# Retries

Build bounded retries with Retry-After, exponential backoff, jitter, and cancellation.

Retry only failures likely to be temporary: `429`, `502`, `503`, and `504`. Do not retry an unchanged `400`, `401`, `403`, `404`, or `413` request.

## Recommended policy

1. Use `Retry-After` when present.
2. Otherwise start near one second and increase exponentially.
3. Add random jitter so clients do not synchronize.
4. Set a total deadline and a small attempt limit.
5. Cancel work the user no longer needs.

```python
import random
import time

delay = retry_after_seconds or min(8.0, 0.8 * (2 ** attempt))
time.sleep(delay * random.uniform(0.8, 1.2))
```

## Replay safety

Text generation is not idempotent: a retry can produce a different answer. A connection can also fail after the model generated output. Before replaying tool-driven or side-effecting workflows, determine whether your application already executed any returned tool call.

KosCompute does not currently expose an idempotency-key contract for generation. Enforce idempotency in your application around external side effects.