@lanxi266/dsh-tool-cluster

Model-facing cluster tools (cluster_view, cluster_send, cluster_spawn) and the cluster Agent persona injection. Registers three tools plus one systemPrompt section so any instance bound to a cluster Agent class sees that class's persona and can route messages across the canvas flow graph.

Model Experience

Request context and condition

What the model sees

The cluster-agent-persona system-prompt section (order -50) is injected for every instance bound to a cluster Agent class. {name} and {prompt} come from the class's agent.json; when prompt is empty the first line is 你是 {name}。. The space line appears only when the class enables its Agent space (spaceEnabled), with {path} the class's absolute directory. The three tools return model-visible text too; see the generated tool catalog.

cluster-agent-persona section

【Agent 设定】
你是 {name}，{prompt}
你正在工作流中，请使用 cluster_view 来确保你的信息流向正确；如果你是最后一环，那么你的 view 将为空。
你的身份空间是【{path}】，请先浏览身份空间。

Token effect

Fixed plus data-dependent: the workflow sentence is fixed per request; {name}/{prompt}/the optional space line scale with the class definition.

KV Cache effect

Prefix-stable: the section registers once at plugin load and its rendered text changes only when the class's agent.json name, prompt, or space toggle changes.

Known Limitations and Deferred Work

• Flow is open by default — with no canvas edges, allowFlow permits every route; deny-by-default applies only once edges exist.
• The graph is browser-synced — nodes and edges are pushed from the canvas via syncGraph and persisted as graph.json; the browser must have synced at least once.