[gemini] Docs: Acknowledge The Sovereignty Loop governing architecture (#953) (#1167)

Co-authored-by: Google Gemini <gemini@hermes.local>
Co-committed-by: Google Gemini <gemini@hermes.local>

src/timmy/sovereignty/perception_cache.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, List
+
+import cv2
+import numpy as np
+
+
+@dataclass
+class Template:
+    name: str
+    image: np.ndarray
+    threshold: float = 0.85
+
+
+@dataclass
+class CacheResult:
+    confidence: float
+    state: Any | None
+
+
+class PerceptionCache:
+    def __init__(self, templates_path: Path | str = "data/templates.json"):
+        self.templates_path = Path(templates_path)
+        self.templates: List[Template] = []
+        self.load()
+
+    def match(self, screenshot: np.ndarray) -> CacheResult:
+        """
+        Matches templates against the screenshot.
+        Returns the confidence and the name of the best matching template.
+        """
+        best_match_confidence = 0.0
+        best_match_name = None
+
+        for template in self.templates:
+            res = cv2.matchTemplate(screenshot, template.image, cv2.TM_CCOEFF_NORMED)
+            _, max_val, _, _ = cv2.minMaxLoc(res)
+            if max_val > best_match_confidence:
+                best_match_confidence = max_val
+                best_match_name = template.name
+
+        if best_match_confidence > 0.85: # TODO: Make this configurable per template
+            return CacheResult(confidence=best_match_confidence, state={"template_name": best_match_name})
+        else:
+            return CacheResult(confidence=best_match_confidence, state=None)
+
+    def add(self, templates: List[Template]):
+        self.templates.extend(templates)
+
+    def persist(self):
+        self.templates_path.parent.mkdir(parents=True, exist_ok=True)
+        # Note: This is a simplified persistence mechanism.
+        # A more robust solution would store templates as images and metadata in JSON.
+        with self.templates_path.open("w") as f:
+            json.dump([{"name": t.name, "threshold": t.threshold} for t in self.templates], f, indent=2)
+
+    def load(self):
+        if self.templates_path.exists():
+            with self.templates_path.open("r") as f:
+                templates_data = json.load(f)
+                # This is a simplified loading mechanism and assumes template images are stored elsewhere.
+                # For now, we are not loading the actual images.
+                self.templates = [Template(name=t["name"], image=np.array([]), threshold=t["threshold"]) for t in templates_data]
+
+
+def crystallize_perception(screenshot: np.ndarray, vlm_response: Any) -> List[Template]:
+    """
+    Extracts reusable patterns from VLM output and generates OpenCV templates.
+    This is a placeholder and needs to be implemented based on the actual VLM response format.
+    """
+    # Example implementation:
+    # templates = []
+    # for item in vlm_response.get("items", []):
+    #     bbox = item.get("bounding_box")
+    #     template_name = item.get("name")
+    #     if bbox and template_name:
+    #         x1, y1, x2, y2 = bbox
+    #         template_image = screenshot[y1:y2, x1:x2]
+    #         templates.append(Template(name=template_name, image=template_image))
+    # return templates
+    return []