App Test

app_test — test apps locally by clicking their buttons

When you build a UI you already have the button images, so an image-matching bot can drive it — no cloud vision tokens needed. A tiny JSON spec → a runnable SikuliX script.

Spec → script → run

python -m skills.app_test.cli gen tests/login_flow.json     # print the SikuliX script
python -m skills.app_test.cli run tests/login_flow.json --out build

{
  "name": "login_flow",
  "image_dir": "tests/images",
  "similarity": 0.8,
  "steps": [
    {"do": "wait",           "image": "login_btn.png", "timeout": 10},
    {"do": "click",          "image": "login_btn.png"},
    {"do": "type",           "text": "user@example.com"},
    {"do": "click",          "image": "submit.png"},
    {"do": "assert_visible", "image": "welcome.png", "timeout": 8},
    {"do": "assert_absent",  "image": "error.png"}
  ]
}

Actions: wait · click · double_click · right_click · type · sleep · assert_visible · assert_absent.

Running it

Uses OculiX (the maintained SikuliX fork: OpenCV matching + embedded Tesseract OCR, Java) — or any SikuliX. The generated -r scripts are drop-in compatible with both.

Install OculiX (Java 11+ required): download the platform "ide" jar from https://github.com/oculix-org/Oculix/releases and place it at ~/.oculix/oculixide.jar (auto-detected), or set OCULIX_JAR=/path/to/jar. runsikulix/oculix on PATH also work. run always generates the .sikuli bundle; it executes it via java -jar -r -c when a runner is found, else it reports how to install one.

Verified: a generated bundle runs end-to-end on OculiX 3.0.4 (oculixide-3.0.4-windows.jar) — Jython executes the steps and returns the exit code (0 = pass).

Why local / economical

Image matching is CPU-cheap and needs no model. For richer checks (is this screen semantically right?), point verification at a Hailo-8/10 NPU or a local vision model via [[llm_backends]] instead of a paid cloud vision API — 0 cloud tokens. The generator is deterministic and unit-tested; the GUI run is local.

Related: [[llm_backends]] (local vision backends), [[metrics]].

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zedarvates
• Source: zedarvates/botte-secrete
• License: MIT
• Homepage: https://github.com/zedarvates/botte-secrete

Install and usage instructions live in the source repository linked above.