- [Stealth Sideloading Execution Mandate Anchor & Instruction Following Fix] (Ref #3997, Thanks to @mqrkin):
- Inject [SYSTEM MANDATE] Directive to Eliminate Attention Drift and Rule Forgetting: Addressed scenarios where complex agents (e.g. Cursor, Cline, agy) with massive custom system instructions experienced reduced prompt compliance (such as indiscriminate
ls -Rcalls or ignoring business constraints) when triggering "Stealth Sideloading". Injected an authoritative[SYSTEM MANDATE]anchor at the head of the<system-instruction priority="critical" role="developer">container. Fully retains 100% of the upstream Google WAF keyword evasion and fake 429 neutralization defenses while forcefully commanding model attention and compliance over all custom directives. (Ref #3997, Thanks to @mqrkin)
- Inject [SYSTEM MANDATE] Directive to Eliminate Attention Drift and Rule Forgetting: Addressed scenarios where complex agents (e.g. Cursor, Cline, agy) with massive custom system instructions experienced reduced prompt compliance (such as indiscriminate
- [Gemini Tool Nested Multi-Dimensional Array Schema Invariant Enforcement] (Fixes #4001, Thanks to @Allenalonel):
- Enforce
type: arrayon Nested Schema Nodes Containingitems: Fixed an issue where client tool declarations with multi-dimensional arrays or matrix parameters (e.g.commandArguments.items.items) omitted explicittype: "array"on intermediateitemsnodes, causing Google's Protobuf Schema validator to reject requests with400 INVALID_ARGUMENT (field predicate failed: $type == Type.ARRAY). Implemented bidirectional invariant enforcement withinclean_json_schema: any schema node containing anitemsproperty is strictly guaranteed to havetype: "array"(subsequently normalized to uppercaseARRAYby the Pipeline), preventing upstream 400 rejections when invoking Gemini models with complex tool schemas. (Fixes #4001, Thanks to @Allenalonel)
- Enforce
- [CacheFirst Sticky Session Unbinding and Failover on Sustained 503/Capacity Exhaustion] (Fixes #4002, Thanks to @vecyang1):
- Automatic Sticky Session Abandonment and Rotation on Retries Exhaustion: Fixed an issue where in
CacheFirstscheduling mode, when a session's bound account encounters persistent503 MODEL_CAPACITY_EXHAUSTEDor server errors (500/503/529) and exhausts its in-place retries, the decision hardcodedabandon_sticky_account: false, causing the session to remain permanently deadlocked to the failing account and starving client requests (especially long autonomous workflows like/goal) for over 50+ minutes. The pipeline now explicitly setsabandon_sticky_account = trueupon exhausting in-place retries and transitions toTerminalNextStep::Rotatewhen healthy candidates exist in the pool (pool_size > 1), achieving immediate, seamless failover to healthy accounts. (Fixes #4002, Thanks to @vecyang1)
- Automatic Sticky Session Abandonment and Rotation on Retries Exhaustion: Fixed an issue where in
- [Mitigate Capacity Exhaustion Auto-Heal Ping-Ponging & Tier Invariant Fix] (Fixes #4000, Thanks to @rainknight27):
- Mutually Exclusive Separation Between Capacity Exhaustion and Budget Degradation: Resolved the issue where upstream
503 MODEL_CAPACITY_EXHAUSTEDfor requests already at-hightier bypassed promotion (already maxed) and inadvertently slipped into the genericis_server_overloadedbranch, getting erroneously downgraded to-mediumand then promoted back to-high, causing an oscillatory ping-pong loop across up to 8 attempts. Enforced explicit mutual exclusion intry_auto_heal(Pipeline First), blocking thinking budget degradation when upstream reports capacity exhaustion or after tier promotion, mitigating excessive retries, preventing upstreamOVERLOADED_TOO_MANY_RETRIES_PER_REQUESTflags, and faithfully retaining the client's original tier in error responses. (Fixes #4000, Thanks to @rainknight27) - Preserve Tier Promotion Invariant Across Multi-Protocol Handler Retries: Updated request repackaging paths across OpenAI, Claude, and Gemini adapters to detect and inherit
retried_with_capacity_promotedstatus, ensuring promoted requests are never silently overwritten or downgraded back tomediumon account switching or backoff retries. (Fixes #4000, Thanks to @rainknight27)
- Mutually Exclusive Separation Between Capacity Exhaustion and Budget Degradation: Resolved the issue where upstream
- [Official Hub Client Fingerprint Baseline Upgrade & Dynamic Detection Fix] (Fixes #3998, Thanks to @uselessbug):
- Upgrade Stable Baseline to Official 2.33.0 (CL 1023160026): Advanced default egress Antigravity Hub User-Agent baseline to Google official release
2.33.0(Piper Changelist1023160026). Ensures version and CL pairs remain strictly co-calibrated, eliminating upstream 404/403 gate rejections caused by outdated fingerprint baselines. - Streaming Co-Calibrated CL Extraction from Local Installations: Implemented
extract_cl_from_binaryandtry_detect_local_hub_fingerprintinhub_version, leveraging high-throughput chunked scanning to automatically extract the Blaze/Piper build Changelist directly from the installedlanguage_servercomponent. Whenever an official client is locally installed, outbound signatures automatically align with the genuine local client (source: local). - Fix False "Synced" Success Messages and Clarify Detection Guidance: Resolved the regression introduced in
v5.0.4-beta.12where clicking "Detect" bypassed remote checks and unconditionally displayed a success toast. Real-time detection now queries official/releasesand differentiates results accurately: if a newer upstream version without verified CL is detected (e.g. future 2.34.0), the UI clarifies that the stable baseline is safely maintained and provides clear update hints; if an official local client is installed, it syncs immediately without risk of Frankenstein User-Agent WAF bans. (Fixes #3998, Thanks to @uselessbug)
- Upgrade Stable Baseline to Official 2.33.0 (CL 1023160026): Advanced default egress Antigravity Hub User-Agent baseline to Google official release
- [隐身侧载执行权威锚点强化与模型指令服从度修复] (Ref #3997, Thanks to @mqrkin):
- 注入 [SYSTEM MANDATE] 权威执行锚点消除降智与遗忘规则: 针对复杂 Agent(如 Cursor/Cline/agy 等)携带大型自定义系统指令(数万字符)在触发「隐身侧载(Stealth Sideloading)」时,模型在首轮 user 消息中注意力权重不足导致盲目执行
ls -R、遗忘业务规则或编造事实的现象,在<system-instruction priority="critical" role="developer">容器首行注入强指令执行锚点([SYSTEM MANDATE])。在 100% 完整保留对抗 Google WAF 敏感词风控与伪 429 拦截能力的同时,强力唤醒模型对自定义规则与工具策略的最高服从度。(Ref #3997, Thanks to @mqrkin)
- 注入 [SYSTEM MANDATE] 权威执行锚点消除降智与遗忘规则: 针对复杂 Agent(如 Cursor/Cline/agy 等)携带大型自定义系统指令(数万字符)在触发「隐身侧载(Stealth Sideloading)」时,模型在首轮 user 消息中注意力权重不足导致盲目执行
- [Gemini 工具多维嵌套数组 Schema 铁律双向守卫修复] (Fixes #4001, Thanks to @Allenalonel):
- 强制补全嵌套 items 节点的
type: array声明: 针对客户端工具声明中包含多维数组/参数矩阵(如commandArguments.items.items)但中间层items节点遗漏type: "array"声明,导致请求进入 Google Gemini API 时被上游 Protobuf 严格校验器拦截并报错400 INVALID_ARGUMENT (field predicate failed: $type == Type.ARRAY)的问题,在clean_json_schema管道中建立终态双向约束守卫。对任何持有items属性的节点强制确保其类型为array(并由流水线大写归一化为ARRAY),彻底消除各类第三方 Agent/CLI 声明复杂工具调用时的 400 崩溃。(Fixes #4001, Thanks to @Allenalonel)
- 强制补全嵌套 items 节点的
- [CacheFirst 粘性会话 503 持续故障/容量枯竭防死锁解绑与轮换] (Fixes #4002, Thanks to @vecyang1):
- 原地重试耗尽主动解绑粘性会话与故障转移: 解决在
CacheFirst缓存优先模式下,当单个绑定账号因上游持续返回503 MODEL_CAPACITY_EXHAUSTED或服务端故障(500/503/529)原地重试耗尽时,因裁决计划硬编码abandon_sticky_account: false导致会话永久死锁绑定在故障账号上、客户端(特别是/goal等长自动化任务)外层连续重试数十次长达 50+ 分钟饿死的问题。在流水线裁决中,同账号原地重试耗尽后主动标记abandon_sticky_account = true解除绑定,并在多账号池(pool_size > 1)下推进换号轮换(TerminalNextStep::Rotate),实现秒级平滑故障转移至池内健康账号。(Fixes #4002, Thanks to @vecyang1)
- 原地重试耗尽主动解绑粘性会话与故障转移: 解决在
- [上游容量枯竭自愈防乒乓震荡与档位保护修复] (Fixes #4000, Thanks to @rainknight27):
- 容量枯竭分支与思考预算降级分支互斥隔离: 解决上游返回
503 MODEL_CAPACITY_EXHAUSTED时,初始为high档位的请求因无法进一步提档而误落入普通is_server_overloaded过载分支被错误降级为medium、进而触发二次提档导致在high ↔ medium之间反复拉扯乒乓的问题。在自愈流水线(Pipeline First)中显式增加互斥阻断条件,对已升档或容量枯竭场景严禁执行思考预算降级,彻底阻断无效重试与上游OVERLOADED_TOO_MANY_RETRIES_PER_REQUEST标记,确保错误体如实反映用户请求档位。(Fixes #4000, Thanks to @rainknight27) - 四协议入口重构请求体提档状态保持: 在 OpenAI、Claude 与 Gemini 适配器换号重试与退避重新包装请求体时,同步感知并继承
retried_with_capacity_promoted提档状态,杜绝已提档请求在后续轮次中被无条件强行冲刷回medium档位。(Fixes #4000, Thanks to @rainknight27)
- 容量枯竭分支与思考预算降级分支互斥隔离: 解决上游返回
- [官方 Hub 客户端指纹基准升级与动态探测真同步修复] (Fixes #3998, Thanks to @uselessbug):
- 内置稳态基准升级至官方最新 2.33.0 (CL 1023160026): 将出站请求原生 User-Agent 稳态基线提升至 Google 官方 Antigravity Hub 最新版本
2.33.0(Piper 构建变更号1023160026),严格保证版本号与 CL 成对同源校准,消除旧基线落后导致的客户端门禁隐患。 - 本地安装客户端二进制同源 CL 流式提取: 在
hub_version模块中实现extract_cl_from_binary与try_detect_local_hub_fingerprint,通过高吞吐流式块扫描从本地已安装的官方language_server核心组件中自动提取完整的构建 Changelist 编号。只要本地安装了官方客户端,无需用户手动维护即可一键同步出站指纹(标记来源为local本地应用)。 - 修复设置页「动态探测」假同步与分流精准提示: 彻底修复
v5.0.4-beta.12以来点击「动态探测」未发起远程拉取且无条件提示“同步成功”的误导性问题。探测时真实拉取官方/releases权威发布源,并根据本地安装、官方发布动态及稳态基准严格分流展示:若检测到官方发布了未经 CL 人工校准的新版本(如未来的 2.34.0),前端明确提示安全沿用当前基线并在设置区展示温馨指引条;若本地已安装官方最新客户端则自动完成同步,杜绝弗兰肯斯坦拼凑 UA 触发 WAF 封号。(Fixes #3998, Thanks to @uselessbug)
- 内置稳态基准升级至官方最新 2.33.0 (CL 1023160026): 将出站请求原生 User-Agent 稳态基线提升至 Google 官方 Antigravity Hub 最新版本
Full Changelog: v5.0.4-beta.17...v5.0.4-beta.18