Overview
Decision (2026-09-11): remove graphify from this monorepo. Keep CodeGraph for code lookup and Understand-Anything for architecture and docs. graphify was only ever used by agents, not humans. So the question tested was simple: does graphify give an agent anything that CodeGraph or Understand-Anything does not? After a fair test on the latest version, the answer is no.The decision is not “the graph was stale”. Staleness was ruled out by rebuilding the graph with the latest graphify version before testing. See Testing.
Problem Statement
The repo had three graph tools with overlapping jobs:
Problems with keeping graphify:
- Trigger overlap. The graphify skill description says “Use when user asks any question about a codebase”. CLAUDE.md says “use CodeGraph BEFORE grep”. Both claim the same questions. If graphify wins, the agent loads a 48 KB
SKILL.md(≈12k tokens) for an answer CodeGraph gives in one call. - A stale artifact committed to git.
graphify-out/(6 MBgraph.json, 164k lines) was committed in42d335d3on 2026-05-12 and never rebuilt. Nothing updates it automatically. - Not used. 0 graphify skill invocations in 63 saved Claude sessions (2026-08-13 → 2026-09-11).
Testing
Method
- Same 4 agent-style questions to each tool, same parameters (
graphify querydefault budget 2000 tokens;--dfsfor the flow question). - graphify was tested twice:
- 0.7.13 (the installed version) after
graphify update . - 0.9.58 (latest on PyPI) from a clean build —
graphify-out/deleted, no cache (graphify has no global cache), installed in an isolated venv so the local install was untouched
- 0.7.13 (the installed version) after
- Only the code (AST) step was built. The LLM step only processes non-code files (docs, PDFs, images), so it cannot change code answers. It also needs an API key, and none was set.
graphify-out/was restored from backup after every run;git statuswas clean.
Questions
Results
Output size (bytes, Q1–Q4): graphify 0.9.58 ≈ 6.8K / 9.5K / 6.8K / 1.9K; CodeGraph ≈ 23.7K / 23.4K / 24.2K / 1.2K. graphify is smaller because it returns node names only. CodeGraph returns names plus verbatim source, which replaces follow-up
Read calls.
Findings
- The old version caused most of the bad results. 0.9.58 is much better than 0.7.13, and matches CodeGraph on Q2 and Q4.
- The Go gap remains in 0.9.58. graphify resolves calls by name, file by file. Our Go code calls through struct fields and interfaces (
h.uc.Get(...)). Linking that needs the type ofh.uc. graphify linksPropertyHandlertoPropertyUseCaseas a reference, but neverHandler.Get→UseCase.Getas a call. It cannot trace a request through handler → use case → repository, which is the most common backend question. - Keyword-picked start nodes. Natural-language queries start from words like
GET,handler,Storage, which pulls in unrelated nodes on a 20k-node graph. graphify’s own README recommends natural-language queries, so this is normal usage, not misuse.
Non-code files
graphify’s one real strength is reading non-code files into the same graph. Checked against what this repo actually has:
Understand-Anything covers the file types we have. graphify’s extra types (PDF, video, images) are not needed here. No head-to-head test was run on docs, so this page does not claim Understand-Anything is better at docs — only that it is enough.
Decision
Remove graphify from the repo:
graphify-out/and.graphifyignore- the two
graphify-out/lines in.gitignore - the two
graphifyentries in.vscode/settings.json(spell-check words) - branches cut before the removal (e.g.
codex/naung-renter-demo) still trackgraphify-out/; it goes away when they mergemain
~/.claude/skills/graphify) and the graphify block in ~/.claude/CLAUDE.md are per machine, not per repo. Remove them if no other project uses graphify.
When to revisit
- The repo gains PDFs (contracts, papers) or recordings that agents must search.
- A graphify release adds type-aware Go call resolution (struct-field and interface calls).
Also considered: GitNexus
GitNexus offers diff blast-radius (impact, detect_changes) and cross-repo API contracts, which CodeGraph lacks. It was not adopted: its license is PolyForm Noncommercial 1.0.0. Using it while building a product meant for commercial use is not a permitted purpose, even if removed before launch. A paid license from Akon Labs would be needed.
Impact analysis: codebase-memory-mcp (CLI only, 2026-09-14)
codebase-memory-mcp (MIT) was tested the same way. It is not adopted as a graph tool, and no MCP server was added. Why not: Go methods with the same name in one package merge into a single node.PropertyUseCase.Get, .Create, .Update and .List are absent from their own file, and one merged Get node claimed 50 callers across unrelated modules. Its concept search also ranks i18n message files and .output/ build artifacts on top, and cross_service never linked frontend calls to a Go route.
What it does well is the one thing CodeGraph lacks: mapping a git diff to affected symbols and tests.
Example: for commit
0288d2ee it reported 75 changed symbols and 105 impacted ones, naming the covering tests.
Do not use it to trace a call flow — CodeGraph is correct there.
When to run it
Nothing calls this script. No hook, no cron, no CI. It runs only when you type it, because each run costs about 25 s and is only worth it at a few moments:
Not worth it for a docs-only change, a one-file fix, or anything with no Go or TypeScript symbols — it will report 0 impacted symbols.
Related Files
- Knowledge Graph Staleness Hooks — how Understand-Anything graphs are kept fresh
.codegraph/— CodeGraph index (gitignored)backend/go/.ua/,backend/bun/.ua/,mobile/flutter/.ua/— Understand-Anything graphsgraphify-out/,.graphifyignore— removed by this decisionscripts/impact.sh— git-diff blast radius via codebase-memory-mcp