Mc Wiki Ingest

Wiki Ingest

A bundled MulmoClaude preset skill (mc- prefix = launcher-managed; do not edit this file in the workspace, it is overwritten on every server boot).

This is the ingest operation from the LLM-Wiki pattern (RFC #1491 Phase A, Karpathy's gist). The first writes-bearing preset in the wiki triad — it modifies data/wiki/ by design. Sister read-only presets: mc-wiki-health-check (structural lint, scheduled) and mc-wiki-deep-lint (LLM lint, on-demand).

Inputs

Accept exactly one of:

• a workspace-relative file path the user attaches — e.g. data/..., sources/..., artifacts/documents/.... Workspace-rooted paths only. Reject (don't Read):
  • absolute paths (/etc/..., /Users/..., /var/...)
  • home-relative paths (~/..., ~user/...)
  • parent-traversal segments (../, embedded .. after normalising)
  • any path whose path.resolve lands outside the agent's workspace root Canonicalise (realpath-aware via Read's own checks) and verify the resolved path is still inside the workspace before reading. A source outside the workspace must be copied into data/sources/ first by the user, then re-ingested from that workspace path.
• pasted text the user sends in the same message (paste an article, notes, an email body, etc.) — use what's in the user turn as the source

Do NOT fetch URLs. If the user gives only a URL, ask them to open it and paste the article text. (URL fetch / batch input are deferred to a later phase — see #1527.)

Size guardrail

If the source exceeds 100 KB (~25k tokens, the safe budget for a single ingest pass), stop early with a clear message:

"Source is N KB, above the 100 KB ingest limit. Summarise it first (paste the summary), or split it into sections and ingest one at a time."

Don't truncate silently. Don't chunk-iterate (that's a later phase).

What to do (in order)

1. Build a slug + summary page

• Derive <slug> from the source title (lowercase, kebab-case ASCII; if the title is non-ASCII, transliterate or use the most prominent English noun phrase).
• If data/wiki/pages/<slug>.md does not exist: write a new page — H1 with the topic name, then a structured body (overview, key points, any dated claims). Each generated bullet must end with the provenance marker (see step 5).
• If data/wiki/pages/<slug>.md already exists (re-ingest): append a new section ## Updated YYYY-MM-DD at the bottom with the new reading. Do not overwrite the existing body. Do not auto-merge via LLM rewriting — appending is the contract (see #1527 Q5).

2. Pick up to 5 related pages for cross-reference

Sources of relevance to combine:

• data/wiki/index.md — categories + tags
• manageWiki graph action (introduced by #1520) — page→page link structure; prefer pages that already link to topics in the source
• LLM judgment over page titles + first H2 sections — semantic proximity

Cap at N=5. This bounds the rollback surface if the run is interrupted: at most 5 page edits to inspect. Pick the 5 most relevant by relevance score; if fewer than 5 pages look relevant, edit fewer.

3. Update the cross-referenced pages

For each of the (up to) 5 picked pages, append a single bullet or a short link reference under the most appropriate H2 — never rewrite the existing body. The bullet ties the existing page to the new source via a [[<new-slug>|<display>]] link plus a one-sentence reason the link is being made.

Display-text normalisation (mandatory — the raw source title can contain characters that break wiki-link parsing or the downstream graph / lint passes):

• Remove every [, ], and | character (these are wiki-link syntax delimiters; even one stray ] corrupts the link)
• Replace newlines / tabs / runs of whitespace with a single space
• Trim leading / trailing whitespace
• Truncate to 80 characters (with a trailing … if cut). Long titles otherwise produce unreadable bullets

If normalisation leaves the display empty (the title was entirely syntax characters), fall back to <new-slug> itself (omit the | section: [[<new-slug>]]).