grep -l 'destroyed\|damaged' results/*
Run details
Every failed and successful run, grouped by scenario.
Each entry starts with what the harness found when it inspected the server afterwards — the authoritative end-state check. On failed runs, the one-sentence summary below it was written by a summariser model and may be wrong. On successful (clean + done) runs, you can inspect the exact commands the careful model ran and its final report. Every command is copied verbatim from the transcript.