steipete/taskflow

TaskFlow

Use TaskFlow when a job needs to outlive one prompt or one detached run, but you still want one owner session, one return context, and one place to inspect or resume the work.

When to use it

• Multi-step background work with one owner
• Work that waits on detached ACP or subagent tasks
• Jobs that may need to emit one clear update back to the owner
• Jobs that need small persisted state between steps
• Plugin or tool work that must survive restarts and revision conflicts cleanly

What TaskFlow owns

• flow identity
• owner session and requester origin
• currentStep, stateJson, and waitJson
• linked child tasks and their parent flow id
• finish, fail, cancel, waiting, and blocked state
• revision tracking for conflict-safe mutations

It does not own branching or business logic. Put that in Lobster, acpx, or the calling code.

Current runtime shape

Canonical plugin/runtime entrypoint:

• api.runtime.tasks.flow
• api.runtime.taskFlow still exists as an alias, but api.runtime.tasks.flow is the canonical shape

Binding:

• api.runtime.tasks.flow.fromToolContext(ctx) when you already have trusted tool context with sessionKey
• api.runtime.tasks.flow.bindSession({ sessionKey, requesterOrigin }) when your binding layer already resolved the session and delivery context

Managed-flow lifecycle:

1. createManaged(...)
2. runTask(...)
3. setWaiting(...) when waiting on a person or an external system
4. resume(...) when work can continue
5. finish(...) or fail(...)
6. requestCancel(...) or cancel(...) when the whole job should stop

Design constraints

• Use managed TaskFlows when your code owns the orchestration.
• One-task mirrored flows are created by core runtime for detached ACP/subagent work; this skill is mainly about managed flows.
• Treat stateJson as the persisted state bag. There is no separate setFlowOutput or appendFlowOutput API.
• Every mutating method after creation is revision-checked. Carry forward the latest flow.revision after each successful mutation.
• runTask(...) links the child task to the flow. Use it instead of manually creating detached tasks when you want parent orchestration.

Example shape

const taskFlow = api.runtime.tasks.flow.fromToolContext(ctx);

const created = taskFlow.createManaged({
  controllerId: "my-plugin/inbox-triage",
  goal: "triage inbox",
  currentStep: "classify",
  stateJson: {
    businessThreads: [],
    personalItems: [],
    eodSummary: [],
  },
});

const classify = taskFlow.runTask({
  flowId: created.flowId,
  runtime: "acp",
  childSessionKey: "agent:main:subagent:classifier",
  runId: "inbox-classify-1",
  task: "Classify inbox messages",
  status: "running",
  startedAt: Date.now(),
  lastEventAt: Date.now(),
});

if (!classify.created) {
  throw new Error(classify.reason);
}

const waiting = taskFlow.setWaiting({
  flowId: created.flowId,
  expectedRevision: created.revision,
  currentStep: "await_business_reply",
  stateJson: {
    businessThreads: ["slack:thread-1"],
    personalItems: [],
    eodSummary: [],
  },
  waitJson: {
    kind: "reply",
    channel: "slack",
    threadKey: "slack:thread-1",
  },
});

if (!waiting.applied) {
  throw new Error(waiting.code);
}

const resumed = taskFlow.resume({
  flowId: waiting.flow.flowId,
  expectedRevision: waiting.flow.revision,
  status: "running",
  currentStep: "finalize",
  stateJson: waiting.flow.stateJson,
});

if (!resumed.applied) {
  throw new Error(resumed.code);
}

taskFlow.finish({
  flowId: resumed.flow.flowId,
  expectedRevision: resumed.flow.revision,
  stateJson: resumed.flow.stateJson,
});

Keep conditionals above the runtime

Use the flow runtime for state and task linkage. Keep decisions in the authoring layer:

• business → post to Slack and wait
• personal → notify the owner now
• later → append to an end-of-day summary bucket

Operational pattern

• Store only the minimum state needed to resume.
• Put human-readable wait reasons in blockedSummary or structured wait metadata in waitJson.
• Use getTaskSummary(flowId) when the orchestrator needs a compact health view of child work.
• Use requestCancel(...) when a caller wants the flow to stop scheduling immediately.
• Use cancel(...) when you also want active linked child tasks cancelled.

Examples

• See skills/taskflow/examples/inbox-triage.lobster
• See skills/taskflow/examples/pr-intake.lobster
• See skills/taskflow-inbox-triage/SKILL.md for a concrete routing pattern