Closes #15174

We originally noticed that last_used_at was constantly ticking upwards even after clients disconnect. After narrowing it down to the web terminal I did some digging and found that agentssh.Bicopy was never exiting even after the client disconnects. This meant our websocket connections on the server were held open forever? and this continually counted as an open connection to the workspace and bumped last_used_at as a result.

https://github.com/coder/coder/blob/f0ssel/last_used_at_inc/coderd/workspaceapps/proxy.go#L702

This is due to a combination of our use of agentssh.Bicopy as the primary reader of the websocket and our lack of timeout on our websocket.Ping attempt.

I've added a timeout equal to the loop interval for httpapi.HeartbeatClose which can now catch the newly failing websocket.Ping call and cancel the entire request context -- leading to the agentssh.Bicopy finally exiting and releasing the connection.

But what makes last_used_at get bumped forever?

Great question, I wasn't sure at first since I was logging the inputs and outputs and clearly saw stats from the terminal session stop coming through on disconnect. When I logged the contents of the workspaceapps.StatsCollector over time I noticed that the stats from the hung websocket connections were never getting cleaned up. After some digging I found that in order for a stats to get cleared out of the stat collector it must get a stat published with a non-zero ended_at time value. The /pty route would send this final stat in a defer block in the handler but since the handler never closed, we never got the stat, so it stayed refreshing forever.

I think there's some improvements we can make to make this safer in the future when we take a look at how to finally merge agent session stats and workspace stats together.