feat(action): local bindings + Output accessors in Actions

[sam-goodwin]

Actions can now bind resources with the current credentials and read resource Outputs, the same way a Cloudflare.Worker does.

const Seed = Alchemy.Action(
  "Seed",
  Effect.gen(function* () {
    const db = yield* Cloudflare.D1.QueryDatabase(database);
    const databaseId = yield* database.databaseId; // accessor captured at init

    return Effect.fn(function* () {
      yield* db.exec("CREATE TABLE IF NOT EXISTS users (id TEXT, name TEXT)");
      yield* db
        .prepare("INSERT INTO users (id, name) VALUES (?, ?)")
        .bind("1", "Ada")
        .run();
      return { databaseId: yield* databaseId };
    });
  }).pipe(Effect.provide(Cloudflare.D1.QueryDatabaseLocal)),
);

*Local binding layers

A third binding variant alongside *Binding (native Worker) and *Http (scoped token). A *Local layer resolves the current CLI credentials (ambient during stack-eval via the stack's providers context), never touches a Worker host or host.bind, and returns the same client backed by the provider's HTTP API. Where an *Http variant already exists, its client builder is refactored to take an injectable { authorize, accountId } auth shared by both the token-scoped Http and the current-creds Local; otherwise the Local layer implements a direct HTTP client (D1's shim is the reference).

Implemented for every Cloudflare capability with an HTTP data-plane — each live-tested against real Cloudflare:

Service	Local layers
D1	QueryDatabaseLocal
KV	ReadNamespaceLocal / WriteNamespaceLocal / ReadWriteNamespaceLocal
R2	ReadBucketLocal / WriteBucketLocal / ReadWriteBucketLocal
Queues	WriteQueueLocal
DNS	ReadDnsLocal / WriteDnsLocal / ReadWriteDnsLocal
Vectorize	SearchIndexLocal
Tunnel	ReadTunnelLocal / WriteTunnelLocal / ReadWriteTunnelLocal
AI Search	QuerySearchLocal / QuerySearchNamespaceLocal
Flagship	ReadFlagsLocal
Workers	BrowserLocal (Browser Rendering REST API)

Worker-runtime-only bindings have no Local variant and are intentionally excluded (no account HTTP data-plane): SecretsStore (write-only secret values), AnalyticsEngine (write-only ingestion), Images (in-Worker transforms), Hyperdrive (connection pooler), Email (send_email), Addressing, Artifacts, WorkersForPlatforms dispatch, AI QueryGateway (inference), RateLimit, VersionMetadata, and service bindings.

Output accessors in Actions

An Action's init runs at stack-eval (before resources exist); its body runs at apply. To make yield* db.databaseId work, RuntimeContext is split into two cooperating halves:

• a capture context provided around the init — records referenced Outputs and hands back deferred accessors
• a resolve context provided around the body — returns values evaluated against the tracker

Captured Outputs also become dependency edges, so the Action waits for the bound resource. Works in both the inline form and the tagged .make form.

Verification

• test/action.test.ts — inline + tagged .make capture/accessor tests.
• One live *Local test per implemented service (deploy resource + Action driving the Local binding).
• Workspace bun tsc -b clean.

Follow-up

Vectorize insertIndex/upsertIndex are generated with a multipart part named body, but the v2 API requires a part named vectors; SearchIndexLocal works around it in-layer. A distilled generator/spec fix would let every consumer get the correct multipart field for free.

[alchemy-version-bot]

Install the packages built from this commit:

alchemy

bun add alchemy@https://pkg.ing/alchemy/031d538

@alchemy.run/better-auth

bun add @alchemy.run/better-auth@https://pkg.ing/@alchemy.run/better-auth/031d538

@alchemy.run/pr-package

bun add @alchemy.run/pr-package@https://pkg.ing/@alchemy.run/pr-package/031d538