I think we will be running every single endpoint as a separate installation (to reduce the attack surface in the single container).

Apart from that, what you described can be covered by a single prometheus histogram, where we should add the following labels:

• Status code of the response (if correctly codified in the code, this will also give us authorization errors and badly formed requests)
• "endpoint" (that is, what program is being requested)
• *maybe* some salient data about the request, like if it contained a pipe?

That's basically all we need to get info about the performance of Shellbox.

But, if we're getting fancy, it would be great to also record the resource usage that php can report in terms of memory used by execution of an external program, although I guess that would require more coding on the shellbox side, and we don't really need that on day 1.