Settings

Sandbox

Docker on this machine, where agents run apart from it. Under Execution.

Card	Does
Docker	Whether it answers, and its version; why not when it does not
Docker Hub	Log in with a username and an access token; log out
Images	The images Nullius pulled or built, per benchmark, with their size; Remove takes one benchmark's

Rule
Login	The token goes to docker login --password-stdin once. Docker keeps it in its own store (~/.docker/config.json, or the credential helper it names), as a login typed in a terminal would; Nullius stores neither value and reads back the username only
Token	A read-only personal access token from Docker Hub's account settings is enough: runs only pull
Why log in	Docker Hub limits anonymous pulls; a benchmark with an image per task (SWE-bench Verified, about 4 GB each) stops part-way without a login
Images kept	Runs keep the images they fetch, so the next run on the same tasks pulls nothing; they go only from here, or with --remove-images (Run a benchmark)
Listed	Only images Nullius fetched (recorded in the standard's images.json); an image that was on the machine before is never offered for removal
Hosts	This machine only; a run host keeps its own login and images
Code	app/nullius/sandbox.py