lbjlaq/Antigravity-Manager v4.9.2-beta.6

Antigravity Tools v4.9.2-beta.6

on GitHub

• [Data-Driven Model Discovery & Hardcoding Removal] Remove Phantom Claude 5.5 and gemini-3.x-flash Heuristics, Refactor Dynamic Derivation and Low-Tier Filtering:
  • Purge Phantom Claude 5.5 & Eliminate 503 Errors: Completely removed unreleased claude-sonnet-5-5 and claude-opus-5-5 from backend public lists (get_supported_models), frontend config presets (modelConfig.ts), and tier router fallbacks. Prevents 503 Service Unavailable errors caused by client requests to phantom models not present in upstream quota catalogs.
  • Remove gemini-3.x-flash Hardcoded Rules: Cleaned up legacy hardcoded interceptors that forcibly converted gemini-3.x-flash (x > 8) to 3.x-flash-tiered, removed hardcoded default mappings and UI badge locks, and added automatic local configuration migration. Replaced entirely with data-driven dynamic discovery via DynamicTierRouter, unlocking standard wildcard (*) mappings.
  • Data-Driven Dynamic Bare Model Derivation (Full Alignment): Strictly grounded in the 27 active models returned by Google's live endpoints (e.g. gemini-3.8-flash-tiered, gemini-3.1-pro-high/low, claude-opus-4-6-thinking, claude-sonnet-4-6, gemini-3.1-flash-image, etc.). Automatically derives clean bare models (gemini-3.1-pro, gemini-3.5-flash, gemini-3.6-flash, gemini-3.7-flash, gemini-3.8-flash) through universal suffix-stripping (["-tiered", "-high", "-medium", "-low", "-extra-low"]) subject to client effort and gateway budget control.
  • Frontend & Backend Baseline Alignment: Harmonized baseline filtering across frontend and backend; strictly filters internal test IDs (chat_20706, chat_23310), retired gemini-2.5-* models, and legacy Claude versions (< 4.6), while correctly preserving official active models (gemini-pro-agent, gemini-3-flash-agent, tab_flash_lite_preview, tab_jump_flash_lite_preview, gemini-3.1-flash-lite, and gemini-3.5-flash-lite).
• [Adaptive Bare Flash Routing & Strict Non-Empty Fallback] Eliminate Bare Flash High Heuristics and Strictly Enforce Catalog-Grounded Decision Chains:
  • Eliminate Bare Flash High Bias: Completely cleared 3 residual heuristics in the tier router, variant mapper, and budget resolver that forcibly defaulted is_bare_gemini_v3_flash to -high when effort was omitted, ensuring all bare models follow the unified adaptive hierarchy.
  • Strict Non-Empty Fallback Validation: Reconstructed pick_optimal_default_tier and effort matching defenses: prioritizes adaptive tiered -> then medium -> then lowest tier strictly greater than low -> picks low ONLY when low actually exists in available tiers -> falls back strictly to the first verified available tier (e.g. extra-low); preserves bare models unchanged when no tiers exist, eliminating fabricated -low / -high suffixes that trigger upstream errors.
  • Authoritative Gateway Budget Synchronization: Ensured custom gateway thinking budgets (> 0 or -1) strictly override upstream model struct defaults and harmonize end-to-end with dynamically derived bare models.

• [纯数据驱动模型发现与消除硬编码] 彻底清理 Claude 5.5 与 gemini-3.x-flash 假模型规则，重构全链路动态衍生与低等模型过滤:
  • 清理 Claude 5.5 幽灵模型与根除 503 报错: 从后端公开模型列表（get_supported_models）、前端配置预设（modelConfig.ts）以及分档路由器保底逻辑中彻底清除未经官方发布的 claude-sonnet-5-5 与 claude-opus-5-5。杜绝因网关虚假暴露该模型导致下游客户端调用后因账号池缺少配额而爆出 503 Service Unavailable 的问题。
  • 全面移除 gemini-3.x-flash 静态特判: 彻底移除旧版中针对 gemini-3.x-flash (x > 8) 强转为 3.x-flash-tiered 的硬编码拦截、默认映射与前端 UI 特殊胶囊标签；加入本地配置自动平滑迁移清洗。全链路统一由 DynamicTierRouter 动态探测真实上游档位自适应接管，前端放开通用通配符（*）自定义配置。
  • 纯数据驱动的动态裸模型衍生 (前后端对齐): 严格以 Google 官方真实下发的可用模型（如 gemini-3.8-flash-tiered、gemini-3.1-pro-high/low、claude-opus-4-6-thinking、claude-sonnet-4-6、gemini-3.1-flash-image 等）为基准源。在此基础上，通过全自动通用后缀剥离扫描（["-tiered", "-high", "-medium", "-low", "-extra-low"]）动态派生出受档位与网关预算控制的裸模型（如 gemini-3.1-pro、gemini-3.5-flash、gemini-3.6-flash、gemini-3.7-flash、gemini-3.8-flash 等），彻底告别静态维护。
  • 前后端基准线过滤精准对齐: 前后端统一基准线过滤机制，坚决过滤 chat_20706 / chat_23310 等内部测试 ID、淘汰的 gemini-2.5-* 旧模型以及 4.6 以下的 Claude 历史旧模型；准确放行官方活跃的 gemini-pro-agent、gemini-3-flash-agent、tab_flash_lite_preview、tab_jump_flash_lite_preview、gemini-3.1-flash-lite 与 gemini-3.5-flash-lite。
• [Flash 裸模型默认路由自适应与防无脑保底] 移除 Flash 强制 High 特判，严格遵循真实存在档位决策链:
  • 彻底移除 Flash 裸模型强制 High 特判: 清理了 is_bare_gemini_v3_flash 未传思考强度时强制路由至 -high 的 3 处硬编码残留（路由器特判、variant_mapping 默认 tier、思考预算默认判定），所有裸模型统一遵循通用自适应决策链。
  • 防无脑保底 low 与可用档位强校验: 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 -> 再次高于 low 的最低档 -> 仅当 low 真实存在于上游可用档位时才选取 low -> 极端保底仅取上游实际存在的首个档位（如 extra-low）；若上游无可用档位则原样保持裸模型，绝不凭空捏造不存在的 -low 或 -high 导致上游报错。
  • 网关预算权威覆盖全链路协同: 确保在网关权威模式下，用户为各思考档位自定义预定的思考预算（> 0 或 -1）具有最高物理优先级，严格覆盖官方结构体默认值，并与动态衍生裸模型全链路联动。

Full Changelog: v4.9.2-beta.5...v4.9.2-beta.6