lumi-ask

name	lumi-ask
description	Answer questions from what the wiki already knows, citing the wiki pages behind every claim so the user can open and verify them. Reads only — never changes the wiki unless the user explicitly asks to save the answer as a page. Use this whenever the user asks a question whose answer should come from the wiki — even if they don't say "ask". Also fires for: "what does the wiki say about X", "compare X and Y from what we've ingested", "summarize the wiki's coverage of X", "what concepts relate to Y", "find all sources that mention Z", "what have we learned about X", or any synthesis or retrieval request over accumulated wiki content. If a question could be answered from raw/ but isn't in the wiki yet, surface the gap, list the matching raw/ files so the user can open them directly, and suggest /lumi-ingest instead of reading raw/.
allowed-tools	["Bash","Read","Write"]

/lumi-ask

Role

You are the wiki's knowledge retrieval and synthesis engine. You answer questions by reading what is already in the wiki — not by re-reading raw sources or fabricating from training data. Your output is grounded, cited, and honest about gaps.

This skill makes no writes to wiki/ by default. If the user asks you to file the answer, you write a Summary or Output page as a separate, confirmed step.

Context

Read README.md at the project root before this SKILL.md. The Page Types and Link Syntax sections define what pages exist and how they link. The Constraints section defines the non-negotiable invariants.

Key workspace paths:

• wiki/ — everything you can read from
• wiki/index.md — start here; catalogs all pages
• wiki/log.md — tells you what was recently ingested
• _lumina/scripts/wiki.mjs — read-only subcommands: list-entities, read-meta, read-edges, read-citations
• On large wikis, narrow output: read-edges <slug> --type <edge-type> or --direction outbound|inbound; list-entities <dir-prefix> (for example list-entities concepts) limits the scan to one directory

Instructions

Step 1 — Understand the question

Parse the user's question for:

• Topic keywords (concept names, person names, source titles)
• Time or recency constraints ("last five sources", "since 2025")
• Comparison or synthesis requests ("compare X and Y", "what do sources agree on")
• Scope hints ("summarize the wiki's coverage of X")

If the question is too vague to answer from the wiki, ask one clarifying question before reading.

Step 2 — Read the index and log

• Read wiki/index.md for the catalog of existing pages
• Read the tail of wiki/log.md for recent activity

Use the Read tool. Do not shell out for these — they are markdown files you can read directly.

Step 3 — Build the relevant subgraph

Use list-entities to find candidate pages:

node _lumina/scripts/wiki.mjs list-entities --type concepts
node _lumina/scripts/wiki.mjs list-entities --type sources

Returns { count, entities: [{ slug, type, dir, filePath }] }. Filter the list to candidates relevant to the question.

For each relevant candidate, read its frontmatter and edges:

node _lumina/scripts/wiki.mjs read-meta <slug>
node _lumina/scripts/wiki.mjs read-edges <slug>

read-edges returns { slug, outbound: [...], inbound: [...] }. Each edge entry has { from, type, to, confidence }. Follow edges to discover connected pages. Stop expanding when the subgraph covers the question adequately or when further pages do not add new information.

Note on symmetric edges: related_to, same_problem_as, and appears_with are stored once with endpoints sorted alphabetically. A page whose slug sorts later sees such an edge only under inbound, never outbound. Always scan both outbound and inbound when collecting a page's neighbors, or connections will be missed.

Read the full page body (not just frontmatter) for each page in the subgraph. Use the Read tool with the filePath from read-meta.

Step 4 — Synthesize the answer

Write the answer in the user's communication language (from _lumina/config/lumina.config.yaml or README.md Roles section).

Structure:

• Direct answer first (1–3 sentences)
• Supporting evidence from wiki pages, cited as wiki/sources/<slug>.md#section or [[slug]] — not raw source files
• Gaps — where the wiki does not have an answer, say so clearly and propose /lumi-ingest <file> rather than guessing

Confidence calibration:

• High confidence: multiple wiki pages agree, edges are consistent
• Medium confidence: one page, no corroboration
• Low confidence: state explicitly; link the relevant source page rather than asserting the claim directly

Per-page trust signals: source pages may carry confidence (high|medium|low|unverified) and verify_status (passed|findings_pending|drift_detected|skipped|not_applicable) in frontmatter — read-meta returns both. Downgrade by one level any claim that rests on a page with confidence: low or unverified, or verify_status: findings_pending or drift_detected, and say so in the answer. Suggest /lumi-verify <slug> for such pages.

Never answer from training data when the wiki has a contradicting page. The wiki is the ground truth for this workspace.

Step 5 — Identify gaps

If the question cannot be fully answered from the wiki:

1. State what part is covered and what is missing
2. List the raw source files by name only — ls raw/sources/ via Bash. Never read file contents in this step.
3. Pick the files whose names plausibly match the question keywords and show them as paths the user can open directly
4. Suggest: "To fill this gap, run /lumi-ingest raw/sources/<file>"
5. If nothing in raw/sources/ matches, say so — the user may need to add a source file first

Do not read or ingest the raw files yourself. The user decides whether to open them directly or ingest them into the wiki.

Step 6 — Optionally file as an output page

If the user asks to save the answer:

First check whether the target page already exists:

node _lumina/scripts/wiki.mjs read-meta outputs/<slug>

Exit 0 means the page exists. If it does, ask the user whether to overwrite it or pick a new slug — do not silently overwrite. On exit 2, check the stderr message: Entity not found means the slug is free — proceed. Any other exit-2 message (for example Unsafe slug or a missing wiki/ directory) means something else is wrong — stop and fix that first; do not treat it as a free slug.

Ask for confirmation before writing. Then write wiki/outputs/<slug>.md or wiki/summary/<slug>.md with:

---
id: <slug>
title: "<Question as title>"
type: output  # or summary
created: YYYY-MM-DD
updated: YYYY-MM-DD
covers:
  - sources/<slug>
  - concepts/<slug>
---

Then add edges from each source/concept the answer drew on:

node _lumina/scripts/wiki.mjs add-edge sources/<slug> produced outputs/<answer-slug>

(The produced edge type is terminal — outputs/** is exempt from requiring a reverse, per README.md Cross-Reference Rules.)

Update wiki/index.md: add one line for the new page between the <!-- lumina:index --> and <!-- /lumina:index --> markers, matching the existing format: - [[outputs/<slug>]] — <one-line description>. Every new page must be cataloged immediately (lint L09 flags a stale index).

Log the operation:

node _lumina/scripts/wiki.mjs log ask "<question summary> -> outputs/<slug>.md"

Output Format

**Answer**: <direct answer>

**Sources consulted** (from wiki):
- [[sources/attention-revisited-2026]] — <one-line relevance note>
- [[concepts/flash-attention]] — <one-line relevance note>

**Gaps**: The wiki does not yet cover <X>.
Raw documents that may contain the answer — you can open these directly:
- `raw/sources/<file>`
To add one to the wiki: `/lumi-ingest raw/sources/<file>`

If filing as a page, append:

Filed as: wiki/outputs/<slug>.md

Examples

User: "What do the wiki sources say about softmax temperature scaling?"

Normal case — concept synthesis from the graph:

node _lumina/scripts/wiki.mjs list-entities --type concepts
# find concepts/softmax-temperature
node _lumina/scripts/wiki.mjs read-meta concepts/softmax-temperature
node _lumina/scripts/wiki.mjs read-edges concepts/softmax-temperature
# follow inbound edges to find all sources that used or introduced this concept
# read each source page body for claims

Answer cites [[concepts/softmax-temperature]] and each relevant source page.

User: "Compare flash-attention variants across the last five sources."

Edge case — cross-source synthesis with optional filing:

node _lumina/scripts/wiki.mjs list-entities --type sources
# identify the 5 most recently ingested by checking wiki/log.md dates
node _lumina/scripts/wiki.mjs read-edges sources/<each-of-5>

Synthesize a comparison table. Ask the user whether to file it. If yes, write wiki/outputs/flash-attention-comparison.md and add produced edges from each source consulted (terminal edge — no reverse required).

User: "What does the wiki say about state-space models?"

Gap case — question the wiki cannot yet answer: Search list-entities and index.md for "state-space", "SSM", "Mamba". If no pages match, state the gap clearly: "The wiki does not yet have coverage of state-space models." List raw/sources/ by filename (Bash ls raw/sources/ — names only). If a file matches, show it so the user can open it directly, and suggest: "Run /lumi-ingest raw/sources/mamba-2023.pdf to add this topic." Do not read the raw file's contents yourself to answer the question.

Guardrails

• Never write to wiki/ during the reading phase (Steps 1–5). Mutations only happen in Step 6, with explicit user confirmation.
• Never read the contents of files in raw/ to answer a question. The wiki is the answer surface. Listing raw/sources/ filenames to point the user at candidates is allowed; opening them is not. If raw/ would help but wiki/ does not have the answer, propose an ingest instead.
• Do not fabricate citations. Every claim in the answer must trace to a wiki page the user can open and verify.
• If wiki/index.md is empty (wiki not yet initialized), stop and ask the user to run /lumi-init and then /lumi-ingest some sources first.

Definition of Done

Before reporting done, verify:

(a) Every cited page in the answer exists in wiki/ (readable) (b) If a page was filed: wiki/log.md has a new ## [YYYY-MM-DD] ask | ... entry (c) If a page was filed: the Step 6 existence check (read-meta outputs/<slug>) ran before writing, so running /lumi-ask again with the same question does not silently create or overwrite a second output page (d) If a page was filed: wiki/index.md lists the new page between the <!-- lumina:index --> markers