Dag

DAG Canvas

The dag renderer provides the display scenario. It validates that every edge references an existing node and that the resulting directed graph is acyclic.

Inline DAG

The Canvas config may be the DAG document itself:

canvas_spawn({
  kind: 'dag',
  scenario: 'display',
  config: JSON.stringify({
    schemaVersion: 1,
    title: 'Research pipeline',
    nodes: [
      { id: 'prices', label: 'Market prices', description: 'Daily OHLCV', color: 'cyan' },
      { id: 'features', label: 'Feature build', color: 'yellow' },
      { id: 'report', label: 'Agent report', color: 'green' },
    ],
    edges: [
      { from: 'prices', to: 'features', label: 'returns' },
      { from: 'features', to: 'report' },
    ],
    layout: {
      direction: 'left-to-right',
    },
  }),
});

File DAG

For larger graphs, create the same DAG document as a JSON file in the user's workspace, then pass exactly one absolute path field. dataFile is an alias of data_file; do not provide both and do not mix file and inline fields.

canvas_spawn({
  kind: 'dag',
  scenario: 'display',
  config: JSON.stringify({
    data_file: '/absolute/path/dag.json',
  }),
});

Document Schema

Required paths:

• schemaVersion: must be 1
• title: non-empty string, at most 160 characters
• nodes: 1 to 500 node objects
• nodes[*].id: unique non-empty string, at most 128 characters
• nodes[*].label: non-empty string, at most 160 characters
• edges: 0 to 2,000 edge objects
• edges[*].from: ID of an existing source node
• edges[*].to: ID of an existing target node

Optional node fields are description, color, and scalar metadata. Optional edge fields are label, color, and scalar metadata. Self-edges, duplicate source/target pairs, unknown node references, and cycles are rejected. Edge labels must be single-line strings. Omit an edge's color to group connections by target: all edges entering one node share a stable generated color, while edges entering different targets use different color slots. Set color only when a specific semantic color is required. Outgoing edges from one node use separate source ports and bend lanes; incoming edges may converge on the same target port and retain their shared target color.

Optional layout:

{
  "layout": {
    "direction": "left-to-right",
    "minNodeWidth": 12,
    "maxNodeWidth": 48,
    "minNodeHeight": 3,
    "maxNodeHeight": 8,
    "layerGap": 8,
    "nodeGap": 2,
    "padding": 2
  }
}

• direction: left-to-right or top-to-bottom
• minNodeWidth / maxNodeWidth: per-node adaptive width bounds for content and vertical-layout ports
• minNodeHeight / maxNodeHeight: per-node adaptive height bounds for content and horizontal-layout ports
• layerGap: minimum of 4 to 24 cells; edge labels and outgoing bend lanes expand the actual gap
• nodeGap: 1 to 12 cells
• padding: 0 to 8 cells

Unknown document fields are rejected.

Node boxes adapt independently to their label, ID, and wrapped description. Edge labels are centered in a reserved section of the connection, and the actual layer gap grows to fit them. Only the bendable connection segments use Braille cells; arrowheads are ordinary → or ↓ terminal characters.

Error Recovery

canvas_spawn and canvas_update wait for parsing, reference validation, and cycle detection. Errors are returned by the originating tool as { success: false, id, status, error }, not as DAG actions. Repair the exact field or edge reported and update the same Canvas ID.

Interaction

• Left-drag the graph viewport to pan when it overflows the pane.
• Click a node without dragging to toggle selection.
• Arrow keys or h/j/k/l pan; the mouse wheel pans vertically.
• Tab and Shift+Tab choose the active node; Space toggles it.
• Home resets the viewport, and c clears selection.
• a attaches the selected neighborhood or visible viewport to LLM context.
• Enter requests DAG analysis; e exports the current context as JSON.

When canvas_wait is available, call it with this Canvas ID while waiting for an interactive selection or analysis request. Otherwise, rely on the host's automatic Canvas event delivery.