Publish tiny_computer_agent_app.py from marimo

tiny_computer_agent_app.py

@@ -1,3 +1,10 @@
+# /// script
+# requires-python = ">=3.13"
+# dependencies = [
+#     "rapidocr-onnxruntime==1.4.4",
+# ]
+# ///
+
 import marimo
 
 __generated_with = "0.25.2"
@@ -211,7 +218,7 @@ def repo_files():
 
     [project]
     name = "tiny-computer-agent"
-    version = "0.1.0"
+    version = "0.2.0"
     description = "An agentic harness that drives a real computer with a tiny, fully local model - no cloud API."
     readme = "README.md"
     requires-python = ">=3.10"
@@ -227,8 +234,9 @@ def repo_files():
     dependencies = []   # the harness core imports nothing outside the standard library
 
     [project.optional-dependencies]
-    desktop = ["mss>=9.0", "pyautogui>=0.9", "pillow>=10"]
-    vision = ["transformers>=4.48", "torch>=2.2", "pillow>=10"]
+    ocr = ["rapidocr-onnxruntime>=1.3"]        # labels from pixels, no system binary
+    desktop = ["mss>=9.0", "pyautogui>=0.9", "pillow>=10"]   # capture + real input
+    vision = ["transformers>=4.48", "torch>=2.2", "pillow>=10"]  # tiny model brains
     dev = ["pytest>=8"]
 
     [project.scripts]
@@ -280,7 +288,8 @@ def repo_files():
     .DS_Store
     '''
 
-    mo.md(f"{len(pyproject_toml) + len(license_txt) + len(gitignore_txt)} chars of packaging ready to publish")
+    mo.md(f"packaging: {len(pyproject_toml) + len(license_txt) + len(gitignore_txt)} chars, "
+          f"v0.2.0 with ocr/desktop/vision extras")
 
     return gitignore_txt, license_txt, pyproject_toml
 
@@ -368,36 +377,39 @@ def publish_status(
 
 @app.cell
 def demo():
-    from tiny_computer_agent import (
-        AgentLoop, DryRunExecutor, Element, FakeScreen, PlannerBrain, Policy,
-    )
-
-    _scripted_screen = FakeScreen(
-        [
-            (Element("Terminal", x=0, y=0, width=40, height=24),
-             Element("Files", x=60, y=0, width=40, height=24)),
-            (Element("Name", kind="input", x=100, y=120, width=200, height=24),
-             Element("Save", x=320, y=120, width=80, height=24)),
-            (Element("Save", x=320, y=120, width=80, height=24),),
-        ],
-        text="Untitled document - unsaved changes",
-    )
+    from tiny_computer_agent import AgentLoop, DryRunExecutor, PlannerBrain, Policy
+    from tiny_computer_agent.demo import fake_desktop
+    from tiny_computer_agent.observe import ImageScreen
+    from tiny_computer_agent.perception import IndexedScreen, OcrDetector
+
+    # A real screenshot -> real OCR -> labelled, clickable elements. No display needed:
+    # this is the same path --image takes, and it is what makes a tiny model viable.
+    _detector = OcrDetector()
+    _observation = IndexedScreen(ImageScreen(fake_desktop()), _detector).capture()
+
+    perception_rows = [
+        {"label": e.label, "kind": e.kind, "x": e.x, "y": e.y, "w": e.width, "h": e.height}
+        for e in _observation.elements
+    ]
 
-    demo_episode = AgentLoop(
-        screen=_scripted_screen,
+    perception_episode = AgentLoop(
+        screen=IndexedScreen(ImageScreen(fake_desktop()), _detector),
         brain=PlannerBrain(),
         executor=DryRunExecutor(),
-        policy=Policy(step_budget=6),
+        policy=Policy(step_budget=4),
     ).run("click \'Save\' to save the document")
 
+    _clicked = next((s.result.note for s in perception_episode.steps if s.action and s.action.kind == "click"), "n/a")
+
     mo.vstack(
         [
             mo.md(
-                "**Scripted episode, dry run** — no model, no display, nothing touched. "
-                "The planner waits for the dialog, clicks the label, and stops."
+                f"**Perception, from pixels.** OCR read **{len(perception_rows)} labels** off a "
+                f"synthetic screenshot; the planner then chose one and the harness resolved it to a real pixel."
             ),
-            mo.ui.table(demo_episode.to_rows(), selection=None),
-            mo.md(f"`{demo_episode.render().splitlines()[1]}`"),
+            mo.ui.table(perception_rows, selection=None),
+            mo.md(f"`{_clicked}`"),
+            mo.md(f"`{perception_episode.render().splitlines()[1]}`"),
         ]
     )