Close a bounded APT build/test/merge/publish mission without an open-ended review loop. Load for Debian package releases, focused re-review after a passing matrix, a user asking whether a release is finished, or scope narrowed to packaging only.
Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
SSkills Directory ProSkills Directory
Get any skill into Claude in one click.
Are you the author of Bounded Apt Release?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/jgwill-bounded-apt-release)
---
name: bounded-apt-release
description: >
Close a bounded APT build/test/merge/publish mission without an open-ended review
loop. Load for Debian package releases, focused re-review after a passing matrix,
a user asking whether a release is finished, or scope narrowed to packaging only.
---
# Bounded APT release: review defects, then finish delivery
Earned in kit #135 / PR #136: the first adversarial review found material
sandbox, privacy and redistribution defects. It was necessary. Re-review then
ran over sixteen minutes despite a focused brief, while the six-container
matrix already passed; Guillaume had to ask whether the mission was finished.
1. Survey once, isolate the branch, preserve other work and record the approved
scope. A later direct correction wins: defer secondary work to an assigned
issue, not a running implementation lane.
2. Build pinned artifacts and run the supported-distro matrix. Require one
adversarial review; do not skip measured security or distribution blockers.
3. For a re-review, name the exact changed files/findings and require an interim
blocker list within five minutes. Budget each additional probe. Do not
silently launch another broad container fleet or repeat an already-passing
matrix without a finding that calls for it.
4. Launch long tests in a separate tmux session or process session, then verify
fresh completion markers. Do not put a `nohup` launch and a timeout-prone
foreground build in the same tool call: a timeout can terminate their process
group together. After any timeout, check actual worker liveness; a stale
`running` file is not a running test and an old exit 0 is not a new pass.
5. At the budget boundary, collect the report. A measured blocker is fixed and
regression-tested; silence is not approval. If the reviewer gives an exact
conditional approval, satisfy those conditions instead of commissioning
another general review. Missing real-host/account instruments are labelled,
never represented by Docker or anonymous sign-in tests.
6. Separate scripted checks from an agent-operated clean-host walkthrough.
Prove pinned artifacts, old-version cleanup, preserved data, signed APT
metadata and a real install. Keep account sign-in manual.
7. Finish the authorized sequence: commit → PR → merge → build/verify artifact
identity → publish in dependency order → signed-index/install readback →
canonical checkout on main → issue/connector receipt. Test success alone is
not shipment. Keep the user informed during a long-running step; never say
"finishing" while opening unrelated scope.
Reviews keep the release safe. A review that never closes is not the release.