Imported from resonatehq/resonate-skills (resonate-human-in-the-loop-pattern-java/SKILL.md). Install upstream with npx skills add resonatehq/resonate-skills --skill resonate-human-in-the-loop-pattern-java. Copyright stays with the author (Apache-2.0).

Resonate Human-in-the-Loop Pattern — Java

Prerelease note. resonate-sdk-java is published on Maven Central — pin io.resonatehq:resonate-sdk-java:0.1.1. The API mirrors the Python SDK and may change before a stable 1.0. Requires Java 21+ (virtual threads, a feature generally available in Java 21). There is no Java HITL example repo yet, so every code block here uses only documented SDK surface, compile-verified against 0.1.1 and cross-checked against develop/java.mdx (docs PR #230) and the SDK source.

Overview

For the language-agnostic mental model, start with resonate-human-in-the-loop-pattern-typescript. The idea is identical: create a latent durable promise, hand its ID to the external actor who will settle it, and await — the workflow parks until settlement arrives, surviving any number of crashes or restarts.

Resolution from Java is direct. The Java SDK ships a top-level r.promises sub-client (it mirrors Python), so settling a promise from outside the workflow is r.promises.resolve(id, new Value(...)). There is no manual base64 encoding — that extra step is a Go-specific workaround for a sub-client Go lacks. Don't copy it into Java.

When to use

• Approval gates (budget, deploy, content moderation)
• Third-party webhook callbacks (Stripe, DocuSign, Twilio)
• Operator unblock steps in runbooks
• Any step where the decision or data originates outside the Resonate worker set

Basic shape

Workflow side — ctx.promise() → promise.id() → publish → await

import io.resonatehq.resonate.Context;
import io.resonatehq.resonate.Context.ResonateFuture;
import java.time.Duration;

public final class Approval {
    private Approval() {}

    public record ReviewRequest(String item, String requester) {}

    /**
     * Parks until an external actor settles the latent promise. promise.id() returns the id once
     * the promise has been created (instantaneous in local mode, a network round-trip against a
     * real server); await suspends the workflow (durably) until it is settled from outside.
     */
    public static String approvalWorkflow(Context ctx, ReviewRequest req) {
        // A latent durable promise — no registered function backs it; it settles only when an
        // external caller resolves it. 24-hour timeout, capped at the workflow deadline.
        ResonateFuture<Object> promise = ctx.promise(Duration.ofHours(24));
        String approvalId = promise.id();

        // Publish the promise ID inside a ctx.run so the side effect is checkpointed and does
        // not re-run on replay. In production: write to a DB, push to a notification queue, etc.
        ctx.run(Approval::publishApprovalId, req.item(), approvalId).await();

        // Suspend until the promise is settled externally; decode the decision.
        Object decision = promise.await();
        return "item " + req.item() + " decided: " + decision;
    }

    /** Checkpointed publication of the promise ID. Replace the println with a real notification. */
    public static String publishApprovalId(Context ctx, String item, String approvalId) {
        System.out.printf("  [workflow] awaiting approval for %s — promise id: %s%n", item, approvalId);
        return "published";
    }
}

Key points:

• ctx.promise() takes an optional Duration timeout; ctx.promise() with no arg uses a 1-day default capped at the parent's remaining deadline.
• promise.id() returns the promise's id once it has been created — instantaneous in local mode, blocking on the create round-trip against a real server.
• Publish the ID inside a ctx.run so the publication is itself durable. A bare side effect above promise.await() re-runs on every replay pass.
• promise.await() with no decode is fine when you only need to know the promise settled; here we read the decision Object.

Resolving from outside

The Java SDK gives you three mechanisms. Prefer the sub-client when resolving from Java.

1. The promises sub-client (preferred when resolving from Java)

import io.resonatehq.resonate.Types.Value;

// approve — the Value's data is decoded into what the workflow's await expects
r.promises.resolve(approvalId, new Value(null, "approved")).join();

// reject
r.promises.reject(approvalId, new Value(null, "rejected: budget exceeded")).join();

// cancel
r.promises.cancel(approvalId, new Value(null, "withdrawn")).join();

new Value(headers, data) — pass null headers and the decision as data. Each call returns a CompletableFuture<PromiseRecord>, so .join() (or compose with thenApply) to wait for it. r.promises also exposes get, create, and search. No base64 encoding is required — the SDK codec handles the durability boundary for you.

2. CLI (simplest — ops gates, manual approval)

resonate promise resolve <approvalId> --data '"approved"'
resonate promise reject  <approvalId> --data '"rejected: budget exceeded"'

The --data value is the JSON payload decoded into the type the workflow's await expects. For a String-shaped decision, a quoted JSON string like '"approved"' is correct. See resonate-cli for the full flag reference.

3. Server HTTP API (any language / tool)

POST to the server's promise-settle endpoint with the promise ID and a JSON body — useful from a webhook handler, an admin script, or a serverless function in any language. Going straight to the server bypasses the SDK codec, so the data field must be the base64 encoding of the JSON-serialized payload (the wire format the SDK stores — value → JSON → base64, Codec.java). "approved" as JSON, base64-encoded, is ImFwcHJvdmVkIg==:

curl -s -X POST "http://localhost:8001/promises/${APPROVAL_ID}/resolve" \
     -H "Content-Type: application/json" \
     -d '{"value":{"data":"ImFwcHJvdmVkIg=="}}'

This base64 step is exactly what r.promises.resolve and the CLI do for you — prefer mechanism 1 or 2 from Java/ops, and reach for the raw endpoint only from a non-Java caller. The ImFwcHJvdmVkIg== value is correct for the default NoopEncryptor; if you configured an Encryptor on the builder, the payload is encrypted before base64, so a raw caller would have to replicate that encryption too — another reason to use the sub-client or CLI.

Hand-off across processes

A realistic deployment has the worker that creates the promise and the process that resolves it in different programs. The pattern:

1. The worker workflow creates the promise and persists promise.id() somewhere the resolver can read it (a DB row, a notification, a ticket).
2. The external actor (an HTTP webhook handler, an operator running the CLI, an admin service) settles it by ID via one of the three mechanisms above.
3. The worker's promise.await() wakes and returns the decoded decision — even if the worker crashed and restarted in between.

Because resolution is by promise ID, the resolver needs nothing but the ID and a connection to the same Resonate server.

Known gaps

• No hand-chosen promise IDs. ctx.promise() generates the ID internally; you cannot pass an application-layer ID like "approval/order-42". Fetch the ID via promise.id() after creation and publish it explicitly. This may change as the API settles toward 1.0.

Distinct Java idioms

• r.promises.resolve(id, new Value(null, data)) — the clean external-resolution path. No base64 layer (that's a Go-only workaround). Value(headers, data) takes null headers and the payload as data.
• CompletableFuture from the sub-client — r.promises.resolve(...) returns a CompletableFuture<PromiseRecord>; call .join() to wait or thenApply to compose.
• Publish the ID inside ctx.run. Java durable functions are not generators (no yield); wrap every observable side effect — DB write, notification, the hand-off that publishes the promise ID — in a ctx.run so it is checkpointed and not repeated on replay.
• ctx.promise(Duration) vs ctx.promise() — the Duration overload sets an explicit timeout; the no-arg form uses the 1-day default capped at the parent deadline.
• promise.await() returns Object — ctx.promise() is untyped; cast or .toString() the decoded decision as needed.

Avoid

• Copying the Go base64 encoding dance. The Java SDK has r.promises.resolve — use new Value(null, data) directly. Manual JSON → base64 → quoted encoding is a Go-only workaround for a sub-client that Go lacks.
• Polling via ctx.sleep + a status check. Defeats park-and-resume; burns checkpoints and wall-clock time. Use ctx.promise() + promise.await().
• Publishing the promise ID outside a ctx.run. A bare write above promise.await() re-runs on every replay. Checkpoint the publication.
• Settling before publishing the ID. A narrow race: if the resolver runs before the ctx.run that publishes the ID is checkpointed, a crash between those lines can lose the ID. Checkpoint the publication first, then await.
• Long-blocking work inside ctx.run. ctx.run functions must return promptly; a blocking call holds the task lease open until TTL expires. Long waits belong in ctx.promise (latent, externally settled).

Related skills

• resonate-basic-durable-world-usage-java — ctx.promise, ResonateFuture.await, the replay model
• resonate-basic-ephemeral-world-usage-java — the r.promises sub-client surface (resolve / reject / cancel / get / create / search)
• resonate-cli — resonate promise resolve / reject / cancel flag reference
• durable-execution — foundational replay semantics; why await survives crashes
• resonate-human-in-the-loop-pattern-typescript — language-agnostic mental model
• resonate-human-in-the-loop-pattern-python — the closest sibling; the Java promises sub-client mirrors Python