Preparation context
Use one private temporary directory per target. Keep the source clone, Python
environment, caches, and .skillroll/runs isolated from other trials. Do not
run repository-provided setup or cleanup scripts merely to prepare SkillRoll.
Pin what will be tested
Record:
- canonical target URL and exact checked-out commit;
- active skills root and discovered skill paths;
- exact local SkillRoll commit and whether its worktree was clean;
- whether remote freshness was actually checked;
- generated configuration and workflow paths; and
- commands run, excluding secret values.
Do not fetch, pull, switch branches, stash, or alter the user's SkillRoll
worktree without authorization. A dirty local checkout is not the same thing as
its main commit; build the recorded commit in a separate clean tree and
isolated environment, or label the trial non-comparable and stop. Do not install
globally.
Keep setup inert
Use skillroll init --skills-path <path> --yes or the equivalent explicit
configuration. Run skillroll validate --all during preparation. Validation
is offline by default and repository commands require explicit opt-in. Unset
the inference credential during both steps and record only its variable name.
Do not call doctor or eval here.
If advisory CI is requested, use SkillRoll's generated workflow. Never replace an existing path. Snapshot and recheck any colliding workflow byte-for-byte. Keep generated CI advisory, review the diff and fork-PR guards, and leave secrets in the configured environment secret rather than repository files. Do not commit or push the temporary target.