lbjlaq/Antigravity-Manager v5.0.4-beta.13 on GitHub

[Neutralize Upstream WAF Prompt Triggers & Decouple Model Quota Normalization] (Fixes #3965):
- Neutralize Agent Tool Directives Against Fake 429: High-precision neutralization for Google Cloud Code upstream WAF heuristic triggers targeting third-party agent templates (e.g. Kelivo, Hermes Agent) with tool guidelines (Prefer read_file, edit_file, and write_file over cat/sed), preserving fenced code blocks while completely eliminating fake 429 RESOURCE_EXHAUSTED upstream responses. (Fixes #3965)
- Decouple Flash and Pro Quota Normalization: Deprecated obsolete mapping of high-tier Flash models (e.g. gemini-3.8-flash-high) to gemini-3-pro-high, ensuring granular base model normalization and preventing 429 quota lookup mismatches from triggering cascading 30-second cooldowns across all accounts. (Fixes #3965)
- Modernize Default Monitored and Warmup Models: Upgraded default warmup, quota protection, and classic auto-switch configurations from legacy gemini-3-flash / gemini-3-pro-high to gemini-3.8-flash and gemini-3.1-pro-high. (Fixes #3965)

[上游 WAF 特征提示词中和与模型配额归一化修复] (Fixes #3965):
- 中和第三方 Agent 工具指令特征防伪 429: 针对 Google Cloud Code 上游 WAF 针对特定第三方 Agent 模板（如 Kelivo、Hermes Agent 等）的启发式指令拦截（Prefer read_file, edit_file, and write_file over cat/sed 等模式），在单趟保护代码块的前提下对非代码区域进行高精度中和，彻底杜绝上游秒级返回伪造的 429 RESOURCE_EXHAUSTED 报错。(Fixes #3965)
- Flash 与 Pro 配额归一化解耦: 废除将高阶 Flash 模型（如 gemini-3.8-flash-high）历史错配至 gemini-3-pro-high 的旧规则，按基底模型分别独立归一化，根除 429 发生时配额查找错位触发全账号 30 秒级联冷却降级的严重隐患。(Fixes #3965)
- 默认监控与预热模型现代化: 将默认预热、配额保护与经典自动切号的过时默认模型从 gemini-3-flash / gemini-3-pro-high 现代化升级为 gemini-3.8-flash 与 gemini-3.1-pro-high。(Fixes #3965)

What's Changed

fix(proxy): neutralize agent tool directives against fake 429 and decouple model quota normalization (Fixes #3965, Release v5.0.4-beta.13) by @cubelikeplayDaniel in #3967

Full Changelog: v5.0.4-beta.12...v5.0.4-beta.13