Who it is forA developer with a test or lint command who wants an agent to keep fixing a change until it passes.
The problemAgents often stop after one attempt, and you have to rerun and re-explain each failure yourself.
The outcomeA loop that repeats each failed step with the failure reason until the check passes, or stops when a check reports blocked or the budget runs out.
What you need first
- A Mac. The release DMG is for Apple Silicon.
- Either a cloud model API key (for example Anthropic or OpenAI) or Ollama or LM Studio running on the Mac.
- A project folder to point the agent at.
- npm, only if you build from source instead of using the DMG.
Steps
- Open a session with the workspace folder set to your project.
- Create a loop with one step per change. Give each step a check: a shell command that must exit 0, such as your own test command.
- Set a spend budget. Long builds have no round cap, so the budget is the ceiling.
- Start the loop and watch the Activity panel as each failed step goes round again with its reason.
- If a check answers blocked, the loop ends at once. Review the reason before you restart it.
Prompt for your agent
Using the repo templetongroup/radiant, set up a loop in a Radiant session on <project folder> where each step must pass a shell test command that exits 0, and run it. Expected output: a loop report showing the rounds each step took and whether each check passed. Verify that the report lists the spend against the budget you set.
Caveat. Loop checks only prove what your command tests, so a weak test command will let a wrong change pass.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for templetongroup/radiant and "Run a build loop that retries until a test command passes"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer