One task, two modes of the same harness

The exercise uses one harness in two modes. Interactive Claude Code builds the fixtures and the schema; headless claude -p does the task with nothing but Read, and its JSON result shows the harness's work: turns, a session, usage and validated output. The raw commands are enough to do the exercise. The small TypeScript wrapper here does nothing new: it saves each result to out/ and prints the fields you are asked to find, and lets a stub stand in for claude so you can dry-run the same commands without a model. The decision that matters most is limiting the headless run to Read, with both --tools and --allowedTools, so whatever the model does, it can only read.