oh-my-wiki 2.22.0

Karpathy-style LLM Wiki skill with multi-vault sqlite registry and Socratic wizard

oh-my-wiki (OMW)

A host-universal LLM-wiki you drive from your AI coding agent (Claude Code / Codex / Gemini).

oh-my-wiki exposes exactly two surfaces. The omw CLI handles deterministic ops — omw setup, omw vault create, omw lint, omw schema list, omw supersede, omw review, omw links, omw fields, omw view, omw doctor — with no LLM required. The omw skill brings natural-language reasoning inside your AI session: ingest, query, autoresearch, and a small set of wiki-maintenance personas (librarian, curator, fact-checker, consistency-checker, terminology-manager). The model is personas propose → you confirm → deterministic ops execute, so every file change is auditable. omw stays focused on the wiki — multi-step orchestration is left to your host AI agent (Claude Code / Codex / Gemini), not reimplemented here.

Short alias: OMW (lowercase omw). Both oh-my-wiki and omw register as skills and respond to the same trigger phrases.

Tutorial: Walk through real dialogs and verified CLI examples in TUTORIAL.md (English) or TUTORIAL.ko.md (한국어).

---

What's new in v3

• Schemas — 13 built-in page types (omw schema list/show), with per-vault overrides in <vault>/schemas/
• Confidence + supersede — confidence frontmatter field; omw supersede retires old pages cleanly
• Review queue (SR) — spaced-repetition via omw review due / omw review done
• Web search — omw search queries an external provider (brave/tavily/exa/…); omw serve exposes vault FTS5 as a local retrieve-only HTTP API on port 8765
• Entity-linking — omw links suggest / omw links link auto-inserts [[slug|Name]] references
• Inline fields — omw fields reads key:: inline syntax alongside frontmatter
• Korean matching — Korean entity names with josa (카르파시가) are suggested and linked correctly
• Note viewers — omw view opens the active vault, a page, or a search in Obsidian or Logseq (URI schemes, no plugin needed); omw setup viewer scaffolds the viewer config
• Visibility (secure-by-default) — omw visibility get/set marks pages public/private; omw serve exposes only public pages
• URL inbox + fetch — omw fetch <url> saves a web page or YouTube transcript to raw/ (tiered urllib → chromium → cloud, SSRF-guarded); omw inbox add/list/run/remove queues URLs for batch fetch

---

Install

Choose whichever path fits your environment. After the PyPI or git path, run omw doctor to confirm everything is wired correctly. After the Skills CLI path, the CLI is set up on first use (then omw doctor).

Path A — PyPI (pip / pipx) — recommended

Install the omw CLI from PyPI without cloning:

pipx install oh-my-wiki        # isolated CLI (recommended)
# or
pip install oh-my-wiki         # into the current environment

Both give you a working omw command (omw status, omw vault create …, omw lint, …). The published wheel is self-contained — it bundles the schemas, personas, backends, and the full skill. To register the bundled skill with your agents afterwards, run:

omw setup agents

Installing straight from GitHub works the same way: pipx install git+https://github.com/dandacompany/oh-my-wiki.

Path B — git clone + install script (developers, Codex CLI users)

git clone https://github.com/dandacompany/oh-my-wiki
cd oh-my-wiki
bash bin/install.sh

The installer checks for Python 3.10+, pip-installs the package editable, creates ~/.claude/skills/oh-my-wiki and ~/.claude/skills/omw symlinks (idempotent), runs pytest -q to verify, and prints next steps. Add --dev to include pytest/ruff extras. Use --force to replace existing symlinks without a prompt; --no-test to skip the test step. Run bash bin/install.sh --help for all flags.

Path C — Skills CLI (Claude Code users)

skills add dandacompany/oh-my-wiki@oh-my-wiki -g -y --copy -a claude-code

This installs the skill into ~/.claude/skills/ and registers both the oh-my-wiki and omw short-alias skill names. This installs the skill only. The omw CLI is installed on first use — open your agent and say set up omw (or omw 셋업 점검해줘); the skill runs its CLI preflight and installs the CLI with your confirmation. Or install it yourself now: pipx install oh-my-wiki.

Verify the install

omw doctor

omw home:   /Users/you/.omw  ok
registry:   /Users/you/.omw/registry.db  ok
  * demo (wiki/markdown) /Users/you/.omw/vaults/demo

---

Quickstart (~60 seconds)

Step 1 — Run the setup wizard

omw setup

Follow the prompts to configure your first vault, search provider, and persona preferences. Accept the defaults for a fast start.

Step 2 — Check status

omw status

{
  "vault_count": 0,
  "active": null,
  "needs": "setup",
  "vaults": []
}

Step 3 — Create your first vault

omw vault create demo --mode wiki

{
  "created": "demo",
  "path": "~/.omw/vaults/demo",
  "mode": "wiki",
  "type": "markdown"
}

omw vault list

[
  {
    "name": "demo",
    "path": "~/.omw/vaults/demo",
    "mode": "wiki",
    "type": "markdown",
    "is_active": true
  }
]

Step 4 — Add a note (in your AI session)

Open Claude Code (or Codex / Gemini) and say:

ingest this

Andrej Karpathy calls the LLM Wiki a "compounding knowledge artifact". Every
source gets saved verbatim to raw/, a summary lands at wiki/summaries/, and
the entities and concepts that appeared get their own pages. 10–15 page touches
per ingest is normal.

Step 5 — Run a lint check

omw lint

{
  "vault_id": 1,
  "vault_path": "~/.omw/vaults/demo",
  "frontmatter_issues": [],
  "drift": { "missing_files": [], "mtime_drift": [] },
  "links": {
    "broken": [],
    "orphans": [],
    "index_drift": { "missing_from_index": [], "dangling_in_index": [] },
    "contradictions": [],
    "supersedes": [],
    "superseded_unmarked": [],
    "link_suggestions": []
  },
  "auto_fix_hints": []
}

→ Full tutorial: TUTORIAL.md · 한국어

---

Architecture

SKILL.md dispatcher → commands/<op>.md (LLM procedure) → scripts/<op>.py (deterministic I/O)
                                                       └─ registry.py → ~/.omw/registry.db (sqlite)
                                                       └─ adapters.py → filesystem (markdown / obsidian)

The 17 CLI subcommands:

Subcommand	Purpose
status	Show active vault and registry state
vault	Create, list, use, forget vaults
lint	Structural health check (frontmatter + links)
search	Web search via the configured external provider (brave/tavily/exa/…)
serve	Local retrieve-only HTTP query API (port 8765) — public pages only
view	Open the vault / a page / a search in Obsidian or Logseq
visibility	Get / set a page's public/private visibility (get / set)
schema	List / inspect page-type schemas
supersede	Mark a page superseded by a newer one
review	Spaced-repetition review queue (due / done)
links	Suggest and insert [[slug]] entity links
fields	Read frontmatter + inline key:: fields
import	Import an existing folder as a vault
fetch	Fetch one URL (web page / YouTube transcript) into raw/
inbox	Queue URLs and batch-fetch them into raw/ (add/list/run/remove)
setup	Interactive setup wizard
doctor	Verify install health

Visibility (secure-by-default): omw serve returns only pages with visibility: public in their frontmatter. Pages without the field are treated as private and never served. Publish pages explicitly with omw visibility set <relpath...> public.

The skill also exposes natural-language ops via your AI session: ingest, query, autoresearch, find, edit, move, delete, and wiki-maintenance persona invocations (fact-check, consistency-check, build glossary).

---

Storage

• The vault registry lives at ~/.omw/registry.db (override with OMW_HOME) as a per-user sqlite database.
• The note index is regenerated by scripts/reindex.py after every mutation.
• Your files stay in the vault path you chose. oh-my-wiki never touches them outside the op you explicitly invoked.

---

Development

• pytest -v runs all tests.
• ruff check scripts/ tests/ runs the linter.
• omw status inspects the registry/vault state.
• python3 -m scripts.lint --vault-id N runs the health check on a specific vault.

Continuous integration runs on GitHub Actions, across a matrix of Python 3.10, 3.11, and 3.12 on both ubuntu-latest and macos-latest.

---

License

Released under the MIT License. See LICENSE for the full text.