Jonas Fackelmeyer Project · Personal project
agentic-workflow
My rules for coding agents as an installable Claude Code plugin. A guard hook enforces what can be checked mechanically, instead of just asking the agent nicely.
- Organisation
- Personal project
- Role
- Author
- Period
- 10/2026
- Team
- solo
- Status
- Public on GitHub
- Stack
- Claude Code plugin · Node.js · Hooks · Codex CLI
Last updated:
Problem
Rules in an AGENTS.md are requests. An agent trying to get a test to pass still reaches for --no-verify or @ts-ignore.
Solution
A guard hook runs before every shell command and file change:
- Blocks: bypassed checks (
--no-verify,@ts-ignore,eslint-disable…), secrets in code and.envfiles. - Asks first: before production dependencies, lockfiles, CI and infrastructure, deletions, force-pushes, commits and pushes.
It also ships a subagent that delegates tasks to Codex in an isolated worktree, my global agent rules and a template for project rules.
How I checked it
The rules are pure functions with tests for both what they must catch and the false positives they must not produce; a commit message that merely mentions --no-verify is not blocked, for example. An end-to-end test with Claude Code shows the difference: with the plugin, git commit --no-verify is rejected; without it, the command runs. CI runs on Linux and macOS with Node 20 and 24 and caught two bugs that would otherwise have made the guard silently useless.
Limits
The guard is a guardrail, not a sandbox: it matches patterns, and an agent can get around it by indirect routes. It does not replace reading every diff.