A proposal is built and tested before its pull request opens, and fixes its own failing CI
A change an automation or the issue assistant writes now runs the repository's own build and test commands in the sandbox, after the static tools: the root package.json's build and test scripts (or the nearest package's), pytest for Python, or the commands you name as fix.build and fix.test. What fails goes back to the agent once, and the pull request says what was run, how it ended, and what wasn't run. Until now only the type checker, linter and Semgrep ran, so a change that broke the tests opened as if checked.
When CI then fails on the pull request the agent opened, the run continues once with the failing job, step and the log around the failure: the agent pushes a fix, or says why the failure isn't its own. Later failures are left to you. Turn it off with Fix a failing check on a proposal's pull request, or fix.ci in .auditor.yaml.
The docs said an implemented issue was checked in the same sandbox as fixes; that overstated it for monorepos and for anything but TypeScript and JavaScript, and now reads as it is.
More in Checking a proposal.