- [Quota Endpoint Host Alignment & Exhausted Model 0% Fallback Fix] (Fixes #3987, PR #3988, PR #3989, Thanks to @vecyang1):
- Drop Prod Host from Quota Endpoints to Match Inference: Removed
cloudcode-pa.googleapis.com(Prod) fromQUOTA_API_ENDPOINTSandQUOTA_SUMMARY_ENDPOINTSso quota polling exclusively usesDaily -> Sandboxalongside inference (V1_INTERNAL_BASE_URL_FALLBACKS), guarded by a host-subset invariant unit test; eliminates100% / 0%Gemini quota flip-flopping and scheduling of depleted accounts into429 QUOTA_EXHAUSTEDupon transient Daily failures. (Fixes #3987, PR #3988, Thanks to @vecyang1) - Treat Omitted
remainingFractionwith ActiveresetTimeas0%: Extractedmodel_quota_percentagesoquotaInfoentries carrying a non-emptyresetTimewithoutremainingFraction(omitted by Protobuf v3 zero-value serialization) resolve toSome(0)while keeping missing/emptyquotaInfoon group-quota fallback, preventing exhausted models from displaying100%and losingresetTimeifretrieveUserQuotaSummaryfails concurrently. (Ref #3987, PR #3989, Thanks to @vecyang1)
- Drop Prod Host from Quota Endpoints to Match Inference: Removed
- [Same-Account Auto-Heal Clamping for Unsupported Extended Thinking Levels] (Fixes #3991, PR #3992, Thanks to @vecyang1):
- Clamp Extended Thinking Level to
HIGHon 400Thinking level is unsupported: Added a single-retry same-account auto-heal arm (try_clamp_thinking_level) in the Pipeline for accounts that advertise-xhigh(thinkingLevel: 100) infetchAvailableModelsbut reject it at generation time with400 INVALID_ARGUMENT: Thinking level is unsupported: THINKING_LEVEL_XHIGH. Clamps extended levels (> 3orXHIGH) toHIGHand downgrades-xhighmodel IDs to-highacross all four protocol handlers without breaking sticky-session KV Cache or tripping family-wide Hydrogen circuit breakers. (Fixes #3991, PR #3992, Thanks to @vecyang1)
- Clamp Extended Thinking Level to
- [OpenAI Adapter Sanitization Timing Fix & Codex Persona Normalization] (Ref #3981, Fixes #3982, Ref #3983, Ref #3984, PR #3986, Thanks to @cubelikeplayDaniel):
- Relocate
sanitize_gemini_payloadin OpenAI Request Mapper: Moved payload sanitization in the OpenAI request mapper to run aftercontents,tools,systemInstruction, andinject_google_search_toolare fully populated and before prefix topology alignment and stealth sideloading (aligned with the Claude adapter), preventing premature sideloading of unnormalized system prompts, duplicate identity blocks, and staleL3-Prefixhashes while preservingL2-Toolscache hot-reload support. (Ref #3981, Ref #3983, Ref #3984, PR #3986, Thanks to @cubelikeplayDaniel) - Normalize Short
CodexClient Declarations: ExtendedRE_KNOWN_CLIENT_DECLARATIONto match standalonecodexassistant/agent preambles without vendor or model attribution clauses (e.g.You are Codex, an AI coding assistant.), normalizing them into official Antigravity native identity. (Fixes #3982, PR #3986, Thanks to @cubelikeplayDaniel) - Harden Concurrent Unit Test Config Isolation: Added
TEST_CONFIG_LOCKsynchronization and explicit default configuration initialization/reset across prompt sanitizer and OpenAI mapper tests, alongside direct regression assertions for tool description sanitization and underscore identifier preservation. (Ref #3981, PR #3986, Thanks to @cubelikeplayDaniel)
- Relocate
- [配额查询端点对齐与耗尽模型 0% 兜底判定修复] (Fixes #3987, PR #3988, PR #3989, Thanks to @vecyang1):
- 配额接口移除 Prod 端点对齐推理链路: 从
QUOTA_API_ENDPOINTS与QUOTA_SUMMARY_ENDPOINTS中移除推理链路早已弃用的cloudcode-pa.googleapis.com(Prod),统一收敛为Daily -> Sandbox并新增跨模块主机集合一致性单测断言,根治 Daily 瞬态波动回退至 Prod 读取失真 Gemini 配额导致面板额度在100% / 0%间反复跳变及误调度已耗尽账号触发429 QUOTA_EXHAUSTED的问题。(Fixes #3987, PR #3988, Thanks to @vecyang1) - 识别 Protobuf 省略零值字段的已耗尽
quotaInfo: 新增model_quota_percentage解析逻辑,当fetchAvailableModels下发的quotaInfo携带非空resetTime但因 Protobuf v3 默认零值省略机制缺失remainingFraction时精准判定为0%(对缺失quotaInfo或无有效resetTime的模型仍维持回退分组配额),防止同轮retrieveUserQuotaSummary失败时将已耗尽模型误判为100%并丢失重置时间。(Ref #3987, PR #3989, Thanks to @vecyang1)
- 配额接口移除 Prod 端点对齐推理链路: 从
- [扩展思考档位不支持(400 XHIGH)同账号降档自愈] (Fixes #3991, PR #3992, Thanks to @vecyang1):
- 400
Thinking level is unsupported单次钳制HIGH原地重试: 针对部分账号虽在目录中声明支持gemini-4-hydrogen-xhigh(thinkingLevel: 100)但推理时返回400 INVALID_ARGUMENT: Thinking level is unsupported: THINKING_LEVEL_XHIGH的问题,在 Pipelinetry_auto_heal中新增同账号单次自愈分支(try_clamp_thinking_level),将扩展档位(数值> 3或XHIGH)安全钳制为HIGH并将-xhigh模型名同步降为-high后原地重试(四协议入口重建请求时同步保持钳制状态),既避免整轮对话失败与跨账号丢失 KV Cache,也避免触发模型熔断连坐冻结全系 Hydrogen 档位。(Fixes #3991, PR #3992, Thanks to @vecyang1)
- 400
- [OpenAI 适配器请求清洗时序修正与 Codex 身份拟态补齐] (Ref #3981, Fixes #3982, Ref #3983, Ref #3984, PR #3986, Thanks to @cubelikeplayDaniel):
- 修正 OpenAI 适配器
sanitize_gemini_payload执行时序: 将 OpenAI 兼容请求映射器中的提示词与工具描述清洗节点下移至contents、tools、systemInstruction及inject_google_search_tool全量装配完成之后、前缀拓扑对齐与隐身侧载执行之前(与 Claude 适配器一致),消除隐身侧载引擎与L3-Prefix缓存哈希读取未归一化原文导致的误侧载及重复身份隐患,同时保留L2-Tools缓存的规则热更新能力。(Ref #3981, Ref #3983, Ref #3984, PR #3986, Thanks to @cubelikeplayDaniel) - 补齐
Codex短开场白原生身份拟态: 在RE_KNOWN_CLIENT_DECLARATION中增补独立codex助手声明匹配分支,将不携带厂牌或模型归属后缀的短开场白(如You are Codex, an AI coding assistant.)无缝归一化为官方 Antigravity 原生身份。(Fixes #3982, PR #3986, Thanks to @cubelikeplayDaniel) - 并发单测配置锁隔离与边界断言加固: 为提示词清洗引擎与 OpenAI 映射器单元测试补齐
TEST_CONFIG_LOCK互斥锁及默认配置初始化复位,补充工具描述脱敏与下划线标识符保留的直接回归测试。(Ref #3981, PR #3986, Thanks to @cubelikeplayDaniel)
- 修正 OpenAI 适配器
What's Changed
- fix(proxy): restore MCP namespace fields for Codex Responses (Fixes #3979) / 修复 Codex Responses 的 MCP namespace 字段 by @Lucas-Qh-Lai in #3980
- fix(oauth): request aicode scope for AGY CLI 1.5.x by @ThermalEng in #3978
- fix(proxy): add full support for Antigravity GCP PAYGO tier and restore project ID commands (Fixes #3960) by @cubelikeplayDaniel in #3970
- fix(proxy): fix OpenAI adapter sanitization timing and normalize Codex client persona (Fixes #3981, #3982, #3983, Ref #3984) by @cubelikeplayDaniel in #3986
- fix(quota): 配额端点移除 Prod,与推理端点保持一致 (Fixes #3987) by @vecyang1 in #3988
- fix(quota): quotaInfo 省略 remainingFraction 时按 0 处理,避免耗尽模型显示 100% (Ref #3987) by @vecyang1 in #3989
- fix(proxy): 400 "Thinking level is unsupported" 时同账号钳制为 HIGH 重试 (Fixes #3991) by @vecyang1 in #3992
New Contributors
- @Lucas-Qh-Lai made their first contribution in #3980
Full Changelog: v5.0.4-beta.14...v5.0.4-beta.15