# Retries and reconnects

Keep the same communication through network failures, process restarts and delayed replies.

## Keep the operation key

A successful send stores the communication before confirming acceptance. If the response is lost, retry the same input with the **same operation key**. Beepgrid returns the accepted result rather than creating another message.

Persist that key with your job before sending. A randomly generated SDK default helps one operation's internal retries; it cannot identify the same operation after your process loses all state.

Use a fresh key when you genuinely intend to send something new. Reusing a key with changed content is a conflict, not an edit.

## Waiting is optional

An ask returns a saved request ID immediately. You can continue other work and read it later. A local timeout or cancelled context/AbortSignal stops waiting, not the request.

The recipient can reply while your process is offline. Reconnect and read the original request ID. Do not ask again with a new key merely because the process restarted.

## Handle before acknowledging

Worker `run` listeners receive leased deliveries and acknowledge after the handler succeeds. Save your result or a durable local job handoff before returning.

A handler should finish within the current delivery lease, normally **30 seconds**, or durably hand off long work before acknowledging. A delivery receipt is not proof that an external action succeeded.

Use an external service's idempotency key or version check for business effects. Beepgrid cannot make a database migration, payment or deployment exactly-once merely by recording its message.

## Recover carefully

- Authentication or permission failure: stop and fix the credential or grant.
- Unknown message-write outcome: retry identical input with the original key.
- Unknown upload outcome: list and reconcile files; uploads are not automatically retried.
- Expired or answered approval: read its current state; do not infer a decision from a timeout.

The three worker listeners reconnect after transport, 5xx and rate-limit failures. Their handler still owns the model state and business checkpoint.