fix(gooddata-eval): count either producing tool in refusal scoring

A refusal is a creation shape, so `_producing_tool` returns
`draft_dashboard` and the flag matched only that name. An agent that routed
to the editor and had a `patch_dashboard` call rejected left the same empty
result and empty part, so it passed as a clean refusal -- the same hole the
previous commit closed, through the other tool.

Both names count now. The flag is read only by the refusal branch, so
widening it tightens that shape and changes nothing else.

Found in review by CodeRabbit.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: Tomkess

packages/gooddata-eval/src/gooddata_eval/core/agentic/dashboard_skill.py

@@ -769,9 +769,9 @@ class DashboardEvaluation:
     titles_matched: bool = True
     forbidden_absent: bool = True
     answer_matched: bool = True
-    # Whether the producing tool was called AT ALL, successful or not. Only a refusal case
-    # reads it: everywhere else a failed call is a retry the agent recovered from, which is
-    # why `_extract_tool_result` skips it.
+    # Whether EITHER dashboard-producing tool was called at all, successful or not. Only a
+    # refusal case reads it: everywhere else a failed call is a retry the agent recovered
+    # from, which is why `_extract_tool_result` skips it.
     producing_tool_called: bool = False
     failures: list[str] = field(default_factory=list)
     notes: list[str] = field(default_factory=list)
@@ -935,7 +935,10 @@ def evaluate_dashboard_response(
                     else []
                 ),
                 *(
-                    [f"the case expects no dashboard, but the agent called {tool} (the call did not succeed)"]
+                    [
+                        "the case expects no dashboard, but the agent called a dashboard-producing "
+                        "tool (the call did not succeed)"
+                    ]
                     if producing_tool_called and not drafted
                     else []
                 ),
@@ -1216,8 +1219,11 @@ def _execute_single_dashboard_run(
             _skill_activated(all_tool_call_events, _required_skill(expected_output)),
             patch_part=patch_part,
             answer_text=answer_text,
-            # Any call, not just a successful one -- see DashboardEvaluation.
-            producing_tool_called=any(tc.function_name == tool for tc in all_tool_call_events),
+            # Any call, not just a successful one -- see DashboardEvaluation. BOTH tools,
+            # not the one this case selected: a refusal is a creation shape, so `tool` is
+            # `draft_dashboard`, and an agent that routed to the editor and had a
+            # `patch_dashboard` call rejected would otherwise leave this false and pass.
+            producing_tool_called=any(tc.function_name in (_DRAFT_TOOL, _PATCH_TOOL) for tc in all_tool_call_events),
         ),
         tool_result=tool_result,
         dashboard_part=dashboard_part,

packages/gooddata-eval/tests/test_agentic_dashboard_skill.py

@@ -1515,6 +1515,20 @@ def test_a_failed_draft_attempt_is_read_off_the_turn_not_the_result(self):
         assert summary.best.evaluation.producing_tool_called
         assert not summary.best.evaluation.strict_pass
 
+    def test_a_rejected_patch_call_counts_too(self):
+        """A refusal is a creation shape, so the case's selected tool is `draft_dashboard`.
+        An agent that routed to the editor instead and had its `patch_dashboard` call
+        rejected leaves the same empty result, so matching only the selected tool would let
+        it pass."""
+        client = MagicMock()
+        client.send_message.return_value = _chat_result(
+            tool_calls=[_tool_call("patch_dashboard", {"status": "error", "message": "no such dashboard"})],
+            text="There is no fraud data in this workspace, so I cannot build that.",
+        )
+        summary = _run_with(client, self._NO_DATA)
+        assert summary.best.evaluation.producing_tool_called
+        assert not summary.best.evaluation.strict_pass
+
     def test_an_answer_alone_still_passes_when_nothing_was_called(self):
         """The control for the two above: the flag must not fail a genuine refusal."""
         client = MagicMock()