Try again by itself after a temporary model error

Agents: claude, codex, opencode · On: phone, desktop

When a turn fails because the model is overloaded, the provider has a server error or the stream drops, the tab waits and tries again by itself: after 30 seconds, then 1, 2 and 5 minutes. After four tries it stops and shows the error.

Why you would use it

You start a long task from your phone and walk away. Codex answers Selected model is at capacity or Claude's API is overloaded for a minute. The work should carry on when the provider comes back, without you noticing the failure and sending it again.

How to use it

Nothing to do. While the tab waits you can:

Tap Retry now on the line to try at once.
Tap Cancel to stop waiting. The error stays, with Retry beside it.
Or type a message. Sending one cancels the wait: that is you taking over.

A Claude tab waiting to try again: the amber line at the end of the transcript with the countdown, Retry now and Cancel

What you see

A line at the end of the transcript with an amber ring:

Claude Code hit a temporary error. Trying again by itself, attempt 2 of 4.

Under it, in grey, the provider's own words, such as Codex's Selected model is at capacity. Please try a different model. On the right, the time left in amber (45s, 4m 05s), Retry now and Cancel.

The line close up: the sentence with the attempt, the provider's words, the countdown, Retry now and Cancel

When the time is up the line reads … Trying again now, attempt 2 of 4. and the next try starts. The tab tells the agent to continue from where it left off; your message is not sent again.

In the status bar, an amber segment counts down: ⟳ retry 45s. Tapping it shows the longer form as a toast, for example Claude Code temporary error · retry 2 of 4 at 14:21 (in 45s).

The toast after tapping the status bar's retry segment

After the fourth try fails, the line turns red and reads Claude Code hit an error. with the provider's words, as any error does.

Options and settings

Option	Default	What it changes
Wait before each try	30 s, 1 min, 2 min, 5 min	Each wait is up to 20% shorter or longer at random, so tabs that failed together do not all come back at the same moment. Fixed; there is no setting

Limits and known gaps

Only errors that go away by themselves are retried. These are:
- Claude: overloaded (529), a server error (5xx), a timeout or a lost connection.
- Codex: Selected model is at capacity, an internal server error, and a stream or connection that dropped. A drop that carries a 4xx status, such as a 401, is a login or request problem and is not retried.
- opencode: a provider error that names a server error, an overload or a timeout.
Never retried: a turn you stopped, a usage limit (that one waits for the reset), a logged-out agent, a bad request, or a prompt too long for the model.
Cursor tabs are not retried: no transient error from Cursor has been seen to tell apart from the others.
Each vendor already retries a few times before it gives up; this waits longer between tries than they do.
The count starts again from one after a try that gets through.
Retry on the red line after the fourth try sends your last message again, as it does for any error.
/clear, a handoff and a stop (v-code tab stop from another tab) cancel the wait.
Only a turn that failed is retried. A turn that hit a temporary error and then finished anyway is left as it is.