Hi, thank you for the great project and all your hard work! 🙏

Background

• current Arbigent workflow, every test step invokes the LLM to decide the “next action,” even if the UI state has no change.

• Re-running identical regression tests on Web, iOS, or Android screens that haven’t been modified causes unnecessary cost.

Problem Statement

• When a screen’s HTML (Web) or UI Automator hierarchy (Android/iOS) and its corresponding “expect” assertion remain exactly the same, repeating the same LLM query wastes tokens without adding any new value.
• We need a lightweight, file-based caching mechanism that detects when “nothing has changed” and reuses the previous LLM response instead of calling the LLM API.

Solution

• Key Generation: Screen DOM Hash + Screen Shot Hash + Expect Text
  • For each test step, capture a snapshot of the displayed UI:
    • Web: Serialize the relevant portion of the DOM as a string (e.g. element.outerHTML, full page HTML, or a trimmed subset).
    • iOS / Android: Serialize the UI Automator tree or accessibility node hierarchy for the current screen.
  • Use this key as the filename to save the LLM response into a JSON file, and retrieve it with getOrPut.

Files Found in Your Project

• https://github.com/takahirom/arbigent/blob/main/arbigent-core/src/main/java/io/github/takahirom/arbigent/ArbigentAgent.kt
• https://github.com/takahirom/arbigent/blob/main/arbigent-ai-openai/src/main/java/io/github/takahirom/arbigent/OpenAIAi.kt