🧠 wikify repo demo

You're on main — the empty template. The actual populated demo (two ingested codebases as grounded, cross-linked wikis + the live graph) lives on the demo branch. main deliberately omits that ingested content so “Use this template” works — GitHub can't generate a template from a repo with git submodules, and the demo pins its sources as submodules. When you use the template, leave “Include all branches” off (the demo branch carries the submodules).

The companion demo for wikify-repo. wikify-repo is the tool — a CLI + agent skill that ingests a code repo into a grounded markdown wiki. This repo is what that tool produces: a live, browseable example so you can see the output before installing anything.

It's a Karpathy LLM-wiki that handles two source types in one knowledge base: code repos — ingested by the wikify-ingest-repo skill (which ships with wikify-repo, not defined here) into a grounded, lint-clean markdown wiki an agent answers internals questions from — and articles / docs / notes, ingested the classic Karpathy way (read → summarize → cross-link). Code is the headline; prose rides along in the same index.md / log.md. See SCHEMA.md for the full workflow.

This repo is both:

• a showcase — browse a real, populated wiki on the demo branch; and
• a template — you're on it: click “Use this template” to start your own code-wiki with the skills and agent conventions already wired up.

The whole wiki at a glance

(click image for interactive view)

A snapshot of the live viewer (92 pages · 384 links) — the populated wiki on the demo branch. Every node is a wiki page; every edge a markdown link between pages. The two dense clusters are the two ingested codebases — mini_pytorch_xla and wikify-repo — joined through the shared root index.md and the prose pages. Code pages are colored by page type (code:concepts, code:catalog, code:doc-concepts, code:overview); prose by folder (topics, sources).

Repo layout (what a populated wiki looks like)

SCHEMA.md              single source of truth — the three layers, ingest/query/lint, retrieval rules.
raw/                   immutable inputs — never modified.
  code/<slug>/           ingested repos as git SUBMODULES, pinned by gitlink (this demo defaults to acquire: submodule).
  sources/               ingested articles / docs / notes — committed source of truth.
wiki/                  LLM-owned markdown — the product.
  index.md               read-first catalog: code repos, cross-repo concepts, topics, sources, notes.
  log.md                 append-only event log (## [date] <op> | <name>).
  code/                  code wikis (one dir per repo) + an auto code catalog (index.md).
    <slug>/                one CODE wiki per repo:
      overview.md            high-level map — main concepts + system diagrams.
      concepts/              grounded mechanism pages (prose + Mermaid + citations).
      catalog/               per-module symbol catalogs — signatures, docstrings, source links, uses-by.
      doc-concepts/          pages derived from the repo's own docs.
  concepts/              cross-repo vocabulary — each concept links down to every repo's implementation (via wikify-connect-repo).
  sources/               one summary page per ingested article.
  topics/                synthesized prose pages — entities, concepts, comparisons.
  notes/                 cross-cutting answers filed back from queries.

On this template branch wiki/ starts empty (just index.md + log.md stubs) and raw/ doesn't exist yet — both fill in on your first ingest. To browse a populated one, switch to the demo branch, whose code sources are git submodules (pinned by gitlink):

git clone --recurse-submodules -b demo https://github.com/vlasenkoalexey/wikify-repo-demo
# already cloned?  git checkout demo && git submodule update --init

Browse it (no install)

On the demo branch, open wiki/index.md and follow links, or grep -ri "<term>" wiki/. Everything is plain markdown — no embeddings, no database. An agent answers cheaply by reading index.md, grepping to the right page, and citing the catalog anchor.

Make your own

1. Use this template (button above) — or git clone this branch. Leave “Include all branches” off (the demo branch's submodules would break generation).

2. Install the tooling once — see wikify-repo (pip install -e . → setup-vendor.sh → install-skill.sh /path/to/this-repo).

3. In Claude Code, Codex, or Antigravity, just say:

   ingest https://github.com/owner/your-repo

   The skill pins + indexes the repo, writes the grounded pages, lints them, assembles wiki/<slug>/, and updates index.md + log.md. Re-running is idempotent; ingest --ref <commit> updates to a newer commit and rebuilds only what changed. From the second repo on, it can connect shared concepts across repos into the cross-repo wiki/concepts/ pages (the wikify-connect-repo skill).

Why a wiki and not just grep over the source?

The wiki sits between raw repos and your questions: the cross-references are already there, the dynamic-dispatch seams a call-graph walk misses are covered by deterministic catalogs, and every mechanism page is grounded in real SCIP symbols. The tedious part — the bookkeeping — is done once and kept current, so answers are fast, grounded, and consistent. Conceptually, the generated wiki is an annotated map of your codebase that an LLM agent can navigate efficiently.

Note on citation anchors: a citation like catalog/<module>.md#Buffer.shape resolves against the catalog's frontmatter symbols: map (that's what the linter gates), not against a rendered heading — so clicking one on github.com opens the right catalog file but won't auto-scroll. Grep the anchor name inside the file, or read the wiki locally.