burst-bug-hunterlisted

# Burst Bug-Hunter (generic engine) ## Why Interactive step-by-step CLI driving costs **one LLM turn per action** — a long hunt is mostly model round-trips, not browser work. This flips the default: **the unit of work is a whole multi-step script (a "burst")** that runs unattended and returns one structured result. You observe at burst *boundaries*, not per action. Typical effect: ~8× fewer model turns per feature vs step-by-step. ## Two modes - **Verify with bursts** (default) — you HAVE the app model; confirm expected behavior fast. - **Explore step-by-step** (rare) — you DON'T have the model yet; build it. **Crystallize whatever flows you discover into the repo's verb library** so the next agent inherits them (see *Answering "doesn't the agent need app knowledge?"* below). ## The engine (this skill, `assets/`, app-agnostic) - `driver-server.mjs <app-url> <workdir> [config.json]` — opens the app ONCE, keeps it warm, runs submitted job scripts, enforces a hard deadline (`<workdir>/deadline` epoch secs → all non-shutdown jobs rejected after). Every result carries `remainingMs`. - `start-hunt.sh <workdir> <app-url> [config.json]` — boots one warm server, waits ready. - `vrun` — submit a job and block for the full result (`VDRIVE=<workdir>` selects the server). Also: `vrun --reset`, `vrun --inline 'CODE'`, `vrun --finding '{json}'` (append a finding line without a browser round-trip). - `example-burst.mjs` — copy/adapt template: a **defensive** burst that records