lbjlaq/Antigravity-Manager v4.9.4

Antigravity Tools v4.9.4

on GitHub

• [Adaptive Thinking Budget Negotiation & User-Intent Recognition Machine] Eliminate Google 400 Invariant Violations, 6x Ratio Fast-Response Compression & Smart Budget Expansion (PR #3599):
  • Bypass Elimination & Ingress Invariant Ingress: Removed legacy early return Some(budget) exits in InboundThinkingPipeline, ensuring all client and gateway mode requests uniformly traverse authoritative safe_limit checks from official model catalogs.
  • 6x Ratio Fast & Concise Intent Recognition: When a client explicitly specifies m and t >= 6 * m (e.g. m = 1024, t = 8192), gateway intelligently detects the intent for an instant, concise reply without lengthy deliberation, compressing thinking budget to m - 1024 (t = 0 when m <= 1024 for instant output).
  • Smart Budget Expansion for Deep Reasoning (Preserve Model IQ): When official physical capacity allows (t + 1024 <= max), gateway refuses to truncate user reasoning, automatically expanding m = t + 1024 to preserve 100% of the thinking budget within official limits.
  • Physical Ceiling Clamping Fallback: Only when thinking budget genuinely breaches the model's physical limit (t + 1024 > max), m is anchored to max with thinking budget clamped to max - 1024 (leaving 1024 tokens for completion), strictly satisfying m > t and eliminating Google 400 INVALID_ARGUMENT.
  • Gemini Adapter Wrapper Invariant & Test Suite: Mirrored invariant enforcement in gemini/wrapper.rs and added full test suites in official_alignment_tests.rs. (Thanks to @cubelikeplayDaniel)

• [自适应思考预算协商状态机与意图识别架构] 彻底杜绝 maxOutputTokens <= thinkingBudget 引发 Google 400，科学识别 6 倍反差极速意图并智能扩充总预算保全推理智商 (PR #3599):
  • 废除提前 return 旁路与全链路贯穿终审校验: 彻底移除 InboundThinkingPipeline 中 Client 模式与 Gateway 模式的过早 return Some(budget) 旁路，统一贯穿官方模型目录权威 safe_limit 物理硬顶终审校验。
  • 6 倍反差极速意图识别 (Fast & Concise Intent): 针对客户端传了明确总输出 m 的场景，当思考预算 t >= 6 * m 时（例如用户传 m = 1024, t = 8192），精准识别用户强烈希望一句话简明直接回答、不希望被漫长思考拖慢的诉求，自动压缩思考预算至 m - 1024（当 m <= 1024 时自动归零关闭思考，实现模型秒回吐字）。
  • 常规深度推理智能自动扩充 (保全思考智商): 未触发极速意图时，若客户端传入的 m 较小但官方物理容量能容纳思考预算（t + 1024 <= max），网关坚决不阉割用户的思考能力，自动提升总输出预算至 m = t + 1024（且不超过官方上限），100% 完整保留用户的深度思考推导能力！
  • 物理天花板硬顶终审兜底: 仅当思考预算极大超限突破官方物理天花板时（t + 1024 > max），终审兜底将总预算锚定官方上限 max，思考预算收敛至 max - 1024（留出 1024 正文空间），从数学与协议层 100% 杜绝 Google HTTP 400 INVALID_ARGUMENT 报错。
  • Gemini Adapter Wrapper 终审兜底与单测覆盖: 在 gemini/wrapper.rs 增加同等协商保护，在 official_alignment_tests.rs 补充 6 倍反差压缩、自动扩充与物理兜底全套单测。 (Thanks to @cubelikeplayDaniel)

What's Changed

• fix(proxy): 修复思考预算超限截断导致的协议硬约束违背 (maxOutputTokens > thinkingBudget) by @cubelikeplayDaniel in #3599

Full Changelog: v4.9.3...v4.9.4