wiki-spaces 1.2.0

Minimal nestable wiki for AI coding harnesses. A folder with `index.md` and a `## Spaces` navigation contract, for any use case (notes, research, recipes, writing, team docs). Spec, conventions catalog, and reference skills. Out of scope: browser-only AI assistants.

wiki-spaces

A minimal nestable wiki — a folder with index.md and a ## Spaces navigation contract, for any use case. Research, recipes, code notes, writing, team docs, a personal life wiki — your shape, your call.

Markdown flavor is Obsidian — wikilinks, frontmatter, callouts, embeds, comments, Bases. One dialect across the spec, the skills, and the tools. Non-Obsidian renderers (GitHub preview, vanilla VS Code, plain markdown viewers) display the content but will not render Obsidian-specific syntax (provenance comments, embeds, .base files) the way Obsidian does — view your wiki in Obsidian for full fidelity.

Audience

wiki-spaces is built for AI coding harnesses with filesystem access — Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Aider, and similar tools that read and write local files. Browser-only AI assistants (ChatGPT in a tab, Claude.ai web) are out of scope: they can't reach the filesystem to read or write your wiki. If your AI lives in a browser, this isn't the tool.

The shape of the wiki — research notes, recipes, journal, team reference, anything — is yours. The harness that drives it is what wiki-spaces assumes you have.

Install

Installing gives your agent the reference skills plus a scaffolded, registered wiki. Letting an agent do it is the recommended path — setup is a short interview, and an agent runs the steps end to end without fat-fingering a path or a flag.

Let your AI agent do it (recommended)

Paste this to your coding agent (Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Antigravity, …):

Install and set up wiki-spaces for me by following the instructions here:
https://raw.githubusercontent.com/anfreire/wiki-spaces/main/references/SETUP.md

The agent reads SETUP.md, asks what the wiki is for and where it should live, infers a layout, links the skills into your harness, scaffolds the wiki, and writes the config — confirming the plan with you before it runs anything.

Manual

uvx wiki-spaces install                  # link skills into detected harnesses
uvx wiki-spaces init ~/Documents/Wiki              # scaffold a fresh wiki + register it
uvx wiki-spaces init ~/notes --adopt     # OR adopt an existing folder of notes
uvx wiki-spaces doctor --no-net          # verify

install auto-detects and links the skills into whichever of Claude Code, Codex, Gemini CLI, Antigravity, Hermes, and Kiro are present (--all pre-positions for every one). Cursor and Windsurf have no global skills directory — they integrate via a generated rule-file snippet (wiki-spaces install --bridge cursor|windsurf). GitHub Copilot and Aider integrate via a manual rule snippet — see HARNESS_INTEGRATION.md for both. For a permanent install, pip install wiki-spaces or uv tool install wiki-spaces, then drop the uvx prefix.

init always emits ## Spaces in the new wiki's index.md from t=0 so the space subcommands work immediately. --adopt walks the existing folder, registers every nested folder containing index.md in its ancestor's ## Spaces, and inserts ## Spaces into any pre-existing index that lacks the heading (the v1 navigation contract — a wiki is a folder with index.md AND ## Spaces). After adoption every folder carries both, so audit reports zero drift on day 1. External subtrees (shared/, foreign submodules, escaping symlinks) are skipped with a per-skip notice; pass --include-external to override.

Once a wiki exists, the space subcommands manage its structure:

uvx wiki-spaces space add projects/foo   # create a space + register it
uvx wiki-spaces space audit              # audit drift, broken links, orphans
uvx wiki-spaces space mount <url> --mode submodule              # mount an external space (default path: shared/<basename>/)

space add, space remove, space mount, and space promote auto-insert ## Spaces when missing — the CLI maintains the contract for you. wiki-spaces init writes ## Spaces from t=0 so the first write command lands clean.

No tooling at all

mkdir -p ~/Documents/Wiki && printf '# My Wiki\n\n## Spaces\n\n' > ~/Documents/Wiki/index.md

A folder with index.md + ## Spaces is already a complete wiki — the whole spec is one page, AGENTS.md. The skills still work on it (they discover the wiki from your current directory), though they lean on the wiki-spaces CLI for cap checks, audit, and structural edits and fall back to manual procedures without it — installing it makes them most effective. Run wiki-spaces init later to register the wiki for config-based discovery.

Skill source paths

Users with their own skill-management scripts bypass install.py and read the source directories directly. The two canonical roots are:

• <repo>/skills/*/ — ws-search, ws-update, ws-tend
• <repo>/vendor/kepano/*/ — obsidian-markdown, obsidian-bases

<repo> is the source checkout or the share dir written by wiki-spaces install (~/.local/share/wiki-spaces/ for packaged installs). Copy, symlink, or aggregate however your tooling prefers.

What you get

Three reference skills your AI agent uses to work with the wiki:

• ws-search — find content
• ws-update — capture / save / sync, with per-file size discipline (hard caps at write time, see CONVENTIONS.md § _meta/limits.md)
• ws-tend — audit, normalize tags, cross-link

Cursor / Windsurf / Copilot / Aider integration is covered under Install above and in HARNESS_INTEGRATION.md.

Search at scale

ws-search works out of the box with grep / ripgrep, which is fine for personal/team wikis up to a few hundred pages. For larger vaults, install qmd — the markdown-aware MCP backend Andrej Karpathy references in the canonical LLM-wiki gist. See CONVENTIONS.md § Recommended search backends.

Learn more

• AGENTS.md — the spec, one page
• CONVENTIONS.md — opt-in conventions catalog
• references/EXAMPLES.md — topology examples per use case
• references/MOUNT.md — mount external spaces into your wiki

Dependencies

Python >=3.11. uv recommended (handles Python provisioning). git optional.

Prior art

• Andrej Karpathy's LLM Wiki gist
• Ar9av/obsidian-wiki — broader framework wiki-spaces extracts from
• kepano — vendored obsidian-markdown and obsidian-bases skills (MIT-licensed; see vendor/kepano/LICENSE)

License

MIT. See LICENSE.