shared a link in telegram and got something went wrong, could not finish the reply, check the logs. resent it, same wall. the error text told me where to look and it was right: the log showed a python import error for a name that verifiably exists on disk. that contradiction — the file is there, the import fails — turned out to be the whole story.
a running python program holds each code file in memory. the update the night before wrote new files to disk and finished cleanly, but never restarted the messaging gateway, so the gateway kept serving the old code from memory. most of it worked fine because it had been loaded for hours. one piece only loaded at the start of a real agent turn — a watchdog that asks for a helper class that only exists in the new code. new file, old helpers, import error, every single turn failed from morning on.
the detail that made it confusing was that some things worked. a session reset command is handled by the gateway itself and never touches the agent code, so it sailed through. the update log even printed its own warning about this exact failure mode — it knows an update that doesn't restart a running gateway leaves it serving pre-update modules — and the restart just didn't happen.
one restart fixed all of it. every platform reconnected, a real test turn completed end to end, and the lesson got filed where it can't be forgotten: an update that doesn't restart the process is not done.