Review an AI bug fix with boundary tests

This fictional function counts pages of search results. Assume totalItems is a nonnegative safe integer and pageSize is a positive safe integer. A partly filled page counts as one page; zero items needs zero pages. The caller validates inputs. Strings, negatives, zero page size and unsafe integers are outside this exercise, not silently supported.

You need basic JavaScript reading skills. Use only the code here in an AI chat you’re allowed to use. Never paste employer repositories, credentials, customer records or private bug reports. Do not connect a repository or grant an agent access for this exercise. Tool access and prices vary; no paid account is required for the manual path.

For 11 items at 10 per page, sketch one full page and one partial page. The expected count is 2. Read the current implementation:

function pageCount(totalItems, pageSize) {
  return Math.floor(totalItems / pageSize);
}

Math.floor rounds the quotient down: 11 / 10 becomes 1.1, then 1. That loses the last page. Keep this failing example before changing anything.

Copy the prompt below into your approved AI chat, with automatic execution and repository access off. Save its proposed patch and predicted tests. Treat both as unverified.

No AI access? Write your own candidate fix and test predictions, then continue. Describe that version as manual practice rather than AI-assisted work.

Write expected results before running the candidate: 0/10 → 0; 1/10 → 1; 10/10 → 1; 11/10 → 2; 20/10 → 2; 21/10 → 3. The two exact multiples help catch a plausible but wrong “add one page” patch.

Optional execution: in an empty local JavaScript file or a blank browser tab’s developer console, copy the original function and the test harness below. Review every line first. It has no packages, files or network calls. Do not run it in a work app’s console or execute unfamiliar AI-generated commands.

const cases = [
  [0, 10, 0], [1, 10, 1], [10, 10, 1],
  [11, 10, 2], [20, 10, 2], [21, 10, 3]
];
let failures = 0;
for (const [total, size, expected] of cases) {
  const actual = pageCount(total, size);
  const ok = actual === expected;
  if (!ok) failures++;
  console.log(
    (ok ? "PASS" : "FAIL") + " " + total + "/" + size +
    ": expected " + expected + ", got " + actual
  );
}
console.log(failures + " failing checks");

The original should report 3 failing checks: 1, 11 and 21 items. Then replace only the function with your reviewed candidate and rerun the same harness. Keep both outputs. If you cannot run code, trace each division and rounding step by hand and label the result “manually checked”, not “tests passed”.

function pageCount(totalItems, pageSize) {
  return Math.ceil(totalItems / pageSize);
}

Compare the patch against the contract line by line. A fix that changes expected values to match the bug, adds a special case only for 11 items, or brings in a dependency has not solved the stated problem cleanly.

These six cases establish their own results, not exhaustive correctness, security or production readiness. We did not test input validation, integration with a UI or large-value behavior. In a real project, confirm who validates inputs and run the relevant project tests and human review before merging.

GitHub’s responsible-use guidance treats AI review as assistance requiring human scrutiny. A confident explanation or “all tests pass” sentence is not an execution log. Separate predictions, manually checked results and results you actually ran.

Keep the synthetic contract, original bug, AI proposal if used, final patch, expected values and actual or manual results together. Write a short note: what failed, why the patch works for the contract, what you rejected and what remains untested.

Example to adapt only after doing it: “The original rounded partial pages down. I checked empty, exact-multiple and partial-page cases and rejected an unconditional +1 fix. The revised function matched all six expected values. Input validation and application integration remain outside this practice.” Add whether you executed the harness or traced it manually.