Claude Function Hooks — Development skill for Claude Code

claude-function-hooks

A reference implementation of the **Function Hooks** algebra proposed for Claude Code in [anthropics/claude-code#91870](https://github.com/anthropics/claude-code/issues/91870), built from the proposal's architecture document so the design could be argued with empirically instead of from taste.

git clone https://github.com/Monte9/claude-function-hooks && cd claude-function-hooks
npm test            # 17 tests, each citing the spec section it checks
npm run demo:redact # redaction: the model never sees PII, the human still does
npm run demo:rules  # a real CLAUDE.md's non-negotiables, enforced

Node ≥ 20, and nothing else: the engine, the tests, the examples and four of the five findings run on a bare clone with no `npm install`. The fifth (`finding:await`) asks TypeScript a question, so it needs the one devDependency — `npm install`, then `npm run finding:await`.

What this is, and is not

**Not an Anthropic project, and not the feature.** Function Hooks is a *proposal*; it does not exist in Claude Code. The design is Alice Poteat's, from *"Function Hooks: Core Architecture"* (Anthropic, August 2026), attached here as [`docs/EXTERNAL.Function.Hooks.Core.Architecture.pdf`](docs/EXTERNAL.Function.Hooks.Core.Architecture.pdf) — the PDF linked from the issue. `docs/architecture-extracted.txt` is its text, extracted by decoding the PDF's font CMaps, so every `§` citation in this repo points at something you can check.

This repo is only the algebra, implemented from that spec by a third party, plus tests and two worked examples. Any mistake in it is mine, not the proposal's.

The proposal in a paragraph

Claude Code's `hooks.json` supports four hook *transports* today — `command`, `prompt`, `agent`, `http` — each a contract with an **environment** (a shell script reading JSON on stdin, signalling by exit code). The proposal adds a fifth: a TypeScript module, a contract with the **engine**.

export function register(on) {