- [Antigravity CLI (agy) Native Evolution, Anti-Telemetry Tracking & Dynamic Model Selection] (Ref #3686):
- Smooth Migration to Antigravity CLI (agy): Officially migrated the CLI synchronization module to support Antigravity CLI (
agy), while maintaining 100% backward compatibility for legacygeminiCLI detection and restoration. (Ref #3686) - Dynamic Model Selection & Live State Inspection: Completely refactored CLI model configuration. Users can now freely choose or customize any model from the frontend dropdown (Gemini 4 Argon series, Gemini 3.8 Flash, Gemini 3.1 Pro, Claude, Codex, etc.). The backend inspects the active model from local configuration in real-time, pre-selecting it on the UI and providing immediate feedback upon synchronization.
- Automated Shell Wrapper & Intelligent Reasoning Effort Splitting: Automatically wraps
~/.local/bin/agyto seamlessly injectGEMINI_API_KEY,GOOGLE_GEMINI_BASE_URL, and telemetry suppression headers. Intelligently decomposes composite model names into base models and thinking effort levels (e.g.gemini-3.8-flash-low->--model gemini-3.8-flash --effort low), preventing known-model resolution errors from silently downgrading to cloud backends. - Google OAuth Credential Isolation & Leakage Prevention: Automatically backs up and isolates legacy
antigravity-oauth-tokencredentials during sync, enforcingmodelProvider: "gemini"and API key authentication to prevent erroneous requests to Google's internaldaily-cloudcode-pa.googleapis.comservice that previously caused stream EOF and connection drops. - Full 12-Language Localization & Changelog Synchronization: Comprehensive localization across all 12 supported languages including Chinese (Simplified & Traditional), English, Japanese, Korean, Spanish, Portuguese, Russian, Turkish, Vietnamese, Arabic, and Malay.
- Smooth Migration to Antigravity CLI (agy): Officially migrated the CLI synchronization module to support Antigravity CLI (
- [Fix Retry-After Header Extraction & Case-Insensitive Matching for 503 Responses] (PR #3687, Fixes #3506, Thanks to @EricZhou05):
- Case-Insensitive Parsing for Rate-Limit Retry-After Duration: Fixed an issue where
extract_retry_after_secondsfailed to parse cooldown seconds after the rate-limit message was updated to"All accounts rate-limited. Cooldown wait Ns.", due to hardcoded uppercase"Wait "matching. Added chained matching for both"Wait "and"wait ", ensuring 503 responses reliably emit standardRetry-After: Nheaders so downstream clients (such as Claude Code and OpenCode) back off accurately without blind retries or prolonged stalls. (PR #3687, Fixes #3506, Thanks to @EricZhou05)
- Case-Insensitive Parsing for Rate-Limit Retry-After Duration: Fixed an issue where
- [Complete Elimination of 0ms Optimistic Reset Storms, Physical Cooldown Backoff & Endpoint Health Decoupling] (PR #3685, Fixes #3506, Ref #3525, Ref #3634, Ref #3666, Ref #3680, Thanks to @cubelikeplayDaniel):
- Eradicate 0ms Optimistic Reset Storms & Introduce Physical Async Cooldown: Completely removed
clear_for_optimistic_reset()fromtoken_manager, stopping the gateway from blindly zero-delay replaying requests across all limited accounts and triggering cascading hard lockouts during concurrent 429 bursts. When the minimum account pool cooldown is<= 5s, executes real physical async sleep (wait_sec * 1000 + 100ms) to allow rate limits to naturally recover; when wait time is> 5s, explicitly returns a cooldown requirement error prompting standard client-side backoff. (PR #3685, Thanks to @cubelikeplayDaniel) - Two-Round Adaptive Paced Rotation & Stepped Backoff Retries: Relaxed Round 1 pool rotation delay from 50ms to 150ms to grant a micro-cooldown window; expanded the request-level 429 fail-fast threshold from 1 round to 2 rounds (
pool_size * 2), introducing linear stepped backoff (1.5s ~ 3.5s) in Round 2 to accommodate large prompt rate limit recovery while preventing indefinite stalls. Verified in production to slash high-concurrency 429 error rates from 37.4% down to 0.0%. (PR #3685, Thanks to @cubelikeplayDaniel) - Strictly Decoupled Background Quota Probing from Primary Inference Health: Created a dedicated, independent
QUOTA_ENDPOINT_HEALTHtracker (60-second observation window) inmodules/quota.rswith probe timeouts extended to 15 seconds, completely preventing background periodic quota andproject_idpolling timeouts from degrading the primary inference gateway's endpoints. (PR #3685, Thanks to @cubelikeplayDaniel) - Adaptive Candidate Endpoint Watchdog & Tightened HTTP/2 Idle Pool: Watchdog fail-safe timeout now dynamically scales based on request body size (15s baseline, plus 1s per 256KB), preventing 11MB+ large prompts from premature failover while uploading; tightened HTTP/2
pool_idle_timeoutto 15 seconds and cappedpool_max_idle_per_hostat 10 to eliminate reusing dead TCP connections silently terminated by Google GFE. (PR #3685, Thanks to @cubelikeplayDaniel)
- Eradicate 0ms Optimistic Reset Storms & Introduce Physical Async Cooldown: Completely removed
- [Official Antigravity Native Identity Alignment & Tool Function Safe Sanitization] (Fixes #3684):
- Official Antigravity Native Identity Alignment: Addressed Google's recent heightened semantic feature rate limiting against external harnesses (e.g. Claude Code CLI, OpenCode, DSH, Pi, Codex), which triggered spurious 429 errors. Introduced the official native Antigravity identity statement (
OFFICIAL_ANTIGRAVITY_IDENTITY) inPromptSanitizer, harmonizing external client preambles and broad attribution statements to mirror the official IDE identity, completely removing harness fingerprints and securing parity with official IDE quota and rate-limit treatment. (Fixes #3684) - Tool Function (Tools) Description Defensive Sanitization: Prevented WAF blocks triggered by billing headers (e.g.
x-anthropic-billing-header:) or SDK fingerprints injected into tool declarations (tools[].functionDeclarations[].description). Added header-track defensive filtering to tool descriptions in the inbound pipeline while 100% preserving function names (name) and parameter schemas (parameters), guaranteeing smooth, error-free autonomous agent tool execution. (Fixes #3684)
- Official Antigravity Native Identity Alignment: Addressed Google's recent heightened semantic feature rate limiting against external harnesses (e.g. Claude Code CLI, OpenCode, DSH, Pi, Codex), which triggered spurious 429 errors. Introduced the official native Antigravity identity statement (
- [Traffic Monitor Detail Redesign: Unified Single-Tab View, Structured Payload Inspection & Dual-Link HAR Export]:
- Unified Postman-Style Single-Tab Container: Refactored the detail modal into "Request", "Forwarded", and "Response" primary tabs, eliminating split-screen compression and maximizing vertical inspection space alongside inline token pills.
- Structured Payload & Raw Code Dual Views: 1-click toggle to intuitively inspect system prompts, multi-turn message bubbles, tool declarations, thinking blocks, and function call receipts.
- Key-Value Headers Table & Masking: Modernized headers and URL query params into interactive 2-column tables with real-time search, hover copy, and
Bearer sk-***masking. - Standard Dual-Link HAR 1.2 Export (All 3 Payloads Complete): Generates dual-link HAR entries (Client->Gateway and Gateway->Upstream), guaranteeing complete retention of inbound request, forwarded payload, and server response; supports 1-click cURL and JSON copying.
- Full 12-Language i18n Support: All new UI components, labels, and feedback messages are fully localized across all 12 supported system languages.
- [Antigravity CLI (agy) 原生全面演进、防遥测封号与模型动态选择导入] (Ref #3686):
- 平滑迁移至 Antigravity CLI (agy): 在 CLI 配置同步模块中正式迁移支持 Antigravity CLI (
agy)，并保持对旧版geminiCLI 探测与配置恢复的 100% 向后兼容。 (Ref #3686) - 前端模型自由选择与状态实时反查回显: 彻底重塑 CLI 模型导入机制，前端下拉列表支持自由选择或自定义输入任意模型（Gemini 4 Argon 全系列、Gemini 3.8 Flash、Gemini 3.1 Pro 及 Claude / Codex 等）；后端在状态检测时实时反查提取本地实际生效的配置模型并精准回显至前端，同步成功提供包含选定模型的即时感知反馈。
- 自动 Shell Wrapper 接管与智能思考档位拆解: 自动识别系统
~/.local/bin/agy二进制并接管写入包装脚本，无缝注入GEMINI_API_KEY、GOOGLE_GEMINI_BASE_URL及防风控遥测标头；底层智能拆解基底模型与思考档位（如gemini-3.8-flash-low自动映射为--model gemini-3.8-flash --effort low），杜绝因复合命名引发已知模型解析失败而静默降级的问题。 - Google OAuth 历史凭据自动隔离与防云端误连: 在同步时自动对历史
antigravity-oauth-token实施安全备份隔离，强制声明modelProvider: "gemini"与 API Key 鉴权模式，从根源杜绝客户端误走 Google 内部账号通道连接daily-cloudcode-pa.googleapis.com引发流式 EOF 与网络中断异常。 - 全量 12 种系统语言与更新日志同步: 覆盖简体中文、繁体中文、英文、日文、韩文、西文、葡文、俄文、土文、越文、阿文、马文等全部 12 国语言标签与提示文案。
- 平滑迁移至 Antigravity CLI (agy): 在 CLI 配置同步模块中正式迁移支持 Antigravity CLI (
- [修复 503 响应 Retry-After 头部提取与大小写兼容] (PR #3687, Fixes #3506, Thanks to @EricZhou05):
- 全账号限流重试秒数解析大小写双向兼顾: 解决历史提交在将全账号限流错误文案升级为
"All accounts rate-limited. Cooldown wait Ns."时，extract_retry_after_seconds内部硬编码检索大写"Wait "导致无法解析冷却秒数的缺陷；支持"Wait "与"wait "大小写双向链式匹配，确保 503 错误响应正常生成Retry-After: N标准头部，避免外部客户端（如 Claude Code / OpenCode 等）因缺失退避指示触发无序重试或长冻结。 (PR #3687, Fixes #3506, Thanks to @EricZhou05)
- 全账号限流重试秒数解析大小写双向兼顾: 解决历史提交在将全账号限流错误文案升级为
- [彻底移除 0ms 乐观重置风暴、物理冷却退避与端点健康解耦] (PR #3685, Fixes #3506, Ref #3525, Ref #3634, Ref #3666, Ref #3680, Thanks to @cubelikeplayDaniel):
- 根除 0ms 乐观重置风暴与物理异步冷却: 彻底移除
token_manager中的clear_for_optimistic_reset()机制，杜绝多会话并发瞬态 429 时网关在 0ms 间隔内将所有账号清空并盲目重放打穿 Google 上游；当号池最短限流等待时间<= 5s时，执行真实的物理异步冷却（wait_sec * 1000 + 100ms），让限流自然恢复；当等待时间> 5s时，主动返回明确的冷却需求错误，触发客户端标准退避。 (PR #3685, Thanks to @cubelikeplayDaniel) - 两轮自适应匀速轮换与阶梯退避重试: 首轮全池轮换延迟由 50ms 放宽至 150ms 匀速轮换，留出微冷却时间；将请求级 429 快速失败门槛从单轮放宽至两轮（
pool_size * 2），并在第二轮引入温和线性阶梯退避（1.5s ~ 3.5s），兼顾突发大 Prompt 限流恢复与防长假死。实测将高并发 429 失败率从 37.4% 彻底降低至 0.0%。 (PR #3685, Thanks to @cubelikeplayDaniel) - 后台配额轮询与推理主力端点健康彻底解耦: 为
modules/quota.rs建立专属独立的QUOTA_ENDPOINT_HEALTH（60 秒观察期）并将探测超时放宽至 15 秒，坚决杜绝后台定时抓取配额与project_id探针超时污染主代理网关的推理主力端点。 (PR #3685, Thanks to @cubelikeplayDaniel) - 候选端点体积自适应看门狗与收紧 HTTP/2 空闲连接: 看门狗防假死超时按请求体体积自适应延展（保底 15 秒，大请求按 256KB/s 动态追加），防止 11MB 等超大 prompt 在上传时因固定超时被误判假死；将 HTTP/2
pool_idle_timeout收紧至 15 秒，pool_max_idle_per_host限制为 10，杜绝复用被 Google GFE 静默掐断的死连接导致流式中断。 (PR #3685, Thanks to @cubelikeplayDaniel)
- 根除 0ms 乐观重置风暴与物理异步冷却: 彻底移除
- [官方原生 Antigravity 身份深度对齐与第三方工具函数安全清洗] (Fixes #3684):
- 官方原生 Antigravity 助手身份深度对齐: 针对 Google 上游近期对非 Antigravity Harness（如 Claude Code CLI、OpenCode、DSH、Pi、Codex 等）加强语义特征风控导致频繁触发假性 429 报错的问题，在
PromptSanitizer中全面引入并对齐官方原生 Antigravity 身份声明（OFFICIAL_ANTIGRAVITY_IDENTITY）；将原先刻意保留的 Claude Code CLI / Agent SDK 等专有前导声明以及广谱归属声明，统一收敛对齐为官方原生开场白，彻底抹去外部 Harness 语义特征，享受官方原生 IDE 同等最高容忍度。 (Fixes #3684) - 第三方工具函数 (Tools) 描述文本防御性清洗: 解决第三方客户端在工具声明（
tools[].functionDeclarations）的description中混入计费伪 Header（如x-anthropic-billing-header:）或 SDK 指纹引发 WAF 拦截的隐患。在入站清洗流水线中增加对工具描述的 Header 轨防御性过滤，同时 100% 捍卫函数名称（name）与参数结构（parameters）的协议契约完好无损，确保各类自主 Agent 工具调用丝滑执行。 (Fixes #3684)
- 官方原生 Antigravity 助手身份深度对齐: 针对 Google 上游近期对非 Antigravity Harness（如 Claude Code CLI、OpenCode、DSH、Pi、Codex 等）加强语义特征风控导致频繁触发假性 429 报错的问题，在
- [流量监控报文详情重构：统一单 Tab 视窗、正文结构化审查与双链路 HAR 导出]:
- 统一 Postman 风格单 Tab 视窗架构: 将报文详情重塑为「请求 (Request)」、「转出 (Forwarded)」、「响应 (Response)」三大主选项卡，告别上下分屏挤压，配合紧凑元数据栏与防换行 Token 消耗状态，大幅提升查阅空间。
- 正文「结构化拆分」与「原始代码」双视图: 支持一键切换视图；智能拆解展示系统提示词、多轮对话气泡、工具函数声明、思考推理块与调用回执，复杂 Agent 报文清晰直观。
- 键值对表格与敏感 Header 脱敏保护: 请求头与 Query 参数升级为两列键值对表格，支持关键词搜索、悬浮复制及
Bearer sk-***掩码脱敏与明文切换。 - 标准双链路 HAR 1.2 导出 (3 报文齐全): 导出包含「客户端->网关」与「网关->上游真实端点」双链路 Entry，确保抓包工具导入时客户端请求、网关转出及响应 3 个报文 100% 齐全；支持一键复制 cURL 命令与全量 JSON。
- 全量 12 种多国语言覆盖: 新增组件、标签与提示文案全面接入并同步至 12 种系统语言。
What's Changed
- fix(proxy,retry): 彻底移除0ms乐观重置风暴、引入物理冷却退避与端点健康隔离 (Fixes #3506, Ref #3525, Ref #3634, Ref #3666, Ref #3680) by @cubelikeplayDaniel in #3685
- fix(retry): 修复 Retry-After 头部解析失败问题 by @EricZhou05 in #3687
Full Changelog: v4.9.8-beta.6...v4.9.8-beta.7