fix(gooddata-eval): persist skill_routing credit across conversation turns

_activated_skills() only looked at the current turn's tool calls, so
skill_routing was recomputed from scratch each turn. The platform keeps a
skill active once set_skills is called, so an agent correctly omits a
redundant set_skills call on a later turn that reuses the same skill --
but the scorer forced skill_routing=False whenever no skill was activated
*this* turn, regardless of whether it was already active.

Found via scripts/authoring/debug_conversation.py replaying
analyst-explores-dynamic-currency-conversion (gdc-mic-ai-evaluation repo):
turns t4/t5 both ran create_adhoc_visualization/create_metric successfully
against an already-active skill, yet scored FAIL solely on this.

Track activated skills in a running set across the whole conversation
instead of resetting it every turn.

Author: Tomkess

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

@@ -300,6 +300,10 @@ def run_agentic_conversation(
     response_id: str | None = None
     conversation_tool_call_events: list[ToolCallEvent] = []
     conversation_reasoning_step_events: list[ReasoningStepEvent] = []
+    # Skills activated by any turn so far. The platform keeps a skill active across
+    # turns once set -- an agent correctly reuses the already-active skill without
+    # re-issuing set_skills, so routing credit must not require a fresh call every turn.
+    activated_skills_so_far: set[str] = set()
     # Every send_message() call (across every logical turn AND every clarification
     # sub-turn within it) restarts call_ts/ts near 0 -- these run across the whole
     # conversation, not reset per logical turn, so every one of those calls shifts them.
@@ -375,7 +379,8 @@ def run_agentic_conversation(
                 current_message = _get_sim_user_response(response_text, resolved_turn, resolved_expected)
 
             activated = _activated_skills(all_tool_calls)
-            skill_routing = turn.expected_skill in activated if activated else False
+            activated_skills_so_far |= set(activated)
+            skill_routing = turn.expected_skill in activated_skills_so_far
             output_present = _check_output_present(resolved_turn, final_result) if final_result else False
             output_correct = (
                 _check_output_correct(resolved_turn, final_result) if (final_result and output_present) else None

packages/gooddata-eval/tests/test_agentic_conversation.py

@@ -329,6 +329,64 @@ def test_run_agentic_conversation_deletes_every_unique_metric_across_turns():
     assert deleted == [("ws1", "extra"), ("ws1", "shared")]
 
 
+def test_run_agentic_conversation_skill_routing_persists_across_turns():
+    """A skill activated in an earlier turn stays credited when a later turn reuses it
+    without re-issuing set_skills -- the platform keeps a skill active once set, so an
+    agent correctly omits a redundant set_skills call. Requiring a fresh call every turn
+    produced false FAILs on turns that did the right thing (found via
+    debug_conversation.py replaying analyst-explores-dynamic-currency-conversion,
+    turns t4/t5: create_adhoc_visualization/create_metric both ran and succeeded, but
+    skill_routing was False solely because set_skills wasn't repeated)."""
+    mock_client = MagicMock()
+    mock_client.create_conversation.return_value = "conv-1"
+    mock_client.send_message.side_effect = [
+        _metric_turn_result([_skills_tc("metric"), _create_metric_tc("m1")]),
+        _metric_turn_result([_create_metric_tc("m2")]),  # no set_skills -- skill already active
+    ]
+
+    with (
+        patch("gooddata_eval.core.agentic.conversation.ChatClient", return_value=mock_client),
+        patch("gooddata_eval.core.agentic.conversation.GoodDataSdk"),
+    ):
+        result = run_agentic_conversation(
+            host="http://host/api/v1/actions/workspaces/ws1/ai",
+            token="tok",
+            workspace_id="ws1",
+            fixture=_two_metric_turn_fixture(),
+        )
+
+    assert result.turn_results[0].skill_routing is True
+    assert result.turn_results[1].skill_routing is True
+
+
+def test_run_agentic_conversation_skill_routing_false_when_skill_never_activated():
+    """Guard against the fix being too lenient: a skill that no turn ever activated
+    must still fail routing, not be credited by the cumulative-set change."""
+    mock_client = MagicMock()
+    mock_client.create_conversation.return_value = "conv-1"
+    mock_client.send_message.return_value = _metric_turn_result([_create_metric_tc("m1")])
+
+    fixture = ConversationFixture(
+        id="test-never-activated",
+        expected_skills=["metric"],
+        turns=[
+            TurnDefinition(turn_id="t1", message="Create x", expected_skill="metric", expected_output_type="metric"),
+        ],
+    )
+    with (
+        patch("gooddata_eval.core.agentic.conversation.ChatClient", return_value=mock_client),
+        patch("gooddata_eval.core.agentic.conversation.GoodDataSdk"),
+    ):
+        result = run_agentic_conversation(
+            host="http://host/api/v1/actions/workspaces/ws1/ai",
+            token="tok",
+            workspace_id="ws1",
+            fixture=fixture,
+        )
+
+    assert result.turn_results[0].skill_routing is False
+
+
 def test_run_agentic_conversation_deletes_metrics_even_when_a_later_turn_raises():
     mock_client = MagicMock()
     mock_client.create_conversation.return_value = "conv-1"