Testing Pipeline Nodes
A Test node is a first-class node you drop into the chain. It carries an assertion the run must satisfy to continue — so you see the whole verification contract before a model touches a single line of code.
An assertion, wired in as a node.
Most agents "test" after the model has written and run code — you find out it was wrong once the damage is done. DISPATCH inverts that. A Test node carries a plain-language instruction the upstream models read, and an assertion the run must satisfy to continue. Four fields:
Plain-language intent the upstream models read as part of their brief — "auth must reject an expired token". They plan toward it.
A concrete check on any output — exit_code == 0, contains("401"), schema.valid, or a graded rubric. Pass or fail, nothing fuzzy.
Which upstream node it inspects — the output of Create, the diff from Apply, the plan from Plan.
On fail it can block, loop back to the producing node, or branch to repair — like any control edge.
Place, instruct, gate.
Output flows in, a verdict flows out — like any node, except its job is to assert, not generate.
Drop it in the chain
Place NODE: TEST #1 after any node you want verified — between Create and Apply, or after a Merge. It wires top-to-bottom like the rest.
Instruct the models
Write the test intent in plain language. Upstream Plan and Create nodes read it as part of their brief and build toward passing it — testable-by-design.
Gate the downstream
At run time the node evaluates its assertion. Pass → control flows on. Fail → it blocks, loops back, or branches to a repair node. Deploy never sees unverified work.
Run the chain. Watch the test gate it.
Toggle whether Create produced passing output, then press Run. The gate either lets Deploy through, or loops back to repair — Deploy never sees unverified work.
Nodes that pair with Test #1.
Advanced Orchestration
Branch, retry, loop, gate, timeout — the control edges a failing test triggers.
openCanvas Planning
Lay out goals and phases on the graph — Test nodes assert the milestones.
openAgentic Streaming
Fan a task across 20 agents — gate each lane's output with its own test.
open