--- type: decision tldr: Schema 1.7 adds agent-built topic pages in wiki/topics/ (type topic) whose ## Sources are member-page digests, so lint flags them when a member changes; a new ckb-consolidate skill writes a proposals-only report (topics, merges, missing links, stale pages, old-log digest) and builds only the topic pages the user approves. status: accepted decided_on: 2026-11-02 decided_by: Michał Kopeć affects: AGENTS.md, wiki/index.md, wiki/topics/index.md, .agents/skills/ckb-consolidate/, .agents/skills/ckb-retrieve/, .agents/skills/ckb-upgrade/, .agents/skills/ckb-init/, .agents/skills/ckb-reset/, .agents/skills/ckb-routine/ confidence: 0.7 quality: 0.8 last_updated: 2026-11-02 freshness_window_days: 365 retention: high --- # D-0016 — Topic summaries in wiki/topics/, and a consolidation report **Status:** Accepted · **Decided:** 2026-11-02 · **Deciders:** Michał Kopeć ## Context Release 3, batch 2: F13 (a background consolidation pass) and F15 (topic summaries for whole-KB questions, after Microsoft GraphRAG's community reports). Both look across the wiki for structure no single ingest sees. Both risk the same failure: an agent rewriting pages on its own judgment while nobody is watching. ## Decision 1. **Topics live in `wiki/topics/`** (the user chose this over `outputs/`), as `type: topic` pages with `## Summary`, `## Members`, `## Sources` and a protected `## Notes`. Their `## Sources` are the member pages with digests, so lint check 12 flags a topic when any member changes. This is schema **1.7**: additive, a new optional directory and type. 2. **A topic is a map, not evidence.** Every claim on it must already be on a member page, and retrieval follows its sources before grounding an answer. "What does the KB say overall about X" questions route to `wiki/topics/index.md` first. 3. **One skill, `ckb-consolidate`, for both.** Its script clusters pages (project scope members, then shared tags, then linked groups of three or more) and writes a report of proposals: topics missing or outdated, merge candidates, missing links, stale pages, and old log entries by month. It never edits a page. The skill builds only the topic pages the user approves; merges and links go to the skills that own those edits. 4. **The log is never compacted.** Rule B's log stays complete; the report may summarise old entries, but it doesn't rewrite them. 5. **As a routine, it asks first** (`--every 14d --confirm`) and never builds topics unattended. ## Alternatives considered - **Topics in `outputs/`** — kept out of the schema, but out of retrieval's first pass too; rejected by the user. - **Two skills** (consolidate, topics) — one more always-loaded description for work that shares one scan. - **Auto-applying safe proposals** (missing links) — rejected: proposals accumulate trust only if a person sees them. ## Consequences - A busy topic is flagged often, because every member edit drifts it. That is intended: the summary may no longer hold. - Clustering is deliberately simple and explainable, not community detection; the report says why each cluster exists.