Phase 01 · Linear intake

The simplest slice of The Archivist: wire a dispatcher, register its nodes and DAGs in dependency order, execute one visitor query, and read the lifecycle result. The full runner is below — this is the real code.

Flow

Code

The #linear-run region covers the dispatcher construction, molecular deep-DAG registration, and the execute call that drives the full flow:

// ── Dispatcher ───────────────────────────────────────────────────────────
const dispatcher = new Dagonizer<ArchivistState, ArchivistServices>({ services });

// ── Deep-DAG node registration (molecular pattern) ───────────────────────
// Each deep-DAG module exports a registerXxxNodes helper that registers
// the nodes it needs. Call it before registerDAG so the validator can
// resolve all node references when the DAG is registered.
registerBookSearchFanoutNodes(dispatcher);
dispatcher.registerDAG(BookSearchFanoutDAG);

registerComposeRetryLoopNodes(dispatcher);
dispatcher.registerDAG(ComposeRetryLoopDAG);

// ── Parent-DAG-only nodes (not used by deep-DAGs) ────────────────────────
for (const node of [
  recallContext,
  classifyIntent,
  // Inlined branch nodes (reviews + describe) — not in the deep-DAGs
  extractQuery,
  decideTools,
  webSearchScout,
  openLibraryScout,
  googleBooksScout,
  subjectScout,
  wikipediaScout,
  rankByRating,
  pickBestMatch,
  mergeCandidates,
  recordFindings,
  hasCitationsGate,
  groupByYear,
  recallPastVisits,
  recommendSimilar,
  // recall-memories branch
  recallMemories,
  composeMemoryResponse,
  respondToVisitor,
  declineOffTopic,
  declineEmpty,
  // empty-result LLM response branch
  composeEmptyResponse,
]) {
  dispatcher.registerNode(node);
}

dispatcher.registerDAG(archivistDAG);

// ── Demo run ─────────────────────────────────────────────────────────────
const visitor = new ArchivistState();
visitor.query = "I'm looking for a book about a strange house and a library";

const result = await dispatcher.execute('the-archivist', visitor);

logger.result(`intent=${result.state.intent}`);
logger.result(`shortlist=${String(result.state.shortlist.length)}`);
logger.result(`draft=${result.state.draft}`);
logger.result(`lifecycle=${result.state.lifecycle.kind}`);
logger.result(`triples=${String(services.memory.size)} written`);

What it demonstrates

• Molecular registration order — deep-DAG nodes must be registered before their DAG is registered (registerBookSearchFanoutNodes → dispatcher.registerDAG(BookSearchFanoutDAG)), and both deep-DAGs before the parent archivistDAG. The dispatcher validates all node references at registration time.
• Single execute call — dispatcher.execute('the-archivist', visitor) drives the entire multi-branch flow. The caller sees one ExecutionResult<ArchivistState> containing the final state and lifecycle.
• Lifecycle result — result.state.lifecycle.kind is 'completed', 'cancelled', or 'timed_out'. Nodes never throw; the dispatcher always returns.
• Services bag — every node receives context.services (LLM, search tools, memory, logger). Nodes never construct their own clients.

See this in action in the Archivist live demo.