lumi-research-prefill

---
name: lumi-research-prefill
description: >
  Seed terminal foundation pages for stable background knowledge so future
  ingestion can link to common concepts without duplicating definitions.
allowed-tools:
  - Bash
  - Read
  - Write
---

# /lumi-research-prefill

## Role

You create stable foundation pages under `wiki/foundations/` for background
concepts. Foundation pages are terminal references: normal wiki pages may link to
them, but they do not require reverse links.

## Context

Read `README.md` first. Use the research pack Wikipedia fetcher when background
material should come from a public encyclopedia:

```bash
python3 _lumina/tools/fetch_wikipedia.py --help
node _lumina/scripts/wiki.mjs read-meta foundations/<slug>
```

When speaking with the user, follow the README language rule exactly. Use the
configured communication language, translate workflow terms, and avoid technical
tool words unless the user asks for them.

## Instructions

1. Turn the requested topic into a slug with:

```bash
node _lumina/scripts/wiki.mjs slug "<topic title>"
```

2. Check whether `wiki/foundations/<slug>.md` already exists:

```bash
node _lumina/scripts/wiki.mjs read-meta foundations/<slug>
```

   - **exit 2 (not found)** — continue to step 3.
   - **exit 0 (exists)** — read the file, then show the user the `title`, `created`
     date, and the most recent line in `wiki/log.md` that references this slug.
     Ask exactly one question with three options in the user's language (no other
     action until answered). Do not expose internal terms; explain the choice as
     "leave it unchanged", "refresh the public background text", or "stop and log
     the decision":

     ```
     Foundation "<title>" already exists.
       [s] skip    — abort, no changes (default)
       [r] refresh — re-fetch from Wikipedia, update non-marked sections and `updated`; preserve `created`, `aliases`, and `<!-- user-edited -->` sections
       [a] abort   — same as skip but log the user's intent
     ```

     Do not proceed without an explicit choice. Map blank/Enter to `skip`.

3. Fetch or handle background material based on exit code from the Wikipedia fetcher:

```bash
python3 _lumina/tools/fetch_wikipedia.py page "<title>"
```

   - **exit 0** — use the JSON output directly.
   - **exit 2 AND stderr is JSON with `kind == "disambiguation"`** — run a search
     to surface candidates:

     ```bash
     python3 _lumina/tools/fetch_wikipedia.py search "<title>" --limit 5
     ```

     Present the numbered results (title + snippet) in the user's language and let
     the user:
     - pick a candidate number,
     - type a more specific title, or
     - type `manual` to paste content directly.

     Re-run `page` with the chosen title, or accept the user-pasted content.

   - **exit 2 for any other reason** (empty title, page not found) — surface the
     `error` field from stderr JSON and abort.
   - **exit 3 (network error)** — tell the user and offer two options: retry, or
     paste content manually.
4. Write `wiki/foundations/<slug>.md` with valid foundation frontmatter:
   `id`, `title`, `type: foundation`, `created`, `updated`.

   Also include the optional `aliases` field — an array of strings listing
   alternative names users or sources might write for this concept (abbreviations,
   expansions, common misspellings). Example: for a foundation titled "Reinforcement
   Learning from Human Feedback", use `aliases: ["RLHF", "human feedback RL"]`.
   Propose a list of 2–5 plausible alternate names, then ask the user to confirm
   or edit before writing. An empty list is fine if nothing obvious applies.
   Alternate names must be unique across all foundations — `lint.mjs` L10 will
   error on collisions.

   Foundation pages do not carry `provenance` or `confidence` frontmatter fields
   (those belong on source pages). When the user later ingests a source that
   references this foundation material, `/lumi-ingest` will set those fields on the
   source page. To help that decision, note in the log entry whether the fetch was
   `replayable` (Wikipedia snapshot saved to `raw/`), `partial` (summary only), or
   `missing` (no fetch — manual entry). This keeps the audit trail clear without
   polluting foundation frontmatter.

   Example log note: `prefilled foundation reinforcement-learning (provenance: replayable — Wikipedia snapshot saved)`
5. Keep the body concise: definition, scope notes, and external references.
6. Log the addition:

```bash
node _lumina/scripts/wiki.mjs log lumi-research-prefill "prefilled foundation <slug>"
```

7. Run lint with fix so `wiki/index.md` and structural checks stay current:

```bash
node _lumina/scripts/lint.mjs --fix --json
```

## Constraints

- Do not create concept pages from this skill; use `/lumi-ingest` for wiki
  knowledge extracted from project sources.
- Do not store secrets or API keys in foundation pages.
- Do not add reverse graph edges for foundations.
- When refreshing an existing foundation, preserve the original `created` date and
  any `<!-- user-edited -->` sections verbatim. Only `updated` and non-marked
  sections may change.
- Aliases must be unique across all foundations. If `lint.mjs --fix --json` reports
  `L10-alias-conflict`, resolve manually before completing the run — there is no
  automatic fix.

## Definition of Done

- Foundation page exists with valid frontmatter and concise source-backed body.
- `node _lumina/scripts/lint.mjs --fix --json` has updated `wiki/index.md` if
  needed and leaves `summary.errors === 0`.
- `wiki/log.md` has an append-only `lumi-research-prefill` entry.
- If the page already existed, the user's choice (skip / refresh / abort) is logged
  in `wiki/log.md` with the actual decision taken.