先问你（收件箱）

付款、对外发送、提交与登录等撤不回的操作先生成确认卡并锁定参数，等你批准后再执行。

在 OpenMousse app 的「今天 → 待处理」、对应对话消息的下方，或手机系统通知中均可查看并处理确认卡。

为了防止 Agent 在未经你同意的情况下对外发信、付款或改动关键数据，OpenMousse 对所有工具调用实行分级管控。凡属撤不回的操作，Agent 一律先停下来生成一张锁定参数的确认卡，交由你批准后才执行。

工具调用三档机制

系统内核将 Agent 的工具调用分为三档即将推出：

1.只读操作：查询世界树、读取日程或查看看板等不改动数据的操作，由 Agent 直接执行。
2.撤得回的操作：在内部新增或修改记录等可逆操作，由 Agent 直接执行，并在该条回复下方附带一个改动编号；你只需回复「撤销」，系统即可将该次改动恢复原状。
3.撤不回的操作：付款、对外发送、提交表单、账号登录等操作，一律停下来生成确认卡，锁定调用参数；经你点击批准后，系统严格照原参数执行一次。

确认卡的来源与处理流程

除第三档工具调用外，以下场景也会在收件箱与「今天 → 待处理」中生成确认卡或提醒：

•代办 Agent 的外部操作：代办 Agent 在隔离沙箱中比价、填表后，涉及提交、发送、登录或付款时生成确认卡。
•管家的对外承诺与 Doorman 复查：好友或好友的管家找你约时间、要东西时，管家不会自行答应，而是生成确认卡交给你定；管家发出的每一句回复若被 Doorman 判定不妥，也会扣下并交由你决定照发、修改或不发。
•赛博的数据修改与世界树写入 即将推出：赛博通过其装载的功能 Agent 请求修改数据，或请求将内容写入世界树时，均会生成确认卡等你确认。
•新建或提议功能 Agent：每晚日结后，主对话提议将你反复做的事建为新 Agent 时，会放进收件箱等你同意。
•外来指令安全提醒：当好友那边发来的消息里夹着试图操纵系统的指令时，系统会在「今天 → 待处理」中生成标签为「安全」的提醒。

你在任意一处（今天页待处理、对话消息下方或通知）批准或拒绝后，该确认卡即处理完毕，并在原处收起为一行回执。

不在这里的东西

•会议纪要不在收件箱中审批：多 Agent 短会的纪要卡直接出现在主对话中，经你在主对话内点击确认后写入日程与各 Agent 看板。
•世界树的「待确认」叶子不在收件箱中：开启世界树的「待确认」开关后，系统拿不准的记忆会停在「设置 → 世界树」（或记忆页年轮环外）的「待确认」列表中，不混入收件箱的确认卡。

另见

•代办与 Doorman：了解代办沙箱与 Doorman 如何配合确认卡管控出网请求。
•今天页与 04:00 日结：了解「待处理」板块与安全提醒的日常展示。
•内核（agent harness） 即将推出：了解发言方权限分档与工具三档的底层实现。

最后更新于 2026-10-19这一页有帮助吗？

For a complete index of the docs, see llms.txt. Append .md to any URL for the Markdown version; this page's Markdown is at approvals.md.

Features

Approvals (Inbox)

Hold irreversible actions like payments, external messages, submissions, and logins in parameter-locked approval cards until you approve.

You can view and act on approval cards in Today → To do in the OpenMousse app, directly beneath the corresponding chat message, or from system notifications on your phone.

To prevent agents from sending external messages, making payments, or modifying critical data without your consent, OpenMousse enforces tiered control over all tool calls. Whenever an action is irreversible, the agent stops and generates an approval card with locked parameters, executing it only after you approve.

Three-tier tool execution

The system engine divides agent tool calls into three tiers Coming soon:

1.Read-only actions: Operations that do not modify data—such as querying the memory tree, reading your schedule, or viewing a board—are executed directly by the agent.
2.Reversible actions: Reversible internal operations—such as adding or updating a record—are executed directly by the agent, with a change ID attached beneath the reply; reply Undo to restore the state prior to that change.
3.Irreversible actions: Operations such as payments, outbound messages, form submissions, and account logins always pause and generate an approval card that locks the call parameters; once you tap approve, the system executes the call exactly once with those locked parameters.

Where approval cards come from and how they are handled

In addition to third-tier tool calls, the following scenarios also generate approval cards or alerts in your inbox and Today → To do:

•External actions by the Errands agent: After the Errands agent compares prices or fills out forms inside its isolated sandbox, any step involving submitting, sending, logging in, or paying generates an approval card.
•External commitments by the Butler and Doorman review: When a friend or a friend's Butler asks to schedule time or request something from you, the Butler does not agree on its own, but instead generates an approval card for you to decide; if any outbound reply from the Butler is flagged by Doorman, it is also held back for you to send as-is, edit, or discard.
•Data modifications and memory tree writes by Cybers Coming soon: When a Cyber requests to modify data through one of its mounted functional agents or requests to write to the memory tree, it generates an approval card for your confirmation.
•Creating or proposing a functional agent: When main chat proposes spinning off a recurring task into a new agent after the nightly settlement, the proposal is placed in your inbox for approval.
•Security alerts for external instructions: When an incoming message from a friend's side contains instructions attempting to manipulate the system, the system generates an alert tagged Security in Today → To do.

Once you approve or reject an approval card in any location (To do in Today, beneath the chat message, or in a notification), the card is resolved and collapses into a single-line receipt in place.

What stays out

•Meeting minutes are not approved in the inbox: Meeting minutes cards from short multi-agent meetings appear directly in main chat and are written to your schedule and agents' boards when you tap confirm inside main chat.
•Memory tree leaves in Needs your approval are not in the inbox: When the Needs your approval toggle is enabled for the memory tree, uncertain memories wait in the Needs your approval list under Settings → Memory tree (or outside the tree ring on the Memory screen) and are not mixed into inbox approval cards.