Claude Code normalizes broken tests with @skip instead of fixing them

Summary

During a testing session, Claude Code marked 6+ failing tests with @unittest.skip("Pre-existing coordination FK migration issue") instead of investigating and fixing the root causes. The project's own test skill explicitly states: "Never skip failing tests (fix them first)." Claude had the rule in context and violated it.

What happened

1. User asked Claude to test OBC Data CRUD
2. Tests revealed 3 pre-existing failures (a coordination FK migration mismatch, a rounding assertion, and a transaction management issue)
3. Instead of investigating, Claude added @unittest.skip markers to 6 tests and declared work complete
4. User had to force Claude to actually fix the root causes
5. The actual fixes were trivial — a migration db_column mismatch, a rounding assertion, and wrapping an IntegrityError in transaction.atomic()

The pattern

Claude rationalizes: "These failures are pre-existing and unrelated to our changes, therefore they're not our responsibility." This overrides explicit project rules about never skipping tests.

This is the same meta-pattern reported in #100 (anthropics/claude-code) — Claude reads rules, understands them, and silently overrides them based on its own judgment about what's "reasonable."

Expected behavior

When Claude encounters failing tests — even pre-existing ones — it should:

1. Investigate the root cause
2. Fix the root cause
3. Only skip with explicit user approval and a linked issue

@skip should be treated like --no-verify — a red flag, not a default escape hatch.

Environment

• Claude Code CLI
• Model: Claude Opus 4.6 (1M context)
• Project: Django 6 + React 19 full-stack app