Use this package when you want to host the OpenTag dispatcher inside another Node or Hono-compatible service instead of running @opentag/dispatcher-app.

Install

pnpm add @opentag/dispatcher

Exports

createDispatcherApp: creates the Hono app that exposes the OpenTag dispatcher API.
UnifiedDeliveryProducer: validates and enqueues configured delivery intents.
DispatcherDeliveryPresentation: provider-neutral rendering input for the configured delivery-intent resolver.

Example

import { createDispatcherApp } from "@opentag/dispatcher";

export const dispatcher = createDispatcherApp({
  databasePath: "opentag.db",
  pairingToken: process.env.OPENTAG_PAIRING_TOKEN
});

API Shape

The app exposes /healthz and /v1/* dispatcher endpoints for runners, Project Target bindings, generic channel bindings, Slack compatibility bindings, runs, progress, heartbeats, completion, and audit event lookup.

Factory-mode callers can additionally create immutable recipe snapshots and WorkThread-only workstreams, submit replay-safe workstream batches, inspect a durable batch receipt, and read workstream metrics/evaluation. Every batch item uses the same managed-channel ownership and Run admission path as /v1/runs; routine per-item source-thread acknowledgements are suppressed in favor of one bounded exception summary.

Governed WorkLoop continuation

An immutable factory recipe may opt into continuation.mode: evidence_driven for selected completion_evidence_changed, human_escalation_resolved, or retryable_run_failure triggers. The dispatcher re-evaluates the canonical WorkLoop and Workstream budgets at the trigger boundary; it never treats a timer, a runner result, or an attention_required metric by itself as authority to continue.

One eligible Workstream creates a normal child Run with triggeredByAction.kind: resume_work_thread, durable parent/trigger/digest metadata, and the same admission, access-profile, routing, and managed-channel checks as any other Run. Managed Slack/Lark admission stores a non-secret application/bot ownership attestation on the parent event; continuation must match that attestation to the current binding, and a request-bound human resolution must also present the current owning adapter principal. Admission- time Slack/Lark escalations persist the server-verified provider, conversation scope, non-secret ownership, and binding digest so acknowledgement and resolution remain fail-closed even though no Run exists yet. Human resolution continuations carry a credential-sanitized immutable decision snapshot in the child context and bind its digest into continuation metadata. Multiple eligible Workstreams, missing recipe authority, a blocked Workstream, a human/reconciliation action, or a concurrent Run all fail closed. The admission check does not queue an implicit follow-up, and Run persistence repeats the active-conversation check inside its transaction to close the admission/write race; direct child Run writers also reject an already-committed automatic continuation, and queued follow-up promotion uses the same transactional fence. Created continuations do not emit a second source acknowledgement; their lineage and control-plane event are the receipt. Deferred, ambiguous, rejected, and failed continuation outcomes remain distinct in direct and source-thread responses so a new task is recommended only for manual-policy outcomes.

When pairingToken is set, every /v1/* endpoint requires:

Authorization: Bearer <pairingToken>

Stability

The Hono app factory and delivery-intent boundary are public API. Delivery is activation-blocked until a resolver and durable submitter are configured.