graph-workflow

Graph Mutation Reference

The live MCP instructions and understanding-work define the fluid medium. This skill only explains how to preserve a chosen update accurately. It does not require orienting, searching, re-entering, or touching every tool in a fixed order.

At a genuine choice point, graph_suggest_next({ task, workflow }) can propose concrete weighted routes. The model may choose, combine, change, reject, or replace them. When a selected route calls graph_understand, keep its proposed stance; workflow and stance are independent.

Mutate atomically

Substantive graph changes go through graph_batch. A batch is one committed encounter: either every operation succeeds or none does. It accepts at most 100 operations; split independent work into separate commits.

graph_batch({
  commit_message: "Connected the new failure model to the observed retry burst",
  agent_name: "current-agent",
  operations: [
    { tool: "graph_add_concept", params: {
        title: "Fixed retries synchronize failure",
        trigger: "surprise",
        understanding: "A recovery rule that protects one client can align the population and intensify overload.",
        why: "Changes the failure model from independent retries to common-mode behavior"
    }},
    { tool: "graph_connect", params: {
        from: "$0.id",
        to: "n_retry_observation",
        type: "learned_from",
        why: "The burst trace is the evidence that occasioned this update"
    }}
  ]
})

$N.id references the result of operation N, counting every operation from zero. Use the exact field names in the live schema. Common ones are:

graph_add_concept  { title, trigger, understanding, why }
graph_note         { about, testimony, title?, trigger?, why?, status?, relations? }
graph_connect      { from, to, type, why }
graph_revise       { node, understanding, before, after, pivot, why }
doc_create         { title, content?, fileType?, isDocRoot?, parentId?, level?, afterId? }
doc_revise         { nodeId, content, why }
doc_move           { nodeId, parentId?, afterId? }
doc_split          { nodeId, mode, lineNumbers?, childLevel? }