Skill v1.0.1
currentAutomated scan100/100+11 new
version: "1.0.1" name: lumi-ingest description: > Turn a raw source file into structured wiki pages: source page, concept and person stubs, wiki.mjs-managed graph edges, citation edges, and a log entry. Pauses at four checkpoints (draft, lint, verify, finalize) so you can review before the work is committed to the wiki — accept, revise, or quit at any checkpoint without losing progress. Resumable across sessions. Use this whenever the user says "ingest", "add", "file", "process", "summarize into the wiki", "create a wiki page for", or drops a filename from raw/sources/. Also fires for: "I added a PDF to raw/sources/", "add this paper to the wiki", "parse this article", "what should I do with raw/sources/X?", or any request to bring a new source into the wiki graph. Also accepts Mode B input — paper title, arxiv ID, or URL, without a local file path. Examples: "ingest paper 2604.03501v2", "ingest arxiv:2604.03501", "ingest https://arxiv.org/abs/2604.03501". The skill fetches the PDF automatically in that case. Also handles whole books, theses, and other 50+ page documents — "ingest this book", "add this textbook to the wiki", "đưa cả cuốn sách này vào wiki" — via a multi-pass reading pipeline (references/long-source.md) that writes page-anchored reading notes instead of a one-pass summary. (A novel the user is still reading belongs to /lumi-reading-chapter-ingest instead.) This is the most-used skill — when in doubt about whether something is an ingest vs an edit, ask the user. allowed-tools:
- Bash
- Read
- Write
- Edit
- Agent
/lumi-ingest
You are the wiki's primary knowledge compiler.
Read README.md at the project root before this SKILL.md. Do this before doing anything else — bidirectional-link discipline drives the whole workflow.
The work is split across four stage files under references/. Each stage ends with a human-in-the-loop gate so the user can review before changes commit. Gate state is durable on the source entry's frontmatter (ingest_status) so the workflow survives session restarts; fine-grained phase resume lives in _lumina/_state/ingest-<file-basename>.json and is handled inside step-01.
| Stage | Step file | Reads | |
|---|---|---|---|
| 1. Draft | references/step-01-draft.md | Resolve input → detect type → write source page + stubs + edges + citations + index | |
| 2. Lint | references/step-02-lint.md | lint.mjs --fix --json; surface residual errors | |
| 3. Verify | references/step-03-verify.md | Invoke /lumi-verify grounding-only; user reviews findings | |
| 4. Finalize | references/step-04-finalize.md | Append log entry, mark ingest_status: finalized |
Stage-1 also relies on references/pdf-preprocessing.md (PDFs / large sources) and references/dedup-policy.md (before creating or updating any source / concept / person / edge).
Resume routing
This is the only logic in this body — everything else lives in step files. The router stays cheap to load whether the user is starting fresh or resuming on the last gate.
- No slug yet, or slug given but no entry on disk. → Read fully and follow
./references/step-01-draft.md.
- Slug given, entry exists. Read its
ingest_status:
``bash node _lumina/scripts/wiki.mjs read-meta sources/<slug> ``
ingest_status | Next | ||
|---|---|---|---|
| absent (legacy v0.8 entry) | Tell the user this entry predates the four-checkpoint workflow and ask: [A] Run lint+verify only (skip draft, the page already exists) \ | [Q] Quit. On [A]: write ingest_status: drafted to mark the entry as adopted, then → ./references/step-02-lint.md. On [Q]: exit cleanly. | |
drafted | → ./references/step-02-lint.md | ||
linted | → ./references/step-03-verify.md | ||
verified | → ./references/step-04-finalize.md | ||
finalized | HALT, ask "this entry is already finalized; restart from scratch?" On [Q] (decline): exit cleanly, exit 0 — declining is not an error. On confirm, do these in order before re-entering step-01: (1) set-meta sources/<slug> ingest_status drafted (so a session crash mid-restart leaves the entry in a coherent stage, not a stale-finalized one), (2) delete the phase checkpoint at _lumina/_state/ingest-<file-basename>.json — at this point you may only know the slug, not the file basename, so derive it first: read raw_paths[0] via read-meta sources/<slug> and take its basename; if raw_paths is empty or unavailable, list _lumina/_state/ingest-*.json and read each one to find the checkpoint whose slug field equals sources/<slug>; if no checkpoint matches either way, there is nothing to delete — simply proceed, (3) → ./references/step-01-draft.md. |
Examples
<example> "/lumi-ingest raw/sources/attention-revisited-2026.pdf" — fresh ingest, all four gates accepted: drafted → linted → verified (passed) → finalized. One log entry, lint 0, verify_status: passed. </example>
<example> "/lumi-ingest attention-revisited-2026" in a new session after quitting yesterday at the verify gate. SKILL.md reads ingest_status: linted, jumps to step-03 without re-running draft or lint. Verify gate presents findings; user accepts; finalize runs. </example>
<example> "/lumi-ingest raw/sources/old-paper.pdf" — entry already finalized. HALT and confirm restart. Decline → no-op. Confirm → checkpoint deleted, step-01 from top. </example>
Next step
After step-04 completes, suggest /lumi-check (structural review) or /lumi-verify --external <slug> (adversarial open-web check) in a fresh session or sub-agent. Blank context catches bias from the reasoning chain that just built these pages.