- [Multi-Client Identity Normalization & Anti-Throttling Prompt Defense] (Ref #3720: Antigravity Rate Limiting Investigation):
- Broad Identity Normalization for Mainstream AI Agents and CLIs: Driven by empirical findings from production rate-limiting and QoS demotion investigations (see Issue #3720 for full telemetry and benchmarks), addressed upstream Google's heightened semantic rate-limiting against third-party harnesses (OpenCode, OpenClaw, Grok Build, Cursor, Cascade, Cline, Aider, KiloCode, DSH, Devin, etc.), which previously demoted high-throughput accounts into degraded best-effort queues (causing TTFT to spike from ~4s to 40s–80s+ under compute pressure). Comprehensively upgraded the
PromptSanitizerengine to detect nominal preambles, host-environment declarations (running inside / within,integrated into), and natural-language Chinese patterns, harmonizing them into the official Antigravity native identity to strip external wrapper fingerprints and restore official IDE-tier fast-track QoS treatment. (Ref #3720) - Adaptive Leading Window with Plugin Metadata Penetration: Extended the system prompt header scan region from a rigid 4-sentence cap to an elastic dual window of up to 8 sentences or 1500 characters, reliably penetrating leading metadata banners (such as OpenClaw's
--- OpenClaw plugin-injected context ---) to prevent real identity declarations from being pushed out of the scanning window. (Ref #3720) - Decimal Version Boundary Lookahead & Core Prompt Immunity: Refactored attribution prefix regexes with lookahead matching for version decimals (e.g.
Grok 4.6,GPT-5.2), preventing decimals from being mistakenly treated as sentence-ending periods. Established an ingress whitelist bypass ensuring core framework prompts (JeikCode), XML delimiters, and neutral developer prompts remain 100% unmutated. (Ref #3720)
- Broad Identity Normalization for Mainstream AI Agents and CLIs: Driven by empirical findings from production rate-limiting and QoS demotion investigations (see Issue #3720 for full telemetry and benchmarks), addressed upstream Google's heightened semantic rate-limiting against third-party harnesses (OpenCode, OpenClaw, Grok Build, Cursor, Cascade, Cline, Aider, KiloCode, DSH, Devin, etc.), which previously demoted high-throughput accounts into degraded best-effort queues (causing TTFT to spike from ~4s to 40s–80s+ under compute pressure). Comprehensively upgraded the
- [多客户端身份归一化与提示词特征流控防御升级] (Ref #3720: antigravity 限速调查):
- 跨平台主流 AI Agent / CLI 身份归一化广谱纳管: 结合近期生产环境限速与低优先级降级调查（详见 Issue #3720《antigravity 限速调查》实测报告），针对 Google 上游对非官方 Antigravity Harness（OpenCode、OpenClaw、Grok Build、Cursor、Cascade、Cline、Aider、KiloCode、DSH、Devin 等）加强语义特征流控、在集群容量紧张时加速将账号打入低优先级排队慢速池的问题（排查证实首字延迟从 4s 恶化至 40s~80s+），全面重构
PromptSanitizer身份归一化引擎；扩展识别原生名谓式开场、宿主集成式声明（running inside / within,integrated into）以及中文自然语序，将其开场白统一收敛对齐为官方原生 Antigravity 身份，彻底抹除外部第三方套壳与竞品生态指纹，享受官方原生 IDE 同等白名单调度与快速通道待遇。(Ref #3720) - 自适应头部探测窗口与前置插件上下文穿透: 将系统提示词头部扫描窗口由原先固定的前 4 句扩展为前 8 句或前 1500 字符的弹性双重窗口，能够可靠穿透如 OpenClaw 等插件在开头自动注入的前置元数据/上下文垫片（
--- OpenClaw plugin-injected context ---），彻底根除因前置垫片消耗句子配额导致真正身份句被挤出扫描范围而漏洗的架构缺陷。(Ref #3720) - 版本号小数点前瞻识别与主要提示词架构级绝对保护: 优化归属动词前导与归属对象匹配正则，支持精准跨越模型版本号小数点（如
Grok 4.6、GPT-5.2），杜绝因小数点被误判为英文句号导致的断句与局部残留；在流水线入口设立最高优先级白名单守卫，确保JeikCode等核心主要提示词、XML 容器及通用无害提示词 100% 豁免原样透传，实现零误伤与架构级安全闭环。(Ref #3720)
- 跨平台主流 AI Agent / CLI 身份归一化广谱纳管: 结合近期生产环境限速与低优先级降级调查（详见 Issue #3720《antigravity 限速调查》实测报告），针对 Google 上游对非官方 Antigravity Harness（OpenCode、OpenClaw、Grok Build、Cursor、Cascade、Cline、Aider、KiloCode、DSH、Devin 等）加强语义特征流控、在集群容量紧张时加速将账号打入低优先级排队慢速池的问题（排查证实首字延迟从 4s 恶化至 40s~80s+），全面重构
Full Changelog: v4.9.9-beta.2...v4.9.9-beta.3