菜单
Keep an LLP corpus in sync with its code. Detects code/doc drift, validates @ref annotations interactively, checks provenance tags, and proposes reconciliations, ref repairs, and status changes. Modes — pre-pr, audit, reconcile, retire-proposal. Proposes; never applies.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| name | llp-maintain |
| description | Keep an LLP corpus in sync with its code. Detects code/doc drift, validates @ref annotations interactively, checks provenance tags, and proposes reconciliations, ref repairs, and status changes. Modes — pre-pr, audit, reconcile, retire-proposal. Proposes; never applies. |
| source | ccheever/llp@v0.3.0 |
The upkeep engine. Living documents rot into misleading guidance unless code and docs co-evolve; this skill finds the drift and proposes fixes. The propose-only line is what lets one skill cover four maintenance moments without becoming a mega-skill.
Invoke as /llp-maintain --intent <pre-pr | audit | reconcile | retire-proposal> (each has a sensible default scope).
Before opening a PR (pre-pr, the common case); periodically or on suspicion (audit); after a pull or against a branch (reconcile); "this doc looks done/stale" (retire-proposal).
[observed]/[confirmed]/[inferred].ref-check's job; this skill is the interactive, judgment-bearing complement.For @refs in scope: broken target/anchor (error — list the anchors that do exist); tombstoned/superseded target (warning, note the replacement); stale gloss (hint); orphaned annotation whose attached construct is gone (warning). For provenance-tagged docs: [inferred] in Accepted/Active (error); [observed] without a stated location, or [confirmed] without name+date (flag for the author). For installed skills: a source: pin that lags upstream releases (note, when network allows). Beyond the mechanical: code whose behavior no longer matches the section it references; new files implementing a documented LLP with no @ref; LLP sections the diff contradicts.
| Intent | Default scope | Artifact |
|---|---|---|
pre-pr | working tree | co-evolution checklist ("update doc or code" / "repair ref" / "consider annotating") + proposed status changes, not applied |
audit | corpus | findings list: drift, broken refs, stale statuses |
reconcile | diff range / branch | per-doc reconciliation proposals (or "the code drifted, not the doc") |
retire-proposal | one named doc | a drafted Superseded/Tombstoned proposal with rationale |
llp-review.llp-orient land here for repair proposals.source: pin noted by audit → /llp-adopt update applies the update after an approved diff (this skill never applies).ref-check.Author-invoked dual-family refine loop for one or more Draft/Review LLPs — each round, two fresh mutually-blind reviews from different model families, orchestrator revision, verdict-lined artifacts — until both families mark the same revision READY, a delta round finds no MATERIAL concern in pre-existing text across families, or a bounded stop rule escalates to the author. Use when the user invokes /llp-super-refine or asks to refine an LLP until independent models agree it is ready.
Create a new LLP document with the next available number, the NNNN-slug.type.md filename convention, and a scaffolded metadata header — or extend an existing LLP when one already covers the topic.
Before editing a subsystem, file, or starting a task, assemble the governing LLP context — read the root LLP, follow the @ref annotations on the code in scope, and find the LLPs covering the relevant systems. Produces a compact context pack of the constraints a change must respect. Read-only.
Run the LLP 0005 review loop for an LLP document — orchestrate reviews from other model families, write one provenance-tracked artifact per review actually received under llp/reviews/, and never fabricate a review or accept on the author's behalf. Review intensity scales with stakes; the author decides when it's enough.
Take one or more tasks (Linear tickets, ticket files, accepted LLPs, free text) end to end — plan lanes, implement in run-namespaced worktrees, integrate on a run branch, verify and review the exact integrated commit, deliver per repo policy (direct push by default), and clean up without destroying recoverable work. Use when the user invokes /llp-ship or hands over tasks to be implemented and shipped.
Bring LLP to a repository and keep its installed skills current — the single entry point for "set up LLP here" and "update the LLP skills". Scaffold mode (fresh structure), retrofit mode (survey the codebase, draft a provenance-tagged root LLP, propose an adoption plan), install step (copy skills from a tagged release with a receipt), and update mode (verified, diff-approved skill updates).