knowledge-wiki

name	knowledge-wiki
description	Query, ingest, maintain, and lint a local Markdown knowledge wiki with immutable raw sources, claim-level provenance, reconciled entity/concept pages, and hybrid GraphRAG retrieval.
license	MIT
metadata	{"version":"4.4.0","compatibility":"Python 3.9+, Ollama by default, WIKI_ROOT env var or config.yaml. Optional: Playwright, lxml_html_clean, trafilatura.","subsystem":["query","ingest","lint","maintain"]}

Knowledge Wiki

Use this skill when the user wants to query, ingest into, repair, or evaluate the local knowledge wiki. For installation and human-facing usage, read README.md.

Invariants

• Treat raw/ Markdown files as immutable evidence after ingest.
• Route canonical wiki writes through the maintainer boundary in scripts/wiki_core.py.
• Store structured claims in wiki/claims/<page-kind>/<slug>.json; Markdown pages are the human-readable projection.
• Reconcile entity and concept pages after claim updates. Preserve human text outside generated sections.
• Ordinary query answers are ephemeral. Persist only when the user explicitly asks to promote/save the answer.
• Prefer compiled wiki pages for query context, then load raw sources for evidence verification or gap filling.
• Do not introduce an external database or hosted vector store for this skill.

Canonical Commands

# Query without writing a synthesis file
python3 scripts/wiki_query.py --question "What do I know about X?"

# Persist a cited answer intentionally
python3 scripts/wiki_query.py --question "What do I know about X?" --promote

# Ingest content and reconcile affected pages
python3 scripts/ingest_source.py --url "https://example.com/post" --category research
python3 scripts/ingest_source.py --file "/path/to/note.md"
python3 scripts/ingest_source.py --text "..." --title "Manual note" --category notes

# Reconcile or migrate claim ledgers
python3 scripts/wiki_reconcile.py agent-memory --kind concepts
python3 scripts/migrate_claim_ledger.py --apply

# Rebuild and inspect derived state
python3 scripts/wiki_graph_builder.py --force
python3 scripts/regen_index.py
python3 scripts/wiki_lint.py
python3 scripts/wiki_maintain.py

# Validate packaging metadata
python3 scripts/validate_skill_metadata.py SKILL.md

Configuration

Resolve the wiki root as CLI --wiki-root, then WIKI_ROOT, then config.yaml, then ~/knowledge. Model/provider settings live in config.yaml: the top-level llm: block is the local Ollama default, and optional llm_profiles: entries can override query, ingest, relevance, categorize, or retrofit independently. Keep remote providers opt-in; use api_key_env: OPENROUTER_API_KEY or another environment-variable name, never a literal key in config or docs. Local .env files are loaded automatically for secrets, but existing process environment variables take precedence.

Operational Notes

• For YouTube or embedded videos, default to transcript-only ingest unless the user explicitly asks for frame or visual analysis.
• For Substack Notes where comments/reactions matter, prefer the normal requests fetch path over --use-browser.
• Run deterministic tests before live smokes. Live URL, browser, Ollama, and embedding checks are optional smoke tests, not the regression oracle.
• Standalone Python scripts cannot call the invoking agent's model directly; all generation goes through scripts/llm_client.py.