Upload app.py

app.py

@@ -542,6 +542,16 @@ def extract_json(s: str):
 def execute_page_expert(expert_instructions: str, page_num: int):
     state.add_log(f'👁️ Spawning Page Expert for page **{page_num}**')
     state.add_analysis(f"👁️ Page Expert searching for {expert_instructions}")
+    state.add_log(f'📄 Attaching full-page context for page **{page_num}**')
+    state.add_analysis(
+        f"📄 Full-page context attached for page `{page_num}`"
+    )
+
+    full_page_img = Image.open(
+        io.BytesIO(image_bytes_list[page_num])
+    )
+    state.add_image(full_page_img)
+    
     client = openai.OpenAI()
 
     tools = [
@@ -937,19 +947,7 @@ def agent_worker(user_question):
                             tile_indexes=tile_idxs,
                             page_num=page_num
                         )
-                        result["high_res_proof_b64"] = base64.b64encode(
-                            stitched_bytes
-                        ).decode("utf-8")
                         
-                        state.add_log(f'📄 Attaching full-page context for page **{page_num}**')
-                        state.add_analysis(
-                            f"📄 Full-page context attached for page `{page_num}`"
-                        )
-    
-                        full_page_img = Image.open(
-                            io.BytesIO(image_bytes_list[page_num])
-                        )
-                        state.add_image(full_page_img)
 
                 tool_responses.append(
                     types.Part.from_function_response(
@@ -957,6 +955,25 @@ def agent_worker(user_question):
                         response={"result": result}
                     )
                 )
+                
+                # If this tool produced visual proof, ALSO send it as an image part
+                if name == "execute_page_expert":
+                    tile_idxs = result.get("visual_pointers", [])
+                    page_num = args.get("page_num")
+                    
+                    if stitched_bytes:
+                        tool_responses.append(
+                            types.Part.from_bytes(
+                                stitched_bytes,
+                                mime_type="image/png"
+                            )
+                        )
+                        tool_responses.append(
+                            types.Part.from_bytes(
+                                image_bytes_list[page_num],
+                                mime_type="image/png"
+                            )
+                        )
 
         state.add_analysis("🧠 Returning tool outputs to planner...")
         response = chat.send_message(tool_responses)