lbjlaq/Antigravity-Manager v4.9.2-beta.7

Antigravity Tools v4.9.2-beta.7

on GitHub

• [Official Claude 5.5 Metadata Injection & 128k Pipeline Ceiling Lift] Inject Full Claude 5.5 Live Traffic Metadata, Dynamic Inbound Pipeline & Cross-Protocol 128k Limit:
  • 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-low/medium/high, claude-opus-5-5-low/medium/high), including internal IDs (MODEL_PLACEHOLDER_M400~M405), 1M context, 128k output limit, thinkingLevel 1/2/3, and vertexModelId — establishing a verified catalog baseline for tier routing.
  • Pipeline Inbound Ingress 128k Ceiling Lifted: Replaced legacy 64k static truncation in InboundThinkingPipeline 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.
  • OpenAI & Claude Protocol Adapter Limit Relaxed: Relaxed safe_limit guards in both OpenAI-compat (openai/request.rs) and Anthropic Claude (claude/request.rs) layers to allow client requests up to 128,000 output tokens without protocol-layer capping.
• [Dynamic Capability Detection, Model-Scoped Circuit Breaker & Auto-Rotation] Quota-Driven Pool Convergence, 404 Evasion & Conversation Continuity:
  • Dynamic Capability Detection & Quota-Refresh Candidate Convergence: Introduced token_supports_advanced_model in TokenManager. Schedulers inspect models dynamically returned from periodic (or manual) quota refreshes, automatically routing advanced model traffic strictly to qualified accounts with top priority while keeping graceful fallbacks for cold starts.
  • Unentitled PRO Account 404/403 Model-Scoped Circuit Break & Rotation Loop: Implemented mark_model_unsupported across handlers (claude.rs, openai.rs, gemini.rs) and TokenManager. When unentitled accounts hit 404/403 on unrolled Claude 5.5 models, the specific model is temporarily isolated (900s) without penalizing other functional models (e.g. Claude 4.6 / Gemini 3), seamlessly rotating to other ready candidates and eliminating client-side 404 errors.
  • Non-Streaming Collection Interruption Retry Closure (Fixes #3593): Fixed a bug in collect_stream_to_json where upstream connection drops during deep thinking silence caused immediate HTTP 500 errors; now sets force_rotate = true and triggers automated rotation retry.
  • Official Catalog & Dynamic Capability Regression Tests: Added OfficialModelCatalog resolution assertions for Claude 5.5 authoritative metadata alongside test_claude_5_5_dynamic_capability_filtering_and_priority verifying capability filtering, single-model evasion, and priority sorting.

• [官方 Claude 5.5 实机元数据注入与全链路 128k 解禁] 对接官方 Antigravity Hub 最新抓包流量，打通 Inbound 进站终审与各协议层 128k 动态上限:
  • 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-low/medium/high 与 claude-opus-5-5-low/medium/high）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，为分档路由器提供真实的官方数据基准。
  • Pipeline 进站终审 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。
  • OpenAI 与 Claude 协议转译层上限放宽: 在 openai/request.rs 与 claude/request.rs 的 safe_limit 逻辑中对齐 Claude 5.5，支持客户端合法请求最高 128,000 输出 Token 不被协议层裁剪。
• [高级模型动态能力探测、未开通账号单模型熔断与重试闭环] 联动配额自动收敛优质号池，隔离 404 账号并保障对话连续性:
  • 联动配额刷新的动态能力探测与号池收敛: 在 TokenManager 引入 token_supports_advanced_model 探测机制。依据后台定时额度刷新（或手动刷新）拉取的实时模型配额，自动将高级模型请求精准导流至已开通该能力的账号，并赋予最高调度优先级；同时针对冷启动与离线状态提供平滑降级兜底。
  • 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（如 Claude 4.6 / Gemini 3），并在账号池有富余候选时自动换号重试，彻底消灭客户端 404 报错。
  • 非流式转流收集断流换号重试闭环 (Fixes #3593): 修复 collect_stream_to_json 非流式转流收集阶段在深度思考静默期发生底层断流时直接抛 500 的缺陷，设置 force_rotate = true 并触发自动换号重试。
  • 官方目录与动态能力回归测试: 补齐 OfficialModelCatalog 对 Claude 5.5 真实元数据的解析断言，以及 TokenManager 动态能力过滤、单模型熔断避让与优先级断言测试。

Full Changelog: v4.9.2-beta.6...v4.9.2-beta.7