Kp Wiki

Instructions

When to activate

Activate when the user invokes /kp-wiki build, /kp-wiki update, /kp-wiki add-page, or /kp-wiki enrich, or asks to build, rebuild, update, or enrich the wiki.

Sub-commands

build

Full rebuild from all staged content.

Run:

<skill-dir>/scripts/wiki_build.py --mode build

The script executes this pipeline in order:

1. Read all staging/<source-id>.json files. Validate schema_version.
2. Read frontmatter from all wiki/queries/ files (feedback layer).
3. Resolve entities - merge aliases across staging files, deduplicate by canonical name.
4. Write wiki/glossary.md - alphabetically sorted, one entry per resolved entity with aliases, definition, sources, and [[wikilinks]] to related terms.
5. Write wiki/<type>/<topic>.md - one page per entity. Pages with generated: false or manual: true frontmatter are skipped.
6. Write wiki/index.md - Obsidian entry point, pages grouped by entity type.
7. Write wiki/index.yaml - agent entry point (see schema below).
8. Validate all [[wikilinks]] and file: entries in index.yaml. Report broken links as warnings; never fail the build for broken links.

update

Incremental rebuild. Only regenerates pages whose staging files have a extracted_at timestamp newer than the page's last_built frontmatter date.

Run:

<skill-dir>/scripts/wiki_build.py --mode update

enrich

Retroactive sweep for enrichment gaps that were not resolved inline by /query (e.g. past questions asked before inline enrichment was added, or gaps where the inline enrichment failed).

Run the automated pipeline:

<skill-dir>/scripts/wiki_build.py --mode enrich

The script outputs:

• Which query files still have enrichment_needed: true
• Which sources to re-stage (read from the target pages' frontmatter)
• Which flags were already cleared (target page now has body content)

After running the script, execute these steps automatically:

1. For each source ID listed in the script output, invoke the kp-staging skill with --force on that source.
2. Run <skill-dir>/scripts/wiki_build.py --mode build to write enriched body content into wiki pages.
3. Run <skill-dir>/scripts/wiki_build.py --mode enrich one more time to clear the flags on query files whose target pages now have body content.

Report a summary: how many gaps were found, which sources were re-extracted, how many flags were cleared.

If there are no gaps, say so and stop.

add-page <topic>

Create a stub page at wiki/topics/<topic-slug>.md with generated: false so it is never overwritten by future builds. Open for the user to edit.

wiki/index.yaml schema

schema_version: "1"
last_built: "<ISO datetime>"
source_count: 12
entity_count: 84

glossary: glossary.md

pages:
  concepts:
    - title: "Large Language Model"
      file: concepts/large-language-model.md
      aliases: ["LLM", "language model"]
      sources: [annual-report-2024, prospectus-q1]
  people:
    - title: "Jane Smith"
      file: people/jane-smith.md
      sources: [annual-report-2024]
  organizations: []
  events: []
  topics: []

gaps:
  - topic: "fine-tuning costs"
    question_count: 3
    max_confidence: low

gaps is populated from wiki/queries/ entries where confidence is low or medium, grouped by recurring topic keywords.

Entity page frontmatter

---
title: Jane Smith
entity_type: person
generated: true
sources:
  - annual-report-2024
last_built: 2026-06-21
---

Entity page body

When an extraction includes a body field on an entity, the page content is:

1. context — the one-liner, written as the lead paragraph under the heading.
2. body — the full markdown content (rules, tables, examples) appended after the lead paragraph.

If the same entity has body content from multiple sources, each body is rendered under a ### From <source-id> sub-heading. Entities with no body render the context one-liner only (unchanged behavior).

Link policy

Links use bare slugs: [[widget|Widget]]. Slugs must be globally unique across all entity subdirectories (concepts/, people/, products/, etc.). Obsidian resolves [[slug]] links by basename; other consumers must implement equivalent lookup.

If two entities produce the same slug (e.g. "Widget" as both a product and a concept), the build disambiguates by appending the entity type: widget-product, widget-concept. A warning is printed and the build continues.

Edge cases

• No staging files: print a helpful message, suggest /kp-staging --all. Exit cleanly.
• Partial staging: build from what exists, warn about sources with extraction.status != "complete" in .meta.json. Legacy extracted: true is treated as complete.
• Broken [[wikilinks]]: report as warnings with originating page. Do not fail.
• schema_version mismatch in an extraction: skip that source and warn.

References

See references/kb-schema.md for full index.yaml schema and page frontmatter spec.