Claude Code false pass: `pytest tests; echo "EXIT=$?"` returns 0

The developer's setup is a single person building business software for small operators and handing development to Claude Code, a coding agent, with test results arriving in the AI's own words. The failure mode he hit was not a wrong test, but a wrong shell pattern. When he wanted to see the exit code he appended `echo "EXIT=$?"`, and when he wanted to trim output he piped it to `tail -n 1`. Both patterns let the last command in the chain own the overall exit code, so a pytest run that actually failed delivered a success signal to the agent, which then reported success. He reproduced the effect with Python 3.14 and pytest 9.1 on Windows Git Bash, and found that adding `set -o pipefail` before the pipe restores the failure signal, though he considers not transforming the output at all the safer route. Two other cases sit outside the success/failure split he had been using. `pytest` returns exit code 5 when it finds no tests, and the developer worries an AI reading only screen text would see no "failed" and call it clean. And a command that cannot start at all, such as npm launched from Python on Windows failing with FileNotFoundError, is neither a pass nor a fail. His rules therefore make reports carry the command and its exit code, treat a run that never started as a third category, and refuse to count a result taken before the code was changed.