Auto continue

Work can stop for two reasons that have nothing to do with the task. The provider’s usage window ran out, or the provider had a passing failure. Auto continue picks the conversation up again when it can.

A provider error

Warshall reads the words Claude and Codex print when a request fails for a passing reason. That covers an overloaded API, a 5xx, 429 or 408 error, a rate limit, a dropped stream, a reset connection and a failed network request.

The first such error on a terminal continues by itself. There is nothing to click. The line shows its own countdown, Auto continue in 58s, and after the minute the terminal goes on with the conversation.

If the run that went on stops on an error again, the terminal does not continue by itself a second time. That line offers an Auto continue button instead. Pick it and the send goes out after a ten second countdown, with the button reading Auto continue · 9s. Click it again to stop the send. A run you start yourself, or a continued run that ends without an error, makes the next error automatic again.

Each terminal keeps its own count. An error in one terminal changes nothing in another.

A usage limit

When a run ends on the provider’s own usage limit line, that line carries an Auto continue button. Click it and the button reads the time it will fire, such as Auto continue · 3:20 AM. Warshall waits for the window to come back and then goes on with the conversation. Click the button again to disarm it.

The time comes from your own account first. Warshall reads the reset of the window that is actually spent when you click the button. Every terminal on that account therefore agrees on the moment, however old the line it was armed on. The send goes out a little after that reset, so the provider has rolled the window over by then.

• If the account shows no spent window, the work can go on now, so the send goes out at once.
• A limit line that arrived moments ago and names a time still ahead keeps that time, since the account’s meter can lag behind the provider.
• The time written in the provider’s line is only used when the meter cannot be read, as on an API key. When neither gives a time, Warshall tries every fifteen minutes until a send gets through.

One turn only. The send goes out once and the terminal disarms with it. If the resumed run hits the limit again, that is a new line with a new button. Before the send, a newer limit line on the same terminal moves the send to the new time, and a run that ends without hitting the limit disarms it.

What gets sent

Neither provider resumes a conversation without a message, so Warshall sends one. Where the session is still on record it is the word continue. Where it is not, it is the terminal’s last prompt again, exactly as you typed it and with its attachments, because continue alone would open on nothing.

The send is the app’s, not yours, and the transcript says so. In place of a prompt the terminal leaves a plain line on the agent’s side, Auto continued at 3:21 PM. The status line under it reads Auto continuing… for the first few seconds and Working… after that. The send never appears as your own words. It is not in the restored transcript, the node panel, the phone or your recently used prompts.

Once and always

Beside Prompts on the composer’s writing row is the terminal’s own Auto continue button. Each click moves it to the next state.

The mode belongs to that terminal alone, is never saved to disk, and is off again after a restart.

Stopping it

• Sending a prompt during a countdown cancels it.
• Escape cancels it. So does closing the terminal.
• Nothing survives a restart. The armed state lives with the line in the transcript, and a line read back after a restart offers the button again, unarmed.