> ## Documentation Index
> Fetch the complete documentation index at: https://docs.onecortex.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Timeouts and retries

> How long a call can run, how to keep a long stream open, which errors to retry and which not, and why Onecortex never retries a call for you.

## How long a call can run

A call runs as long as your agent takes, up to the life of its session's instance: an hour at most. There is no shorter limit on a single call.

Your client is usually the limit that matters. Most HTTP clients give up on a response that sends nothing for a while, and a proxy in front of your app may do the same.

* **Stream long calls.** With `"stream": true`, Onecortex sends a `: ping` comment every 15 seconds while your agent is silent, so the connection never looks idle.
* **Set your client's read timeout above 15 seconds**, or turn it off, for a streamed call. httpx's default is 5 seconds.
* **A JSON call sends nothing until the run ends**, so its read timeout has to cover the whole run.

## Onecortex never retries a call

A call reaches your agent once. If it fails, Onecortex does not call your agent again, so a tool with side effects, such as sending an email or charging a card, runs once per request you make. Whether to retry is yours to decide, with that in mind.

## What to retry

| Status and code | Retry? |
| - | - |
| `429` `rate_limited` | Yes, after the `Retry-After` header's seconds |
| `503` `agent_unavailable`, `service_unavailable`, `capacity_exceeded` | Yes, with backoff |
| `502` `upstream_error` | Yes, with backoff, unless the message says to redeploy |
| `409` `agent_not_ready` | Yes, once the first build succeeds |
| `500` `internal_error` | Once or twice; then send the `requestId` to support |
| `502` `agent_error` | Only if your agent's failure is transient: it is your code that raised |
| `400`, `401`, `403`, `404`, `409` `agent_not_deployed`, `410`, `413` | No. The request or the setup is wrong, and the same call fails the same way |

An `event: error` in a stream carries the same codes, and the same advice applies.

Back off exponentially from about a second, with jitter, and stop after a few attempts. Every response carries `X-Request-Id`: log it, so a failure can be matched to your agent's [logs](/observe/logs) and to support.

## Rate limits

Calls are limited per organization and per agent. See [Limits](/production/limits#invoke). A retry loop that ignores `Retry-After` spends the budget faster and fails for longer.