SLOPSHOPPER

e4-static-ok

A clean module for comparison

newguardtoast
v0.1.0MITupdated 2026-10-12RobGruhl/claude-mod-learnings/experiments/e4-static-ok
A shopper browsing a rack in a slop shop
Preview · a replayed session in a sandbox
claude · ~/work/app · e4-static-ok
› fix the failing auth test and add an audit log call ╭────────────────────────────────────────────╮ │ e4-static-ok │ ⏺ Read(src/auth.ts) │ rm -rf build && git push --force origin │ ⎿ Read 6 lines │ main │ ⏺ Update(src/auth.ts) ╰────────────────────────────────────────────╯ ⎿ Added 2 lines, removed 1 line ╭──────────────╮ ⏺ Bash(bun test) │ e4-static-ok │ ⎿ 3 pass, 1 fail │ cat .env │ ╰──────────────╯ ● Done. refresh now rejects expired claims and logs an audit event. ✻ Worked for 42s · done 4:20 PM ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts
README

Hook Chain: a field guide to Claude Code mods

Read it on the web: robgruhl.github.io/claude-mod-learnings

Hook Chain: a conveyor of small robots inspecting, rewriting, stopping and decorating cards on their way to a brass engine

Claude Code mods are TypeScript or JavaScript plugins that run inside Claude Code and hook its own event chain: tool calls, prompts, model requests, subagents and what the interface draws. They shipped on 1 October 2026 in v2.1.287. This repo teaches them from an explain-like-I'm-five version up to policy engineering, backs the claims with experiments we ran, and includes a facilitated tutorial you run in Claude Code.

It's the sister site of Decision Tier (repo), a field guide to decision models. The two meet on Mods meet the cascade. This is an independent study, not an Anthropic project. Everything was checked against Claude Code 2.1.296 on 2026-10-12. The API is early access, so trust the declarations your own build writes over anything here.

What's here

PathWhat it is
site/pages/The page fragments: ELI5, 101, 201, 301, four architecture pages, three evidence pages and the tutorial page
site/build.py, site/template.htmlStandard-library build that wraps each fragment and writes docs/, which GitHub Pages serves. Run python3 site/build.py.
experiments/Small mods that each test one claim, run with claude plugin test (and one live headless probe). Results in experiments/RESULTS.md. Also cascade-guard, a tiered Bash guard.
tutorial/The Mod Skeleton, a facilitated tutorial. cd tutorial && claude and Claude becomes your facilitator: one module at a time, predict before each check, every check run for real, one commit per module. Modelled on walkingskeleton.org.
tutorial/.reference/skeleton-guard/The tutorial's finished mod, validated and tested on 2.1.296

The pages

The ten-line version

  1. Pick the lightest extension point that can do the job. Text goes in CLAUDE.md or a skill, tools come from MCP, and a script you already have goes in a settings hook. Reach for a mod when you need to draw, rewrite an event or run code on a command.
  2. A mod is a plugin whose hooks/hooks.json names one module exporting register(on, options). Every hook is ($, e, next).
  3. A hook has three moves. It can observe (return next(e)), rewrite (next({ ...e, … })) or answer without calling next, so nothing beneath it runs.
  4. Hooks run in a fixed chain: prepend (org policy), then user, append and builtin, then Claude Code itself. The result comes back up the same way.
  5. A hook that throws or runs past its 10-second budget is skipped. The call goes ahead. A guard must register .catch to fail closed. We confirmed both cases on 2.1.296.
  6. Mods aren't sandboxed. They act with your permissions. Run claude plugin validate on a mod before you install it and read its calls: line.
  7. Everything outside the module goes through $, which is why the loader can list a mod's calls. It refuses a module that hides $ in a destructure, an alias or a helper it can't follow.
  8. There are three lifetimes. Module variables last until a reload, $.state until /clear, and $.store across sessions. session.start doesn't fire again after /clear, so reload from the store in classic.SessionStart.
  9. In claude plugin test your test's hooks sit beneath the mod and stand in for Claude Code, so answer every event.
  10. A tool.call guard doesn't see @-mentioned files. Hook prompt.mention too.

Where the docs and our results differ (2.1.296)

  • setTimeout and setInterval exist in the hooks runtime, though the docs say there are no timer globals. Waiting on one counts against a hook's budget.
  • Two hooks on the same event with no matcher are refused, but the same matcher twice loads, and both run.
  • A wrong-shape answer reaches .catch as kind: 'throw'. If the handler also answers in the wrong shape, the guard fails open anyway.
  • Reserved plugin names go beyond the claude- prefix: anthropic-, anthropics- and cc-plugin- prefixes, plus exact names like claude-code.

Details and outputs are in experiments/RESULTS.md.

License

Text and illustrations © Rob Gruhl, CC BY 4.0. Code (the build, experiments, tutorial reference mod and code samples) is MIT; see LICENSE. Illustrations were made with gpt-image-2, and the site was built with Claude Code.

Source 1 files
hooks/register.ts 15 lines
1import type { EngineInterface, Register } from 'claude-code'
2
3async function note($: EngineInterface, text: string) {
4  await $.ui.toast(text)
5}
6
7export const register: Register = on => {
8  on('tool.call', { tool: 'Bash' }, async ($, e, next) => {
9    await note($, e.command)
10    const home = await $.env.get('HOME')
11    await $.store.set('last', { command: e.command, home })
12    return next(e)
13  })
14}
15