How-to guides

Each how-to is a recipe for one task. They assume you've read the Concepts (or are happy looking things up as you go).

• Wire a non-idempotent upstream Wires, payments, one-shot side effects. The outbox pattern, end-to-end.

• Configure reactors Enable, disable, scale, swap polling for event-driven.

• Pick a storage backend SQLite, Postgres, AlloyDB, Bigtable, Spanner. Trade-offs and operating notes.

• Add observability Structured logs, OTel spans, log-based metrics, the bundled dashboard.

• Configure tenancy Single, trusted-multi-app, hard-multi-tenant — what each mode means today.

• Cancel & timeout patterns tape.cancel_run, gates with gate_timeout, tape.heartbeat, cooperative cancellation.

• Write a custom connector Implement EffectConnector — dispatch, observe, compensate.

• Run the outbox dispatcher in any language tape-outbox-ts, cmd/tape-outbox, dev.tape.cli.TapeOutbox — one dispatch loop, one safety contract, four languages.

• Fan the journal out (sinks) LogSink / WebhookSink / PubSubSink in every SDK — exactly-once-effective delivery with consumer-side dedup.

• Cross-SDK parity harness One scenario, four languages, identical journal projection on every PR.

How to read these

Each how-to:

• Starts with what you want to accomplish.
• Gives the smallest code change that does it.
• Lists the gotchas that bite people in production.
• Links to the relevant concept page and reference page.

If you find yourself reading a how-to for context, switch to the concept page. If you find yourself reading a concept page for "which exact parameter," switch to the reference. The three are different modes of documentation, on purpose.

Don't see your task?

• Check the FAQ — many "how do I" questions are one-liners.
• File an issue at vamsiramakrishnan/durable-agents describing what you wanted to do and what the docs didn't help with. That's the fastest way to make this section better.