- [Zero-Tampering Finish & Usage Preservation for Thought-Only Streams] (Fixes #4008):
- Eliminate Synthetic "task ready" Injection on Auto-Heal Collapse: Resolved an issue where the streaming gate injected a fallback
"task ready"text when an upstream model produced only thinking blocks without visible content and subsequent auto-heal continuation failed, violating downstream agent expectations. - Canonical STOP Chunk & Token Usage Preservation: Adheres strictly to the gateway's zero-content-tampering contract by avoiding any synthetic text injection; instead constructs a canonical
finishReason: "STOP"candidate chunk followed by[DONE], accurately preserving tokenusageMetadataaccumulated across the stream. (Fixes #4008) - Transparent Monitor & HAR Diagnostics: Guarantees that proxy traffic monitors and HAR export files faithfully record upstream thinking blocks alongside real empty content, facilitating upstream behavior diagnosis and debugging.
- Eliminate Synthetic "task ready" Injection on Auto-Heal Collapse: Resolved an issue where the streaming gate injected a fallback
- [Accurate Stream Truncation Semantics & Capacity Promotion Boundary Guard] (Fixes #4006, PR #4007, Thanks to @cubelikeplayDaniel):
- Restrict Empty-Thought Streaming Auto-Heal to
finishReason: "STOP": Fixed the streaming auto-heal gate intercepting all thought-containing chunks indiscriminately. Auto-heal continuation is now strictly triggered only when a thought-only stream prematurely ends withSTOP; normal terminations such as token truncation (MAX_TOKENS) or safety filters (SAFETY) pass through untouched, preserving standard upstream truncation semantics. (Fixes #4006, PR #4007, Thanks to @cubelikeplayDaniel) - Capacity Exhaustion Promotion Boundary Guards: Added strict model family checks in
try_promote_to_high_capacity, preventing invalid tier promotion for lightweight models (flash-lite), image models (imagen), embedding models, and requests already operating at elevated thinking tiers (-high,-xhigh,-max,-extreme), eliminating upstream 400 parameter validation errors. (Fixes #4006, PR #4007, Thanks to @cubelikeplayDaniel)
- Restrict Empty-Thought Streaming Auto-Heal to
- [Degraded Empty-Thought Record Rejection & Gemini 4+ Quota Migration to Neon] (Fixes #3997, Fixes #3745, PR #4005, Thanks to @cubelikeplayDaniel):
- Reject Degraded Empty-Thought Cache Records: Enhanced
ThinkingStorevalidation to reject and discard degraded empty-thought cache records, preventing corrupted signatures from polluting session stores and causing multi-turn context drift. (Fixes #3997, PR #4005, Thanks to @cubelikeplayDaniel) - Forward-Compatible Gemini 4+ Dynamic Codename Extraction & Neon Quota UI: Implemented regex-based codename extraction (
extractGemini4Codename) to dynamically support future models (neon, carbon, etc.) without manual code changes while filtering generic role tokens; migrated quota capsule rendering, account tables, and badge views to display Neon labels, supported dual live-limit detection for both Neon and Argon, and enforced multi-tier flagship eligibility to eliminate false-positive badges on Free-tier accounts. (Fixes #3745, Ref #3997, PR #4005, Thanks to @cubelikeplayDaniel)
- Reject Degraded Empty-Thought Cache Records: Enhanced
- [Clear 403 Button Visibility Gate & Transparent Warmup Error Surfacing] (Ref #3995, PR #4010, Thanks to @vecyang1):
- Precise 403 Recovery Action Visibility Guard: Extracted
canClearForbiddenLockinaccountStatus.ts, strictly excluding disabled accounts (disabled: true, e.g. OAuthinvalid_grant), upstream banned accounts (banned), and pure manual proxy-disabled accounts, preventing invalid accounts from mistakenly showing the "Clear 403 Lock & Re-probe" action inAccountErrorDialog. (Ref #3995, PR #4010, Thanks to @vecyang1) - Overreach Defense & Complete Validation Lock Cleanup in Backend: Updated
clear_account_forbidden_lockinaccount.rsto rejectdisabledaccounts and preserve intentionalmanual_proxy_disabledflags; fully resetsvalidation_blocked = falseandvalidation_url = None, preventing temporary validation locks from being misinterpreted as permanent locks. (Ref #3995, PR #4010, Thanks to @vecyang1) - Transparent Probe Error Reporting & Headless API Parity: Resolved error swallowing in
handleClearAndWarmUpby surfacing real failure causes viaformatWarmUpErrorDetailacross 12 languages; added thePOST /api/accounts/:accountId/clear-forbiddenroute in Axum server and mapped it inrequest.tsfor full Headless/Web API parity. (Ref #3995, PR #4010, Thanks to @vecyang1)
- Precise 403 Recovery Action Visibility Guard: Extracted
- [Antigravity IDE Seamless Hot-Switch & Onboarding Terms Suppression]:
- IDE Client Seamless Hot-Switch with Supervisor Auto-Respawn: Extended the hot-switch architecture to Antigravity IDE. When the IDE is active, credentials are safely dual-written to SQLite
state.vscdb, macOS Keychain, and local standalone token files, while terminating only the backgroundagy/language_serversubprocess attached to the IDE process tree. The IDE's built-in supervisor automatically respawns the AI engine within ~1.5s, completely eliminating window restarts, buffer loss, and screen flickering; seamlessly downgrades to a full restart for older IDE builds without decoupled subprocesses or on respawn timeouts, and boots the IDE automatically when switching while the IDE is offline. - SQLite Onboarding Terms Alignment & Modal Dialog Suppression: Updated SQLite
state.vscdbinjection to persistagy.onboarding.lastLoginUsername,antigravityOnboarding, and TOS flags, completely neutralizing repetitive "Security Notice & Data Use" onboarding modal cards on newly switched accounts.
- IDE Client Seamless Hot-Switch with Supervisor Auto-Respawn: Extended the hot-switch architecture to Antigravity IDE. When the IDE is active, credentials are safely dual-written to SQLite
- [Warmup Transparency Notice & Small-Token Proxy Request Dialogs] (Ref #4009):
- Full Warmup Confirmation Dialog Coverage & Mechanism Transparency: Added explicit confirmation notices across single-account manual warmup, batch warmup, and full warmup in the Accounts page, clarifying that warmup sends a small-token request through the reverse proxy to activate sessions and quota cycles; added an accompanying descriptive note under "7-Day Weekly Reset Warmup" in Settings. (Ref #4009)
- [纯思考空回复流式收尾零篡改与真实用量保全] (Fixes #4008):
- 移除流式门禁自愈未果注入假正文
"task ready": 解决在纯思考输出场景下,当上游流式响应因仅包含思考块、未产出可见正文且二次续写自愈未果时,网关在终止前强行注入"task ready"假正文篡改真实响应导致下游 Agent 逻辑受扰的问题。 - 零篡改下发标准终止帧与用量保全: 严格遵循网关零正文篡改规范,在无有效正文产出时不下发任何虚构内容,改为构建携带规范
finishReason: "STOP"的候选帧与[DONE]终态终止帧,并在终止帧中完整保留流式阶段积累的usageMetadataToken 统计指标。(Fixes #4008) - 流量日志与 HAR 抓包完整可视: 监控与 HAR 导出中如实记录思考块(
thinking)与真实空正文,便于开发者清晰诊断上游模型表现与流式通信细节。
- 移除流式门禁自愈未果注入假正文
- [流式截断语义判定精准化与容量枯竭升档边界防护] (Fixes #4006, PR #4007, Thanks to @cubelikeplayDaniel):
- 流式思考空回复判定精准收敛至
finishReason: "STOP": 修复流式自愈门禁无差别拦截所有包含思考但尚未输出正文的终止帧的问题。精准判定仅当候选帧显式以STOP异常截断时才暂缓发射[DONE]启动自愈;对因达到最大输出限制(MAX_TOKENS)或安全拦截(SAFETY)等合法流式终止原因原样透传,保留正常的截断语义与终止帧。(Fixes #4006, PR #4007, Thanks to @cubelikeplayDaniel) - 容量枯竭升档边界安全守卫: 在
try_promote_to_high_capacity提档逻辑中增加模型系列守卫,严禁对不支持-high档位的轻量模型(flash-lite)、图像生成模型(imagen)、向量嵌入模型(embedding)以及已处于高思考档位(-high、-xhigh、-max、-extreme)的请求执行盲目提档,杜绝无效提档引发上游 400 参数校验崩溃。(Fixes #4006, PR #4007, Thanks to @cubelikeplayDaniel)
- 流式思考空回复判定精准收敛至
- [思考缓存签名污染拦截与 Gemini 4+ 配额动态迁移 Neon] (Fixes #3997, Fixes #3745, PR #4005, Thanks to @cubelikeplayDaniel):
- 拒绝退化空思考记录防止会话签名污染: 在
ThinkingStore存取管道中增强签名合法性校验,主动丢弃降级或异常为空思考的缓存条目,防止空思考指纹污染会话存储引发后续多轮对话的签名错位与上下文断裂。(Fixes #3997, PR #4005, Thanks to @cubelikeplayDaniel) - Gemini 4+ 代号动态正则提取与 Neon 配额界面迁移: 实现前向兼容的代号动态提取引擎
extractGemini4Codename,自动兼容未来模型代号(neon、carbon 等)并过滤通用角色标识;将配额分组胶囊、账号表格卡片等 UI 视图平滑迁移至 Neon 标识与动态旗舰标签,支持 Neon 与 Argon 双旗舰实时限额探测,并增加多层级旗舰资格校验防止免费层级账号误标旗舰徽章。(Fixes #3745, Ref #3997, PR #4005, Thanks to @cubelikeplayDaniel)
- 拒绝退化空思考记录防止会话签名污染: 在
- [非 403/停用账号清除 403 按钮显隐门禁与真实测活错误透出] (Ref #3995, PR #4010, Thanks to @vecyang1):
- 前端清 403 按钮显隐精准门禁: 抽取
canClearForbiddenLock,严格排除已停用(disabled,如 OAuthinvalid_grant)、已封禁(banned)及纯手动/非 403 停代账号,杜绝失效账号误显「清除 403 标记并重新测活」按钮。(Ref #3995, PR #4010, Thanks to @vecyang1) - 后端
clear_account_forbidden_lock越权防护与验证锁彻底清理: 明确拒绝disabled账号,严禁清除用户主动设置的manual_proxy_disabled停代意图;同时完整重置validation_blocked = false与validation_url = None,根除因残留标记导致临时验证锁被前端视作永久死锁的问题。(Ref #3995, PR #4010, Thanks to @vecyang1) - 测活失败真实原因透出与 Headless API 补齐: 修复
handleClearAndWarmUp丢弃真实错误并一律提示排查代理的问题,新增formatWarmUpErrorDetail与 12 种语言错误详情模板;并在后端补充POST /api/accounts/:accountId/clear-forbidden路由与request.ts映射,实现 GUI 与 Headless/Web 模式功能对齐。(Ref #3995, PR #4010, Thanks to @vecyang1)
- 前端清 403 按钮显隐精准门禁: 抽取
- [Antigravity IDE 无感热切号与条款声明免弹窗支持]:
- IDE 客户端无感热切号支持与 supervisor 自愈闭环: 将经典原生版的热切号架构扩展至 Antigravity IDE。当 IDE 处于运行态时,优先将凭据双写至 SQLite
state.vscdb、系统 Keychain 与本地凭据,仅安全终止挂靠在 IDE 树下的后台agy/language_server语言服务子进程,由宿主 supervisor 原地拉起并挂载新凭据,彻底消除此前因杀死 IDE 主窗口导致的会话中断、未保存缓冲丢失与窗口闪烁问题;针对无解耦子进程的旧版 IDE 或拉起超时场景,自动平滑降级为完整安全重启,未运行状态下则自动唤醒启动 IDE 客户端。 - SQLite 安全条款确认状态对齐与免弹窗守护: 在注入 IDE
state.vscdb数据库时,同步写入agy.onboarding.lastLoginUsername、antigravityOnboarding与 TOS 状态,彻底消除切换账号后 IDE 内嵌页面反复弹出“Security Notice & Data Use”条款向导卡片的问题。
- IDE 客户端无感热切号支持与 supervisor 自愈闭环: 将经典原生版的热切号架构扩展至 Antigravity IDE。当 IDE 处于运行态时,优先将凭据双写至 SQLite
- [账号手动预热弹窗确认与反代小额 Token 提示] (Ref #4009):
- 预热确认弹窗全场景覆盖与机制透明化: 在账号管理页中为单账号手动预热、批量勾选预热与全量预热弹窗统一增加明确提示,告知使用预热功能将通过反向代理发起一次小额 Token 的请求以激活会话与周期,避免误触与对底层原理的误解;并在设置页「7天周配额智能预热」卡片下方补充对应技术说明。(Ref #4009)
What's Changed
- fix(proxy): reject degraded empty-thought records & feat(quota): migrate quota UI to Neon (Fixes #3997, Fixes #3745) by @cubelikeplayDaniel in #4005
- fix(proxy): preserve max_tokens stream truncation & guard high-capacity promotion (Fixes #4006) by @cubelikeplayDaniel in #4007
- fix(account,ui): 非 403/停用账号不再误显清除 403 按钮并透出真实测活错误 (Ref #3995) by @vecyang1 in #4010
Full Changelog: v5.0.4-beta.18...v5.0.4-beta.19