---
title: QA with Codex: the skill in skills/kaloko and a pointer in AGENTS.md
description: How Codex runs Kaloko: where the skill lives, how AGENTS.md points to it, the prompts that work, and how Codex reads reviewer feedback before it continues.
seo_title: QA with Codex: the skill and a pointer in AGENTS.md
query: QA for OpenAI Codex
kind: agent
order: 11
demo: true
---
Codex reads `AGENTS.md` and follows it. `kaloko init --agent codex` puts the skill into `skills/kaloko` and adds one paragraph to `AGENTS.md`, so acceptance walkthroughs become part of how the agent works in the repository.

## The problem

Codex is fast at code and silent about screens. It will happily merge a form that overflows on phones, because nothing asked it to look.

## What changes with Kaloko

`AGENTS.md` says: for acceptance plans, scenario runs and the canvas use `skills/kaloko/SKILL.md`. The skill says what done means, what is allowed without asking and what is not, and links the references for writing the plan, executing steps, reading evaluations and handling environments.

## How it works

1. **Install**: `npm install --save-dev kaloko`, `npx kaloko init --agent codex --org <your-org>`, `npx kaloko doctor`. The skill lands in `skills/kaloko`, `AGENTS.md` gets the pointer, `kaloko.config.yml` is created.
2. **Ask**: the same prompts as for any agent; Codex runs the CLI commands from the skill in its sandbox with network access to your local app and to the Kaloko service.
3. **Steps Codex cannot script**: mark them `agent` in the scenario; Codex follows the instructions, then `kaloko capture --step <id>` records what it reached.
4. **Feedback**: `kaloko feedback` and `kaloko inbox` in the terminal, or the MCP server as an HTTP tool source, before the next round.

## Skills and prompts

```
Follow skills/kaloko/SKILL.md: walk the acceptance plan of TASK-123 on local, evaluate, build and share it, then post the canvas link into the PR.
```

```
Turn this task description into ACCEPTANCE.md and qa/scenario.yml; one criterion per check, semantic questions only for judgement.
```

Expected outcome: the files in the task folder, a validated scenario (`kaloko validate`), a shared run and one PR comment.

## What you get

- Repository-level rules Codex respects every time, not a prompt someone has to remember.
- A read-only production policy the CLI enforces even if the agent tries otherwise.
- The same canvas, verdicts and inbox as with any other agent.

## FAQ

### Sandbox and network?

The CLI talks to your local app, to the shared Chrome over CDP and to `kaloko.app` over HTTPS. Allow those in the Codex sandbox; nothing else is needed.

### Can Codex and Claude Code share a repository?

Yes: `kaloko init --agent all` installs both locations; runs are pinned per agent with `--run <id>`.
