lbjlaq/Antigravity-Manager v4.9.2-beta.5

Antigravity Tools v4.9.2-beta.5

on GitHub

• [Dynamic Tier Router & Claude 5.5 Spec Architecture Upgrade] Data-Driven DynamicTierRouter, Automated Tier Scanning, and Full Claude 5.5 Support (Inspired by PR #3594, Thanks to @CarlitoDon):
  • Pure Data-Driven Dynamic Tier Router (DynamicTierRouter): Fully decoupled from model-specific heuristics; implemented collect_tiers_for_base in OfficialModelCatalog to dynamically discover all available tier suffixes for any base model from the official upstream catalog. Whether for Gemini 3.x Flash, Claude 5.5, or future models, automatically strips suffixes to derive bare models and routes requests dynamically based on tier weight hierarchies (lite < low < default < medium < high < xhigh < max) and client effort hints.
  • Claude 5.5 Model Specs & Factory: Fully transitioned to dynamic RealModelSpec construction, dynamically fetching maxOutputTokens (128k) and thinking budgets from OfficialModelCatalog without static constant maintenance.
  • Permission Gate & Soft-Sorting: Enforced semver-based subscription gating in TokenManager (Claude >= 5.0) to filter out Free accounts, giving scheduling priority to PRO/ULTRA accounts that have synchronized the model catalog.
  • Thinking Level Normalization: Added support for snake_case levels such as extra_low, x_high, and flash_lite, seamlessly normalizing them with kebab-case counterparts.
  • Unified Thinking Budget & Default Alignment: Unified Gemini and Claude thinking configuration, aligning the default budget to -1 (representing native adaptive defaults) and simplifying settings panels.
• [Long-Generation & Deep-Thinking Stream Interruption Fix] Relax HTTP/2 Keep-Alive Timeout to 90s and Fix Claude Peek Error Detection (Fixes #3593, Thanks to @Hubitski):
  • HTTP/2 Keep-Alive Timeout Relaxed: Extended keep_alive_timeout in UpstreamClient to 90s (aligned with pool_idle_timeout) and adjusted probe interval to 5s, completely resolving 10s client connection drops during deep thinking or long generation pauses (Stream interrupted before completion / error reading a body from connection).
  • Claude Streaming Peek Error Detection: Completed error event detection (claude_stream_chunk_has_error_event) during the peek pre-read phase in claude.rs, preventing stream head errors from being mistaken for valid responses and bypassing account pool rotation.

• [通用分档路由器与 Claude 5.5 动态规格架构升级] 纯数据驱动的通用 DynamicTierRouter、可用档位自动扫描与 Claude 5.5 全链路接入 (PR #3594 启发推进, Thanks to @CarlitoDon):
  • 纯数据驱动通用分档路由器 (DynamicTierRouter): 从硬编码特判中彻底解耦，在 OfficialModelCatalog 中实现 collect_tiers_for_base，支持从官方目录动态探测任意模型的所有可用分档后缀（available_tiers 列表）。无论是 Gemini 3.x Flash 还是 Claude 5.5 或未来的新模型，自动剥离后缀派生裸模型并根据权重梯队（lite < low < default < medium < high < xhigh < max）与客户端 effort 动态智能路由。
  • Claude 5.5 模型规格与动态工厂: 全面对接 RealModelSpec 动态构建，优先从 OfficialModelCatalog 动态挂载 maxOutputTokens（128k）与思考预算，彻底告别静态常量维护。
  • 权限门禁与优先调度: 在 TokenManager 中基于语义版本（Claude >= 5.0）自动拦截 Free 账号，优先调度已同步该模型目录的 PRO/ULTRA 账号。
  • 思考等级归一化扩展: 支持 extra_low、x_high、flash_lite 等下划线蛇形命名，与连字符形态无缝兼容归一化。
  • 思考预算统一精简与默认值对齐: 统合 Gemini 与 Claude 思考配置，默认预算统一为 -1（语义为遵循官方默认自适应档位预算），精简冗余分区。
• [长生成与深度思考断流根治及流式错误识别] 放宽 HTTP/2 PING 超时至 90s 并修复 Claude Peek 错误检测 (Fixes #3593, Thanks to @Hubitski):
  • HTTP/2 保持长连接与超时放宽: 将 UpstreamClient 中底层 HTTP/2 keep_alive_timeout 放宽至 90s（与 pool_idle_timeout 对齐），探测间隔优化为 5s，彻底根除模型在深度思考或超长生成静默期时底层连接在 10s 被主动掐断断流的报错（Stream interrupted before completion / error reading a body from connection）。
  • Claude 流式 Peek 错误检测补齐: 补齐 claude.rs 的 Peek 预读阶段错误事件识别逻辑（claude_stream_chunk_has_error_event），防止流式首包错误被误判为合法正文而绕过账号池轮换重试机制。

Full Changelog: v4.9.2-beta.4...v4.9.2-beta.5