lbjlaq/Antigravity-Manager v4.9.5-beta.4

Antigravity Tools v4.9.5-beta.4

on GitHub

• [Thinking Budget Root-Cause Fix & Model Suffix Absolute Priority] Eliminate 32768 Overwrite, Enforce Named Suffix Priority & Migrate Legacy Configs (PR #3611, Fixes #3610):
  • Enforce Named Model Suffix Absolute Priority: Established absolute priority for explicit suffixes (-low, -medium, -high) in resolve_custom_budget, restricting client effort mapping strictly to bare models (is_bare), preventing named models (such as gemini-3.8-flash-medium 4000, gemini-3.8-flash-low 1000) from being overridden to inflated 32768 budgets by client client_effort: high. (Thanks to @cubelikeplayDaniel)
  • Restore Default Mode Fallback Branch: Added missing flash_mode == ThinkingBudgetMode::Default interception for named non-tiered models, allowing gateway default mode to seamlessly fall back to official catalog preset values (Medium 4000, Low 1000, High -1). (Thanks to @cubelikeplayDaniel)
  • Decouple Claude Adapter (Pipeline First): Removed .or_else(|| tb_config.effort.as_ref()) global fallback from the Claude protocol adapter, ensuring adapters purely pass through client parameters and leaving arbitration entirely to the pipeline. (Thanks to @cubelikeplayDaniel)
  • Converge Inbound Capacity Expansion: Eliminated two premature and redundant min_overhead (+8192) expansions from early branching, converging all headroom coordination into the single-point terminal invariant state machine. (Thanks to @cubelikeplayDaniel)
  • Smooth Legacy 32k Config Migration: Introduced thinking_budget_32k_legacy_migrated startup migration to reset legacy default values (32768/16384) back to official adaptive -1 without affecting custom configurations. (Thanks to @cubelikeplayDaniel)

• [网关思考预算根治与模型后缀绝对优先级保障] 根治思考预算无差别覆盖 32768、确立具名后缀最高优先级并平滑迁移旧版配置 (PR #3611, Fixes #3610):
  • 确立模型后缀绝对最高优先级: 在 resolve_custom_budget 中建立显式后缀（-low、-medium、-high）绝对优先级判定，仅当纯净无后缀裸模型（is_bare）时才允许客户端 effort 映射，彻底阻断具名中低档位模型（如 gemini-3.8-flash-medium 4000、gemini-3.8-flash-low 1000）被客户端 client_effort: high 越权覆盖为 32768 满血预算。 (Thanks to @cubelikeplayDaniel)
  • 补齐 Default 模式回退分支: 为具名非 Tiered 模型补齐 flash_mode == ThinkingBudgetMode::Default 拦截分支，使网关默认模式能无损回退到官方模型结构体权威默认值（如 Medium 4000、Low 1000、High -1）。 (Thanks to @cubelikeplayDaniel)
  • 落实 Pipeline First 协议适配层解耦: 彻底移除 Claude 协议适配器中 .or_else(|| tb_config.effort.as_ref()) 全局配置脑补行为，适配器仅纯粹透传客户端原始参数，统一由流水线集中仲裁。 (Thanks to @cubelikeplayDaniel)
  • 收敛 Inbound 流水线上限扩充逻辑: 移除模式分流中两处超前且重复的 min_overhead (+8192) 强制扩充，所有预算与容量协商统一收敛至 Inbound 尾部的「终审上限保护与不变量协调状态机」单点裁决。 (Thanks to @cubelikeplayDaniel)
  • 旧版出厂脏配置平滑迁移: 新增 thinking_budget_32k_legacy_migrated 启动迁移逻辑，在应用启动时自动将 Default 模式下残留的历史 32768/16384 脏数据重置为官方自适应值 -1，彻底消除历史遗留锁定。 (Thanks to @cubelikeplayDaniel)

What's Changed

• fix(pipeline): 彻底根治思考预算统一覆盖为32768根因并保证后缀与默认模式优先级 (#3610) by @cubelikeplayDaniel in #3611

Full Changelog: v4.9.5-beta.3...v4.9.5-beta.4