Tip: Receipts reruns an agent's new tests with the fix reverted, and found 10% of agent PRs had no test that exercised the old behavior

GitHub·low signalXBlueskyLinkedInCopy link

Receipts (created 09-25, Claude Code plugin, GitHub Action and CLI) runs each changed pytest, vitest or jest test twice: once with the change, once with only the source files reverted. Each test gets a label: PROVEN, THEATER (passes both ways) or WEAK. Across 181 real changes, 82% of agent PRs were proven, against 90% of maintainer fixes. In 10% of agent PRs every test failed on old code only because it imported a new name. Its prove-fix skill makes Claude rewrite the test, not the fix, until it is proven.