sailfork / t3code-github / .repos/ alchemy-effect/ website/ src/ content/ docs/ blog/ 2026-05-12-beta-37.md

v2.0.0-beta.37 is the biggest beta in a while. The headline addition is cross-stack and cross-stage references — the missing piece that lets PR-preview stages share a database with staging instead of re-provisioning a whole Postgres cluster every time someone opens a draft PR. On top of that: fully-typed Worker-to-Worker bindings, typed Workflow I/O, the new Alchemy.Secret / Alchemy.Variable one-liners, cron triggers, and an Analytics Engine binding.

A bunch of this came from outside the core team — props to the contributors throughout, and full credits at the bottom of the post.

Cross-stack and cross-stage references

Lazy, typed references to resources deployed by a different stack or stage. The use case it was built for: ephemeral PR-preview stages that need a Neon project, shouldn't pay for their own, and should instead share one that's owned by staging.

Pointing a PR stage at staging's database

Same alchemy.run.ts, conditional on the stage. If the stage looks like a PR preview (pr-147, pr-148, …), Neon.Project.ref reaches into the staging stage's state file and pulls out the already-deployed project. Otherwise the stage creates its own.

// src/Db.ts
import * as Alchemy from "alchemy";
import * as Drizzle from "alchemy/Drizzle";
import * as Neon from "alchemy/Neon";
import * as Effect from "effect/Effect";

export const NeonDb = Effect.gen(function* () {
  const { stage } = yield* Alchemy.Stack;

  const schema = yield* Drizzle.Schema("app-schema", {
    schema: "./src/schema.ts",
    out: "./migrations",
  });

  // PR previews share the long-lived staging project.
  // Every other stage gets its own.
  const project = stage.startsWith("pr-")
    ? yield* Neon.Project.ref("app-db", { stage: "staging" })
    : yield* Neon.Project("app-db", { region: "aws-us-east-1" });

  // Branches are cheap and per-stage either way.
  const branch = yield* Neon.Branch("app-branch", {
    project,
    migrationsDir: schema.out,
  });

  return { project, branch, schema };
});

Three things to note:

  • Same logical id, same type. "app-db" matches the id staging uses to create the project; project is Neon.Project either way, so downstream (Neon.Branch({ project })) doesn't know or care whether it's real or referenced.
  • Resolved at plan time. Alchemy reads the project's attributes (id, host, etc.) out of staging's persisted state store. If staging hasn't been deployed yet, plan fails loudly with InvalidReferenceError.
  • PR teardown stays scoped. alchemy destroy --stage pr-147 deletes the per-PR Neon.Branch but doesn't touch the shared project — this stage doesn't own it.

Deploy staging once, then PR stages can point at it:

alchemy deploy --stage staging        # creates the project once
alchemy deploy --stage pr-147         # references it, creates only the branch

The full file lives in examples/cloudflare-neon-drizzle/src/Db.ts; the guide is at Guides › Shared database across stages.

Referencing an entire stack's outputs

The example above pulls one resource across stages. The other shape — pulling a whole stack's outputs — is what you reach for in a monorepo where the frontend package wants to read the backend stack's deployed URL.

Declare a typed stack handle once:

// backend/src/Stack.ts
import * as Alchemy from "alchemy";

export class Backend extends Alchemy.Stack<
  Backend,
  { url: string }
>()("Backend") {}

Deploy the backend with Backend.make(...) (the typed shorthand for Alchemy.Stack), then yield* Backend from the frontend's stack to get its outputs back, type-checked:

// frontend/alchemy.run.ts
import * as Alchemy from "alchemy";
import * as Cloudflare from "alchemy/Cloudflare";
import { Backend } from "backend";
import * as Effect from "effect/Effect";

export default Alchemy.Stack(
  "Frontend",
  { providers: Cloudflare.providers(), state: Cloudflare.state() },
  Effect.gen(function* () {
    // Resolves Backend's outputs from the same stage of the same
    // stack name. `pr-42` frontend reads `pr-42` backend.
    const backend = yield* Backend;
    //    ^? { url: string }

    return yield* Cloudflare.Website.Vite("Website", {
      env: { VITE_API_URL: backend.url },
    });
  }),
);

yield* Backend defaults to "same stage as the consumer". When you need to pin — say, the prod frontend always reads the prod backend regardless of which branch deploys it — use Backend.stage.<name>:

const backend = yield* Backend.stage.prod;        // always pin to prod
const backend = yield* Backend.stage["pr-42"];    // arbitrary stage name

Under the hood, both shapes are Output.stackRef / Resource.ref reading the state store. The new APIs just make them ergonomic.

Concepts › References · Guides › Shared database · Guides › Monorepos · Tutorial › Branch from a shared database

Alchemy.Secret and Alchemy.Variable

The boring part of a stack — wiring an env var into a deploy target — used to leak across three files. One yield now collapses it into a line that's also a typed runtime accessor.

// alchemy.run.ts — declare once on the Worker
export default Cloudflare.Worker("Api",
  { main: import.meta.filename },
  Effect.gen(function* () {
    const apiKey = yield* Alchemy.Secret("OPENAI_API_KEY");
    //    ^? Output<Redacted<string>>

    return {
      fetch: Effect.gen(function* () {
        // …and read the bound value inside the handler.
        const key = yield* apiKey; // Redacted<string>
        return HttpServerResponse.text(
          `key has ${Redacted.value(key).length} chars`,
        );
      }),
    };
  }),
);

Alchemy.Variable is the same shape without Redacted:

const port  = yield* Alchemy.Variable("PORT", 3000);
const flags = yield* Alchemy.Variable("FLAGS", { beta: true });

// inside fetch
const p = yield* port;   // number — 3000
const f = yield* flags;  // { beta: true }

Both accept a literal, an Effect, a Config, or default to reading the value from the active ConfigProvider under the same name. The same call routes to the platform's native secret/variable binding — Cloudflare secret_text for Secret, Lambda encrypted env vars on AWS — and the runtime accessor decodes back to the original type.

The docs are at Concepts › Secrets and Secrets & env.

Worker-to-Worker bindings, fully typed

A Worker is now bindable as a binding on another Worker, and the caller gets full RPC types on the other side — no codegen, no manual interfaces, no as casts on env. Three call shapes are supported, all on the same deployment.

The example: a Backend Worker exposing both an RPC method and an HTTP route, plus a TanStack Start frontend calling it three different ways. (This is the examples/cloudflare-tanstack project — cut down to the relevant pieces.)

The backend Worker:

// src/backend.ts
import * as Cloudflare from "alchemy/Cloudflare";
import * as Effect from "effect/Effect";
import { HttpServerRequest } from "effect/unstable/http/HttpServerRequest";
import * as HttpServerResponse from "effect/unstable/http/HttpServerResponse";

export const Bucket = Cloudflare.R2.Bucket("Bucket");

export default class Backend extends Cloudflare.Worker<Backend>()(
  "Backend",
  { main: import.meta.filename },
  Effect.gen(function* () {
    const bucket = yield* Cloudflare.R2.ReadWriteBucket(Bucket);

    return {
      // RPC method — callable via `backend.hello(key)` on the other side.
      hello: Effect.fn("Backend.hello")(function* (key: string) {
        const object = yield* bucket.get(key);
        return object === null ? null : yield* object.text();
      }),

      // HTTP handler — callable via `env.BACKEND.fetch(...)`.
      fetch: Effect.gen(function* () {
        const request = yield* HttpServerRequest;
        const key = new URL(request.url, "http://backend").searchParams.get("key");
        if (!key) return HttpServerResponse.text("missing key", { status: 400 });

        if (request.method === "GET") {
          const object = yield* bucket.get(key);
          return object === null
            ? HttpServerResponse.text("not found", { status: 404 })
            : HttpServerResponse.stream(object.body);
        }
        return HttpServerResponse.text("method not allowed", { status: 405 });
      }),
    };
  }).pipe(Effect.provide(Cloudflare.R2.ReadWriteBucketBinding)),
) {}

Wire it into another Worker as a binding:

// alchemy.run.ts
import Backend, { Bucket } from "./src/backend.ts";

export const Website = Cloudflare.Website.Vite("Website", {
  bindings: {
    BUCKET: Bucket,       // R2 binding
    BACKEND: Backend,     // Worker-to-Worker binding
  },
});

Now the caller has three ways to talk to the backend. All three are real, all three are typed, all three work in the same handler — pick whichever fits the call site.

// frontend route handler
import * as Cloudflare from "alchemy/Cloudflare";
import type Backend from "../backend.ts";
import { env } from "../env.ts";

// Option 1 — async binding (just call the platform API directly).
const object = await env.BUCKET.get(key);

// Option 2 — Worker-to-Worker fetch over the service binding.
const res = await env.BACKEND.fetch(`https://backend/?key=${encodeURIComponent(key)}`);

// Option 3 — typed RPC. `toPromiseApi<Backend>` wraps the wire-shape
// binding into a Promise<T> view that throws on `Effect.fail` and
// unwraps stream envelopes — full method signatures from `Backend`.
const backend = Cloudflare.toPromiseApi<Backend>(env.BACKEND);
const value = await backend.hello(key);
//    ^? string | null  (typed end-to-end)

Effect-native callers also get a fourth path — yield* Backend.bind(env.BACKEND) returns the same RPC surface without the Promise envelope. The HTTP option (#2) is the right one when you need request/response semantics with streaming bodies; the RPC option (#3) is the right one when you want typed method calls.

Tutorial › Vite SPA + Worker bridge · Example › cloudflare-tanstack

Workflows with typed input and output

Cloudflare.Workflow is now generic over input and output types. The body is an Effect.fn that takes the typed input directly; workflow.create(input) is type-checked end to end; the returned value flows through to instance.status().output.

A realistic example — a notifier workflow that touches KV, reads an Alchemy.Secret, broadcasts through a Durable Object, sleeps, and finalizes — with each side effect wrapped in a task so a crash + replay returns the persisted result instead of re-running:

// src/NotifyWorkflow.ts
import * as Alchemy from "alchemy";
import * as Cloudflare from "alchemy/Cloudflare";
import * as Effect from "effect/Effect";
import * as Redacted from "effect/Redacted";
import { KV } from "./KV.ts";
import Room from "./Room.ts";

export default class NotifyWorkflow extends Cloudflare.Workflow<NotifyWorkflow>()(
  "Notifier",
  Effect.gen(function* () {
    // Outer init phase: resolve shared dependencies once.
    const rooms = yield* Room;
    const kv = yield* Cloudflare.KV.ReadWriteNamespace(KV);
    const secret = yield* Alchemy.Secret("WORKFLOW_SECRET");

    return Effect.fn(function* (input: { roomId: string; message: string }) {
      const { roomId, message } = input;

      // Each `task` is a checkpoint — replay-safe.
      const stored = yield* Cloudflare.Workflows.task("kv-roundtrip",
        Effect.gen(function* () {
          const key = `notify:${roomId}`;
          yield* kv.put(key, message);
          return (yield* kv.get(key)) ?? message;
        }).pipe(Effect.orDie),
      );

      const value = Redacted.value(yield* secret);
      const processed = yield* Cloudflare.Workflows.task("process",
        Effect.succeed({ text: `Processed: ${stored}`, secret: value }),
      );

      yield* Cloudflare.Workflows.task("broadcast",
        rooms.getByName(roomId).broadcast(`[workflow] ${processed.text}`),
      );

      yield* Cloudflare.Workflows.sleep("cooldown", "2 seconds");
      yield* Cloudflare.Workflows.task("finalize",
        rooms.getByName(roomId).broadcast(`[workflow] complete for ${roomId}`),
      );

      return processed;
    });
  }),
) {}

Start it from a Worker — create is typed against the input shape, instance.status() reports the typed output:

// inside a Worker's fetch handler
const notifier = yield* NotifyWorkflow;

const instance = yield* notifier.create({
  roomId: "room-42",
  message: "hello",
});
// instance.id: string

const status = yield* (yield* notifier.get(instance.id)).status();
// status.output: { text: string; secret: string } | undefined

Tutorial › Workflows

Cron triggers via Cloudflare.Workers.cron(...)

Subscribe to a Cloudflare Cron Trigger with an Effect handler. The deploy-time half attaches the cron expression to the host Worker; the runtime half registers a scheduled listener. Use it alongside any other binding you've already wired up — the handler runs inside the same Worker context.

export default Cloudflare.Worker("Reporter",
  { main: import.meta.filename },
  Effect.gen(function* () {
    const kv = yield* Cloudflare.KV.ReadWriteNamespace(Counters);

    // Fires once at the top of every hour.
    yield* Cloudflare.Workers.cron("0 * * * *").subscribe((controller) =>
      Effect.gen(function* () {
        yield* kv.put(`tick:${controller.scheduledTime}`, "ok");
        yield* Effect.log(`tick at ${new Date(controller.scheduledTime).toISOString()}`);
      }),
    );

    return { fetch: Effect.succeed(HttpServerResponse.text("ok")) };
  }),
);

Multiple cron("…") calls register multiple schedules on the same Worker; the cheapest cron granularity Cloudflare offers is one minute (* * * * *).

— Thanks to Dawson (#288) for the contribution.

Analytics Engine binding

Cloudflare Workers Analytics Engine, exposed as a zero-provisioning Worker binding. Declare a dataset resource, bind it on a Worker, and call writeDataPoint from the handler — same Effect-error channel as every other Alchemy binding.

// alchemy.run.ts
export const Events = Cloudflare.AnalyticsEngine.Dataset("Events", {
  dataset: "app-events",
});

// inside the Worker
export default Cloudflare.Worker("Api",
  { main: import.meta.filename },
  Effect.gen(function* () {
    const analytics = yield* Cloudflare.AnalyticsEngineDataset.bind(Events);

    return {
      fetch: Effect.gen(function* () {
        yield* analytics.writeDataPoint({
          indexes: ["account-1"],   // queryable, low-cardinality
          blobs: ["signup"],        // arbitrary string columns
          doubles: [1],             // numeric metrics
        });
        return HttpServerResponse.text("recorded");
      }),
    };
  }).pipe(Effect.provide(Cloudflare.AnalyticsEngineDatasetBindingLive)),
);

— Thanks to Dawson (#286) for the contribution.

R2 buckets empty themselves on destroy

destroy on an Bucket now drains the contents before deleting the bucket. No more BucketNotEmpty failures during teardown — ephemeral PR previews and integration tests tear down cleanly with a single alchemy destroy.

const Photos = Cloudflare.R2.Bucket("Photos");
// `alchemy destroy` empties Photos and deletes it in one go

If you want the old behavior (fail if non-empty) for production guards, opt out per-bucket:

const Photos = Cloudflare.R2.Bucket("Photos", { emptyOnDestroy: false });

— Thanks to Michael K (#276) for the contribution.

Fixes worth knowing about

  • D1 prepare() / bind() are synchronous now. Matches the upstream Cloudflare Workers API — no more yield* on trivial statement construction.
  • WASM modules in the local sidecar bundle. bun alchemy dev now correctly bundles .wasm modules into the local sidecar, fixing a class of "module not found" errors for Workers that depend on WASM. Thanks to Baptiste Arnaud (#305).
  • Unresolved Output JS-coercion throws. Accidentally using an unresolved Output<string> in a template literal (e.g. `${bucket.bucketName}` outside an Effect) previously coerced to "[object Output]" and shipped garbage to the cloud. It now throws. Thanks to Zé Yuri (#306).
  • Deprecated libsodium wrapper types removed. No public API impact; if you were importing internal types they're gone. Thanks to 齐天大圣 (#311).

Contributors

Big thank-you to everyone who shipped code in this beta:

  • Michael K — R2 empty-on-destroy (#276)
  • Dawson — Worker cron triggers (#288)
  • Dawson — Analytics Engine binding (#286)
  • Baptiste Arnaud — WASM in local sidecar bundle (#305)
  • Zé Yuri — throw on unresolved Output JS-coercion (#306)
  • 齐天大圣 — remove deprecated libsodium wrapper types (#311)

Where to go next