Get started
Install the check in your repository.
One command creates the team’s record next to the code. Another reads it again at every delivery, and says what is missing.
To see it first
Try it on an example
In an empty folder, with git initialised:
The command writes a complete example and runs the check: everything is green. Then rename the function it points to, run the tests (in code/, after npm install) and then the check: the tests pass, the check refuses and names the two lines of the record that cite it. The scene, on a public repository, shows the same gesture.
Before you start
What you need
- Node.js 22.12 or later, and git.
- The repository of an application.
- If you ship with a coding agent: the command creates its instructions, in an
AGENTS.mdfile.
Step 1
Install
At the root of your repository:
The command asks a few questions, then creates an AGENTS.md file, a runward/ folder and an .agents/skills/ folder for your agents. It overwrites no existing file. The runward/ folder is the team’s record: the team sets in it what it requires before production, and its agents draft most of it, with what proves it.
| The main files | What it holds |
|---|---|
runward/framing.md | The problem, the expected value, the success criterion. |
runward/architecture.md | The system’s boundaries. |
runward/adr/ | The structural decisions, written and dated. |
runward/floor.md | The first-version measurements. |
runward/governance/threat-model.md | The security assessment. |
runward/runbook.md | The recovery procedure. |
runward/handover.md | The handover note. |
AGENTS.md | The team’s instructions for your coding agents. |
Step 2
Run the check
The check says which document of the record is missing or not filled in. It exits with code 1 while a gap remains, and with code 0 when nothing is blocking.
With --strict, it also verifies that each piece of evidence cited in the record still exists: a file, a test, a written decision. When it refuses, it names the file and the line.
Step 3
Wire it into your delivery chain
Add the same command to your delivery chain, with the version you chose:
It fails while a gap remains. You decide whether that failure blocks the release to production.
With a coding agent
Your agents read the same record.
Codex, Cursor, Copilot and the other tools that read AGENTS.md find the team’s instructions there. Claude Code reads it too when the repository has no CLAUDE.md. For Gemini CLI, add its profile: npx runward init --tools gemini. Ask your agents to run runward check before they say they are done: the check answers them as it answers you.
Without installing anything in your repository
First see a refusal, on a public repository.
This command runs the check again on the commit of the scene of the home page. It refuses, and says which lines.
To go further
The documentation
Try it on an example, in one command.
Try it on an example