Ask — unified knowledge query

One query for the whole registry. lib/registry.py picks what to read; you answer from it, grounding every claim in raw/.

Core rule — wiki is the index, raw is the truth

• raw/ ($CLAUDE_DOC_HOME) is the source of truth. The doc-* commands write it directly, so it is always current. Every claim must cite a raw/ file.
• wiki/ (a sibling of raw/, when present) is a compiled navigation index of [[backlinks]], refreshed by /sdd:doc-ingest. It can lag raw/. Navigate it; never answer from it alone.

---

Step 1 — Retrieve

Run the engine with the arguments verbatim — do not add type: or service: yourself; without filters it searches everything and ranks by relevance:

python3 ${CLAUDE_PLUGIN_ROOT}/lib/registry.py ask "$ARGUMENTS"

It prints, in order:

Block	What it is
WIKI: (+ ⚠ The wiki is behind)	Only when a wiki/ layer exists: its path, last ingest, and raw docs changed since
INDEX	Every in-scope doc: path, title, date — always complete
LOADED	The best-matching docs (large ones trimmed to their matching ## sections) up to a budget; or every doc when the scope fits; or only unresolved Open Questions for "which specs have open questions?"
ALSO MATCHED	Other matches by path, not loaded
HINTS	How to read each registry type that loaded
tip:	When matches concentrate in one registry, the type: that would narrow it

Step 2 — Fill gaps

• An INDEX title or ALSO MATCHED path looks relevant but was not loaded → cat it (<Registry>/<path>). Do this especially when LOADED says none matched: term search misses synonyms and Spanish questions over English docs; titles don't.
• A section you need was [omitted …] from a trimmed doc → read the file.

Step 3 — Navigate the wiki (only if a WIKI: line was printed)

For cross-cutting questions (spanning catalog + ADR + spec + investigation):

1. Read wiki/wiki.md and the folder-notes (entities.md, concepts.md, sources.md) to find the pages for the question's entities.
2. Read those pages and follow their [[wikilinks]] one hop.
3. Their ## Source links point at raw/ files — read any that were not loaded.

A missing wiki page means "not compiled yet", not "no data" — raw/ still decides.

Step 4 — Answer and cite

Answer $ARGUMENTS, citing a raw/ path for every claim (and the wiki pages you navigated through, when they added cross-links). If the engine printed ⚠ The wiki is behind, relay it in one line and suggest /sdd:doc-ingest.

Question: Who consumes LeadCreated, and why did we design it that way?

Consumed by (service-catalog/marketing-service.md, service-catalog/consent-service.md):
  marketing-service ← SQS marketing-lead-events-queue
  consent-service   ← SQS consent-lead-events-queue
Design rationale (adr-registry/leads-service/PROJ-12-ADR-002-sns-fanout.md):
  Fan-out via SNS chosen over direct calls to decouple consumers …

If the registry does not cover the question, say what you read (the INDEX size and what you opened) and offer to widen — never invent.

---

Related Skills

• /sdd:doc-ingest — builds/refreshes the wiki/ layer this skill navigates.
• /sdd:doc-catalog · /sdd:doc-adr · /sdd:doc-spec · /sdd:doc-investigation — write the raw/ registries answers are grounded in.