lbjlaq/Antigravity-Manager v4.9.8-beta.2

Antigravity Tools v4.9.8-beta.2

on GitHub

• [Full fetchAvailableModels Dynamic Catalog Alignment, Removal of Static Hardcoded Models & Adaptive Support] (Ref #3665):
  • Native Authorized Model Catalog (available_models): Added available_models: HashSet<String> to ProxyToken, mirroring the raw model catalog delivered live by Google's fetchAvailableModels. Acts as the single source of truth for account capabilities, enabling fully adaptive model dispatching. (Ref #3665)
  • Abolish Static Gate & 3-Step Dynamic Matching: Completely eliminated the legacy model_quotas.contains_key(normalized) veto gate. Replaced it with a 3-step dynamic matching pipeline (Step A exact matching on raw ID, Step B tier suffix stripping via strip_tier_suffix to match base model variants, Step C fallback to generic wildcard patterns only when all accounts are in cold start), completely eliminating false-positive 503 rejection on unindexed models. (Ref #3665)
  • Prune Redundant CLAUDE_TO_GEMINI Self-Mappings: Purged all redundant X -> X self-mappings for native gemini-* and claude-* models from CLAUDE_TO_GEMINI, delegating transparent routing to dynamic pass-through and DYNAMIC_MODEL_FORWARDING_RULES; kept only true cross-vendor aliases (e.g., gpt-*) and deprecated version transitions. (Ref #3665)
  • Completely Remove Specific Codename (argon) Hardcodings: Refactored is_gemini_v4_or_above in model_specs.rs into a pure numeric/semantic version parser (gemini-[4-9], >= 4.0), removing all hardcoded "argon" string checks. Upgraded get_thinking_budget and get_max_output_tokens to generic tier deductions and base-name dynamic fallbacks, enabling zero-code-change support for future frontier models. (Ref #3665)
• [Imminent Quota Floor to Eliminate False 429s & Stream Truncation Golden Triplet Contract] (PR #3666, Fixes #429, Fixes #458, Fixes #3635, Fixes #3661, Ref #3626, Ref #3639, Thanks to @cubelikeplayDaniel):
  • Imminent Quota Floor (IMMINENT_QUOTA_FLOOR): Enforced a quota floor (<= 10%) in EDF and P2C scheduling algorithms where critically depleted accounts are stripped of imminent reset deadline boosting and strictly fall back to sorting by absolute quota. Eradicates false-positive RESOURCE_EXHAUSTED errors caused by 1%~2% accounts placed at head-of-line that instantly trigger 429 and exhaust circuit-breaker retry budgets. (PR #3666, Fixes #3661, Thanks to @cubelikeplayDaniel)
  • Atomic Proxy Switch File and Index Synchronization: Refactored toggle_proxy_status to atomically sync individual account credentials and accounts.json under an ACCOUNT_INDEX_LOCK mutex, preventing state drift. (PR #3666, Thanks to @cubelikeplayDaniel)
  • Golden Triplet Termination Contract & Token Preservation: Added content generation detection (has_content()) in stream_lifecycle.rs. When deep reasoning or long-running streams encounter network disconnection or idle timeouts after generating content, the gateway gracefully emits content_block_stop, message_delta(stop_reason: end_turn), and message_stop golden triplet sequence to satisfy client completion state machines (such as Claude Code), activating client-side keepPartial preservation and eradicating token wastage and infinite re-thinking loops. (PR #3666, Fixes #429, Fixes #458, Fixes #3635, Thanks to @cubelikeplayDaniel)

• [全面接入 fetchAvailableModels 动态目录、彻底清理静态模型硬编码与新模型自适应支持] (Ref #3665):
  • 原生授权模型集合 (available_models): 在 ProxyToken 中新增 available_models: HashSet<String>，直接镜像 Google 上游 fetchAvailableModels 实时下发的授权列表。作为账号模型能力的唯一权威来源，实现模型全生命周期自适应判定。 (Ref #3665)
  • 废除静态门禁与三步自适应匹配: 彻底废除旧的 model_quotas.contains_key(normalized) 一票否决门禁，重构为动态三步匹配（Step A 精确匹配完整模型 ID、Step B 基于 strip_tier_suffix 剥离 -high/-medium/-low 档位后缀后匹配基名变体、Step C 仅在所有账号尚未刷新的冷启动期退回通用通配），彻底杜绝由于本地字典未登记新模型导致的误杀报错。 (Ref #3665)
  • 清理 CLAUDE_TO_GEMINI 冗余自映射: 移除 CLAUDE_TO_GEMINI 映射表中所有 gemini-* 与 claude-* 原生模型的自映射条目，全面由动态透传分支与 DYNAMIC_MODEL_FORWARDING_RULES 原生接管；仅保留跨供应商别名（如 gpt-*）及必要的已下线版本平滑降级。 (Ref #3665)
  • 彻底移除特定模型代号 (argon) 硬编码: 将 model_specs.rs 的 is_gemini_v4_or_above 重构为通用的纯数值语义版本解析（gemini-[4-9]、>= 4.0），彻底移除了对代号 "argon" 的硬编码判断；思考预算 get_thinking_budget 与 Token 限额 get_max_output_tokens 升级为通用分档推导与基名动态回退，未来 Google 推出任意新模型（无论名称如何）均能零代码修改自适应调度。 (Ref #3665)
• [引入濒危额度过滤消除假性 429 报错与补齐客户端断流终结契约] (PR #3666, Fixes #429, Fixes #458, Fixes #3635, Fixes #3661, Ref #3626, Ref #3639, Thanks to @cubelikeplayDaniel):
  • 濒危额度提权剥离 (IMMINENT_QUOTA_FLOOR): 在 EDF（最早截止时间优先）与 P2C 调度算法中，将剩余配额 <= 10% 的濒危账号强制剥离临期截止时间提权，严格按额度绝对值降级排序；彻底根除仅剩 1%~2% 极低额度账号因临期被排在队首、请求打过去瞬间耗尽配额触发 429 并击穿自适应重试容错导致误报 RESOURCE_EXHAUSTED 的缺陷。 (PR #3666, Fixes #3661, Thanks to @cubelikeplayDaniel)
  • 反代开关实体文件与索引原子同步: 彻底重构 toggle_proxy_status，统一由 modules::account 在 ACCOUNT_INDEX_LOCK 排他锁保护下原子同步账号实体文件与 accounts.json 索引，杜绝单向覆写引发的状态持久化漂移。 (PR #3666, Thanks to @cubelikeplayDaniel)
  • 黄金三元组终结契约与断流 Token 保全: 在 stream_lifecycle.rs 统一流生命周期中增加内容产出检测（has_content()）；长思考或深度推理中途遭遇网络抖动或空闲超时时，网关优雅补发 content_block_stop、message_delta(stop_reason: end_turn) 和 message_stop 黄金三元组终结序列，满足 Claude Code 等终端客户端完成态契约，激活客户端 keepPartial 原生保全机制，彻底杜绝丢弃已生成 Token 与无限重新思考死循环。 (PR #3666, Fixes #429, Fixes #458, Fixes #3635, Thanks to @cubelikeplayDaniel)

What's Changed

• fix(scheduler,stream): 引入濒危额度过滤消除假性429并补齐断流三元组终结契约 (Fixes #429, #458, #3635, Ref #3626, #3639) by @cubelikeplayDaniel in #3666

Full Changelog: v4.9.8-beta.1...v4.9.8-beta.2