| name | wiki-clean |
| description | Read-only wiki hygiene sweep. Runs the mechanical scanner (scripts/wiki-clean.mjs), dispatches subagents to triage the candidates (contradictions, stale claims, near-duplicates, thin stubs, dead links, orphans), and returns a prioritized CANDIDATE list for the operator to decide on. NEVER edits or deletes — proposal only. Trigger: /wiki-clean. |
Read-only. This skill proposes; the operator disposes. Do NOT edit or delete any wiki content while running it.
Division of labour: the mechanical scan and the page reading run on a cheap model (breadth); the main loop only collates and judges. Spend expensive tokens on the decision, not on reading every page.
-
Mechanical scan (deterministic, no model). Run node scripts/wiki-clean.mjs --json (add --today=YYYY-MM-DD from date for a fixed day). It emits candidates the LLM should not have to guess at — orphans, brokenLinks, stubs, staleReview, indexMissing, dupeTitles — with paths and line numbers. The script never writes. If it returns near-empty, say so; a clean wiki is a valid result, not a reason to invent work.
-
Semantic triage (read-only). Dispatch subagents — preferably a structurally read-only agent type (no Edit/Write) — pinned to a cheap model. Shard by domain (wiki/index/<domain>) when the candidate list is large; a single subagent when small. Each reads the flagged pages + their neighbours and judges what the script cannot:
- contradictions — a claim conflicting with another page,
- staleness — a claim superseded by a newer source/page,
- near-duplicates — two pages that should be one (name the merge target + which survives),
- stubs — too thin: grow, merge, or drop?,
- orphans / indexMissing — wire into a router/sub-index, or let die,
- dead links — fix the target, create the missing page, or remove the link,
- missing cross-refs — related pages that should link each other.
Each finding returns:
{ class, pages[], issue, evidence (quote + path:line), suggestedAction, confidence: high|med|low }. Subagents must mutate nothing.
-
Collate + prioritize (main loop). Merge subagent findings with the mechanical signals into ONE candidate report, grouped by action class (MERGE · DELETE? · UPDATE-STALE · FIX-LINK · ADD-XREF · GROW/DROP-STUB · WIRE-INDEX). De-duplicate overlap; sort by confidence × impact.
-
Present, do not act. Show the operator the candidate list as a decision block, then STOP. Only if the operator approves specific items do you act — and acting goes through the wiki-write skill (one atomic commit per logical cleanup) plus a lint entry appended to wiki/log.md. Approval is the operator's call, never the skill's.
Hard rules:
- Zero mutations to
wiki/**, outputs/** during the sweep. The only allowed side effect is an OPTIONAL report saved to outputs/wiki-clean/<date>.md, and only if the operator asks to persist it.
- Never delete on your own initiative — deletion is always operator-approved, even for a clear orphan or stub.
- Scale to the ask: "quick pass" → one subagent, high-confidence findings only; "deep clean" → per-domain subagents + a completeness critic asking "what did the scanner miss?".
