Solution · Project 2 — Make the Tests Pass, Then Stop

Reference solution. Read the project first.

Worked example

Target: a FastAPI service, pytest suite of 340 tests, one broken test after a dependency bump changed a date-parsing library's default timezone handling — tests/test_reports.py::test_weekly_report_boundary fails because the report's week boundary is now computed in UTC instead of the service's configured local timezone.

fix-state.md after the run:

# Spine: fix-loop (Project 2)

**Target test:** tests/test_reports.py::test_weekly_report_boundary
**Attempts:** 2 / 5
**Last error (attempt 1):** AssertionError: expected 2026-07-20, got 2026-07-21
**Attempt 2 result:** PASS — pytest exit 0, full suite (340/340)

Body diff (attempt 2, the one that passed): exactly one file, app/reports/scheduling.py — the boundary computation now converts to the configured timezone before truncating to the week start, matching the pre-bump behavior. No test files, no fixtures, no unrelated modules touched.

Checker invocation — a separate run, never the maker's own read of the diff:

$ pytest -q
340 passed in 18.42s
exit code: 0

The loop's stop was gated on this exit code, not on the model summarizing "this should fix it."

Why this satisfies "done when"

• Attempt 1's fix (correcting the assertion in the test itself instead of the implementation) was rejected — it made the suite pass by weakening the check, which isn't what "fix the failing test" means. The checker still said pass, but a human reviewing the diff caught that the test's boundary shifted, not the bug — this is why a checker script alone isn't sufficient for semantic correctness, only for "does it build/pass"; a human still reads the accepted diff before it ships.
• Attempt count is truthful (2, not "eventually") and stayed under the 5-attempt cap.

Common wrong turns

• Marking success from the model's own "the fix looks right" — the exact failure mode Step 11 warns against; always re-run the real suite as the checker.
• Editing the assertion to match broken behavior instead of fixing the behavior — passes the checker, fails the actual intent. A green suite is necessary, not sufficient.