API: Repo Ablations - SelfBench

An environment runs one script in every approved task’s image and builds each task with the change. Pass its id in environments when you start a comparison, alone or with each saved harness picked.

Method	Route	Does
`GET`, `POST`	`{repo}/environments`	List environments, or create one and build every approved task in it
`GET`, `DELETE`	`{repo}/environments/{id}`	One environment with each task’s check, or delete it
`POST`	`{repo}/environments/{id}/build`	Build the approved tasks it lacks or that failed

Request Body

Route	Body
`POST {repo}/environments`	`{ environment: { name, script }, generation }`

script is Bash of at most 64 KB, run in /app as the image is built, before the agent starts: "rm -f AGENTS.md CLAUDE.md\npnpm add next@15.5.0". It stops at the first command that fails, and that task’s build fails with it.