From e62283f1b6c05fd440641a5fbd8c5b314456dbb0 Mon Sep 17 00:00:00 2001
From: Jan Tychtl <jan.tychtl@gooddata.com>
Date: Fri, 6 Nov 2026 15:29:25 +0100
Subject: [PATCH] feat(gooddata-eval): keep workflowStatus parts out of scored
 answers

gen-ai is adding a stored workflowStatus multipart part that tells the
browser a dashboard workflow ended with nothing left open. It rides on
the turn's last visible answer, next to the text.

The SSE client put every non-text part it did not handle into
unhandled_parts, and render_answer_text appends those to the text the
evaluators score, so dashboard evals would grade a JSON blob as part of
the answer and log an unknown-part warning on every decline turn.

Register the type and skip it in _handle_multipart: it carries nothing
an evaluator scores.

jira: GDAI-2733
risk: low
---
 .../src/gooddata_eval/core/chat/sse_client.py       |  4 ++++
 packages/gooddata-eval/tests/test_chat_render.py    | 13 +++++++++++++
 2 files changed, 17 insertions(+)

diff --git a/packages/gooddata-eval/src/gooddata_eval/core/chat/sse_client.py b/packages/gooddata-eval/src/gooddata_eval/core/chat/sse_client.py
index ce7ead5c1..b8389c056 100644
--- a/packages/gooddata-eval/src/gooddata_eval/core/chat/sse_client.py
+++ b/packages/gooddata-eval/src/gooddata_eval/core/chat/sse_client.py
@@ -71,6 +71,7 @@
         "searchResults",
         "alertProposal",
         "clarifyingQuestions",
+        "workflowStatus",
     }
 )
 
@@ -246,6 +247,9 @@ def _handle_multipart(content: dict[str, Any], acc: _SseAccumulator, item_id: st
             acc.alert_proposals.append(part.get("alertProposal") or {})
         elif ptype == "searchResults":
             acc.search_results.append(part)
+        elif ptype == "workflowStatus":
+            # A UI signal that the turn ended a workflow; it carries nothing the answer is scored on.
+            continue
         else:
             if ptype not in _KNOWN_PART_TYPES:
                 _log.warning("unknown multipart part type %r; captured as an unhandled part", ptype)
diff --git a/packages/gooddata-eval/tests/test_chat_render.py b/packages/gooddata-eval/tests/test_chat_render.py
index 29904158f..ef7d7bdbb 100644
--- a/packages/gooddata-eval/tests/test_chat_render.py
+++ b/packages/gooddata-eval/tests/test_chat_render.py
@@ -106,6 +106,18 @@ def test_alert_proposal_part_still_goes_to_its_own_field():
     assert result.unhandled_parts == []
 
 
+def test_workflow_status_part_stays_out_of_the_answer():
+    result = parse_sse_lines(
+        _multipart_lines(
+            {"type": "workflowStatus", "skills": ["dashboard_editor"], "status": "ended"},
+            {"type": "text", "text": "Done."},
+        )
+    )
+    assert result.unhandled_parts == []
+    assert result.search_results == []
+    assert render_answer_text(result) == "Done."
+
+
 def test_known_part_types_matches_the_documented_gen_ai_union():
     gen_ai_message_part_union = {
         "text",
@@ -118,6 +130,7 @@ def test_known_part_types_matches_the_documented_gen_ai_union():
         "searchResults",
         "alertProposal",
         "clarifyingQuestions",
+        "workflowStatus",
     }
     assert gen_ai_message_part_union == _KNOWN_PART_TYPES