Test a workflow without a model

Use MockHarness to run the real executor with scripted node results. You can check routing, preconditions, and evaluations without an API key or model call.

import assert from "node:assert/strict";
import { execute, type Workflow } from "@sweny-ai/core";
import { MockHarness } from "@sweny-ai/core/testing";

const workflow: Workflow = {
  id: "review",
  name: "Review",
  description: "Check a scripted finding",
  entry: "inspect",
  nodes: {
    inspect: { name: "Inspect", instruction: "Inspect the repository", skills: [] },
  },
  edges: [],
};

const harness = new MockHarness({
  strict: true,
  responses: { inspect: { data: { finding: "Missing timeout" } } },
});

const run = await execute(workflow, {}, {
  harness,
  skills: new Map(),
  env: {},
  offline: true,
});

assert.equal(run.results.get("inspect")?.status, "success");
assert.equal(run.results.get("inspect")?.data.finding, "Missing timeout");
assert.deepEqual(harness.executedNodes, ["inspect"]);

Catch incomplete fixtures

With strict: true, a missing node response or an unavailable scripted tool returns a failed result naming the problem. fail_soft cannot turn these fixture errors into success. All scripted tool names are checked before any handler runs for that attempt.

The executor supplies the node ID on every attempt. Duplicate instructions, workflow rules, and retry feedback therefore select the same node fixture. Each retry reuses that node’s response. For a manual harness.run() call, pass nodeId, or supply workflow to the constructor and use an exact, unique instruction. Strict mode refuses ambiguous instructions instead of guessing.

Without strict, older manual callers retain instruction and sequential matching. MockClaude remains an alias for MockHarness.

Tools still execute their handlers

Scripted toolCalls invoke the handlers supplied to the executor. Strict mode checks fixtures; it does not isolate file writes or network requests. Use in-memory tool handlers when a test must have no external effects.

The example above supplies no skills or tool calls. offline: true also prevents remote source resolution; it does not sandbox a custom tool handler. For conditional branches, supply routes: { sourceNodeId: "targetNodeId" } and assert the executed nodes as well as the final result. Checking only a successful final status can miss the wrong branch or an incomplete fixture.