Chat streaming that survives sleep and dropped Wi-Fi

Long agent runs used to look broken at exactly the wrong moments: a laptop waking from sleep, a train going through a tunnel, a flaky office network. The chat stream is rebuilt so reconnecting is clean. Your run keeps working on our servers the whole time, and the chat catches up exactly where it left off.

How it works

• Every event is numbered. After a reconnect, the chat applies only newer events, each exactly once, so text never appears twice.
• Fewer, larger updates. Text is batched into about 25 updates per second instead of one per token, in strict order with tool and status updates.
• Dead connections are caught fast. The server checks in every 15 seconds, and 40 seconds of silence forces a reconnect even when the browser thinks the connection is fine. A network change or a tab waking up reconnects immediately.
• An outage isn't a failure. During a long drop, the chat keeps retrying and says the connection is out while your run continues.
• Specific errors. A run that ends on a provider problem now says why, such as rate limited or provider unavailable, with a next step.

Also shipped

• One memory-hungry command no longer takes down an agent's whole sandbox; only that command is stopped, with a note to process the data in smaller pieces.