Lesson 3: The bounded change, reviewed hunk by hunk
Objective
The starter passes all nine tests through a sequence of bounded changes, and
you can explain every diff the agent produced without reading the solution
first.
Why this lesson exists
This is the lesson people skip — and the one that builds the actual skill.
Letting a thread rewrite the file in one shot teaches you nothing and gives
you a diff you cannot honestly review. Supervising small diffs, each pinned
to a contract line, is how the code stays yours. The Codex app’s approval
model exists for exactly this: commands ask before they run, and edits land
as reviewable diffs.
The lesson
Start a fresh thread (a finished task’s thread carries stale context) and
say:
“Work through TASK.md against starter/report_tool.py. One failing test
group at a time: JSON loading, then validation isolation, then rounding,
then atomic writes, then the end-to-end tests. After each group, run
python verify.py starter and show me the diff before continuing.”
Hold it to the contract’s boundaries:
- Only
starter/report_tool.py changes. If a diff touches tests/,
solution/, or scenario/, stop and ask why.
- No new imports outside the standard library — and no import that is not
needed by the change being made.
- Each change should move toward one contract line. Reject drive-by refactors
(“while I was here I renamed…”).
- If the agent wants to change a test, the answer is no. Tests are the
contract; the code moves.
Use the approval prompts deliberately. When the agent proposes a command,
read it — python verify.py starter is safe; pip install anything is a
stop sign. When it edits a file, open the diff view and read every hunk
before accepting. A diff you cannot explain is a diff you reject.
Expect the failing count to drop group by group: 7 → 5 → 4 → 3 → 2 → 0.
Exercise
Pick one hunk the agent produced — ideally the validation-isolation change —
and explain it back in the thread: “line N does X; that satisfies contract
line Y.” If you cannot, ask the agent to walk you through its own diff before
accepting the next change.
Checkpoint
Run python verify.py. This checkpoint’s code appears only when the
starter suite is green. You pass the lesson when you can answer:
- How many turns did the bounded change take, and what made each turn bounded?
- Which approval prompt did you reject or narrow, and why?
- Point at the hunk that implements “isolate invalid rows with reasons” —
where is it?
Expected evidence
A green python verify.py starter run, and your hunk-by-hunk explanation of
the change that earned it.
