| name | doc-query |
| description | Answer a question using the Memex wiki. Read the index, find relevant pages, read them, synthesise a cited answer, optionally file the answer back as a new wiki page for compounding value. Invoked when the user asks a question referring to "the wiki", "my notes", "the docs", or asks a question whose answer likely lives in the wiki. |
| triggers | ["what does the wiki say about","based on my notes","check the docs for","search the wiki","synthesise from the wiki"] |
| allowed-tools | Read, Write, Grep, Glob, Bash |
Query the Memex wiki
1. Load index
- Read
memex.config.json#/index.path (default index.md) under the ops root
- Scan for sections relevant to the question
2. Search
- If
search.engine is qmd and qmd is on PATH, shell out: qmd search --path <ops-root> --top 8 -- "<question>"
- Otherwise, use
Grep across the ops root with keywords from the question (drop stopwords)
- Collect up to 8 candidate pages
Note. The docsite's GET /search?q=... endpoint uses the same precedence (qmd → grep). The skill stays grep-based intentionally — it must work without the [docsite] extras installed and without a running server.
3. Read
- Read the top candidate pages in full (not chunks)
- Note the frontmatter
updated: date — if stale, flag it
4. Synthesise
Draft an answer with:
- A direct response in 1–3 paragraphs
- Citations as inline links —
[page-title](<root>/<path>.md) — for every claim
- A short
## Caveats section noting where the wiki is thin, stale, or contradictory on this topic
- A
## Next steps section suggesting what to investigate or ingest to sharpen the answer
5. Offer to file the answer back
Ask the user: "Would you like me to file this as <root>/wiki/analyses/<slug>.md so it compounds?"
If yes:
- Write the analysis page with full frontmatter (
type: analysis or equivalent)
- Cross-link to every page cited
- Update
index.md under the appropriate analyses / syntheses section
- Append a
log.md entry with event: query | <question snippet>
6. Handle no-match
If nothing relevant is found:
- Say so directly — don't manufacture an answer from training data
- Suggest which
raw/ or external source might fill the gap
- Offer to file the question as
.open-questions/<slug>.md. When filing, populate ## Related files with markdown links to whatever pages you did find in the search — even partial matches are useful context for whoever resolves the question. Bare paths do not count; the link graph only picks up real markdown links.
