AGNT Workflow

Renders a ticket + issue type into an AGNT.gg workflow graph — a { name, description, nodes, edges } JSON document AGNT can import and run.

operator workflow export FEAT-1234 --format agnt   # writes FEAT-1234.agnt.workflow.json

Or over REST (also used by the operator-export-workflow plugin node):

curl -X POST "http://localhost:7008/api/v1/tickets/FEAT-1234/workflow-export?format=agnt"

Output shape

Each node carries { id, type, text, x, y, parameters } (AGNT’s runnable node shape — text is the canvas label, x/y are coordinates, and parameters is the bag AGNT resolves into the node’s tool). Each issue-type step becomes one node:

operator-run-step — the generic node, carrying { ticket, step, prompt, model, … } in its parameters; it calls Operator’s launch endpoint. (The prompt is an inert annotation — Operator owns the prompt internally; the tool reads only ticket/model.)
agnt-agent — AGNT’s native agent-chat node, emitted when a delegator is an AGNT-hosted remote agent (remote_agent.platform == "agnt"), carrying agentId (the agent’s UUID) and message (the prompt) so AGNT runs the step itself instead of calling back into Operator.

The next_step chain becomes edges, each { id, start: { id }, end: { id } }. Fan-out shapes (MultiModel / Matrixed / Pipeline) flatten to a single node, and human review gates, RAG, and MCP requirements are recorded in a gap field — lossy conversions are annotated, not dropped silently.

This is one half of the broader AGNT integration; AGNT-the-plugin (the operator-* node vocabulary that drives Operator) is the other.