What problem does it solve? Building a predictive command palette requires tuning target aliases so a ranking model surfaces the right launch target per keystroke, and doing this blind leads to palettes where the model guesses or never re-ranks. ## Core Features & Use Cases - Palette authoring: Edit launcher.json with titles, prompts, categorized targets, and fuzzy-match aliases while the OpenHarness viewer updates Jev's ranking live with no restart. - Schema validation: Run node "$JEV_DSH/toolchain/check.mjs" to reject invalid palettes (missing title, fewer than two targets, duplicate names) before shipping. - Ranking evaluation: Query Jev directly through toolchain/jev.mjs to check which target it picks for a given query, using a deterministic mock without TYPESAFE_API_KEY or the live model with it. - Use Case: Add a Deploy target with aliases "ship", "release", and "deploy", then type partial queries in the viewer to confirm Jev's top pick moves toward Deploy as the query narrows. ## Quick Start Ask the agent to add a new launch target with aliases to launcher.json and verify the palette with the check script while watching Jev's ranking update in the viewer.
