Workflows

Open Workflows — Automations to use the visual workflow builder. The large
canvas supports dragging steps, drawing connections, zooming and panning with
the canvas controls, and finding steps in the minimap.

When the canvas is empty, it says Start building a workflow.

Create and save a workflow

1. Select New Workflow to clear the canvas.
2. Select Add Step. A New Step node appears.
3. Select the node to open Edit Step.
4. Configure the step, then select Update.
5. Drag the node to the position where it should run.
6. Select Save.

The first save uses the default name New Workflow and confirms Workflow
created successfully. Open Expand workflow library, then use Rename
workflow to enter a better Workflow name. Later saves confirm Workflow
updated successfully.

Always confirm that the workflow appears in the library. Some workflow API
failures do not have an inline error notice in the builder, so the absence of a
success message does not mean the save worked.

Use the workflow library

The expanded Workflows library provides:

• New Workflow;
• Search workflows;
• a Saved workflows count; and
• the Workflow library list.

It can show No workflows yet — Create a workflow to build a reusable local
automation. or No matching workflows — Try a different search. Select a
saved name to load it on the canvas. Press Escape while renaming to cancel.

There is no delete-workflow, workflow import, or workflow export action in
v1.0. You can delete an individual step in its editor.

Edit a step

Select a node to open Edit Step. Its fields are:

• Step Name
• Description
• Step Type
• Command/Configuration

The actions are Update, Delete, and Cancel. Delete removes the
step and any lines attached to it from the current canvas.

Available step types are Trigger (Input), Map, Filter, Reduce,
Expand, Custom, LLM, Agent, and Adapter (Output).

Current step behavior

The Command/Configuration field accepts a small JSON configuration or, for
some functional steps, a limited expression. It is not a general code editor.

• Trigger (Input) preserves the run input and records the trigger
  configuration.
• Map can rename fields with a field_mapping object or extract one field
  from each list item with an expression such as x.name.
• Filter can compare one field to one value with a condition object or a
  simple equality expression such as x.status == 'ready'.
• Reduce supports count, or sum, min, and max for one field. With no
  recognized operation, a list is reduced to its count.
• Expand can expand a named list field into one output item per value.
• Custom does not execute arbitrary code. A JSON message creates a small
  wrapper around the input; other content passes the data through.
• LLM, Agent, and Adapter (Output) are placeholder processors in
  v1.0. They return clearly marked simulated output and do not call a language
  model, run an agent, or send data to another system.

For example, a Reduce step that sums an amount field uses:

{
  "operation": "sum",
  "field": "amount"
}

Unrecognized configuration often passes the input through unchanged. Inspect
Output Data after every experimental run.

Connections are visual only

You can drag a line from one step handle to another, but current connections
are not saved or restored and do not control execution.

The server runs every saved step sequentially by its canvas position: top to
bottom, then left to right when steps share the same height. Arrange nodes in
that order before selecting Save. Do not rely on arrows to define branches,
parallel work, or skipped steps.

Run a workflow

Only a saved workflow has a Run button.

1. Select Run.
2. Enter an optional JSON object under Input Data (JSON), or leave it
   blank.
3. Select Run again. The button changes to Running....

Invalid JSON produces Invalid JSON input data. An unsaved workflow produces
Please save the workflow first.

The Workflow Execution Result panel shows Status, Run ID, Started,
Completed, an optional Error, and formatted Output Data. Select
Close Results when finished.

Because execution is synchronous and the LLM/Agent/Adapter processors are
placeholders, treat Workflows as a limited local data-transformation prototype,
not a production automation or graph-execution system.