"Done" means the checks pass
Solo buildersSolo developer tired of 'done' that isn't
The agent kept saying 'done' after editing code it had never run. The next session spent its first half hour fixing the type errors and failing tests the last one left behind.
How do I make Claude Code run typecheck and tests before it says it's done?
With ArcRoom, write the commands that must pass into .arcroom.json, such as typecheck and tests, then approve them once with arcroom checks approve. When the agent tries to stop, ArcRoom runs the checks. If one fails, the run continues with the failure shown to the agent. It is only as good as your checks: if the tests don't cover it, done still isn't done.
Also asked as: Why does my AI agent say done when the code doesn't even compile? · How do I stop an agent from finishing with failing tests? · Can I force an AI agent to pass my checks before it stops?
Write the commands that must pass into .arcroom.json, for example typecheck and tests.
- Done-checks
Done-check failed: pnpm typecheck
Step by step
- Write the commands that must pass into .arcroom.json, for example typecheck and tests.
- Approve them once with arcroom checks approve.
- When the agent tries to stop, ArcRoom runs the checks; if one fails, the run continues with the failure shown to the agent.
What the agent sees
Done-check failed: pnpm typecheck
src/billing/upgrade.ts(42): Property 'tier' does not exist…
Keep going until the checks pass.Compared with today's tools
Claude Code Stop hook (DIY)
Write a Stop hook that runs typecheck and tests and refuses to let the agent finish while they fail.
- Setup
- 1–2 hours per agent
- Ongoing
- keep one script per agent in sync
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 5 min (init + checks in .arcroom.json)
- Ongoing
- None
In one agent a hand-written hook does the same job; ArcRoom sets the same checks up for every supported agent from one .arcroom.json.
Where it stops
Only as good as your checks. If the tests don't cover it, 'done' still isn't done.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.