Workflows and steps - Momor Workflows docs

TriggerTriggersThe request and its schema
DeclareOperationsTools, workflows, waits
StepsStepsIn order, nested where a step holds others
OutputOutputWhat the workflow returns

The four parts every workflow file has, in the order they are usually written.

A complete workflow

SummarizePage.yaml

Workflow: SummarizePage
Description: Fetches a page and answers a question about it.
Triggers:
  - New: Ask
    Schema:
      type: object
      required: [Prompt, Url]
      properties:
        Prompt: { type: string }
        Url: { type: string }
Instructions: Answer the question using the fetched page.
Operations:
  - Tool: Fetch
    Locator: embedded://momor.ai/services/tools/fetchPage
Output: Content
Steps:
  - ExecuteTool: Read
    Operations: [Fetch]
    Input:
      url: { Field: Trigger.Url }
  - DeliverResults: HandOver

ExecuteTool fetches the page, DeliverResults puts it into the run’s conversation, and because the conversation then ends on a result, the engine asks the model to reply to it. That reply is the workflow’s Content answer.

Writing a step

A step is written under its kind, and the value of that key is the step’s name: - ModelCall: Triage. The name is optional, and only needed where a later step reads the step’s output. Names are unique within a workflow. The same form names every list entry: - New: Ask for a trigger, - Tool: Fetch for an operation, - Hook: Title for a hook.

Step kinds

KindWhat it doesStep reference

ModelCall

Asks the model for one reply.

Ask the model

ExecuteTool

Runs one or more declared tools the author chose.

Call a tool

ExecuteModelToolCall

Runs the calls a model chose.

Call a tool

DeliverResults

Puts results into the run’s conversation.

Call a tool

Branch

Sends the flow down one of several paths.

ForEach

Runs work once per item of a list: one operation per item, or a body of steps per item.

For each

Wait

Asks a person a question.

Wait for approval

WorkflowCall

Runs one or more declared workflows.

Run a workflow

Repeat

Runs a body of steps more than once, up to a number of rounds.

—

While

Runs a body of steps for as long as a condition holds.

—

Exit

Ends the enclosing Repeat or While, or one item’s pass inside a ForEach body.

—

Guard

Wraps one step and serves an earlier run’s output where the key matches.

—

RefineLoop

Invokes an operation until what the run holds is good enough.

—

AgentCall

Gives a task to one or more agents and takes their answers.

—

Supervise

Runs agents in rounds and judges their answers between rounds.

—

Debate

Gives the same instructions to two models and crosses their positions in rounds until they agree.

—

Scope

Charges every call its steps make against a budget of its own.

Limits

The fields of every kind are listed in the YAML reference. A step kind the engine does not know is refused when the workflow is uploaded.

Data between steps

The feed. Each step receives the value the step before it produced. The request that opened the turn is the first value in the feed.
Named outputs. A named step’s output stays readable for the rest of the turn: { Field: Triage.Topic } reads the Topic field of the step named Triage. Trigger reads the request: { Field: Trigger.Url }.
Input mappings. A step with an Input mapping reads that instead of the feed. The keys are the target’s own field names, so they follow the tool’s or the called workflow’s schema.
The conversation. Model steps read the run’s conversation, not the feed. The opening request is its first message; DeliverResults adds tool results to it.

Names inside a branch or a loopA step named inside a branch case can be read only inside that case, and a step named inside a loop body only inside that body. The branch or loop’s own name carries its result out.