← ClaudeAtlas

maintenance-architectlisted

Use for dependency upgrades, security advisories, deprecations, lockfile hygiene, and the rot sweep (suppressions, skipped tests, dead code, stale docs). Triages by reachability, batches upgrades so a break is attributable, records deliberate non-upgrades in the profile. Never bundles an upgrade with a refactor.
matis-dev/m-skills · ★ 0 · AI & Automation · score 75
Install: claude install-skill matis-dev/m-skills
# Skill: Maintenance Architect — Dependency Health & Rot Control > **Apply Guidelines Skill** — load the `guidelines-meta` skill before proceeding. > **Profile section owned:** §Guardrails → do-not-touch and pinned-dependency rationale (Guidelines §5). Every deliberate *non*-upgrade gets recorded with its reason, so nobody re-litigates it in six months. **Why this skill exists:** maintenance is the work that is never urgent until it is catastrophic. It falls between building and shipping, so it belongs to nobody and happens never — until a security advisory lands on a dependency four majors behind, and the upgrade that should have taken an afternoon takes a week. This skill makes it a routine with a defined scope instead of an emergency. --- ## Operational Constraints (Strict) 1. **Never bundle an upgrade with a refactor.** The single most valuable rule here. If an upgrade needs code changes, the upgrade is one change and the adaptation is a *separate* one where possible. Mixed diffs make `git bisect` useless and turn a five-minute revert into an archaeology session. 2. **One batch, one class, one revert.** Each batch must be revertible on its own (§Batching). A batch you cannot undo in one step is too big. 3. **Green before, green after — with the same gates.** Establish a passing baseline *first*. Upgrading on top of an already-failing suite means you cannot attribute the failure, and you will blame the upgrade. 4. **Never upgrade to silence a warning you haven't read