grinding-until-pass（研磨直至通过）

Grind Until Pass

Use this skill when you want the agent to keep working autonomously until a specific goal is met — all tests pass, the build succeeds, or linting is clean. Instead of stopping after one attempt, the agent loops until done.

Steps

1. Define the goal command — the command whose exit code determines success:

- Tests: npm test or npx vitest run - Build: npm run build - Lint: npm run lint - Type-check: npx tsc --noEmit - All of the above: npm run lint && npx tsc --noEmit && npm test && npm run build

1. Run the command — execute it and capture the output.
2. If it fails — analyze and fix:

- Read the error output carefully. - Identify the root cause: failing test assertion, type error, lint violation, import error, etc. - Make the minimal fix. Don't refactor — just fix the error. - Go back to step 2.

1. If it passes — stop and report:

- Report what was fixed and how many iterations it took. - Summarize the changes made.

Rules for the Loop

• Maximum 10 iterations — if after 10 attempts the command still fails, stop and report what's blocking progress. Something fundamental is wrong and needs human input.
• Fix one thing at a time — don't try to fix all errors at once. Fix the first error, re-run, and see if the fix resolves downstream errors too.
• Don't delete tests — if a test is failing, fix the code to make it pass. Don't modify the test unless the test itself is clearly wrong (testing old behavior that was intentionally changed).
• Don't suppress errors — don't add @ts-ignore, eslint-disable, or any types to silence errors. Fix the actual problem.
• Track progress — if the number of errors is increasing instead of decreasing, stop and reassess the approach.

When to Use This

• After a large refactor that broke multiple tests
• After upgrading a dependency that introduced type errors
• After merging a branch with conflicts that need resolution
• When you want to "just make it green" and trust the agent to grind through it

Advanced: Cursor Hooks Integration

You can automate this with a Cursor hook in .cursor/hooks.json that triggers after the agent's turn ends, checks if tests pass, and sends a follow-up message if they don't:

{
  "hooks": [
    {
      "event": "stop",
      "command": "bash .cursor/scripts/check-tests.sh",
      "description": "Re-run tests after agent stops and send follow-up if failing"
    }
  ]
}

The script checks the exit code and returns a followup_message if tests are still failing.

Notes

• This works best with fast test suites. If your tests take 5+ minutes, the loop will be slow.
• Use --bail or --fail-fast flags to stop at the first failure for faster iteration.
• The agent will be thorough but not creative — if the fix requires a design change, it'll need human guidance.