Sessions, failover, and a full disk

A session in the terminal client is mirrored into your account as a chat while it runs, and what you add on the site comes back into the client when you open that session again. Both directions carry a pointer, so nothing is written twice.

Failover used to react only to an error a provider returned. A provider that never answered at all was a dead request. Network failures and timeouts move the request to the next candidate now, and when every candidate is exhausted the message names the real cause instead of reporting an empty balance.

The design engine

Design runs had produced nothing usable for days, for five separate reasons: a variable declared outside the scope that read it, a package manager that could not authenticate to the egress broker and reported success anyway, a file reader that truncated silently at 100 kilobytes, a text replacement that ate dollar signs inside the bundle, and a content policy that blocked the inline script. All five are fixed, and a run renders again.

Disk

The release script left a 1.4 gigabyte directory behind on every run. Eighteen of them had filled the disk to 98 percent, and the sandbox refused to start new containers. The script cleans up after itself now. The warning that should have reported this was reachable only from the one code path a full disk prevents; it runs on the periodic check now.

Reloading a chat while it streams keeps streaming. The client had been reconnecting correctly the whole time and never told the view about it.