Join a Knowledge Graph

Join the engines: your function as a graph task in an Active Knowledge Graph.

At a glance

• What — a graph.task node names your route; the same declarative map from the flow chapter points it at your host.
• Requires — engine v4.11.11 or later (the release that taught the deployed-graph guard about the declarative map).

The graph names the route — nothing more

In MiniGraph, behavior lives on nodes as skills. The graph.task skill invokes a composable function by route name — and since v4.11.11 that route may resolve through yaml.event.over.http to a Python host. A minimal node:

{
  "types": ["Task"],
  "alias": "python-step",
  "properties": {
    "skill": "graph.task",
    "task": "hello.python",
    "input": ["input.body -> *"],
    "output": ["result -> output.body"]
  }
}

The graph model carries no URL and no language — deployment stays in the engine's configuration, exactly as the composable theme demands. Everything about authoring graphs (nodes, skills, the compile gate, the Playground) is the engine's domain: start at Build your first graph and the built-in skills reference.

Deadlines: two distinct clocks

• The graph's ttl bounds the event call to your function — on breach the engine sees the standard 408 envelope.
• If your function itself calls onward over HTTP, that client call carries its own timeout — the same decoupling the engines document for async.http.request (headers.x-ttl in milliseconds).

Design functions so their own outbound work respects the deadline they were given: read the caller's intent from your ttl-bounded world and fail fast rather than overstay (Design).

Errors: the graph's error.* contract

When your function raises AppException(code, message) — or fails unexpectedly with a 500 — the walker stages the graph's generic error context: error.source (the failing node), error.code, error.message, and error.stack when one traveled. A graph can therefore route around a Python failure with the same IF/THEN patterns it uses for any other task — see the engine's knowledge-graph guides for the retry and recovery idioms.

Worked demo

The quickest end-to-end proof mirrors the engines' own conformance test: a graph with one graph.task node whose task names a route served by the Python demo app (hello.python in Getting Started), with the declarative map on the engine pointing at http://127.0.0.1:8086/api/event. Deploy the graph (or dry-run it in the Playground), POST /api/graph/{graph-id}, and the reply body is your function's output — trace id visible in the Python host's log line.

Checklist

1. Engine at v4.11.11+ with yaml.event.over.http configured.
2. The map entry's route == the node's task value == your @preload route.
3. The graph passes the compile gate and is listed in the manifest (deployed lane) — "compiled or 404" is the engine's rule.
4. Your host is healthy; /info/routes lists the route as public (a private route answers 403 on the wire by design).