dashboard: simplify execution log view and highlight tool calls (#74)

Author: simonvanlaak

src/cyberagent/ui/dashboard.py

@@ -167,40 +167,58 @@ def _render_execution_log(st: Any, execution_log: str | None) -> None:
         st.dataframe(rows, width="stretch", hide_index=True)
     else:
         st.write(rows)
-    _render_execution_log_details(st, parsed)
 
 
 def _build_execution_log_rows(entries: list[object]) -> list[dict[str, object]]:
     rows: list[dict[str, object]] = []
     for index, entry in enumerate(entries, 1):
         if isinstance(entry, dict):
             step_type = str(entry.get("type", "-"))
+            tool_name = (
+                _extract_tool_name(entry) if _is_tool_call_event(step_type) else "-"
+            )
             source = str(entry.get("source", "-"))
             summary = _summarize_execution_entry(entry)
         else:
             step_type = type(entry).__name__
+            tool_name = "-"
             source = "-"
             summary = _truncate_text(str(entry), 200)
         rows.append(
             {
                 "step": index,
                 "type": step_type,
+                "tool": tool_name,
                 "source": source,
                 "summary": summary,
             }
         )
     return rows
 
 
-def _render_execution_log_details(st: Any, entries: list[object]) -> None:
-    if not hasattr(st, "expander"):
-        return
-    for index, entry in enumerate(entries, 1):
-        with st.expander(f"Step {index} details"):
-            if hasattr(st, "code"):
-                st.code(json.dumps(entry, indent=2, ensure_ascii=True, default=str))
-            else:
-                st.write(entry)
+def _is_tool_call_event(step_type: str) -> bool:
+    normalized = step_type.strip().lower()
+    return "toolcall" in normalized or ("tool" in normalized and "call" in normalized)
+
+
+def _extract_tool_name(entry: dict[str, object]) -> str:
+    name = entry.get("name")
+    if isinstance(name, str) and name.strip():
+        return name.strip()
+
+    content = entry.get("content")
+    if isinstance(content, dict):
+        content_name = content.get("name")
+        if isinstance(content_name, str) and content_name.strip():
+            return content_name.strip()
+    if isinstance(content, list):
+        for item in content:
+            if isinstance(item, dict):
+                item_name = item.get("name")
+                if isinstance(item_name, str) and item_name.strip():
+                    return item_name.strip()
+
+    return "-"
 
 
 def _summarize_execution_entry(entry: dict[str, object]) -> str:

tests/cyberagent/test_dashboard_ui.py

@@ -392,9 +392,7 @@ def _fake_resolve(
 
 def test_render_task_details_shows_status_reasoning(monkeypatch) -> None:
     writes: list[object] = []
-    code_values: list[str] = []
     dataframe_data: list[object] = []
-    expander_labels: list[str] = []
 
     class _TitleCol:
         def __enter__(self):
@@ -428,23 +426,9 @@ def markdown(self, _text: str) -> None:
         def write(self, text: object) -> None:
             writes.append(text)
 
-        def code(self, text: str) -> None:
-            code_values.append(text)
-
         def dataframe(self, data: object, **_kwargs: object) -> None:
             dataframe_data.append(data)
 
-        class _ExpanderCtx:
-            def __enter__(self):
-                return None
-
-            def __exit__(self, exc_type, exc, tb):
-                return False
-
-        def expander(self, label: str):
-            expander_labels.append(label)
-            return self._ExpanderCtx()
-
         def button(self, _text: str) -> bool:
             return False
 
@@ -483,9 +467,8 @@ def rerun(self) -> None:
     rows = dataframe_data[0]
     assert isinstance(rows, list)
     assert rows[0]["type"] == "ToolCallExecutionEvent"
+    assert rows[0]["tool"] == "task_search"
     assert "task_search" in str(rows[0]["summary"])
-    assert len(expander_labels) == 1
-    assert len(code_values) == 1
 
 
 def test_render_execution_log_fallbacks_to_code_for_invalid_json() -> None: