Evaluate Graphify as an opt-in code-comprehension aid for modernization/aidlc (brownfield understanding)

[devfloor9]

Summary

Evaluate bundling Graphify (safishamsi/graphify, MIT) as an optional code-comprehension aid for the modernization and aidlc plugins — distinct from the Knowledge Wiki grounding layer (#54).

Why this is separate from #54

#54 (Knowledge Wiki) is about pre-generation grounding over a narrow curated corpus (schemas/ontology/ + ADR/Spec) to prevent ontology drift. That is a correctness-axis layer and should keep a minimal first-party implementation (see the prior-art evaluation on #54).

This issue is the other value proposition Graphify surfaced: understanding an unfamiliar whole codebase. That is exactly the brownfield problem modernization (6R: assess → strategize → containerize → cutover) and aidlc construction already face — and where Graphify's whole-repo graph (god nodes, community detection, cross-file/cross-modal "surprise" edges, rationale_for edges) is a natural fit rather than an over-scoped one.

What Graphify offers here

• Multi-modal extraction: code (19 langs via Tree-sitter), Markdown, PDFs, diagrams — useful for legacy repos with mixed docs.
• Whole-repo map: "god nodes" (highest-degree concepts) and surprising cross-module edges to prioritize assessment.
• Reported ~71.5× token reduction vs naive context loading on a mixed corpus — relevant for large legacy estates.
• Ships as a slash command + optional MCP server; AST pass is local (no code leaves the machine), only semantic descriptions hit the configured model.

Open questions to resolve before adoption

1. Opt-in only. Must be a non-default, clearly-flagged optional skill — not wired into the core install path. OMA's "install a plugin → dual-axis" promise should not depend on a third-party graph builder.
2. Maturity / sustainability. Released 2026-04-03, single maintainer. Acceptable for an optional aid; would need a vendoring/pinning story before any deeper reliance.
3. Config-ownership overlap. Graphify installs its own PreToolUse hook + CLAUDE.md directives; OMA's harness owns a PreToolUse hook. PreToolUse arrays are additive, but the composition must be tested (both hooks fire, neither suppresses the other).
4. Dependency weight. NetworkX + Tree-sitter + optional vision models — confirm this is acceptable to bundle or whether it stays a documented external install.

Acceptance criteria (draft)

• Decision recorded (adopt as optional skill / document-only / decline) with rationale
• If adopted: lands as an opt-in skill in modernization (and/or aidlc), not the default path
• PreToolUse hook composition with the OMA harness verified (both fire; no suppression)
• Dependency + sustainability story documented (pin/vendor or external-install)

References

• graphify.net · safishamsi/graphify (MIT)
• feat: implement Knowledge Wiki retrieval layer (pre-generation grounding) #54 — Knowledge Wiki grounding layer (the correctness-axis use; this issue is the comprehension use)

Filed from the plugin-UX re-verification thread. OMA currently has zero in-repo references to Graphify.