Introduction

TestDog is a desktop test case management tool: generate UI test scripts from natural language or manual recording, then replay them as-written with Playwright to verify (everyday regression costs no model tokens). Built for end-to-end regression testing of web applications.

Core capabilities

• AI script generation — Describe the test in natural language (attachments and login state supported). The model pre-splits a step plan, then executes it in a real browser via a tool-calling loop, turning each step into a semantic-locator script. Pause/resume, AI repair and manual takeover are all supported.
• Manual recording — Operate the browser through Playwright codegen; the recording is parsed into editable structured steps.
• Deterministic replay — Playwright replay consumes no model tokens. UI / API / WebSocket assertions included; broken selectors trigger automatic AI self-healing with one-click write-back.
• Plugin system — Semantic-action plugins for dropdowns, tree selects and date pickers, dispatched per component-library variant, eliminating "clicked but didn't select" false successes.

Key terms

Don't worry if these are new — each page explains them in context; quick reference:

Term	Meaning
AI gateway	The service endpoint for calling LLMs; anything speaking the OpenAI-compatible protocol works (see Settings)
Pre-split (step plan)	Before executing, the model turns your natural-language description into a step-by-step plan for confirmation
Semantic locator	Step locators written the way humans describe elements (e.g. role=button name=Login) — more resilient to page changes than raw CSS/XPath
Replay	Re-executes the script step by step, exactly as written, to verify the result
Self-healing	When a selector breaks (page redesign), the AI re-finds the element using the step's natural-language instruction and continues
Headless mode	Replay runs silently in the background without showing a browser window; off by default so you can watch the process
Login config	A login state recorded once, reused by generation and replay — no need to test login in every case
Preset	An execution group of plugins: only projects linked to a preset get its plugins injected during generation/replay
Token usage	The billing unit of model calls made by AI generation/self-healing; cache-hit portions cost far less than normal calls

How it works

Tauri 2 desktop shell (Rust + system webview)
  └ React frontend (Ant Design 5 + Tailwind CSS 4, zh/en)
      · fetch  → backend REST
      · WebSocket → live progress (generate / run / record)
Node backend (Fastify 5 + Prisma 7 + SQLite)
  · LLM tool-calling loop (OpenAI-compatible gateway)
  · Stagehand 4 (generation execution / self-healing relocation)
  · Playwright (replay / recording / precise verification)

The LLM only participates in generation and self-healing relocation; everyday regression replay is fully deterministic and offline.

UI map

Menu	Purpose	Documentation
Projects	Manage projects, test cases and script versions	Getting started
Run records	Replay results, failure screenshots and diagnostics	Replay
Generation logs	Step-level logs and token accounting	AI generation
Plugins	Built-in/custom plugins and preset orchestration	Plugin system
Settings	AI gateway, model, browser path, etc.	Settings

When to use TestDog

• Regression testing: capture repetitive smoke/regression flows as scripts and replay them on demand.
• Component-heavy apps: dropdowns, cascaders and date pickers on Ant Design / Element UI / Element Plus.
• Team collaboration: share cases as .testcase files, or have coding agents generate them from the bundled skill.