Replay the run that broke the customer's project

The ticket says: your agent wrecked my project around 4 PM. On most stacks, that opens an archaeology dig through logs.

Here you reconstruct the scene. The customer's container snapshots on a schedule; you take the 3:50 PM snapshot and POST /copy with source_snapshot into a scratch container. That's the customer's machine, ten minutes before the damage, running at its own URL. Then you replay: dispatch the same headless run: same prompt, same model: and watch it go wrong, live, in a copy, against the customer's real state.

The fix gets verified the same way: patch, replay, confirm, before you touch their actual container. The scratch copy dies afterward.