SKILL.md
Code Graph
Build, query, and incrementally update a persistent knowledge graph for the target project. Infrastructure — other skills consume it via query; Flow Conductor checks freshness on session start.
Two-tier extraction
| Tier | Tool | Role |
|---|---|---|
| Structural | graphify | AST nodes, edges, communities, confidence — deterministic |
| Semantic | Model | Purpose, depth, impact — contextual |
graphify parses; the model annotates. Never replace graphify structural data with grep guesses.
Prerequisites
[prerequisites.md](prerequisites.md) — install graphifyy via uv/pipx/pip. Non-blocking if user declines.
Storage
| Path | Source | Purpose |
|---|---|---|
graphify-out/graph.json |
graphify | Structural graph |
.scratch/graph/annotations.json |
Model | Semantic layer |
Do not copy graph.json into .scratch/. Schema: [graph-schema.md](graph-schema.md).
Operations
Detailed steps: [build-steps.md](build-steps.md).
/code-graph build— full scan + semantic annotation/code-graph query <subcommand>— read-only; patterns in [query-patterns.md](query-patterns.md)/code-graph update— incremental via graphify cache
Constraints
- graphify is the structural backbone — no grep-based dependency guessing
- Model adds semantic layer only; never override graphify edges
- Prefer
updateoverbuildwhen graph exists - Non-blocking on graphify failure — warn and continue
- Confidence-aware: EXTRACTED (1.0) vs INFERRED (0.5) vs AMBIGUOUS (0.2)
Flow Conductor
aiops-graph.js hook suggests /code-graph build when stale. Architecture health runs graph_build phase first (phases.py) — build when missing/stale, skip when fresh, user may decline (organic fallback in /improve-codebase-architecture).