Drydock

自然语言验收测试 Agent：把 QA 用例编译成可版本化计划，在真实浏览器里确定性执行、采证、断言，并在定位失效时生成需要人工审批的 self-healing 补丁提案。

Natural-language acceptance testing agent: compile QA cases into versioned plans, run them deterministically in a real browser, collect evidence, assert outcomes, and later propose human-approved self-healing patches when locators drift.

Status

M0 local foundation is implemented:

• Pydantic data models for cases, plans, steps, assertions, runs, run events, and proposals.
• SQLite schema with append-only run_events enforced by triggers.
• Deterministic plan runner with an injectable browser session.
• Playwright adapter plus a local TodoMVC fixture.
• M0 acceptance script that reruns the same hardcoded plan three times and checks stable event signatures.

Public GitHub creation and first push require explicit Wesley confirmation per PROJECT_PLAN.md.

Quickstart

uv sync
uv run pytest
uv run python scripts/acceptance/run_todomvc.py --runs 3

Trace artifacts are written under runs/<run_id>/ with screenshots, DOM snapshots, and console logs for each step.