fix(chat): preserve answers with invalid legacy handoff metadata

loopx/chat.py

@@ -837,10 +837,13 @@ def parse_agent_response(
                 # A bad action envelope must not discard an already generated
                 # answer or turn a formatting failure into a business retry.
                 # Preserve prose, but admit none of the unvalidated actions.
-                if not isinstance(payload.get("message"), str) or not payload["message"].strip():
+                answer = payload.get("message")
+                if not isinstance(answer, str) or not answer.strip():
+                    answer = raw_text[:start].split(CHAT_REVIEW_OPEN_TAG, 1)[0].strip()
+                if not answer:
                     raise
                 summary = "答复正文已保留，但操作或委派元数据未通过校验；相关操作和委派尚未提交。"
-                return normalize_agent_response({"message": payload["message"] + "\n\n待处理：" + summary,
+                return normalize_agent_response({"message": answer + "\n\n待处理：" + summary,
                     "gate": {"kind": "response_metadata_invalid", "summary": summary}},
                     protected_paths=protected)
         key = re.search(r'"message"\s*:\s*', body)

tests/test_chat_ordinary_project.py

@@ -65,6 +65,41 @@ def request(path, body=None):
     thread.join(timeout=2)
 
 
+def test_http_keeps_native_answer_when_legacy_handoff_metadata_is_invalid(ordinary):
+    store, runtime, contexts, request, capture, fake, workspace = ordinary
+    # Use the real protocol process and HTTP queue, including its completed
+    # upstream Turn; only the provider's final response is a fixture.
+    source = fake.read_text()
+    marker = '            "message": "Runtime response.",'
+    assert marker in source
+    fake.write_text(source.replace(marker,
+        '            "context_handoff": [{"goal_id": "existing-work", '
+        '"agent_id": "existing-worker", "brief": "invalid legacy brief"}],', 1))
+    project = request("/api/chat/projects")[1]["projects"][0]
+    status, opened = request("/api/chat/sessions", {
+        "context_kind": "project", "project_ref": project["project_ref"]})
+    assert status == 201
+    sid = opened["session_id"]
+    try:
+        status, accepted = request(f"/api/chat/sessions/{sid}/turns", {
+            "message": "Recover the original answer", "client_turn_id": "legacy-answer"})
+        assert status == 202
+        ended = runtime.wait_for_turn(session_id=sid, turn_id=accepted["turn_id"], timeout_sec=10)
+        assert ended["status"] == "completed", ended
+        assert ended["response"]["gate"]["kind"] == "response_metadata_invalid"
+        assert not ended["response"].get("context_handoff")
+        snapshot = request(f"/api/chat/sessions/{sid}")[1]
+        answer = next(m["text"] for m in snapshot["messages"] if m["role"] == "agent")
+        assert answer.startswith("Visible runtime response.\n\n待处理：")
+        assert "相关操作和委派尚未提交" in answer
+        assert "invalid legacy brief" not in answer
+        calls = [json.loads(line) for line in capture.read_text().splitlines()]
+        assert sum(c.get("method") == "turn/start" for c in calls) == 1
+        assert store.load_session(sid).get("active_turn_id") is None
+    finally:
+        runtime.close()
+
+
 def test_http_ordinary_project_continues_native_session_without_goal_or_portfolio(ordinary):
     store, runtime, contexts, request, capture, fake, workspace = ordinary
     status, projects = request("/api/chat/projects")

tests/test_chat_response_parsing.py

@@ -24,6 +24,41 @@ def test_invalid_action_metadata_preserves_answer_without_admitting_actions(meta
     assert response["proposals"] == [] and response["protected_action"] is None
 
 
+@pytest.mark.parametrize("message_field", [{}, {"message": None}, {"message": " "}])
+def test_invalid_legacy_handoff_keeps_visible_answer_when_envelope_has_no_answer(message_field):
+    visible = "已核对 [报告](/custom-volume/project/report.md)。结论已保留。"
+    payload = {**message_field, "GlobalSummary": "hidden metadata", "context_handoff": [{
+        "goal_id": "existing-work", "agent_id": "existing-worker", "brief": "invalid brief",
+    }], "proposals": [{"kind": "todo", "text": "must not run"}]}
+    response = parse_agent_response(visible + "\n" + CHAT_REVIEW_OPEN_TAG
+        + json.dumps(payload) + CHAT_REVIEW_CLOSE_TAG, protected_paths=["/custom-volume/project"])
+
+    assert response["message"].startswith("已核对 报告。结论已保留。\n\n待处理：")
+    assert response["gate"]["kind"] == "response_metadata_invalid"
+    assert not response.get("context_handoff") and not response.get("goal_draft")
+    assert response["proposals"] == [] and response["protected_action"] is None
+    assert "hidden metadata" not in response["message"]
+    assert "invalid brief" not in response["message"]
+
+
+def test_invalid_handoff_without_any_answer_still_fails_closed():
+    raw = CHAT_REVIEW_OPEN_TAG + json.dumps({"context_handoff": [{
+        "goal_id": "existing-work", "agent_id": "existing-worker", "brief": "invalid brief",
+    }]}) + CHAT_REVIEW_CLOSE_TAG
+    with pytest.raises(ValueError):
+        parse_agent_response(raw)
+
+
+def test_invalid_handoff_never_recovers_an_earlier_hidden_envelope_as_prose():
+    raw = 'Answer.\n<loopx-review-json>{"message":"hidden"}</loopx-review-json>\n'
+    raw += CHAT_REVIEW_OPEN_TAG + json.dumps({"context_handoff": [{
+        "goal_id": "existing-work", "agent_id": "existing-worker", "brief": "invalid brief",
+    }]}) + CHAT_REVIEW_CLOSE_TAG
+    response = parse_agent_response(raw)
+    assert response["message"].startswith("Answer.\n\n待处理：")
+    assert "hidden" not in response["message"] and "loopx-review-json" not in response["message"]
+
+
 @pytest.mark.parametrize("change", [None, {"effects_settled": False}, {"effects_settled": "true"},
     {"remaining": ""}, {"remaining": "x" * 1001}, {"dependency": 4}, {"grant": "workspace_write"}])
 def test_unfinished_work_is_explicit_bounded_data_never_inferred_from_prose(change):