wiki-ingestlisted

# wiki-ingest — incremental ingest & update A wiki is a **compiled knowledge representation, not a retrieval dump.** New material doesn't get appended blindly — it routes to the page it belongs on, merges in, and only spawns a new page when nothing covers it. The failure mode to avoid: dumping raw content that leaves the real pages stale. A smaller, curated wiki beats a larger unvetted one. **Agent topology (required):** - **Root = opus-tier** (the strongest model the harness offers). Splits source material into atomic claims, decides route vs. merge vs. overwrite vs. new, judges contradictions, and writes the final entries. Every judgment call lives here. - **Fan-out = haiku-tier** (the cheapest model). One sub-agent per candidate claim/topic: runs `ground` to find the nearest existing page, reads it, and returns the match, its similarity score, and the relevant existing lines. Retrieval and reading are fanned out; decisions are not. ## Phase 1 — Atomize (root / opus) - Take the source (a file path, pasted doc, a conversation takeaway, a decision) and split it into **atomic claims** — one topic each, the same granularity as a wiki page section. Don't ingest a 10-page doc as one blob. - Read `wiki-conventions.md` (the wiki's constitution) for slug/voice/merge rules. If absent, fall back to hallouminate's authoring conventions (one topic per file, H1 first line, kebab slug). - Pick the corpus (`repo:{name}:wiki` or ask). ## Phase 2 — Locate (haiku,