Skip to content

DAG Nodes

Workflows in the HQ visual editor are built from typed nodes connected by edges. Each node has its own execution semantics and configuration surface.

TypePurpose
StartEntry sentinel. Receives the trigger payload.
EndExit sentinel. Marks where a graph run finishes.
SkillExecutes a skill (LLM task) on a minion.
ReviewHuman or minion approval gate with revision loop.
Fan-outExpands a sub-graph over an input array.
JoinN-input barrier that aggregates upstream outputs.
ConditionalRoutes execution down one branch based on a condition.
TransformLLM-driven contract bridge between two edge contracts.
ScriptRun an inline Python or Node.js script — deterministic, no LLM.
flowchart LR
S(("Start")) --> A["Skill: fetch"]
A --> C{{"Conditional"}}
C -->|ok| T["Transform"]
C -->|error| E(("End"))
T --> F[["Fan-out"]]
F --> J(("Join"))
J --> R{"Review"}
R -->|approved| E
R -. revision .-> A

Start and End are sentinels — the editor inserts them automatically and they do not appear in the node palette. Every other node type is drag-and-drop from the palette on the left of the editor.

Every node has these fields in the property panel (rendered by the common header, not by the type-specific section):

FieldDescription
idStable identifier, unique within the graph (and within each template).
labelHuman-readable display name on the canvas.

Every step runs on any minion that is a member of the project. To have a particular minion run a step, pin it under Who runs each step on the workflow’s page in the project (the pin only applies to that project). Workflows made before project roles were retired may still name a role (assigned_role) on a step; it is ignored, and it is removed when the workflow is saved or published.

For type-specific configuration, see each node’s page.

A failed step shows why it failed on the run’s page:

Shown asWhat happened
Timed outThe step ran longer than its time limit.
LLM authentication failedThe minion’s LLM CLI is not signed in, or its key was rejected. Someone has to sign it in again.
LLM unavailableThe LLM was busy, hit a rate or usage limit, or ran out of credit.
Run errorThe LLM CLI exited with an error.
No outputThe step ended without handing in an output.
Contract violationThe output did not match the edge’s Contract.
Declared not completableThe step itself reported that the task could not be done.
Failed to startThe step’s session could not be started.
Script failedA Script node exited with an error, printed something that is not JSON, or timed out.
Minion stopped respondingThe minion running the step went silent for 5 minutes. The run moves on instead of waiting forever.
Marked as failed manuallySomeone pressed Mark as Failed.
Execution stoppedSomeone stopped the run.

Turn on Retry failed steps automatically in the editor’s Settings tab to have a workflow retry steps that failed for a reason another attempt can fix. Choose the maximum number of attempts, from 2 to 5, counting the first run. Retries are off by default, because every retry uses the LLM again.

  • Retried: minion stopped responding, timed out, LLM unavailable, run error, failed to start, no output, Contract violation.
  • Not retried: LLM authentication failed (it fails the same way until someone signs in), declared not completable, script failed, marked as failed manually, execution stopped.
  • The next attempt starts 1 minute after the first failure, then 5 minutes after each later one. The step is told why its previous attempt failed.
  • A retried step shows Attempt 2/3 on the run’s page, and its earlier attempts are listed under it.
  • For a step that must not run twice, such as sending mail, tick Never retry this step in its properties.
  • The setting is saved with the workflow, so it applies to runs started after you publish.

You can publish a workflow’s design so people outside your workspace can see it. The editor’s Publish button is a different thing: it fixes the version your workspace runs, and shows nothing to the outside. For publishing outside, see Public sharing, and What gets published for what people can see.