Execution

The backend plans a tree from its stored parent-child relationships. Nodes whose dependencies are ready can run in the same concurrent batch; dependent nodes wait for their parent outputs.

Node lifecycle

For each node, runtime processing can:

1. assemble parent output and run input
2. load reviewed skills and connector context
3. resolve the node’s AI connection
4. render prompt blocks and call the configured model
5. validate and store output
6. write a node trace and make the result available to dependants

Retries, timeouts, model parameters, and output contracts are controlled by the node execution policy.

Start a run

Provide JSON input and run the complete tree from the editor or Runs area. The GraphQL runTree mutation is non-blocking: it returns a pending run while execution continues in a background task.

Use runNode when testing one node synchronously through GraphQL.

Run states

A run can finish as completed, failed, partial, or cancelled. Partial means some nodes completed while other branches failed. Cancellation is best-effort; work already in flight may finish before the run reaches its terminal state.

Configuration failures

Before debugging model output, check that:

• the tree has reachable nodes and valid parent relationships
• an AI connection resolves for every model-backed node
• required connectors and skills can load
• input and output data satisfy configured schemas

Continue with Runs and traces to inspect an execution.