在 Armadra 里，连线即上下文：一个 Agent 能读取哪些节点，由画布上的连线决定。

读取范围

相连的 Agent 可以按需读取对方的：

• 摘要与对话转录；
• 终端画面；
• 便签、文件、差异等相关资料，以及节点上未解决的评论。

读取按预算裁剪，不会把整段长对话塞进另一个 Agent 的会话。没有连线的节点读不到；跨工作空间的节点即使残留连线也不生效。删除连线就收回了读取权限。

两种关系

点击连线可以设置它的关系：

• 对等：两个 Agent 平级协作，可以互相读取、互发消息。
• 主 → 从：主节点可以向从节点派发任务、读取其进度；常用于协调者和成员。

同级消息是资料，不是指令

从另一个 Agent 读到的内容、收到的消息，都作为资料交给接收方，而不是用户指令。接收方的 CLI 仍按它自己的权限策略行事，Armadra 不会替你回答它的权限提示。

只在画布内生效

Armadra 让 Agent 知道怎么使用画布，靠的是在从画布启动时才注入的 Hook、技能与画布说明。你在画布外自己启动的同一个 CLI 不受任何影响，Armadra 也不会改写 CLI 的全局配置。

In Armadra, links are context: which nodes an agent can read is decided by the links on the canvas.

What can be read

A linked agent can read, on demand, the other node’s:

• Summary and conversation transcript;
• Terminal screen;
• Notes, files, diffs and other related material, plus the node’s unresolved comments.

Reads are trimmed to a budget, so a long conversation is never dumped into another agent’s session. Unlinked nodes are out of reach, and a leftover link to another workspace has no effect. Deleting a link revokes access.

Two relationships

Click a link to set its relationship:

• Peers: two agents work side by side, reading each other and exchanging messages.
• Lead → member: the lead can assign tasks to the member and read its progress; typical for a coordinator and its members.

Peer messages are material, not instructions

Whatever an agent reads from another agent, and every message it receives, is handed over as material, not as a user instruction. The receiving CLI still follows its own permission policy, and Armadra never answers its permission prompts for you.

Only inside the canvas

Agents learn to use the canvas through hooks, skills and canvas notes that are injected only when launched from the canvas. The same CLI started by you outside the canvas is untouched, and Armadra never rewrites a CLI’s global configuration.