The RADIUS Kit — Stop Letting Your AI Read the Whole Repo: the 5-Minute code-review-graph Setup (26.5k-Star Tool, Live-Tested)

The companion to the full video. code-review-graph is the 26,500-star, MIT-licensed, local-first tool that parses your codebase into a graph with Tree-sitter — files, classes, functions and tests as nodes; calls, imports, inheritance and test coverage as edges — stores it in one SQLite file inside your repo, and serves it to your AI coding tools over MCP so they read only what a change can actually touch. We live-tested it on Flask for the video: 94 files became 1,446 nodes and 8,337 edges in about 2.4 seconds, and for a real 3-function change the review context dropped from 25,578 tokens to 643 (~97%), cross-checked with OpenAI's own tokenizer at 22,183 → 732. INSIDE THE KIT: (1) SETUP ON ONE PAGE — the three commands (pip install, install, build), what the auto-installer wires up across 14+ platforms including Claude Code, Cursor, Codex, Gemini CLI and Copilot, and the shallow-clone gotcha that silently blanks change detection until you fetch history. (2) THE DAILY WORKFLOW — detect-changes --brief before every review (read-only, ~1 s, prints the Token Savings panel with your own numbers), update --brief after a rebase, and the --verify flag that recounts everything with the cl100k tokenizer. (3) BLAST RADIUS — on our deliberate break of Flask's make_response(): 41 nodes directly changed, ~500 impacted within two hops, 59 other files affected, plus 3 test gaps it flagged unasked. (4) THE GITHUB ACTION — the exact 14-line workflow YAML for risk-scored sticky PR comments with an optional fail-on-risk merge gate, built entirely on your own runner. (5) THE HONEST LIMITS, quoted from the project's own README — why its 'recall 1.0' is a circular upper bound, why tiny single-file edits can cost MORE through the graph than reading the file, search MRR 0.35, and the deliberate over-flagging trade-off. (6) BONUS TOOLS — dead-code detection (97 leads found in Flask on first pass), 13 auto-named communities, 77 traced execution flows, the in-browser D3 graph, and exports to Obsidian, Neo4j, GraphML and SVG. Our 97% figure is our run on our change — the kit shows you how to print yours. Independent field kit from Hyperautomation Labs — not affiliated with the code-review-graph project.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.