jev-slalom

# Jev Slalom course design Jev Slalom runs a live course: a line of gates sweeps down a valley and Jev (TypeSafe's System One model) reads its position and the next gate each tick and steers left/right to thread each gap. The agent shapes `slalom.json`: the descent speed, the valley width, and how many gates. ## The loop 1. Update `slalom.json` (title, speed, gates, valleyWidth, optional gateGap, a `style` line that tells Jev a line strategy). The viewer watches it and Jev adapts live — no restart, no second server. 2. `node "$JEV_DSH/toolchain/check.mjs"` verifies the workspace's `slalom.json` is valid. Run it before you call a run done. 3. Watch the descent. Does Jev line up on the next gate, commit as it arrives, and thread a clean run — or never fall (dull) or fall even at calm speed (bad)? That observation is the finding. ## Reading the run The viewer shows a top-down course: the skier descends with a snow trail, gates sweep toward it and light up green as they're threaded, a dashed line marks the target gate, and a finished run shows a CLEAN RUN or FELL banner plus a run log. Good courses produce a readable loop: Jev carves across the valley to the next gate, threads it, and builds a rhythm; `speed` is the difficulty dial — crank it up and Jev's aim wobbles and it clips a gate and falls. ## Verifying a course `node "$JEV_DSH/toolchain/check.mjs"` returns non-zero when `slalom.json` is invalid (no title, a non-positive speed, out-of-range gates/valleyWidt...