Fix your first bug

This walkthrough uses the pagination task from aop-mode’s evaluation suite. The first page should return the first two items, but the starting implementation skips them. You will run one repair and inspect the evidence it produces.

You need this checkout installed with pnpm install --frozen-lockfile, plus an authenticated Codex CLI. The live run consumes provider usage. For installation in your own project, follow the installation guide.

Understand the failure

The task starts with this function in src/page.ts:

export function page<T>(items: readonly T[], number: number, size: number): T[] {
  if (!Number.isInteger(number) || number < 1 || !Number.isInteger(size) || size < 1) {
    throw new RangeError('positive integers required');
  }
  return items.slice(number * size, number * size + size);
}

For page(['a', 'b', 'c'], 1, 2), the slice starts at index 2. It returns ['c']; the expected result is ['a', 'b']. Page numbers start at one, while array indices start at zero.

Run the repair task

From the aop-mode checkout, run:

pnpm test:harnesses --harness codex --task repair --timeout 180

The runner creates a temporary project containing the broken function and a test, installs the skills there, and invokes Codex. You do not need to copy the fixture into this repository.

The task asks the agent to reproduce the failure, preserve the existing test and input validation, repair the function, and run the test again.

Two separate checks. The agent runs the task's test. After the agent finishes, the runner independently checks the resulting behavior and whether the original test stayed unchanged.

Diagram text

flowchart TD
  A[Prepare broken pagination task] --> B[Agent reproduces and repairs]
  B --> C[Agent runs the existing test]
  C --> D[Runner checks the outcome]
  D --> E[Record result and evidence]

Inspect the result

The reference repair changes the slice to:

return items.slice((number - 1) * size, number * size);

This is the expected solution, not a transcript of your run. Inspect the run’s artifacts under .eval-artifacts/. Each task’s result.json records its candidate workspace; open that workspace to review the actual diff.

Check that the original test is unchanged, the first page returns ['a', 'b'], and the agent records both the failing and passing checks. Independent grading also checks page boundaries, empty and partial pages, invalid arguments, and input mutation.

A passed result means the run satisfied the task’s checks. A failed result needs investigation. A blocked result, such as an authentication error, means the run did not establish a successful repair. Read the evaluation guide for artifact details and limitations.

Use the workflow in your project

After installing the skills, start a fresh coding-agent session in your project. Give it a bounded request, naming the behavior and checks:

$aop-mode Fix the first page skipping items. Reproduce the failure before
editing. Preserve validation and the public API. Keep existing tests intact,
and show the checks you ran before and after the change.

Use /aop-mode in Claude Code or interactive Copilot CLI. Your normal session does not include the evaluation suite’s independent grader: review the diff and run your project’s checks yourself.

Next, read how aop-mode works or review a change without editing files.