Without MCP
Drive runs from your own code, such as a web form, a Slack button, a queue or a test.
MCP is one way to drive a run. workflowRuns is the same thing for your own code: start a
run, read its open question, answer it later.
import { workflowRuns } from '@onedef/elicitation-workflow'
const runs = workflowRuns(release, { runStore })
const state = await runs.start({ version: '2.1.0' })
// {
// runId: 'run_…',
// status: 'waiting',
// waitingFor: { kind: 'human', name: 'elicit-1', message: 'Publish 2.1.0 (build-1)?', requestedSchema },
// }Show the question however you like, and send the answer back with next. It can come
minutes or days later, from another process:
await runs.next(state.runId, { action: 'accept', content: { channel: 'beta' } })
// → { runId, status: 'done', output: { released: true, … } }What next takes depends on waitingFor.kind:
| kind | from | answer |
|---|---|---|
human | ctx.elicitHumanInput | { action: 'accept' | 'decline' | 'cancel', content? } |
approval | ctx.approve | { approved: boolean, feedback?: string } |
agent | ctx.elicitAgentInput | a value matching waitingFor.schema |
next(runId) with no answer just reports where the run is. A wrong answer throws
RunError with code INVALID_INPUT, and the run stays where it was.
Examples of drivers
- Web app: one route starts the run and saves the
runId. A page renders a form fromrequestedSchema(a flat object of primitive fields). The submit handler callsnext. - Slack or email: post the question with buttons. The button's webhook calls
next, hours later, on whichever instance receives it. - Tests: start, assert on
waitingFor, answer, assert onoutput.
Advanced: headless runs
A headless run is one that nobody will answer, such as a nightly job. MCP servers don't need this, because an MCP client can always answer.
await runs.start(input, { headless: true })In a headless run:
| result | |
|---|---|
ctx.elicitHumanInput | { action: 'cancel' }, as from a client that cannot show forms |
ctx.elicitAgentInput | the run fails: nobody can answer it |
ctx.approve | waits as usual; start returns status: 'waiting', and next can still answer it |
To finish anyway, branch on ctx.headless and do the agent's work yourself:
const plan = ctx.headless
? await ctx.step('plan', () => myModel.plan(ctx.input.topic))
: await ctx.elicitAgentInput('plan', { schema, instructions, data })