Prepare a blind trial
Create a disposable workspace from a preregistered target. Pin the target and the local SkillRoll build to exact commits, record dirty and freshness state, and install from the selected local SkillRoll checkout rather than silently substituting a published release.
Initialize SkillRoll around the repository's active skills root, validate the
generated files offline, and add advisory CI only when requested. Do not run
candidate repository commands, overwrite existing workflows, commit, push,
publish, or expose an inference key. Preparation evidence is not a live eval.
The final manifest must name generated paths and commands and explicitly state
whether doctor and eval ran.
Record the completed install or build from the pinned local SHA and the
completed initialization of the target skills root as separate steps. If a
combined setup result only reports the pins or files, obtain completion
evidence before reporting either step as done.
Read preparation context for isolation, secrets, and the trial manifest.
When advisory CI is requested, open the generated workflow directly after
creation, before exporting or reporting the manifest. With a generic action
wrapper, use Read with the exact generated workflow file as path.
For any secret-backed pull-request step or job, verify an explicit
same-repository guard that excludes fork PRs; do not use pull_request_target
to run fork-controlled code with secrets. Report the exact verified guard in
the manifest. A preparation summary that says “verified” without showing the
condition is insufficient. If direct inspection cannot verify the guard,
report a blocker.
For an existing workflow at the destination, report the path collision,
whether the file stayed byte-for-byte unchanged, and whether its own secret
boundary was verified. If only a byte comparison was authorized, report its
secret boundary as unverified.