repo-dependency-graph

Turn a skill repo (or a whole campaign package) into a clean, interactive use-dependency graph: read the original SKILL.md + design docs, reconstruct the real use edges, and render one offline HTML per repo that looks and behaves like an Obsidian graph (force-directed, draggable, neighbourhood-highlight, HTML hover tooltips).

This skill exists because a repo's frontmatter alone gives a broken graph — real routing lives in prose, in references/*-index.md, and in same-layer escalation handoffs. Reconstructing the true graph takes a careful read of each repo, then a deterministic render. This skill captures both halves so the result is consistent every time and reusable across all repos.

The model (read references/layer-rules.md for the full rules)

• 5 vertex types: campaign (red) · strategy (cyan) · tactic (yellow) · sop (purple) · references (gray dashed, = a .py/.md helper file).
• One edge type use: A -->|use| B = A invokes/orchestrates B (caller → callee).
• Layer comes from frontmatter type: / layer: (NOT execution:). Infer from body+README only if absent, and record the reasoning.
• Escalation = same-layer sop → sop use edge (locked decision): "escalate to X" / "for deeper analysis use X" / "import X" are all drawn as use edges. Do NOT promote a skill's layer just to make the edge look legal — same-layer handoff is a first-class edge here.

Workflow

1. Read the repo and reconstruct the dependencies

For each skills/*/SKILL.md: read the full body + frontmatter (layer field, and any prose that invokes/escalates to another skill or points at a references/ helper). Also read README.md, docs/, assets/, and any *-index.md — real routing often lives there, not in frontmatter.

Only draw an edge the design files actually justify. If two skills are independent siblings, leave them unconnected — do not invent edges. A pointer to an external MCP tool (alphaxiv, brave) is a tool, not a vertex. A broken / never-used file pointer does not justify a references vertex — verify the live reference.

For a large package (50+ skills), dispatch one subagent per package to read its skills/ + docs/ and return the reconstructed nodes/edges — they are independent and parallelize cleanly. Give each subagent references/layer-rules.md and references/graph-schema.md so its output is consistent and directly usable.

2. Write the graph JSON

Emit one JSON file per repo following references/graph-schema.md:

{
  "name": "<repo-name>",
  "nodes": [ {"id": "<full-skill-name>", "layer": "sop", "desc": "<bilingual hover>"} ],
  "edges": [ {"from": "...", "to": "...", "tip": "<trigger scenario hover>"} ]
}

• Node IDs are full skill names — never abbreviate.
• desc (中英混合) = what the SKILL does + its depth / HARD-GATE. HTML allowed.
• tip = the condition under which the caller hands off to the callee; quote the source skill's own escalation language where possible. HTML allowed.
• Both desc and tip are optional — omit and the tooltip falls back to name [layer] / use.

3. Render the HTML

python scripts/render_graph.py --data <repo>.json --out <repo>.html
# or batch a whole directory of graph JSONs:
python scripts/render_graph.py --data-dir ./data --out-dir ./graphs

The script prints a [warn] for any edge that breaks layer legality (sop → sop escalation never warns). It produces a fully offline HTML (vis-network inlined, external CDN stripped) with:

• 5 vertex-type colors, dashed border for references.
• Force-directed, draggable, scroll-zoom, click-to-highlight neighbourhood.
• A self-managed #dare-tip hover layer that renders the HTML in desc/tip (vis's native tooltip shows a string title as plain text, leaking <b> tags — the script bypasses it with an event-driven layer; this is why HTML renders).
• No navigation buttons, no bootstrap card frame — just the graph on a dark canvas.

4. Verify

Open the HTML in a browser and confirm: nodes colored by layer, edges arrow caller→callee, hovering a node/edge shows the bilingual tooltip with HTML rendered (bold + <hr> divider, not raw tags). Requires pyvis (>=0.3.2).

Key constraints

• HTML directly, no intermediate mermaid .md. The HTML is the deliverable.
• Full skill names as node IDs and labels — no abbreviations.
• Privacy: the HTML embeds only node names + layer + the desc/tip you write. Never emit filesystem paths, log paths, or session IDs into the graph data or output.
• Read-only on the source repo — reconstructing a graph never edits skills.

See examples/literature-engine.json for a complete, working graph JSON.