Finished runs let go straight away

Every run the portal does for you — a script, a notebook cell, a compile, a build, a test — writes its progress to an activity. While the run is going, the portal keeps a small live connection open so the page showing you that progress updates as it happens. That connection is not free: it checks in every 45 seconds, on purpose, to keep the machinery behind it from being cleaned up mid-run.

When the run ends, that connection should end with it. The end of a run is a fact, not a guess, so there is nothing left to wait for.

For most runs it already worked that way. Three kinds did not — including, awkwardly, the most common one in the whole platform: the ordinary log a script or a test writes as it goes. Those runs finished, and their connection stayed open for up to ten more minutes, until a periodic sweep noticed nobody was using it. One stale run is nothing. A busy portal doing this all day is carrying ten minutes of finished work at all times, for no reason.

All three now release the moment the run reaches its end state.

There was a related problem at one of the three, and it is fixed in the same change: when a code run failed because nothing was there to execute it, the step meant to mark it failed could quietly do nothing at all, depending on the exact shape the data happened to arrive in. The run then sat at Running for ever, showing a spinner for something that had already given up. It is now marked failed reliably.

One detail worth stating, because it is what makes this safe: a finished run whose page you are still looking at keeps its connection. The release asks first, and simply declines while anyone is watching.