Reproduce it first

Goal#

Turn a bug report into a test that fails, fix the code, and watch the same test pass.

Why first#

A bug you cannot reproduce is a bug you cannot prove you fixed. Writing the test first gives you three things:

• Proof the bug exists, in a form anyone can run.
• Proof your fix works, because the same test turns green.
• A guard forever, so the bug cannot quietly come back.

The test must fail before the fix. A test that passes on broken code tests nothing.

The scene of the crime#

uv run python -m detective run 2

2500 cents with HALF: pay 1250 cents
2500 cents with TENOFF: pay 1500 cents
680 cents with TENOFF: pay -320 cents

A customer bought a 6.80 coffee and cake, used the code TENOFF, and the till owed them 3.20.

Write the witness#

Create lab/cases/test_case02.py with a test that states what should happen:

from cases.case02_reproduce import apply_discount


def test_a_discount_never_makes_the_price_negative():
    assert apply_discount(680, "TENOFF") == 0

Now check, before fixing anything:

uv run python -m detective check 2

  TODO  a small order with TENOFF costs 0, not less
  PASS  TENOFF still takes 10 euros off a big order
  PASS  HALF still halves
  PASS  you wrote cases/test_case02.py
  PASS  your test fails on the original bug
  TODO  your test passes on your fix

The checker runs your test against the original code and confirms it fails. That is a real witness. Now fix apply_discount, and run the check again.

Make the smallest reproduction#

Real bug reports are long: "on Tuesday, after logging in with a voucher, the checkout page showed a negative total". The skill is cutting it down to the smallest input that still fails. Here, that is one function call with two arguments. Every detail you can remove is a detail you no longer have to think about.

Check it works#

• uv run python -m detective check 2 shows six PASS lines.
• Your test file is the first thing someone will read if this bug ever returns.

Exercise#

Add a second test for an edge you can think of: a total of exactly 1,000 cents, or no code at all. Does it pass? Should it?

More to try

• The lab already has pytest. Run your test with uv run pytest cases/test_case02.py before and after the fix, and read how pytest shows the failing values.
• Name your tests as sentences about behaviour, like the one above. Which is clearer six months from now: test_discount_2 or this one?