Reliable by Design · 05.20

Node Reference: Data, Compute & IO

transform, code, data_query, http_request, aggregate_window, knowledge, browser, artifact_save, mcp, human_input — every data/compute/IO node explained.

NodesReference

Deterministic data & compute

Node	Purpose
`transform` / `filter`	Reshape or select data via safe expressions — never `eval`.
`code`	Run sandboxed JS (or Python) with a call cap and timeout.
`data_query` / `data_mutate`	Read or write the App datastore.
`http_request` / `graphql`	Call external APIs — SSRF-guarded, cancellable.
`aggregate_window`	Windowed aggregation over a stream of items.
`spreadsheet` / `json_schema_validate`	Parse or emit CSV; validate a payload against a schema.
`workflow_store` / `workspace_store` / `scratchpad`	Durable per-workflow, per-workspace, or per-run key-value storage.

Knowledge, IO & human-in-the-loop

Node	Purpose
`knowledge` / `knowledge_ingest`	Search a knowledge base, or ingest documents into one.
`browser`	Headless Playwright navigation, extraction, or screenshot.
`artifact_save` / `artifact_collect`	Persist a referenceable artifact; gather several for delivery.
`mcp` / `integration` / `extension_task`	Call an MCP tool, a connector operation, or an agent-authored extension.
`human_input`	Suspend the run for a human answer, then resume exactly where it paused.

Adding an entirely new node kind touches roughly 4–6 wiring points: the type union, its config schema, dispatch, the handler itself, validation, and canvas metadata. The node executors live behind a typed host facade in engine/executors/, not scattered through the core engine file.

Continue

Node Reference: Intelligence Nodes

agent_task, agent_session, agent_swarm, converge, pursue, planner, evaluator, guardrails — every intelligence node explained.

NodesReference

The Runtime Abstraction Layer: Adapters & the Normalized Contract

How AdapterManager drives Claude Code, Codex, Cursor, Hermes, Antigravity, and HTTP runtimes through one normalized task/result contract.

RALAdapters