Build, validate, stop
The agent builds, checks its own work against written rules, lists what it touched and stops. Here's what that check can and can't catch.
1Guess first
The agent has finished a build. The last lines of its reply:
Does that mean the report opens and looks right in Desktop?
The is the agent's own checklist, run after every build. It's part of the : the tools and checks wrapped around the model so it can act, and be caught.
2Break it on purpose
Break the build on purpose, then run the validator. Catch three faults. One of them it will never catch.
3Every bug becomes a check
The check count didn't stay at 710. It grew to 1,533. Almost every new check started as a real bug.
Clicking a tile brought it to the front and hid its button. Grouping each tile with its button fixed it. Which check was added?
That's the pattern: you find the problem in Desktop, the fix goes into the build, and a check goes into the validator so the fix sticks.
4The file list
The build ends with a list of every file it rewrote. You expected 52. It lists 60.
Ignoring whitespace, only 52 differ. What are the other 8?