@memberjunction/task-graph

Durable execution of task graphs: submission, the dispatcher, the claim protocol, and the runners that turn a node into work.

New to workflows? Start with the Workflows and Task Graphs Guide, which covers what a workflow is, when to use one, and every rule that decides what happens next. To step a live run, see the Workflow Debugger Guide. This README is the package tour.

What this package is for

A task graph is work that has to outlive the thing that asked for it. Before durable execution, a multi-step agent plan lived inside one agent run: a page reload lost it, a server restart orphaned it, and no channel other than the one that started it could see it.

This package is the other half of that split. Something else produces a graph; this package makes it durable and runs it.

graph TB
    subgraph Producers["Producers — anyone"]
        P1["Flow agent
compiled"]
        P2["Loop agent
emitted"]
        P3["Entity action"]
        P4["TaskGraph.Submit
remote op"]
    end

    P1 & P2 & P3 & P4 --> SVC["TaskGraphService.Submit
validate · resolve · persist"]
    SVC --> ROWS["Task + TaskDependency"]
    ROWS --> DISP["TaskGraphDispatcher
poll · claim · execute · settle"]

    DISP --> R1["TaskAgentRunner"]
    DISP --> R2["TaskActionRunner"]
    DISP --> R3["TaskLoopExecutor"]
    DISP --> R4["a person"]

    style SVC fill:#7c5295,stroke:#563a6b,color:#fff
    style ROWS fill:#2d8659,stroke:#1a5c3a,color:#fff
    style DISP fill:#b8762f,stroke:#8a5722,color:#fff
Copy

Submission never waits for execution. Submit returns as soon as the graph is durable. That split is what makes the engine invocation-agnostic: an agent, a scheduled job, a Slack message and a manual UI all call the same method, and whichever dispatcher instance is running picks the work up.

What it deliberately does not do

It does not decide graph semantics. Eligibility, failure propagation, parent rollup, skip cascade, exclusive-group resolution and stall detection all come from pure, dependency-free functions in @memberjunction/ai-core-plus. That is not tidiness — it is what stops the in-run executor and the durable executor from drifting apart. Neither owns the rules.

It does not import MJServer. Provider minting and agent execution arrive as injected seams (ProviderFactory, TaskAgentRunner, TaskActionRunner), so the dependency runs MJServer → task-graph and never the reverse. That is also what keeps the dispatcher unit-testable without standing up the agent framework.

The pieces

Claiming: why a task runs exactly once

Multiple dispatcher instances poll the same table. Correctness comes from a guarded write, not from coordination: claiming is an update that only succeeds if the row still looks the way the claimer expects.

sequenceDiagram
    participant A as Instance A
    participant B as Instance B
    participant DB as Task row

    A->>DB: claim if unclaimed / expired
    B->>DB: claim if unclaimed / expired
    DB-->>A: 1 row updated ✅
    DB-->>B: 0 rows updated ❌
    Note over B: defers — no lock, no retry storm
    loop while running
        A->>DB: heartbeat (extends ClaimExpiresAt)
    end
    A->>DB: complete IF still owned by A
Copy

If A dies, its claim expires and reconciliation makes the task claimable again. If A finishes but the row changed underneath it (cancelled, reassigned, reclaimed), the guarded completion refuses and A defers — overwriting would undo a newer, deliberate decision.

Runners are seams, not implementations

A host with no runner is limited, not broken. "Nobody here can run this" is not the same as "this ran and did not work", so those tasks stay visible and claimable by an instance that can.

An Agent node starts a brand-new AIAgentRun — a root run, linked from Task.AgentRunID rather than nested under the submitting run. The relationship is expressed by the Task row.

Starting a dispatcher

MJAPI does this for you (StartTaskGraphDispatcher). For a custom host:

import { TaskGraphDispatcher, LoadTaskGraphOperations } from '@memberjunction/task-graph';

LoadTaskGraphOperations();   // registers the remote operations + the durable submitter

const dispatcher = new TaskGraphDispatcher(
    providerFactory,
    agentRunner,
    contextUser,
    { InstanceID: `worker-${process.pid}` },
    continuationDeliverer,   // optional
    observer,                // optional
    actionRunner,            // optional
);
await dispatcher.Start();
Copy

Without LoadTaskGraphOperations() nothing registers the submitter, and GetTaskGraphSubmitter() returns null — which callers must report rather than swallow.

Start() registers a kick and fires one immediately. Do not wait for the first setInterval tick — that is a 5s dead pause after Debug/Run. Submit cannot import the dispatcher instance (submit and execute are separate halves), so it calls KickTaskGraphDispatchers() instead.

Debugging a live graph

Every control is a gate on claiming, never on running. State lives under $.debug on the parent task's InputPayload. The debugger UI and the dispatcher never talk directly — they rendezvous on the row. See debug-state.ts and the Workflow Debugger Guide.

• Start-paused. Submit writes paused: true before any dispatcher has seen the graph.
• Breakpoints. The claim gate treats an armed, eligible task like a hold. Resume from a breakpoint stamps skipBreakpointTaskID so that one claim is allowed through; without it the next poll re-hits the same Pending row. ForEach is one task — one Continue runs the loop.
• Edge overrides. Authored 'true' / 'false' on a TaskDependency id, honored by every instance, survive a restart.
• Kick. After Submit (and on Start()), KickTaskGraphDispatchers() runs a pass now.

The Angular drop-in is <mj-task-graph-debugger> in @memberjunction/ng-task-graph-editor.

Cost rollup

When a graph settles, the dispatcher credits its spending back to the run that submitted it, writing the …Rollup columns on AIAgentRun:

• TotalCost — what the submitting run itself spent. Never rewritten here.
• TotalCostRollup — that run plus everything it caused.

This cannot happen during the run: a submitting run ends at submission, so at the moment it computes its own totals the graph has not spent anything yet. See the guide's cost section.

Testing

cd packages/TaskGraph && pnpm test
Copy

Unit tests cover the pure pieces — loop semantics, dispatchable kinds, configuration persistence, parent metadata. The dispatcher driving real rows against SQL Server is covered by the IT74 bundle in @memberjunction/integration-test-suite, which uses a stub agent runner so it stays in the deterministic tier: no model calls, no tokens, real claim protocol, real condition evaluator, real rollup.

Related

• Workflows and Task Graphs Guide — start here
• Workflow Debugger Guide — step a live run
• @memberjunction/ng-task-graph-editor — canvas and drop-in debugger
• @memberjunction/ai-core-plus — the spec, validator, compiler, pure algorithms, payload mapping and layout
• @memberjunction/ai-agents — the agent framework and FlowAgentType
• packages/Actions/CLAUDE.md — Actions as boundaries