GitHub - eersnington/sideffect: Build Cloudflare Workflows from composable typed steps, with Vite-generated workflow config. Define workflows in effect style code and let the Vite plugin create Wrangler workflow bindings and env types automatically. You use workflow bindings in your Worker as normal, without touching wrangler.toml/json file.

Define reusable and readable Cloudflare Workflows in effect inspired code and let the Vite plugin create Wrangler workflow bindings and env types upon wrangler dev and wrangler deploy step.

You use workflow bindings in your Worker as usual, and not deal with the hassle of configuring wrangler.toml/json file.

⚠️ Warning: The API is experimental and is subject to change.

Why Sideffect

Reusable typed steps — Define workflow steps as schema-backed, reusable activities.
No manual Wrangler config — The Vite plugin discovers your workflow files and generates Wrangler bindings and env types automatically.
Use bindings as normal — Workflow bindings are available on env like any other Cloudflare Worker binding.
Cloudflare-native runtime — Sideffect generates native WorkflowEntrypoint classes; nothing is emulated.

Writing Workflows

Define your workflows in src/workflows. Each file exports a workflow layer: a typed description of the workflow's payload, its steps, and the logic that connects them.

// src/workflows/my-workflow.ts
import { Schema, Step, Workflow } from "sideffect";

const workflow = Workflow.make({
  name: "image-processing",
  payload: Schema.Struct({
    imageKey: Schema.String,
  }),
});

const fetchImageStep = Step.make("fetch image", {
  payload: Schema.Struct({ imageKey: Schema.String }),
  result: Schema.Struct({ data: Schema.Uint8Array }),
  run: async (payload, ctx) => {
    const object = await ctx.env.BUCKET.get(payload.imageKey);
    const data = new Uint8Array(await object.arrayBuffer());
    return { data };
  },
});

const generateDescriptionStep = Step.make("generate description", {
  payload: Schema.Struct({ imageData: Schema.Uint8Array }),
  result: Schema.Struct({ description: Schema.String }),
  run: async ({ imageData }, ctx) => {
    const imageArray = Array.from(imageData);
    const result = await ctx.env.AI.run("@cf/llava-hf/llava-1.5-7b-hf", {
      image: imageArray,
      prompt: "Describe this image in one sentence",
      max_tokens: 50,
    });
    return { description: result.description };
  },
});

const publishImageStep = Step.make("publish", {
  payload: Schema.Struct({ imageKey: Schema.String, imageData: Schema.Uint8Array }),
  result: Schema.Void,
  run: async ({ imageKey, imageData }, ctx) => {
    await ctx.env.BUCKET.put(`public/${imageKey}`, imageData);
  },
});

export const myWorkflowLayer = workflow.toLayer(async (event, step) => {
  const image = await step.do(fetchImageStep, { imageKey: event.payload.imageKey });
  const description = await step.do(generateDescriptionStep, { imageData: image.data });

  await step.sleep("wait briefly", "1 second");

  await step.waitForEvent("await approval", { type: "approved", timeout: "24 hours" });

  await step.do(publishImageStep, { imageKey: event.payload.imageKey, imageData: image.data });

  return description;
});

The workflow name controls how Sideffect and Cloudflare refer to the workflow. For image-processing, the Cloudflare class name is ImageProcessing and the Worker binding is IMAGE_PROCESSING.

Step Context

Step.run receives the same Cloudflare WorkflowStepContext fields that native step.do callbacks receive:

ctx.env is typed from your project's Cloudflare.Env; use wrangler types or augment it in src/env.d.ts.

const describeImageStep = Step.make("describe image", {
  payload: Schema.Struct({ imageData: Schema.Uint8Array }),
  result: Schema.Struct({ description: Schema.String }),
  run: async ({ imageData }, ctx) => {
    if (ctx.attempt > 1) {
      console.warn(`Retrying ${ctx.step.name}, attempt ${ctx.attempt}`);
    }

    const result = await ctx.env.AI.run("@cf/llava-hf/llava-1.5-7b-hf", {
      image: Array.from(imageData),
      prompt: "Describe this image in one sentence",
      max_tokens: 50,
    });

    return { description: result.description };
  },
});

Rollback

Rollback is a Cloudflare-native feature. Sideffect lets you attach rollback handlers and config per step, but Cloudflare owns execution and ordering.

const publishImageStep = Step.make("publish image", {
  payload: Schema.Struct({ imageKey: Schema.String, imageData: Schema.Uint8Array }),
  result: Schema.Void,
  run: async ({ imageKey, imageData }, ctx) => {
    await ctx.env.BUCKET.put(`public/${imageKey}`, imageData);
  },
}).pipe(
  Rollback.with((_result, ctx) => {
    return ctx.env.BUCKET.delete(`public/${ctx.payload.imageKey}`);
  }),
);

Vite Adapter

The recommended setup uses Cloudflare's Vite plugin alongside Sideffect's adapter. Wrap cloudflare with withCloudflareWorkflows and the rest is automatic — Sideffect discovers your workflow layers and injects the generated entrypoints and Wrangler bindings into the build output.

import { cloudflare } from "@cloudflare/vite-plugin";
import { defineConfig } from "vite";
import { withCloudflareWorkflows } from "sideffect/vite";

export default defineConfig({
  plugins: [
    withCloudflareWorkflows(cloudflare, {
      workflowPaths: ["src/jobs", "src/features/billing/workflows"],
    }),
  ],
});

By default Sideffect scans src/workflows. If your workflow files live elsewhere, pass workflowPaths.

Your source wrangler.jsonc does not need a workflows field. Sideffect writes the workflow config into the Vite build output and generates env types for the workflow bindings it creates.

Static Discovery

The Vite adapter discovers workflow layers with TypeScript AST analysis before Worker modules run. It does not execute your source files during discovery.

Discovery supports statically knowable local workflow layers, including direct Workflow.make(...).toLayer(...) exports, local workflow definitions followed by workflow.toLayer(...), local relative imports of workflow definitions, and default-exported workflow layers.

Use the generated binding from your Worker as usual:

export default {
  async fetch(_req: Request, env: Env): Promise<Response> {
    const instance = await env.IMAGE_PROCESSING.create({
      params: { imageKey: "uploaded-photo-123" },
    });

    return Response.json({ id: instance.id });
  },
};

Plain Wrangler

Without the Vite adapter, Wrangler needs two things you provide manually: the native workflow class exported from your Worker entry, and the matching binding in wrangler.jsonc. Sideffect creates the class from your workflow layer via WorkflowEntrypoints.make.

Export the native workflow class alongside your Worker:

// src/index.ts
import { WorkflowEntrypoints } from "sideffect/cloudflare";
import { myWorkflowLayer } from "./workflows/my-workflow";

type Params = {
  imageKey: string;
};

declare global {
  namespace Cloudflare {
    interface Env {
      BUCKET: R2Bucket;
      AI: Ai;
      IMAGE_PROCESSING: Workflow<Params>;
    }
  }

  interface Env extends Cloudflare.Env {}
}

export const { ImageProcessing } = WorkflowEntrypoints.make({
  ImageProcessing: myWorkflowLayer,
});

export default {
  async fetch(_req: Request, env: Env): Promise<Response> {
    const instance = await env.IMAGE_PROCESSING.create({
      params: {
        imageKey: "uploaded-photo-123",
      },
    });

    return Response.json({ id: instance.id });
  },
};

Then register the workflow in your Wrangler config. The class_name must match the key passed to WorkflowEntrypoints.make, and binding is the property available on env:

LICENSE

Apache-2.0

推荐订阅源

Show HN