Maintain Skills

Maintain: Skills + Plugins (shared agent extensions)

The shared agent-extension libraries — the agentic counterpart to the CI building blocks in the github-actions card. Both are checked-in submodules, so a per-run worktree can give you a working copy to build and validate in — but only once the submodule is populated. Populate it with .claude/scripts/submodule-init.sh <path>, never a bare git submodule update --init, which silently collapses every parallel session into one physical tree; if that script fails, stop rather than work against an unpopulated gitlink. Each repo's canonical task menu lives in its own AGENTS.md ## Maintenance, and this card is a thin pointer to it:

• devantler-tech/agent-skills — generic, cross-tool agent skills at submodule path libraries/agent-skills. Task menu: https://github.com/devantler-tech/agent-skills/blob/main/AGENTS.md.
• devantler-tech/agent-plugins — the tool-neutral marketplace at submodule path libraries/agent-plugins. Task menu: https://github.com/devantler-tech/agent-plugins/blob/main/AGENTS.md. This monorepo consumes agentic-engineering@devantler-plugins via .claude/settings.json (see monorepo#2363). Keep plugins additive & backward-compatible; generic role improvements land here so every consumer inherits them.

These are shared libraries — design for reuse, not for one product. Keep them generic and industry-standard (the portability principle: a Claude→Copilot/ChatGPT switch should stay painless), additive & backward-compatible (every product is a potential consumer), and well-tested. Don't put product-specific logic here.

Advance them via the holistic review (contract Holistic review; product-engineering §7): when a generic skill/convention has emerged across 2+ products, extract it here so every product inherits it, then migrate consumers. Triage/label issues, drive actionable trusted-author PRs to merge, intervene when dependency automation cannot finish its PR, and keep dependency automation & docs current. Programmed chore(deps): update agent skills PRs are the no-review exception defined in the root contract only when .claude/scripts/programmed-bot-review-exemption.sh exits 0: let required CI and auto-merge decide accepted exemptions, and never spend a review lane on exit-0 exemptions. Installed skill roots hold copies from several upstreams, so exit 0 also requires every changed skill to be listed in the reviewed .claude/skill-ownership-allowlist.tsv; an unlisted or third-party skill returns 3. Never authorize that from the skill's own metadata.github-repo — its upstream writes that line, so it can grant itself the carve-out; passing it to the classifier only corroborates. A genuine agent-plugins marketplace update is trusted but review-bearing when the classifier exits 3; request semantic review because bundled skill prose is executable agent instruction. Route classifier exit 1 and non-matching lookalikes through the external/static-only path; exit 2 is a fail-closed error.

Shared cross-repo rules are in the monorepo AGENTS.md.

Roadmap & enhancement

The two libraries keep separate issue queues — do not conflate them:

• Skills: GitHub Issues on devantler-tech/agent-skills (roadmap label)
• Plugins: GitHub Issues on devantler-tech/agent-plugins (roadmap label)

Strategic frame for both: what generic capability would most help the suite next? — surface candidates from the holistic review.