wiki-query

name: wiki-query description: Answer operator questions by reading the LLM-maintained wiki at $WALTER_OS_HOME/wiki/ before falling back to web search or model knowledge. Read $WALTER_OS_HOME/wiki/index.md first, grep for relevant terms, load matching pages, answer with `[[wikilink]]` citations. If the wiki doesn't have the answer, derive it via web/MCP search and propose a new wiki page (concepts/ or decisions/) so future queries don't re-derive. Triggered any time the operator asks a substantive question — the agent should self-decide to apply this skill, not wait for an explicit "/query".

Wiki query

When this skill applies

Whenever the operator asks a question that's:

About a person, company, tool, concept, or decision relevant to the operator's contexts (work / projects-personal / personal).
Likely answered already (operator is asking for the second time, or the area's been discussed before).
Substantive — not a passing chat-style question (don't query the wiki for "what's 2+2?").

The agent should proactively check the wiki before answering, then optionally file the answer back if it's a novel synthesis.

Workflow

1. Identify wiki-relevant terms in the operator's question

Extract proper nouns, tool names, technical concepts, project names. Skip stop-words and generic verbs.

2. Read `wiki/index.md` first

index.md is the catalog. Grep it for the relevant terms.

grep -iE "<term1>|<term2>|<term3>" wiki/index.md

If hits, load those pages.

3. If `index.md` has nothing — broaden

ls wiki/people/ wiki/companies/ wiki/tools/ wiki/concepts/ \
    wiki/decisions/ wiki/topics/

then grep -ril "<term>" wiki/{people,companies,tools,concepts,decisions,topics}.

4. Load candidates + answer

Read the matching pages. Cite with [[wikilinks]] in the answer so the operator sees what the agent based its claim on.

If the wiki gives a confident answer, return it. Mention which pages were consulted.

5. Wiki-miss flow

If the wiki doesn't have the answer:

Note this to the operator: "Wiki doesn't have this — going to derive."
Use web/MCP search / model knowledge / current-codebase reading.
Answer.
Propose a new wiki page if the answer is reusable. Pattern:

"I derived this from . Worth filing as concepts/<slug>.md? It would link from topics/<topic> and capture: <2-line summary of the keep-able knowledge>."
If operator says yes → invoke the wiki-ingest skill flow to file the answer.

6. Don't refile what's already filed

If the wiki page already exists and you just confirmed the same fact, don't duplicate. If the page is stale (last_verified: > 90d for time-sensitive content), update it AND log: "refreshed [[page]] — verified as of ".

Citation format

In the answer:

The operator's CCR deployment uses @musistudio/claude-code-router@2.0.0, bound to claude-code-router:3456 inside litellm_net. Per [[tools/claude-code-router]] (last verified 2026-05-05).

Don't paraphrase wiki content as if you knew it. Make the source explicit so the operator can audit.

Failure modes to avoid

Hallucinating wiki content: never invent [[wikilink]] paths. If you cite [[tools/foo]], that file must exist. If it doesn't, you propose it via the wiki-ingest flow.
Skipping the wiki: even when the answer "feels obvious", a quick grep is cheap. The wiki is the operator's accumulated context — use it.
Over-filing: not every conversation produces a wiki-worthy fact. Only file when the answer is reusable (will help a future session) AND specific (not generic LLM knowledge).

Performance

wiki/index.md is the cache. Greps against it are O(file size). Loading 3-5 candidate pages is cheap. Don't grep the whole tree on every question — the index exists for a reason.

If the wiki grows past 200 pages, the spec migrates to qmd (BM25+vector). Until then, plain grep is fine.

skills/wiki-ingest/SKILL.md — write direction
skills/wiki-lint/SKILL.md — health checks (catches answers that weren't filed back)
wiki/SCHEMA.md — the contract