Done-checks in CI without a terminal
CI and automationTeam that wants the same 'done' in CI as on laptops
On laptops a developer approves the done-checks once. In CI there's nobody to type 'yes'.
How do I approve an AI agent's done checks in CI when nobody is there to type yes?
Read the approval code with ArcRoom's arcroom checks command on a laptop, store it as a CI secret, and set ARCROOM_APPROVE_CHECKS in the job. If someone changes the checks in a pull request, the code changes and CI refuses to run them until a person approves again. Someone has to update the secret when the checks change on purpose.
Also asked as: Our agent's checks need approval but CI has no terminal. What do we do? · How do I use the same definition of done for agents in CI and on laptops? · Can someone sneak new agent checks into CI through a pull request?
Read the approval code with arcroom checks on a laptop.
- Done-checks
$ arcroom checks
Step by step
- Read the approval code with arcroom checks on a laptop.
- Store it as a CI secret and set ARCROOM_APPROVE_CHECKS in the job.
- If someone changes the checks in a pull request, the code changes and CI refuses to run them until a person approves again.
What the agent sees
$ arcroom checks
Checks in .arcroom.json: pnpm typecheck, pnpm test
Approve with: arcroom checks approve 4KD9-2PCompared with today's tools
Test steps in the workflow plus branch protection
The workflow runs typecheck and tests as their own steps after the agent, and branch protection stops a red run from merging.
- Setup
- 30 min
- Ongoing
- None once set up
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 2 min plus one CI secret
- Ongoing
- None
Workflow steps catch failures after the agent stops but don't make it keep working until green. If you only need to block bad merges, branch protection alone is enough.
Where it stops
Someone has to update the secret when the checks change on purpose.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.