| name | wiki-desk |
| description | Path-reference librarian over the project's docs/ work-history (devlog, testlog, plan, simlog) and the sub-doc mirror. Builds a deterministic relationship graph (cites/realizes/evidences edges) and surfaces authority-ranked, edge-traversable context WITHOUT copying source bodies. Use when starting an Ouroboros interview, writing a plan, devising a model serving strategy, bumping a vLLM version, or adapting topology — whenever grounded prior context or faster Seed convergence helps; also when shelving new docs into the library or asking "what does this project already know about X". Also triggers on Korean instructions like "위키 조회", "이전에 뭐 했는지/뭐였지", "근거 문서 찾아줘", "관련 plan/devlog 찾아줘", "문서 등록/갱신/입고". Run init once to bootstrap __llm-wiki, then warm-start incrementally on each invocation. |
wiki-desk — path-reference librarian (self-improvement loop)
wiki-desk is the librarian for a project-local __llm-wiki/: a metadata-first library of
path references + sha256 + roles + authority ranks + deterministic relationship edges +
concise summaries. It is not a raw source archive — original docs stay in place; the
library only points at them. This closes the self-improvement loop: relationship building →
refined context surfacing → faster Ouroboros interview/Seed convergence for version/topology/model
adaptation.
Building-block, main-only (one central librarian; sub-node knowledge is recovered via the
doc-mirror, not a second library). The live __llm-wiki/ is untracked; this skill is tracked.
When to use (invocation timing)
Invoke the librarian — read its metadata first, then act — when:
- starting an Ouroboros interview / writing a
docs/plan/ → warm-start the relevant thread,
surface prior devlog/testlog evidence so the Seed converges faster;
- devising a model serving strategy / bumping a vLLM version / changing topology → query the
library for what already worked (and what failed) on that model/version;
- a new doc is written (plan/devlog/testlog/simlog/benchmark) → shelve it (warm-start incremental);
- asked "what do we know about X / where is the evidence for Y" → topical query with edges.
Operating model
__llm-wiki/ lives at the project root, untracked. It indexes docs/{devlog,testlog,plan,simlog}
sync_staging/sub_docs + seed (graceful skip if absent). The project ROOT constitution
(/CLAUDE.md, /.claude/) is deliberately NOT indexed (exclusion is by-root — those paths are
simply not document_roots; vendored subtrees under seed/ are excluded too) — the library is a
neutral evidence base that can surface patterns the constitution has not yet codified
(anti-confirmation-bias); the constitution is the human-authored output of reviewing the logs, not an input.
- Authority = execution-truth > plan-intent:
devlog:100 > testlog:85 > sub-doc:70 > plan:55 > simlog:40 > seed:25. A devlog of what actually ran outranks the plan that was only intended.
- Cycle:
init (once, adaptive bootstrap) → loop (new plan/serving/version) → warm-start
(incremental shelving on invocation). Detection is inline (no cron). The script's contract is the
fixed document_roots: a doc under those roots auto-shelves (deterministic); a doc OUTSIDE them is
not emitted by the script — proposing a contract amendment for it is an agent/HITL judgment, not
a script behavior.
Mandatory orientation
Before answering a project question or updating the library, read in this order:
__llm-wiki/index.md (entrypoint) 2. __llm-wiki/SCHEMA.md 3. __llm-wiki/sources/source-registry.md__llm-wiki/relationships/edge-graph.md 5. __llm-wiki/concepts/document-universe-hierarchy-map.md.
Scripts read the machine source of truth __llm-wiki/sources/registry.json.
Commands (deterministic scripts — scripts/)
ROOT=.; WIKI=__llm-wiki; ANS=.claude/skills/wiki-desk/fixtures/project_init_answers.yaml
python3 .claude/skills/wiki-desk/scripts/init_wiki_desk.py --project-root $ROOT --wiki-root $WIKI --answers $ANS
python3 .claude/skills/wiki-desk/scripts/init_wiki_desk.py --project-root $ROOT --wiki-root $WIKI --answers $ANS --incremental
python3 .claude/skills/wiki-desk/scripts/smoke_query.py --wiki-root $WIKI --query "<question or doc stem>"
python3 .claude/skills/wiki-desk/scripts/lint_wiki.py --wiki-root $WIKI --project-root $ROOT [--fix]
python3 .claude/skills/wiki-desk/scripts/scan_forbidden_strings.py .claude/skills/wiki-desk
python3 .claude/skills/wiki-desk/scripts/scan_raw_copy.py $WIKI
If the target environment has no document system yet, init is interview-led: settle
document_roots / authority_order / retrieval with the user first (see reference/INIT_GUIDE.md).
Query loop
- Resolve intent: completion ("did it work") vs intent ("what approach"). 2. Search registry
metadata + edges first, never raw bodies. 3. Rank by intent → authority → relevance. 4. Traverse
realizes/evidences/same-thread edges to surface supporting plan/testlog/simlog. 5. Open a
source body only when the summary is insufficient. Negative honesty: if nothing clears the
authority/score gate, answer "no evidence" — never fabricate a path.
Relationship edges (deterministic, grep-grounded — no inference)
cites (explicit cross-ref) · realizes (devlog/testlog → plan) · evidences (simlog/testlog →
the higher tier it supports; the docs chain simlog→testlog→devlog) · part-of (file under a
directory-node, e.g. a simlog run) · same-thread (identical topic slug). Citations are resolved
by stem with prefix-glob, brace-glob {1,2,3}, and _* wildcard expansion. (Semantic
contradicts/supersedes edges are intentionally not v1 — they need judgment; parked.)
Verification (release gate — done = all pass on real data)
- init bootstraps the live library; 2. query smoke — direct lookup returns a node + its
realizes→plan edge; topical returns ≥2 authority-ranked sources with edge traversal; a no-evidence
query returns the negative verdict (exit 2); 3. scan_forbidden_strings + scan_raw_copy pass;
- lint reports 0 critical. Commit only the tracked skill (the live
__llm-wiki/ is untracked).
Provenance
Derived from the operator's own wiki-desk-operating-pattern v1.0.0-candidate and the open-source
llm-wiki (nvk) — adopting its _index-first navigation, lint, lightweight confidence, dual-link,
and concepts/ synthesis-node ideas while keeping strict path-reference (no raw-copy / web-research /
multi-runtime packaging). MIT. Constitution-adapted for easy_vllm_simulator.
Non-goals
No raw __llm-wiki/raw/ copies · no sub-node library instance · no constitution/.claude/ indexing ·
no cron/background ingestion · no web research · no semantic-judgment edges (v1) · no multi-runtime packaging.
Deeper references (read on demand)
reference/INIT_GUIDE.md — adaptive-init interview contract (slots, edge slot, authority design).reference/MANUAL.md — install, init, warm-start, query, lint, scans, field report.templates/__llm-wiki/SCHEMA.md — the metadata + edge schema (copied into the live library).
