wiki-graph

wiki-graph

Structured graph queries over the wiki — without a persistent graph store.

The wiki already has a knowledge graph: every [[wikilink]] is an edge, every page is a node, every frontmatter field is a property. This skill just computes over that graph on demand. Each invocation scans the markdown files, builds an in-memory graph, runs the query, prints the answer, and exits. Nothing is cached — the files are the source of truth, always.

Karpathy's wiki treats the filesystem as the database. wiki-graph treats the filesystem as the graph.

When to use

• Structured Dataview-style questions: "all type: model pages tagged transformer updated since March"
• Traversal questions: "what's within 2 hops of [[claude-3]]?"
• Comparison discovery: "what's the shortest path between [[gpt-4]] and [[llama-3]] — do any concepts bridge them?"
• Topology questions: "which pages are hubs? which are orphans?"
• Visualizing a subgraph: render a Mermaid diagram of a cluster or neighborhood

wiki-search is for "find pages about X" (ranked text relevance). wiki-graph is for "pages where property P holds" and "pages connected to X" (structural queries).

Pre-flight

Orientation is not required — this skill runs cold against the file layout. It reads SCHEMA.md only to know which frontmatter fields are reserved types vs. free-form strings.

read_file {wiki_path}/SCHEMA.md  (optional — for type field names)

Steps

① Parse arguments. Exactly one of --query, --neighbors, --shortest-path, --hubs, --orphans, --cluster should be set (they are different query modes). --format and --limit apply to all modes.

② Scan the wiki (shared by all modes): Walk every .md file outside raw/. For each file extract:

• Path (relative to wiki root) — the node id
• YAML frontmatter — the node properties (title, type, tags, created, updated, published_at, ingested_at, tier_override, and every SCHEMA.md custom dimension)
• Outbound [[wikilinks]] in the body — the outbound edges
• Computed tier — per SCHEMA.md memory_tiers: pick the driving field, compare age to thresholds, apply any tier_override: pin

Tier filtering depends on mode:

• --query, --hubs, --orphans, --cluster — filter nodes by --tier (default active when tiers are enabled, else all). These are ambient-state queries and should reflect the "focus" surface.
• --neighbors, --shortest-path — include all tiers by default. The user named a specific seed, so connectivity matters more than surface focus. Each result annotates its tier ([active]/[archived]/[frozen]) so the user can eyeball where archived bridge concepts live. Pass --tier=active explicitly to restrict.

Build three in-memory structures:

nodes: {id → {properties}}
out_edges: {id → [target_id, ...]}
in_edges:  {id → [source_id, ...]}   # inverted while walking

Resolve [[wikilinks]] to node ids by matching the link text against page titles (frontmatter title) first, then filename stems. Unresolved links become dangling edges — reported in the --orphans / wiki-lint modes but not treated as nodes.

③ Dispatch to the query mode.

Mode: --query "<expression>"

Dataview-style filter over frontmatter. Supports:

• Field equality: type: entity, tags contains transformer
• Comparisons on dates/numbers: updated > 2025-01, sources >= 3
• Boolean combinators: AND, OR, NOT
• Parentheses for grouping

--query "type: entity AND tags contains model AND updated > 2025-01"
--query "type: comparison OR (type: concept AND tags contains attention)"
--query "NOT (tags contains stub) AND sources >= 3"

Evaluation: iterate nodes, evaluate expression against each node's properties, collect matches. Sort by updated descending by default (override with --sort=<field>).

Mode: --neighbors <page> [--depth=N]

BFS from the starting page over out_edges ∪ in_edges (undirected traversal — a wiki link is a bidirectional relation in practice, even if stored one-way). Default depth is 1; cap at 4 to keep output readable.

Return the layered neighborhood:

depth 0: [[claude-3]]                         (seed)
depth 1: [[anthropic]], [[constitutional-ai]], [[rlhf]]
depth 2: [[alignment]], [[dario-amodei]], ...

Each neighbor line also carries its type and top 2 tags so the user can scan for the ones that matter.

Mode: --shortest-path <a>,<b>

Undirected BFS from a to b. If no path exists, say so and suggest the closest candidates (smallest BFS frontier intersection). If multiple shortest paths of the same length exist, return up to 3.

Output the path with edge labels (frontmatter type of each node):

[[claude-3]] (entity) → [[rlhf]] (concept) → [[instruct-gpt]] (entity) → [[gpt-4]] (entity)

This is often the most useful mode — it surfaces bridge concepts between two entities the user didn't realize were connected.

Mode: --hubs [--limit=20]

Sort all nodes by |in_edges| + 0.5·|out_edges| descending (inbound links matter more — being linked to is a stronger signal than linking out). Return the top --limit with their link counts.

Mode: --orphans

Two classes:

• True orphans — no inbound and no outbound links (isolated nodes)
• Leaves — inbound links only, no outbound (dead-ends — not broken, but worth reviewing)

Report both separately. True orphans are almost always bugs (the ingest forgot to cross-reference). Leaves are sometimes intentional (a stub page waiting to be fleshed out).

Mode: --cluster=<tag>

Collect all nodes carrying <tag> in frontmatter. Report:

• Member count
• Most-linked member (the cluster anchor)
• Intra-cluster edges (links between members) vs. external edges (links leaving the cluster)
• Density = intra_edges / (N·(N-1)/2) — high density means the cluster is self-referential; low density means it's a loose grouping

Offers a quick way to ask "is this tag actually a coherent topic, or just a shelf?"

④ Format output.

• --format=text (default) — human-readable, excerpt above
• --format=json — machine-readable, for piping to other tools (e.g. jq)
• --format=mermaid — a graph TD block the user can paste into Obsidian or GitHub markdown. Best with --neighbors, --shortest-path, or --cluster.

Mermaid example for --neighbors claude-3 --depth=2 --format=mermaid:

graph TD
    claude-3[claude-3] --> anthropic
    claude-3 --> constitutional-ai
    claude-3 --> rlhf
    constitutional-ai --> alignment
    rlhf --> instruct-gpt
    anthropic --> dario-amodei

⑤ Suggested next actions. Based on what the query found:

• Orphans detected → suggest wiki-lint --fix to surface cross-reference candidates
• A cluster with zero synthesis page → suggest wiki-query --file "synthesis of {tag}"
• A shortest path through a surprising bridge → suggest wiki-query to write that connection up as a comparison page

Output format

[Operation] wiki-graph | {mode}: {argument}

[Scanned] {N} pages, {M} edges ({K} dangling)

[Results]
{mode-specific body}

[Summary]
{1–2 sentence interpretation of the result}

[Suggested next]
→ {next skill invocation}

---

Why there is no persistent graph

A persistent graph store (Neo4j, an index file, a pickled networkx) would be faster for very large wikis, but it introduces a second source of truth. The moment the agent updates a page and forgets to update the graph, queries silently lie.

Karpathy's design keeps the filesystem as the only source of truth. Every query is computed from the current state of the files. wiki-graph honors that rule: it builds and throws away its graph every call. A scan of several hundred pages is milliseconds — cheaper than the overhead of keeping an index coherent.

If the wiki grows past the size where ad-hoc scans feel slow (typically several thousand pages), the right upgrade is qmd in MCP mode — same model, but with a purpose-built search engine. A dedicated graph DB is rarely worth it.

Notes for the agent

• Respect SCHEMA.md field names — do not invent frontmatter keys. If the user's wiki uses kind: instead of type:, honor that.
• [[wikilinks]] may use aliases ([[claude-3|Claude 3 Opus]]) — resolve by the pre-pipe id.
• Treat heading-anchor links ([[page#section]]) as edges to page.
• Embedded images (![[image.png]]) are not edges — they belong to the asset layer, not the concept graph.
• For large wikis, do not load every file body into memory — only read frontmatter blocks and grep wikilinks out of the body. This is the hot path.
• This skill is read-only. It does not write pages, does not update index.md, and does not append to log.md.