Scheduling a Project#

This guide covers how to run the scheduler through the raw WebAssembly exports. It explains when to call the one-shot schedule function and when to keep a JsonProject handle. It also covers what the solve options change, how forward and backward planning place tasks, and what converged means in a result.

How scheduling works#

There are two entry points, and they run the same planner:

schedule(data_json, options_json) takes a ProjectData snapshot, schedules it once and returns a ScheduleResult. It builds a default project with strict validation, solves it and drops it. There's no handle to free.
JsonProject#solve(data_json, options_json) does the same work on a long-lived handle. The project keeps the scheduled dates as its accepted baseline, so later calls can edit, preview, level and evaluate against it.

Both take JSON text and return JSON text: an envelope, {"ok": true, "result": …} or {"ok": false, "error": …}. The envelope and the field spelling are on Wire format.

Every solve follows the same steps. The engine parses the JSON, imports the snapshot into its own normalized model and validates it. It builds the scheduling graph from links and summaries, checks for dependency cycles and resolves the scope (the whole project, or one branch with subtree). Then it runs the planner until the dates stop changing. The result lists only what moved.

A few rules shape what you get back:

Updates are sparse. result.updates holds one entry per task whose dates changed, with only the changed fields set. The entries use the task-data names (start_date, end_date, duration), so you can write them straight back into your own model. To get every task's dates, set includeUnchanged.
Problems are data, not errors. Conflicts, cycles, dropped input and non-convergence come back inside an ok: true result. An error envelope means invalid input (kind: "parse" or "validation") or a call that doesn't fit the project's state (kind: "project"). See Error handling.
Some tasks never move. Tasks with auto_scheduling: false, unscheduled tasks and completed tasks (use_progress with progress == 1) stay where they are. That's how you lock work in place.
Calls are synchronous and CPU-only. There's no I/O and no clock. A large solve blocks the calling thread, so run it in a Web Worker or a worker thread. See Run in a worker.

One-shot scheduling#

Use schedule in batch jobs, request handlers and edge functions that schedule a snapshot and throw it away.

const { schedule } = require("./pkg-node/plen_wasm.js");

const data = {
  tasks: [
    { id: 1, start_date: "2026-08-10 08:00", duration: 2 },
    { id: 2, start_date: "2026-08-10 08:00", duration: 3 },
  ],
  links: [{ id: 1, source: 1, target: 2, type: "fs" }],
};

const response = JSON.parse(schedule(JSON.stringify(data), "{}"));
if (!response.ok) {
  const { kind, message, path } = response.error;
  throw new Error(`${kind}: ${message} (at ${path})`);
}
for (const update of response.result.updates) {
  console.log(`task ${update.id}: ${update.start_date} .. ${update.end_date}`);
}

Task 2 has to wait for task 1, so it's the only update: task 2: 2026-08-12 08:00 .. 2026-08-17 08:00. Task 1 is already where it belongs and isn't reported. The default "global" calendar works Mon-Fri, 08:00-12:00 and 13:00-17:00, which is why the dates skip the weekend.

"{}" means default SolveOptions. schedule never throws for bad input. It returns an error envelope with kind parse or validation when the snapshot is rejected.

Note

tasks is required in ProjectData; send "tasks": [] for an empty project.

A server that already holds the request body as JSON text can pass it to schedule and return the envelope text without parsing either. See Forward JSON from a server.

Solving a project#

Keep a JsonProject when the plan lives on: an editor backend with one project per open document, or a service that applies edits over time.

const { JsonProject } = require("./pkg-node/plen_wasm.js");

function unwrap(text) {
  const response = JSON.parse(text);
  if (!response.ok) {
    const { kind, code, message, path } = response.error;
    throw new Error(`${kind} ${code ?? ""}: ${message} (at ${path})`);
  }
  return response.result;
}

const data = {
  tasks: [
    { id: 1, start_date: "2026-08-10 08:00", duration: 2 },
    { id: 2, start_date: "2026-08-10 08:00", duration: 3 },
  ],
  links: [{ id: 1, source: 1, target: 2, type: "fs" }],
};

const project = new JsonProject("{}");
try {
  // Install the data, schedule it and accept the dates as the baseline.
  const loaded = unwrap(project.solve(JSON.stringify(data), "{}"));

  // Nothing changed since: the cached result comes back.
  const cached = unwrap(project.solve(undefined, "{}"));
  console.log(cached.baselineId === loaded.baselineId); // true

  // Edit, then reschedule.
  const report = unwrap(project.apply(JSON.stringify({ tasks: { update: [{ id: 1, duration: 4 }] } })));
  if (report.needsResync) throw new Error("edit rejected; resync with solve(dataJson, ...)");

  const after = unwrap(project.solve(undefined, "{}"));
  for (const update of after.updates) {
    console.log(`task ${update.id}: ${update.start_date} .. ${update.end_date}`);
  }
} finally {
  project.free();
}

The last solve prints both tasks:

task 1: 2026-08-10 08:00 .. 2026-08-14 08:00
task 2: 2026-08-14 08:00 .. 2026-08-19 08:00

The two forms of solve do different jobs:

solve(dataJson, …) validates the replacement data, installs it, schedules it and accepts the dates as the new baseline. Use it on the first load and to resync the whole project. If validation fails, you get an error envelope and the previous state and cached result stay as they were.
solve(undefined, …) (or null) schedules the current state. After an apply it plans the edits. Tasks the edit touched (task 1 above) are reported too, even when the planner leaves them where the edit put them. On a clean project with the same options it returns the cached result without planning again.

The cache key is the whole SolveOptions object. Calling solve(undefined, …) with different options replans from the accepted dates. That doesn't change the data, so the baselineId stays the same unless a task moves.

After a rejected apply operation, solve(undefined, …) returns a project error with code needs-resync until you send the full data again with solve(dataJson, …). Editing is covered in Editing a project.

The constructor is the one call that throws on bad input, and it throws the error envelope as a string. The handle lives in WebAssembly memory until free(). See Module and handles.

Solve options#

SolveOptions changes what a solve schedules and what it reports. The defaults schedule the whole project and return only the moves. Option names are camelCase.

const { JsonProject } = require("./pkg-node/plen_wasm.js");

function unwrap(text) {
  const response = JSON.parse(text);
  if (!response.ok) {
    const { kind, code, message, path } = response.error;
    throw new Error(`${kind} ${code ?? ""}: ${message} (at ${path})`);
  }
  return response.result;
}

const data = {
  tasks: [
    { id: 1, start_date: "2026-08-10 08:00", duration: 2 },
    { id: 2, start_date: "2026-08-10 08:00", duration: 3 },
    { id: 3, start_date: "2026-08-10 08:00", duration: 1 },
  ],
  links: [{ id: 1, source: 2, target: 3, type: "fs" }],
};

const project = new JsonProject("{}");
try {
  // Initial load: every task's dates, why each one sits where it does,
  // plus the critical path and slack.
  const options = {
    includeUnchanged: true,
    reasons: true,
    analytics: { criticalPath: true, slack: true },
  };
  const result = unwrap(project.solve(JSON.stringify(data), JSON.stringify(options)));
  for (const update of result.updates) {
    console.log(update.id, update.reason.kind);
  }

  // Later: reschedule only task 2 and what depends on it.
  unwrap(project.solve(undefined, JSON.stringify({ subtree: { task: 2 } })));
} finally {
  project.free();
}

The first solve reports all three tasks. Tasks 1 and 2 stay at their dates (preferred), and task 3 is pushed by link 1 (asap):

{
  "duration": 1,
  "end_date": "2026-08-14 08:00",
  "id": 3,
  "kind": "task",
  "reason": { "kind": "asap", "linkId": 1, "sourceId": 2 },
  "start_date": "2026-08-13 08:00"
}

includeUnchanged also reports tasks that didn't move, each with start_date, end_date and duration set. The result alone is then the full normalized date list, which is what an editor needs on first load. Unscheduled tasks are never reported. See Load all task dates.
reasons adds a reason to each update: pushed by a predecessor link (asap), pulled by a successor (alap), moved with its summary block (summary-translation), placed by its constraint (constraint) or kept at its preferred date (preferred).
analytics asks for the critical path, slack, or both. They're off by default. See Critical path and slack.
subtree limits the run to one task and what depends on it. inclusive defaults to true, so the root task is scheduled too. The subtree applies to this call only. A scoped solve accepts only its own moves, and a later unscoped solve picks up the rest. An unknown or unscheduled subtree task is a validation error at subtree.task, and the cached result is kept. See Reschedule a branch.

With subtree and includeUnchanged together, the result holds every task in the subtree (with their ancestors) plus the out-of-subtree tasks that changed or were touched by an edit.

level and preview don't inherit these options: level solves with defaults first, and preview takes its own PreviewOptions.

Forward and backward planning#

By default the planner works forward: each task starts as early as its links allow. Set config.schedule_from_end to plan backward instead: each task ends as late as its successors allow, and the project is anchored at its end.

const { schedule } = require("./pkg-node/plen_wasm.js");

const data = {
  config: {
    schedule_from_end: true,
    gap_behavior: "compress",
    project_end: "2026-08-28 17:00",
  },
  tasks: [
    { id: 1, start_date: "2026-08-10 08:00", duration: 2 },
    { id: 2, start_date: "2026-08-12 08:00", duration: 3 },
  ],
  links: [{ id: 1, source: 1, target: 2, type: "fs" }],
};

const response = JSON.parse(schedule(JSON.stringify(data), "{}"));
if (!response.ok) throw new Error(response.error.message);
for (const update of response.result.updates) {
  console.log(`task ${update.id}: ${update.start_date} .. ${update.end_date}`);
}

Task 2 has no successor, so project_end caps it and it ends at 2026-08-28 17:00. Task 1 ends where task 2 starts, on 2026-08-25 17:00.

The project bounds act only on tasks at the open ends of the graph:

project_start (forward) applies to tasks without predecessors. With gap_behavior: "compress" it pulls them onto the project start. With "preserve" (the default) it only raises tasks that start earlier and leaves later ones at their dates. For a task with predecessors it's a hard floor only when apply_constraints is on.
project_end (backward) caps tasks without successors, in either gap mode. Without it, such a task keeps its current end.

The full config reference is on SchedulingConfig.

A task whose constraint points the other way gets a secondary pass in the opposite direction. In a forward project with apply_constraints, an alap task moves as late as its successors allow:

const { schedule } = require("./pkg-node/plen_wasm.js");

// "order" can happen any time before "install"; "build" sets the pace.
const data = {
  config: { apply_constraints: true },
  tasks: [
    { id: "build", start_date: "2026-08-10 08:00", duration: 4 },
    { id: "order", start_date: "2026-08-10 08:00", duration: 1, constraint_type: "alap" },
    { id: "install", start_date: "2026-08-10 08:00", duration: 1 },
  ],
  links: [
    { id: "l1", source: "build", target: "install", type: "fs" },
    { id: "l2", source: "order", target: "install", type: "fs" },
  ],
};

const response = JSON.parse(schedule(JSON.stringify(data), "{}"));
if (!response.ok) throw new Error(response.error.message);
for (const update of response.result.updates) {
  console.log(`${update.id}: ${update.start_date} .. ${update.end_date}`);
}

install waits for build and starts on 2026-08-14 08:00. order moves from Monday to 2026-08-13 08:00, so it ends right when install starts.

Leveling is forward-only. level on a project with schedule_from_end: true returns a validation error at config.schedule_from_end. Constraint types and summary behavior are covered in Links, constraints and summaries.

Determinism and convergence#

The same input and options always give the same result. The engine reads no clock and uses no randomness. It iterates in input order and topological order, never in hash order. That makes results safe to cache, diff and compare in tests.

The one exception is baselineId. Tokens are opaque, and two schedule calls on the same data can return different ones. Never compare them across runs or store them. Take the current one from the latest solve result. To compare two envelopes, drop result.baselineId first.

The planner runs passes until the dates stop changing. result.converged tells you whether it reached that point within config.max_iterations passes:

An unscoped run converges in two passes by design. The second pass reproduces the first.
A scoped run (subtree) can need more passes, because tasks outside the scope feed into it.
max_iterations defaults to 8 and must be at least 1. 0 is a validation error at config.max_iterations.

const { schedule } = require("./pkg-node/plen_wasm.js");

const data = {
  config: { max_iterations: 4 },
  tasks: [{ id: 1, start_date: "2026-08-10 08:00", duration: 1 }],
};

const response = JSON.parse(schedule(JSON.stringify(data), "{}"));
if (!response.ok) throw new Error(response.error.message);
const result = response.result;
if (result.cycles.length > 0) {
  console.log(`dependency cycles: ${result.cycles.length}`);
} else if (!result.converged) {
  console.log("planner stopped after max_iterations");
}

converged: false isn't an error, and the result is still valid.

A dependency cycle stops the run before planning. Dates stay unchanged, updates is empty and cycles lists each unique cycle, for example [{"links": [1, 2], "tasks": [1, 2]}]. For everything else a result carries, see Reading results.