Antigravity Tools v4.9.3

• [Upstream HTTP/2 Protocol Hardening & Stream Drop Fix] Completely remove destructive HTTP/2 PING injection, restore native Chrome 123 protocol baseline, and eliminate 10s deep-thinking disconnects (Fixes #3593):
  • Root Cause Elimination for Destructive HTTP/2 PING: Protocol frame tracing confirmed that Google GFE / Cloudcode endpoints silently drop client-initiated HTTP/2 PING frames without returning PING ACKs. In v4.8.9 (PR #3571), the configured keep_alive_timeout = 10s caused local hyper2 to unilaterally terminate the connection at exactly 10s when waiting for an ACK, throwing error reading a body from connection (buffer_bytes=0). The destructive .http2(|mut h2| ...) configuration has been completely removed from UpstreamClient, eliminating the local self-abort timer.
  • Restoration of Native Chrome 123 Behavior & L4 Keep-Alive: Strictly aligned with real Chrome 123 browser protocol behavior (browsers never send HTTP/2 PING frames at the application layer), relying on standard L4 TCP Keep-Alive (60s). Seamlessly collaborates with PR #3578's downstream SSE 3s comment heartbeats (: ping\n\n) to prevent client-side idle timeouts without jeopardizing upstream stability.
  • Eliminate Connection Pool Pollution & 503 Token Pool Exhaustion: Removed idle PINGs from keep_alive_while_idle(true) that poisoned pooled connections, resolving the failure on consecutive requests and preventing cascading account lockouts that led to 503 Token pool is empty.
  • Deep Stress Testing Verification: Rigorously verified with long-generation benchmarks; models successfully sustained over 15.30s of initial silent thinking pauses and generated over 250KB across 125s+ with 100% connection reuse success.
• [Cross-Protocol SSE Byte-Level Line Buffering & Multibyte Splitting Defense] Eliminate Deserialization Failures & Stream Interruptions on CJK/Cyrillic/Emoji/Non-ASCII TCP Chunk Fragmentation (Fixes #3593):
  • SSE Collectors Upgraded with BytesMut Line Buffering: Refactored claude/collector.rs, openai/collector.rs, and gemini/collector.rs. Completely eliminated naive String::from_utf8_lossy(&chunk).lines() chunk-by-chunk splitting in favor of BytesMut accumulating byte streams with strict newline (\n) slicing.
  • Multilingual UTF-8 Boundary Splitting Protected: When CJK (3-byte), Cyrillic/Arabic (2-byte), Emoji/special symbols (4-byte), or oversized tool outputs are fragmented across TCP MTU boundaries, legacy chunk slicing produced corrupt \u{FFFD} replacement bytes that broke JSON parsing; the new architecture guarantees UTF-8 parsing only takes place on fully assembled lines. (Thanks to @Hubitski)
• [Official Claude 5.5 Live Metadata Injection, 128k Pipeline Ceiling Lift & Dynamic Scheduling] Dynamic Capability Detection, 404/403 Circuit Breaker & Evasion:
  • Authoritative Claude 5.5 Metadata Populated: Captured live traffic from Antigravity Hub to populate official_models.json with all 6 Claude 5.5 models (claude-sonnet-5-5-* and claude-opus-5-5-*), including internal IDs (MODEL_PLACEHOLDER_M400~M405), 1M context, 128k output limit, thinkingLevel 1/2/3, and vertexModelId.
  • Pipeline Inbound Ingress & Adapter 128k Ceiling Lifted: Replaced legacy 64k static truncation in InboundThinkingPipeline and protocol adapters with catalog-driven resolution, granting Claude 5.5 models up to 128,000 output tokens to prevent long reasoning answers from being truncated at ingress.
  • Account Capability Convergence & 404/403 Graceful Rotation: Introduced token_supports_advanced_model to route advanced model traffic strictly to qualified accounts. Added mark_model_unsupported to temporarily isolate unsupported accounts for 900s on 404/403 without locking out standard models, rotating seamlessly within the pool.
• [Pure Data-Driven DynamicTierRouter & Available Tier Scanning] Eliminate Phantom Models & Hardcoded Fallbacks:
  • Data-Driven Bare Model Derivation: Base models are dynamically discovered and derived from live upstream catalogs via suffix stripping (["-tiered", "-high", "-medium", "-low", "-extra-low"]), eliminating static maintenance.
  • Removed Flash High Hardcoding with Adaptive Decision Chains: Replaced hardcoded Flash high routing with adaptive tier matching (adaptive tiered -> medium -> actual available lowest tier -> first available tier), guaranteeing non-existent tiers are never synthesized.
• [Claude Cowork Deep Context Compaction & Decoupled Patch Utility] (PR #3580, PR #3582):
  • Opt-in Deep Compaction Mechanism: Implemented opt-in deep context compaction and decoupled patching utilities aligned with official 50% thresholds and latency optimization.
  • Cross-Version Lifecycle & Patch Lock Defenses: Hardened deep archive lifecycle and atomic patch locks against deadlocks; provided empirical verification across 21 official versions. (Thanks to @cubelikeplayDaniel)
• [Mini Window Account Pool Matrix & Single View Toggle] (PR #3589, PR #3591):
  • Account Pool Matrix View: Introduced multi-account status matrix and single-view toggling in Mini View, supporting one-click active account switching.
  • UI Visual & Interaction Enhancements: Automatically downgrade quota bars below 50% to warning colors, compacted multi-language toggle labels, and resolved GroupedSelect dropdown overflow near viewport edges. (Thanks to @EricZhou05, @zepeng-jin)
• [Cross-Platform Compatibility, Gateway Security & Tool Signature Pruning] (PR #3590):
  • Cross-Platform Display Fixes: Resolved WebKit DMA-BUF compatibility on Linux KDE Wayland.
  • Tool Signature Storage Governance: Enforced hard row bounds and TTL pruning on tool_signatures.db to prevent disk bloat; strictly isolated Classic client and IDE plugin processes and database paths. (Thanks to @cubelikeplayDaniel)

---

• [上游底层 HTTP/2 协议栈重构与根治假死断流] 彻底移除破坏性 HTTP/2 PING 注入，回归真实 Chrome 123 原生协议基线，终结深度思考与长上下文 10s 自杀式断流 (Fixes #3593):
  • 根因溯源与彻底移除有毒 HTTP/2 PING: 深入抓包与底层协议状态机实测证实，Google GFE / Cloudcode 生产端点对客户端主动发送的 HTTP/2 PING 帧静默丢弃、100% 不回复 PING ACK。v4.8.9 (PR #3571) 配置的 keep_alive_timeout = 10s 导致本地 hyper2 在思考静默期等不到 ACK 时于第 10 秒整准时在客户端本地主动杀死连接，抛出 error reading a body from connection (buffer_bytes=0)。现彻底移除 UpstreamClient 中破坏性的 .http2(|mut h2| ...) 配置，拔除本地等待 ACK 的自杀定时炸弹。
  • 回归 Chrome 123 原生协议行为与稳态 L4 保活: 严格对齐真实 Chrome 123 浏览器的协议行为（真实浏览器在 HTTP/2 层从不发应用层 PING），完全依赖操作系统的标准 L4 TCP Keep-Alive (60s)；同时与 PR #3578 在下游 SSE 提供的 3s 纯文本注释心跳（: ping\n\n）形成完美协同，既防止下游客户端/IDE 空闲超时，又保证上游长思考连接绝不断流。
  • 消除空闲连接池污染与根除 503 Token pool is empty 雪崩: 彻底清除后台 keep_alive_while_idle(true) 对连接池空闲连接乱发得不到 ACK 的 PING 所造成的连接坏死，消灭“连续请求两次就失败”的连接复用崩溃，避免网关因连接断开误触发重试熔断将可用账号锁死，彻底解决账号池雪崩问题。
  • 超长任务与长思考深度压测验证: 经真实凭据长任务压测，模型在经历长达 15.30 秒的首包思考完全静默期依然稳如磐石，连续跑满 125 秒（25 万字节）大生成且多轮连接池复用 100% 成功，连接零断流。
• [三大协议 SSE 字节级行缓冲加固与多字节防撕裂] 彻底解决中文/俄语/Emoji 等非 ASCII 多字节字符与长工具输出跨 TCP Chunk 拆包导致反序列化崩溃与断流 (Fixes #3593):
  • 流收集器全量升级 BytesMut 字节级行缓冲区: 重构 claude/collector.rs、openai/collector.rs 与 gemini/collector.rs。彻底废除旧版直接在单个原始数据块上调用 String::from_utf8_lossy(&chunk).lines() 的朴素切片逻辑，全面引入 BytesMut 字节流累积与 \n 定位切割。
  • 杜绝多语言 UTF-8 跨 Chunk 撕裂: 无论是中日韩（3字节）、俄语/阿拉伯语（2字节）、Emoji/特殊符号（4字节）还是超长工具调用结果，当数据恰好落在 TCP MTU 拆包边界时，旧版切片会在边界产生无效替换符 \u{FFFD} 破坏 JSON 语法；新版保证必须在接收到完整换行后才进行 UTF-8 校验与分发。 (Thanks to @Hubitski)
• [官方 Claude 5.5 实机元数据全量入库、128k 解禁与动态能力调度] 动态能力探测、404/403 单模型熔断避让与换号重试:
  • 官方 Claude 5.5 真实模型元数据全量入库: 依据官方实机流量首次为 official_models.json 注入全部 6 款 Claude 5.5（claude-sonnet-5-5-* 与 claude-opus-5-5-*）权威内部代号（MODEL_PLACEHOLDER_M400~M405）、1M 上下文上限、128k 输出上限、thinkingLevel 1/2/3 档位与 vertexModelId。
  • Pipeline 进站终审与各协议层 128k 限制解禁: InboundThinkingPipeline 终审上限保护由旧版静态截断（64k）升级为优先遵循官方目录权威设定，并为 Claude 5.5 放宽至 128,000 输出 Token，彻底解决超长生成被网关终审截断的问题。
  • 未获权限 PRO 账号 404/403 单模型临时熔断与换号重试: 在协议层与 TokenManager 落地 mark_model_unsupported。当未灰度到 Claude 5.5 的 PRO 账号报错 404/403 时，仅对该账号隔离特定模型（900s 临时避让），绝不连坐该账号调用其他正常模型，并在账号池有富余候选时自动换号重试。
• [纯数据驱动的通用 DynamicTierRouter 与可用档位扫描机制] 彻底清理假模型与硬编码，对齐官方真实可用模型:
  • 纯数据驱动的动态裸模型衍生: 严格以 Google 官方真实下发的可用模型为基准源，通过全自动通用后缀剥离扫描（["-tiered", "-high", "-medium", "-low", "-extra-low"]）动态派生裸模型。
  • 移除 Flash 强制 High 特判，严格遵循真实存在档位自适应决策链: 重构 pick_optimal_default_tier 与显式 effort 匹配防御：优先匹配自适应 tiered -> 其次 medium -> 再次真实存在的最低档 -> 极端保底仅取上游实际存在的首个档位，绝不凭空捏造不存在的档位导致上游报错。
• [Claude Cowork 深度压缩 (Deep Compact) 与原子化补丁引擎] (PR #3580, PR #3582):
  • 深度上下文压缩机制: 实现 opt-in 的 Claude Cowork 深度上下文压缩机制与解耦补丁工具，对齐官方 50% 阈值与延迟优化；
  • 跨版本生命周期与补丁锁安全: 健全深度归档链路与原子补丁锁，防范死锁；补齐覆盖 21 个官方版本的跨版本实测证据文档。 (Thanks to @cubelikeplayDaniel)
• [迷你窗口多账号矩阵与单账号切换视图] (PR #3589, PR #3591):
  • 多账号池矩阵展示与一键切换: 迷你窗口引入账号池多账号状态矩阵展示与单账号简洁视图切换，支持点击一键切换活跃账号；
  • UI 视觉与交互体验优化: 进度条低于 50% 自动降级为警示色，精简多国语言紧凑标签；修复 GroupedSelect 下拉菜单靠近视口底部时的溢出与滚动问题。 (Thanks to @EricZhou05, @zepeng-jin)
• [跨平台兼容性、网关安全与工具签名持久化治理] (PR #3590):
  • 跨平台显示修复: 修复 Linux KDE Wayland 平台下 WebKit DMA-BUF 兼容性问题；
  • 工具签名数据库治理: 加固工具签名数据库（tool_signatures.db）的行数硬上限与 TTL 清理机制，防范磁盘膨胀；严格隔离 Classic 客户端与 IDE 插件的进程与数据库路径。 (Thanks to @cubelikeplayDaniel)

What's Changed

• fix(cowork): 健全深度归档链路、自适应客户端版本探测与官方 Token 估算对齐 by @cubelikeplayDaniel in #3580
• docs(cowork): 补充 Deep Compact 核心架构 RFC 与深度开发细节规范 by @cubelikeplayDaniel in #3582
• fix(ui): 修复 GroupedSelect 下拉菜单靠近视口底部时被截断与无法滚动的溢出问题 by @zepeng-jin in #3591
• fix(cowork,proxy): harden deep compact lifecycle, eliminate 503 timeouts, isolate clients and align retry strategy by @cubelikeplayDaniel in #3584
• fix(ui): sync active account on init and resolve mini view refresh deadlock by @EricZhou05 in #3589
• fix(proxy): harden pipeline schema normalization, model versioning and tool signature pruning by @cubelikeplayDaniel in #3590
• feat(ui): support account pool matrix and single view toggle in mini window by @EricZhou05 in #3592
• fix(proxy): 修复 Claude 5.5 原始 Protobuf 思考签名识别与出站格式校验 by @cubelikeplayDaniel in #3595

New Contributors

• @cubelikeplayDaniel made their first contribution in #3580
• @zepeng-jin made their first contribution in #3591

Full Changelog: v4.9.1...v4.9.3