only print local ui URL once

Author: Dylan Huang

eval_protocol/pytest/plugin.py

@@ -327,16 +327,16 @@ def _print_local_ui_results_urls(session):
             RESULTS_URLS_STASH_KEY = None
 
         # Get URLs from pytest stash
-        urls = []
+        urls_dict = {}
         if RESULTS_URLS_STASH_KEY is not None and RESULTS_URLS_STASH_KEY in session.stash:
-            urls = session.stash[RESULTS_URLS_STASH_KEY]
+            urls_dict = session.stash[RESULTS_URLS_STASH_KEY]
 
-        if urls:
+        if urls_dict:
             print("\n" + "=" * 80, file=sys.__stderr__)
             print("📊 LOCAL UI EVALUATION RESULTS", file=sys.__stderr__)
             print("=" * 80, file=sys.__stderr__)
 
-            for url_data in urls:
+            for url_data in urls_dict.values():
                 print(f"📊 Invocation {url_data['invocation_id']}:", file=sys.__stderr__)
                 print(f"  📊 Aggregate scores: {url_data['pivot_url']}", file=sys.__stderr__)
                 print(f"  📋 Trajectories: {url_data['table_url']}", file=sys.__stderr__)

eval_protocol/pytest/store_results_url.py

@@ -8,7 +8,7 @@ class ResultsUrl(TypedDict):
     table_url: str
 
 
-RESULTS_URLS_STASH_KEY = StashKey[list[ResultsUrl]]()
+RESULTS_URLS_STASH_KEY = StashKey[dict[str, ResultsUrl]]()
 
 
 def _store_local_ui_url_in_stash(invocation_id: str, pivot_url: str, table_url: str):
@@ -29,11 +29,14 @@ def _store_local_ui_url_in_stash(invocation_id: str, pivot_url: str, table_url:
             global RESULTS_URLS_STASH_KEY
 
             if RESULTS_URLS_STASH_KEY not in session.stash:  # pyright: ignore[reportAny]
-                session.stash[RESULTS_URLS_STASH_KEY] = []  # pyright: ignore[reportAny]
-
-            session.stash[RESULTS_URLS_STASH_KEY].append(  # pyright: ignore[reportAny]
-                {"invocation_id": invocation_id, "pivot_url": pivot_url, "table_url": table_url}
-            )
+                session.stash[RESULTS_URLS_STASH_KEY] = {}  # pyright: ignore[reportAny]
+
+            # Store by invocation_id as key - automatically handles deduplication
+            session.stash[RESULTS_URLS_STASH_KEY][invocation_id] = {  # pyright: ignore[reportAny]
+                "invocation_id": invocation_id,
+                "pivot_url": pivot_url,
+                "table_url": table_url,
+            }
         else:
             pass