English
If you have any questions or suggestions about the software, please reach us through the QQ group or GitHub Issues and we will get back to you as soon as possible.
Added
- ACP access (Antigravity): the OAuth page has a new ACP Access tab. Sign in to the Antigravity CLI with official Google authorization, save several accounts and switch between them — all accounts share one CLI while login state is stored per account. After connecting you can send one test message to verify a model; enabling API access registers the account's models as
acp/<model-id>under a reserved OpenAI-compatible provider, callable with your existing CPA endpoint and key. The tab also carries CLI detection and install entry points, the current account identity and connection details. - Thinking levels for ACP:
reasoning_effort,reasoning.effortandthinking.effortare resolved to the official model's matching level variant, with levels detected from official model ID suffixes. A model without declared levels, an unknown level or conflicting values returns an explicit error instead of being ignored. The ACP provider does not write thinking-level declarations automatically; levels stay configured manually the usual CPA way. - Multi-turn chat for ACP: the bridge now accepts
system,userandassistantmessages and flattens the history into one prompt asSystem:/User:/Assistant:blocks, because each request uses a separate official CLI session. The last message must come from the user. A single user message is still passed through verbatim.
Improved
- ACP card layout: the two cards on the access page are now equal height, and the disconnected state is a compact lock state with an icon and explanation instead of a dashed box inside a box. When connected or with the API enabled, the heading icon switches to the success colour, the card lifts slightly on hover, and buttons wrap on narrow widths.
Fixed
- Health probe budget: the probe request budget is higher, and an exhausted generation budget is reported as its own distinct reason instead of being lumped in with other failures.
- New Gemini Interactions event names: the health check only recognised the legacy
content.delta/interaction.completeevents and only read text frominteraction.outputs, so a valid stream usingstep.delta/interaction.completed/stepswas reported as missing its first token. Both event generations are accepted now, completion text is read fromstepsoroutputs, a missing status counts as success, and failed or cancelled statuses are rejected explicitly. - Alias sync with multi-word keys: alias and agent YAML updates went through
yaml_serde_edit, whose transitive parser truncated mappings at unquoted multi-word plain keys, so legitimate edits during One-Click Connect were reported as a compatibility-check failure. Every related path now usesserde_norwayparsing and theyaml-editrenderer, and multi-word alias keys sync correctly again.
简体中文
- 关于软件,如果您有任何问题或建议,请通过 QQ群 或 GitHub Issues 反馈,我们会尽快处理。
新增
- ACP 接入(Antigravity):OAuth 页面新增「ACP 接入」标签页,可用 Google 官方授权登录 Antigravity CLI,并保存多个账号按需切换,账号共用同一份 CLI、登录状态分别保存。连接后可发送一条测试消息验证模型,启用 API 后账号下的模型会以
acp/<模型ID>注册成保留的 OpenAI 兼容 provider,用现有的 CPA 地址和密钥即可调用。该标签页也提供 CLI 检测与安装入口、当前账号身份和连接详情。 - ACP 的思考等级:请求里的
reasoning_effort、reasoning.effort或thinking.effort会被解析到官方模型对应的等级变体,等级从官方模型 ID 后缀识别。未声明等级、等级不存在或取值冲突时返回明确错误,不会静默忽略。ACP provider 不会自动写入思考等级声明,等级仍需按 CPA 的常规方式手动配置。 - ACP 的多轮对话:桥接现在接受
system、user、assistant消息,历史会按System:/User:/Assistant:合并进单次提示(官方 CLI 每次请求使用独立会话),最后一条必须是用户消息。单条用户消息保持原样直传,不受影响。
改进
- ACP 卡片布局:接入页的两张卡片改为等高排列,未连接时用图标加说明的紧凑锁定态,去掉了框内套框的虚线。已连接或 API 已启用时,卡片标题图标转为成功色,鼠标悬停有轻微阴影反馈,窄屏下按钮自动换行。
修复
- 健康检查的探针预算:探测请求的预算上调,并把「生成额度已用尽」作为一种独立原因上报,不再与其它失败混在一起。
- Gemini Interactions 的新事件名:健康检查此前只识别旧版
content.delta/interaction.complete并只从interaction.outputs取文本,使用step.delta/interaction.completed/steps的合法流会被误判为缺少首字。现在两代事件名都支持,完成文本从steps或outputs读取,缺少状态视为成功,失败和取消状态明确拒绝。 - 模型别名同步遇到多词键:别名与智能体的 YAML 更新此前经由
yaml_serde_edit,其传递依赖的解析器会在未加引号的多词键处截断映射,导致一键连接时合法修改被判定为兼容性检查失败。现在所有相关路径改用serde_norway解析与yaml-edit渲染,多词键的别名同步恢复正常。