Reproduce Then Fix

• Start from an observable failure: the exact command, its output, and the input that caused it
• Reduce that to a minimal, deterministic case before editing production code
• Turn the case into a test or a recorded command that fails for the same reason
• Change the code so that case passes, then run the surrounding suite
• If you cannot reproduce it, say so and stop; do not “fix” a failure you have not seen
• Keep the reproduction after the fix, so the same hole cannot close silently