git-better

---

Why

Your AI coding agent is a superb pair-programmer, but left alone it optimizes for getting the change in, not for the conventions that make a change safe to keep. You end up re-typing the same review feedback every session: split that file, don't swallow that error, that test is all mocks, document the export, don't push that unreviewed.

toolu moves those rules out of your head and into the tool:

• Hooks enforce on every edit — a post-edit quality gate checks each file the agent touches and blocks the session from moving on while any error, warning, or test failure exists — even in unrelated files.
• Skills enforce a process — an opinionated 8-phase workflow with a write/review checkpoint at every step, so design happens before code and review happens before "done."
• A registry keeps it modular — drop in a domain plugin (Rust rules, TypeScript rules, structural search) and its hook modules register themselves into the core engine, fail-closed, with zero wiring.

It's a personal bundle, built in the open, MIT-licensed. Take the whole thing or lift the pieces you like — on Claude Code or pi.

The quality gate

The headline feature. When rust-quality and/or ts-quality is installed, every Rust/TypeScript file the agent edits is checked on the spot. Limits are config-driven (project/user override → the active native linter's max-lines → built-in default), and the gate is multi-slot: a failing test command and a failing file check are tracked independently, so fixing one never silently masks the other.

TypeScript	Rust
File / function line limits No ../ relative imports — use the @/ alias No as type assertions — use a type guard No hand-rolled type guards — use a Zod schema Tests colocated in a flat __tests__/ Duplicate-type detection across the tree "Does too much" / too-many-factories heuristics	File / function / impl line limits No .unwrap() / .expect() — use ? or match No unsafe blocks No #[allow] / #[expect] lint suppression Tests in tests/, never inline #[cfg(test)] Flat tests/ layout enforced

The rule isn't "warn and move on" — it's a hard gate: no new task while the gate is red. Found a real problem? Fix it in code. (There's no "disable this check" escape hatch by design.)

Install

toolu runs on two hosts — install it the way that matches yours.

Claude Code

Install from the public marketplace in any Claude Code session:

# 1. Add the upstream marketplaces the plugins depend on
/plugin marketplace add anthropics/claude-plugins-official
/plugin marketplace add JuliusBrussee/caveman

# 2. Add this marketplace and install the core bundle
/plugin marketplace add Falconiere/toolu
/plugin install toolu@toolu

Add the language gates, search, and docs tooling too:

/plugin install rust-quality@toolu   # Rust quality gates
/plugin install ts-quality@toolu     # TypeScript quality gates
/plugin install ast-grep@toolu       # structural code search & rewrite
/plugin install comemory@toolu       # persistent cross-session memory
/plugin install context7@toolu       # live library documentation lookup
/plugin install exa-search@toolu     # web / code / URL search + research