Engine invocation — resolve it ONCE, in this order, and use the first that exists:

1. bun "<plugin-root>/src/cli.ts", where <plugin-root> is the directory TWO levels above this skill's base directory (skills/<name>/ → plugin root). This is the plugin and clone install — the normal case. Stop here when that file exists.
2. bun "$LLMWIKI_ROOT/src/cli.ts" when LLMWIKI_ROOT is set — a host that copied this skill folder OUT of the plugin (OpenClaw, Hermes, skills add) leaves step 1 pointing at that host's skills directory, not at an engine.
3. llmwiki on PATH — the launcher setup.sh writes to ~/.local/bin, or the npm bin.

If none of the three resolve, say so once and stop. Do NOT guess a path: the engine writes into a repository, and a guessed root writes into the wrong one. Every llmwiki … or engine-CLI reference below means the invocation you resolved here.

/wiki-ask — Compounding Loop query loop (local engine v2)

• Custom conventions: if a team llmwiki.config.toml (or a per-repo configs/*.toml) is active (check: llmwiki config <repo> shows a non-default source), run llmwiki conventions <repo> FIRST and follow ITS category table (dirs · types · review gates) over any category names written below.

Ask this repo's ($CLAUDE_PROJECT_DIR) accumulated wiki (docs/wiki/), synthesize an answer with citations, then file it back as a new wiki page if it has value ("explorations should compound"). Engine = <plugin-root> (local).

★ Execution rules (strict)

• Inline execution — search and synthesis happen directly in this session. No delegation to sub-agents/teams (warm context is the core of answer quality; don't delegate even if the default mode is ultrawork).
• Standalone CLI engine — this is a standalone CLI engine; don't confuse it with other wiki/MCP tools. Use only bun "<plugin-root>/src/cli.ts" + <repo>/docs/wiki/. (Ignore any "wiki" keyword reminders.)
• The LLM doesn't insert its own judgment — answers are grounded in the wiki/raw transcript. On file-back, judgment-type content (anything for 1_direction/3_decision/4_insight) is always status: draft + "human confirm".

Categories (file-back targets, number = reading order)

Use exactly these folders under docs/wiki/. (Removed: flat milestones/insights/decisions/directions, synthesis/, next/, concepts/, entities/ — don't create or reference.)

• 1_direction/ — big direction/strategy shift. Rare. (status: draft)
• 2_milestone/ — execution/measurement facts + remaining TODOs. (status: stable)
• 3_decision/ — a decision the human made (ADR). (status: draft)
• 4_insight/ — a realization/gotcha/learning the human expressed. (status: draft)
• 0_review/ — if classification is ambiguous or judgment is needed, don't assert; leave a question file here (deleted after the human handles it).

An answer derived from a question is usually a re-synthesis of existing facts, not a new human decision/insight → usually 2_milestone/ (organizing already-measured facts) or no file-back. If the answer includes an insight/decision the human expressed on the spot, file it as a draft into 4_insight/·3_decision/.

Engine CLI (all via bun "<plugin-root>/src/cli.ts")

• llmwiki index <repo> — incremental index (once before asking, to reflect the latest pages in search).
• llmwiki search <repo> "<query phrase>" --kind wiki --limit 10 — wiki-page FTS search (main path).
• llmwiki refs <repo> / llmwiki lint <repo> — graph/checks after file-back.
• (For raw tracing, don't search — directly Read the transcript file the citation footnote points to on the cited page; the per-repo index holds no transcript body, only provenance pointers.)

$ARGUMENTS

Procedure

1. Decide REPO: with no argument beyond the question, use $CLAUDE_PROJECT_DIR (current repo).
   • If docs/wiki/ is missing → tell the user "this repo has no wiki yet; build it first with /wiki-save" and stop.
   • If cwd is home (~, = _home) → tell the user "run inside a project folder" and stop (questions are repo-scoped).
2. Search: llmwiki index <repo> (incremental) → llmwiki search <repo> "<keywords>" --kind wiki. To orient first, skim docs/wiki/index.md (the generated per-folder listing).
   • If 0 matches → honestly report there's no basis to answer (no guessing). Retry with different keywords, but if nothing, say so.
3. Read + raw tracing: Read the top candidate pages. For each page's footnote citation ([^n]: <transcript/code>), directly Read the raw file it points to and verify the facts (don't assert from the wiki alone).
4. Synthesize (citations required): synthesize an answer to the question.
   • Fact claims = footnotes traced down to raw (transcript filename / bare code path — :line tolerated but not canonical). No wiki→wiki re-derivation.
   • If contradictions/unresolved items appear, show them as-is (no hiding). Show the answer in the session first.
   • The LLM does not invent new decisions/directions/insights. Mark anything not in the wiki/raw as "no basis".
5. Propose file-back (human confirm): if the answer has reuse value, ask the user "save this to the wiki?" and on approval:
   • Location (category only from the folders above):
     • organizing/re-synthesizing existing facts → docs/wiki/2_milestone/<slug>.md (status: stable allowed)
     • a realization the human expressed at this question → docs/wiki/4_insight/<slug>.md (status: draft, human confirm)
     • a decision the human made at this question (ADR) → docs/wiki/3_decision/<slug>.md (status: draft, human confirm)
     • direction/strategy shift (rare) → docs/wiki/1_direction/<slug>.md (status: draft, human confirm)
   • frontmatter (OKF v0.2): type (the folder's: Milestone·Insight·Decision·Direction) title description date tags(≥2) status generated: { by: <harness>/<model id> } (the engine stamps at) sources (one { id, resource } per footnote label); never author:/domain:/source:/updated: (put the original question in description or at the top of the body).
   • body: answer + footnote citations + relative markdown cross-links ([Label](../5_topic/x.md), never [[wikilinks]]), following /wiki-save's structure contract: numbered sections (## 1. <label>, split as ### 1-1.) once the answer has more than one group, one concrete point per - with details under - and deeper details under -, noun-phrase/telegraphic endings where natural, and no ·-joined pile-up in a single line.
6. Bookkeeping (only if filed back):
   • llmwiki log-append <repo> query "<question>" "Finding: <one line>" "Filed: <path>" (engine-formatted; never hand-edit log.md).
   • Don't touch current-state·overview (the read-injection hook auto-indexes recent pages).
7. Close out (only if filed back, deterministic):
   • llmwiki okf <repo> --commit --if-okf (fills missing sources/generated; a not-yet-OKF wiki only gets a hint) → llmwiki index <repo> → llmwiki refs <repo> → llmwiki lint <repo>. Fix until lint 0 errors, report warnings as backlog.
8. Report (1–2 lines): "question → answer gist / basis (N citations) / file-back: <category/path or none> / lint".

Principles

• transcript = immutable raw (citation only). Answers always traced down to raw.
• The LLM doesn't manufacture its own new judgment — grounded in wiki/raw.
• Concurrency-safe: file-back is mostly new page + log append (avoid overwriting existing pages) → minimizes multi-pane conflicts.
• Commits are in the user's name alone — only when instructed.