inngest-realtime

---

name: inngest-realtime description: Use when streaming durable workflow updates to a UI in real time — live order status pages that animate as steps complete, AI agent token streaming from a function to the browser, log tailing for long-running jobs, or human-in-the-loop approval flows that publish a prompt and wait for a user reply. Covers Inngest v4 native realtime: defining typed channels, publishing from inside step.run, minting subscription tokens via server actions, and consuming the stream from React/Next.js client components.

Inngest Realtime

Stream updates from durable Inngest functions to live UIs. Use channels and topics to broadcast progress, render workflow execution as it happens, or build bi-directional human-in-the-loop flows.

These skills are focused on TypeScript. For Python or Go, refer to the Inngest documentation for language-specific guidance. Core concepts apply across all languages.

⚠ CRITICAL: v3 vs v4 package selection

Realtime in Inngest v4 lives at the SDK subpath inngest/realtime. The standalone @inngest/realtime npm package is a v3-era package and is NOT compatible with inngest@4.x. If your project is on v4 (the npm default), do not install @inngest/realtime. Use the imports below.

Symptoms of using the wrong package on v4: TypeError: Cls is not a constructor on every PUT /api/inngest, 401 on subscription tokens, type incompatibility on new Inngest({ middleware: [...] }). Verify your package.json shows "inngest": "^4.x" before reading further.

Prerequisites

• Inngest v4 SDK installed (npm install inngest) — see the inngest-setup skill
• INNGEST_DEV=1 set in .env.local for local development (without it, the SDK demands cloud signing keys and 401s on token requests)
• Local Inngest dev server running (npx inngest-cli@latest dev)
• Optional: zod for schema validation on topics

When to use Realtime

Problem shape	Pattern
Order status page animates as durable workflow steps complete	Per-run channel, publish per step, client subscribes
AI agent streams tokens to a chat UI	Per-conversation channel, publish chunks, stream to browser
Log tail for a long-running job	Single channel, log topic, append to UI
Human-in-the-loop approval	Channel + waitForEvent, publish prompt, wait for response
Admin dashboard with live order list	Global admin channel, fan-out from each function

Architecture

Three pieces:

1. Channel definition — a typed contract for what gets published. Lives in shared module so both server and client can reference the same channel name.
2. Publishing — call step.realtime.publish between steps to wrap a durable publish, or inngest.realtime.publish inside step.run because you're already inside a memoized step. See "Which publish method to use" below.
3. Subscribing — server action mints a subscription token; React client uses the useRealtime hook (or the lower-level subscribe() API for non-React consumers).

Step 1: Define a channel

Channels are pure data — no class hierarchy, no zod runtime required (but recommended for type safety). Define them once and import where needed.

// src/inngest/channels.ts
import { channel } from "inngest/realtime";
import { z } from "zod";

// Per-run channel: each fulfill-order run publishes step updates to its own channel.
export const orderChannel = channel({
  name: (orderId: string) => `order:${orderId}`,
  topics: {
    step: {
      schema: z.object({
        name: z.string(),
        status: z.enum(["running", "complete", "failed"]),
        output: z.record(z.string(), z.unknown()).optional(),
        ts: z.number(),
      }),
    },
  },
});

// Global admin channel: fan-out for cross-cutting visibility.
export const adminChannel = channel({
  name: "admin",
  topics: {
    order: {
      schema: z.object({
        orderId: z.string(),
        step: z.string(),
        status: z.enum(["running", "complete", "failed"]),
        ts: z.number(),
      }),
    },
  },
});

Two channel name shapes:

• name: 'admin' — static channel, accessed as adminChannel.order (topic ref)
• name: (id) => 'channel:${id}' — parametric, accessed as orderChannel(id).step (call the channel def with the id, then access topic)

Step 2: Publish from inside a function

Inngest v4 ships realtime support natively — no middleware required. But where you call publish matters: it determines whether the publish is durable, and it's the most common place to get realtime wrong.

Which publish method to use

Where you are	Use this	Why
Outside a step (top-level handler code, between step.run calls)	step.realtime.publish(id, topicRef, data)	Wraps the publish in its own step so it's durable, deduplicated by id, and retry-safe.
Inside a step (inside the callback passed to step.run)	inngest.realtime.publish(topicRef, data)	You're already inside a memoized step. step.realtime.publish would create a step inside a step. The bare client publish is the right call here.
Outside a function (one-off route, script, etc.)	inngest.realtime.publish(topicRef, data)	Allowed, but not retry-safe — your client receiver must handle duplicates.

The 90% rule: if you're writing handler code and you reach for publish, use step.realtime.publish. If you're writing code inside a step.run block and you reach for publish, use inngest.realtime.publish.

Example: both patterns in one function

// src/inngest/functions/fulfill-order.ts
import { inngest } from "../client";
import { orderChannel, adminChannel } from "../channels";

export const fulfillOrder = inngest.createFunction(
  {
    id: "fulfill-order",
    retries: 3,
    triggers: [{ event: "store/order.placed" }],
  },
  async ({ event, step }) => {
    const { orderId, customerEmail, lineItems } = event.data;

    // Outside any step.run — use step.realtime.publish for a durable wrapper.
    const emit = async (
      name: string,
      status: "running" | "complete" | "failed",
      output?: Record<string, unknown>,
    ) => {
      const ts = Date.now();
      await step.realtime.publish(
        `emit-order-${name}-${status}`,
        orderChannel(orderId).step,
        { name, status, output, ts },
      );
      await step.realtime.publish(
        `emit-admin-${name}-${status}`,
        adminChannel.order,
        { orderId, step: name, status, ts },
      );
    };

    await emit("capture-payment", "running");

    // Inside step.run — use inngest.realtime.publish (already in a memoized step).
    const payment = await step.run("capture-payment", async () => {
      const intent = await stripe.paymentIntents.create({
        /* ... */
      });

      // Stream a partial update mid-step. No step-in-step wrapping needed.
      await inngest.realtime.publish(orderChannel(orderId).step, {
        name: "capture-payment",
        status: "running",
        output: { stage: "intent-created", intentId: intent.id },
        ts: Date.now(),
      });

      return await stripe.paymentIntents.confirm(intent.id);
    });

    await emit("capture-payment", "complete", payment);

    await emit("reserve-inventory", "running");
    const inventory = await step.run("reserve-inventory", async () => {
      // ...
    });
    await emit("reserve-inventory", "complete", inventory);

    // ...
  },
);

Why no middleware: Earlier versions used @inngest/realtime's realtimeMiddleware() to inject a publish arg into the handler. v4 puts it on step.realtime and inngest.realtime directly.

Step 3: Mint a subscription token (server action)

In Next.js App Router, use a Server Action to securely mint a short-lived token for the React hook in Step 4. Without a token, clients can't subscribe.

// src/app/orders/[orderId]/actions.ts
"use server";

import { getClientSubscriptionToken } from "inngest/react";
import { inngest } from "@/inngest/client";
import { orderChannel } from "@/inngest/channels";

export async function fetchOrderSubscriptionToken(orderId: string) {
  // ⚠ AUTHORIZATION GATE: verify the current user owns this orderId
  // before minting a token. Channels are addressable by ID, so without
  // an ownership check, anyone can subscribe to any order's stream by
  // guessing IDs.
  //
  //   const session = await getServerSession();
  //   if (!session) throw new Error('Unauthenticated');
  //   const order = await db.order.findUnique({ where: { id: orderId } });
  //   if (order?.userId !== session.userId) throw new Error('Forbidden');

  return getClientSubscriptionToken(inngest, {
    channel: orderChannel(orderId),
    topics: ["step"],
  });
}

getClientSubscriptionToken from inngest/react returns a token shape that the useRealtime hook in Step 4 consumes directly. No ChannelInstance stripping needed — that gotcha only applies to the lower-level getSubscriptionToken + manual subscribe() path (see "Pattern: Manual subscribe" below).

Step 4: Subscribe with the useRealtime hook

The recommended consumer for React/Next.js is the useRealtime hook from inngest/react. It handles the subscription lifecycle, reconnect, type narrowing per topic, and cleanup.

// src/components/OrderStatusClient.tsx
'use client';

import { useRealtime } from 'inngest/react';
import { orderChannel } from '@/inngest/channels';
import { fetchOrderSubscriptionToken } from '@/app/orders/[orderId]/actions';

export function OrderStatusClient({ orderId }: { orderId: string }) {
  const { messages, connectionStatus, error } = useRealtime({
    channel: orderChannel(orderId),
    topics: ['step'] as const,
    token: () => fetchOrderSubscriptionToken(orderId),
  });

  if (error) return <div>Error: {error.message}</div>;

  return (
    <div>
      <div>Status: {connectionStatus}</div>
      <ul>
        {messages.all.map((m, i) => (
          <li key={i}>
            {(m.data as { name: string }).name}: {(m.data as { status: string }).status}
          </li>
        ))}
      </ul>
    </div>
  );
}

Useful options on the hook:

Option	Default	Use it when
enabled	true	Delay the subscription until you have an ID (e.g., enabled: !!runId).
bufferInterval	0	Batch updates from a fast stream so React doesn't re-render per message.
pauseOnHidden	false	Pause the stream when the tab isn't visible (saves bandwidth).
autoCloseOnTerminal	true	Disconnect when the run completes — turn off to keep the stream open for fan-out channels.
historyLimit	unbounded	Cap how many messages are retained in messages.all.

The hook returns messages.byTopic (latest per topic), messages.all (full history), messages.last (most recent), and messages.delta (new since last render).

Pattern: Manual subscribe (non-React or custom transport)

The useRealtime hook covers the React case. If you're not using React, or you need a custom subscription lifecycle (server-side streaming, background workers, custom protocols), use the lower-level subscribe() API directly.

Server action: mint a token with the lower-level helper

// src/app/orders/[orderId]/actions.ts
"use server";

import { getSubscriptionToken } from "inngest/realtime";
import { inngest } from "@/inngest/client";
import { orderChannel } from "@/inngest/channels";

export async function fetchOrderSubscriptionTokenLowLevel(orderId: string) {
  // ⚠ AUTHORIZATION GATE: same as Step 3 — verify ownership before minting.

  const token = await getSubscriptionToken(inngest, {
    channel: orderChannel(orderId),
    topics: ["step"],
  });

  // ⚠ CRITICAL: strip the ChannelInstance from the response.
  // getSubscriptionToken returns { channel: ChannelInstance, ... } where
  // ChannelInstance contains zod schema methods (a class with prototypes).
  // Next.js refuses to serialize classes across the server-action → client-component
  // boundary, so return ONLY primitives.
  return {
    channel: orderChannel(orderId).name as string,
    topics: ["step"] as const,
    key: token.key,
    apiBaseUrl: token.apiBaseUrl,
  };
}

Manual client subscription

// src/components/OrderStatusManual.tsx
"use client";

import * as React from "react";
import { subscribe } from "inngest/realtime";
import { fetchOrderSubscriptionTokenLowLevel } from "@/app/orders/[orderId]/actions";

export function OrderStatusManual({ orderId }: { orderId: string }) {
  const [messages, setMessages] = React.useState<unknown[]>([]);

  React.useEffect(() => {
    let cancelled = false;
    let sub: { close?: (reason?: string) => void } | undefined;

    (async () => {
      const token = await fetchOrderSubscriptionTokenLowLevel(orderId);
      if (cancelled) return;

      sub = await subscribe(
        {
          channel: token.channel,
          topics: [...token.topics],
          key: token.key,
          apiBaseUrl: token.apiBaseUrl,
        },
        (message) => {
          if (cancelled) return;
          setMessages((prev) => [...prev, message.data]);
        },
      );
    })();

    return () => {
      cancelled = true;
      sub?.close?.("unmount");
    };
  }, [orderId]);

  // ... render ...
}

SSE streaming from a route handler

Subscribe inside a Next.js API route and pipe the stream to the client via SSE:

// src/app/api/orders/[orderId]/stream/route.ts
import { inngest } from "@/inngest/client";
import { subscribe } from "inngest/realtime";
import { orderChannel } from "@/inngest/channels";

export async function GET(
  req: Request,
  { params }: { params: { orderId: string } },
) {
  // ⚠ AUTHORIZATION GATE: same rule as the server-action token mint above.
  // Authenticate the request and confirm the caller owns params.orderId
  // before opening the SSE stream. Skipping this leaks every order's
  // step events to anyone with a URL.

  const stream = await subscribe({
    app: inngest,
    channel: orderChannel(params.orderId),
    topics: ["step"],
  });

  return new Response(stream.getEncodedStream(), {
    headers: {
      "Content-Type": "text/event-stream",
      "Cache-Control": "no-cache",
      Connection: "keep-alive",
    },
  });
}

Client consumes via fetch().getReader() rather than the subscribe() callback. Use this when you want the SSE behavior or when the client-side subscribe() API doesn't fit your component lifecycle.

Pattern: Human-in-the-loop

Combine step.realtime.publish with step.waitForEvent:

import crypto from "crypto";

export const reviewWorkflow = inngest.createFunction(
  { id: "review-workflow", triggers: [{ event: "review/start" }] },
  async ({ event, step }) => {
    const confirmationId = await step.run("gen-id", () => crypto.randomUUID());

    // Publish a prompt — the client subscribes and renders an approval UI
    await step.realtime.publish("publish-prompt", reviewChannel.message, {
      message: "Confirm to proceed?",
      confirmationId,
    });

    // Wait up to 15 minutes for the user to send the matching event back
    const confirmation = await step.waitForEvent("await-confirmation", {
      event: "review/confirmation",
      timeout: "15m",
      if: `async.data.confirmationId == "${confirmationId}"`,
    });

    if (!confirmation) {
      // user didn't respond — abort or escalate
      return { decision: "timed_out" };
    }
    // continue workflow...
  },
);

The confirmationId links the published prompt to the matching reply, so the workflow knows which response to act on.

Common pitfalls

Don't use @inngest/realtime on v4

The standalone @inngest/realtime package is for Inngest v3 only. On v4, all realtime APIs are in the SDK subpath inngest/realtime. Mixing them produces:

• TypeError: Cls is not a constructor on PUT /api/inngest (v3 middleware class signature mismatch)
• 401 Unauthorized on subscription tokens
• TypeScript errors casting middleware

Verify with: grep '"inngest"' package.json — if it's ^4.x, use inngest/realtime. Period.

Don't return ChannelInstance from a Next.js server action (manual subscribe path only)

getSubscriptionToken returns { channel: ChannelInstance, ... } where ChannelInstance has zod schema methods (a class). Next.js refuses to serialize classes across the server-action → client-component boundary. Strip to primitives before returning. See "Pattern: Manual subscribe" above.

This gotcha does not apply when you use getClientSubscriptionToken from inngest/react (Step 3 — the recommended path). That helper returns a serialization-safe shape directly.

INNGEST_DEV=1 is required for local dev

Without it, the SDK assumes cloud mode and demands INNGEST_SIGNING_KEY + INNGEST_EVENT_KEY. All realtime operations 401 / 500. Add to .env.local. Hard restart the dev server (Next.js does not hot-reload .env.local changes).

Channel topic schemas validate on publish, not on consume

If your published payload doesn't match the zod schema, the publish fails server-side. Subscriber receives nothing. Catch publish errors during step execution, or run with validate: false in subscribe() if you have a reason to skip schema validation client-side.

Reference

• v4 entry points:
  • import { channel } from 'inngest/realtime' — channel definitions
  • import { useRealtime, getClientSubscriptionToken } from 'inngest/react' — React hook + matching token helper (Step 3 + Step 4)
  • import { getSubscriptionToken, subscribe } from 'inngest/realtime' — lower-level helpers for non-React or custom transport
• Publish methods:
  • Outside a step: step.realtime.publish(id, topicRef, data) — wraps in a durable step
  • Inside step.run: inngest.realtime.publish(topicRef, data) — already inside a memoized step, no wrapping needed
  • Outside a function: inngest.realtime.publish(topicRef, data) — allowed but not retry-safe
• Subscribe overloads: subscribe(token) returns a stream; subscribe(token, callback) invokes callback per message
• Next.js Server Action gotcha (manual path only): strip ChannelInstance → return { channel: string, topics, key, apiBaseUrl }. Not needed with getClientSubscriptionToken.

This Repository

These upstream Inngest instructions are vendored for agent tooling and integration work in this monorepo.

Repository Triggers

Use this skill when inngest-realtime matches the current Inngest task. If the right skill is unclear, start with docs/ai/skills/inngest/SKILL.md.

Repository Workflow

1. Confirm whether the request is agent-tooling guidance or product runtime integration.
2. Use inngest-brownfield-audit before changing existing app workflows or fragile background work.
3. Follow this upstream guidance under OpenSpec, root AGENTS.md, repo rulebooks, framework docs, and runtime evidence.
4. Keep runtime packages, app code, migrations, and INNGEST_* env requirements out of agent-tooling-only changes.

Repository Checklist

• The task has explicit product-runtime scope before adding Inngest app code or dependencies.
• Existing workflows were audited before introducing or changing durable workflow behavior.
• Any MCP usage is backed by a running Inngest dev server on the configured port.
• Upstream source and license attribution remain documented in docs/ai/skills/inngest/references/upstream.md.