I shy away from CL scripting due to runtime startup time, though if you compile it down to binary it's manageable (if one doesn't mind a 50MB executable).

That said, I wonder why nobody tried a "hybrid" solution, in which the multi-call binary is implemented as a resident, self-restarting process, that executes each request in a lightweight sandbox. Actual shell commands would then be implemented as a request to that process, and incur only the costs of setting up IO streams on startup.

what CL implementation ? sbcl ? I don't know what's the leanest CL out there (even if not the most featured, since we're talking about short scripts)

SBCL mostly. I probably should give GCL a try here. SBCL is fast as hell once it starts up, and bare SBCL starts up quickly, but once you add Quicklisp and start loading external systems, the startup time gets noticeable - though again, this gets reduced quickly iff you dump a binary of your image with dependencies already loaded. The one CL script I wrote that I actually use[0] (for controlling Hue lamps) I have dumped into binary just to cut down startup time to below the point I notice.

--

[0] - https://github.com/TeMPOraL/hju/

SBCL does have a tiny delay (akin to a raw JVM). CLISP is near interactive though. But that's without quicklisp.