graphify-on-graphify — interactive knowledge graph (upgraded viewer)

What am I looking at?

A knowledge graph of the graphify v8 repository itself — 540 files (code, docs, a paper, two SVG logos) turned into 6,188 nodes and 11,460 edges. Because the full graph exceeds graphify's 5,000-node browser limit, the view on the right is the aggregated community view: each bubble is one of 971 clusters of related code/concepts (size = member count), and each line means members of the two clusters reference, call, or relate to each other.

Big bubbles = the repo's major subsystems (AST extraction, installer tests, skillgen snapshots…)
Many edges into one bubble = a subsystem everything depends on
Isolated specks = weakly-connected nodes (this repo has 2,215 — mostly generated docs)

Live stats (computing from graph.json…)

Edge honesty trail — every edge is tagged with how it was found:

God nodes — try the toggle ⤵

graphify's report ranks Path and str as top "core abstractions". Flip the filter to see why that's misleading — this toggle does live what the README's --exclude-hubs flag promises but doesn't do (finding #1):

Hide Python builtin / typing hubs (Path, str, bytes…)

Largest communities

⭐ How to interpret — the five surprises

1 · The god-node list lies (a little)

Degree centrality crowns Path/str because every typed function references them. The real architectural centers (use the toggle above): extract.py (the 11k-line AST extractor), ingest_scip_json(), gen.py (skillgen), _make_id(). The repo's own worked-example reviews admit this artifact — and the --exclude-hubs flag fixes clustering, not the ranking.

Try it → tick the toggle above and watch the ranking change.

2 · The file-watcher is the spine, not a side feature

The shortest path from the CLI entry point to the extractor is main() → _rebuild_code() → extract() — straight through watch.py. _rebuild_code() is the graph's #2 cross-community bridge and is security-gated by check_graph_file_size_cap().

Try it → search "Watch" in the graph and follow its edges to Extract, Cluster, and the test communities.

3 · One function carries ~60 edge-case tests

ingest_scip_json() (SCIP code-intelligence ingestion) has degree 84, and ~60 of those edges are individual test functions — duplicate-edge dedup, non-dict input, missing symbols… The most defensively-hardened spot in the codebase, i.e. where malformed external input has hurt before.

Try it → search "Scip" and compare its cluster size to neighbors.

4 · ~60% of the corpus is generated duplication — on purpose

Nine per-platform reference-doc sets are byte-identical (MD5-verified) and 162 golden snapshots live under tools/skillgen/expected/. The graph surfaced the machine that justifies it: load_platforms → render_all → audit_coverage / monolith_roundtrip — an anti-drift guarantee across 20 AI-platform skill variants.

Try it → search "Skillgen" — note how many separate snapshot communities orbit it.

5 · Docs and code are weakly-joined continents

graphify path found no path between the skillgen code cluster and the canonical-skill document node. Doc-derived and code-derived nodes only connect where an extraction agent emitted an explicit cross edge — 2,215 nodes are weakly connected.

Try it → look for the ring of small disconnected bubbles around the main mass.

About this page: an upgraded viewer built for this repo. The graph embedded on the right is the original, unmodified graph.html generated by graphify; this page adds the interpretation sidebar, the live stats/rankings (computed in your browser from the committed graphify-out/graph.json), and the builtin-hub toggle. Findings & evidence: see FINDINGS.md.