The workflow graph

Flow contains a workflow UUID, name, optional brief and banner, task categories, an optional coordinator, nodes, edges and an ISO modified timestamp. Each node uses type: "step" and a Canvas position. Its data contains the label, kind, instructions in note, and the selected catalog entry.

{
  "id": "analyze",
  "type": "step",
  "position": { "x": 0, "y": 0 },
  "data": {
    "label": "Analyze the quote",
    "kind": "model",
    "note": "Extract the main idea and tone."
  }
}

The selected entry supplies the resource identity and capability metadata. Its kind must match the step. A model’s display name helps the reader; its entry.id identifies the actual resource for execution.

Dependencies

An edge stores its own id, source and target node IDs, with optional handle IDs and type: "wire". The source output becomes context for the target step. Several target nodes can consume the same source. A node with multiple incoming edges receives the earlier results together. The coordinator’s plan identifies these dependencies and groups ready nodes into topological waves. The backend accepts up to 500 nodes and 2,000 edges. Node IDs, edge IDs and source-to-target pairs must be unique. Connections must refer to existing nodes, and the graph must be acyclic.

Saved identity

id stays stable across edits. The server returns a new modified timestamp when it saves a workflow; send that value in If-Match when updating or publishing it. Developer workflow requests show the save flow.

How it fits together

Catalog and agents

Execution

The workflow graph

Dependencies

Saved identity