Waits and input

Helmr exposes three durable waiting patterns:

Need	Primitive
Continuing commands, corrections, and progressive output	Actor Session
One external result such as an approval or callback	Token
Resume after a duration or timestamp	Timer

Actor input is an ordered log owned by one Session. The Actor calls session.input.receive() and external callers append JSON with an optional idempotency key. Output is a separate ordered log with independent readers. This is the right shape for a conversation or a workflow that can be steered more than once.

A Token represents one pending value. tokens.create() returns the Token plus a callback URL and public access credential. The owning Run waits with token.wait(), optionally validating the completion through a Standard Schema. Completed, cancelled, and expired are terminal Token states. The narrow public completion capability is preferable when an outside user should not gain access to an Actor channel.

Timers park only on time:

import { timers } from "@helmr/sdk"

await timers.waitFor("15m")
await timers.waitUntil(new Date("2026-08-10T09:00:00Z"))

A waiting Run has released active execution but retains durable intent to resume. Set application-level timeouts for input and Tokens, and handle wait_timeout, closed Sessions, cancelled Tokens, and expired Tokens as normal branches. Avoid sending secret data through input, output, Token results, or wait metadata.