片容器 dark:bg-base-100、药丸与分段控制器 dark:bg-base-200、边框 dark:border-base-200/300），完美消除粘性导航栏与页面接缝色差，统一经典看板与新看板的明暗调性。
    • Token 统计全 12 种多国语言完备支持: 全量同步并补齐 token_stats i18n 键值至所有 12 种语言本地化文件（包含英语、西语、俄语、阿语、日语、韩语、葡语、土耳其语、越南语、缅甸语及简繁中文），解耦自然周周数格式化硬编码。
    • 图表 Tooltip 与价格配置弹窗暗色调谐: 优化趋势图/饼图浮层 Tooltip 与模型自定义价格配置弹窗（PricingModal）在深色模式下的底色透明度、输入框及重置按钮对比度，保证全局交互体验自洽舒适。
  • [智能多维调度增强：周配额临期优先、零配额长对话无感迁移、跨模型隔离与 Claude 压缩自愈加固] (PR #3640, PR #3641, PR #3642, PR #3644, PR #3645, PR #3647, PR #3650, Thanks to @modded-soldier-9, @cubelikeplayDaniel):
    • 周配额临期优先调度算法 (EDF / Use-It-Or-Lose-It): 从账号配额详情中精准提取 Google 7d/weekly 桶配额重置时间戳，在多账号调度与 P2C 随机选号中引入最早截止时间优先规则（Priority 0.5）。当账号处于临期窗口（周配额 <= 24h 或 5h 滚动配额 <= 2h 且额度 > 0%）时赋予优先调度权，在被 Google 周期重置前优先消耗释放，彻底根除高配额优先策略导致的临期剩余资产白白作废清零问题。 (PR #3645, Fixes #549, Ref #900, Thanks to @modded-soldier-9)
    • 持续会话与固定账号零额度无感自动迁移: 在粘性会话（Sticky Session）与固定优先账号（Preferred Account）选号阶段增加前置配额感知。当绑定账号的目标模型配额降至 0% 且号池存在具备正配额的健康账号时，立即原子解绑并在选号完成后原子重绑新会话，避免继续死锁复用 0% 账号导致客户端碰壁报错中断，保障多轮 Agent 编程持续顺畅。 (PR #3642, Fixes #3639, Ref #3509, Thanks to @modded-soldier-9)
    • 配额重置时间跨模型精准隔离: 彻底重构 get_quota_reset_time 与 fetch_and_lock_with_realtime_quota。增加模型名称归一化精准匹配门禁，仅当目标模型实际配额耗尽 (percentage <= 0) 时才提取长周期重置时间；当额度充足的模型遭遇并发 429 时，坚决拒绝盗用账号内其他耗尽模型（如 Claude Opus）数十小时的重置时间，平滑回退至秒级指数退避阶梯，根除误报 503 调度受限与号池假死。 (PR #3640, Fixes #3627, Thanks to @modded-soldier-9)
    • Google Imagen 3 微间隙原地 GraceRetry 与限流队列救援: 针对生图模型上游 429 短周期重置延迟（通常约 1.17s），在多账号第 1 轮调度中引入微间隙优先原地自愈机制（delay <= 2500ms 原地等待重试），彻底根除 50ms 盲目快切在 300ms 内遍历打穿全池导致全军覆没的缺陷；同时在 get_image_token 中引入限流冷却预算救援，只要请求超时预算充足自动等待瞬态冷却后重试，生图成功率大幅提升。 (PR #3641, Fixes #3329, Fixes #3409, Thanks to @modded-soldier-9)
    • Claude Cowork 深度压缩状态机与接续标记穿透加固: 将压缩后接续免死租约滑动窗口（COMPACTION_IMMUNITY_WINDOW_SECS）进一步放宽调优至 90 秒，稳固覆盖慢推理思考与超长耗时 MCP 工具链，彻底消除 Double-400 假报警死循环；将接续标记扫描深度拓展至前 5 轮消息，穿透多轮 <system-reminder> 与环境详情块稳定捕获摘要标记，并补齐 /compact. 与 /compact <args> 命令变体。 (PR #3644, PR #3647, Ref #3576, Thanks to @modded-soldier-9, @cubelikeplayDaniel)
    • Claude Code 会话用量缩放隔离与原生 95% 自动压缩探针自愈: 统一抽离 is_code_mode 精准识别 Claude Code CLI 与 Desktop Code Tab；在开启 enable_usage_scaling 时对 Code 模式强制旁路缩放以如实反馈物理 Token 消耗，杜绝因缩放后虚假低用量麻痹客户端探针导致永远无法触碰 95% 自动压缩水位线、进而在超限后突发断连报错的问题。 (PR #3650, Fixes #867, Fixes #1066, Fixes #1079, Thanks to @cubelikeplayDaniel)
  • [Gemini 思考档位别名规范化与防空转标点惰性降级] (PR #3646, PR #3649, Fixes #3629, Ref #2208, Thanks to @ThermalEng, @modded-soldier-9):
    • Pro 系列思考档位别名规范化为上游真实 ID: 统一复用模型兼容映射表，确保裸模型 gemini-3.1-pro 在 high、medium 回退或缺省路由阶段派生出的公开别名，以及显式指定的 gemini-3.1-pro-high 别名，均统一规范化映射至上游真实物理模型 ID gemini-pro-agent，彻底修复直接向上游请求 gemini-3.1-pro-high 导致的 400 INVALID_ARGUMENT 报错。 (PR #3646, Ref #2208, Thanks to @ThermalEng)
    • Gemini 空轮次保底垫片降级为惰性标点: 将通用防御层 ensure_gemini_payload_ends_with_user 中针对空文本/无实质内容用户轮次的保底填充文本（TRANSIT_DEFENSE_FALLBACK_TEXT）由口语短句 "ok go on" 彻底降级为中性惰性标点 "."。在完全满足 Google Gemini 协议 parts 非空校验的同时，杜绝被 AI Agent（如 Cline、Claude Code、OpenCode 等）误判为人类明确授权指令，根治人机安全确认门禁被击穿与上下文死循环缺陷。 (PR #3649, Fixes #3629, Thanks to @modded-soldier-9)
    • 零思考预算严格关闭上游思维链与面板软偏好特性释义: 在进站流水线中，当用户显式配置或自适应极速压缩至 0 思考预算时，严格双向注入 includeThoughts: false 与 thinkingBudget: 0，对齐官方抓包协议规范彻底断绝思维链；并在设置面板与全 12 国语言中明晰标注 Google Gemini 上游将 thinkingBudget 视为启发式软偏好（Soft Target）之自适应特性，杜绝长 Prompt 深思场景下的误解。 (Ref #3648)
  • [修复 HTTP API 验证字段丢失、静默启动支持、storage.json 缺失自愈与管理端热同步] (PR #3651, Fixes #3631, Fixes #3630, Fixes #1411, Fixes #3270, Fixes #3275, Fixes #3280, Fixes #3260, Thanks to @modded-soldier-9):
    • HTTP / Headless API 元数据字段全量补齐: 在 modules/http_api.rs 的 AccountResponse 中全面补充 validation_url、validation_blocked*、priority、disabled*、proxy_disabled* 与 protected_models 等全量元数据字段，彻底消除前端与第三方 API 消费者的死代码，确保外部集成与无头面板可无缝获取解封验证链接与全景状态。 (PR #3651, Fixes #3631, Thanks to @modded-soldier-9)
    • 账号验证封禁彻底自愈清理: 在 modules/account.rs 的 clear_validation_blocked 中同步清空 validation_blocked_until、validation_url、validation_blocked_reason 并彻底重置 quota.is_forbidden = false，放宽清理防护判据，彻底根除手动解封后因残留脏标记导致的环形锁死问题。 (PR #3651, Fixes #3630, Thanks to @modded-soldier-9)
    • 开机静默启动参数全面支持: 为 apply_login_launch_policy 增加 --silent、--hidden、-s 与 /minimized 等常用命令行参数检测，支持 Windows 计划任务、批处理或快捷方式传入参数时静默启动至托盘。 (PR #3651, Fixes #1411, Thanks to @modded-soldier-9)
    • 全新安装初次切号 storage.json 缺失自愈: 在 modules/device.rs 的 get_storage_path 与 write_profile 中，当全新安装的 Antigravity IDE 尚未生成 storage.json 时自动递归创建目录并初始化设备指纹配置，彻底根治初次切号报 storage_json_not_found 错误。 (PR #3651, Fixes #3270, Thanks to @modded-soldier-9)
    • Codex / ChatGPT CLI 目录双向同步兼容: 适配 Codex 更名为 ChatGPT 后的配置路径变化，当 .chatgpt 与 .codex 目录同时或单独存在时执行双向全量同步。 (PR #3651, Fixes #3275, Thanks to @modded-soldier-9)
    • 管理服务 (Admin Server) 配置同步热更新: 在 commands/mod.rs 的 save_config 中，当用户保存配置时同步热更新 admin_server.axum_server 的模型过滤与代理策略，确保仅运行管理后台时修改配置即刻生效。 (PR #3651, Fixes #3280, Fixes #3285, Thanks to @modded-soldier-9)
    • Linux RPM 发行版安装路径补充: 在 modules/process.rs 的 Linux 安装路径探测中补充 Fedora / openSUSE / RHEL 规范的 /usr/lib64 与 /usr/lib 路径。 (PR #3651, Fixes #3260, Thanks to @modded-soldier-9)

What's Changed

  • fix(proxy): 修复配额重置时间跨模型污染导致额度充足时误报 503 调度受限 (Fixes #3627) by @modded-soldier-9 in #3640
  • fix(image,proxy): 优化 Imagen 3 微间隙原地 GraceRetry 与限流冷却队列救援 (Fixes #3329, Fixes #3409) by @modded-soldier-9 in #3641
  • feat(proxy): 支持持续会话与固定账号零额度无感自动切换 (Fixes #3639, Ref #3509) by @modded-soldier-9 in #3642
  • fix(cowork,compact): 加固 Claude Cowork 深度压缩状态机与接续标记穿透识别 (Ref #3576, RFC Deep Compact) by @modded-soldier-9 in #3644
  • feat(proxy): 引入周配额临期优先调度算法 (EDF / Use-It-Or-Lose-It)，防止周期重置作废剩余额度 (Fixes #549, Ref #900) by @modded-soldier-9 in #3645
  • fix(defense): 将 Gemini 空轮次保底垫片调整为惰性标点，根治 AI Agent 确认状态机被 'ok go on' 劫持缺陷 (Fixes #3629) by @modded-soldier-9 in #3649

New Contributors

Full Changelog: v4.9.7-beta.7...v4.9.7-beta.8

Check out latest releases or
releases around lbjlaq/Antigravity-Manager v4.9.7-beta.8

Don't miss a new Antigravity-Manager release

NewReleases is sending notifications on new releases.

Get notifications