Publish tests/ from marimo

tests/test_live_desktop.py

@@ -6,7 +6,7 @@ synthetic screenshot. It is opt-in because it moves the real mouse and clicks:
     TINY_COMPUTER_AGENT_LIVE_TEST=1 python -m pytest tests/test_live_desktop.py
 
 Run it on a throwaway desktop (VM, container with Xvfb), never your working
-session. With Xvfb it needs no hardware:
+session. With Xvfb it needs no hardware at all:
 
     Xvfb :99 -screen 0 1280x800x24 &
     DISPLAY=:99 TINY_COMPUTER_AGENT_LIVE_TEST=1 python -m pytest tests/test_live_desktop.py
@@ -21,10 +21,11 @@ import time
 import pytest
 
 from tiny_computer_agent.act import PyAutoGuiExecutor
-from tiny_computer_agent.actions import Action
 from tiny_computer_agent.desktop import display_available
 from tiny_computer_agent.observe import GrabScreen
-from tiny_computer_agent.perception import OcrDetector, detector_available
+from tiny_computer_agent.perception import (
+    IndexedScreen, OcrDetector, detector_available,
+)
 
 OPT_IN = os.environ.get("TINY_COMPUTER_AGENT_LIVE_TEST") == "1"
 
@@ -33,15 +34,18 @@ pytestmark = pytest.mark.skipif(
     reason="moves the real mouse: set TINY_COMPUTER_AGENT_LIVE_TEST=1 on a throwaway desktop",
 )
 
-APP = '''import pathlib, tkinter as tk
-
-MARKER = pathlib.Path("%(marker)s")
 WINDOW = (120, 120)
 BUTTON = (20, 155, 110, 40)          # x, y, width, height inside the window
+READY_TIMEOUT = 20.0
+
+APP = '''import pathlib, tkinter as tk
+
+MARKER = pathlib.Path(r"__MARKER__")
+READY = pathlib.Path(str(MARKER) + ".ready")
 
 root = tk.Tk()
 root.title("Document")
-root.geometry("420x230+%%d+%%d" %% WINDOW)
+root.geometry("420x230+120+120")
 root.configure(bg="#f5f6f8")
 title = tk.Label(root, text="Untitled document - unsaved changes", bg="#f5f6f8")
 title.place(x=20, y=18)
@@ -57,8 +61,9 @@ def chosen(what):
         widget.place_forget()
 
 buttons = [tk.Button(root, text="Save", command=lambda: chosen("Save"))]
-buttons[0].place(x=BUTTON[0], y=BUTTON[1], width=BUTTON[2], height=BUTTON[3])
+buttons[0].place(x=20, y=155, width=110, height=40)
 root.update()
+READY.write_text("mapped")
 root.mainloop()
 '''
 
@@ -70,43 +75,69 @@ def save_dialog(tmp_path):
         pytest.skip("no display session to drive")
     if not detector_available("ocr")[0]:
         pytest.skip("this test needs OCR to find the button")
+
     marker = tmp_path / "clicked.txt"
+    ready = pathlib.Path(str(marker) + ".ready")
     app = tmp_path / "app.py"
-    app.write_text(APP % {"marker": marker})
+    app.write_text(APP.replace("__MARKER__", str(marker)))
+
     process = subprocess.Popen([sys.executable, str(app)])
     try:
-        time.sleep(2.5)                      # let Tk map the window
-        yield marker, (120, 120), (20, 155, 110, 40)
+        deadline = time.time() + READY_TIMEOUT
+        while time.time() < deadline and not ready.exists():
+            if process.poll() is not None:
+                pytest.fail(f"the test window exited immediately (code {process.returncode})")
+            time.sleep(0.2)
+        assert ready.exists(), "the test window never mapped"
+        yield marker
     finally:
+        # Wait for the window to actually disappear: a ghost window left on the
+        # display makes the *next* test's OCR see two identical buttons.
         process.terminate()
+        try:
+            process.wait(timeout=10)
+        except subprocess.TimeoutExpired:  # pragma: no cover
+            process.kill()
+            process.wait(timeout=5)
+
+
+def _look() -> tuple:
+    """(observation, elements) - one real capture, one real OCR pass."""
+    observation = GrabScreen(monitor=1).capture()
+    return observation, OcrDetector().detect(observation.frame)
 
 
 def test_ocr_finds_a_real_button_at_the_right_pixels(save_dialog):
     """Pixels in, a label whose coordinates really are the button's."""
-    marker, window, (bx, by, bw, bh) = save_dialog
-    observation = GrabScreen(monitor=1).capture()
-    elements = OcrDetector().detect(observation.frame)
+    _, elements = _look()
 
     save = [e for e in elements if e.matches("Save")]
     assert save, f"OCR found no Save button, only {[e.label for e in elements]}"
-    centre = save[0].center
-    assert window[0] + bx <= centre[0] <= window[0] + bx + bw, f"{centre} outside the button"
-    assert window[1] + by <= centre[1] <= window[1] + by + bh, f"{centre} outside the button"
+    x, y = save[0].center
+    assert WINDOW[0] + BUTTON[0] <= x <= WINDOW[0] + BUTTON[0] + BUTTON[2], f"{x} outside the button"
+    assert WINDOW[1] + BUTTON[1] <= y <= WINDOW[1] + BUTTON[1] + BUTTON[3], f"{y} outside the button"
 
 
 def test_a_real_click_lands_on_the_button(save_dialog):
     """The whole loop: see it, choose it, click it, and the app reacts."""
     from tiny_computer_agent import AgentLoop, PlannerBrain, Policy
 
-    marker, _, _ = save_dialog
+    _, elements = _look()
+    labels = [e.label for e in elements]
+    save = [e for e in elements if e.matches("Save")]
+    assert save, f"nothing named Save on screen before clicking: {labels}"
+
+    # IndexedScreen, not a bare GrabScreen: capture gives pixels, the detector
+    # gives labels, and without labels there is nothing for a brain to choose.
     episode = AgentLoop(
-        screen=GrabScreen(monitor=1),
+        screen=IndexedScreen(GrabScreen(monitor=1), OcrDetector()),
         brain=PlannerBrain(),
         executor=PyAutoGuiExecutor(),
         policy=Policy(dry_run=False, step_budget=2),
         verify=lambda observation: "saved" in observation.text.lower(),
     ).run("click 'Save' to save the document")
 
-    assert marker.exists(), "no click landed on the button"
-    assert marker.read_text() == "Save"
-    assert episode.success, episode.render()
+    detail = episode.render() + f"\nclicked at {save[0].center}, labels were {labels}"
+    assert save_dialog.exists(), f"no click landed on the button\n{detail}"
+    assert save_dialog.read_text() == "Save"
+    assert episode.success, detail