agent-config

Installs into your user-level AI tool configuration; your repository only gets an agents/overrides/ folder for your own changes. Works with 20 AI coding tools .

Agent Config — teach your AI coding agents to check their work before they say done

Install once. Your AI coding tools are told to plan, test and verify before they say done; on Claude Code a turn that changed files without checking them can be stopped.

npx -y @event4u/agent-config init

Free and MIT-licensed. Node 22.6 or newer on macOS, Linux or WSL2. Costs more tokens: 7.8× a run with no rules on Claude Haiku as installed, 2.9× with only the core rules. What it costs →

On a desktop this opens a browser wizard: pick your role and your tools, click Finish. Nothing is written before you do.

Try it in 30 seconds

Not ready to install anything? Drop one read-only checker into your repo and ask it whether your branch is really done:

mkdir -p .claude/agents

curl -fsSL https://raw.githubusercontent.com/event4u-app/agent-config/main/docs/wedge/production-validator/production-validator.md \ -o .claude/agents/production-validator.md

then in Claude Code: @production-validator check this branch is actually done

production-validator reads the code that would ship, hunts mocks and placeholders on that path and asks for proof that it ran. It changes nothing. What it checks →

What changes

Before. You ask your agent to add a CSV export. It writes a handler that returns a hard-coded placeholder, the existing tests stay green and it replies "Done — export implemented." Nothing ever ran the new code.

After. In Claude Code you ask @production-validator whether the branch is really done. It reads the code that would ship, finds the placeholder and answers NOT READY with the exact file and line. On a branch that is really finished it answers READY and reports nothing. We tested this on two cases: one with a placeholder left in, one finished.

On Claude Code the package can also refuse to let a turn end on "done" when the agent changed files and ran nothing to check them. On every other tool that is an instruction the agent is asked to follow, not a block.

What your agent is asked to do

It changes what the agent is told to do, not how smart the model is. Re-measured in November 2026 after a scorer repair, the full install's discipline gain on Claude Haiku was not significant and Claude Sonnet showed no gain; only the small core rule set kept a significant gain on Haiku. We publish results like that one instead of hiding them. The whole story →

How it fits together

One install gives the agent you already use a set of skills, agent rules and slash commands. Rules that matter on every task are always on; the narrower ones load when a task needs them. See every slash command →

What it costs

• More tokens. On Claude Haiku, on two scope-discipline tasks, the install as shipped cost 7.8× a run with no rules; the core rules alone (kernel plus downstream-changes) cost 2.9×. The discipline_profile setting does not yet change what a Claude install loads.
• Blocking on Claude Code only and only while its hooks run. A host setting can switch them off and a user-installed mod can lift a refusal; agent-config doctor neighbours lists installed mods. Elsewhere the rules are instructions: they cut violations but do not remove them.
• Network. By default nothing leaves your machine except the daily update check against the npm registry and a daily question to GitHub, through your own gh login, whether a pull request you pushed has merged. The update check is off in CI, without a terminal or with AGENT_CONFIG_NO_UPDATE_CHECK=1. Telemetry is off. The second-opinion feature is off until you set it up; then it sends work to other AI providers with your own API keys, which bill you. What leaves your machine →
• Your machine. Node 22.6 or newer. Standing cost of the always-on rules per tool: table.
• Your way out. agent-config uninstall --global removes only its own entries and leaves other tools' config untouched.

Benchmark and cost, explained →

See what a change touches

agent-config code-graph map --open draws a local map of your repository: who calls what, which tests cover it, data, interfaces and delivery. Pick a file and every area it reaches lights up. It runs offline and keeps file:line for every fact.

Open the demo map → (sample project, not your code) · Code map →

Pick your profile

The wizard asks for your role and installs a focused set for it instead of the whole catalog. Most of them are not engineering roles.

Choose a profile →

Workflows

You start a piece of work, not a list of commands. These cover most days:

When a request is unclear the agent stops and asks with numbered options instead of guessing. Implement a change →

Use it in your project

Start with the Quickstart below; the full walk-through is on the site. Install →

Prove it

Every measured figure in this README carries a marker that points at a recorded, re-runnable piece of evidence. The build fails when one stops resolving. The proof page lists each claim with its evidence, including the runs where the package changed nothing, plus commands you run yourself on a fresh checkout. Verify it yourself →

Contribute

Skills, rules and commands are plain Markdown under src/; one command rebuilds every tool's copy and runs the checks. Development has the loop. Contributing guide →

Quickstart

npx -y @event4u/agent-config init                              # browser wizard on a desktop

npx -