dsh-mp-automator

The scariest failures in LLM-driven Mini Program testing are silent: element UIDs get quietly renumbered across reconnects, compiled artifacts lag behind the source, and a text-only model describes a screenshot it never actually saw. Prompts can remind, but they cannot enforce. dsh-mp-automator bakes the discipline into the tools themselves: selector-based addressing ensures no UID ever crosses a call boundary, a freshness gate runs a real build check before every action and hard-rejects stale artifacts, and a dual-path screenshot pipeline guarantees that both text-only and vision models receive an honest, self-contained context, with an image budget enforced in code.

Eight mp_* tools cover the full loop: session management, project health check, page inspection, selector query with geometry flags, element actions (tap, input, longpress, nav, switchTab, reLaunch, camera-free scan injection), screenshot with geometry facts, console reading with errors-first ordering, and a gated JS eval escape hatch. Every result is byte-clamped and self-contained, so long sessions survive context compaction without degrading into "see above." A companion skill file adds a judgment layer: three-beat inspect-act-verify rhythm, selector discipline, assertion recipes, and screenshot economy awareness.

Built for macOS developers iterating on WeChat Mini Programs who want to fold dsh agents into their daily QA workflow without writing Playwright scripts or trusting the model to remember a wall of caveats. Open a dsh session in the project directory, describe the test in plain language, and the tools hold the line for you.